From 3a73624b7e7ab165952f492bbbcaebc7af7778a0 Mon Sep 17 00:00:00 2001 From: nillerusr Date: Sun, 15 May 2022 21:09:59 +0300 Subject: [PATCH] misaligment fixes --- bitmap/colorconversion.cpp | 5 +- common/sse2neon.h | 13043 +++++++++++--------- engine/OcclusionSystem.cpp | 6 +- engine/gl_rsurf.cpp | 4 +- game/client/c_rumble.cpp | 4 +- game/client/c_vote_controller.cpp | 4 +- game/client/client_hl1mp.vpc | 109 + game/client/hud_vote.h | 3 + game/client/viewrender.cpp | 2 +- game/client/wscript | 2 +- game/server/AI_Criteria.h | 4 +- game/server/AI_ResponseSystem.cpp | 2 - game/server/server_hl1mp.vpc | 179 + game/server/wscript | 2 +- game/shared/saverestore.cpp | 3 +- gameui/BaseSaveGameDialog.cpp | 17 +- public/bone_setup.cpp | 14 +- public/dt_send.cpp | 2 +- public/mathlib/compressed_vector.h | 4 +- public/mathlib/lightdesc.h | 3 +- public/mathlib/vector4d.h | 16 +- public/mathlib/vmatrix.h | 6 + public/studio.h | 4 +- public/togles/linuxwin/dxabstract_types.h | 4 +- public/togles/linuxwin/glentrypoints.h | 96 +- serverbrowser/ServerBrowserDialog.cpp | 7 +- studiorender/studiorendercontext.cpp | 4 +- tier0/cpu.cpp | 28 +- tier1/checksum_crc.cpp | 12 +- tier1/processor_detect_linux.cpp | 6 +- tier1/snappy-stubs-internal.h | 3 +- togles/linuxwin/cglmbuffer.cpp | 5 +- togles/linuxwin/cglmprogram.cpp | 5 +- togles/linuxwin/cglmtex.cpp | 4 +- togles/linuxwin/decompress.o | Bin 35548 -> 0 bytes togles/linuxwin/dx9asmtogl2.cpp | 21 +- togles/linuxwin/glentrypoints.cpp | 6 +- togles/linuxwin/glmgr.cpp | 2 +- vphysics/physics_collide.cpp | 7 +- 39 files changed, 7592 insertions(+), 6056 deletions(-) create mode 100644 game/client/client_hl1mp.vpc create mode 100644 game/server/server_hl1mp.vpc delete mode 100644 togles/linuxwin/decompress.o diff --git a/bitmap/colorconversion.cpp b/bitmap/colorconversion.cpp index ac033fd2..9b9432fe 100644 --- a/bitmap/colorconversion.cpp +++ b/bitmap/colorconversion.cpp @@ -435,7 +435,8 @@ static inline void DecodeAlpha3BitLinear( CDestPixel *pImPos, DXTAlphaBlock3BitL // pRows = (Alpha3BitRows*) & ( pAlphaBlock->stuff[0] ); const DWORD mask = 0x00000007; // bits = 00 00 01 11 - DWORD bits = *( (DWORD*) & ( pAlphaBlock->stuff[0] )); + DWORD bits; + memcpy( &bits, &(pAlphaBlock->stuff[0]), sizeof(DWORD) ); gBits[0][0] = (BYTE)( bits & mask ); bits >>= 3; @@ -454,7 +455,7 @@ static inline void DecodeAlpha3BitLinear( CDestPixel *pImPos, DXTAlphaBlock3BitL gBits[1][3] = (BYTE)( bits & mask ); // now for last two rows: - bits = *( (DWORD*) & ( pAlphaBlock->stuff[3] )); // last 3 bytes + memcpy( &bits, &(pAlphaBlock->stuff[3]), sizeof(DWORD) ); gBits[2][0] = (BYTE)( bits & mask ); bits >>= 3; diff --git a/common/sse2neon.h b/common/sse2neon.h index 9e512acf..490c0a45 100644 --- a/common/sse2neon.h +++ b/common/sse2neon.h @@ -52,9 +52,9 @@ /* Enable precise implementation of math operations * This would slow down the computation a bit, but gives consistent result with - * x86 SSE2. (e.g. would solve a hole or NaN pixel in the rendering result) + * x86 SSE. (e.g. would solve a hole or NaN pixel in the rendering result) */ -/* _mm_min_ps and _mm_max_ps */ +/* _mm_min|max_ps|ss|pd|sd */ #ifndef SSE2NEON_PRECISE_MINMAX #define SSE2NEON_PRECISE_MINMAX (0) #endif @@ -66,33 +66,36 @@ #ifndef SSE2NEON_PRECISE_SQRT #define SSE2NEON_PRECISE_SQRT (0) #endif +/* _mm_dp_pd */ +#ifndef SSE2NEON_PRECISE_DP +#define SSE2NEON_PRECISE_DP (0) +#endif +/* compiler specific definitions */ #if defined(__GNUC__) || defined(__clang__) #pragma push_macro("FORCE_INLINE") #pragma push_macro("ALIGN_STRUCT") #define FORCE_INLINE static inline __attribute__((always_inline)) #define ALIGN_STRUCT(x) __attribute__((aligned(x))) -#ifndef likely -#define likely(x) __builtin_expect(!!(x), 1) -#endif -#ifndef unlikely -#define unlikely(x) __builtin_expect(!!(x), 0) -#endif -#else -#error "Macro name collisions may happen with unsupported compiler." -#ifdef FORCE_INLINE -#undef FORCE_INLINE -#endif +#define _sse2neon_likely(x) __builtin_expect(!!(x), 1) +#define _sse2neon_unlikely(x) __builtin_expect(!!(x), 0) +#else /* non-GNU / non-clang compilers */ +#warning "Macro name collisions may happen with unsupported compiler." +#ifndef FORCE_INLINE #define FORCE_INLINE static inline +#endif #ifndef ALIGN_STRUCT #define ALIGN_STRUCT(x) __declspec(align(x)) #endif +#define _sse2neon_likely(x) (x) +#define _sse2neon_unlikely(x) (x) #endif -#ifndef likely -#define likely(x) (x) -#endif -#ifndef unlikely -#define unlikely(x) (x) + +/* C language does not allow initializing a variable with a function call. */ +#ifdef __cplusplus +#define _sse2neon_const static const +#else +#define _sse2neon_const const #endif #include @@ -118,12 +121,25 @@ #pragma GCC push_options #pragma GCC target("+simd") #endif +#elif __ARM_ARCH == 8 +#if !defined(__ARM_NEON) || !defined(__ARM_NEON__) +#error \ + "You must enable NEON instructions (e.g. -mfpu=neon-fp-armv8) to use SSE2NEON." +#endif +#if !defined(__clang__) +#pragma GCC push_options +#endif #else #error "Unsupported target. Must be either ARMv7-A+NEON or ARMv8-A." #endif #endif #include +#if !defined(__aarch64__) && (__ARM_ARCH == 8) +#if defined __has_include && __has_include() +#include +#endif +#endif /* Rounding functions require either Aarch64 instructions or libm failback */ #if !defined(__aarch64__) @@ -135,7 +151,7 @@ */ #ifndef __has_builtin /* GCC prior to 10 or non-clang compilers */ /* Compatibility with gcc <= 9 */ -#if __GNUC__ <= 9 +#if defined(__GNUC__) && (__GNUC__ <= 9) #define __has_builtin(x) HAS##x #define HAS__builtin_popcount 1 #define HAS__builtin_popcountll 1 @@ -162,10 +178,25 @@ #define _MM_FROUND_TO_ZERO 0x03 #define _MM_FROUND_CUR_DIRECTION 0x04 #define _MM_FROUND_NO_EXC 0x08 +#define _MM_FROUND_RAISE_EXC 0x00 +#define _MM_FROUND_NINT (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_FLOOR (_MM_FROUND_TO_NEG_INF | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_CEIL (_MM_FROUND_TO_POS_INF | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_TRUNC (_MM_FROUND_TO_ZERO | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_RINT (_MM_FROUND_CUR_DIRECTION | _MM_FROUND_RAISE_EXC) +#define _MM_FROUND_NEARBYINT (_MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC) #define _MM_ROUND_NEAREST 0x0000 #define _MM_ROUND_DOWN 0x2000 #define _MM_ROUND_UP 0x4000 #define _MM_ROUND_TOWARD_ZERO 0x6000 +/* Flush zero mode macros. */ +#define _MM_FLUSH_ZERO_MASK 0x8000 +#define _MM_FLUSH_ZERO_ON 0x8000 +#define _MM_FLUSH_ZERO_OFF 0x0000 +/* Denormals are zeros mode macros. */ +#define _MM_DENORMALS_ZERO_MASK 0x0040 +#define _MM_DENORMALS_ZERO_ON 0x0040 +#define _MM_DENORMALS_ZERO_OFF 0x0000 /* indicate immediate constant argument in a given range */ #define __constrange(a, b) const @@ -188,6 +219,16 @@ typedef float32x4_t __m128d; #endif typedef int64x2_t __m128i; /* 128-bit vector containing integers */ +// __int64 is defined in the Intrinsics Guide which maps to different datatype +// in different data model +#if !(defined(_WIN32) || defined(_WIN64) || defined(__int64)) +#if (defined(__x86_64__) || defined(__i386__)) +#define __int64 long long +#else +#define __int64 int64_t +#endif +#endif + /* type-safe casting between types */ #define vreinterpretq_m128_f16(x) vreinterpretq_f32_f16(x) @@ -301,10 +342,10 @@ typedef int64x2_t __m128i; /* 128-bit vector containing integers */ #endif // A struct is defined in this header file called 'SIMDVec' which can be used -// by applications which attempt to access the contents of an _m128 struct +// by applications which attempt to access the contents of an __m128 struct // directly. It is important to note that accessing the __m128 struct directly // is bad coding practice by Microsoft: @see: -// https://msdn.microsoft.com/en-us/library/ayeb3ayc.aspx +// https://docs.microsoft.com/en-us/cpp/cpp/m128 // // However, some legacy source code may try to access the contents of an __m128 // struct directly so the developer can use the SIMDVec as an alias for it. Any @@ -340,13 +381,48 @@ typedef union ALIGN_STRUCT(16) SIMDVec { #define vreinterpretq_nth_u32_m128i(x, n) (((SIMDVec *) &x)->m128_u32[n]) #define vreinterpretq_nth_u8_m128i(x, n) (((SIMDVec *) &x)->m128_u8[n]) +/* SSE macros */ +#define _MM_GET_FLUSH_ZERO_MODE _sse2neon_mm_get_flush_zero_mode +#define _MM_SET_FLUSH_ZERO_MODE _sse2neon_mm_set_flush_zero_mode +#define _MM_GET_DENORMALS_ZERO_MODE _sse2neon_mm_get_denormals_zero_mode +#define _MM_SET_DENORMALS_ZERO_MODE _sse2neon_mm_set_denormals_zero_mode + +// Function declaration +// SSE +FORCE_INLINE unsigned int _MM_GET_ROUNDING_MODE(); +FORCE_INLINE __m128 _mm_move_ss(__m128, __m128); +FORCE_INLINE __m128 _mm_or_ps(__m128, __m128); +FORCE_INLINE __m128 _mm_set_ps1(float); +FORCE_INLINE __m128 _mm_setzero_ps(void); +// SSE2 +FORCE_INLINE __m128i _mm_and_si128(__m128i, __m128i); +FORCE_INLINE __m128i _mm_castps_si128(__m128); +FORCE_INLINE __m128i _mm_cmpeq_epi32(__m128i, __m128i); +FORCE_INLINE __m128i _mm_cvtps_epi32(__m128); +FORCE_INLINE __m128d _mm_move_sd(__m128d, __m128d); +FORCE_INLINE __m128i _mm_or_si128(__m128i, __m128i); +FORCE_INLINE __m128i _mm_set_epi32(int, int, int, int); +FORCE_INLINE __m128i _mm_set_epi64x(int64_t, int64_t); +FORCE_INLINE __m128d _mm_set_pd(double, double); +FORCE_INLINE __m128i _mm_set1_epi32(int); +FORCE_INLINE __m128i _mm_setzero_si128(); +// SSE4.1 +FORCE_INLINE __m128d _mm_ceil_pd(__m128d); +FORCE_INLINE __m128 _mm_ceil_ps(__m128); +FORCE_INLINE __m128d _mm_floor_pd(__m128d); +FORCE_INLINE __m128 _mm_floor_ps(__m128); +FORCE_INLINE __m128d _mm_round_pd(__m128d, int); +FORCE_INLINE __m128 _mm_round_ps(__m128, int); +// SSE4.2 +FORCE_INLINE uint32_t _mm_crc32_u8(uint32_t, uint8_t); + /* Backwards compatibility for compilers with lack of specific type support */ // Older gcc does not define vld1q_u8_x4 type -#if defined(__GNUC__) && !defined(__clang__) && \ - ((__GNUC__ == 10 && (__GNUC_MINOR__ <= 1)) || \ - (__GNUC__ == 9 && (__GNUC_MINOR__ <= 3)) || \ - (__GNUC__ == 8 && (__GNUC_MINOR__ <= 4)) || __GNUC__ <= 7) +#if defined(__GNUC__) && !defined(__clang__) && \ + ((__GNUC__ <= 10 && defined(__arm__)) || \ + (__GNUC__ == 10 && __GNUC_MINOR__ < 3 && defined(__aarch64__)) || \ + (__GNUC__ <= 9 && defined(__aarch64__))) FORCE_INLINE uint8x16x4_t _sse2neon_vld1q_u8_x4(const uint8_t *p) { uint8x16x4_t ret; @@ -441,8 +517,6 @@ FORCE_INLINE uint8x16x4_t _sse2neon_vld1q_u8_x4(const uint8_t *p) +------+------+------+------+------+------+-------------+ */ -/* Set/get methods */ - /* Constants for use with _mm_prefetch. */ enum _mm_hint { _MM_HINT_NTA = 0, /* load data to L1 and L2 cache, mark it as NTA */ @@ -455,1323 +529,18 @@ enum _mm_hint { _MM_HINT_ET2 = 7 /* exclusive version of _MM_HINT_T2 */ }; -// Loads one cache line of data from address p to a location closer to the -// processor. https://msdn.microsoft.com/en-us/library/84szxsww(v=vs.100).aspx -FORCE_INLINE void _mm_prefetch(const void *p, int i) -{ - (void) i; - __builtin_prefetch(p); -} - -// Pause the processor. This is typically used in spin-wait loops and depending -// on the x86 processor typical values are in the 40-100 cycle range. The -// 'yield' instruction isn't a good fit beacuse it's effectively a nop on most -// Arm cores. Experience with several databases has shown has shown an 'isb' is -// a reasonable approximation. -FORCE_INLINE void _mm_pause() -{ - __asm__ __volatile__("isb\n"); -} - -// Copy the lower single-precision (32-bit) floating-point element of a to dst. -// -// dst[31:0] := a[31:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_f32 -FORCE_INLINE float _mm_cvtss_f32(__m128 a) -{ - return vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); -} - -// Convert the lower single-precision (32-bit) floating-point element in b to a -// double-precision (64-bit) floating-point element, store the result in the -// lower element of dst, and copy the upper element from a to the upper element -// of dst. -// -// dst[63:0] := Convert_FP32_To_FP64(b[31:0]) -// dst[127:64] := a[127:64] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_sd -FORCE_INLINE __m128d _mm_cvtss_sd(__m128d a, __m128 b) -{ - double d = (double) vgetq_lane_f32(vreinterpretq_f32_m128(b), 0); +// The bit field mapping to the FPCR(floating-point control register) +typedef struct { + uint16_t res0; + uint8_t res1 : 6; + uint8_t bit22 : 1; + uint8_t bit23 : 1; + uint8_t bit24 : 1; + uint8_t res2 : 7; #if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vsetq_lane_f64(d, vreinterpretq_f64_m128d(a), 0)); -#else - return vreinterpretq_m128d_s64( - vsetq_lane_s64(*(int64_t *) &d, vreinterpretq_s64_m128d(a), 0)); + uint32_t res3; #endif -} - -// Convert the lower single-precision (32-bit) floating-point element in a to a -// 32-bit integer, and store the result in dst. -// -// dst[31:0] := Convert_FP32_To_Int32(a[31:0]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_si32 -#define _mm_cvtss_si32(a) _mm_cvt_ss2si(a) - -// Convert the lower single-precision (32-bit) floating-point element in a to a -// 64-bit integer, and store the result in dst. -// -// dst[63:0] := Convert_FP32_To_Int64(a[31:0]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_si64 -FORCE_INLINE int _mm_cvtss_si64(__m128 a) -{ -#if defined(__aarch64__) - return vgetq_lane_s64( - vreinterpretq_s64_s32(vcvtnq_s32_f32(vreinterpretq_f32_m128(a))), 0); -#else - float32_t data = vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); - float32_t diff = data - floor(data); - if (diff > 0.5) - return (int64_t) ceil(data); - if (unlikely(diff == 0.5)) { - int64_t f = (int64_t) floor(data); - int64_t c = (int64_t) ceil(data); - return c & 1 ? f : c; - } - return (int64_t) floor(data); -#endif -} - -// Convert packed single-precision (32-bit) floating-point elements in a to -// packed 32-bit integers with truncation, and store the results in dst. -// -// FOR j := 0 to 1 -// i := 32*j -// dst[i+31:i] := Convert_FP32_To_Int32_Truncate(a[i+31:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtt_ps2pi -FORCE_INLINE __m64 _mm_cvtt_ps2pi(__m128 a) -{ - return vreinterpret_m64_s32( - vget_low_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a)))); -} - -// Convert the lower single-precision (32-bit) floating-point element in a to a -// 32-bit integer with truncation, and store the result in dst. -// -// dst[31:0] := Convert_FP32_To_Int32_Truncate(a[31:0]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtt_ss2si -FORCE_INLINE int _mm_cvtt_ss2si(__m128 a) -{ - return vgetq_lane_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a)), 0); -} - -// Convert packed single-precision (32-bit) floating-point elements in a to -// packed 32-bit integers with truncation, and store the results in dst. -// -// FOR j := 0 to 1 -// i := 32*j -// dst[i+31:i] := Convert_FP32_To_Int32_Truncate(a[i+31:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttps_pi32 -#define _mm_cvttps_pi32(a) _mm_cvtt_ps2pi(a) - -// Convert the lower single-precision (32-bit) floating-point element in a to a -// 32-bit integer with truncation, and store the result in dst. -// -// dst[31:0] := Convert_FP32_To_Int32_Truncate(a[31:0]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttss_si32 -#define _mm_cvttss_si32(a) _mm_cvtt_ss2si(a) - -// Convert the lower single-precision (32-bit) floating-point element in a to a -// 64-bit integer with truncation, and store the result in dst. -// -// dst[63:0] := Convert_FP32_To_Int64_Truncate(a[31:0]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttss_si64 -FORCE_INLINE int64_t _mm_cvttss_si64(__m128 a) -{ - return vgetq_lane_s64( - vmovl_s32(vget_low_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a)))), 0); -} - -// Sets the 128-bit value to zero -// https://msdn.microsoft.com/en-us/library/vstudio/ys7dw0kh(v=vs.100).aspx -FORCE_INLINE __m128i _mm_setzero_si128(void) -{ - return vreinterpretq_m128i_s32(vdupq_n_s32(0)); -} - -// Clears the four single-precision, floating-point values. -// https://msdn.microsoft.com/en-us/library/vstudio/tk1t2tbz(v=vs.100).aspx -FORCE_INLINE __m128 _mm_setzero_ps(void) -{ - return vreinterpretq_m128_f32(vdupq_n_f32(0)); -} - -// Return vector of type __m128d with all elements set to zero. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_setzero_pd -FORCE_INLINE __m128d _mm_setzero_pd(void) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64(vdupq_n_f64(0)); -#else - return vreinterpretq_m128d_f32(vdupq_n_f32(0)); -#endif -} - -// Sets the four single-precision, floating-point values to w. -// -// r0 := r1 := r2 := r3 := w -// -// https://msdn.microsoft.com/en-us/library/vstudio/2x1se8ha(v=vs.100).aspx -FORCE_INLINE __m128 _mm_set1_ps(float _w) -{ - return vreinterpretq_m128_f32(vdupq_n_f32(_w)); -} - -// Sets the four single-precision, floating-point values to w. -// https://msdn.microsoft.com/en-us/library/vstudio/2x1se8ha(v=vs.100).aspx -FORCE_INLINE __m128 _mm_set_ps1(float _w) -{ - return vreinterpretq_m128_f32(vdupq_n_f32(_w)); -} - -// Sets the four single-precision, floating-point values to the four inputs. -// https://msdn.microsoft.com/en-us/library/vstudio/afh0zf75(v=vs.100).aspx -FORCE_INLINE __m128 _mm_set_ps(float w, float z, float y, float x) -{ - float ALIGN_STRUCT(16) data[4] = {x, y, z, w}; - return vreinterpretq_m128_f32(vld1q_f32(data)); -} - -// Copy single-precision (32-bit) floating-point element a to the lower element -// of dst, and zero the upper 3 elements. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_ss -FORCE_INLINE __m128 _mm_set_ss(float a) -{ - float ALIGN_STRUCT(16) data[4] = {a, 0, 0, 0}; - return vreinterpretq_m128_f32(vld1q_f32(data)); -} - -// Sets the four single-precision, floating-point values to the four inputs in -// reverse order. -// https://msdn.microsoft.com/en-us/library/vstudio/d2172ct3(v=vs.100).aspx -FORCE_INLINE __m128 _mm_setr_ps(float w, float z, float y, float x) -{ - float ALIGN_STRUCT(16) data[4] = {w, z, y, x}; - return vreinterpretq_m128_f32(vld1q_f32(data)); -} - -// Sets the 8 signed 16-bit integer values in reverse order. -// -// Return Value -// r0 := w0 -// r1 := w1 -// ... -// r7 := w7 -FORCE_INLINE __m128i _mm_setr_epi16(short w0, - short w1, - short w2, - short w3, - short w4, - short w5, - short w6, - short w7) -{ - int16_t ALIGN_STRUCT(16) data[8] = {w0, w1, w2, w3, w4, w5, w6, w7}; - return vreinterpretq_m128i_s16(vld1q_s16((int16_t *) data)); -} - -// Sets the 4 signed 32-bit integer values in reverse order -// https://technet.microsoft.com/en-us/library/security/27yb3ee5(v=vs.90).aspx -FORCE_INLINE __m128i _mm_setr_epi32(int i3, int i2, int i1, int i0) -{ - int32_t ALIGN_STRUCT(16) data[4] = {i3, i2, i1, i0}; - return vreinterpretq_m128i_s32(vld1q_s32(data)); -} - -// Set packed 64-bit integers in dst with the supplied values in reverse order. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_setr_epi64 -FORCE_INLINE __m128i _mm_setr_epi64(__m64 e1, __m64 e0) -{ - return vreinterpretq_m128i_s64(vcombine_s64(e1, e0)); -} - -// Sets the 16 signed 8-bit integer values to b. -// -// r0 := b -// r1 := b -// ... -// r15 := b -// -// https://msdn.microsoft.com/en-us/library/6e14xhyf(v=vs.100).aspx -FORCE_INLINE __m128i _mm_set1_epi8(signed char w) -{ - return vreinterpretq_m128i_s8(vdupq_n_s8(w)); -} - -// Broadcast double-precision (64-bit) floating-point value a to all elements of -// dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set1_pd -FORCE_INLINE __m128d _mm_set1_pd(double d) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64(vdupq_n_f64(d)); -#else - return vreinterpretq_m128d_s64(vdupq_n_s64(*(int64_t *) &d)); -#endif -} - -// Sets the 8 signed 16-bit integer values to w. -// -// r0 := w -// r1 := w -// ... -// r7 := w -// -// https://msdn.microsoft.com/en-us/library/k0ya3x0e(v=vs.90).aspx -FORCE_INLINE __m128i _mm_set1_epi16(short w) -{ - return vreinterpretq_m128i_s16(vdupq_n_s16(w)); -} - -// Sets the 16 signed 8-bit integer values. -// https://msdn.microsoft.com/en-us/library/x0cx8zd3(v=vs.90).aspx -FORCE_INLINE __m128i _mm_set_epi8(signed char b15, - signed char b14, - signed char b13, - signed char b12, - signed char b11, - signed char b10, - signed char b9, - signed char b8, - signed char b7, - signed char b6, - signed char b5, - signed char b4, - signed char b3, - signed char b2, - signed char b1, - signed char b0) -{ - int8_t ALIGN_STRUCT(16) - data[16] = {(int8_t) b0, (int8_t) b1, (int8_t) b2, (int8_t) b3, - (int8_t) b4, (int8_t) b5, (int8_t) b6, (int8_t) b7, - (int8_t) b8, (int8_t) b9, (int8_t) b10, (int8_t) b11, - (int8_t) b12, (int8_t) b13, (int8_t) b14, (int8_t) b15}; - return (__m128i) vld1q_s8(data); -} - -// Sets the 8 signed 16-bit integer values. -// https://msdn.microsoft.com/en-au/library/3e0fek84(v=vs.90).aspx -FORCE_INLINE __m128i _mm_set_epi16(short i7, - short i6, - short i5, - short i4, - short i3, - short i2, - short i1, - short i0) -{ - int16_t ALIGN_STRUCT(16) data[8] = {i0, i1, i2, i3, i4, i5, i6, i7}; - return vreinterpretq_m128i_s16(vld1q_s16(data)); -} - -// Sets the 16 signed 8-bit integer values in reverse order. -// https://msdn.microsoft.com/en-us/library/2khb9c7k(v=vs.90).aspx -FORCE_INLINE __m128i _mm_setr_epi8(signed char b0, - signed char b1, - signed char b2, - signed char b3, - signed char b4, - signed char b5, - signed char b6, - signed char b7, - signed char b8, - signed char b9, - signed char b10, - signed char b11, - signed char b12, - signed char b13, - signed char b14, - signed char b15) -{ - int8_t ALIGN_STRUCT(16) - data[16] = {(int8_t) b0, (int8_t) b1, (int8_t) b2, (int8_t) b3, - (int8_t) b4, (int8_t) b5, (int8_t) b6, (int8_t) b7, - (int8_t) b8, (int8_t) b9, (int8_t) b10, (int8_t) b11, - (int8_t) b12, (int8_t) b13, (int8_t) b14, (int8_t) b15}; - return (__m128i) vld1q_s8(data); -} - -// Sets the 4 signed 32-bit integer values to i. -// -// r0 := i -// r1 := i -// r2 := i -// r3 := I -// -// https://msdn.microsoft.com/en-us/library/vstudio/h4xscxat(v=vs.100).aspx -FORCE_INLINE __m128i _mm_set1_epi32(int _i) -{ - return vreinterpretq_m128i_s32(vdupq_n_s32(_i)); -} - -// Sets the 2 signed 64-bit integer values to i. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/whtfzhzk(v=vs.100) -FORCE_INLINE __m128i _mm_set1_epi64(__m64 _i) -{ - return vreinterpretq_m128i_s64(vdupq_n_s64((int64_t) _i)); -} - -// Sets the 2 signed 64-bit integer values to i. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set1_epi64x -FORCE_INLINE __m128i _mm_set1_epi64x(int64_t _i) -{ - return vreinterpretq_m128i_s64(vdupq_n_s64(_i)); -} - -// Sets the 4 signed 32-bit integer values. -// https://msdn.microsoft.com/en-us/library/vstudio/019beekt(v=vs.100).aspx -FORCE_INLINE __m128i _mm_set_epi32(int i3, int i2, int i1, int i0) -{ - int32_t ALIGN_STRUCT(16) data[4] = {i0, i1, i2, i3}; - return vreinterpretq_m128i_s32(vld1q_s32(data)); -} - -// Returns the __m128i structure with its two 64-bit integer values -// initialized to the values of the two 64-bit integers passed in. -// https://msdn.microsoft.com/en-us/library/dk2sdw0h(v=vs.120).aspx -FORCE_INLINE __m128i _mm_set_epi64x(int64_t i1, int64_t i2) -{ - return vreinterpretq_m128i_s64( - vcombine_s64(vcreate_s64(i2), vcreate_s64(i1))); -} - -// Returns the __m128i structure with its two 64-bit integer values -// initialized to the values of the two 64-bit integers passed in. -// https://msdn.microsoft.com/en-us/library/dk2sdw0h(v=vs.120).aspx -FORCE_INLINE __m128i _mm_set_epi64(__m64 i1, __m64 i2) -{ - return _mm_set_epi64x((int64_t) i1, (int64_t) i2); -} - -// Set packed double-precision (64-bit) floating-point elements in dst with the -// supplied values. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_pd -FORCE_INLINE __m128d _mm_set_pd(double e1, double e0) -{ - double ALIGN_STRUCT(16) data[2] = {e0, e1}; -#if defined(__aarch64__) - return vreinterpretq_m128d_f64(vld1q_f64((float64_t *) data)); -#else - return vreinterpretq_m128d_f32(vld1q_f32((float32_t *) data)); -#endif -} - -// Set packed double-precision (64-bit) floating-point elements in dst with the -// supplied values in reverse order. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_setr_pd -FORCE_INLINE __m128d _mm_setr_pd(double e1, double e0) -{ - return _mm_set_pd(e0, e1); -} - -// Copy double-precision (64-bit) floating-point element a to the lower element -// of dst, and zero the upper element. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_sd -FORCE_INLINE __m128d _mm_set_sd(double a) -{ - return _mm_set_pd(0, a); -} - -// Broadcast double-precision (64-bit) floating-point value a to all elements of -// dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_pd1 -#define _mm_set_pd1 _mm_set1_pd - -// Stores four single-precision, floating-point values. -// https://msdn.microsoft.com/en-us/library/vstudio/s3h4ay6y(v=vs.100).aspx -FORCE_INLINE void _mm_store_ps(float *p, __m128 a) -{ - vst1q_f32(p, vreinterpretq_f32_m128(a)); -} - -// Store the lower single-precision (32-bit) floating-point element from a into -// 4 contiguous elements in memory. mem_addr must be aligned on a 16-byte -// boundary or a general-protection exception may be generated. -// -// MEM[mem_addr+31:mem_addr] := a[31:0] -// MEM[mem_addr+63:mem_addr+32] := a[31:0] -// MEM[mem_addr+95:mem_addr+64] := a[31:0] -// MEM[mem_addr+127:mem_addr+96] := a[31:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store_ps1 -FORCE_INLINE void _mm_store_ps1(float *p, __m128 a) -{ - float32_t a0 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); - vst1q_f32(p, vdupq_n_f32(a0)); -} - -// Store the lower single-precision (32-bit) floating-point element from a into -// 4 contiguous elements in memory. mem_addr must be aligned on a 16-byte -// boundary or a general-protection exception may be generated. -// -// MEM[mem_addr+31:mem_addr] := a[31:0] -// MEM[mem_addr+63:mem_addr+32] := a[31:0] -// MEM[mem_addr+95:mem_addr+64] := a[31:0] -// MEM[mem_addr+127:mem_addr+96] := a[31:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store1_ps -#define _mm_store1_ps _mm_store_ps1 - -// Store 4 single-precision (32-bit) floating-point elements from a into memory -// in reverse order. mem_addr must be aligned on a 16-byte boundary or a -// general-protection exception may be generated. -// -// MEM[mem_addr+31:mem_addr] := a[127:96] -// MEM[mem_addr+63:mem_addr+32] := a[95:64] -// MEM[mem_addr+95:mem_addr+64] := a[63:32] -// MEM[mem_addr+127:mem_addr+96] := a[31:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storer_ps -FORCE_INLINE void _mm_storer_ps(float *p, __m128 a) -{ - float32x4_t tmp = vrev64q_f32(vreinterpretq_f32_m128(a)); - float32x4_t rev = vextq_f32(tmp, tmp, 2); - vst1q_f32(p, rev); -} - -// Stores four single-precision, floating-point values. -// https://msdn.microsoft.com/en-us/library/44e30x22(v=vs.100).aspx -FORCE_INLINE void _mm_storeu_ps(float *p, __m128 a) -{ - vst1q_f32(p, vreinterpretq_f32_m128(a)); -} - -// Stores four 32-bit integer values as (as a __m128i value) at the address p. -// https://msdn.microsoft.com/en-us/library/vstudio/edk11s13(v=vs.100).aspx -FORCE_INLINE void _mm_store_si128(__m128i *p, __m128i a) -{ - vst1q_s32((int32_t *) p, vreinterpretq_s32_m128i(a)); -} - -// Stores 128-bits of integer data a at the address p. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si128 -FORCE_INLINE void _mm_storeu_si128(__m128i *p, __m128i a) -{ - vst1q_s32((int32_t *) p, vreinterpretq_s32_m128i(a)); -} - -// Stores 64-bits of integer data a at the address p. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si64 -FORCE_INLINE void _mm_storeu_si64(void *p, __m128i a) -{ - vst1q_lane_s64((int64_t *) p, vreinterpretq_s64_m128i(a), 0); -} - -// Stores 32-bits of integer data a at the address p. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si32 -FORCE_INLINE void _mm_storeu_si32(void *p, __m128i a) -{ - vst1q_lane_s32((int32_t *) p, vreinterpretq_s32_m128i(a), 0); -} - -// Stores 16-bits of integer data a at the address p. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si16 -FORCE_INLINE void _mm_storeu_si16(void *p, __m128i a) -{ - vst1q_lane_s16((int16_t *) p, vreinterpretq_s16_m128i(a), 0); -} - -// Stores the lower single - precision, floating - point value. -// https://msdn.microsoft.com/en-us/library/tzz10fbx(v=vs.100).aspx -FORCE_INLINE void _mm_store_ss(float *p, __m128 a) -{ - vst1q_lane_f32(p, vreinterpretq_f32_m128(a), 0); -} - -// Store 128-bits (composed of 2 packed double-precision (64-bit) floating-point -// elements) from a into memory. mem_addr must be aligned on a 16-byte boundary -// or a general-protection exception may be generated. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store_pd -FORCE_INLINE void _mm_store_pd(double *mem_addr, __m128d a) -{ -#if defined(__aarch64__) - vst1q_f64((float64_t *) mem_addr, vreinterpretq_f64_m128d(a)); -#else - vst1q_f32((float32_t *) mem_addr, vreinterpretq_f32_m128d(a)); -#endif -} - -// Store the upper double-precision (64-bit) floating-point element from a into -// memory. -// -// MEM[mem_addr+63:mem_addr] := a[127:64] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeh_pd -FORCE_INLINE void _mm_storeh_pd(double *mem_addr, __m128d a) -{ -#if defined(__aarch64__) - vst1_f64((float64_t *) mem_addr, vget_high_f64(vreinterpretq_f64_m128d(a))); -#else - vst1_f32((float32_t *) mem_addr, vget_high_f32(vreinterpretq_f32_m128d(a))); -#endif -} - -// Store the lower double-precision (64-bit) floating-point element from a into -// memory. -// -// MEM[mem_addr+63:mem_addr] := a[63:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storel_pd -FORCE_INLINE void _mm_storel_pd(double *mem_addr, __m128d a) -{ -#if defined(__aarch64__) - vst1_f64((float64_t *) mem_addr, vget_low_f64(vreinterpretq_f64_m128d(a))); -#else - vst1_f32((float32_t *) mem_addr, vget_low_f32(vreinterpretq_f32_m128d(a))); -#endif -} - -// Store 2 double-precision (64-bit) floating-point elements from a into memory -// in reverse order. mem_addr must be aligned on a 16-byte boundary or a -// general-protection exception may be generated. -// -// MEM[mem_addr+63:mem_addr] := a[127:64] -// MEM[mem_addr+127:mem_addr+64] := a[63:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storer_pd -FORCE_INLINE void _mm_storer_pd(double *mem_addr, __m128d a) -{ - float32x4_t f = vreinterpretq_f32_m128d(a); - _mm_store_pd(mem_addr, vreinterpretq_m128d_f32(vextq_f32(f, f, 2))); -} - -// Store the lower double-precision (64-bit) floating-point element from a into -// 2 contiguous elements in memory. mem_addr must be aligned on a 16-byte -// boundary or a general-protection exception may be generated. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store_pd1 -FORCE_INLINE void _mm_store_pd1(double *mem_addr, __m128d a) -{ -#if defined(__aarch64__) - float64x1_t a_low = vget_low_f64(vreinterpretq_f64_m128d(a)); - vst1q_f64((float64_t *) mem_addr, - vreinterpretq_f64_m128d(vcombine_f64(a_low, a_low))); -#else - float32x2_t a_low = vget_low_f32(vreinterpretq_f32_m128d(a)); - vst1q_f32((float32_t *) mem_addr, - vreinterpretq_f32_m128d(vcombine_f32(a_low, a_low))); -#endif -} - -// Store the lower double-precision (64-bit) floating-point element from a into -// memory. mem_addr does not need to be aligned on any particular boundary. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_store_sd -FORCE_INLINE void _mm_store_sd(double *mem_addr, __m128d a) -{ -#if defined(__aarch64__) - vst1_f64((float64_t *) mem_addr, vget_low_f64(vreinterpretq_f64_m128d(a))); -#else - vst1_u64((uint64_t *) mem_addr, vget_low_u64(vreinterpretq_u64_m128d(a))); -#endif -} - -// Store the lower double-precision (64-bit) floating-point element from a into -// 2 contiguous elements in memory. mem_addr must be aligned on a 16-byte -// boundary or a general-protection exception may be generated. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=9,526,5601&text=_mm_store1_pd -#define _mm_store1_pd _mm_store_pd1 - -// Store 128-bits (composed of 2 packed double-precision (64-bit) floating-point -// elements) from a into memory. mem_addr does not need to be aligned on any -// particular boundary. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_pd -FORCE_INLINE void _mm_storeu_pd(double *mem_addr, __m128d a) -{ - _mm_store_pd(mem_addr, a); -} - -// Reads the lower 64 bits of b and stores them into the lower 64 bits of a. -// https://msdn.microsoft.com/en-us/library/hhwf428f%28v=vs.90%29.aspx -FORCE_INLINE void _mm_storel_epi64(__m128i *a, __m128i b) -{ - uint64x1_t hi = vget_high_u64(vreinterpretq_u64_m128i(*a)); - uint64x1_t lo = vget_low_u64(vreinterpretq_u64_m128i(b)); - *a = vreinterpretq_m128i_u64(vcombine_u64(lo, hi)); -} - -// Stores the lower two single-precision floating point values of a to the -// address p. -// -// *p0 := a0 -// *p1 := a1 -// -// https://msdn.microsoft.com/en-us/library/h54t98ks(v=vs.90).aspx -FORCE_INLINE void _mm_storel_pi(__m64 *p, __m128 a) -{ - *p = vreinterpret_m64_f32(vget_low_f32(a)); -} - -// Stores the upper two single-precision, floating-point values of a to the -// address p. -// -// *p0 := a2 -// *p1 := a3 -// -// https://msdn.microsoft.com/en-us/library/a7525fs8(v%3dvs.90).aspx -FORCE_INLINE void _mm_storeh_pi(__m64 *p, __m128 a) -{ - *p = vreinterpret_m64_f32(vget_high_f32(a)); -} - -// Loads a single single-precision, floating-point value, copying it into all -// four words -// https://msdn.microsoft.com/en-us/library/vstudio/5cdkf716(v=vs.100).aspx -FORCE_INLINE __m128 _mm_load1_ps(const float *p) -{ - return vreinterpretq_m128_f32(vld1q_dup_f32(p)); -} - -// Load a single-precision (32-bit) floating-point element from memory into all -// elements of dst. -// -// dst[31:0] := MEM[mem_addr+31:mem_addr] -// dst[63:32] := MEM[mem_addr+31:mem_addr] -// dst[95:64] := MEM[mem_addr+31:mem_addr] -// dst[127:96] := MEM[mem_addr+31:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_ps1 -#define _mm_load_ps1 _mm_load1_ps - -// Sets the lower two single-precision, floating-point values with 64 -// bits of data loaded from the address p; the upper two values are passed -// through from a. -// -// Return Value -// r0 := *p0 -// r1 := *p1 -// r2 := a2 -// r3 := a3 -// -// https://msdn.microsoft.com/en-us/library/s57cyak2(v=vs.100).aspx -FORCE_INLINE __m128 _mm_loadl_pi(__m128 a, __m64 const *p) -{ - return vreinterpretq_m128_f32( - vcombine_f32(vld1_f32((const float32_t *) p), vget_high_f32(a))); -} - -// Load 4 single-precision (32-bit) floating-point elements from memory into dst -// in reverse order. mem_addr must be aligned on a 16-byte boundary or a -// general-protection exception may be generated. -// -// dst[31:0] := MEM[mem_addr+127:mem_addr+96] -// dst[63:32] := MEM[mem_addr+95:mem_addr+64] -// dst[95:64] := MEM[mem_addr+63:mem_addr+32] -// dst[127:96] := MEM[mem_addr+31:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadr_ps -FORCE_INLINE __m128 _mm_loadr_ps(const float *p) -{ - float32x4_t v = vrev64q_f32(vld1q_f32(p)); - return vreinterpretq_m128_f32(vextq_f32(v, v, 2)); -} - -// Sets the upper two single-precision, floating-point values with 64 -// bits of data loaded from the address p; the lower two values are passed -// through from a. -// -// r0 := a0 -// r1 := a1 -// r2 := *p0 -// r3 := *p1 -// -// https://msdn.microsoft.com/en-us/library/w92wta0x(v%3dvs.100).aspx -FORCE_INLINE __m128 _mm_loadh_pi(__m128 a, __m64 const *p) -{ - return vreinterpretq_m128_f32( - vcombine_f32(vget_low_f32(a), vld1_f32((const float32_t *) p))); -} - -// Loads four single-precision, floating-point values. -// https://msdn.microsoft.com/en-us/library/vstudio/zzd50xxt(v=vs.100).aspx -FORCE_INLINE __m128 _mm_load_ps(const float *p) -{ - return vreinterpretq_m128_f32(vld1q_f32(p)); -} - -// Loads four single-precision, floating-point values. -// https://msdn.microsoft.com/en-us/library/x1b16s7z%28v=vs.90%29.aspx -FORCE_INLINE __m128 _mm_loadu_ps(const float *p) -{ - // for neon, alignment doesn't matter, so _mm_load_ps and _mm_loadu_ps are - // equivalent for neon - return vreinterpretq_m128_f32(vld1q_f32(p)); -} - -// Load unaligned 16-bit integer from memory into the first element of dst. -// -// dst[15:0] := MEM[mem_addr+15:mem_addr] -// dst[MAX:16] := 0 -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_si16 -FORCE_INLINE __m128i _mm_loadu_si16(const void *p) -{ - return vreinterpretq_m128i_s16( - vsetq_lane_s16(*(const int16_t *) p, vdupq_n_s16(0), 0)); -} - -// Load unaligned 64-bit integer from memory into the first element of dst. -// -// dst[63:0] := MEM[mem_addr+63:mem_addr] -// dst[MAX:64] := 0 -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_si64 -FORCE_INLINE __m128i _mm_loadu_si64(const void *p) -{ - return vreinterpretq_m128i_s64( - vcombine_s64(vld1_s64((const int64_t *) p), vdup_n_s64(0))); -} - -// Load a double-precision (64-bit) floating-point element from memory into the -// lower of dst, and zero the upper element. mem_addr does not need to be -// aligned on any particular boundary. -// -// dst[63:0] := MEM[mem_addr+63:mem_addr] -// dst[127:64] := 0 -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_sd -FORCE_INLINE __m128d _mm_load_sd(const double *p) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64(vsetq_lane_f64(*p, vdupq_n_f64(0), 0)); -#else - const float *fp = (const float *) p; - float ALIGN_STRUCT(16) data[4] = {fp[0], fp[1], 0, 0}; - return vreinterpretq_m128d_f32(vld1q_f32(data)); -#endif -} - -// Loads two double-precision from 16-byte aligned memory, floating-point -// values. -// -// dst[127:0] := MEM[mem_addr+127:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_pd -FORCE_INLINE __m128d _mm_load_pd(const double *p) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64(vld1q_f64(p)); -#else - const float *fp = (const float *) p; - float ALIGN_STRUCT(16) data[4] = {fp[0], fp[1], fp[2], fp[3]}; - return vreinterpretq_m128d_f32(vld1q_f32(data)); -#endif -} - -// Loads two double-precision from unaligned memory, floating-point values. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_pd -FORCE_INLINE __m128d _mm_loadu_pd(const double *p) -{ - return _mm_load_pd(p); -} - -// Loads an single - precision, floating - point value into the low word and -// clears the upper three words. -// https://msdn.microsoft.com/en-us/library/548bb9h4%28v=vs.90%29.aspx -FORCE_INLINE __m128 _mm_load_ss(const float *p) -{ - return vreinterpretq_m128_f32(vsetq_lane_f32(*p, vdupq_n_f32(0), 0)); -} - -// Load 64-bit integer from memory into the first element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadl_epi64 -FORCE_INLINE __m128i _mm_loadl_epi64(__m128i const *p) -{ - /* Load the lower 64 bits of the value pointed to by p into the - * lower 64 bits of the result, zeroing the upper 64 bits of the result. - */ - return vreinterpretq_m128i_s32( - vcombine_s32(vld1_s32((int32_t const *) p), vcreate_s32(0))); -} - -// Load a double-precision (64-bit) floating-point element from memory into the -// lower element of dst, and copy the upper element from a to dst. mem_addr does -// not need to be aligned on any particular boundary. -// -// dst[63:0] := MEM[mem_addr+63:mem_addr] -// dst[127:64] := a[127:64] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadl_pd -FORCE_INLINE __m128d _mm_loadl_pd(__m128d a, const double *p) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vcombine_f64(vld1_f64(p), vget_high_f64(vreinterpretq_f64_m128d(a)))); -#else - return vreinterpretq_m128d_f32( - vcombine_f32(vld1_f32((const float *) p), - vget_high_f32(vreinterpretq_f32_m128d(a)))); -#endif -} - -// Load 2 double-precision (64-bit) floating-point elements from memory into dst -// in reverse order. mem_addr must be aligned on a 16-byte boundary or a -// general-protection exception may be generated. -// -// dst[63:0] := MEM[mem_addr+127:mem_addr+64] -// dst[127:64] := MEM[mem_addr+63:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadr_pd -FORCE_INLINE __m128d _mm_loadr_pd(const double *p) -{ -#if defined(__aarch64__) - float64x2_t v = vld1q_f64(p); - return vreinterpretq_m128d_f64(vextq_f64(v, v, 1)); -#else - int64x2_t v = vld1q_s64((const int64_t *) p); - return vreinterpretq_m128d_s64(vextq_s64(v, v, 1)); -#endif -} - -// Sets the low word to the single-precision, floating-point value of b -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/35hdzazd(v=vs.100) -FORCE_INLINE __m128 _mm_move_ss(__m128 a, __m128 b) -{ - return vreinterpretq_m128_f32( - vsetq_lane_f32(vgetq_lane_f32(vreinterpretq_f32_m128(b), 0), - vreinterpretq_f32_m128(a), 0)); -} - -// Move the lower double-precision (64-bit) floating-point element from b to the -// lower element of dst, and copy the upper element from a to the upper element -// of dst. -// -// dst[63:0] := b[63:0] -// dst[127:64] := a[127:64] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_move_sd -FORCE_INLINE __m128d _mm_move_sd(__m128d a, __m128d b) -{ - return vreinterpretq_m128d_f32( - vcombine_f32(vget_low_f32(vreinterpretq_f32_m128d(b)), - vget_high_f32(vreinterpretq_f32_m128d(a)))); -} - -// Copy the lower 64-bit integer in a to the lower element of dst, and zero the -// upper element. -// -// dst[63:0] := a[63:0] -// dst[127:64] := 0 -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_move_epi64 -FORCE_INLINE __m128i _mm_move_epi64(__m128i a) -{ - return vreinterpretq_m128i_s64( - vsetq_lane_s64(0, vreinterpretq_s64_m128i(a), 1)); -} - -// Return vector of type __m128 with undefined elements. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_undefined_ps -FORCE_INLINE __m128 _mm_undefined_ps(void) -{ -#if defined(__GNUC__) || defined(__clang__) -#pragma GCC diagnostic push -#pragma GCC diagnostic ignored "-Wuninitialized" -#endif - __m128 a; - return a; -#if defined(__GNUC__) || defined(__clang__) -#pragma GCC diagnostic pop -#endif -} - -/* Logic/Binary operations */ - -// Computes the bitwise AND-NOT of the four single-precision, floating-point -// values of a and b. -// -// r0 := ~a0 & b0 -// r1 := ~a1 & b1 -// r2 := ~a2 & b2 -// r3 := ~a3 & b3 -// -// https://msdn.microsoft.com/en-us/library/vstudio/68h7wd02(v=vs.100).aspx -FORCE_INLINE __m128 _mm_andnot_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_s32( - vbicq_s32(vreinterpretq_s32_m128(b), - vreinterpretq_s32_m128(a))); // *NOTE* argument swap -} - -// Compute the bitwise NOT of packed double-precision (64-bit) floating-point -// elements in a and then AND with b, and store the results in dst. -// -// FOR j := 0 to 1 -// i := j*64 -// dst[i+63:i] := ((NOT a[i+63:i]) AND b[i+63:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_andnot_pd -FORCE_INLINE __m128d _mm_andnot_pd(__m128d a, __m128d b) -{ - // *NOTE* argument swap - return vreinterpretq_m128d_s64( - vbicq_s64(vreinterpretq_s64_m128d(b), vreinterpretq_s64_m128d(a))); -} - -// Computes the bitwise AND of the 128-bit value in b and the bitwise NOT of the -// 128-bit value in a. -// -// r := (~a) & b -// -// https://msdn.microsoft.com/en-us/library/vstudio/1beaceh8(v=vs.100).aspx -FORCE_INLINE __m128i _mm_andnot_si128(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s32( - vbicq_s32(vreinterpretq_s32_m128i(b), - vreinterpretq_s32_m128i(a))); // *NOTE* argument swap -} - -// Computes the bitwise AND of the 128-bit value in a and the 128-bit value in -// b. -// -// r := a & b -// -// https://msdn.microsoft.com/en-us/library/vstudio/6d1txsa8(v=vs.100).aspx -FORCE_INLINE __m128i _mm_and_si128(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s32( - vandq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -} - -// Computes the bitwise AND of the four single-precision, floating-point values -// of a and b. -// -// r0 := a0 & b0 -// r1 := a1 & b1 -// r2 := a2 & b2 -// r3 := a3 & b3 -// -// https://msdn.microsoft.com/en-us/library/vstudio/73ck1xc5(v=vs.100).aspx -FORCE_INLINE __m128 _mm_and_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_s32( - vandq_s32(vreinterpretq_s32_m128(a), vreinterpretq_s32_m128(b))); -} - -// Compute the bitwise AND of packed double-precision (64-bit) floating-point -// elements in a and b, and store the results in dst. -// -// FOR j := 0 to 1 -// i := j*64 -// dst[i+63:i] := a[i+63:i] AND b[i+63:i] -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_and_pd -FORCE_INLINE __m128d _mm_and_pd(__m128d a, __m128d b) -{ - return vreinterpretq_m128d_s64( - vandq_s64(vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b))); -} - -// Computes the bitwise OR of the four single-precision, floating-point values -// of a and b. -// https://msdn.microsoft.com/en-us/library/vstudio/7ctdsyy0(v=vs.100).aspx -FORCE_INLINE __m128 _mm_or_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_s32( - vorrq_s32(vreinterpretq_s32_m128(a), vreinterpretq_s32_m128(b))); -} - -// Computes bitwise EXOR (exclusive-or) of the four single-precision, -// floating-point values of a and b. -// https://msdn.microsoft.com/en-us/library/ss6k3wk8(v=vs.100).aspx -FORCE_INLINE __m128 _mm_xor_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_s32( - veorq_s32(vreinterpretq_s32_m128(a), vreinterpretq_s32_m128(b))); -} - -// Compute the bitwise XOR of packed double-precision (64-bit) floating-point -// elements in a and b, and store the results in dst. -// -// FOR j := 0 to 1 -// i := j*64 -// dst[i+63:i] := a[i+63:i] XOR b[i+63:i] -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_xor_pd -FORCE_INLINE __m128d _mm_xor_pd(__m128d a, __m128d b) -{ - return vreinterpretq_m128d_s64( - veorq_s64(vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b))); -} - -// Compute the bitwise OR of packed double-precision (64-bit) floating-point -// elements in a and b, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_or_pd -FORCE_INLINE __m128d _mm_or_pd(__m128d a, __m128d b) -{ - return vreinterpretq_m128d_s64( - vorrq_s64(vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b))); -} - -// Computes the bitwise OR of the 128-bit value in a and the 128-bit value in b. -// -// r := a | b -// -// https://msdn.microsoft.com/en-us/library/vstudio/ew8ty0db(v=vs.100).aspx -FORCE_INLINE __m128i _mm_or_si128(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s32( - vorrq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -} - -// Computes the bitwise XOR of the 128-bit value in a and the 128-bit value in -// b. https://msdn.microsoft.com/en-us/library/fzt08www(v=vs.100).aspx -FORCE_INLINE __m128i _mm_xor_si128(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s32( - veorq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -} - -// Duplicate the low double-precision (64-bit) floating-point element from a, -// and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movedup_pd -FORCE_INLINE __m128d _mm_movedup_pd(__m128d a) -{ -#if (__aarch64__) - return vreinterpretq_m128d_f64( - vdupq_laneq_f64(vreinterpretq_f64_m128d(a), 0)); -#else - return vreinterpretq_m128d_u64( - vdupq_n_u64(vgetq_lane_u64(vreinterpretq_u64_m128d(a), 0))); -#endif -} - -// Duplicate odd-indexed single-precision (32-bit) floating-point elements -// from a, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movehdup_ps -FORCE_INLINE __m128 _mm_movehdup_ps(__m128 a) -{ -#if __has_builtin(__builtin_shufflevector) - return vreinterpretq_m128_f32(__builtin_shufflevector( - vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a), 1, 1, 3, 3)); -#else - float32_t a1 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 1); - float32_t a3 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 3); - float ALIGN_STRUCT(16) data[4] = {a1, a1, a3, a3}; - return vreinterpretq_m128_f32(vld1q_f32(data)); -#endif -} - -// Duplicate even-indexed single-precision (32-bit) floating-point elements -// from a, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_moveldup_ps -FORCE_INLINE __m128 _mm_moveldup_ps(__m128 a) -{ -#if __has_builtin(__builtin_shufflevector) - return vreinterpretq_m128_f32(__builtin_shufflevector( - vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a), 0, 0, 2, 2)); -#else - float32_t a0 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); - float32_t a2 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 2); - float ALIGN_STRUCT(16) data[4] = {a0, a0, a2, a2}; - return vreinterpretq_m128_f32(vld1q_f32(data)); -#endif -} - -// Moves the upper two values of B into the lower two values of A. -// -// r3 := a3 -// r2 := a2 -// r1 := b3 -// r0 := b2 -FORCE_INLINE __m128 _mm_movehl_ps(__m128 __A, __m128 __B) -{ - float32x2_t a32 = vget_high_f32(vreinterpretq_f32_m128(__A)); - float32x2_t b32 = vget_high_f32(vreinterpretq_f32_m128(__B)); - return vreinterpretq_m128_f32(vcombine_f32(b32, a32)); -} - -// Moves the lower two values of B into the upper two values of A. -// -// r3 := b1 -// r2 := b0 -// r1 := a1 -// r0 := a0 -FORCE_INLINE __m128 _mm_movelh_ps(__m128 __A, __m128 __B) -{ - float32x2_t a10 = vget_low_f32(vreinterpretq_f32_m128(__A)); - float32x2_t b10 = vget_low_f32(vreinterpretq_f32_m128(__B)); - return vreinterpretq_m128_f32(vcombine_f32(a10, b10)); -} - -// Create mask from the most significant bit of each 8-bit element in a, and -// store the result in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movemask_pi8 -FORCE_INLINE int _mm_movemask_pi8(__m64 a) -{ - uint8x8_t input = vreinterpret_u8_m64(a); -#if defined(__aarch64__) - static const int8x8_t shift = {0, 1, 2, 3, 4, 5, 6, 7}; - uint8x8_t tmp = vshr_n_u8(input, 7); - return vaddv_u8(vshl_u8(tmp, shift)); -#else - // Refer the implementation of `_mm_movemask_epi8` - uint16x4_t high_bits = vreinterpret_u16_u8(vshr_n_u8(input, 7)); - uint32x2_t paired16 = - vreinterpret_u32_u16(vsra_n_u16(high_bits, high_bits, 7)); - uint8x8_t paired32 = - vreinterpret_u8_u32(vsra_n_u32(paired16, paired16, 14)); - return vget_lane_u8(paired32, 0) | ((int) vget_lane_u8(paired32, 4) << 4); -#endif -} - -// Compute the absolute value of packed signed 32-bit integers in a, and store -// the unsigned results in dst. -// -// FOR j := 0 to 3 -// i := j*32 -// dst[i+31:i] := ABS(a[i+31:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_epi32 -FORCE_INLINE __m128i _mm_abs_epi32(__m128i a) -{ - return vreinterpretq_m128i_s32(vabsq_s32(vreinterpretq_s32_m128i(a))); -} - -// Compute the absolute value of packed signed 16-bit integers in a, and store -// the unsigned results in dst. -// -// FOR j := 0 to 7 -// i := j*16 -// dst[i+15:i] := ABS(a[i+15:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_epi16 -FORCE_INLINE __m128i _mm_abs_epi16(__m128i a) -{ - return vreinterpretq_m128i_s16(vabsq_s16(vreinterpretq_s16_m128i(a))); -} - -// Compute the absolute value of packed signed 8-bit integers in a, and store -// the unsigned results in dst. -// -// FOR j := 0 to 15 -// i := j*8 -// dst[i+7:i] := ABS(a[i+7:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_epi8 -FORCE_INLINE __m128i _mm_abs_epi8(__m128i a) -{ - return vreinterpretq_m128i_s8(vabsq_s8(vreinterpretq_s8_m128i(a))); -} - -// Compute the absolute value of packed signed 32-bit integers in a, and store -// the unsigned results in dst. -// -// FOR j := 0 to 1 -// i := j*32 -// dst[i+31:i] := ABS(a[i+31:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_pi32 -FORCE_INLINE __m64 _mm_abs_pi32(__m64 a) -{ - return vreinterpret_m64_s32(vabs_s32(vreinterpret_s32_m64(a))); -} - -// Compute the absolute value of packed signed 16-bit integers in a, and store -// the unsigned results in dst. -// -// FOR j := 0 to 3 -// i := j*16 -// dst[i+15:i] := ABS(a[i+15:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_pi16 -FORCE_INLINE __m64 _mm_abs_pi16(__m64 a) -{ - return vreinterpret_m64_s16(vabs_s16(vreinterpret_s16_m64(a))); -} - -// Compute the absolute value of packed signed 8-bit integers in a, and store -// the unsigned results in dst. -// -// FOR j := 0 to 7 -// i := j*8 -// dst[i+7:i] := ABS(a[i+7:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_pi8 -FORCE_INLINE __m64 _mm_abs_pi8(__m64 a) -{ - return vreinterpret_m64_s8(vabs_s8(vreinterpret_s8_m64(a))); -} - -// Concatenate 16-byte blocks in a and b into a 32-byte temporary result, shift -// the result right by imm8 bytes, and store the low 16 bytes in dst. -// -// tmp[255:0] := ((a[127:0] << 128)[255:0] OR b[127:0]) >> (imm8*8) -// dst[127:0] := tmp[127:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_alignr_epi8 -#define _mm_alignr_epi8(a, b, imm) \ - __extension__({ \ - __m128i ret; \ - if (unlikely((imm) >= 32)) { \ - ret = _mm_setzero_si128(); \ - } else { \ - uint8x16_t tmp_low, tmp_high; \ - if (imm >= 16) { \ - const int idx = imm - 16; \ - tmp_low = vreinterpretq_u8_m128i(a); \ - tmp_high = vdupq_n_u8(0); \ - ret = \ - vreinterpretq_m128i_u8(vextq_u8(tmp_low, tmp_high, idx)); \ - } else { \ - const int idx = imm; \ - tmp_low = vreinterpretq_u8_m128i(b); \ - tmp_high = vreinterpretq_u8_m128i(a); \ - ret = \ - vreinterpretq_m128i_u8(vextq_u8(tmp_low, tmp_high, idx)); \ - } \ - } \ - ret; \ - }) - -// Concatenate 8-byte blocks in a and b into a 16-byte temporary result, shift -// the result right by imm8 bytes, and store the low 8 bytes in dst. -// -// tmp[127:0] := ((a[63:0] << 64)[127:0] OR b[63:0]) >> (imm8*8) -// dst[63:0] := tmp[63:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_alignr_pi8 -#define _mm_alignr_pi8(a, b, imm) \ - __extension__({ \ - __m64 ret; \ - if (unlikely((imm) >= 16)) { \ - ret = vreinterpret_m64_s8(vdup_n_s8(0)); \ - } else { \ - uint8x8_t tmp_low, tmp_high; \ - if (imm >= 8) { \ - const int idx = imm - 8; \ - tmp_low = vreinterpret_u8_m64(a); \ - tmp_high = vdup_n_u8(0); \ - ret = vreinterpret_m64_u8(vext_u8(tmp_low, tmp_high, idx)); \ - } else { \ - const int idx = imm; \ - tmp_low = vreinterpret_u8_m64(b); \ - tmp_high = vreinterpret_u8_m64(a); \ - ret = vreinterpret_m64_u8(vext_u8(tmp_low, tmp_high, idx)); \ - } \ - } \ - ret; \ - }) +} fpcr_bitfield; // Takes the upper 64 bits of a and places it in the low end of the result // Takes the lower 64 bits of b and places it into the high end of the result. @@ -1908,6 +677,255 @@ FORCE_INLINE __m128 _mm_shuffle_ps_2032(__m128 a, __m128 b) return vreinterpretq_m128_f32(vcombine_f32(a32, b20)); } +// Kahan summation for accurate summation of floating-point numbers. +// http://blog.zachbjornson.com/2019/08/11/fast-float-summation.html +FORCE_INLINE void _sse2neon_kadd_f32(float *sum, float *c, float y) +{ + y -= *c; + float t = *sum + y; + *c = (t - *sum) - y; + *sum = t; +} + +#if defined(__ARM_FEATURE_CRYPTO) && \ + (defined(__aarch64__) || __has_builtin(__builtin_arm_crypto_vmullp64)) +// Wraps vmull_p64 +FORCE_INLINE uint64x2_t _sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b) +{ + poly64_t a = vget_lane_p64(vreinterpret_p64_u64(_a), 0); + poly64_t b = vget_lane_p64(vreinterpret_p64_u64(_b), 0); + return vreinterpretq_u64_p128(vmull_p64(a, b)); +} +#else // ARMv7 polyfill +// ARMv7/some A64 lacks vmull_p64, but it has vmull_p8. +// +// vmull_p8 calculates 8 8-bit->16-bit polynomial multiplies, but we need a +// 64-bit->128-bit polynomial multiply. +// +// It needs some work and is somewhat slow, but it is still faster than all +// known scalar methods. +// +// Algorithm adapted to C from +// https://www.workofard.com/2017/07/ghash-for-low-end-cores/, which is adapted +// from "Fast Software Polynomial Multiplication on ARM Processors Using the +// NEON Engine" by Danilo Camara, Conrado Gouvea, Julio Lopez and Ricardo Dahab +// (https://hal.inria.fr/hal-01506572) +static uint64x2_t _sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b) +{ + poly8x8_t a = vreinterpret_p8_u64(_a); + poly8x8_t b = vreinterpret_p8_u64(_b); + + // Masks + uint8x16_t k48_32 = vcombine_u8(vcreate_u8(0x0000ffffffffffff), + vcreate_u8(0x00000000ffffffff)); + uint8x16_t k16_00 = vcombine_u8(vcreate_u8(0x000000000000ffff), + vcreate_u8(0x0000000000000000)); + + // Do the multiplies, rotating with vext to get all combinations + uint8x16_t d = vreinterpretq_u8_p16(vmull_p8(a, b)); // D = A0 * B0 + uint8x16_t e = + vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 1))); // E = A0 * B1 + uint8x16_t f = + vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 1), b)); // F = A1 * B0 + uint8x16_t g = + vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 2))); // G = A0 * B2 + uint8x16_t h = + vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 2), b)); // H = A2 * B0 + uint8x16_t i = + vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 3))); // I = A0 * B3 + uint8x16_t j = + vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 3), b)); // J = A3 * B0 + uint8x16_t k = + vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 4))); // L = A0 * B4 + + // Add cross products + uint8x16_t l = veorq_u8(e, f); // L = E + F + uint8x16_t m = veorq_u8(g, h); // M = G + H + uint8x16_t n = veorq_u8(i, j); // N = I + J + + // Interleave. Using vzip1 and vzip2 prevents Clang from emitting TBL + // instructions. +#if defined(__aarch64__) + uint8x16_t lm_p0 = vreinterpretq_u8_u64( + vzip1q_u64(vreinterpretq_u64_u8(l), vreinterpretq_u64_u8(m))); + uint8x16_t lm_p1 = vreinterpretq_u8_u64( + vzip2q_u64(vreinterpretq_u64_u8(l), vreinterpretq_u64_u8(m))); + uint8x16_t nk_p0 = vreinterpretq_u8_u64( + vzip1q_u64(vreinterpretq_u64_u8(n), vreinterpretq_u64_u8(k))); + uint8x16_t nk_p1 = vreinterpretq_u8_u64( + vzip2q_u64(vreinterpretq_u64_u8(n), vreinterpretq_u64_u8(k))); +#else + uint8x16_t lm_p0 = vcombine_u8(vget_low_u8(l), vget_low_u8(m)); + uint8x16_t lm_p1 = vcombine_u8(vget_high_u8(l), vget_high_u8(m)); + uint8x16_t nk_p0 = vcombine_u8(vget_low_u8(n), vget_low_u8(k)); + uint8x16_t nk_p1 = vcombine_u8(vget_high_u8(n), vget_high_u8(k)); +#endif + // t0 = (L) (P0 + P1) << 8 + // t1 = (M) (P2 + P3) << 16 + uint8x16_t t0t1_tmp = veorq_u8(lm_p0, lm_p1); + uint8x16_t t0t1_h = vandq_u8(lm_p1, k48_32); + uint8x16_t t0t1_l = veorq_u8(t0t1_tmp, t0t1_h); + + // t2 = (N) (P4 + P5) << 24 + // t3 = (K) (P6 + P7) << 32 + uint8x16_t t2t3_tmp = veorq_u8(nk_p0, nk_p1); + uint8x16_t t2t3_h = vandq_u8(nk_p1, k16_00); + uint8x16_t t2t3_l = veorq_u8(t2t3_tmp, t2t3_h); + + // De-interleave +#if defined(__aarch64__) + uint8x16_t t0 = vreinterpretq_u8_u64( + vuzp1q_u64(vreinterpretq_u64_u8(t0t1_l), vreinterpretq_u64_u8(t0t1_h))); + uint8x16_t t1 = vreinterpretq_u8_u64( + vuzp2q_u64(vreinterpretq_u64_u8(t0t1_l), vreinterpretq_u64_u8(t0t1_h))); + uint8x16_t t2 = vreinterpretq_u8_u64( + vuzp1q_u64(vreinterpretq_u64_u8(t2t3_l), vreinterpretq_u64_u8(t2t3_h))); + uint8x16_t t3 = vreinterpretq_u8_u64( + vuzp2q_u64(vreinterpretq_u64_u8(t2t3_l), vreinterpretq_u64_u8(t2t3_h))); +#else + uint8x16_t t1 = vcombine_u8(vget_high_u8(t0t1_l), vget_high_u8(t0t1_h)); + uint8x16_t t0 = vcombine_u8(vget_low_u8(t0t1_l), vget_low_u8(t0t1_h)); + uint8x16_t t3 = vcombine_u8(vget_high_u8(t2t3_l), vget_high_u8(t2t3_h)); + uint8x16_t t2 = vcombine_u8(vget_low_u8(t2t3_l), vget_low_u8(t2t3_h)); +#endif + // Shift the cross products + uint8x16_t t0_shift = vextq_u8(t0, t0, 15); // t0 << 8 + uint8x16_t t1_shift = vextq_u8(t1, t1, 14); // t1 << 16 + uint8x16_t t2_shift = vextq_u8(t2, t2, 13); // t2 << 24 + uint8x16_t t3_shift = vextq_u8(t3, t3, 12); // t3 << 32 + + // Accumulate the products + uint8x16_t cross1 = veorq_u8(t0_shift, t1_shift); + uint8x16_t cross2 = veorq_u8(t2_shift, t3_shift); + uint8x16_t mix = veorq_u8(d, cross1); + uint8x16_t r = veorq_u8(mix, cross2); + return vreinterpretq_u64_u8(r); +} +#endif // ARMv7 polyfill + +// C equivalent: +// __m128i _mm_shuffle_epi32_default(__m128i a, +// __constrange(0, 255) int imm) { +// __m128i ret; +// ret[0] = a[imm & 0x3]; ret[1] = a[(imm >> 2) & 0x3]; +// ret[2] = a[(imm >> 4) & 0x03]; ret[3] = a[(imm >> 6) & 0x03]; +// return ret; +// } +#define _mm_shuffle_epi32_default(a, imm) \ + __extension__({ \ + int32x4_t ret; \ + ret = vmovq_n_s32( \ + vgetq_lane_s32(vreinterpretq_s32_m128i(a), (imm) & (0x3))); \ + ret = vsetq_lane_s32( \ + vgetq_lane_s32(vreinterpretq_s32_m128i(a), ((imm) >> 2) & 0x3), \ + ret, 1); \ + ret = vsetq_lane_s32( \ + vgetq_lane_s32(vreinterpretq_s32_m128i(a), ((imm) >> 4) & 0x3), \ + ret, 2); \ + ret = vsetq_lane_s32( \ + vgetq_lane_s32(vreinterpretq_s32_m128i(a), ((imm) >> 6) & 0x3), \ + ret, 3); \ + vreinterpretq_m128i_s32(ret); \ + }) + +// Takes the upper 64 bits of a and places it in the low end of the result +// Takes the lower 64 bits of a and places it into the high end of the result. +FORCE_INLINE __m128i _mm_shuffle_epi_1032(__m128i a) +{ + int32x2_t a32 = vget_high_s32(vreinterpretq_s32_m128i(a)); + int32x2_t a10 = vget_low_s32(vreinterpretq_s32_m128i(a)); + return vreinterpretq_m128i_s32(vcombine_s32(a32, a10)); +} + +// takes the lower two 32-bit values from a and swaps them and places in low end +// of result takes the higher two 32 bit values from a and swaps them and places +// in high end of result. +FORCE_INLINE __m128i _mm_shuffle_epi_2301(__m128i a) +{ + int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); + int32x2_t a23 = vrev64_s32(vget_high_s32(vreinterpretq_s32_m128i(a))); + return vreinterpretq_m128i_s32(vcombine_s32(a01, a23)); +} + +// rotates the least significant 32 bits into the most significant 32 bits, and +// shifts the rest down +FORCE_INLINE __m128i _mm_shuffle_epi_0321(__m128i a) +{ + return vreinterpretq_m128i_s32( + vextq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(a), 1)); +} + +// rotates the most significant 32 bits into the least significant 32 bits, and +// shifts the rest up +FORCE_INLINE __m128i _mm_shuffle_epi_2103(__m128i a) +{ + return vreinterpretq_m128i_s32( + vextq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(a), 3)); +} + +// gets the lower 64 bits of a, and places it in the upper 64 bits +// gets the lower 64 bits of a and places it in the lower 64 bits +FORCE_INLINE __m128i _mm_shuffle_epi_1010(__m128i a) +{ + int32x2_t a10 = vget_low_s32(vreinterpretq_s32_m128i(a)); + return vreinterpretq_m128i_s32(vcombine_s32(a10, a10)); +} + +// gets the lower 64 bits of a, swaps the 0 and 1 elements, and places it in the +// lower 64 bits gets the lower 64 bits of a, and places it in the upper 64 bits +FORCE_INLINE __m128i _mm_shuffle_epi_1001(__m128i a) +{ + int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); + int32x2_t a10 = vget_low_s32(vreinterpretq_s32_m128i(a)); + return vreinterpretq_m128i_s32(vcombine_s32(a01, a10)); +} + +// gets the lower 64 bits of a, swaps the 0 and 1 elements and places it in the +// upper 64 bits gets the lower 64 bits of a, swaps the 0 and 1 elements, and +// places it in the lower 64 bits +FORCE_INLINE __m128i _mm_shuffle_epi_0101(__m128i a) +{ + int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); + return vreinterpretq_m128i_s32(vcombine_s32(a01, a01)); +} + +FORCE_INLINE __m128i _mm_shuffle_epi_2211(__m128i a) +{ + int32x2_t a11 = vdup_lane_s32(vget_low_s32(vreinterpretq_s32_m128i(a)), 1); + int32x2_t a22 = vdup_lane_s32(vget_high_s32(vreinterpretq_s32_m128i(a)), 0); + return vreinterpretq_m128i_s32(vcombine_s32(a11, a22)); +} + +FORCE_INLINE __m128i _mm_shuffle_epi_0122(__m128i a) +{ + int32x2_t a22 = vdup_lane_s32(vget_high_s32(vreinterpretq_s32_m128i(a)), 0); + int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); + return vreinterpretq_m128i_s32(vcombine_s32(a22, a01)); +} + +FORCE_INLINE __m128i _mm_shuffle_epi_3332(__m128i a) +{ + int32x2_t a32 = vget_high_s32(vreinterpretq_s32_m128i(a)); + int32x2_t a33 = vdup_lane_s32(vget_high_s32(vreinterpretq_s32_m128i(a)), 1); + return vreinterpretq_m128i_s32(vcombine_s32(a32, a33)); +} + +// FORCE_INLINE __m128i _mm_shuffle_epi32_splat(__m128i a, __constrange(0,255) +// int imm) +#if defined(__aarch64__) +#define _mm_shuffle_epi32_splat(a, imm) \ + __extension__({ \ + vreinterpretq_m128i_s32( \ + vdupq_laneq_s32(vreinterpretq_s32_m128i(a), (imm))); \ + }) +#else +#define _mm_shuffle_epi32_splat(a, imm) \ + __extension__({ \ + vreinterpretq_m128i_s32( \ + vdupq_n_s32(vgetq_lane_s32(vreinterpretq_s32_m128i(a), (imm)))); \ + }) +#endif + // NEON does not support a general purpose permute intrinsic // Selects four specific single-precision, floating-point values from a and b, // based on the mask i. @@ -1939,6 +957,1595 @@ FORCE_INLINE __m128 _mm_shuffle_ps_2032(__m128 a, __m128 b) vreinterpretq_m128_f32(ret); \ }) +// Shuffles the lower 4 signed or unsigned 16-bit integers in a as specified +// by imm. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/y41dkk37(v=vs.100) +// FORCE_INLINE __m128i _mm_shufflelo_epi16_function(__m128i a, +// __constrange(0,255) int +// imm) +#define _mm_shufflelo_epi16_function(a, imm) \ + __extension__({ \ + int16x8_t ret = vreinterpretq_s16_m128i(a); \ + int16x4_t lowBits = vget_low_s16(ret); \ + ret = vsetq_lane_s16(vget_lane_s16(lowBits, (imm) & (0x3)), ret, 0); \ + ret = vsetq_lane_s16(vget_lane_s16(lowBits, ((imm) >> 2) & 0x3), ret, \ + 1); \ + ret = vsetq_lane_s16(vget_lane_s16(lowBits, ((imm) >> 4) & 0x3), ret, \ + 2); \ + ret = vsetq_lane_s16(vget_lane_s16(lowBits, ((imm) >> 6) & 0x3), ret, \ + 3); \ + vreinterpretq_m128i_s16(ret); \ + }) + +// Shuffles the upper 4 signed or unsigned 16-bit integers in a as specified +// by imm. +// https://msdn.microsoft.com/en-us/library/13ywktbs(v=vs.100).aspx +// FORCE_INLINE __m128i _mm_shufflehi_epi16_function(__m128i a, +// __constrange(0,255) int +// imm) +#define _mm_shufflehi_epi16_function(a, imm) \ + __extension__({ \ + int16x8_t ret = vreinterpretq_s16_m128i(a); \ + int16x4_t highBits = vget_high_s16(ret); \ + ret = vsetq_lane_s16(vget_lane_s16(highBits, (imm) & (0x3)), ret, 4); \ + ret = vsetq_lane_s16(vget_lane_s16(highBits, ((imm) >> 2) & 0x3), ret, \ + 5); \ + ret = vsetq_lane_s16(vget_lane_s16(highBits, ((imm) >> 4) & 0x3), ret, \ + 6); \ + ret = vsetq_lane_s16(vget_lane_s16(highBits, ((imm) >> 6) & 0x3), ret, \ + 7); \ + vreinterpretq_m128i_s16(ret); \ + }) + +/* MMX */ + +//_mm_empty is a no-op on arm +FORCE_INLINE void _mm_empty(void) {} + +/* SSE */ + +// Adds the four single-precision, floating-point values of a and b. +// +// r0 := a0 + b0 +// r1 := a1 + b1 +// r2 := a2 + b2 +// r3 := a3 + b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/c9848chc(v=vs.100).aspx +FORCE_INLINE __m128 _mm_add_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_f32( + vaddq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// adds the scalar single-precision floating point values of a and b. +// https://msdn.microsoft.com/en-us/library/be94x2y6(v=vs.100).aspx +FORCE_INLINE __m128 _mm_add_ss(__m128 a, __m128 b) +{ + float32_t b0 = vgetq_lane_f32(vreinterpretq_f32_m128(b), 0); + float32x4_t value = vsetq_lane_f32(b0, vdupq_n_f32(0), 0); + // the upper values in the result must be the remnants of . + return vreinterpretq_m128_f32(vaddq_f32(a, value)); +} + +// Computes the bitwise AND of the four single-precision, floating-point values +// of a and b. +// +// r0 := a0 & b0 +// r1 := a1 & b1 +// r2 := a2 & b2 +// r3 := a3 & b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/73ck1xc5(v=vs.100).aspx +FORCE_INLINE __m128 _mm_and_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_s32( + vandq_s32(vreinterpretq_s32_m128(a), vreinterpretq_s32_m128(b))); +} + +// Computes the bitwise AND-NOT of the four single-precision, floating-point +// values of a and b. +// +// r0 := ~a0 & b0 +// r1 := ~a1 & b1 +// r2 := ~a2 & b2 +// r3 := ~a3 & b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/68h7wd02(v=vs.100).aspx +FORCE_INLINE __m128 _mm_andnot_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_s32( + vbicq_s32(vreinterpretq_s32_m128(b), + vreinterpretq_s32_m128(a))); // *NOTE* argument swap +} + +// Average packed unsigned 16-bit integers in a and b, and store the results in +// dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := (a[i+15:i] + b[i+15:i] + 1) >> 1 +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_avg_pu16 +FORCE_INLINE __m64 _mm_avg_pu16(__m64 a, __m64 b) +{ + return vreinterpret_m64_u16( + vrhadd_u16(vreinterpret_u16_m64(a), vreinterpret_u16_m64(b))); +} + +// Average packed unsigned 8-bit integers in a and b, and store the results in +// dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := (a[i+7:i] + b[i+7:i] + 1) >> 1 +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_avg_pu8 +FORCE_INLINE __m64 _mm_avg_pu8(__m64 a, __m64 b) +{ + return vreinterpret_m64_u8( + vrhadd_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); +} + +// Compares for equality. +// https://msdn.microsoft.com/en-us/library/vstudio/36aectz5(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmpeq_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for equality. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/k423z28e(v=vs.100) +FORCE_INLINE __m128 _mm_cmpeq_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpeq_ps(a, b)); +} + +// Compares for greater than or equal. +// https://msdn.microsoft.com/en-us/library/vstudio/fs813y2t(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmpge_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vcgeq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for greater than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/kesh3ddc(v=vs.100) +FORCE_INLINE __m128 _mm_cmpge_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpge_ps(a, b)); +} + +// Compares for greater than. +// +// r0 := (a0 > b0) ? 0xffffffff : 0x0 +// r1 := (a1 > b1) ? 0xffffffff : 0x0 +// r2 := (a2 > b2) ? 0xffffffff : 0x0 +// r3 := (a3 > b3) ? 0xffffffff : 0x0 +// +// https://msdn.microsoft.com/en-us/library/vstudio/11dy102s(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmpgt_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vcgtq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for greater than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/1xyyyy9e(v=vs.100) +FORCE_INLINE __m128 _mm_cmpgt_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpgt_ps(a, b)); +} + +// Compares for less than or equal. +// +// r0 := (a0 <= b0) ? 0xffffffff : 0x0 +// r1 := (a1 <= b1) ? 0xffffffff : 0x0 +// r2 := (a2 <= b2) ? 0xffffffff : 0x0 +// r3 := (a3 <= b3) ? 0xffffffff : 0x0 +// +// https://msdn.microsoft.com/en-us/library/vstudio/1s75w83z(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmple_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vcleq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for less than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/a7x0hbhw(v=vs.100) +FORCE_INLINE __m128 _mm_cmple_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmple_ps(a, b)); +} + +// Compares for less than +// https://msdn.microsoft.com/en-us/library/vstudio/f330yhc8(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmplt_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32( + vcltq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Compares for less than +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/fy94wye7(v=vs.100) +FORCE_INLINE __m128 _mm_cmplt_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmplt_ps(a, b)); +} + +// Compares for inequality. +// https://msdn.microsoft.com/en-us/library/sf44thbx(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cmpneq_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for inequality. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/ekya8fh4(v=vs.100) +FORCE_INLINE __m128 _mm_cmpneq_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpneq_ps(a, b)); +} + +// Compares for not greater than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/wsexys62(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnge_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vcgeq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for not greater than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/fk2y80s8(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnge_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpnge_ps(a, b)); +} + +// Compares for not greater than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/d0xh7w0s(v=vs.100) +FORCE_INLINE __m128 _mm_cmpngt_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vcgtq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for not greater than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/z7x9ydwh(v=vs.100) +FORCE_INLINE __m128 _mm_cmpngt_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpngt_ps(a, b)); +} + +// Compares for not less than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/6a330kxw(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnle_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vcleq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for not less than or equal. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/z7x9ydwh(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnle_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpnle_ps(a, b)); +} + +// Compares for not less than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/4686bbdw(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnlt_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_u32(vmvnq_u32( + vcltq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); +} + +// Compares for not less than. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/56b9z2wf(v=vs.100) +FORCE_INLINE __m128 _mm_cmpnlt_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpnlt_ps(a, b)); +} + +// Compares the four 32-bit floats in a and b to check if any values are NaN. +// Ordered compare between each value returns true for "orderable" and false for +// "not orderable" (NaN). +// https://msdn.microsoft.com/en-us/library/vstudio/0h9w00fx(v=vs.100).aspx see +// also: +// http://stackoverflow.com/questions/8627331/what-does-ordered-unordered-comparison-mean +// http://stackoverflow.com/questions/29349621/neon-isnanval-intrinsics +FORCE_INLINE __m128 _mm_cmpord_ps(__m128 a, __m128 b) +{ + // Note: NEON does not have ordered compare builtin + // Need to compare a eq a and b eq b to check for NaN + // Do AND of results to get final + uint32x4_t ceqaa = + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); + uint32x4_t ceqbb = + vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_u32(vandq_u32(ceqaa, ceqbb)); +} + +// Compares for ordered. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/343t62da(v=vs.100) +FORCE_INLINE __m128 _mm_cmpord_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpord_ps(a, b)); +} + +// Compares for unordered. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/khy6fk1t(v=vs.100) +FORCE_INLINE __m128 _mm_cmpunord_ps(__m128 a, __m128 b) +{ + uint32x4_t f32a = + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); + uint32x4_t f32b = + vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_u32(vmvnq_u32(vandq_u32(f32a, f32b))); +} + +// Compares for unordered. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/2as2387b(v=vs.100) +FORCE_INLINE __m128 _mm_cmpunord_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_cmpunord_ps(a, b)); +} + +// Compares the lower single-precision floating point scalar values of a and b +// using an equality operation. : +// https://msdn.microsoft.com/en-us/library/93yx2h2b(v=vs.100).aspx +FORCE_INLINE int _mm_comieq_ss(__m128 a, __m128 b) +{ + uint32x4_t a_eq_b = + vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_eq_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using a greater than or equal operation. : +// https://msdn.microsoft.com/en-us/library/8t80des6(v=vs.100).aspx +FORCE_INLINE int _mm_comige_ss(__m128 a, __m128 b) +{ + uint32x4_t a_ge_b = + vcgeq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_ge_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using a greater than operation. : +// https://msdn.microsoft.com/en-us/library/b0738e0t(v=vs.100).aspx +FORCE_INLINE int _mm_comigt_ss(__m128 a, __m128 b) +{ + uint32x4_t a_gt_b = + vcgtq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_gt_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using a less than or equal operation. : +// https://msdn.microsoft.com/en-us/library/1w4t7c57(v=vs.90).aspx +FORCE_INLINE int _mm_comile_ss(__m128 a, __m128 b) +{ + uint32x4_t a_le_b = + vcleq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_le_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using a less than operation. : +// https://msdn.microsoft.com/en-us/library/2kwe606b(v=vs.90).aspx Important +// note!! The documentation on MSDN is incorrect! If either of the values is a +// NAN the docs say you will get a one, but in fact, it will return a zero!! +FORCE_INLINE int _mm_comilt_ss(__m128 a, __m128 b) +{ + uint32x4_t a_lt_b = + vcltq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); + return vgetq_lane_u32(a_lt_b, 0) & 0x1; +} + +// Compares the lower single-precision floating point scalar values of a and b +// using an inequality operation. : +// https://msdn.microsoft.com/en-us/library/bafh5e0a(v=vs.90).aspx +FORCE_INLINE int _mm_comineq_ss(__m128 a, __m128 b) +{ + return !_mm_comieq_ss(a, b); +} + +// Convert packed signed 32-bit integers in b to packed single-precision +// (32-bit) floating-point elements, store the results in the lower 2 elements +// of dst, and copy the upper 2 packed elements from a to the upper elements of +// dst. +// +// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) +// dst[63:32] := Convert_Int32_To_FP32(b[63:32]) +// dst[95:64] := a[95:64] +// dst[127:96] := a[127:96] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_pi2ps +FORCE_INLINE __m128 _mm_cvt_pi2ps(__m128 a, __m64 b) +{ + return vreinterpretq_m128_f32( + vcombine_f32(vcvt_f32_s32(vreinterpret_s32_m64(b)), + vget_high_f32(vreinterpretq_f32_m128(a)))); +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 32-bit integers, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// dst[i+31:i] := Convert_FP32_To_Int32(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_ps2pi +FORCE_INLINE __m64 _mm_cvt_ps2pi(__m128 a) +{ +#if defined(__aarch64__) || defined(__ARM_FEATURE_DIRECTED_ROUNDING) + return vreinterpret_m64_s32( + vget_low_s32(vcvtnq_s32_f32(vrndiq_f32(vreinterpretq_f32_m128(a))))); +#else + return vreinterpret_m64_s32(vcvt_s32_f32(vget_low_f32( + vreinterpretq_f32_m128(_mm_round_ps(a, _MM_FROUND_CUR_DIRECTION))))); +#endif +} + +// Convert the signed 32-bit integer b to a single-precision (32-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper 3 packed elements from a to the upper elements of dst. +// +// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_si2ss +FORCE_INLINE __m128 _mm_cvt_si2ss(__m128 a, int b) +{ + return vreinterpretq_m128_f32( + vsetq_lane_f32((float) b, vreinterpretq_f32_m128(a), 0)); +} + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 32-bit integer, and store the result in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_ss2si +FORCE_INLINE int _mm_cvt_ss2si(__m128 a) +{ +#if defined(__aarch64__) || defined(__ARM_FEATURE_DIRECTED_ROUNDING) + return vgetq_lane_s32(vcvtnq_s32_f32(vrndiq_f32(vreinterpretq_f32_m128(a))), + 0); +#else + float32_t data = vgetq_lane_f32( + vreinterpretq_f32_m128(_mm_round_ps(a, _MM_FROUND_CUR_DIRECTION)), 0); + return (int32_t) data; +#endif +} + +// Convert packed 16-bit integers in a to packed single-precision (32-bit) +// floating-point elements, and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// m := j*32 +// dst[m+31:m] := Convert_Int16_To_FP32(a[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi16_ps +FORCE_INLINE __m128 _mm_cvtpi16_ps(__m64 a) +{ + return vreinterpretq_m128_f32( + vcvtq_f32_s32(vmovl_s16(vreinterpret_s16_m64(a)))); +} + +// Convert packed 32-bit integers in b to packed single-precision (32-bit) +// floating-point elements, store the results in the lower 2 elements of dst, +// and copy the upper 2 packed elements from a to the upper elements of dst. +// +// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) +// dst[63:32] := Convert_Int32_To_FP32(b[63:32]) +// dst[95:64] := a[95:64] +// dst[127:96] := a[127:96] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi32_ps +FORCE_INLINE __m128 _mm_cvtpi32_ps(__m128 a, __m64 b) +{ + return vreinterpretq_m128_f32( + vcombine_f32(vcvt_f32_s32(vreinterpret_s32_m64(b)), + vget_high_f32(vreinterpretq_f32_m128(a)))); +} + +// Convert packed signed 32-bit integers in a to packed single-precision +// (32-bit) floating-point elements, store the results in the lower 2 elements +// of dst, then convert the packed signed 32-bit integers in b to +// single-precision (32-bit) floating-point element, and store the results in +// the upper 2 elements of dst. +// +// dst[31:0] := Convert_Int32_To_FP32(a[31:0]) +// dst[63:32] := Convert_Int32_To_FP32(a[63:32]) +// dst[95:64] := Convert_Int32_To_FP32(b[31:0]) +// dst[127:96] := Convert_Int32_To_FP32(b[63:32]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi32x2_ps +FORCE_INLINE __m128 _mm_cvtpi32x2_ps(__m64 a, __m64 b) +{ + return vreinterpretq_m128_f32(vcvtq_f32_s32( + vcombine_s32(vreinterpret_s32_m64(a), vreinterpret_s32_m64(b)))); +} + +// Convert the lower packed 8-bit integers in a to packed single-precision +// (32-bit) floating-point elements, and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*8 +// m := j*32 +// dst[m+31:m] := Convert_Int8_To_FP32(a[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi8_ps +FORCE_INLINE __m128 _mm_cvtpi8_ps(__m64 a) +{ + return vreinterpretq_m128_f32(vcvtq_f32_s32( + vmovl_s16(vget_low_s16(vmovl_s8(vreinterpret_s8_m64(a)))))); +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 16-bit integers, and store the results in dst. Note: this intrinsic +// will generate 0x7FFF, rather than 0x8000, for input values between 0x7FFF and +// 0x7FFFFFFF. +// +// FOR j := 0 to 3 +// i := 16*j +// k := 32*j +// IF a[k+31:k] >= FP32(0x7FFF) && a[k+31:k] <= FP32(0x7FFFFFFF) +// dst[i+15:i] := 0x7FFF +// ELSE +// dst[i+15:i] := Convert_FP32_To_Int16(a[k+31:k]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pi16 +FORCE_INLINE __m64 _mm_cvtps_pi16(__m128 a) +{ + const __m128 i16Min = _mm_set_ps1((float) INT16_MIN); + const __m128 i16Max = _mm_set_ps1((float) INT16_MAX); + const __m128 i32Max = _mm_set_ps1((float) INT32_MAX); + const __m128i maxMask = _mm_castps_si128( + _mm_and_ps(_mm_cmpge_ps(a, i16Max), _mm_cmple_ps(a, i32Max))); + const __m128i betweenMask = _mm_castps_si128( + _mm_and_ps(_mm_cmpgt_ps(a, i16Min), _mm_cmplt_ps(a, i16Max))); + const __m128i minMask = _mm_cmpeq_epi32(_mm_or_si128(maxMask, betweenMask), + _mm_setzero_si128()); + __m128i max = _mm_and_si128(maxMask, _mm_set1_epi32(INT16_MAX)); + __m128i min = _mm_and_si128(minMask, _mm_set1_epi32(INT16_MIN)); + __m128i cvt = _mm_and_si128(betweenMask, _mm_cvtps_epi32(a)); + __m128i res32 = _mm_or_si128(_mm_or_si128(max, min), cvt); + return vreinterpret_m64_s16(vmovn_s32(vreinterpretq_s32_m128i(res32))); +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 32-bit integers, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// dst[i+31:i] := Convert_FP32_To_Int32(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pi32 +#define _mm_cvtps_pi32(a) _mm_cvt_ps2pi(a) + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 8-bit integers, and store the results in lower 4 elements of dst. +// Note: this intrinsic will generate 0x7F, rather than 0x80, for input values +// between 0x7F and 0x7FFFFFFF. +// +// FOR j := 0 to 3 +// i := 8*j +// k := 32*j +// IF a[k+31:k] >= FP32(0x7F) && a[k+31:k] <= FP32(0x7FFFFFFF) +// dst[i+7:i] := 0x7F +// ELSE +// dst[i+7:i] := Convert_FP32_To_Int8(a[k+31:k]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pi8 +FORCE_INLINE __m64 _mm_cvtps_pi8(__m128 a) +{ + const __m128 i8Min = _mm_set_ps1((float) INT8_MIN); + const __m128 i8Max = _mm_set_ps1((float) INT8_MAX); + const __m128 i32Max = _mm_set_ps1((float) INT32_MAX); + const __m128i maxMask = _mm_castps_si128( + _mm_and_ps(_mm_cmpge_ps(a, i8Max), _mm_cmple_ps(a, i32Max))); + const __m128i betweenMask = _mm_castps_si128( + _mm_and_ps(_mm_cmpgt_ps(a, i8Min), _mm_cmplt_ps(a, i8Max))); + const __m128i minMask = _mm_cmpeq_epi32(_mm_or_si128(maxMask, betweenMask), + _mm_setzero_si128()); + __m128i max = _mm_and_si128(maxMask, _mm_set1_epi32(INT8_MAX)); + __m128i min = _mm_and_si128(minMask, _mm_set1_epi32(INT8_MIN)); + __m128i cvt = _mm_and_si128(betweenMask, _mm_cvtps_epi32(a)); + __m128i res32 = _mm_or_si128(_mm_or_si128(max, min), cvt); + int16x4_t res16 = vmovn_s32(vreinterpretq_s32_m128i(res32)); + int8x8_t res8 = vmovn_s16(vcombine_s16(res16, res16)); + static const uint32_t bitMask[2] = {0xFFFFFFFF, 0}; + int8x8_t mask = vreinterpret_s8_u32(vld1_u32(bitMask)); + + return vreinterpret_m64_s8(vorr_s8(vand_s8(mask, res8), vdup_n_s8(0))); +} + +// Convert packed unsigned 16-bit integers in a to packed single-precision +// (32-bit) floating-point elements, and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// m := j*32 +// dst[m+31:m] := Convert_UInt16_To_FP32(a[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpu16_ps +FORCE_INLINE __m128 _mm_cvtpu16_ps(__m64 a) +{ + return vreinterpretq_m128_f32( + vcvtq_f32_u32(vmovl_u16(vreinterpret_u16_m64(a)))); +} + +// Convert the lower packed unsigned 8-bit integers in a to packed +// single-precision (32-bit) floating-point elements, and store the results in +// dst. +// +// FOR j := 0 to 3 +// i := j*8 +// m := j*32 +// dst[m+31:m] := Convert_UInt8_To_FP32(a[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpu8_ps +FORCE_INLINE __m128 _mm_cvtpu8_ps(__m64 a) +{ + return vreinterpretq_m128_f32(vcvtq_f32_u32( + vmovl_u16(vget_low_u16(vmovl_u8(vreinterpret_u8_m64(a)))))); +} + +// Convert the signed 32-bit integer b to a single-precision (32-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper 3 packed elements from a to the upper elements of dst. +// +// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi32_ss +#define _mm_cvtsi32_ss(a, b) _mm_cvt_si2ss(a, b) + +// Convert the signed 64-bit integer b to a single-precision (32-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper 3 packed elements from a to the upper elements of dst. +// +// dst[31:0] := Convert_Int64_To_FP32(b[63:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi64_ss +FORCE_INLINE __m128 _mm_cvtsi64_ss(__m128 a, int64_t b) +{ + return vreinterpretq_m128_f32( + vsetq_lane_f32((float) b, vreinterpretq_f32_m128(a), 0)); +} + +// Copy the lower single-precision (32-bit) floating-point element of a to dst. +// +// dst[31:0] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_f32 +FORCE_INLINE float _mm_cvtss_f32(__m128 a) +{ + return vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); +} + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 32-bit integer, and store the result in dst. +// +// dst[31:0] := Convert_FP32_To_Int32(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_si32 +#define _mm_cvtss_si32(a) _mm_cvt_ss2si(a) + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 64-bit integer, and store the result in dst. +// +// dst[63:0] := Convert_FP32_To_Int64(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_si64 +FORCE_INLINE int64_t _mm_cvtss_si64(__m128 a) +{ +#if defined(__aarch64__) || defined(__ARM_FEATURE_DIRECTED_ROUNDING) + return (int64_t) vgetq_lane_f32(vrndiq_f32(vreinterpretq_f32_m128(a)), 0); +#else + float32_t data = vgetq_lane_f32( + vreinterpretq_f32_m128(_mm_round_ps(a, _MM_FROUND_CUR_DIRECTION)), 0); + return (int64_t) data; +#endif +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 32-bit integers with truncation, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// dst[i+31:i] := Convert_FP32_To_Int32_Truncate(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtt_ps2pi +FORCE_INLINE __m64 _mm_cvtt_ps2pi(__m128 a) +{ + return vreinterpret_m64_s32( + vget_low_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a)))); +} + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 32-bit integer with truncation, and store the result in dst. +// +// dst[31:0] := Convert_FP32_To_Int32_Truncate(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtt_ss2si +FORCE_INLINE int _mm_cvtt_ss2si(__m128 a) +{ + return vgetq_lane_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a)), 0); +} + +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed 32-bit integers with truncation, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// dst[i+31:i] := Convert_FP32_To_Int32_Truncate(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttps_pi32 +#define _mm_cvttps_pi32(a) _mm_cvtt_ps2pi(a) + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 32-bit integer with truncation, and store the result in dst. +// +// dst[31:0] := Convert_FP32_To_Int32_Truncate(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttss_si32 +#define _mm_cvttss_si32(a) _mm_cvtt_ss2si(a) + +// Convert the lower single-precision (32-bit) floating-point element in a to a +// 64-bit integer with truncation, and store the result in dst. +// +// dst[63:0] := Convert_FP32_To_Int64_Truncate(a[31:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttss_si64 +FORCE_INLINE int64_t _mm_cvttss_si64(__m128 a) +{ + return (int64_t) vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); +} + +// Divides the four single-precision, floating-point values of a and b. +// +// r0 := a0 / b0 +// r1 := a1 / b1 +// r2 := a2 / b2 +// r3 := a3 / b3 +// +// https://msdn.microsoft.com/en-us/library/edaw8147(v=vs.100).aspx +FORCE_INLINE __m128 _mm_div_ps(__m128 a, __m128 b) +{ +#if defined(__aarch64__) && !SSE2NEON_PRECISE_DIV + return vreinterpretq_m128_f32( + vdivq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#else + float32x4_t recip = vrecpeq_f32(vreinterpretq_f32_m128(b)); + recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(b))); +#if SSE2NEON_PRECISE_DIV + // Additional Netwon-Raphson iteration for accuracy + recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(b))); +#endif + return vreinterpretq_m128_f32(vmulq_f32(vreinterpretq_f32_m128(a), recip)); +#endif +} + +// Divides the scalar single-precision floating point value of a by b. +// https://msdn.microsoft.com/en-us/library/4y73xa49(v=vs.100).aspx +FORCE_INLINE __m128 _mm_div_ss(__m128 a, __m128 b) +{ + float32_t value = + vgetq_lane_f32(vreinterpretq_f32_m128(_mm_div_ps(a, b)), 0); + return vreinterpretq_m128_f32( + vsetq_lane_f32(value, vreinterpretq_f32_m128(a), 0)); +} + +// Extract a 16-bit integer from a, selected with imm8, and store the result in +// the lower element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_extract_pi16 +#define _mm_extract_pi16(a, imm) \ + (int32_t) vget_lane_u16(vreinterpret_u16_m64(a), (imm)) + +// Free aligned memory that was allocated with _mm_malloc. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_free +FORCE_INLINE void _mm_free(void *addr) +{ + free(addr); +} + +// Macro: Get the flush zero bits from the MXCSR control and status register. +// The flush zero may contain any of the following flags: _MM_FLUSH_ZERO_ON or +// _MM_FLUSH_ZERO_OFF +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_MM_GET_FLUSH_ZERO_MODE +FORCE_INLINE unsigned int _sse2neon_mm_get_flush_zero_mode() +{ + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + __asm__ __volatile__("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + __asm__ __volatile__("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + return r.field.bit24 ? _MM_FLUSH_ZERO_ON : _MM_FLUSH_ZERO_OFF; +} + +// Macro: Get the rounding mode bits from the MXCSR control and status register. +// The rounding mode may contain any of the following flags: _MM_ROUND_NEAREST, +// _MM_ROUND_DOWN, _MM_ROUND_UP, _MM_ROUND_TOWARD_ZERO +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_MM_GET_ROUNDING_MODE +FORCE_INLINE unsigned int _MM_GET_ROUNDING_MODE() +{ + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + __asm__ __volatile__("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + __asm__ __volatile__("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + if (r.field.bit22) { + return r.field.bit23 ? _MM_ROUND_TOWARD_ZERO : _MM_ROUND_UP; + } else { + return r.field.bit23 ? _MM_ROUND_DOWN : _MM_ROUND_NEAREST; + } +} + +// Copy a to dst, and insert the 16-bit integer i into dst at the location +// specified by imm8. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_insert_pi16 +#define _mm_insert_pi16(a, b, imm) \ + __extension__({ \ + vreinterpret_m64_s16( \ + vset_lane_s16((b), vreinterpret_s16_m64(a), (imm))); \ + }) + +// Loads four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/vstudio/zzd50xxt(v=vs.100).aspx +FORCE_INLINE __m128 _mm_load_ps(const float *p) +{ + return vreinterpretq_m128_f32(vld1q_f32(p)); +} + +// Load a single-precision (32-bit) floating-point element from memory into all +// elements of dst. +// +// dst[31:0] := MEM[mem_addr+31:mem_addr] +// dst[63:32] := MEM[mem_addr+31:mem_addr] +// dst[95:64] := MEM[mem_addr+31:mem_addr] +// dst[127:96] := MEM[mem_addr+31:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_ps1 +#define _mm_load_ps1 _mm_load1_ps + +// Loads an single - precision, floating - point value into the low word and +// clears the upper three words. +// https://msdn.microsoft.com/en-us/library/548bb9h4%28v=vs.90%29.aspx +FORCE_INLINE __m128 _mm_load_ss(const float *p) +{ + return vreinterpretq_m128_f32(vsetq_lane_f32(*p, vdupq_n_f32(0), 0)); +} + +// Loads a single single-precision, floating-point value, copying it into all +// four words +// https://msdn.microsoft.com/en-us/library/vstudio/5cdkf716(v=vs.100).aspx +FORCE_INLINE __m128 _mm_load1_ps(const float *p) +{ + return vreinterpretq_m128_f32(vld1q_dup_f32(p)); +} + +// Sets the upper two single-precision, floating-point values with 64 +// bits of data loaded from the address p; the lower two values are passed +// through from a. +// +// r0 := a0 +// r1 := a1 +// r2 := *p0 +// r3 := *p1 +// +// https://msdn.microsoft.com/en-us/library/w92wta0x(v%3dvs.100).aspx +FORCE_INLINE __m128 _mm_loadh_pi(__m128 a, __m64 const *p) +{ + return vreinterpretq_m128_f32( + vcombine_f32(vget_low_f32(a), vld1_f32((const float32_t *) p))); +} + +// Sets the lower two single-precision, floating-point values with 64 +// bits of data loaded from the address p; the upper two values are passed +// through from a. +// +// Return Value +// r0 := *p0 +// r1 := *p1 +// r2 := a2 +// r3 := a3 +// +// https://msdn.microsoft.com/en-us/library/s57cyak2(v=vs.100).aspx +FORCE_INLINE __m128 _mm_loadl_pi(__m128 a, __m64 const *p) +{ + return vreinterpretq_m128_f32( + vcombine_f32(vld1_f32((const float32_t *) p), vget_high_f32(a))); +} + +// Load 4 single-precision (32-bit) floating-point elements from memory into dst +// in reverse order. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// dst[31:0] := MEM[mem_addr+127:mem_addr+96] +// dst[63:32] := MEM[mem_addr+95:mem_addr+64] +// dst[95:64] := MEM[mem_addr+63:mem_addr+32] +// dst[127:96] := MEM[mem_addr+31:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadr_ps +FORCE_INLINE __m128 _mm_loadr_ps(const float *p) +{ + float32x4_t v = vrev64q_f32(vld1q_f32(p)); + return vreinterpretq_m128_f32(vextq_f32(v, v, 2)); +} + +// Loads four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/x1b16s7z%28v=vs.90%29.aspx +FORCE_INLINE __m128 _mm_loadu_ps(const float *p) +{ + // for neon, alignment doesn't matter, so _mm_load_ps and _mm_loadu_ps are + // equivalent for neon + return vreinterpretq_m128_f32(vld1q_f32(p)); +} + +// Load unaligned 16-bit integer from memory into the first element of dst. +// +// dst[15:0] := MEM[mem_addr+15:mem_addr] +// dst[MAX:16] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_si16 +FORCE_INLINE __m128i _mm_loadu_si16(const void *p) +{ + return vreinterpretq_m128i_s16( + vsetq_lane_s16(*(const int16_t *) p, vdupq_n_s16(0), 0)); +} + +// Load unaligned 64-bit integer from memory into the first element of dst. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[MAX:64] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_si64 +FORCE_INLINE __m128i _mm_loadu_si64(const void *p) +{ + return vreinterpretq_m128i_s64( + vcombine_s64(vld1_s64((const int64_t *) p), vdup_n_s64(0))); +} + +// Allocate aligned blocks of memory. +// https://software.intel.com/en-us/ +// cpp-compiler-developer-guide-and-reference-allocating-and-freeing-aligned-memory-blocks +FORCE_INLINE void *_mm_malloc(size_t size, size_t align) +{ + void *ptr; + if (align == 1) + return malloc(size); + if (align == 2 || (sizeof(void *) == 8 && align == 4)) + align = sizeof(void *); + if (!posix_memalign(&ptr, align, size)) + return ptr; + return NULL; +} + +// Conditionally store 8-bit integer elements from a into memory using mask +// (elements are not stored when the highest bit is not set in the corresponding +// element) and a non-temporal memory hint. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_maskmove_si64 +FORCE_INLINE void _mm_maskmove_si64(__m64 a, __m64 mask, char *mem_addr) +{ + int8x8_t shr_mask = vshr_n_s8(vreinterpret_s8_m64(mask), 7); + __m128 b = _mm_load_ps((const float *) mem_addr); + int8x8_t masked = + vbsl_s8(vreinterpret_u8_s8(shr_mask), vreinterpret_s8_m64(a), + vreinterpret_s8_u64(vget_low_u64(vreinterpretq_u64_m128(b)))); + vst1_s8((int8_t *) mem_addr, masked); +} + +// Conditionally store 8-bit integer elements from a into memory using mask +// (elements are not stored when the highest bit is not set in the corresponding +// element) and a non-temporal memory hint. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_maskmovq +#define _m_maskmovq(a, mask, mem_addr) _mm_maskmove_si64(a, mask, mem_addr) + +// Compare packed signed 16-bit integers in a and b, and store packed maximum +// values in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := MAX(a[i+15:i], b[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pi16 +FORCE_INLINE __m64 _mm_max_pi16(__m64 a, __m64 b) +{ + return vreinterpret_m64_s16( + vmax_s16(vreinterpret_s16_m64(a), vreinterpret_s16_m64(b))); +} + +// Computes the maximums of the four single-precision, floating-point values of +// a and b. +// https://msdn.microsoft.com/en-us/library/vstudio/ff5d607a(v=vs.100).aspx +FORCE_INLINE __m128 _mm_max_ps(__m128 a, __m128 b) +{ +#if SSE2NEON_PRECISE_MINMAX + float32x4_t _a = vreinterpretq_f32_m128(a); + float32x4_t _b = vreinterpretq_f32_m128(b); + return vreinterpretq_m128_f32(vbslq_f32(vcgtq_f32(_a, _b), _a, _b)); +#else + return vreinterpretq_m128_f32( + vmaxq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#endif +} + +// Compare packed unsigned 8-bit integers in a and b, and store packed maximum +// values in dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := MAX(a[i+7:i], b[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pu8 +FORCE_INLINE __m64 _mm_max_pu8(__m64 a, __m64 b) +{ + return vreinterpret_m64_u8( + vmax_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); +} + +// Computes the maximum of the two lower scalar single-precision floating point +// values of a and b. +// https://msdn.microsoft.com/en-us/library/s6db5esz(v=vs.100).aspx +FORCE_INLINE __m128 _mm_max_ss(__m128 a, __m128 b) +{ + float32_t value = vgetq_lane_f32(_mm_max_ps(a, b), 0); + return vreinterpretq_m128_f32( + vsetq_lane_f32(value, vreinterpretq_f32_m128(a), 0)); +} + +// Compare packed signed 16-bit integers in a and b, and store packed minimum +// values in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := MIN(a[i+15:i], b[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pi16 +FORCE_INLINE __m64 _mm_min_pi16(__m64 a, __m64 b) +{ + return vreinterpret_m64_s16( + vmin_s16(vreinterpret_s16_m64(a), vreinterpret_s16_m64(b))); +} + +// Computes the minima of the four single-precision, floating-point values of a +// and b. +// https://msdn.microsoft.com/en-us/library/vstudio/wh13kadz(v=vs.100).aspx +FORCE_INLINE __m128 _mm_min_ps(__m128 a, __m128 b) +{ +#if SSE2NEON_PRECISE_MINMAX + float32x4_t _a = vreinterpretq_f32_m128(a); + float32x4_t _b = vreinterpretq_f32_m128(b); + return vreinterpretq_m128_f32(vbslq_f32(vcltq_f32(_a, _b), _a, _b)); +#else + return vreinterpretq_m128_f32( + vminq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#endif +} + +// Compare packed unsigned 8-bit integers in a and b, and store packed minimum +// values in dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := MIN(a[i+7:i], b[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pu8 +FORCE_INLINE __m64 _mm_min_pu8(__m64 a, __m64 b) +{ + return vreinterpret_m64_u8( + vmin_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); +} + +// Computes the minimum of the two lower scalar single-precision floating point +// values of a and b. +// https://msdn.microsoft.com/en-us/library/0a9y7xaa(v=vs.100).aspx +FORCE_INLINE __m128 _mm_min_ss(__m128 a, __m128 b) +{ + float32_t value = vgetq_lane_f32(_mm_min_ps(a, b), 0); + return vreinterpretq_m128_f32( + vsetq_lane_f32(value, vreinterpretq_f32_m128(a), 0)); +} + +// Sets the low word to the single-precision, floating-point value of b +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/35hdzazd(v=vs.100) +FORCE_INLINE __m128 _mm_move_ss(__m128 a, __m128 b) +{ + return vreinterpretq_m128_f32( + vsetq_lane_f32(vgetq_lane_f32(vreinterpretq_f32_m128(b), 0), + vreinterpretq_f32_m128(a), 0)); +} + +// Moves the upper two values of B into the lower two values of A. +// +// r3 := a3 +// r2 := a2 +// r1 := b3 +// r0 := b2 +FORCE_INLINE __m128 _mm_movehl_ps(__m128 __A, __m128 __B) +{ + float32x2_t a32 = vget_high_f32(vreinterpretq_f32_m128(__A)); + float32x2_t b32 = vget_high_f32(vreinterpretq_f32_m128(__B)); + return vreinterpretq_m128_f32(vcombine_f32(b32, a32)); +} + +// Moves the lower two values of B into the upper two values of A. +// +// r3 := b1 +// r2 := b0 +// r1 := a1 +// r0 := a0 +FORCE_INLINE __m128 _mm_movelh_ps(__m128 __A, __m128 __B) +{ + float32x2_t a10 = vget_low_f32(vreinterpretq_f32_m128(__A)); + float32x2_t b10 = vget_low_f32(vreinterpretq_f32_m128(__B)); + return vreinterpretq_m128_f32(vcombine_f32(a10, b10)); +} + +// Create mask from the most significant bit of each 8-bit element in a, and +// store the result in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movemask_pi8 +FORCE_INLINE int _mm_movemask_pi8(__m64 a) +{ + uint8x8_t input = vreinterpret_u8_m64(a); +#if defined(__aarch64__) + static const int8x8_t shift = {0, 1, 2, 3, 4, 5, 6, 7}; + uint8x8_t tmp = vshr_n_u8(input, 7); + return vaddv_u8(vshl_u8(tmp, shift)); +#else + // Refer the implementation of `_mm_movemask_epi8` + uint16x4_t high_bits = vreinterpret_u16_u8(vshr_n_u8(input, 7)); + uint32x2_t paired16 = + vreinterpret_u32_u16(vsra_n_u16(high_bits, high_bits, 7)); + uint8x8_t paired32 = + vreinterpret_u8_u32(vsra_n_u32(paired16, paired16, 14)); + return vget_lane_u8(paired32, 0) | ((int) vget_lane_u8(paired32, 4) << 4); +#endif +} + +// NEON does not provide this method +// Creates a 4-bit mask from the most significant bits of the four +// single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/vstudio/4490ys29(v=vs.100).aspx +FORCE_INLINE int _mm_movemask_ps(__m128 a) +{ + uint32x4_t input = vreinterpretq_u32_m128(a); +#if defined(__aarch64__) + static const int32x4_t shift = {0, 1, 2, 3}; + uint32x4_t tmp = vshrq_n_u32(input, 31); + return vaddvq_u32(vshlq_u32(tmp, shift)); +#else + // Uses the exact same method as _mm_movemask_epi8, see that for details. + // Shift out everything but the sign bits with a 32-bit unsigned shift + // right. + uint64x2_t high_bits = vreinterpretq_u64_u32(vshrq_n_u32(input, 31)); + // Merge the two pairs together with a 64-bit unsigned shift right + add. + uint8x16_t paired = + vreinterpretq_u8_u64(vsraq_n_u64(high_bits, high_bits, 31)); + // Extract the result. + return vgetq_lane_u8(paired, 0) | (vgetq_lane_u8(paired, 8) << 2); +#endif +} + +// Multiplies the four single-precision, floating-point values of a and b. +// +// r0 := a0 * b0 +// r1 := a1 * b1 +// r2 := a2 * b2 +// r3 := a3 * b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/22kbk6t9(v=vs.100).aspx +FORCE_INLINE __m128 _mm_mul_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_f32( + vmulq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Multiply the lower single-precision (32-bit) floating-point element in a and +// b, store the result in the lower element of dst, and copy the upper 3 packed +// elements from a to the upper elements of dst. +// +// dst[31:0] := a[31:0] * b[31:0] +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mul_ss +FORCE_INLINE __m128 _mm_mul_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_mul_ps(a, b)); +} + +// Multiply the packed unsigned 16-bit integers in a and b, producing +// intermediate 32-bit integers, and store the high 16 bits of the intermediate +// integers in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mulhi_pu16 +FORCE_INLINE __m64 _mm_mulhi_pu16(__m64 a, __m64 b) +{ + return vreinterpret_m64_u16(vshrn_n_u32( + vmull_u16(vreinterpret_u16_m64(a), vreinterpret_u16_m64(b)), 16)); +} + +// Computes the bitwise OR of the four single-precision, floating-point values +// of a and b. +// https://msdn.microsoft.com/en-us/library/vstudio/7ctdsyy0(v=vs.100).aspx +FORCE_INLINE __m128 _mm_or_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_s32( + vorrq_s32(vreinterpretq_s32_m128(a), vreinterpretq_s32_m128(b))); +} + +// Average packed unsigned 8-bit integers in a and b, and store the results in +// dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := (a[i+7:i] + b[i+7:i] + 1) >> 1 +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pavgb +#define _m_pavgb(a, b) _mm_avg_pu8(a, b) + +// Average packed unsigned 16-bit integers in a and b, and store the results in +// dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := (a[i+15:i] + b[i+15:i] + 1) >> 1 +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pavgw +#define _m_pavgw(a, b) _mm_avg_pu16(a, b) + +// Extract a 16-bit integer from a, selected with imm8, and store the result in +// the lower element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pextrw +#define _m_pextrw(a, imm) _mm_extract_pi16(a, imm) + +// Copy a to dst, and insert the 16-bit integer i into dst at the location +// specified by imm8. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=m_pinsrw +#define _m_pinsrw(a, i, imm) _mm_insert_pi16(a, i, imm) + +// Compare packed signed 16-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmaxsw +#define _m_pmaxsw(a, b) _mm_max_pi16(a, b) + +// Compare packed unsigned 8-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmaxub +#define _m_pmaxub(a, b) _mm_max_pu8(a, b) + +// Compare packed signed 16-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pminsw +#define _m_pminsw(a, b) _mm_min_pi16(a, b) + +// Compare packed unsigned 8-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pminub +#define _m_pminub(a, b) _mm_min_pu8(a, b) + +// Create mask from the most significant bit of each 8-bit element in a, and +// store the result in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmovmskb +#define _m_pmovmskb(a) _mm_movemask_pi8(a) + +// Multiply the packed unsigned 16-bit integers in a and b, producing +// intermediate 32-bit integers, and store the high 16 bits of the intermediate +// integers in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmulhuw +#define _m_pmulhuw(a, b) _mm_mulhi_pu16(a, b) + +// Loads one cache line of data from address p to a location closer to the +// processor. https://msdn.microsoft.com/en-us/library/84szxsww(v=vs.100).aspx +FORCE_INLINE void _mm_prefetch(const void *p, int i) +{ + (void) i; + __builtin_prefetch(p); +} + +// Compute the absolute differences of packed unsigned 8-bit integers in a and +// b, then horizontally sum each consecutive 8 differences to produce four +// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low +// 16 bits of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=m_psadbw +#define _m_psadbw(a, b) _mm_sad_pu8(a, b) + +// Shuffle 16-bit integers in a using the control in imm8, and store the results +// in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pshufw +#define _m_pshufw(a, imm) _mm_shuffle_pi16(a, imm) + +// Compute the approximate reciprocal of packed single-precision (32-bit) +// floating-point elements in a, and store the results in dst. The maximum +// relative error for this approximation is less than 1.5*2^-12. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_rcp_ps +FORCE_INLINE __m128 _mm_rcp_ps(__m128 in) +{ + float32x4_t recip = vrecpeq_f32(vreinterpretq_f32_m128(in)); + recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(in))); +#if SSE2NEON_PRECISE_DIV + // Additional Netwon-Raphson iteration for accuracy + recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(in))); +#endif + return vreinterpretq_m128_f32(recip); +} + +// Compute the approximate reciprocal of the lower single-precision (32-bit) +// floating-point element in a, store the result in the lower element of dst, +// and copy the upper 3 packed elements from a to the upper elements of dst. The +// maximum relative error for this approximation is less than 1.5*2^-12. +// +// dst[31:0] := (1.0 / a[31:0]) +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_rcp_ss +FORCE_INLINE __m128 _mm_rcp_ss(__m128 a) +{ + return _mm_move_ss(a, _mm_rcp_ps(a)); +} + +// Computes the approximations of the reciprocal square roots of the four +// single-precision floating point values of in. +// The current precision is 1% error. +// https://msdn.microsoft.com/en-us/library/22hfsh53(v=vs.100).aspx +FORCE_INLINE __m128 _mm_rsqrt_ps(__m128 in) +{ + float32x4_t out = vrsqrteq_f32(vreinterpretq_f32_m128(in)); +#if SSE2NEON_PRECISE_SQRT + // Additional Netwon-Raphson iteration for accuracy + out = vmulq_f32( + out, vrsqrtsq_f32(vmulq_f32(vreinterpretq_f32_m128(in), out), out)); + out = vmulq_f32( + out, vrsqrtsq_f32(vmulq_f32(vreinterpretq_f32_m128(in), out), out)); +#endif + return vreinterpretq_m128_f32(out); +} + +// Compute the approximate reciprocal square root of the lower single-precision +// (32-bit) floating-point element in a, store the result in the lower element +// of dst, and copy the upper 3 packed elements from a to the upper elements of +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_rsqrt_ss +FORCE_INLINE __m128 _mm_rsqrt_ss(__m128 in) +{ + return vsetq_lane_f32(vgetq_lane_f32(_mm_rsqrt_ps(in), 0), in, 0); +} + +// Compute the absolute differences of packed unsigned 8-bit integers in a and +// b, then horizontally sum each consecutive 8 differences to produce four +// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low +// 16 bits of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sad_pu8 +FORCE_INLINE __m64 _mm_sad_pu8(__m64 a, __m64 b) +{ + uint64x1_t t = vpaddl_u32(vpaddl_u16( + vpaddl_u8(vabd_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))))); + return vreinterpret_m64_u16( + vset_lane_u16(vget_lane_u64(t, 0), vdup_n_u16(0), 0)); +} + +// Macro: Set the flush zero bits of the MXCSR control and status register to +// the value in unsigned 32-bit integer a. The flush zero may contain any of the +// following flags: _MM_FLUSH_ZERO_ON or _MM_FLUSH_ZERO_OFF +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_MM_SET_FLUSH_ZERO_MODE +FORCE_INLINE void _sse2neon_mm_set_flush_zero_mode(unsigned int flag) +{ + // AArch32 Advanced SIMD arithmetic always uses the Flush-to-zero setting, + // regardless of the value of the FZ bit. + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + __asm__ __volatile__("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + __asm__ __volatile__("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + r.field.bit24 = (flag & _MM_FLUSH_ZERO_MASK) == _MM_FLUSH_ZERO_ON; + +#if defined(__aarch64__) + __asm__ __volatile__("msr FPCR, %0" ::"r"(r)); /* write */ +#else + __asm__ __volatile__("vmsr FPSCR, %0" ::"r"(r)); /* write */ +#endif +} + +// Sets the four single-precision, floating-point values to the four inputs. +// https://msdn.microsoft.com/en-us/library/vstudio/afh0zf75(v=vs.100).aspx +FORCE_INLINE __m128 _mm_set_ps(float w, float z, float y, float x) +{ + float ALIGN_STRUCT(16) data[4] = {x, y, z, w}; + return vreinterpretq_m128_f32(vld1q_f32(data)); +} + +// Sets the four single-precision, floating-point values to w. +// https://msdn.microsoft.com/en-us/library/vstudio/2x1se8ha(v=vs.100).aspx +FORCE_INLINE __m128 _mm_set_ps1(float _w) +{ + return vreinterpretq_m128_f32(vdupq_n_f32(_w)); +} + +// Macro: Set the rounding mode bits of the MXCSR control and status register to +// the value in unsigned 32-bit integer a. The rounding mode may contain any of +// the following flags: _MM_ROUND_NEAREST, _MM_ROUND_DOWN, _MM_ROUND_UP, +// _MM_ROUND_TOWARD_ZERO +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_MM_SET_ROUNDING_MODE +FORCE_INLINE void _MM_SET_ROUNDING_MODE(int rounding) +{ + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + __asm__ __volatile__("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + __asm__ __volatile__("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + switch (rounding) { + case _MM_ROUND_TOWARD_ZERO: + r.field.bit22 = 1; + r.field.bit23 = 1; + break; + case _MM_ROUND_DOWN: + r.field.bit22 = 0; + r.field.bit23 = 1; + break; + case _MM_ROUND_UP: + r.field.bit22 = 1; + r.field.bit23 = 0; + break; + default: //_MM_ROUND_NEAREST + r.field.bit22 = 0; + r.field.bit23 = 0; + } + +#if defined(__aarch64__) + __asm__ __volatile__("msr FPCR, %0" ::"r"(r)); /* write */ +#else + __asm__ __volatile__("vmsr FPSCR, %0" ::"r"(r)); /* write */ +#endif +} + +// Copy single-precision (32-bit) floating-point element a to the lower element +// of dst, and zero the upper 3 elements. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_ss +FORCE_INLINE __m128 _mm_set_ss(float a) +{ + float ALIGN_STRUCT(16) data[4] = {a, 0, 0, 0}; + return vreinterpretq_m128_f32(vld1q_f32(data)); +} + +// Sets the four single-precision, floating-point values to w. +// +// r0 := r1 := r2 := r3 := w +// +// https://msdn.microsoft.com/en-us/library/vstudio/2x1se8ha(v=vs.100).aspx +FORCE_INLINE __m128 _mm_set1_ps(float _w) +{ + return vreinterpretq_m128_f32(vdupq_n_f32(_w)); +} + +// FIXME: _mm_setcsr() implementation supports changing the rounding mode only. +FORCE_INLINE void _mm_setcsr(unsigned int a) +{ + _MM_SET_ROUNDING_MODE(a); +} + +// FIXME: _mm_getcsr() implementation supports reading the rounding mode only. +FORCE_INLINE unsigned int _mm_getcsr() +{ + return _MM_GET_ROUNDING_MODE(); +} + +// Sets the four single-precision, floating-point values to the four inputs in +// reverse order. +// https://msdn.microsoft.com/en-us/library/vstudio/d2172ct3(v=vs.100).aspx +FORCE_INLINE __m128 _mm_setr_ps(float w, float z, float y, float x) +{ + float ALIGN_STRUCT(16) data[4] = {w, z, y, x}; + return vreinterpretq_m128_f32(vld1q_f32(data)); +} + +// Clears the four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/vstudio/tk1t2tbz(v=vs.100).aspx +FORCE_INLINE __m128 _mm_setzero_ps(void) +{ + return vreinterpretq_m128_f32(vdupq_n_f32(0)); +} + +// Shuffle 16-bit integers in a using the control in imm8, and store the results +// in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_pi16 +#if __has_builtin(__builtin_shufflevector) +#define _mm_shuffle_pi16(a, imm) \ + __extension__({ \ + vreinterpret_m64_s16(__builtin_shufflevector( \ + vreinterpret_s16_m64(a), vreinterpret_s16_m64(a), (imm & 0x3), \ + ((imm >> 2) & 0x3), ((imm >> 4) & 0x3), ((imm >> 6) & 0x3))); \ + }) +#else +#define _mm_shuffle_pi16(a, imm) \ + __extension__({ \ + int16x4_t ret; \ + ret = \ + vmov_n_s16(vget_lane_s16(vreinterpret_s16_m64(a), (imm) & (0x3))); \ + ret = vset_lane_s16( \ + vget_lane_s16(vreinterpret_s16_m64(a), ((imm) >> 2) & 0x3), ret, \ + 1); \ + ret = vset_lane_s16( \ + vget_lane_s16(vreinterpret_s16_m64(a), ((imm) >> 4) & 0x3), ret, \ + 2); \ + ret = vset_lane_s16( \ + vget_lane_s16(vreinterpret_s16_m64(a), ((imm) >> 6) & 0x3), ret, \ + 3); \ + vreinterpret_m64_s16(ret); \ + }) +#endif + +// Guarantees that every preceding store is globally visible before any +// subsequent store. +// https://msdn.microsoft.com/en-us/library/5h2w73d1%28v=vs.90%29.aspx +FORCE_INLINE void _mm_sfence(void) +{ + __sync_synchronize(); +} + // FORCE_INLINE __m128 _mm_shuffle_ps(__m128 a, __m128 b, __constrange(0,255) // int imm) #if __has_builtin(__builtin_shufflevector) @@ -2015,737 +2622,2127 @@ FORCE_INLINE __m128 _mm_shuffle_ps_2032(__m128 a, __m128 b) }) #endif -// Takes the upper 64 bits of a and places it in the low end of the result -// Takes the lower 64 bits of a and places it into the high end of the result. -FORCE_INLINE __m128i _mm_shuffle_epi_1032(__m128i a) +// Computes the approximations of square roots of the four single-precision, +// floating-point values of a. First computes reciprocal square roots and then +// reciprocals of the four values. +// +// r0 := sqrt(a0) +// r1 := sqrt(a1) +// r2 := sqrt(a2) +// r3 := sqrt(a3) +// +// https://msdn.microsoft.com/en-us/library/vstudio/8z67bwwk(v=vs.100).aspx +FORCE_INLINE __m128 _mm_sqrt_ps(__m128 in) { - int32x2_t a32 = vget_high_s32(vreinterpretq_s32_m128i(a)); - int32x2_t a10 = vget_low_s32(vreinterpretq_s32_m128i(a)); - return vreinterpretq_m128i_s32(vcombine_s32(a32, a10)); -} +#if SSE2NEON_PRECISE_SQRT + float32x4_t recip = vrsqrteq_f32(vreinterpretq_f32_m128(in)); -// takes the lower two 32-bit values from a and swaps them and places in low end -// of result takes the higher two 32 bit values from a and swaps them and places -// in high end of result. -FORCE_INLINE __m128i _mm_shuffle_epi_2301(__m128i a) -{ - int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); - int32x2_t a23 = vrev64_s32(vget_high_s32(vreinterpretq_s32_m128i(a))); - return vreinterpretq_m128i_s32(vcombine_s32(a01, a23)); -} + // Test for vrsqrteq_f32(0) -> positive infinity case. + // Change to zero, so that s * 1/sqrt(s) result is zero too. + const uint32x4_t pos_inf = vdupq_n_u32(0x7F800000); + const uint32x4_t div_by_zero = + vceqq_u32(pos_inf, vreinterpretq_u32_f32(recip)); + recip = vreinterpretq_f32_u32( + vandq_u32(vmvnq_u32(div_by_zero), vreinterpretq_u32_f32(recip))); -// rotates the least significant 32 bits into the most signficant 32 bits, and -// shifts the rest down -FORCE_INLINE __m128i _mm_shuffle_epi_0321(__m128i a) -{ - return vreinterpretq_m128i_s32( - vextq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(a), 1)); -} + // Additional Netwon-Raphson iteration for accuracy + recip = vmulq_f32( + vrsqrtsq_f32(vmulq_f32(recip, recip), vreinterpretq_f32_m128(in)), + recip); + recip = vmulq_f32( + vrsqrtsq_f32(vmulq_f32(recip, recip), vreinterpretq_f32_m128(in)), + recip); -// rotates the most significant 32 bits into the least signficant 32 bits, and -// shifts the rest up -FORCE_INLINE __m128i _mm_shuffle_epi_2103(__m128i a) -{ - return vreinterpretq_m128i_s32( - vextq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(a), 3)); -} - -// gets the lower 64 bits of a, and places it in the upper 64 bits -// gets the lower 64 bits of a and places it in the lower 64 bits -FORCE_INLINE __m128i _mm_shuffle_epi_1010(__m128i a) -{ - int32x2_t a10 = vget_low_s32(vreinterpretq_s32_m128i(a)); - return vreinterpretq_m128i_s32(vcombine_s32(a10, a10)); -} - -// gets the lower 64 bits of a, swaps the 0 and 1 elements, and places it in the -// lower 64 bits gets the lower 64 bits of a, and places it in the upper 64 bits -FORCE_INLINE __m128i _mm_shuffle_epi_1001(__m128i a) -{ - int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); - int32x2_t a10 = vget_low_s32(vreinterpretq_s32_m128i(a)); - return vreinterpretq_m128i_s32(vcombine_s32(a01, a10)); -} - -// gets the lower 64 bits of a, swaps the 0 and 1 elements and places it in the -// upper 64 bits gets the lower 64 bits of a, swaps the 0 and 1 elements, and -// places it in the lower 64 bits -FORCE_INLINE __m128i _mm_shuffle_epi_0101(__m128i a) -{ - int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); - return vreinterpretq_m128i_s32(vcombine_s32(a01, a01)); -} - -FORCE_INLINE __m128i _mm_shuffle_epi_2211(__m128i a) -{ - int32x2_t a11 = vdup_lane_s32(vget_low_s32(vreinterpretq_s32_m128i(a)), 1); - int32x2_t a22 = vdup_lane_s32(vget_high_s32(vreinterpretq_s32_m128i(a)), 0); - return vreinterpretq_m128i_s32(vcombine_s32(a11, a22)); -} - -FORCE_INLINE __m128i _mm_shuffle_epi_0122(__m128i a) -{ - int32x2_t a22 = vdup_lane_s32(vget_high_s32(vreinterpretq_s32_m128i(a)), 0); - int32x2_t a01 = vrev64_s32(vget_low_s32(vreinterpretq_s32_m128i(a))); - return vreinterpretq_m128i_s32(vcombine_s32(a22, a01)); -} - -FORCE_INLINE __m128i _mm_shuffle_epi_3332(__m128i a) -{ - int32x2_t a32 = vget_high_s32(vreinterpretq_s32_m128i(a)); - int32x2_t a33 = vdup_lane_s32(vget_high_s32(vreinterpretq_s32_m128i(a)), 1); - return vreinterpretq_m128i_s32(vcombine_s32(a32, a33)); -} - -// Shuffle packed 8-bit integers in a according to shuffle control mask in the -// corresponding 8-bit element of b, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_epi8 -FORCE_INLINE __m128i _mm_shuffle_epi8(__m128i a, __m128i b) -{ - int8x16_t tbl = vreinterpretq_s8_m128i(a); // input a - uint8x16_t idx = vreinterpretq_u8_m128i(b); // input b - uint8x16_t idx_masked = - vandq_u8(idx, vdupq_n_u8(0x8F)); // avoid using meaningless bits -#if defined(__aarch64__) - return vreinterpretq_m128i_s8(vqtbl1q_s8(tbl, idx_masked)); -#elif defined(__GNUC__) - int8x16_t ret; - // %e and %f represent the even and odd D registers - // respectively. - __asm__ __volatile__( - "vtbl.8 %e[ret], {%e[tbl], %f[tbl]}, %e[idx]\n" - "vtbl.8 %f[ret], {%e[tbl], %f[tbl]}, %f[idx]\n" - : [ret] "=&w"(ret) - : [tbl] "w"(tbl), [idx] "w"(idx_masked)); - return vreinterpretq_m128i_s8(ret); + // sqrt(s) = s * 1/sqrt(s) + return vreinterpretq_m128_f32(vmulq_f32(vreinterpretq_f32_m128(in), recip)); +#elif defined(__aarch64__) + return vreinterpretq_m128_f32(vsqrtq_f32(vreinterpretq_f32_m128(in))); #else - // use this line if testing on aarch64 - int8x8x2_t a_split = {vget_low_s8(tbl), vget_high_s8(tbl)}; + float32x4_t recipsq = vrsqrteq_f32(vreinterpretq_f32_m128(in)); + float32x4_t sq = vrecpeq_f32(recipsq); + return vreinterpretq_m128_f32(sq); +#endif +} + +// Computes the approximation of the square root of the scalar single-precision +// floating point value of in. +// https://msdn.microsoft.com/en-us/library/ahfsc22d(v=vs.100).aspx +FORCE_INLINE __m128 _mm_sqrt_ss(__m128 in) +{ + float32_t value = + vgetq_lane_f32(vreinterpretq_f32_m128(_mm_sqrt_ps(in)), 0); + return vreinterpretq_m128_f32( + vsetq_lane_f32(value, vreinterpretq_f32_m128(in), 0)); +} + +// Stores four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/vstudio/s3h4ay6y(v=vs.100).aspx +FORCE_INLINE void _mm_store_ps(float *p, __m128 a) +{ + vst1q_f32(p, vreinterpretq_f32_m128(a)); +} + +// Store the lower single-precision (32-bit) floating-point element from a into +// 4 contiguous elements in memory. mem_addr must be aligned on a 16-byte +// boundary or a general-protection exception may be generated. +// +// MEM[mem_addr+31:mem_addr] := a[31:0] +// MEM[mem_addr+63:mem_addr+32] := a[31:0] +// MEM[mem_addr+95:mem_addr+64] := a[31:0] +// MEM[mem_addr+127:mem_addr+96] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store_ps1 +FORCE_INLINE void _mm_store_ps1(float *p, __m128 a) +{ + float32_t a0 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); + vst1q_f32(p, vdupq_n_f32(a0)); +} + +// Stores the lower single - precision, floating - point value. +// https://msdn.microsoft.com/en-us/library/tzz10fbx(v=vs.100).aspx +FORCE_INLINE void _mm_store_ss(float *p, __m128 a) +{ + vst1q_lane_f32(p, vreinterpretq_f32_m128(a), 0); +} + +// Store the lower single-precision (32-bit) floating-point element from a into +// 4 contiguous elements in memory. mem_addr must be aligned on a 16-byte +// boundary or a general-protection exception may be generated. +// +// MEM[mem_addr+31:mem_addr] := a[31:0] +// MEM[mem_addr+63:mem_addr+32] := a[31:0] +// MEM[mem_addr+95:mem_addr+64] := a[31:0] +// MEM[mem_addr+127:mem_addr+96] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store1_ps +#define _mm_store1_ps _mm_store_ps1 + +// Stores the upper two single-precision, floating-point values of a to the +// address p. +// +// *p0 := a2 +// *p1 := a3 +// +// https://msdn.microsoft.com/en-us/library/a7525fs8(v%3dvs.90).aspx +FORCE_INLINE void _mm_storeh_pi(__m64 *p, __m128 a) +{ + *p = vreinterpret_m64_f32(vget_high_f32(a)); +} + +// Stores the lower two single-precision floating point values of a to the +// address p. +// +// *p0 := a0 +// *p1 := a1 +// +// https://msdn.microsoft.com/en-us/library/h54t98ks(v=vs.90).aspx +FORCE_INLINE void _mm_storel_pi(__m64 *p, __m128 a) +{ + *p = vreinterpret_m64_f32(vget_low_f32(a)); +} + +// Store 4 single-precision (32-bit) floating-point elements from a into memory +// in reverse order. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// MEM[mem_addr+31:mem_addr] := a[127:96] +// MEM[mem_addr+63:mem_addr+32] := a[95:64] +// MEM[mem_addr+95:mem_addr+64] := a[63:32] +// MEM[mem_addr+127:mem_addr+96] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storer_ps +FORCE_INLINE void _mm_storer_ps(float *p, __m128 a) +{ + float32x4_t tmp = vrev64q_f32(vreinterpretq_f32_m128(a)); + float32x4_t rev = vextq_f32(tmp, tmp, 2); + vst1q_f32(p, rev); +} + +// Stores four single-precision, floating-point values. +// https://msdn.microsoft.com/en-us/library/44e30x22(v=vs.100).aspx +FORCE_INLINE void _mm_storeu_ps(float *p, __m128 a) +{ + vst1q_f32(p, vreinterpretq_f32_m128(a)); +} + +// Stores 16-bits of integer data a at the address p. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si16 +FORCE_INLINE void _mm_storeu_si16(void *p, __m128i a) +{ + vst1q_lane_s16((int16_t *) p, vreinterpretq_s16_m128i(a), 0); +} + +// Stores 64-bits of integer data a at the address p. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si64 +FORCE_INLINE void _mm_storeu_si64(void *p, __m128i a) +{ + vst1q_lane_s64((int64_t *) p, vreinterpretq_s64_m128i(a), 0); +} + +// Store 64-bits of integer data from a into memory using a non-temporal memory +// hint. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_pi +FORCE_INLINE void _mm_stream_pi(__m64 *p, __m64 a) +{ + vst1_s64((int64_t *) p, vreinterpret_s64_m64(a)); +} + +// Store 128-bits (composed of 4 packed single-precision (32-bit) floating- +// point elements) from a into memory using a non-temporal memory hint. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_ps +FORCE_INLINE void _mm_stream_ps(float *p, __m128 a) +{ +#if __has_builtin(__builtin_nontemporal_store) + __builtin_nontemporal_store(a, (float32x4_t *) p); +#else + vst1q_f32(p, vreinterpretq_f32_m128(a)); +#endif +} + +// Subtracts the four single-precision, floating-point values of a and b. +// +// r0 := a0 - b0 +// r1 := a1 - b1 +// r2 := a2 - b2 +// r3 := a3 - b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/1zad2k61(v=vs.100).aspx +FORCE_INLINE __m128 _mm_sub_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_f32( + vsubq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +} + +// Subtract the lower single-precision (32-bit) floating-point element in b from +// the lower single-precision (32-bit) floating-point element in a, store the +// result in the lower element of dst, and copy the upper 3 packed elements from +// a to the upper elements of dst. +// +// dst[31:0] := a[31:0] - b[31:0] +// dst[127:32] := a[127:32] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_ss +FORCE_INLINE __m128 _mm_sub_ss(__m128 a, __m128 b) +{ + return _mm_move_ss(a, _mm_sub_ps(a, b)); +} + +// Macro: Transpose the 4x4 matrix formed by the 4 rows of single-precision +// (32-bit) floating-point elements in row0, row1, row2, and row3, and store the +// transposed matrix in these vectors (row0 now contains column 0, etc.). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=MM_TRANSPOSE4_PS +#define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \ + do { \ + float32x4x2_t ROW01 = vtrnq_f32(row0, row1); \ + float32x4x2_t ROW23 = vtrnq_f32(row2, row3); \ + row0 = vcombine_f32(vget_low_f32(ROW01.val[0]), \ + vget_low_f32(ROW23.val[0])); \ + row1 = vcombine_f32(vget_low_f32(ROW01.val[1]), \ + vget_low_f32(ROW23.val[1])); \ + row2 = vcombine_f32(vget_high_f32(ROW01.val[0]), \ + vget_high_f32(ROW23.val[0])); \ + row3 = vcombine_f32(vget_high_f32(ROW01.val[1]), \ + vget_high_f32(ROW23.val[1])); \ + } while (0) + +// according to the documentation, these intrinsics behave the same as the +// non-'u' versions. We'll just alias them here. +#define _mm_ucomieq_ss _mm_comieq_ss +#define _mm_ucomige_ss _mm_comige_ss +#define _mm_ucomigt_ss _mm_comigt_ss +#define _mm_ucomile_ss _mm_comile_ss +#define _mm_ucomilt_ss _mm_comilt_ss +#define _mm_ucomineq_ss _mm_comineq_ss + +// Return vector of type __m128i with undefined elements. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_undefined_si128 +FORCE_INLINE __m128i _mm_undefined_si128(void) +{ +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wuninitialized" +#endif + __m128i a; + return a; +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic pop +#endif +} + +// Return vector of type __m128 with undefined elements. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_undefined_ps +FORCE_INLINE __m128 _mm_undefined_ps(void) +{ +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wuninitialized" +#endif + __m128 a; + return a; +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic pop +#endif +} + +// Selects and interleaves the upper two single-precision, floating-point values +// from a and b. +// +// r0 := a2 +// r1 := b2 +// r2 := a3 +// r3 := b3 +// +// https://msdn.microsoft.com/en-us/library/skccxx7d%28v=vs.90%29.aspx +FORCE_INLINE __m128 _mm_unpackhi_ps(__m128 a, __m128 b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32( + vzip2q_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#else + float32x2_t a1 = vget_high_f32(vreinterpretq_f32_m128(a)); + float32x2_t b1 = vget_high_f32(vreinterpretq_f32_m128(b)); + float32x2x2_t result = vzip_f32(a1, b1); + return vreinterpretq_m128_f32(vcombine_f32(result.val[0], result.val[1])); +#endif +} + +// Selects and interleaves the lower two single-precision, floating-point values +// from a and b. +// +// r0 := a0 +// r1 := b0 +// r2 := a1 +// r3 := b1 +// +// https://msdn.microsoft.com/en-us/library/25st103b%28v=vs.90%29.aspx +FORCE_INLINE __m128 _mm_unpacklo_ps(__m128 a, __m128 b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32( + vzip1q_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#else + float32x2_t a1 = vget_low_f32(vreinterpretq_f32_m128(a)); + float32x2_t b1 = vget_low_f32(vreinterpretq_f32_m128(b)); + float32x2x2_t result = vzip_f32(a1, b1); + return vreinterpretq_m128_f32(vcombine_f32(result.val[0], result.val[1])); +#endif +} + +// Computes bitwise EXOR (exclusive-or) of the four single-precision, +// floating-point values of a and b. +// https://msdn.microsoft.com/en-us/library/ss6k3wk8(v=vs.100).aspx +FORCE_INLINE __m128 _mm_xor_ps(__m128 a, __m128 b) +{ + return vreinterpretq_m128_s32( + veorq_s32(vreinterpretq_s32_m128(a), vreinterpretq_s32_m128(b))); +} + +/* SSE2 */ + +// Adds the 8 signed or unsigned 16-bit integers in a to the 8 signed or +// unsigned 16-bit integers in b. +// https://msdn.microsoft.com/en-us/library/fceha5k4(v=vs.100).aspx +FORCE_INLINE __m128i _mm_add_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vaddq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Adds the 4 signed or unsigned 32-bit integers in a to the 4 signed or +// unsigned 32-bit integers in b. +// +// r0 := a0 + b0 +// r1 := a1 + b1 +// r2 := a2 + b2 +// r3 := a3 + b3 +// +// https://msdn.microsoft.com/en-us/library/vstudio/09xs4fkk(v=vs.100).aspx +FORCE_INLINE __m128i _mm_add_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vaddq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Adds the 4 signed or unsigned 64-bit integers in a to the 4 signed or +// unsigned 32-bit integers in b. +// https://msdn.microsoft.com/en-us/library/vstudio/09xs4fkk(v=vs.100).aspx +FORCE_INLINE __m128i _mm_add_epi64(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s64( + vaddq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b))); +} + +// Adds the 16 signed or unsigned 8-bit integers in a to the 16 signed or +// unsigned 8-bit integers in b. +// https://technet.microsoft.com/en-us/subscriptions/yc7tcyzs(v=vs.90) +FORCE_INLINE __m128i _mm_add_epi8(__m128i a, __m128i b) +{ return vreinterpretq_m128i_s8( - vcombine_s8(vtbl2_s8(a_split, vget_low_u8(idx_masked)), - vtbl2_s8(a_split, vget_high_u8(idx_masked)))); + vaddq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Add packed double-precision (64-bit) floating-point elements in a and b, and +// store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_add_pd +FORCE_INLINE __m128d _mm_add_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vaddq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + double *da = (double *) &a; + double *db = (double *) &b; + double c[2]; + c[0] = da[0] + db[0]; + c[1] = da[1] + db[1]; + return vld1q_f32((float32_t *) c); #endif } -// C equivalent: -// __m128i _mm_shuffle_epi32_default(__m128i a, -// __constrange(0, 255) int imm) { -// __m128i ret; -// ret[0] = a[imm & 0x3]; ret[1] = a[(imm >> 2) & 0x3]; -// ret[2] = a[(imm >> 4) & 0x03]; ret[3] = a[(imm >> 6) & 0x03]; -// return ret; -// } -#define _mm_shuffle_epi32_default(a, imm) \ - __extension__({ \ - int32x4_t ret; \ - ret = vmovq_n_s32( \ - vgetq_lane_s32(vreinterpretq_s32_m128i(a), (imm) & (0x3))); \ - ret = vsetq_lane_s32( \ - vgetq_lane_s32(vreinterpretq_s32_m128i(a), ((imm) >> 2) & 0x3), \ - ret, 1); \ - ret = vsetq_lane_s32( \ - vgetq_lane_s32(vreinterpretq_s32_m128i(a), ((imm) >> 4) & 0x3), \ - ret, 2); \ - ret = vsetq_lane_s32( \ - vgetq_lane_s32(vreinterpretq_s32_m128i(a), ((imm) >> 6) & 0x3), \ - ret, 3); \ - vreinterpretq_m128i_s32(ret); \ - }) - -// FORCE_INLINE __m128i _mm_shuffle_epi32_splat(__m128i a, __constrange(0,255) -// int imm) +// Add the lower double-precision (64-bit) floating-point element in a and b, +// store the result in the lower element of dst, and copy the upper element from +// a to the upper element of dst. +// +// dst[63:0] := a[63:0] + b[63:0] +// dst[127:64] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_add_sd +FORCE_INLINE __m128d _mm_add_sd(__m128d a, __m128d b) +{ #if defined(__aarch64__) -#define _mm_shuffle_epi32_splat(a, imm) \ - __extension__({ \ - vreinterpretq_m128i_s32( \ - vdupq_laneq_s32(vreinterpretq_s32_m128i(a), (imm))); \ - }) + return _mm_move_sd(a, _mm_add_pd(a, b)); #else -#define _mm_shuffle_epi32_splat(a, imm) \ - __extension__({ \ - vreinterpretq_m128i_s32( \ - vdupq_n_s32(vgetq_lane_s32(vreinterpretq_s32_m128i(a), (imm)))); \ - }) + double *da = (double *) &a; + double *db = (double *) &b; + double c[2]; + c[0] = da[0] + db[0]; + c[1] = da[1]; + return vld1q_f32((float32_t *) c); #endif +} -// Shuffles the 4 signed or unsigned 32-bit integers in a as specified by imm. -// https://msdn.microsoft.com/en-us/library/56f67xbk%28v=vs.90%29.aspx -// FORCE_INLINE __m128i _mm_shuffle_epi32(__m128i a, -// __constrange(0,255) int imm) -#if __has_builtin(__builtin_shufflevector) -#define _mm_shuffle_epi32(a, imm) \ - __extension__({ \ - int32x4_t _input = vreinterpretq_s32_m128i(a); \ - int32x4_t _shuf = __builtin_shufflevector( \ - _input, _input, (imm) & (0x3), ((imm) >> 2) & 0x3, \ - ((imm) >> 4) & 0x3, ((imm) >> 6) & 0x3); \ - vreinterpretq_m128i_s32(_shuf); \ - }) -#else // generic -#define _mm_shuffle_epi32(a, imm) \ - __extension__({ \ - __m128i ret; \ - switch (imm) { \ - case _MM_SHUFFLE(1, 0, 3, 2): \ - ret = _mm_shuffle_epi_1032((a)); \ - break; \ - case _MM_SHUFFLE(2, 3, 0, 1): \ - ret = _mm_shuffle_epi_2301((a)); \ - break; \ - case _MM_SHUFFLE(0, 3, 2, 1): \ - ret = _mm_shuffle_epi_0321((a)); \ - break; \ - case _MM_SHUFFLE(2, 1, 0, 3): \ - ret = _mm_shuffle_epi_2103((a)); \ - break; \ - case _MM_SHUFFLE(1, 0, 1, 0): \ - ret = _mm_shuffle_epi_1010((a)); \ - break; \ - case _MM_SHUFFLE(1, 0, 0, 1): \ - ret = _mm_shuffle_epi_1001((a)); \ - break; \ - case _MM_SHUFFLE(0, 1, 0, 1): \ - ret = _mm_shuffle_epi_0101((a)); \ - break; \ - case _MM_SHUFFLE(2, 2, 1, 1): \ - ret = _mm_shuffle_epi_2211((a)); \ - break; \ - case _MM_SHUFFLE(0, 1, 2, 2): \ - ret = _mm_shuffle_epi_0122((a)); \ - break; \ - case _MM_SHUFFLE(3, 3, 3, 2): \ - ret = _mm_shuffle_epi_3332((a)); \ - break; \ - case _MM_SHUFFLE(0, 0, 0, 0): \ - ret = _mm_shuffle_epi32_splat((a), 0); \ - break; \ - case _MM_SHUFFLE(1, 1, 1, 1): \ - ret = _mm_shuffle_epi32_splat((a), 1); \ - break; \ - case _MM_SHUFFLE(2, 2, 2, 2): \ - ret = _mm_shuffle_epi32_splat((a), 2); \ - break; \ - case _MM_SHUFFLE(3, 3, 3, 3): \ - ret = _mm_shuffle_epi32_splat((a), 3); \ - break; \ - default: \ - ret = _mm_shuffle_epi32_default((a), (imm)); \ - break; \ - } \ - ret; \ - }) -#endif - -// Shuffles the lower 4 signed or unsigned 16-bit integers in a as specified -// by imm. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/y41dkk37(v=vs.100) -// FORCE_INLINE __m128i _mm_shufflelo_epi16_function(__m128i a, -// __constrange(0,255) int -// imm) -#define _mm_shufflelo_epi16_function(a, imm) \ - __extension__({ \ - int16x8_t ret = vreinterpretq_s16_m128i(a); \ - int16x4_t lowBits = vget_low_s16(ret); \ - ret = vsetq_lane_s16(vget_lane_s16(lowBits, (imm) & (0x3)), ret, 0); \ - ret = vsetq_lane_s16(vget_lane_s16(lowBits, ((imm) >> 2) & 0x3), ret, \ - 1); \ - ret = vsetq_lane_s16(vget_lane_s16(lowBits, ((imm) >> 4) & 0x3), ret, \ - 2); \ - ret = vsetq_lane_s16(vget_lane_s16(lowBits, ((imm) >> 6) & 0x3), ret, \ - 3); \ - vreinterpretq_m128i_s16(ret); \ - }) - -// FORCE_INLINE __m128i _mm_shufflelo_epi16(__m128i a, -// __constrange(0,255) int imm) -#if __has_builtin(__builtin_shufflevector) -#define _mm_shufflelo_epi16(a, imm) \ - __extension__({ \ - int16x8_t _input = vreinterpretq_s16_m128i(a); \ - int16x8_t _shuf = __builtin_shufflevector( \ - _input, _input, ((imm) & (0x3)), (((imm) >> 2) & 0x3), \ - (((imm) >> 4) & 0x3), (((imm) >> 6) & 0x3), 4, 5, 6, 7); \ - vreinterpretq_m128i_s16(_shuf); \ - }) -#else // generic -#define _mm_shufflelo_epi16(a, imm) _mm_shufflelo_epi16_function((a), (imm)) -#endif - -// Shuffles the upper 4 signed or unsigned 16-bit integers in a as specified -// by imm. -// https://msdn.microsoft.com/en-us/library/13ywktbs(v=vs.100).aspx -// FORCE_INLINE __m128i _mm_shufflehi_epi16_function(__m128i a, -// __constrange(0,255) int -// imm) -#define _mm_shufflehi_epi16_function(a, imm) \ - __extension__({ \ - int16x8_t ret = vreinterpretq_s16_m128i(a); \ - int16x4_t highBits = vget_high_s16(ret); \ - ret = vsetq_lane_s16(vget_lane_s16(highBits, (imm) & (0x3)), ret, 4); \ - ret = vsetq_lane_s16(vget_lane_s16(highBits, ((imm) >> 2) & 0x3), ret, \ - 5); \ - ret = vsetq_lane_s16(vget_lane_s16(highBits, ((imm) >> 4) & 0x3), ret, \ - 6); \ - ret = vsetq_lane_s16(vget_lane_s16(highBits, ((imm) >> 6) & 0x3), ret, \ - 7); \ - vreinterpretq_m128i_s16(ret); \ - }) - -// FORCE_INLINE __m128i _mm_shufflehi_epi16(__m128i a, -// __constrange(0,255) int imm) -#if __has_builtin(__builtin_shufflevector) -#define _mm_shufflehi_epi16(a, imm) \ - __extension__({ \ - int16x8_t _input = vreinterpretq_s16_m128i(a); \ - int16x8_t _shuf = __builtin_shufflevector( \ - _input, _input, 0, 1, 2, 3, ((imm) & (0x3)) + 4, \ - (((imm) >> 2) & 0x3) + 4, (((imm) >> 4) & 0x3) + 4, \ - (((imm) >> 6) & 0x3) + 4); \ - vreinterpretq_m128i_s16(_shuf); \ - }) -#else // generic -#define _mm_shufflehi_epi16(a, imm) _mm_shufflehi_epi16_function((a), (imm)) -#endif - -// Shuffle double-precision (64-bit) floating-point elements using the control -// in imm8, and store the results in dst. +// Add 64-bit integers a and b, and store the result in dst. // -// dst[63:0] := (imm8[0] == 0) ? a[63:0] : a[127:64] -// dst[127:64] := (imm8[1] == 0) ? b[63:0] : b[127:64] +// dst[63:0] := a[63:0] + b[63:0] // -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_pd -#if __has_builtin(__builtin_shufflevector) -#define _mm_shuffle_pd(a, b, imm8) \ - vreinterpretq_m128d_s64(__builtin_shufflevector( \ - vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b), imm8 & 0x1, \ - ((imm8 & 0x2) >> 1) + 2)) -#else -#define _mm_shuffle_pd(a, b, imm8) \ - _mm_castsi128_pd(_mm_set_epi64x( \ - vgetq_lane_s64(vreinterpretq_s64_m128d(b), (imm8 & 0x2) >> 1), \ - vgetq_lane_s64(vreinterpretq_s64_m128d(a), imm8 & 0x1))) -#endif +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_add_si64 +FORCE_INLINE __m64 _mm_add_si64(__m64 a, __m64 b) +{ + return vreinterpret_m64_s64( + vadd_s64(vreinterpret_s64_m64(a), vreinterpret_s64_m64(b))); +} -// Blend packed 16-bit integers from a and b using control mask imm8, and store -// the results in dst. +// Adds the 8 signed 16-bit integers in a to the 8 signed 16-bit integers in b +// and saturates. // -// FOR j := 0 to 7 -// i := j*16 -// IF imm8[j] -// dst[i+15:i] := b[i+15:i] -// ELSE -// dst[i+15:i] := a[i+15:i] -// FI -// ENDFOR -// FORCE_INLINE __m128i _mm_blend_epi16(__m128i a, __m128i b, -// __constrange(0,255) int imm) -#define _mm_blend_epi16(a, b, imm) \ - __extension__({ \ - const uint16_t _mask[8] = {((imm) & (1 << 0)) ? (uint16_t) -1 : 0x0, \ - ((imm) & (1 << 1)) ? (uint16_t) -1 : 0x0, \ - ((imm) & (1 << 2)) ? (uint16_t) -1 : 0x0, \ - ((imm) & (1 << 3)) ? (uint16_t) -1 : 0x0, \ - ((imm) & (1 << 4)) ? (uint16_t) -1 : 0x0, \ - ((imm) & (1 << 5)) ? (uint16_t) -1 : 0x0, \ - ((imm) & (1 << 6)) ? (uint16_t) -1 : 0x0, \ - ((imm) & (1 << 7)) ? (uint16_t) -1 : 0x0}; \ - uint16x8_t _mask_vec = vld1q_u16(_mask); \ - uint16x8_t _a = vreinterpretq_u16_m128i(a); \ - uint16x8_t _b = vreinterpretq_u16_m128i(b); \ - vreinterpretq_m128i_u16(vbslq_u16(_mask_vec, _b, _a)); \ - }) +// r0 := SignedSaturate(a0 + b0) +// r1 := SignedSaturate(a1 + b1) +// ... +// r7 := SignedSaturate(a7 + b7) +// +// https://msdn.microsoft.com/en-us/library/1a306ef8(v=vs.100).aspx +FORCE_INLINE __m128i _mm_adds_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vqaddq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} -// Blend packed double-precision (64-bit) floating-point elements from a and b -// using control mask imm8, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blend_pd -#define _mm_blend_pd(a, b, imm) \ - __extension__({ \ - const uint64_t _mask[2] = { \ - ((imm) & (1 << 0)) ? ~UINT64_C(0) : UINT64_C(0), \ - ((imm) & (1 << 1)) ? ~UINT64_C(0) : UINT64_C(0)}; \ - uint64x2_t _mask_vec = vld1q_u64(_mask); \ - uint64x2_t _a = vreinterpretq_u64_m128d(a); \ - uint64x2_t _b = vreinterpretq_u64_m128d(b); \ - vreinterpretq_m128d_u64(vbslq_u64(_mask_vec, _b, _a)); \ - }) - -// Blend packed 8-bit integers from a and b using mask, and store the results in -// dst. +// Add packed signed 8-bit integers in a and b using saturation, and store the +// results in dst. // // FOR j := 0 to 15 -// i := j*8 -// IF mask[i+7] -// dst[i+7:i] := b[i+7:i] -// ELSE -// dst[i+7:i] := a[i+7:i] -// FI -// ENDFOR -FORCE_INLINE __m128i _mm_blendv_epi8(__m128i _a, __m128i _b, __m128i _mask) -{ - // Use a signed shift right to create a mask with the sign bit - uint8x16_t mask = - vreinterpretq_u8_s8(vshrq_n_s8(vreinterpretq_s8_m128i(_mask), 7)); - uint8x16_t a = vreinterpretq_u8_m128i(_a); - uint8x16_t b = vreinterpretq_u8_m128i(_b); - return vreinterpretq_m128i_u8(vbslq_u8(mask, b, a)); -} - -/* Shifts */ - - -// Shift packed 16-bit integers in a right by imm while shifting in sign -// bits, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srai_epi16 -FORCE_INLINE __m128i _mm_srai_epi16(__m128i a, int imm) -{ - const int count = (imm & ~15) ? 15 : imm; - return (__m128i) vshlq_s16((int16x8_t) a, vdupq_n_s16(-count)); -} - -// Shifts the 8 signed or unsigned 16-bit integers in a left by count bits while -// shifting in zeros. -// -// r0 := a0 << count -// r1 := a1 << count -// ... -// r7 := a7 << count -// -// https://msdn.microsoft.com/en-us/library/es73bcsy(v=vs.90).aspx -#define _mm_slli_epi16(a, imm) \ - __extension__({ \ - __m128i ret; \ - if (unlikely((imm)) <= 0) { \ - ret = a; \ - } \ - if (unlikely((imm) > 15)) { \ - ret = _mm_setzero_si128(); \ - } else { \ - ret = vreinterpretq_m128i_s16( \ - vshlq_n_s16(vreinterpretq_s16_m128i(a), (imm))); \ - } \ - ret; \ - }) - -// Shifts the 4 signed or unsigned 32-bit integers in a left by count bits while -// shifting in zeros. : -// https://msdn.microsoft.com/en-us/library/z2k3bbtb%28v=vs.90%29.aspx -// FORCE_INLINE __m128i _mm_slli_epi32(__m128i a, __constrange(0,255) int imm) -FORCE_INLINE __m128i _mm_slli_epi32(__m128i a, int imm) -{ - if (unlikely(imm <= 0)) /* TODO: add constant range macro: [0, 255] */ - return a; - if (unlikely(imm > 31)) - return _mm_setzero_si128(); - return vreinterpretq_m128i_s32( - vshlq_s32(vreinterpretq_s32_m128i(a), vdupq_n_s32(imm))); -} - -// Shift packed 64-bit integers in a left by imm8 while shifting in zeros, and -// store the results in dst. -FORCE_INLINE __m128i _mm_slli_epi64(__m128i a, int imm) -{ - if (unlikely(imm <= 0)) /* TODO: add constant range macro: [0, 255] */ - return a; - if (unlikely(imm > 63)) - return _mm_setzero_si128(); - return vreinterpretq_m128i_s64( - vshlq_s64(vreinterpretq_s64_m128i(a), vdupq_n_s64(imm))); -} - -// Shift packed 16-bit integers in a right by imm8 while shifting in zeros, and -// store the results in dst. -// -// FOR j := 0 to 7 -// i := j*16 -// IF imm8[7:0] > 15 -// dst[i+15:i] := 0 -// ELSE -// dst[i+15:i] := ZeroExtend16(a[i+15:i] >> imm8[7:0]) -// FI +// i := j*8 +// dst[i+7:i] := Saturate8( a[i+7:i] + b[i+7:i] ) // ENDFOR // -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_epi16 -#define _mm_srli_epi16(a, imm) \ - __extension__({ \ - __m128i ret; \ - if (unlikely(imm) == 0) { \ - ret = a; \ - } \ - if (likely(0 < (imm) && (imm) < 16)) { \ - ret = vreinterpretq_m128i_u16( \ - vshlq_u16(vreinterpretq_u16_m128i(a), vdupq_n_s16(-imm))); \ - } else { \ - ret = _mm_setzero_si128(); \ - } \ - ret; \ - }) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_adds_epi8 +FORCE_INLINE __m128i _mm_adds_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vqaddq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} -// Shift packed 32-bit integers in a right by imm8 while shifting in zeros, and -// store the results in dst. -// -// FOR j := 0 to 3 -// i := j*32 -// IF imm8[7:0] > 31 -// dst[i+31:i] := 0 -// ELSE -// dst[i+31:i] := ZeroExtend32(a[i+31:i] >> imm8[7:0]) -// FI -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_epi32 -// FORCE_INLINE __m128i _mm_srli_epi32(__m128i a, __constrange(0,255) int imm) -#define _mm_srli_epi32(a, imm) \ - __extension__({ \ - __m128i ret; \ - if (unlikely((imm) == 0)) { \ - ret = a; \ - } \ - if (likely(0 < (imm) && (imm) < 32)) { \ - ret = vreinterpretq_m128i_u32( \ - vshlq_u32(vreinterpretq_u32_m128i(a), vdupq_n_s32(-imm))); \ - } else { \ - ret = _mm_setzero_si128(); \ - } \ - ret; \ - }) +// Add packed unsigned 16-bit integers in a and b using saturation, and store +// the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_adds_epu16 +FORCE_INLINE __m128i _mm_adds_epu16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vqaddq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); +} -// Shift packed 64-bit integers in a right by imm8 while shifting in zeros, and -// store the results in dst. +// Adds the 16 unsigned 8-bit integers in a to the 16 unsigned 8-bit integers in +// b and saturates.. +// https://msdn.microsoft.com/en-us/library/9hahyddy(v=vs.100).aspx +FORCE_INLINE __m128i _mm_adds_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vqaddq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +// Compute the bitwise AND of packed double-precision (64-bit) floating-point +// elements in a and b, and store the results in dst. // // FOR j := 0 to 1 // i := j*64 -// IF imm8[7:0] > 63 -// dst[i+63:i] := 0 -// ELSE -// dst[i+63:i] := ZeroExtend64(a[i+63:i] >> imm8[7:0]) -// FI +// dst[i+63:i] := a[i+63:i] AND b[i+63:i] // ENDFOR // -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_epi64 -#define _mm_srli_epi64(a, imm) \ - __extension__({ \ - __m128i ret; \ - if (unlikely((imm) == 0)) { \ - ret = a; \ - } \ - if (likely(0 < (imm) && (imm) < 64)) { \ - ret = vreinterpretq_m128i_u64( \ - vshlq_u64(vreinterpretq_u64_m128i(a), vdupq_n_s64(-imm))); \ - } else { \ - ret = _mm_setzero_si128(); \ - } \ - ret; \ - }) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_and_pd +FORCE_INLINE __m128d _mm_and_pd(__m128d a, __m128d b) +{ + return vreinterpretq_m128d_s64( + vandq_s64(vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b))); +} -// Shift packed 32-bit integers in a right by imm8 while shifting in sign bits, -// and store the results in dst. +// Computes the bitwise AND of the 128-bit value in a and the 128-bit value in +// b. // -// FOR j := 0 to 3 -// i := j*32 -// IF imm8[7:0] > 31 -// dst[i+31:i] := (a[i+31] ? 0xFFFFFFFF : 0x0) -// ELSE -// dst[i+31:i] := SignExtend32(a[i+31:i] >> imm8[7:0]) -// FI +// r := a & b +// +// https://msdn.microsoft.com/en-us/library/vstudio/6d1txsa8(v=vs.100).aspx +FORCE_INLINE __m128i _mm_and_si128(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vandq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compute the bitwise NOT of packed double-precision (64-bit) floating-point +// elements in a and then AND with b, and store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// dst[i+63:i] := ((NOT a[i+63:i]) AND b[i+63:i]) // ENDFOR // -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srai_epi32 -// FORCE_INLINE __m128i _mm_srai_epi32(__m128i a, __constrange(0,255) int imm) -#define _mm_srai_epi32(a, imm) \ - __extension__({ \ - __m128i ret; \ - if (unlikely((imm) == 0)) { \ - ret = a; \ - } \ - if (likely(0 < (imm) && (imm) < 32)) { \ - ret = vreinterpretq_m128i_s32( \ - vshlq_s32(vreinterpretq_s32_m128i(a), vdupq_n_s32(-imm))); \ - } else { \ - ret = vreinterpretq_m128i_s32( \ - vshrq_n_s32(vreinterpretq_s32_m128i(a), 31)); \ - } \ - ret; \ - }) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_andnot_pd +FORCE_INLINE __m128d _mm_andnot_pd(__m128d a, __m128d b) +{ + // *NOTE* argument swap + return vreinterpretq_m128d_s64( + vbicq_s64(vreinterpretq_s64_m128d(b), vreinterpretq_s64_m128d(a))); +} -// Shifts the 128 - bit value in a right by imm bytes while shifting in -// zeros.imm must be an immediate. +// Computes the bitwise AND of the 128-bit value in b and the bitwise NOT of the +// 128-bit value in a. // -// r := srl(a, imm*8) +// r := (~a) & b // -// https://msdn.microsoft.com/en-us/library/305w28yz(v=vs.100).aspx -// FORCE_INLINE _mm_srli_si128(__m128i a, __constrange(0,255) int imm) -#define _mm_srli_si128(a, imm) \ - __extension__({ \ - __m128i ret; \ - if (unlikely((imm) <= 0)) { \ - ret = a; \ - } \ - if (unlikely((imm) > 15)) { \ - ret = _mm_setzero_si128(); \ - } else { \ - ret = vreinterpretq_m128i_s8( \ - vextq_s8(vreinterpretq_s8_m128i(a), vdupq_n_s8(0), (imm))); \ - } \ - ret; \ - }) +// https://msdn.microsoft.com/en-us/library/vstudio/1beaceh8(v=vs.100).aspx +FORCE_INLINE __m128i _mm_andnot_si128(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vbicq_s32(vreinterpretq_s32_m128i(b), + vreinterpretq_s32_m128i(a))); // *NOTE* argument swap +} -// Shifts the 128-bit value in a left by imm bytes while shifting in zeros. imm -// must be an immediate. +// Computes the average of the 8 unsigned 16-bit integers in a and the 8 +// unsigned 16-bit integers in b and rounds. // -// r := a << (imm * 8) +// r0 := (a0 + b0) / 2 +// r1 := (a1 + b1) / 2 +// ... +// r7 := (a7 + b7) / 2 // -// https://msdn.microsoft.com/en-us/library/34d3k2kt(v=vs.100).aspx -// FORCE_INLINE __m128i _mm_slli_si128(__m128i a, __constrange(0,255) int imm) -#define _mm_slli_si128(a, imm) \ - __extension__({ \ - __m128i ret; \ - if (unlikely((imm) <= 0)) { \ - ret = a; \ - } \ - if (unlikely((imm) > 15)) { \ - ret = _mm_setzero_si128(); \ - } else { \ - ret = vreinterpretq_m128i_s8(vextq_s8( \ - vdupq_n_s8(0), vreinterpretq_s8_m128i(a), 16 - (imm))); \ - } \ - ret; \ - }) +// https://msdn.microsoft.com/en-us/library/vstudio/y13ca3c8(v=vs.90).aspx +FORCE_INLINE __m128i _mm_avg_epu16(__m128i a, __m128i b) +{ + return (__m128i) vrhaddq_u16(vreinterpretq_u16_m128i(a), + vreinterpretq_u16_m128i(b)); +} -// Compute the square root of packed double-precision (64-bit) floating-point -// elements in a, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sqrt_pd -FORCE_INLINE __m128d _mm_sqrt_pd(__m128d a) +// Computes the average of the 16 unsigned 8-bit integers in a and the 16 +// unsigned 8-bit integers in b and rounds. +// +// r0 := (a0 + b0) / 2 +// r1 := (a1 + b1) / 2 +// ... +// r15 := (a15 + b15) / 2 +// +// https://msdn.microsoft.com/en-us/library/vstudio/8zwh554a(v%3dvs.90).aspx +FORCE_INLINE __m128i _mm_avg_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vrhaddq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +// Shift a left by imm8 bytes while shifting in zeros, and store the results in +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_bslli_si128 +#define _mm_bslli_si128(a, imm) _mm_slli_si128(a, imm) + +// Shift a right by imm8 bytes while shifting in zeros, and store the results in +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_bsrli_si128 +#define _mm_bsrli_si128(a, imm) _mm_srli_si128(a, imm) + +// Cast vector of type __m128d to type __m128. This intrinsic is only used for +// compilation and does not generate any instructions, thus it has zero latency. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castpd_ps +FORCE_INLINE __m128 _mm_castpd_ps(__m128d a) +{ + return vreinterpretq_m128_s64(vreinterpretq_s64_m128d(a)); +} + +// Cast vector of type __m128d to type __m128i. This intrinsic is only used for +// compilation and does not generate any instructions, thus it has zero latency. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castpd_si128 +FORCE_INLINE __m128i _mm_castpd_si128(__m128d a) +{ + return vreinterpretq_m128i_s64(vreinterpretq_s64_m128d(a)); +} + +// Cast vector of type __m128 to type __m128d. This intrinsic is only used for +// compilation and does not generate any instructions, thus it has zero latency. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castps_pd +FORCE_INLINE __m128d _mm_castps_pd(__m128 a) +{ + return vreinterpretq_m128d_s32(vreinterpretq_s32_m128(a)); +} + +// Applies a type cast to reinterpret four 32-bit floating point values passed +// in as a 128-bit parameter as packed 32-bit integers. +// https://msdn.microsoft.com/en-us/library/bb514099.aspx +FORCE_INLINE __m128i _mm_castps_si128(__m128 a) +{ + return vreinterpretq_m128i_s32(vreinterpretq_s32_m128(a)); +} + +// Cast vector of type __m128i to type __m128d. This intrinsic is only used for +// compilation and does not generate any instructions, thus it has zero latency. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castsi128_pd +FORCE_INLINE __m128d _mm_castsi128_pd(__m128i a) { #if defined(__aarch64__) - return vreinterpretq_m128d_f64(vsqrtq_f64(vreinterpretq_f64_m128d(a))); + return vreinterpretq_m128d_f64(vreinterpretq_f64_m128i(a)); #else - double a0 = sqrt(((double *) &a)[0]); - double a1 = sqrt(((double *) &a)[1]); + return vreinterpretq_m128d_f32(vreinterpretq_f32_m128i(a)); +#endif +} + +// Applies a type cast to reinterpret four 32-bit integers passed in as a +// 128-bit parameter as packed 32-bit floating point values. +// https://msdn.microsoft.com/en-us/library/bb514029.aspx +FORCE_INLINE __m128 _mm_castsi128_ps(__m128i a) +{ + return vreinterpretq_m128_s32(vreinterpretq_s32_m128i(a)); +} + +// Cache line containing p is flushed and invalidated from all caches in the +// coherency domain. : +// https://msdn.microsoft.com/en-us/library/ba08y07y(v=vs.100).aspx +FORCE_INLINE void _mm_clflush(void const *p) +{ + (void) p; + // no corollary for Neon? +} + +// Compares the 8 signed or unsigned 16-bit integers in a and the 8 signed or +// unsigned 16-bit integers in b for equality. +// https://msdn.microsoft.com/en-us/library/2ay060te(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmpeq_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vceqq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Compare packed 32-bit integers in a and b for equality, and store the results +// in dst +FORCE_INLINE __m128i _mm_cmpeq_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u32( + vceqq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compares the 16 signed or unsigned 8-bit integers in a and the 16 signed or +// unsigned 8-bit integers in b for equality. +// https://msdn.microsoft.com/en-us/library/windows/desktop/bz5xk21a(v=vs.90).aspx +FORCE_INLINE __m128i _mm_cmpeq_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vceqq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for equality, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpeq_pd +FORCE_INLINE __m128d _mm_cmpeq_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + // (a == b) -> (a_lo == b_lo) && (a_hi == b_hi) + uint32x4_t cmp = + vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(b)); + uint32x4_t swapped = vrev64q_u32(cmp); + return vreinterpretq_m128d_u32(vandq_u32(cmp, swapped)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for equality, store the result in the lower element of dst, and copy the +// upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpeq_sd +FORCE_INLINE __m128d _mm_cmpeq_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpeq_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for greater-than-or-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpge_pd +FORCE_INLINE __m128d _mm_cmpge_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vcgeq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) >= (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = (*(double *) &a1) >= (*(double *) &b1) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for greater-than-or-equal, store the result in the lower element of dst, +// and copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpge_sd +FORCE_INLINE __m128d _mm_cmpge_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmpge_pd(a, b)); +#else + // expand "_mm_cmpge_pd()" to reduce unnecessary operations + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) >= (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compares the 8 signed 16-bit integers in a and the 8 signed 16-bit integers +// in b for greater than. +// +// r0 := (a0 > b0) ? 0xffff : 0x0 +// r1 := (a1 > b1) ? 0xffff : 0x0 +// ... +// r7 := (a7 > b7) ? 0xffff : 0x0 +// +// https://technet.microsoft.com/en-us/library/xd43yfsa(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmpgt_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vcgtq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Compares the 4 signed 32-bit integers in a and the 4 signed 32-bit integers +// in b for greater than. +// https://msdn.microsoft.com/en-us/library/vstudio/1s9f2z0y(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmpgt_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u32( + vcgtq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compares the 16 signed 8-bit integers in a and the 16 signed 8-bit integers +// in b for greater than. +// +// r0 := (a0 > b0) ? 0xff : 0x0 +// r1 := (a1 > b1) ? 0xff : 0x0 +// ... +// r15 := (a15 > b15) ? 0xff : 0x0 +// +// https://msdn.microsoft.com/zh-tw/library/wf45zt2b(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmpgt_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vcgtq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for greater-than, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpgt_pd +FORCE_INLINE __m128d _mm_cmpgt_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vcgtq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) > (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = (*(double *) &a1) > (*(double *) &b1) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for greater-than, store the result in the lower element of dst, and copy +// the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpgt_sd +FORCE_INLINE __m128d _mm_cmpgt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmpgt_pd(a, b)); +#else + // expand "_mm_cmpge_pd()" to reduce unnecessary operations + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) > (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for less-than-or-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmple_pd +FORCE_INLINE __m128d _mm_cmple_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vcleq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) <= (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = (*(double *) &a1) <= (*(double *) &b1) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for less-than-or-equal, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmple_sd +FORCE_INLINE __m128d _mm_cmple_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmple_pd(a, b)); +#else + // expand "_mm_cmpge_pd()" to reduce unnecessary operations + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) <= (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compares the 8 signed 16-bit integers in a and the 8 signed 16-bit integers +// in b for less than. +// +// r0 := (a0 < b0) ? 0xffff : 0x0 +// r1 := (a1 < b1) ? 0xffff : 0x0 +// ... +// r7 := (a7 < b7) ? 0xffff : 0x0 +// +// https://technet.microsoft.com/en-us/library/t863edb2(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmplt_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vcltq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + + +// Compares the 4 signed 32-bit integers in a and the 4 signed 32-bit integers +// in b for less than. +// https://msdn.microsoft.com/en-us/library/vstudio/4ak0bf5d(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cmplt_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u32( + vcltq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Compares the 16 signed 8-bit integers in a and the 16 signed 8-bit integers +// in b for lesser than. +// https://msdn.microsoft.com/en-us/library/windows/desktop/9s46csht(v=vs.90).aspx +FORCE_INLINE __m128i _mm_cmplt_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vcltq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for less-than, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmplt_pd +FORCE_INLINE __m128d _mm_cmplt_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64( + vcltq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) < (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = (*(double *) &a1) < (*(double *) &b1) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for less-than, store the result in the lower element of dst, and copy the +// upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmplt_sd +FORCE_INLINE __m128d _mm_cmplt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmplt_pd(a, b)); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) < (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpneq_pd +FORCE_INLINE __m128d _mm_cmpneq_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_s32(vmvnq_s32(vreinterpretq_s32_u64( + vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))))); +#else + // (a == b) -> (a_lo == b_lo) && (a_hi == b_hi) + uint32x4_t cmp = + vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(b)); + uint32x4_t swapped = vrev64q_u32(cmp); + return vreinterpretq_m128d_u32(vmvnq_u32(vandq_u32(cmp, swapped))); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-equal, store the result in the lower element of dst, and copy the +// upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpneq_sd +FORCE_INLINE __m128d _mm_cmpneq_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpneq_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-greater-than-or-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnge_pd +FORCE_INLINE __m128d _mm_cmpnge_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64(veorq_u64( + vcgeq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)), + vdupq_n_u64(UINT64_MAX))); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = + !((*(double *) &a0) >= (*(double *) &b0)) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = + !((*(double *) &a1) >= (*(double *) &b1)) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-greater-than-or-equal, store the result in the lower element of +// dst, and copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnge_sd +FORCE_INLINE __m128d _mm_cmpnge_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpnge_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-greater-than, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_cmpngt_pd +FORCE_INLINE __m128d _mm_cmpngt_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64(veorq_u64( + vcgtq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)), + vdupq_n_u64(UINT64_MAX))); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = + !((*(double *) &a0) > (*(double *) &b0)) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = + !((*(double *) &a1) > (*(double *) &b1)) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-greater-than, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpngt_sd +FORCE_INLINE __m128d _mm_cmpngt_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpngt_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-less-than-or-equal, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnle_pd +FORCE_INLINE __m128d _mm_cmpnle_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64(veorq_u64( + vcleq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)), + vdupq_n_u64(UINT64_MAX))); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = + !((*(double *) &a0) <= (*(double *) &b0)) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = + !((*(double *) &a1) <= (*(double *) &b1)) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-less-than-or-equal, store the result in the lower element of dst, +// and copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnle_sd +FORCE_INLINE __m128d _mm_cmpnle_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpnle_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// for not-less-than, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnlt_pd +FORCE_INLINE __m128d _mm_cmpnlt_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_u64(veorq_u64( + vcltq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)), + vdupq_n_u64(UINT64_MAX))); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = + !((*(double *) &a0) < (*(double *) &b0)) ? ~UINT64_C(0) : UINT64_C(0); + d[1] = + !((*(double *) &a1) < (*(double *) &b1)) ? ~UINT64_C(0) : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b for not-less-than, store the result in the lower element of dst, and copy +// the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnlt_sd +FORCE_INLINE __m128d _mm_cmpnlt_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_cmpnlt_pd(a, b)); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// to see if neither is NaN, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpord_pd +FORCE_INLINE __m128d _mm_cmpord_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + // Excluding NaNs, any two floating point numbers can be compared. + uint64x2_t not_nan_a = + vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(a)); + uint64x2_t not_nan_b = + vceqq_f64(vreinterpretq_f64_m128d(b), vreinterpretq_f64_m128d(b)); + return vreinterpretq_m128d_u64(vandq_u64(not_nan_a, not_nan_b)); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = ((*(double *) &a0) == (*(double *) &a0) && + (*(double *) &b0) == (*(double *) &b0)) + ? ~UINT64_C(0) + : UINT64_C(0); + d[1] = ((*(double *) &a1) == (*(double *) &a1) && + (*(double *) &b1) == (*(double *) &b1)) + ? ~UINT64_C(0) + : UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b to see if neither is NaN, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpord_sd +FORCE_INLINE __m128d _mm_cmpord_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmpord_pd(a, b)); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t d[2]; + d[0] = ((*(double *) &a0) == (*(double *) &a0) && + (*(double *) &b0) == (*(double *) &b0)) + ? ~UINT64_C(0) + : UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b +// to see if either is NaN, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpunord_pd +FORCE_INLINE __m128d _mm_cmpunord_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + // Two NaNs are not equal in comparison operation. + uint64x2_t not_nan_a = + vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(a)); + uint64x2_t not_nan_b = + vceqq_f64(vreinterpretq_f64_m128d(b), vreinterpretq_f64_m128d(b)); + return vreinterpretq_m128d_s32( + vmvnq_s32(vreinterpretq_s32_u64(vandq_u64(not_nan_a, not_nan_b)))); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = ((*(double *) &a0) == (*(double *) &a0) && + (*(double *) &b0) == (*(double *) &b0)) + ? UINT64_C(0) + : ~UINT64_C(0); + d[1] = ((*(double *) &a1) == (*(double *) &a1) && + (*(double *) &b1) == (*(double *) &b1)) + ? UINT64_C(0) + : ~UINT64_C(0); + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b to see if either is NaN, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpunord_sd +FORCE_INLINE __m128d _mm_cmpunord_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_cmpunord_pd(a, b)); +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t d[2]; + d[0] = ((*(double *) &a0) == (*(double *) &a0) && + (*(double *) &b0) == (*(double *) &b0)) + ? UINT64_C(0) + : ~UINT64_C(0); + d[1] = a1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for greater-than-or-equal, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comige_sd +FORCE_INLINE int _mm_comige_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vcgeq_f64(a, b), 0) & 0x1; +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + + return (*(double *) &a0 >= *(double *) &b0); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for greater-than, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comigt_sd +FORCE_INLINE int _mm_comigt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vcgtq_f64(a, b), 0) & 0x1; +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + + return (*(double *) &a0 > *(double *) &b0); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for less-than-or-equal, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comile_sd +FORCE_INLINE int _mm_comile_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vcleq_f64(a, b), 0) & 0x1; +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + + return (*(double *) &a0 <= *(double *) &b0); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for less-than, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comilt_sd +FORCE_INLINE int _mm_comilt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vcltq_f64(a, b), 0) & 0x1; +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + + return (*(double *) &a0 < *(double *) &b0); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for equality, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comieq_sd +FORCE_INLINE int _mm_comieq_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vgetq_lane_u64(vceqq_f64(a, b), 0) & 0x1; +#else + uint32x4_t a_not_nan = + vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(a)); + uint32x4_t b_not_nan = + vceqq_u32(vreinterpretq_u32_m128d(b), vreinterpretq_u32_m128d(b)); + uint32x4_t a_and_b_not_nan = vandq_u32(a_not_nan, b_not_nan); + uint32x4_t a_eq_b = + vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(b)); + uint64x2_t and_results = vandq_u64(vreinterpretq_u64_u32(a_and_b_not_nan), + vreinterpretq_u64_u32(a_eq_b)); + return vgetq_lane_u64(and_results, 0) & 0x1; +#endif +} + +// Compare the lower double-precision (64-bit) floating-point element in a and b +// for not-equal, and return the boolean result (0 or 1). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comineq_sd +FORCE_INLINE int _mm_comineq_sd(__m128d a, __m128d b) +{ + return !_mm_comieq_sd(a, b); +} + +// Convert packed signed 32-bit integers in a to packed double-precision +// (64-bit) floating-point elements, and store the results in dst. +// +// FOR j := 0 to 1 +// i := j*32 +// m := j*64 +// dst[m+63:m] := Convert_Int32_To_FP64(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtepi32_pd +FORCE_INLINE __m128d _mm_cvtepi32_pd(__m128i a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcvtq_f64_s64(vmovl_s32(vget_low_s32(vreinterpretq_s32_m128i(a))))); +#else + double a0 = (double) vgetq_lane_s32(vreinterpretq_s32_m128i(a), 0); + double a1 = (double) vgetq_lane_s32(vreinterpretq_s32_m128i(a), 1); return _mm_set_pd(a1, a0); #endif } -// Compute the square root of the lower double-precision (64-bit) floating-point -// element in b, store the result in the lower element of dst, and copy the -// upper element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sqrt_sd -FORCE_INLINE __m128d _mm_sqrt_sd(__m128d a, __m128d b) +// Converts the four signed 32-bit integer values of a to single-precision, +// floating-point values +// https://msdn.microsoft.com/en-us/library/vstudio/36bwxcx5(v=vs.100).aspx +FORCE_INLINE __m128 _mm_cvtepi32_ps(__m128i a) +{ + return vreinterpretq_m128_f32(vcvtq_f32_s32(vreinterpretq_s32_m128i(a))); +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed 32-bit integers, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// k := 64*j +// dst[i+31:i] := Convert_FP64_To_Int32(a[k+63:k]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpd_epi32 +FORCE_INLINE __m128i _mm_cvtpd_epi32(__m128d a) +{ + __m128d rnd = _mm_round_pd(a, _MM_FROUND_CUR_DIRECTION); + double d0 = ((double *) &rnd)[0]; + double d1 = ((double *) &rnd)[1]; + return _mm_set_epi32(0, 0, (int32_t) d1, (int32_t) d0); +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed 32-bit integers, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// k := 64*j +// dst[i+31:i] := Convert_FP64_To_Int32(a[k+63:k]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpd_pi32 +FORCE_INLINE __m64 _mm_cvtpd_pi32(__m128d a) +{ + __m128d rnd = _mm_round_pd(a, _MM_FROUND_CUR_DIRECTION); + double d0 = ((double *) &rnd)[0]; + double d1 = ((double *) &rnd)[1]; + int32_t ALIGN_STRUCT(16) data[2] = {(int32_t) d0, (int32_t) d1}; + return vreinterpret_m64_s32(vld1_s32(data)); +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed single-precision (32-bit) floating-point elements, and store the +// results in dst. +// +// FOR j := 0 to 1 +// i := 32*j +// k := 64*j +// dst[i+31:i] := Convert_FP64_To_FP32(a[k+64:k]) +// ENDFOR +// dst[127:64] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpd_ps +FORCE_INLINE __m128 _mm_cvtpd_ps(__m128d a) { #if defined(__aarch64__) - return _mm_move_sd(a, _mm_sqrt_pd(b)); + float32x2_t tmp = vcvt_f32_f64(vreinterpretq_f64_m128d(a)); + return vreinterpretq_m128_f32(vcombine_f32(tmp, vdup_n_f32(0))); #else - return _mm_set_pd(((double *) &a)[1], sqrt(((double *) &b)[0])); + float a0 = (float) ((double *) &a)[0]; + float a1 = (float) ((double *) &a)[1]; + return _mm_set_ps(0, 0, a1, a0); #endif } -// Shifts the 8 signed or unsigned 16-bit integers in a left by count bits while -// shifting in zeros. +// Convert packed signed 32-bit integers in a to packed double-precision +// (64-bit) floating-point elements, and store the results in dst. // -// r0 := a0 << count -// r1 := a1 << count -// ... -// r7 := a7 << count +// FOR j := 0 to 1 +// i := j*32 +// m := j*64 +// dst[m+63:m] := Convert_Int32_To_FP64(a[i+31:i]) +// ENDFOR // -// https://msdn.microsoft.com/en-us/library/c79w388h(v%3dvs.90).aspx -FORCE_INLINE __m128i _mm_sll_epi16(__m128i a, __m128i count) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi32_pd +FORCE_INLINE __m128d _mm_cvtpi32_pd(__m64 a) { - uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); - if (unlikely(c > 15)) - return _mm_setzero_si128(); - - int16x8_t vc = vdupq_n_s16((int16_t) c); - return vreinterpretq_m128i_s16(vshlq_s16(vreinterpretq_s16_m128i(a), vc)); +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcvtq_f64_s64(vmovl_s32(vreinterpret_s32_m64(a)))); +#else + double a0 = (double) vget_lane_s32(vreinterpret_s32_m64(a), 0); + double a1 = (double) vget_lane_s32(vreinterpret_s32_m64(a), 1); + return _mm_set_pd(a1, a0); +#endif } -// Shifts the 4 signed or unsigned 32-bit integers in a left by count bits while -// shifting in zeros. +// Converts the four single-precision, floating-point values of a to signed +// 32-bit integer values. // -// r0 := a0 << count -// r1 := a1 << count -// r2 := a2 << count -// r3 := a3 << count +// r0 := (int) a0 +// r1 := (int) a1 +// r2 := (int) a2 +// r3 := (int) a3 // -// https://msdn.microsoft.com/en-us/library/6fe5a6s9(v%3dvs.90).aspx -FORCE_INLINE __m128i _mm_sll_epi32(__m128i a, __m128i count) +// https://msdn.microsoft.com/en-us/library/vstudio/xdc42k5e(v=vs.100).aspx +// *NOTE*. The default rounding mode on SSE is 'round to even', which ARMv7-A +// does not support! It is supported on ARMv8-A however. +FORCE_INLINE __m128i _mm_cvtps_epi32(__m128 a) { - uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); - if (unlikely(c > 31)) - return _mm_setzero_si128(); - - int32x4_t vc = vdupq_n_s32((int32_t) c); - return vreinterpretq_m128i_s32(vshlq_s32(vreinterpretq_s32_m128i(a), vc)); +#if defined(__aarch64__) || defined(__ARM_FEATURE_DIRECTED_ROUNDING) + switch (_MM_GET_ROUNDING_MODE()) { + case _MM_ROUND_NEAREST: + return vreinterpretq_m128i_s32(vcvtnq_s32_f32(a)); + case _MM_ROUND_DOWN: + return vreinterpretq_m128i_s32(vcvtmq_s32_f32(a)); + case _MM_ROUND_UP: + return vreinterpretq_m128i_s32(vcvtpq_s32_f32(a)); + default: // _MM_ROUND_TOWARD_ZERO + return vreinterpretq_m128i_s32(vcvtq_s32_f32(a)); + } +#else + float *f = (float *) &a; + switch (_MM_GET_ROUNDING_MODE()) { + case _MM_ROUND_NEAREST: { + uint32x4_t signmask = vdupq_n_u32(0x80000000); + float32x4_t half = vbslq_f32(signmask, vreinterpretq_f32_m128(a), + vdupq_n_f32(0.5f)); /* +/- 0.5 */ + int32x4_t r_normal = vcvtq_s32_f32(vaddq_f32( + vreinterpretq_f32_m128(a), half)); /* round to integer: [a + 0.5]*/ + int32x4_t r_trunc = vcvtq_s32_f32( + vreinterpretq_f32_m128(a)); /* truncate to integer: [a] */ + int32x4_t plusone = vreinterpretq_s32_u32(vshrq_n_u32( + vreinterpretq_u32_s32(vnegq_s32(r_trunc)), 31)); /* 1 or 0 */ + int32x4_t r_even = vbicq_s32(vaddq_s32(r_trunc, plusone), + vdupq_n_s32(1)); /* ([a] + {0,1}) & ~1 */ + float32x4_t delta = vsubq_f32( + vreinterpretq_f32_m128(a), + vcvtq_f32_s32(r_trunc)); /* compute delta: delta = (a - [a]) */ + uint32x4_t is_delta_half = + vceqq_f32(delta, half); /* delta == +/- 0.5 */ + return vreinterpretq_m128i_s32( + vbslq_s32(is_delta_half, r_even, r_normal)); + } + case _MM_ROUND_DOWN: + return _mm_set_epi32(floorf(f[3]), floorf(f[2]), floorf(f[1]), + floorf(f[0])); + case _MM_ROUND_UP: + return _mm_set_epi32(ceilf(f[3]), ceilf(f[2]), ceilf(f[1]), + ceilf(f[0])); + default: // _MM_ROUND_TOWARD_ZERO + return _mm_set_epi32((int32_t) f[3], (int32_t) f[2], (int32_t) f[1], + (int32_t) f[0]); + } +#endif } -// Shifts the 2 signed or unsigned 64-bit integers in a left by count bits while -// shifting in zeros. +// Convert packed single-precision (32-bit) floating-point elements in a to +// packed double-precision (64-bit) floating-point elements, and store the +// results in dst. // -// r0 := a0 << count -// r1 := a1 << count +// FOR j := 0 to 1 +// i := 64*j +// k := 32*j +// dst[i+63:i] := Convert_FP32_To_FP64(a[k+31:k]) +// ENDFOR // -// https://msdn.microsoft.com/en-us/library/6ta9dffd(v%3dvs.90).aspx -FORCE_INLINE __m128i _mm_sll_epi64(__m128i a, __m128i count) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pd +FORCE_INLINE __m128d _mm_cvtps_pd(__m128 a) { - uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); - if (unlikely(c > 63)) - return _mm_setzero_si128(); - - int64x2_t vc = vdupq_n_s64((int64_t) c); - return vreinterpretq_m128i_s64(vshlq_s64(vreinterpretq_s64_m128i(a), vc)); +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcvt_f64_f32(vget_low_f32(vreinterpretq_f32_m128(a)))); +#else + double a0 = (double) vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); + double a1 = (double) vgetq_lane_f32(vreinterpretq_f32_m128(a), 1); + return _mm_set_pd(a1, a0); +#endif } -// Shifts the 8 signed or unsigned 16-bit integers in a right by count bits -// while shifting in zeros. +// Copy the lower double-precision (64-bit) floating-point element of a to dst. // -// r0 := srl(a0, count) -// r1 := srl(a1, count) -// ... -// r7 := srl(a7, count) +// dst[63:0] := a[63:0] // -// https://msdn.microsoft.com/en-us/library/wd5ax830(v%3dvs.90).aspx -FORCE_INLINE __m128i _mm_srl_epi16(__m128i a, __m128i count) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_f64 +FORCE_INLINE double _mm_cvtsd_f64(__m128d a) { - uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); - if (unlikely(c > 15)) - return _mm_setzero_si128(); - - int16x8_t vc = vdupq_n_s16(-(int16_t) c); - return vreinterpretq_m128i_u16(vshlq_u16(vreinterpretq_u16_m128i(a), vc)); +#if defined(__aarch64__) + return (double) vgetq_lane_f64(vreinterpretq_f64_m128d(a), 0); +#else + return ((double *) &a)[0]; +#endif } -// Shifts the 4 signed or unsigned 32-bit integers in a right by count bits -// while shifting in zeros. +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 32-bit integer, and store the result in dst. // -// r0 := srl(a0, count) -// r1 := srl(a1, count) -// r2 := srl(a2, count) -// r3 := srl(a3, count) +// dst[31:0] := Convert_FP64_To_Int32(a[63:0]) // -// https://msdn.microsoft.com/en-us/library/a9cbttf4(v%3dvs.90).aspx -FORCE_INLINE __m128i _mm_srl_epi32(__m128i a, __m128i count) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_si32 +FORCE_INLINE int32_t _mm_cvtsd_si32(__m128d a) { - uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); - if (unlikely(c > 31)) - return _mm_setzero_si128(); - - int32x4_t vc = vdupq_n_s32(-(int32_t) c); - return vreinterpretq_m128i_u32(vshlq_u32(vreinterpretq_u32_m128i(a), vc)); +#if defined(__aarch64__) + return (int32_t) vgetq_lane_f64(vrndiq_f64(vreinterpretq_f64_m128d(a)), 0); +#else + __m128d rnd = _mm_round_pd(a, _MM_FROUND_CUR_DIRECTION); + double ret = ((double *) &rnd)[0]; + return (int32_t) ret; +#endif } -// Shifts the 2 signed or unsigned 64-bit integers in a right by count bits -// while shifting in zeros. +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 64-bit integer, and store the result in dst. // -// r0 := srl(a0, count) -// r1 := srl(a1, count) +// dst[63:0] := Convert_FP64_To_Int64(a[63:0]) // -// https://msdn.microsoft.com/en-us/library/yf6cf9k8(v%3dvs.90).aspx -FORCE_INLINE __m128i _mm_srl_epi64(__m128i a, __m128i count) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_si64 +FORCE_INLINE int64_t _mm_cvtsd_si64(__m128d a) { - uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); - if (unlikely(c > 63)) - return _mm_setzero_si128(); +#if defined(__aarch64__) + return (int64_t) vgetq_lane_f64(vrndiq_f64(vreinterpretq_f64_m128d(a)), 0); +#else + __m128d rnd = _mm_round_pd(a, _MM_FROUND_CUR_DIRECTION); + double ret = ((double *) &rnd)[0]; + return (int64_t) ret; +#endif +} - int64x2_t vc = vdupq_n_s64(-(int64_t) c); - return vreinterpretq_m128i_u64(vshlq_u64(vreinterpretq_u64_m128i(a), vc)); +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 64-bit integer, and store the result in dst. +// +// dst[63:0] := Convert_FP64_To_Int64(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_si64x +#define _mm_cvtsd_si64x _mm_cvtsd_si64 + +// Convert the lower double-precision (64-bit) floating-point element in b to a +// single-precision (32-bit) floating-point element, store the result in the +// lower element of dst, and copy the upper 3 packed elements from a to the +// upper elements of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_ss +FORCE_INLINE __m128 _mm_cvtsd_ss(__m128 a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32(vsetq_lane_f32( + vget_lane_f32(vcvt_f32_f64(vreinterpretq_f64_m128d(b)), 0), + vreinterpretq_f32_m128(a), 0)); +#else + return vreinterpretq_m128_f32(vsetq_lane_f32((float) ((double *) &b)[0], + vreinterpretq_f32_m128(a), 0)); +#endif +} + +// Copy the lower 32-bit integer in a to dst. +// +// dst[31:0] := a[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si32 +FORCE_INLINE int _mm_cvtsi128_si32(__m128i a) +{ + return vgetq_lane_s32(vreinterpretq_s32_m128i(a), 0); +} + +// Copy the lower 64-bit integer in a to dst. +// +// dst[63:0] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si64 +FORCE_INLINE int64_t _mm_cvtsi128_si64(__m128i a) +{ + return vgetq_lane_s64(vreinterpretq_s64_m128i(a), 0); +} + +// Copy the lower 64-bit integer in a to dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si64x +#define _mm_cvtsi128_si64x(a) _mm_cvtsi128_si64(a) + +// Convert the signed 32-bit integer b to a double-precision (64-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi32_sd +FORCE_INLINE __m128d _mm_cvtsi32_sd(__m128d a, int32_t b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vsetq_lane_f64((double) b, vreinterpretq_f64_m128d(a), 0)); +#else + double bf = (double) b; + return vreinterpretq_m128d_s64( + vsetq_lane_s64(*(int64_t *) &bf, vreinterpretq_s64_m128d(a), 0)); +#endif +} + +// Copy the lower 64-bit integer in a to dst. +// +// dst[63:0] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si64x +#define _mm_cvtsi128_si64x(a) _mm_cvtsi128_si64(a) + +// Moves 32-bit integer a to the least significant 32 bits of an __m128 object, +// zero extending the upper bits. +// +// r0 := a +// r1 := 0x0 +// r2 := 0x0 +// r3 := 0x0 +// +// https://msdn.microsoft.com/en-us/library/ct3539ha%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_cvtsi32_si128(int a) +{ + return vreinterpretq_m128i_s32(vsetq_lane_s32(a, vdupq_n_s32(0), 0)); +} + +// Convert the signed 64-bit integer b to a double-precision (64-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi64_sd +FORCE_INLINE __m128d _mm_cvtsi64_sd(__m128d a, int64_t b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vsetq_lane_f64((double) b, vreinterpretq_f64_m128d(a), 0)); +#else + double bf = (double) b; + return vreinterpretq_m128d_s64( + vsetq_lane_s64(*(int64_t *) &bf, vreinterpretq_s64_m128d(a), 0)); +#endif +} + +// Moves 64-bit integer a to the least significant 64 bits of an __m128 object, +// zero extending the upper bits. +// +// r0 := a +// r1 := 0x0 +FORCE_INLINE __m128i _mm_cvtsi64_si128(int64_t a) +{ + return vreinterpretq_m128i_s64(vsetq_lane_s64(a, vdupq_n_s64(0), 0)); +} + +// Copy 64-bit integer a to the lower element of dst, and zero the upper +// element. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi64x_si128 +#define _mm_cvtsi64x_si128(a) _mm_cvtsi64_si128(a) + +// Convert the signed 64-bit integer b to a double-precision (64-bit) +// floating-point element, store the result in the lower element of dst, and +// copy the upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi64x_sd +#define _mm_cvtsi64x_sd(a, b) _mm_cvtsi64_sd(a, b) + +// Convert the lower single-precision (32-bit) floating-point element in b to a +// double-precision (64-bit) floating-point element, store the result in the +// lower element of dst, and copy the upper element from a to the upper element +// of dst. +// +// dst[63:0] := Convert_FP32_To_FP64(b[31:0]) +// dst[127:64] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtss_sd +FORCE_INLINE __m128d _mm_cvtss_sd(__m128d a, __m128 b) +{ + double d = (double) vgetq_lane_f32(vreinterpretq_f32_m128(b), 0); +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vsetq_lane_f64(d, vreinterpretq_f64_m128d(a), 0)); +#else + return vreinterpretq_m128d_s64( + vsetq_lane_s64(*(int64_t *) &d, vreinterpretq_s64_m128d(a), 0)); +#endif +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed 32-bit integers with truncation, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttpd_epi32 +FORCE_INLINE __m128i _mm_cvttpd_epi32(__m128d a) +{ + double a0 = ((double *) &a)[0]; + double a1 = ((double *) &a)[1]; + return _mm_set_epi32(0, 0, (int32_t) a1, (int32_t) a0); +} + +// Convert packed double-precision (64-bit) floating-point elements in a to +// packed 32-bit integers with truncation, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttpd_pi32 +FORCE_INLINE __m64 _mm_cvttpd_pi32(__m128d a) +{ + double a0 = ((double *) &a)[0]; + double a1 = ((double *) &a)[1]; + int32_t ALIGN_STRUCT(16) data[2] = {(int32_t) a0, (int32_t) a1}; + return vreinterpret_m64_s32(vld1_s32(data)); +} + +// Converts the four single-precision, floating-point values of a to signed +// 32-bit integer values using truncate. +// https://msdn.microsoft.com/en-us/library/vstudio/1h005y6x(v=vs.100).aspx +FORCE_INLINE __m128i _mm_cvttps_epi32(__m128 a) +{ + return vreinterpretq_m128i_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a))); +} + +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 32-bit integer with truncation, and store the result in dst. +// +// dst[63:0] := Convert_FP64_To_Int32_Truncate(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttsd_si32 +FORCE_INLINE int32_t _mm_cvttsd_si32(__m128d a) +{ + double ret = *((double *) &a); + return (int32_t) ret; +} + +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 64-bit integer with truncation, and store the result in dst. +// +// dst[63:0] := Convert_FP64_To_Int64_Truncate(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttsd_si64 +FORCE_INLINE int64_t _mm_cvttsd_si64(__m128d a) +{ +#if defined(__aarch64__) + return vgetq_lane_s64(vcvtq_s64_f64(vreinterpretq_f64_m128d(a)), 0); +#else + double ret = *((double *) &a); + return (int64_t) ret; +#endif +} + +// Convert the lower double-precision (64-bit) floating-point element in a to a +// 64-bit integer with truncation, and store the result in dst. +// +// dst[63:0] := Convert_FP64_To_Int64_Truncate(a[63:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttsd_si64x +#define _mm_cvttsd_si64x(a) _mm_cvttsd_si64(a) + +// Divide packed double-precision (64-bit) floating-point elements in a by +// packed elements in b, and store the results in dst. +// +// FOR j := 0 to 1 +// i := 64*j +// dst[i+63:i] := a[i+63:i] / b[i+63:i] +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_div_pd +FORCE_INLINE __m128d _mm_div_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vdivq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + double *da = (double *) &a; + double *db = (double *) &b; + double c[2]; + c[0] = da[0] / db[0]; + c[1] = da[1] / db[1]; + return vld1q_f32((float32_t *) c); +#endif +} + +// Divide the lower double-precision (64-bit) floating-point element in a by the +// lower double-precision (64-bit) floating-point element in b, store the result +// in the lower element of dst, and copy the upper element from a to the upper +// element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_div_sd +FORCE_INLINE __m128d _mm_div_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + float64x2_t tmp = + vdivq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)); + return vreinterpretq_m128d_f64( + vsetq_lane_f64(vgetq_lane_f64(vreinterpretq_f64_m128d(a), 1), tmp, 1)); +#else + return _mm_move_sd(a, _mm_div_pd(a, b)); +#endif +} + +// Extracts the selected signed or unsigned 16-bit integer from a and zero +// extends. +// https://msdn.microsoft.com/en-us/library/6dceta0c(v=vs.100).aspx +// FORCE_INLINE int _mm_extract_epi16(__m128i a, __constrange(0,8) int imm) +#define _mm_extract_epi16(a, imm) \ + vgetq_lane_u16(vreinterpretq_u16_m128i(a), (imm)) + +// Inserts the least significant 16 bits of b into the selected 16-bit integer +// of a. +// https://msdn.microsoft.com/en-us/library/kaze8hz1%28v=vs.100%29.aspx +// FORCE_INLINE __m128i _mm_insert_epi16(__m128i a, int b, +// __constrange(0,8) int imm) +#define _mm_insert_epi16(a, b, imm) \ + __extension__({ \ + vreinterpretq_m128i_s16( \ + vsetq_lane_s16((b), vreinterpretq_s16_m128i(a), (imm))); \ + }) + +// Loads two double-precision from 16-byte aligned memory, floating-point +// values. +// +// dst[127:0] := MEM[mem_addr+127:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_pd +FORCE_INLINE __m128d _mm_load_pd(const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vld1q_f64(p)); +#else + const float *fp = (const float *) p; + float ALIGN_STRUCT(16) data[4] = {fp[0], fp[1], fp[2], fp[3]}; + return vreinterpretq_m128d_f32(vld1q_f32(data)); +#endif +} + +// Load a double-precision (64-bit) floating-point element from memory into both +// elements of dst. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_pd1 +#define _mm_load_pd1 _mm_load1_pd + +// Load a double-precision (64-bit) floating-point element from memory into the +// lower of dst, and zero the upper element. mem_addr does not need to be +// aligned on any particular boundary. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_sd +FORCE_INLINE __m128d _mm_load_sd(const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vsetq_lane_f64(*p, vdupq_n_f64(0), 0)); +#else + const float *fp = (const float *) p; + float ALIGN_STRUCT(16) data[4] = {fp[0], fp[1], 0, 0}; + return vreinterpretq_m128d_f32(vld1q_f32(data)); +#endif +} + +// Loads 128-bit value. : +// https://msdn.microsoft.com/en-us/library/atzzad1h(v=vs.80).aspx +FORCE_INLINE __m128i _mm_load_si128(const __m128i *p) +{ + return vreinterpretq_m128i_s32(vld1q_s32((const int32_t *) p)); +} + +// Load a double-precision (64-bit) floating-point element from memory into both +// elements of dst. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load1_pd +FORCE_INLINE __m128d _mm_load1_pd(const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vld1q_dup_f64(p)); +#else + return vreinterpretq_m128d_s64(vdupq_n_s64(*(const int64_t *) p)); +#endif +} + +// Load a double-precision (64-bit) floating-point element from memory into the +// upper element of dst, and copy the lower element from a to dst. mem_addr does +// not need to be aligned on any particular boundary. +// +// dst[63:0] := a[63:0] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadh_pd +FORCE_INLINE __m128d _mm_loadh_pd(__m128d a, const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcombine_f64(vget_low_f64(vreinterpretq_f64_m128d(a)), vld1_f64(p))); +#else + return vreinterpretq_m128d_f32(vcombine_f32( + vget_low_f32(vreinterpretq_f32_m128d(a)), vld1_f32((const float *) p))); +#endif +} + +// Load 64-bit integer from memory into the first element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadl_epi64 +FORCE_INLINE __m128i _mm_loadl_epi64(__m128i const *p) +{ + /* Load the lower 64 bits of the value pointed to by p into the + * lower 64 bits of the result, zeroing the upper 64 bits of the result. + */ + return vreinterpretq_m128i_s32( + vcombine_s32(vld1_s32((int32_t const *) p), vcreate_s32(0))); +} + +// Load a double-precision (64-bit) floating-point element from memory into the +// lower element of dst, and copy the upper element from a to dst. mem_addr does +// not need to be aligned on any particular boundary. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadl_pd +FORCE_INLINE __m128d _mm_loadl_pd(__m128d a, const double *p) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vcombine_f64(vld1_f64(p), vget_high_f64(vreinterpretq_f64_m128d(a)))); +#else + return vreinterpretq_m128d_f32( + vcombine_f32(vld1_f32((const float *) p), + vget_high_f32(vreinterpretq_f32_m128d(a)))); +#endif +} + +// Load 2 double-precision (64-bit) floating-point elements from memory into dst +// in reverse order. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// dst[63:0] := MEM[mem_addr+127:mem_addr+64] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadr_pd +FORCE_INLINE __m128d _mm_loadr_pd(const double *p) +{ +#if defined(__aarch64__) + float64x2_t v = vld1q_f64(p); + return vreinterpretq_m128d_f64(vextq_f64(v, v, 1)); +#else + int64x2_t v = vld1q_s64((const int64_t *) p); + return vreinterpretq_m128d_s64(vextq_s64(v, v, 1)); +#endif +} + +// Loads two double-precision from unaligned memory, floating-point values. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_pd +FORCE_INLINE __m128d _mm_loadu_pd(const double *p) +{ + return _mm_load_pd(p); +} + +// Loads 128-bit value. : +// https://msdn.microsoft.com/zh-cn/library/f4k12ae8(v=vs.90).aspx +FORCE_INLINE __m128i _mm_loadu_si128(const __m128i *p) +{ + return vreinterpretq_m128i_s32(vld1q_s32((const int32_t *) p)); +} + +// Load unaligned 32-bit integer from memory into the first element of dst. +// +// dst[31:0] := MEM[mem_addr+31:mem_addr] +// dst[MAX:32] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_si32 +FORCE_INLINE __m128i _mm_loadu_si32(const void *p) +{ + return vreinterpretq_m128i_s32( + vsetq_lane_s32(*(const int32_t *) p, vdupq_n_s32(0), 0)); +} + +// Multiplies the 8 signed 16-bit integers from a by the 8 signed 16-bit +// integers from b. +// +// r0 := (a0 * b0) + (a1 * b1) +// r1 := (a2 * b2) + (a3 * b3) +// r2 := (a4 * b4) + (a5 * b5) +// r3 := (a6 * b6) + (a7 * b7) +// https://msdn.microsoft.com/en-us/library/yht36sa6(v=vs.90).aspx +FORCE_INLINE __m128i _mm_madd_epi16(__m128i a, __m128i b) +{ + int32x4_t low = vmull_s16(vget_low_s16(vreinterpretq_s16_m128i(a)), + vget_low_s16(vreinterpretq_s16_m128i(b))); + int32x4_t high = vmull_s16(vget_high_s16(vreinterpretq_s16_m128i(a)), + vget_high_s16(vreinterpretq_s16_m128i(b))); + + int32x2_t low_sum = vpadd_s32(vget_low_s32(low), vget_high_s32(low)); + int32x2_t high_sum = vpadd_s32(vget_low_s32(high), vget_high_s32(high)); + + return vreinterpretq_m128i_s32(vcombine_s32(low_sum, high_sum)); +} + +// Conditionally store 8-bit integer elements from a into memory using mask +// (elements are not stored when the highest bit is not set in the corresponding +// element) and a non-temporal memory hint. mem_addr does not need to be aligned +// on any particular boundary. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_maskmoveu_si128 +FORCE_INLINE void _mm_maskmoveu_si128(__m128i a, __m128i mask, char *mem_addr) +{ + int8x16_t shr_mask = vshrq_n_s8(vreinterpretq_s8_m128i(mask), 7); + __m128 b = _mm_load_ps((const float *) mem_addr); + int8x16_t masked = + vbslq_s8(vreinterpretq_u8_s8(shr_mask), vreinterpretq_s8_m128i(a), + vreinterpretq_s8_m128(b)); + vst1q_s8((int8_t *) mem_addr, masked); +} + +// Computes the pairwise maxima of the 8 signed 16-bit integers from a and the 8 +// signed 16-bit integers from b. +// https://msdn.microsoft.com/en-us/LIBRary/3x060h7c(v=vs.100).aspx +FORCE_INLINE __m128i _mm_max_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vmaxq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Computes the pairwise maxima of the 16 unsigned 8-bit integers from a and the +// 16 unsigned 8-bit integers from b. +// https://msdn.microsoft.com/en-us/library/st6634za(v=vs.100).aspx +FORCE_INLINE __m128i _mm_max_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vmaxq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b, +// and store packed maximum values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pd +FORCE_INLINE __m128d _mm_max_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) +#if SSE2NEON_PRECISE_MINMAX + float64x2_t _a = vreinterpretq_f64_m128d(a); + float64x2_t _b = vreinterpretq_f64_m128d(b); + return vreinterpretq_m128d_f64(vbslq_f64(vcgtq_f64(_a, _b), _a, _b)); +#else + return vreinterpretq_m128d_f64( + vmaxq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#endif +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) > (*(double *) &b0) ? a0 : b0; + d[1] = (*(double *) &a1) > (*(double *) &b1) ? a1 : b1; + + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b, store the maximum value in the lower element of dst, and copy the upper +// element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_sd +FORCE_INLINE __m128d _mm_max_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_max_pd(a, b)); +#else + double *da = (double *) &a; + double *db = (double *) &b; + double c[2] = {da[0] > db[0] ? da[0] : db[0], da[1]}; + return vreinterpretq_m128d_f32(vld1q_f32((float32_t *) c)); +#endif +} + +// Computes the pairwise minima of the 8 signed 16-bit integers from a and the 8 +// signed 16-bit integers from b. +// https://msdn.microsoft.com/en-us/library/vstudio/6te997ew(v=vs.100).aspx +FORCE_INLINE __m128i _mm_min_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vminq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Computes the pairwise minima of the 16 unsigned 8-bit integers from a and the +// 16 unsigned 8-bit integers from b. +// https://msdn.microsoft.com/ko-kr/library/17k8cf58(v=vs.100).aspxx +FORCE_INLINE __m128i _mm_min_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vminq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +// Compare packed double-precision (64-bit) floating-point elements in a and b, +// and store packed minimum values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pd +FORCE_INLINE __m128d _mm_min_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) +#if SSE2NEON_PRECISE_MINMAX + float64x2_t _a = vreinterpretq_f64_m128d(a); + float64x2_t _b = vreinterpretq_f64_m128d(b); + return vreinterpretq_m128d_f64(vbslq_f64(vcltq_f64(_a, _b), _a, _b)); +#else + return vreinterpretq_m128d_f64( + vminq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#endif +#else + uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); + uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); + uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); + uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); + uint64_t d[2]; + d[0] = (*(double *) &a0) < (*(double *) &b0) ? a0 : b0; + d[1] = (*(double *) &a1) < (*(double *) &b1) ? a1 : b1; + return vreinterpretq_m128d_u64(vld1q_u64(d)); +#endif +} + +// Compare the lower double-precision (64-bit) floating-point elements in a and +// b, store the minimum value in the lower element of dst, and copy the upper +// element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_sd +FORCE_INLINE __m128d _mm_min_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_min_pd(a, b)); +#else + double *da = (double *) &a; + double *db = (double *) &b; + double c[2] = {da[0] < db[0] ? da[0] : db[0], da[1]}; + return vreinterpretq_m128d_f32(vld1q_f32((float32_t *) c)); +#endif +} + +// Copy the lower 64-bit integer in a to the lower element of dst, and zero the +// upper element. +// +// dst[63:0] := a[63:0] +// dst[127:64] := 0 +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_move_epi64 +FORCE_INLINE __m128i _mm_move_epi64(__m128i a) +{ + return vreinterpretq_m128i_s64( + vsetq_lane_s64(0, vreinterpretq_s64_m128i(a), 1)); +} + +// Move the lower double-precision (64-bit) floating-point element from b to the +// lower element of dst, and copy the upper element from a to the upper element +// of dst. +// +// dst[63:0] := b[63:0] +// dst[127:64] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_move_sd +FORCE_INLINE __m128d _mm_move_sd(__m128d a, __m128d b) +{ + return vreinterpretq_m128d_f32( + vcombine_f32(vget_low_f32(vreinterpretq_f32_m128d(b)), + vget_high_f32(vreinterpretq_f32_m128d(a)))); } // NEON does not provide a version of this function. @@ -2865,88 +4862,1193 @@ FORCE_INLINE __m128i _mm_movpi64_epi64(__m64 a) vcombine_s64(vreinterpret_s64_m64(a), vdup_n_s64(0))); } -// NEON does not provide this method -// Creates a 4-bit mask from the most significant bits of the four -// single-precision, floating-point values. -// https://msdn.microsoft.com/en-us/library/vstudio/4490ys29(v=vs.100).aspx -FORCE_INLINE int _mm_movemask_ps(__m128 a) +// Multiply the low unsigned 32-bit integers from each packed 64-bit element in +// a and b, and store the unsigned 64-bit results in dst. +// +// r0 := (a0 & 0xFFFFFFFF) * (b0 & 0xFFFFFFFF) +// r1 := (a2 & 0xFFFFFFFF) * (b2 & 0xFFFFFFFF) +FORCE_INLINE __m128i _mm_mul_epu32(__m128i a, __m128i b) +{ + // vmull_u32 upcasts instead of masking, so we downcast. + uint32x2_t a_lo = vmovn_u64(vreinterpretq_u64_m128i(a)); + uint32x2_t b_lo = vmovn_u64(vreinterpretq_u64_m128i(b)); + return vreinterpretq_m128i_u64(vmull_u32(a_lo, b_lo)); +} + +// Multiply packed double-precision (64-bit) floating-point elements in a and b, +// and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mul_pd +FORCE_INLINE __m128d _mm_mul_pd(__m128d a, __m128d b) { - uint32x4_t input = vreinterpretq_u32_m128(a); #if defined(__aarch64__) - static const int32x4_t shift = {0, 1, 2, 3}; - uint32x4_t tmp = vshrq_n_u32(input, 31); - return vaddvq_u32(vshlq_u32(tmp, shift)); + return vreinterpretq_m128d_f64( + vmulq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); #else - // Uses the exact same method as _mm_movemask_epi8, see that for details. - // Shift out everything but the sign bits with a 32-bit unsigned shift - // right. - uint64x2_t high_bits = vreinterpretq_u64_u32(vshrq_n_u32(input, 31)); - // Merge the two pairs together with a 64-bit unsigned shift right + add. - uint8x16_t paired = - vreinterpretq_u8_u64(vsraq_n_u64(high_bits, high_bits, 31)); - // Extract the result. - return vgetq_lane_u8(paired, 0) | (vgetq_lane_u8(paired, 8) << 2); + double *da = (double *) &a; + double *db = (double *) &b; + double c[2]; + c[0] = da[0] * db[0]; + c[1] = da[1] * db[1]; + return vld1q_f32((float32_t *) c); #endif } -// Compute the bitwise NOT of a and then AND with a 128-bit vector containing -// all 1's, and return 1 if the result is zero, otherwise return 0. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_test_all_ones -FORCE_INLINE int _mm_test_all_ones(__m128i a) +// Multiply the lower double-precision (64-bit) floating-point element in a and +// b, store the result in the lower element of dst, and copy the upper element +// from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_mul_sd +FORCE_INLINE __m128d _mm_mul_sd(__m128d a, __m128d b) { - return (uint64_t)(vgetq_lane_s64(a, 0) & vgetq_lane_s64(a, 1)) == - ~(uint64_t) 0; + return _mm_move_sd(a, _mm_mul_pd(a, b)); } -// Compute the bitwise AND of 128 bits (representing integer data) in a and -// mask, and return 1 if the result is zero, otherwise return 0. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_test_all_zeros -FORCE_INLINE int _mm_test_all_zeros(__m128i a, __m128i mask) +// Multiply the low unsigned 32-bit integers from a and b, and store the +// unsigned 64-bit result in dst. +// +// dst[63:0] := a[31:0] * b[31:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mul_su32 +FORCE_INLINE __m64 _mm_mul_su32(__m64 a, __m64 b) { - int64x2_t a_and_mask = - vandq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(mask)); - return (vgetq_lane_s64(a_and_mask, 0) | vgetq_lane_s64(a_and_mask, 1)) ? 0 - : 1; + return vreinterpret_m64_u64(vget_low_u64( + vmull_u32(vreinterpret_u32_m64(a), vreinterpret_u32_m64(b)))); } -/* Math operations */ - -// Subtracts the four single-precision, floating-point values of a and b. +// Multiplies the 8 signed 16-bit integers from a by the 8 signed 16-bit +// integers from b. // -// r0 := a0 - b0 -// r1 := a1 - b1 -// r2 := a2 - b2 -// r3 := a3 - b3 +// r0 := (a0 * b0)[31:16] +// r1 := (a1 * b1)[31:16] +// ... +// r7 := (a7 * b7)[31:16] // -// https://msdn.microsoft.com/en-us/library/vstudio/1zad2k61(v=vs.100).aspx -FORCE_INLINE __m128 _mm_sub_ps(__m128 a, __m128 b) +// https://msdn.microsoft.com/en-us/library/vstudio/59hddw1d(v=vs.100).aspx +FORCE_INLINE __m128i _mm_mulhi_epi16(__m128i a, __m128i b) { - return vreinterpretq_m128_f32( - vsubq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); + /* FIXME: issue with large values because of result saturation */ + // int16x8_t ret = vqdmulhq_s16(vreinterpretq_s16_m128i(a), + // vreinterpretq_s16_m128i(b)); /* =2*a*b */ return + // vreinterpretq_m128i_s16(vshrq_n_s16(ret, 1)); + int16x4_t a3210 = vget_low_s16(vreinterpretq_s16_m128i(a)); + int16x4_t b3210 = vget_low_s16(vreinterpretq_s16_m128i(b)); + int32x4_t ab3210 = vmull_s16(a3210, b3210); /* 3333222211110000 */ + int16x4_t a7654 = vget_high_s16(vreinterpretq_s16_m128i(a)); + int16x4_t b7654 = vget_high_s16(vreinterpretq_s16_m128i(b)); + int32x4_t ab7654 = vmull_s16(a7654, b7654); /* 7777666655554444 */ + uint16x8x2_t r = + vuzpq_u16(vreinterpretq_u16_s32(ab3210), vreinterpretq_u16_s32(ab7654)); + return vreinterpretq_m128i_u16(r.val[1]); } -// Subtract the lower single-precision (32-bit) floating-point element in b from -// the lower single-precision (32-bit) floating-point element in a, store the -// result in the lower element of dst, and copy the upper 3 packed elements from -// a to the upper elements of dst. -// -// dst[31:0] := a[31:0] - b[31:0] -// dst[127:32] := a[127:32] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_ss -FORCE_INLINE __m128 _mm_sub_ss(__m128 a, __m128 b) +// Multiply the packed unsigned 16-bit integers in a and b, producing +// intermediate 32-bit integers, and store the high 16 bits of the intermediate +// integers in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mulhi_epu16 +FORCE_INLINE __m128i _mm_mulhi_epu16(__m128i a, __m128i b) { - return _mm_move_ss(a, _mm_sub_ps(a, b)); + uint16x4_t a3210 = vget_low_u16(vreinterpretq_u16_m128i(a)); + uint16x4_t b3210 = vget_low_u16(vreinterpretq_u16_m128i(b)); + uint32x4_t ab3210 = vmull_u16(a3210, b3210); +#if defined(__aarch64__) + uint32x4_t ab7654 = + vmull_high_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b)); + uint16x8_t r = vuzp2q_u16(vreinterpretq_u16_u32(ab3210), + vreinterpretq_u16_u32(ab7654)); + return vreinterpretq_m128i_u16(r); +#else + uint16x4_t a7654 = vget_high_u16(vreinterpretq_u16_m128i(a)); + uint16x4_t b7654 = vget_high_u16(vreinterpretq_u16_m128i(b)); + uint32x4_t ab7654 = vmull_u16(a7654, b7654); + uint16x8x2_t r = + vuzpq_u16(vreinterpretq_u16_u32(ab3210), vreinterpretq_u16_u32(ab7654)); + return vreinterpretq_m128i_u16(r.val[1]); +#endif } -// Subtract 2 packed 64-bit integers in b from 2 packed 64-bit integers in a, -// and store the results in dst. -// r0 := a0 - b0 -// r1 := a1 - b1 -FORCE_INLINE __m128i _mm_sub_epi64(__m128i a, __m128i b) +// Multiplies the 8 signed or unsigned 16-bit integers from a by the 8 signed or +// unsigned 16-bit integers from b. +// +// r0 := (a0 * b0)[15:0] +// r1 := (a1 * b1)[15:0] +// ... +// r7 := (a7 * b7)[15:0] +// +// https://msdn.microsoft.com/en-us/library/vstudio/9ks1472s(v=vs.100).aspx +FORCE_INLINE __m128i _mm_mullo_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vmulq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} + +// Compute the bitwise OR of packed double-precision (64-bit) floating-point +// elements in a and b, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_or_pd +FORCE_INLINE __m128d _mm_or_pd(__m128d a, __m128d b) +{ + return vreinterpretq_m128d_s64( + vorrq_s64(vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b))); +} + +// Computes the bitwise OR of the 128-bit value in a and the 128-bit value in b. +// +// r := a | b +// +// https://msdn.microsoft.com/en-us/library/vstudio/ew8ty0db(v=vs.100).aspx +FORCE_INLINE __m128i _mm_or_si128(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vorrq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Packs the 16 signed 16-bit integers from a and b into 8-bit integers and +// saturates. +// https://msdn.microsoft.com/en-us/library/k4y4f7w5%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_packs_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vcombine_s8(vqmovn_s16(vreinterpretq_s16_m128i(a)), + vqmovn_s16(vreinterpretq_s16_m128i(b)))); +} + +// Packs the 8 signed 32-bit integers from a and b into signed 16-bit integers +// and saturates. +// +// r0 := SignedSaturate(a0) +// r1 := SignedSaturate(a1) +// r2 := SignedSaturate(a2) +// r3 := SignedSaturate(a3) +// r4 := SignedSaturate(b0) +// r5 := SignedSaturate(b1) +// r6 := SignedSaturate(b2) +// r7 := SignedSaturate(b3) +// +// https://msdn.microsoft.com/en-us/library/393t56f9%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_packs_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vcombine_s16(vqmovn_s32(vreinterpretq_s32_m128i(a)), + vqmovn_s32(vreinterpretq_s32_m128i(b)))); +} + +// Packs the 16 signed 16 - bit integers from a and b into 8 - bit unsigned +// integers and saturates. +// +// r0 := UnsignedSaturate(a0) +// r1 := UnsignedSaturate(a1) +// ... +// r7 := UnsignedSaturate(a7) +// r8 := UnsignedSaturate(b0) +// r9 := UnsignedSaturate(b1) +// ... +// r15 := UnsignedSaturate(b7) +// +// https://msdn.microsoft.com/en-us/library/07ad1wx4(v=vs.100).aspx +FORCE_INLINE __m128i _mm_packus_epi16(const __m128i a, const __m128i b) +{ + return vreinterpretq_m128i_u8( + vcombine_u8(vqmovun_s16(vreinterpretq_s16_m128i(a)), + vqmovun_s16(vreinterpretq_s16_m128i(b)))); +} + +// Pause the processor. This is typically used in spin-wait loops and depending +// on the x86 processor typical values are in the 40-100 cycle range. The +// 'yield' instruction isn't a good fit because it's effectively a nop on most +// Arm cores. Experience with several databases has shown has shown an 'isb' is +// a reasonable approximation. +FORCE_INLINE void _mm_pause() +{ + __asm__ __volatile__("isb\n"); +} + +// Compute the absolute differences of packed unsigned 8-bit integers in a and +// b, then horizontally sum each consecutive 8 differences to produce two +// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low +// 16 bits of 64-bit elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sad_epu8 +FORCE_INLINE __m128i _mm_sad_epu8(__m128i a, __m128i b) +{ + uint16x8_t t = vpaddlq_u8(vabdq_u8((uint8x16_t) a, (uint8x16_t) b)); + return vreinterpretq_m128i_u64(vpaddlq_u32(vpaddlq_u16(t))); +} + +// Sets the 8 signed 16-bit integer values. +// https://msdn.microsoft.com/en-au/library/3e0fek84(v=vs.90).aspx +FORCE_INLINE __m128i _mm_set_epi16(short i7, + short i6, + short i5, + short i4, + short i3, + short i2, + short i1, + short i0) +{ + int16_t ALIGN_STRUCT(16) data[8] = {i0, i1, i2, i3, i4, i5, i6, i7}; + return vreinterpretq_m128i_s16(vld1q_s16(data)); +} + +// Sets the 4 signed 32-bit integer values. +// https://msdn.microsoft.com/en-us/library/vstudio/019beekt(v=vs.100).aspx +FORCE_INLINE __m128i _mm_set_epi32(int i3, int i2, int i1, int i0) +{ + int32_t ALIGN_STRUCT(16) data[4] = {i0, i1, i2, i3}; + return vreinterpretq_m128i_s32(vld1q_s32(data)); +} + +// Returns the __m128i structure with its two 64-bit integer values +// initialized to the values of the two 64-bit integers passed in. +// https://msdn.microsoft.com/en-us/library/dk2sdw0h(v=vs.120).aspx +FORCE_INLINE __m128i _mm_set_epi64(__m64 i1, __m64 i2) +{ + return _mm_set_epi64x((int64_t) i1, (int64_t) i2); +} + +// Returns the __m128i structure with its two 64-bit integer values +// initialized to the values of the two 64-bit integers passed in. +// https://msdn.microsoft.com/en-us/library/dk2sdw0h(v=vs.120).aspx +FORCE_INLINE __m128i _mm_set_epi64x(int64_t i1, int64_t i2) { return vreinterpretq_m128i_s64( - vsubq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b))); + vcombine_s64(vcreate_s64(i2), vcreate_s64(i1))); +} + +// Sets the 16 signed 8-bit integer values. +// https://msdn.microsoft.com/en-us/library/x0cx8zd3(v=vs.90).aspx +FORCE_INLINE __m128i _mm_set_epi8(signed char b15, + signed char b14, + signed char b13, + signed char b12, + signed char b11, + signed char b10, + signed char b9, + signed char b8, + signed char b7, + signed char b6, + signed char b5, + signed char b4, + signed char b3, + signed char b2, + signed char b1, + signed char b0) +{ + int8_t ALIGN_STRUCT(16) + data[16] = {(int8_t) b0, (int8_t) b1, (int8_t) b2, (int8_t) b3, + (int8_t) b4, (int8_t) b5, (int8_t) b6, (int8_t) b7, + (int8_t) b8, (int8_t) b9, (int8_t) b10, (int8_t) b11, + (int8_t) b12, (int8_t) b13, (int8_t) b14, (int8_t) b15}; + return (__m128i) vld1q_s8(data); +} + +// Set packed double-precision (64-bit) floating-point elements in dst with the +// supplied values. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_pd +FORCE_INLINE __m128d _mm_set_pd(double e1, double e0) +{ + double ALIGN_STRUCT(16) data[2] = {e0, e1}; +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vld1q_f64((float64_t *) data)); +#else + return vreinterpretq_m128d_f32(vld1q_f32((float32_t *) data)); +#endif +} + +// Broadcast double-precision (64-bit) floating-point value a to all elements of +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_pd1 +#define _mm_set_pd1 _mm_set1_pd + +// Copy double-precision (64-bit) floating-point element a to the lower element +// of dst, and zero the upper element. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set_sd +FORCE_INLINE __m128d _mm_set_sd(double a) +{ + return _mm_set_pd(0, a); +} + +// Sets the 8 signed 16-bit integer values to w. +// +// r0 := w +// r1 := w +// ... +// r7 := w +// +// https://msdn.microsoft.com/en-us/library/k0ya3x0e(v=vs.90).aspx +FORCE_INLINE __m128i _mm_set1_epi16(short w) +{ + return vreinterpretq_m128i_s16(vdupq_n_s16(w)); +} + +// Sets the 4 signed 32-bit integer values to i. +// +// r0 := i +// r1 := i +// r2 := i +// r3 := I +// +// https://msdn.microsoft.com/en-us/library/vstudio/h4xscxat(v=vs.100).aspx +FORCE_INLINE __m128i _mm_set1_epi32(int _i) +{ + return vreinterpretq_m128i_s32(vdupq_n_s32(_i)); +} + +// Sets the 2 signed 64-bit integer values to i. +// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/whtfzhzk(v=vs.100) +FORCE_INLINE __m128i _mm_set1_epi64(__m64 _i) +{ + return vreinterpretq_m128i_s64(vdupq_n_s64((int64_t) _i)); +} + +// Sets the 2 signed 64-bit integer values to i. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set1_epi64x +FORCE_INLINE __m128i _mm_set1_epi64x(int64_t _i) +{ + return vreinterpretq_m128i_s64(vdupq_n_s64(_i)); +} + +// Sets the 16 signed 8-bit integer values to b. +// +// r0 := b +// r1 := b +// ... +// r15 := b +// +// https://msdn.microsoft.com/en-us/library/6e14xhyf(v=vs.100).aspx +FORCE_INLINE __m128i _mm_set1_epi8(signed char w) +{ + return vreinterpretq_m128i_s8(vdupq_n_s8(w)); +} + +// Broadcast double-precision (64-bit) floating-point value a to all elements of +// dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_set1_pd +FORCE_INLINE __m128d _mm_set1_pd(double d) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vdupq_n_f64(d)); +#else + return vreinterpretq_m128d_s64(vdupq_n_s64(*(int64_t *) &d)); +#endif +} + +// Sets the 8 signed 16-bit integer values in reverse order. +// +// Return Value +// r0 := w0 +// r1 := w1 +// ... +// r7 := w7 +FORCE_INLINE __m128i _mm_setr_epi16(short w0, + short w1, + short w2, + short w3, + short w4, + short w5, + short w6, + short w7) +{ + int16_t ALIGN_STRUCT(16) data[8] = {w0, w1, w2, w3, w4, w5, w6, w7}; + return vreinterpretq_m128i_s16(vld1q_s16((int16_t *) data)); +} + +// Sets the 4 signed 32-bit integer values in reverse order +// https://technet.microsoft.com/en-us/library/security/27yb3ee5(v=vs.90).aspx +FORCE_INLINE __m128i _mm_setr_epi32(int i3, int i2, int i1, int i0) +{ + int32_t ALIGN_STRUCT(16) data[4] = {i3, i2, i1, i0}; + return vreinterpretq_m128i_s32(vld1q_s32(data)); +} + +// Set packed 64-bit integers in dst with the supplied values in reverse order. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_setr_epi64 +FORCE_INLINE __m128i _mm_setr_epi64(__m64 e1, __m64 e0) +{ + return vreinterpretq_m128i_s64(vcombine_s64(e1, e0)); +} + +// Sets the 16 signed 8-bit integer values in reverse order. +// https://msdn.microsoft.com/en-us/library/2khb9c7k(v=vs.90).aspx +FORCE_INLINE __m128i _mm_setr_epi8(signed char b0, + signed char b1, + signed char b2, + signed char b3, + signed char b4, + signed char b5, + signed char b6, + signed char b7, + signed char b8, + signed char b9, + signed char b10, + signed char b11, + signed char b12, + signed char b13, + signed char b14, + signed char b15) +{ + int8_t ALIGN_STRUCT(16) + data[16] = {(int8_t) b0, (int8_t) b1, (int8_t) b2, (int8_t) b3, + (int8_t) b4, (int8_t) b5, (int8_t) b6, (int8_t) b7, + (int8_t) b8, (int8_t) b9, (int8_t) b10, (int8_t) b11, + (int8_t) b12, (int8_t) b13, (int8_t) b14, (int8_t) b15}; + return (__m128i) vld1q_s8(data); +} + +// Set packed double-precision (64-bit) floating-point elements in dst with the +// supplied values in reverse order. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_setr_pd +FORCE_INLINE __m128d _mm_setr_pd(double e1, double e0) +{ + return _mm_set_pd(e0, e1); +} + +// Return vector of type __m128d with all elements set to zero. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_setzero_pd +FORCE_INLINE __m128d _mm_setzero_pd(void) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vdupq_n_f64(0)); +#else + return vreinterpretq_m128d_f32(vdupq_n_f32(0)); +#endif +} + +// Sets the 128-bit value to zero +// https://msdn.microsoft.com/en-us/library/vstudio/ys7dw0kh(v=vs.100).aspx +FORCE_INLINE __m128i _mm_setzero_si128(void) +{ + return vreinterpretq_m128i_s32(vdupq_n_s32(0)); +} + +// Shuffles the 4 signed or unsigned 32-bit integers in a as specified by imm. +// https://msdn.microsoft.com/en-us/library/56f67xbk%28v=vs.90%29.aspx +// FORCE_INLINE __m128i _mm_shuffle_epi32(__m128i a, +// __constrange(0,255) int imm) +#if __has_builtin(__builtin_shufflevector) +#define _mm_shuffle_epi32(a, imm) \ + __extension__({ \ + int32x4_t _input = vreinterpretq_s32_m128i(a); \ + int32x4_t _shuf = __builtin_shufflevector( \ + _input, _input, (imm) & (0x3), ((imm) >> 2) & 0x3, \ + ((imm) >> 4) & 0x3, ((imm) >> 6) & 0x3); \ + vreinterpretq_m128i_s32(_shuf); \ + }) +#else // generic +#define _mm_shuffle_epi32(a, imm) \ + __extension__({ \ + __m128i ret; \ + switch (imm) { \ + case _MM_SHUFFLE(1, 0, 3, 2): \ + ret = _mm_shuffle_epi_1032((a)); \ + break; \ + case _MM_SHUFFLE(2, 3, 0, 1): \ + ret = _mm_shuffle_epi_2301((a)); \ + break; \ + case _MM_SHUFFLE(0, 3, 2, 1): \ + ret = _mm_shuffle_epi_0321((a)); \ + break; \ + case _MM_SHUFFLE(2, 1, 0, 3): \ + ret = _mm_shuffle_epi_2103((a)); \ + break; \ + case _MM_SHUFFLE(1, 0, 1, 0): \ + ret = _mm_shuffle_epi_1010((a)); \ + break; \ + case _MM_SHUFFLE(1, 0, 0, 1): \ + ret = _mm_shuffle_epi_1001((a)); \ + break; \ + case _MM_SHUFFLE(0, 1, 0, 1): \ + ret = _mm_shuffle_epi_0101((a)); \ + break; \ + case _MM_SHUFFLE(2, 2, 1, 1): \ + ret = _mm_shuffle_epi_2211((a)); \ + break; \ + case _MM_SHUFFLE(0, 1, 2, 2): \ + ret = _mm_shuffle_epi_0122((a)); \ + break; \ + case _MM_SHUFFLE(3, 3, 3, 2): \ + ret = _mm_shuffle_epi_3332((a)); \ + break; \ + case _MM_SHUFFLE(0, 0, 0, 0): \ + ret = _mm_shuffle_epi32_splat((a), 0); \ + break; \ + case _MM_SHUFFLE(1, 1, 1, 1): \ + ret = _mm_shuffle_epi32_splat((a), 1); \ + break; \ + case _MM_SHUFFLE(2, 2, 2, 2): \ + ret = _mm_shuffle_epi32_splat((a), 2); \ + break; \ + case _MM_SHUFFLE(3, 3, 3, 3): \ + ret = _mm_shuffle_epi32_splat((a), 3); \ + break; \ + default: \ + ret = _mm_shuffle_epi32_default((a), (imm)); \ + break; \ + } \ + ret; \ + }) +#endif + +// Shuffle double-precision (64-bit) floating-point elements using the control +// in imm8, and store the results in dst. +// +// dst[63:0] := (imm8[0] == 0) ? a[63:0] : a[127:64] +// dst[127:64] := (imm8[1] == 0) ? b[63:0] : b[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_pd +#if __has_builtin(__builtin_shufflevector) +#define _mm_shuffle_pd(a, b, imm8) \ + vreinterpretq_m128d_s64(__builtin_shufflevector( \ + vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b), imm8 & 0x1, \ + ((imm8 & 0x2) >> 1) + 2)) +#else +#define _mm_shuffle_pd(a, b, imm8) \ + _mm_castsi128_pd(_mm_set_epi64x( \ + vgetq_lane_s64(vreinterpretq_s64_m128d(b), (imm8 & 0x2) >> 1), \ + vgetq_lane_s64(vreinterpretq_s64_m128d(a), imm8 & 0x1))) +#endif + +// FORCE_INLINE __m128i _mm_shufflehi_epi16(__m128i a, +// __constrange(0,255) int imm) +#if __has_builtin(__builtin_shufflevector) +#define _mm_shufflehi_epi16(a, imm) \ + __extension__({ \ + int16x8_t _input = vreinterpretq_s16_m128i(a); \ + int16x8_t _shuf = __builtin_shufflevector( \ + _input, _input, 0, 1, 2, 3, ((imm) & (0x3)) + 4, \ + (((imm) >> 2) & 0x3) + 4, (((imm) >> 4) & 0x3) + 4, \ + (((imm) >> 6) & 0x3) + 4); \ + vreinterpretq_m128i_s16(_shuf); \ + }) +#else // generic +#define _mm_shufflehi_epi16(a, imm) _mm_shufflehi_epi16_function((a), (imm)) +#endif + +// FORCE_INLINE __m128i _mm_shufflelo_epi16(__m128i a, +// __constrange(0,255) int imm) +#if __has_builtin(__builtin_shufflevector) +#define _mm_shufflelo_epi16(a, imm) \ + __extension__({ \ + int16x8_t _input = vreinterpretq_s16_m128i(a); \ + int16x8_t _shuf = __builtin_shufflevector( \ + _input, _input, ((imm) & (0x3)), (((imm) >> 2) & 0x3), \ + (((imm) >> 4) & 0x3), (((imm) >> 6) & 0x3), 4, 5, 6, 7); \ + vreinterpretq_m128i_s16(_shuf); \ + }) +#else // generic +#define _mm_shufflelo_epi16(a, imm) _mm_shufflelo_epi16_function((a), (imm)) +#endif + +// Shift packed 16-bit integers in a left by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF count[63:0] > 15 +// dst[i+15:i] := 0 +// ELSE +// dst[i+15:i] := ZeroExtend16(a[i+15:i] << count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sll_epi16 +FORCE_INLINE __m128i _mm_sll_epi16(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~15)) + return _mm_setzero_si128(); + + int16x8_t vc = vdupq_n_s16((int16_t) c); + return vreinterpretq_m128i_s16(vshlq_s16(vreinterpretq_s16_m128i(a), vc)); +} + +// Shift packed 32-bit integers in a left by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF count[63:0] > 31 +// dst[i+31:i] := 0 +// ELSE +// dst[i+31:i] := ZeroExtend32(a[i+31:i] << count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sll_epi32 +FORCE_INLINE __m128i _mm_sll_epi32(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~31)) + return _mm_setzero_si128(); + + int32x4_t vc = vdupq_n_s32((int32_t) c); + return vreinterpretq_m128i_s32(vshlq_s32(vreinterpretq_s32_m128i(a), vc)); +} + +// Shift packed 64-bit integers in a left by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF count[63:0] > 63 +// dst[i+63:i] := 0 +// ELSE +// dst[i+63:i] := ZeroExtend64(a[i+63:i] << count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sll_epi64 +FORCE_INLINE __m128i _mm_sll_epi64(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~63)) + return _mm_setzero_si128(); + + int64x2_t vc = vdupq_n_s64((int64_t) c); + return vreinterpretq_m128i_s64(vshlq_s64(vreinterpretq_s64_m128i(a), vc)); +} + +// Shift packed 16-bit integers in a left by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF imm8[7:0] > 15 +// dst[i+15:i] := 0 +// ELSE +// dst[i+15:i] := ZeroExtend16(a[i+15:i] << imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_slli_epi16 +FORCE_INLINE __m128i _mm_slli_epi16(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~15)) + return _mm_setzero_si128(); + return vreinterpretq_m128i_s16( + vshlq_s16(vreinterpretq_s16_m128i(a), vdupq_n_s16(imm))); +} + +// Shift packed 32-bit integers in a left by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF imm8[7:0] > 31 +// dst[i+31:i] := 0 +// ELSE +// dst[i+31:i] := ZeroExtend32(a[i+31:i] << imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_slli_epi32 +FORCE_INLINE __m128i _mm_slli_epi32(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~31)) + return _mm_setzero_si128(); + return vreinterpretq_m128i_s32( + vshlq_s32(vreinterpretq_s32_m128i(a), vdupq_n_s32(imm))); +} + +// Shift packed 64-bit integers in a left by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF imm8[7:0] > 63 +// dst[i+63:i] := 0 +// ELSE +// dst[i+63:i] := ZeroExtend64(a[i+63:i] << imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_slli_epi64 +FORCE_INLINE __m128i _mm_slli_epi64(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~63)) + return _mm_setzero_si128(); + return vreinterpretq_m128i_s64( + vshlq_s64(vreinterpretq_s64_m128i(a), vdupq_n_s64(imm))); +} + +// Shift a left by imm8 bytes while shifting in zeros, and store the results in +// dst. +// +// tmp := imm8[7:0] +// IF tmp > 15 +// tmp := 16 +// FI +// dst[127:0] := a[127:0] << (tmp*8) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_slli_si128 +FORCE_INLINE __m128i _mm_slli_si128(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~15)) + return _mm_setzero_si128(); + uint8x16_t tmp[2] = {vdupq_n_u8(0), vreinterpretq_u8_m128i(a)}; + return vreinterpretq_m128i_u8( + vld1q_u8(((uint8_t const *) tmp) + (16 - imm))); +} + +// Compute the square root of packed double-precision (64-bit) floating-point +// elements in a, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sqrt_pd +FORCE_INLINE __m128d _mm_sqrt_pd(__m128d a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vsqrtq_f64(vreinterpretq_f64_m128d(a))); +#else + double a0 = sqrt(((double *) &a)[0]); + double a1 = sqrt(((double *) &a)[1]); + return _mm_set_pd(a1, a0); +#endif +} + +// Compute the square root of the lower double-precision (64-bit) floating-point +// element in b, store the result in the lower element of dst, and copy the +// upper element from a to the upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sqrt_sd +FORCE_INLINE __m128d _mm_sqrt_sd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return _mm_move_sd(a, _mm_sqrt_pd(b)); +#else + return _mm_set_pd(((double *) &a)[1], sqrt(((double *) &b)[0])); +#endif +} + +// Shift packed 16-bit integers in a right by count while shifting in sign bits, +// and store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF count[63:0] > 15 +// dst[i+15:i] := (a[i+15] ? 0xFFFF : 0x0) +// ELSE +// dst[i+15:i] := SignExtend16(a[i+15:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sra_epi16 +FORCE_INLINE __m128i _mm_sra_epi16(__m128i a, __m128i count) +{ + int64_t c = (int64_t) vget_low_s64((int64x2_t) count); + if (_sse2neon_unlikely(c & ~15)) + return _mm_cmplt_epi16(a, _mm_setzero_si128()); + return vreinterpretq_m128i_s16(vshlq_s16((int16x8_t) a, vdupq_n_s16(-c))); +} + +// Shift packed 32-bit integers in a right by count while shifting in sign bits, +// and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF count[63:0] > 31 +// dst[i+31:i] := (a[i+31] ? 0xFFFFFFFF : 0x0) +// ELSE +// dst[i+31:i] := SignExtend32(a[i+31:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sra_epi32 +FORCE_INLINE __m128i _mm_sra_epi32(__m128i a, __m128i count) +{ + int64_t c = (int64_t) vget_low_s64((int64x2_t) count); + if (_sse2neon_unlikely(c & ~31)) + return _mm_cmplt_epi32(a, _mm_setzero_si128()); + return vreinterpretq_m128i_s32(vshlq_s32((int32x4_t) a, vdupq_n_s32(-c))); +} + +// Shift packed 16-bit integers in a right by imm8 while shifting in sign +// bits, and store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF imm8[7:0] > 15 +// dst[i+15:i] := (a[i+15] ? 0xFFFF : 0x0) +// ELSE +// dst[i+15:i] := SignExtend16(a[i+15:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srai_epi16 +FORCE_INLINE __m128i _mm_srai_epi16(__m128i a, int imm) +{ + const int count = (imm & ~15) ? 15 : imm; + return (__m128i) vshlq_s16((int16x8_t) a, vdupq_n_s16(-count)); +} + +// Shift packed 32-bit integers in a right by imm8 while shifting in sign bits, +// and store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF imm8[7:0] > 31 +// dst[i+31:i] := (a[i+31] ? 0xFFFFFFFF : 0x0) +// ELSE +// dst[i+31:i] := SignExtend32(a[i+31:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srai_epi32 +// FORCE_INLINE __m128i _mm_srai_epi32(__m128i a, __constrange(0,255) int imm) +#define _mm_srai_epi32(a, imm) \ + __extension__({ \ + __m128i ret; \ + if (_sse2neon_unlikely((imm) == 0)) { \ + ret = a; \ + } else if (_sse2neon_likely(0 < (imm) && (imm) < 32)) { \ + ret = vreinterpretq_m128i_s32( \ + vshlq_s32(vreinterpretq_s32_m128i(a), vdupq_n_s32(-imm))); \ + } else { \ + ret = vreinterpretq_m128i_s32( \ + vshrq_n_s32(vreinterpretq_s32_m128i(a), 31)); \ + } \ + ret; \ + }) + +// Shift packed 16-bit integers in a right by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF count[63:0] > 15 +// dst[i+15:i] := 0 +// ELSE +// dst[i+15:i] := ZeroExtend16(a[i+15:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srl_epi16 +FORCE_INLINE __m128i _mm_srl_epi16(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~15)) + return _mm_setzero_si128(); + + int16x8_t vc = vdupq_n_s16(-(int16_t) c); + return vreinterpretq_m128i_u16(vshlq_u16(vreinterpretq_u16_m128i(a), vc)); +} + +// Shift packed 32-bit integers in a right by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF count[63:0] > 31 +// dst[i+31:i] := 0 +// ELSE +// dst[i+31:i] := ZeroExtend32(a[i+31:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srl_epi32 +FORCE_INLINE __m128i _mm_srl_epi32(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~31)) + return _mm_setzero_si128(); + + int32x4_t vc = vdupq_n_s32(-(int32_t) c); + return vreinterpretq_m128i_u32(vshlq_u32(vreinterpretq_u32_m128i(a), vc)); +} + +// Shift packed 64-bit integers in a right by count while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF count[63:0] > 63 +// dst[i+63:i] := 0 +// ELSE +// dst[i+63:i] := ZeroExtend64(a[i+63:i] >> count[63:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srl_epi64 +FORCE_INLINE __m128i _mm_srl_epi64(__m128i a, __m128i count) +{ + uint64_t c = vreinterpretq_nth_u64_m128i(count, 0); + if (_sse2neon_unlikely(c & ~63)) + return _mm_setzero_si128(); + + int64x2_t vc = vdupq_n_s64(-(int64_t) c); + return vreinterpretq_m128i_u64(vshlq_u64(vreinterpretq_u64_m128i(a), vc)); +} + +// Shift packed 16-bit integers in a right by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// IF imm8[7:0] > 15 +// dst[i+15:i] := 0 +// ELSE +// dst[i+15:i] := ZeroExtend16(a[i+15:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_epi16 +#define _mm_srli_epi16(a, imm) \ + __extension__({ \ + __m128i ret; \ + if (_sse2neon_unlikely((imm) & ~15)) { \ + ret = _mm_setzero_si128(); \ + } else { \ + ret = vreinterpretq_m128i_u16( \ + vshlq_u16(vreinterpretq_u16_m128i(a), vdupq_n_s16(-(imm)))); \ + } \ + ret; \ + }) + +// Shift packed 32-bit integers in a right by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// IF imm8[7:0] > 31 +// dst[i+31:i] := 0 +// ELSE +// dst[i+31:i] := ZeroExtend32(a[i+31:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_epi32 +// FORCE_INLINE __m128i _mm_srli_epi32(__m128i a, __constrange(0,255) int imm) +#define _mm_srli_epi32(a, imm) \ + __extension__({ \ + __m128i ret; \ + if (_sse2neon_unlikely((imm) & ~31)) { \ + ret = _mm_setzero_si128(); \ + } else { \ + ret = vreinterpretq_m128i_u32( \ + vshlq_u32(vreinterpretq_u32_m128i(a), vdupq_n_s32(-(imm)))); \ + } \ + ret; \ + }) + +// Shift packed 64-bit integers in a right by imm8 while shifting in zeros, and +// store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF imm8[7:0] > 63 +// dst[i+63:i] := 0 +// ELSE +// dst[i+63:i] := ZeroExtend64(a[i+63:i] >> imm8[7:0]) +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_epi64 +#define _mm_srli_epi64(a, imm) \ + __extension__({ \ + __m128i ret; \ + if (_sse2neon_unlikely((imm) & ~63)) { \ + ret = _mm_setzero_si128(); \ + } else { \ + ret = vreinterpretq_m128i_u64( \ + vshlq_u64(vreinterpretq_u64_m128i(a), vdupq_n_s64(-(imm)))); \ + } \ + ret; \ + }) + +// Shift a right by imm8 bytes while shifting in zeros, and store the results in +// dst. +// +// tmp := imm8[7:0] +// IF tmp > 15 +// tmp := 16 +// FI +// dst[127:0] := a[127:0] >> (tmp*8) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_srli_si128 +FORCE_INLINE __m128i _mm_srli_si128(__m128i a, int imm) +{ + if (_sse2neon_unlikely(imm & ~15)) + return _mm_setzero_si128(); + uint8x16_t tmp[2] = {vreinterpretq_u8_m128i(a), vdupq_n_u8(0)}; + return vreinterpretq_m128i_u8(vld1q_u8(((uint8_t const *) tmp) + imm)); +} + +// Store 128-bits (composed of 2 packed double-precision (64-bit) floating-point +// elements) from a into memory. mem_addr must be aligned on a 16-byte boundary +// or a general-protection exception may be generated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store_pd +FORCE_INLINE void _mm_store_pd(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + vst1q_f64((float64_t *) mem_addr, vreinterpretq_f64_m128d(a)); +#else + vst1q_f32((float32_t *) mem_addr, vreinterpretq_f32_m128d(a)); +#endif +} + +// Store the lower double-precision (64-bit) floating-point element from a into +// 2 contiguous elements in memory. mem_addr must be aligned on a 16-byte +// boundary or a general-protection exception may be generated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_store_pd1 +FORCE_INLINE void _mm_store_pd1(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + float64x1_t a_low = vget_low_f64(vreinterpretq_f64_m128d(a)); + vst1q_f64((float64_t *) mem_addr, + vreinterpretq_f64_m128d(vcombine_f64(a_low, a_low))); +#else + float32x2_t a_low = vget_low_f32(vreinterpretq_f32_m128d(a)); + vst1q_f32((float32_t *) mem_addr, + vreinterpretq_f32_m128d(vcombine_f32(a_low, a_low))); +#endif +} + +// Store the lower double-precision (64-bit) floating-point element from a into +// memory. mem_addr does not need to be aligned on any particular boundary. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_store_sd +FORCE_INLINE void _mm_store_sd(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + vst1_f64((float64_t *) mem_addr, vget_low_f64(vreinterpretq_f64_m128d(a))); +#else + vst1_u64((uint64_t *) mem_addr, vget_low_u64(vreinterpretq_u64_m128d(a))); +#endif +} + +// Stores four 32-bit integer values as (as a __m128i value) at the address p. +// https://msdn.microsoft.com/en-us/library/vstudio/edk11s13(v=vs.100).aspx +FORCE_INLINE void _mm_store_si128(__m128i *p, __m128i a) +{ + vst1q_s32((int32_t *) p, vreinterpretq_s32_m128i(a)); +} + +// Store the lower double-precision (64-bit) floating-point element from a into +// 2 contiguous elements in memory. mem_addr must be aligned on a 16-byte +// boundary or a general-protection exception may be generated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=9,526,5601&text=_mm_store1_pd +#define _mm_store1_pd _mm_store_pd1 + +// Store the upper double-precision (64-bit) floating-point element from a into +// memory. +// +// MEM[mem_addr+63:mem_addr] := a[127:64] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeh_pd +FORCE_INLINE void _mm_storeh_pd(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + vst1_f64((float64_t *) mem_addr, vget_high_f64(vreinterpretq_f64_m128d(a))); +#else + vst1_f32((float32_t *) mem_addr, vget_high_f32(vreinterpretq_f32_m128d(a))); +#endif +} + +// Reads the lower 64 bits of b and stores them into the lower 64 bits of a. +// https://msdn.microsoft.com/en-us/library/hhwf428f%28v=vs.90%29.aspx +FORCE_INLINE void _mm_storel_epi64(__m128i *a, __m128i b) +{ + vst1_u64((uint64_t *) a, vget_low_u64(vreinterpretq_u64_m128i(b))); +} + +// Store the lower double-precision (64-bit) floating-point element from a into +// memory. +// +// MEM[mem_addr+63:mem_addr] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storel_pd +FORCE_INLINE void _mm_storel_pd(double *mem_addr, __m128d a) +{ +#if defined(__aarch64__) + vst1_f64((float64_t *) mem_addr, vget_low_f64(vreinterpretq_f64_m128d(a))); +#else + vst1_f32((float32_t *) mem_addr, vget_low_f32(vreinterpretq_f32_m128d(a))); +#endif +} + +// Store 2 double-precision (64-bit) floating-point elements from a into memory +// in reverse order. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// MEM[mem_addr+63:mem_addr] := a[127:64] +// MEM[mem_addr+127:mem_addr+64] := a[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storer_pd +FORCE_INLINE void _mm_storer_pd(double *mem_addr, __m128d a) +{ + float32x4_t f = vreinterpretq_f32_m128d(a); + _mm_store_pd(mem_addr, vreinterpretq_m128d_f32(vextq_f32(f, f, 2))); +} + +// Store 128-bits (composed of 2 packed double-precision (64-bit) floating-point +// elements) from a into memory. mem_addr does not need to be aligned on any +// particular boundary. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_pd +FORCE_INLINE void _mm_storeu_pd(double *mem_addr, __m128d a) +{ + _mm_store_pd(mem_addr, a); +} + +// Stores 128-bits of integer data a at the address p. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si128 +FORCE_INLINE void _mm_storeu_si128(__m128i *p, __m128i a) +{ + vst1q_s32((int32_t *) p, vreinterpretq_s32_m128i(a)); +} + +// Stores 32-bits of integer data a at the address p. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_storeu_si32 +FORCE_INLINE void _mm_storeu_si32(void *p, __m128i a) +{ + vst1q_lane_s32((int32_t *) p, vreinterpretq_s32_m128i(a), 0); +} + +// Store 128-bits (composed of 2 packed double-precision (64-bit) floating-point +// elements) from a into memory using a non-temporal memory hint. mem_addr must +// be aligned on a 16-byte boundary or a general-protection exception may be +// generated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_pd +FORCE_INLINE void _mm_stream_pd(double *p, __m128d a) +{ +#if __has_builtin(__builtin_nontemporal_store) + __builtin_nontemporal_store(a, (float32x4_t *) p); +#elif defined(__aarch64__) + vst1q_f64(p, vreinterpretq_f64_m128d(a)); +#else + vst1q_s64((int64_t *) p, vreinterpretq_s64_m128d(a)); +#endif +} + +// Stores the data in a to the address p without polluting the caches. If the +// cache line containing address p is already in the cache, the cache will be +// updated. +// https://msdn.microsoft.com/en-us/library/ba08y07y%28v=vs.90%29.aspx +FORCE_INLINE void _mm_stream_si128(__m128i *p, __m128i a) +{ +#if __has_builtin(__builtin_nontemporal_store) + __builtin_nontemporal_store(a, p); +#else + vst1q_s64((int64_t *) p, vreinterpretq_s64_m128i(a)); +#endif +} + +// Store 32-bit integer a into memory using a non-temporal hint to minimize +// cache pollution. If the cache line containing address mem_addr is already in +// the cache, the cache will be updated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_si32 +FORCE_INLINE void _mm_stream_si32(int *p, int a) +{ + vst1q_lane_s32((int32_t *) p, vdupq_n_s32(a), 0); +} + +// Store 64-bit integer a into memory using a non-temporal hint to minimize +// cache pollution. If the cache line containing address mem_addr is already in +// the cache, the cache will be updated. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_si64 +FORCE_INLINE void _mm_stream_si64(__int64 *p, __int64 a) +{ + vst1_s64((int64_t *) p, vdup_n_s64((int64_t) a)); +} + +// Subtract packed 16-bit integers in b from packed 16-bit integers in a, and +// store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_epi16 +FORCE_INLINE __m128i _mm_sub_epi16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s16( + vsubq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); } // Subtracts the 4 signed or unsigned 32-bit integers of b from the 4 signed or @@ -2964,13 +6066,14 @@ FORCE_INLINE __m128i _mm_sub_epi32(__m128i a, __m128i b) vsubq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); } -// Subtract packed 16-bit integers in b from packed 16-bit integers in a, and -// store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_epi16 -FORCE_INLINE __m128i _mm_sub_epi16(__m128i a, __m128i b) +// Subtract 2 packed 64-bit integers in b from 2 packed 64-bit integers in a, +// and store the results in dst. +// r0 := a0 - b0 +// r1 := a1 - b1 +FORCE_INLINE __m128i _mm_sub_epi64(__m128i a, __m128i b) { - return vreinterpretq_m128i_s16( - vsubq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); + return vreinterpretq_m128i_s64( + vsubq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b))); } // Subtract packed 8-bit integers in b from packed 8-bit integers in a, and @@ -2982,78 +6085,6 @@ FORCE_INLINE __m128i _mm_sub_epi8(__m128i a, __m128i b) vsubq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); } -// Subtract 64-bit integer b from 64-bit integer a, and store the result in dst. -// -// dst[63:0] := a[63:0] - b[63:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_si64 -FORCE_INLINE __m64 _mm_sub_si64(__m64 a, __m64 b) -{ - return vreinterpret_m64_s64( - vsub_s64(vreinterpret_s64_m64(a), vreinterpret_s64_m64(b))); -} - -// Subtracts the 8 unsigned 16-bit integers of bfrom the 8 unsigned 16-bit -// integers of a and saturates.. -// https://technet.microsoft.com/en-us/subscriptions/index/f44y0s19(v=vs.90).aspx -FORCE_INLINE __m128i _mm_subs_epu16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u16( - vqsubq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); -} - -// Subtracts the 16 unsigned 8-bit integers of b from the 16 unsigned 8-bit -// integers of a and saturates. -// -// r0 := UnsignedSaturate(a0 - b0) -// r1 := UnsignedSaturate(a1 - b1) -// ... -// r15 := UnsignedSaturate(a15 - b15) -// -// https://technet.microsoft.com/en-us/subscriptions/yadkxc18(v=vs.90) -FORCE_INLINE __m128i _mm_subs_epu8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u8( - vqsubq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); -} - -#define _mm_ucomieq_sd _mm_comieq_sd -#define _mm_ucomige_sd _mm_comige_sd -#define _mm_ucomigt_sd _mm_comigt_sd -#define _mm_ucomile_sd _mm_comile_sd -#define _mm_ucomilt_sd _mm_comilt_sd -#define _mm_ucomineq_sd _mm_comineq_sd - -// Subtracts the 16 signed 8-bit integers of b from the 16 signed 8-bit integers -// of a and saturates. -// -// r0 := SignedSaturate(a0 - b0) -// r1 := SignedSaturate(a1 - b1) -// ... -// r15 := SignedSaturate(a15 - b15) -// -// https://technet.microsoft.com/en-us/subscriptions/by7kzks1(v=vs.90) -FORCE_INLINE __m128i _mm_subs_epi8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s8( - vqsubq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -} - -// Subtracts the 8 signed 16-bit integers of b from the 8 signed 16-bit integers -// of a and saturates. -// -// r0 := SignedSaturate(a0 - b0) -// r1 := SignedSaturate(a1 - b1) -// ... -// r7 := SignedSaturate(a7 - b7) -// -// https://technet.microsoft.com/en-us/subscriptions/3247z5b8(v=vs.90) -FORCE_INLINE __m128i _mm_subs_epi16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s16( - vqsubq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -} - // Subtract packed double-precision (64-bit) floating-point elements in b from // packed double-precision (64-bit) floating-point elements in a, and store the // results in dst. @@ -3089,52 +6120,975 @@ FORCE_INLINE __m128d _mm_sub_sd(__m128d a, __m128d b) return _mm_move_sd(a, _mm_sub_pd(a, b)); } -// Add packed unsigned 16-bit integers in a and b using saturation, and store -// the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_adds_epu16 -FORCE_INLINE __m128i _mm_adds_epu16(__m128i a, __m128i b) +// Subtract 64-bit integer b from 64-bit integer a, and store the result in dst. +// +// dst[63:0] := a[63:0] - b[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sub_si64 +FORCE_INLINE __m64 _mm_sub_si64(__m64 a, __m64 b) { - return vreinterpretq_m128i_u16( - vqaddq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); + return vreinterpret_m64_s64( + vsub_s64(vreinterpret_s64_m64(a), vreinterpret_s64_m64(b))); } -// Negate packed 8-bit integers in a when the corresponding signed -// 8-bit integer in b is negative, and store the results in dst. -// Element in dst are zeroed out when the corresponding element -// in b is zero. +// Subtracts the 8 signed 16-bit integers of b from the 8 signed 16-bit integers +// of a and saturates. // -// for i in 0..15 -// if b[i] < 0 -// r[i] := -a[i] -// else if b[i] == 0 -// r[i] := 0 -// else -// r[i] := a[i] -// fi -// done -FORCE_INLINE __m128i _mm_sign_epi8(__m128i _a, __m128i _b) +// r0 := SignedSaturate(a0 - b0) +// r1 := SignedSaturate(a1 - b1) +// ... +// r7 := SignedSaturate(a7 - b7) +// +// https://technet.microsoft.com/en-us/subscriptions/3247z5b8(v=vs.90) +FORCE_INLINE __m128i _mm_subs_epi16(__m128i a, __m128i b) { - int8x16_t a = vreinterpretq_s8_m128i(_a); - int8x16_t b = vreinterpretq_s8_m128i(_b); + return vreinterpretq_m128i_s16( + vqsubq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +} - // signed shift right: faster than vclt - // (b < 0) ? 0xFF : 0 - uint8x16_t ltMask = vreinterpretq_u8_s8(vshrq_n_s8(b, 7)); +// Subtracts the 16 signed 8-bit integers of b from the 16 signed 8-bit integers +// of a and saturates. +// +// r0 := SignedSaturate(a0 - b0) +// r1 := SignedSaturate(a1 - b1) +// ... +// r15 := SignedSaturate(a15 - b15) +// +// https://technet.microsoft.com/en-us/subscriptions/by7kzks1(v=vs.90) +FORCE_INLINE __m128i _mm_subs_epi8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s8( + vqsubq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +} - // (b == 0) ? 0xFF : 0 -#if defined(__aarch64__) - int8x16_t zeroMask = vreinterpretq_s8_u8(vceqzq_s8(b)); -#else - int8x16_t zeroMask = vreinterpretq_s8_u8(vceqq_s8(b, vdupq_n_s8(0))); +// Subtracts the 8 unsigned 16-bit integers of bfrom the 8 unsigned 16-bit +// integers of a and saturates.. +// https://technet.microsoft.com/en-us/subscriptions/index/f44y0s19(v=vs.90).aspx +FORCE_INLINE __m128i _mm_subs_epu16(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vqsubq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); +} + +// Subtracts the 16 unsigned 8-bit integers of b from the 16 unsigned 8-bit +// integers of a and saturates. +// +// r0 := UnsignedSaturate(a0 - b0) +// r1 := UnsignedSaturate(a1 - b1) +// ... +// r15 := UnsignedSaturate(a15 - b15) +// +// https://technet.microsoft.com/en-us/subscriptions/yadkxc18(v=vs.90) +FORCE_INLINE __m128i _mm_subs_epu8(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u8( + vqsubq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); +} + +#define _mm_ucomieq_sd _mm_comieq_sd +#define _mm_ucomige_sd _mm_comige_sd +#define _mm_ucomigt_sd _mm_comigt_sd +#define _mm_ucomile_sd _mm_comile_sd +#define _mm_ucomilt_sd _mm_comilt_sd +#define _mm_ucomineq_sd _mm_comineq_sd + +// Return vector of type __m128d with undefined elements. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_undefined_pd +FORCE_INLINE __m128d _mm_undefined_pd(void) +{ +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wuninitialized" #endif + __m128d a; + return a; +#if defined(__GNUC__) || defined(__clang__) +#pragma GCC diagnostic pop +#endif +} - // bitwise select either a or nagative 'a' (vnegq_s8(a) return nagative 'a') - // based on ltMask - int8x16_t masked = vbslq_s8(ltMask, vnegq_s8(a), a); - // res = masked & (~zeroMask) - int8x16_t res = vbicq_s8(masked, zeroMask); +// Interleaves the upper 4 signed or unsigned 16-bit integers in a with the +// upper 4 signed or unsigned 16-bit integers in b. +// +// r0 := a4 +// r1 := b4 +// r2 := a5 +// r3 := b5 +// r4 := a6 +// r5 := b6 +// r6 := a7 +// r7 := b7 +// +// https://msdn.microsoft.com/en-us/library/03196cz7(v=vs.100).aspx +FORCE_INLINE __m128i _mm_unpackhi_epi16(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s16( + vzip2q_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +#else + int16x4_t a1 = vget_high_s16(vreinterpretq_s16_m128i(a)); + int16x4_t b1 = vget_high_s16(vreinterpretq_s16_m128i(b)); + int16x4x2_t result = vzip_s16(a1, b1); + return vreinterpretq_m128i_s16(vcombine_s16(result.val[0], result.val[1])); +#endif +} - return vreinterpretq_m128i_s8(res); +// Interleaves the upper 2 signed or unsigned 32-bit integers in a with the +// upper 2 signed or unsigned 32-bit integers in b. +// https://msdn.microsoft.com/en-us/library/65sa7cbs(v=vs.100).aspx +FORCE_INLINE __m128i _mm_unpackhi_epi32(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s32( + vzip2q_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +#else + int32x2_t a1 = vget_high_s32(vreinterpretq_s32_m128i(a)); + int32x2_t b1 = vget_high_s32(vreinterpretq_s32_m128i(b)); + int32x2x2_t result = vzip_s32(a1, b1); + return vreinterpretq_m128i_s32(vcombine_s32(result.val[0], result.val[1])); +#endif +} + +// Interleaves the upper signed or unsigned 64-bit integer in a with the +// upper signed or unsigned 64-bit integer in b. +// +// r0 := a1 +// r1 := b1 +FORCE_INLINE __m128i _mm_unpackhi_epi64(__m128i a, __m128i b) +{ + int64x1_t a_h = vget_high_s64(vreinterpretq_s64_m128i(a)); + int64x1_t b_h = vget_high_s64(vreinterpretq_s64_m128i(b)); + return vreinterpretq_m128i_s64(vcombine_s64(a_h, b_h)); +} + +// Interleaves the upper 8 signed or unsigned 8-bit integers in a with the upper +// 8 signed or unsigned 8-bit integers in b. +// +// r0 := a8 +// r1 := b8 +// r2 := a9 +// r3 := b9 +// ... +// r14 := a15 +// r15 := b15 +// +// https://msdn.microsoft.com/en-us/library/t5h7783k(v=vs.100).aspx +FORCE_INLINE __m128i _mm_unpackhi_epi8(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s8( + vzip2q_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +#else + int8x8_t a1 = + vreinterpret_s8_s16(vget_high_s16(vreinterpretq_s16_m128i(a))); + int8x8_t b1 = + vreinterpret_s8_s16(vget_high_s16(vreinterpretq_s16_m128i(b))); + int8x8x2_t result = vzip_s8(a1, b1); + return vreinterpretq_m128i_s8(vcombine_s8(result.val[0], result.val[1])); +#endif +} + +// Unpack and interleave double-precision (64-bit) floating-point elements from +// the high half of a and b, and store the results in dst. +// +// DEFINE INTERLEAVE_HIGH_QWORDS(src1[127:0], src2[127:0]) { +// dst[63:0] := src1[127:64] +// dst[127:64] := src2[127:64] +// RETURN dst[127:0] +// } +// dst[127:0] := INTERLEAVE_HIGH_QWORDS(a[127:0], b[127:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_unpackhi_pd +FORCE_INLINE __m128d _mm_unpackhi_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vzip2q_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + return vreinterpretq_m128d_s64( + vcombine_s64(vget_high_s64(vreinterpretq_s64_m128d(a)), + vget_high_s64(vreinterpretq_s64_m128d(b)))); +#endif +} + +// Interleaves the lower 4 signed or unsigned 16-bit integers in a with the +// lower 4 signed or unsigned 16-bit integers in b. +// +// r0 := a0 +// r1 := b0 +// r2 := a1 +// r3 := b1 +// r4 := a2 +// r5 := b2 +// r6 := a3 +// r7 := b3 +// +// https://msdn.microsoft.com/en-us/library/btxb17bw%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_unpacklo_epi16(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s16( + vzip1q_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); +#else + int16x4_t a1 = vget_low_s16(vreinterpretq_s16_m128i(a)); + int16x4_t b1 = vget_low_s16(vreinterpretq_s16_m128i(b)); + int16x4x2_t result = vzip_s16(a1, b1); + return vreinterpretq_m128i_s16(vcombine_s16(result.val[0], result.val[1])); +#endif +} + +// Interleaves the lower 2 signed or unsigned 32 - bit integers in a with the +// lower 2 signed or unsigned 32 - bit integers in b. +// +// r0 := a0 +// r1 := b0 +// r2 := a1 +// r3 := b1 +// +// https://msdn.microsoft.com/en-us/library/x8atst9d(v=vs.100).aspx +FORCE_INLINE __m128i _mm_unpacklo_epi32(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s32( + vzip1q_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +#else + int32x2_t a1 = vget_low_s32(vreinterpretq_s32_m128i(a)); + int32x2_t b1 = vget_low_s32(vreinterpretq_s32_m128i(b)); + int32x2x2_t result = vzip_s32(a1, b1); + return vreinterpretq_m128i_s32(vcombine_s32(result.val[0], result.val[1])); +#endif +} + +FORCE_INLINE __m128i _mm_unpacklo_epi64(__m128i a, __m128i b) +{ + int64x1_t a_l = vget_low_s64(vreinterpretq_s64_m128i(a)); + int64x1_t b_l = vget_low_s64(vreinterpretq_s64_m128i(b)); + return vreinterpretq_m128i_s64(vcombine_s64(a_l, b_l)); +} + +// Interleaves the lower 8 signed or unsigned 8-bit integers in a with the lower +// 8 signed or unsigned 8-bit integers in b. +// +// r0 := a0 +// r1 := b0 +// r2 := a1 +// r3 := b1 +// ... +// r14 := a7 +// r15 := b7 +// +// https://msdn.microsoft.com/en-us/library/xf7k860c%28v=vs.90%29.aspx +FORCE_INLINE __m128i _mm_unpacklo_epi8(__m128i a, __m128i b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128i_s8( + vzip1q_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); +#else + int8x8_t a1 = vreinterpret_s8_s16(vget_low_s16(vreinterpretq_s16_m128i(a))); + int8x8_t b1 = vreinterpret_s8_s16(vget_low_s16(vreinterpretq_s16_m128i(b))); + int8x8x2_t result = vzip_s8(a1, b1); + return vreinterpretq_m128i_s8(vcombine_s8(result.val[0], result.val[1])); +#endif +} + +// Unpack and interleave double-precision (64-bit) floating-point elements from +// the low half of a and b, and store the results in dst. +// +// DEFINE INTERLEAVE_QWORDS(src1[127:0], src2[127:0]) { +// dst[63:0] := src1[63:0] +// dst[127:64] := src2[63:0] +// RETURN dst[127:0] +// } +// dst[127:0] := INTERLEAVE_QWORDS(a[127:0], b[127:0]) +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_unpacklo_pd +FORCE_INLINE __m128d _mm_unpacklo_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vzip1q_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + return vreinterpretq_m128d_s64( + vcombine_s64(vget_low_s64(vreinterpretq_s64_m128d(a)), + vget_low_s64(vreinterpretq_s64_m128d(b)))); +#endif +} + +// Compute the bitwise XOR of packed double-precision (64-bit) floating-point +// elements in a and b, and store the results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// dst[i+63:i] := a[i+63:i] XOR b[i+63:i] +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_xor_pd +FORCE_INLINE __m128d _mm_xor_pd(__m128d a, __m128d b) +{ + return vreinterpretq_m128d_s64( + veorq_s64(vreinterpretq_s64_m128d(a), vreinterpretq_s64_m128d(b))); +} + +// Computes the bitwise XOR of the 128-bit value in a and the 128-bit value in +// b. https://msdn.microsoft.com/en-us/library/fzt08www(v=vs.100).aspx +FORCE_INLINE __m128i _mm_xor_si128(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + veorq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +/* SSE3 */ + +// Alternatively add and subtract packed double-precision (64-bit) +// floating-point elements in a to/from packed elements in b, and store the +// results in dst. +// +// FOR j := 0 to 1 +// i := j*64 +// IF ((j & 1) == 0) +// dst[i+63:i] := a[i+63:i] - b[i+63:i] +// ELSE +// dst[i+63:i] := a[i+63:i] + b[i+63:i] +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_addsub_pd +FORCE_INLINE __m128d _mm_addsub_pd(__m128d a, __m128d b) +{ + _sse2neon_const __m128d mask = _mm_set_pd(1.0f, -1.0f); +#if defined(__aarch64__) + return vreinterpretq_m128d_f64(vfmaq_f64(vreinterpretq_f64_m128d(a), + vreinterpretq_f64_m128d(b), + vreinterpretq_f64_m128d(mask))); +#else + return _mm_add_pd(_mm_mul_pd(b, mask), a); +#endif +} + +// Alternatively add and subtract packed single-precision (32-bit) +// floating-point elements in a to/from packed elements in b, and store the +// results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=addsub_ps +FORCE_INLINE __m128 _mm_addsub_ps(__m128 a, __m128 b) +{ + _sse2neon_const __m128 mask = _mm_setr_ps(-1.0f, 1.0f, -1.0f, 1.0f); +#if defined(__aarch64__) || defined(__ARM_FEATURE_FMA) /* VFPv4+ */ + return vreinterpretq_m128_f32(vfmaq_f32(vreinterpretq_f32_m128(a), + vreinterpretq_f32_m128(mask), + vreinterpretq_f32_m128(b))); +#else + return _mm_add_ps(_mm_mul_ps(b, mask), a); +#endif +} + +// Horizontally add adjacent pairs of double-precision (64-bit) floating-point +// elements in a and b, and pack the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadd_pd +FORCE_INLINE __m128d _mm_hadd_pd(__m128d a, __m128d b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vpaddq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); +#else + double *da = (double *) &a; + double *db = (double *) &b; + double c[] = {da[0] + da[1], db[0] + db[1]}; + return vreinterpretq_m128d_u64(vld1q_u64((uint64_t *) c)); +#endif +} + +// Computes pairwise add of each argument as single-precision, floating-point +// values a and b. +// https://msdn.microsoft.com/en-us/library/yd9wecaa.aspx +FORCE_INLINE __m128 _mm_hadd_ps(__m128 a, __m128 b) +{ +#if defined(__aarch64__) + return vreinterpretq_m128_f32( + vpaddq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); +#else + float32x2_t a10 = vget_low_f32(vreinterpretq_f32_m128(a)); + float32x2_t a32 = vget_high_f32(vreinterpretq_f32_m128(a)); + float32x2_t b10 = vget_low_f32(vreinterpretq_f32_m128(b)); + float32x2_t b32 = vget_high_f32(vreinterpretq_f32_m128(b)); + return vreinterpretq_m128_f32( + vcombine_f32(vpadd_f32(a10, a32), vpadd_f32(b10, b32))); +#endif +} + +// Horizontally subtract adjacent pairs of double-precision (64-bit) +// floating-point elements in a and b, and pack the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_pd +FORCE_INLINE __m128d _mm_hsub_pd(__m128d _a, __m128d _b) +{ +#if defined(__aarch64__) + float64x2_t a = vreinterpretq_f64_m128d(_a); + float64x2_t b = vreinterpretq_f64_m128d(_b); + return vreinterpretq_m128d_f64( + vsubq_f64(vuzp1q_f64(a, b), vuzp2q_f64(a, b))); +#else + double *da = (double *) &_a; + double *db = (double *) &_b; + double c[] = {da[0] - da[1], db[0] - db[1]}; + return vreinterpretq_m128d_u64(vld1q_u64((uint64_t *) c)); +#endif +} + +// Horizontally subtract adjacent pairs of single-precision (32-bit) +// floating-point elements in a and b, and pack the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_ps +FORCE_INLINE __m128 _mm_hsub_ps(__m128 _a, __m128 _b) +{ + float32x4_t a = vreinterpretq_f32_m128(_a); + float32x4_t b = vreinterpretq_f32_m128(_b); +#if defined(__aarch64__) + return vreinterpretq_m128_f32( + vsubq_f32(vuzp1q_f32(a, b), vuzp2q_f32(a, b))); +#else + float32x4x2_t c = vuzpq_f32(a, b); + return vreinterpretq_m128_f32(vsubq_f32(c.val[0], c.val[1])); +#endif +} + +// Load 128-bits of integer data from unaligned memory into dst. This intrinsic +// may perform better than _mm_loadu_si128 when the data crosses a cache line +// boundary. +// +// dst[127:0] := MEM[mem_addr+127:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_lddqu_si128 +#define _mm_lddqu_si128 _mm_loadu_si128 + +// Load a double-precision (64-bit) floating-point element from memory into both +// elements of dst. +// +// dst[63:0] := MEM[mem_addr+63:mem_addr] +// dst[127:64] := MEM[mem_addr+63:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loaddup_pd +#define _mm_loaddup_pd _mm_load1_pd + +// Duplicate the low double-precision (64-bit) floating-point element from a, +// and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movedup_pd +FORCE_INLINE __m128d _mm_movedup_pd(__m128d a) +{ +#if defined(__aarch64__) + return vreinterpretq_m128d_f64( + vdupq_laneq_f64(vreinterpretq_f64_m128d(a), 0)); +#else + return vreinterpretq_m128d_u64( + vdupq_n_u64(vgetq_lane_u64(vreinterpretq_u64_m128d(a), 0))); +#endif +} + +// Duplicate odd-indexed single-precision (32-bit) floating-point elements +// from a, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_movehdup_ps +FORCE_INLINE __m128 _mm_movehdup_ps(__m128 a) +{ +#if __has_builtin(__builtin_shufflevector) + return vreinterpretq_m128_f32(__builtin_shufflevector( + vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a), 1, 1, 3, 3)); +#else + float32_t a1 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 1); + float32_t a3 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 3); + float ALIGN_STRUCT(16) data[4] = {a1, a1, a3, a3}; + return vreinterpretq_m128_f32(vld1q_f32(data)); +#endif +} + +// Duplicate even-indexed single-precision (32-bit) floating-point elements +// from a, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_moveldup_ps +FORCE_INLINE __m128 _mm_moveldup_ps(__m128 a) +{ +#if __has_builtin(__builtin_shufflevector) + return vreinterpretq_m128_f32(__builtin_shufflevector( + vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a), 0, 0, 2, 2)); +#else + float32_t a0 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); + float32_t a2 = vgetq_lane_f32(vreinterpretq_f32_m128(a), 2); + float ALIGN_STRUCT(16) data[4] = {a0, a0, a2, a2}; + return vreinterpretq_m128_f32(vld1q_f32(data)); +#endif +} + +/* SSSE3 */ + +// Compute the absolute value of packed signed 16-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// dst[i+15:i] := ABS(a[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_epi16 +FORCE_INLINE __m128i _mm_abs_epi16(__m128i a) +{ + return vreinterpretq_m128i_s16(vabsq_s16(vreinterpretq_s16_m128i(a))); +} + +// Compute the absolute value of packed signed 32-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 3 +// i := j*32 +// dst[i+31:i] := ABS(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_epi32 +FORCE_INLINE __m128i _mm_abs_epi32(__m128i a) +{ + return vreinterpretq_m128i_s32(vabsq_s32(vreinterpretq_s32_m128i(a))); +} + +// Compute the absolute value of packed signed 8-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 15 +// i := j*8 +// dst[i+7:i] := ABS(a[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_epi8 +FORCE_INLINE __m128i _mm_abs_epi8(__m128i a) +{ + return vreinterpretq_m128i_s8(vabsq_s8(vreinterpretq_s8_m128i(a))); +} + +// Compute the absolute value of packed signed 16-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 3 +// i := j*16 +// dst[i+15:i] := ABS(a[i+15:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_pi16 +FORCE_INLINE __m64 _mm_abs_pi16(__m64 a) +{ + return vreinterpret_m64_s16(vabs_s16(vreinterpret_s16_m64(a))); +} + +// Compute the absolute value of packed signed 32-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 1 +// i := j*32 +// dst[i+31:i] := ABS(a[i+31:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_pi32 +FORCE_INLINE __m64 _mm_abs_pi32(__m64 a) +{ + return vreinterpret_m64_s32(vabs_s32(vreinterpret_s32_m64(a))); +} + +// Compute the absolute value of packed signed 8-bit integers in a, and store +// the unsigned results in dst. +// +// FOR j := 0 to 7 +// i := j*8 +// dst[i+7:i] := ABS(a[i+7:i]) +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_abs_pi8 +FORCE_INLINE __m64 _mm_abs_pi8(__m64 a) +{ + return vreinterpret_m64_s8(vabs_s8(vreinterpret_s8_m64(a))); +} + +// Concatenate 16-byte blocks in a and b into a 32-byte temporary result, shift +// the result right by imm8 bytes, and store the low 16 bytes in dst. +// +// tmp[255:0] := ((a[127:0] << 128)[255:0] OR b[127:0]) >> (imm8*8) +// dst[127:0] := tmp[127:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_alignr_epi8 +FORCE_INLINE __m128i _mm_alignr_epi8(__m128i a, __m128i b, int imm) +{ + if (_sse2neon_unlikely(imm & ~31)) + return _mm_setzero_si128(); + int idx; + uint8x16_t tmp[2]; + if (imm >= 16) { + idx = imm - 16; + tmp[0] = vreinterpretq_u8_m128i(a); + tmp[1] = vdupq_n_u8(0); + } else { + idx = imm; + tmp[0] = vreinterpretq_u8_m128i(b); + tmp[1] = vreinterpretq_u8_m128i(a); + } + return vreinterpretq_m128i_u8(vld1q_u8(((uint8_t const *) tmp) + idx)); +} + +// Concatenate 8-byte blocks in a and b into a 16-byte temporary result, shift +// the result right by imm8 bytes, and store the low 8 bytes in dst. +// +// tmp[127:0] := ((a[63:0] << 64)[127:0] OR b[63:0]) >> (imm8*8) +// dst[63:0] := tmp[63:0] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_alignr_pi8 +#define _mm_alignr_pi8(a, b, imm) \ + __extension__({ \ + __m64 ret; \ + if (_sse2neon_unlikely((imm) >= 16)) { \ + ret = vreinterpret_m64_s8(vdup_n_s8(0)); \ + } else { \ + uint8x8_t tmp_low, tmp_high; \ + if ((imm) >= 8) { \ + const int idx = (imm) -8; \ + tmp_low = vreinterpret_u8_m64(a); \ + tmp_high = vdup_n_u8(0); \ + ret = vreinterpret_m64_u8(vext_u8(tmp_low, tmp_high, idx)); \ + } else { \ + const int idx = (imm); \ + tmp_low = vreinterpret_u8_m64(b); \ + tmp_high = vreinterpret_u8_m64(a); \ + ret = vreinterpret_m64_u8(vext_u8(tmp_low, tmp_high, idx)); \ + } \ + } \ + ret; \ + }) + +// Computes pairwise add of each argument as a 16-bit signed or unsigned integer +// values a and b. +FORCE_INLINE __m128i _mm_hadd_epi16(__m128i _a, __m128i _b) +{ + int16x8_t a = vreinterpretq_s16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); +#if defined(__aarch64__) + return vreinterpretq_m128i_s16(vpaddq_s16(a, b)); +#else + return vreinterpretq_m128i_s16( + vcombine_s16(vpadd_s16(vget_low_s16(a), vget_high_s16(a)), + vpadd_s16(vget_low_s16(b), vget_high_s16(b)))); +#endif +} + +// Computes pairwise add of each argument as a 32-bit signed or unsigned integer +// values a and b. +FORCE_INLINE __m128i _mm_hadd_epi32(__m128i _a, __m128i _b) +{ + int32x4_t a = vreinterpretq_s32_m128i(_a); + int32x4_t b = vreinterpretq_s32_m128i(_b); + return vreinterpretq_m128i_s32( + vcombine_s32(vpadd_s32(vget_low_s32(a), vget_high_s32(a)), + vpadd_s32(vget_low_s32(b), vget_high_s32(b)))); +} + +// Horizontally add adjacent pairs of 16-bit integers in a and b, and pack the +// signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadd_pi16 +FORCE_INLINE __m64 _mm_hadd_pi16(__m64 a, __m64 b) +{ + return vreinterpret_m64_s16( + vpadd_s16(vreinterpret_s16_m64(a), vreinterpret_s16_m64(b))); +} + +// Horizontally add adjacent pairs of 32-bit integers in a and b, and pack the +// signed 32-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadd_pi32 +FORCE_INLINE __m64 _mm_hadd_pi32(__m64 a, __m64 b) +{ + return vreinterpret_m64_s32( + vpadd_s32(vreinterpret_s32_m64(a), vreinterpret_s32_m64(b))); +} + +// Computes saturated pairwise sub of each argument as a 16-bit signed +// integer values a and b. +FORCE_INLINE __m128i _mm_hadds_epi16(__m128i _a, __m128i _b) +{ +#if defined(__aarch64__) + int16x8_t a = vreinterpretq_s16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); + return vreinterpretq_s64_s16( + vqaddq_s16(vuzp1q_s16(a, b), vuzp2q_s16(a, b))); +#else + int32x4_t a = vreinterpretq_s32_m128i(_a); + int32x4_t b = vreinterpretq_s32_m128i(_b); + // Interleave using vshrn/vmovn + // [a0|a2|a4|a6|b0|b2|b4|b6] + // [a1|a3|a5|a7|b1|b3|b5|b7] + int16x8_t ab0246 = vcombine_s16(vmovn_s32(a), vmovn_s32(b)); + int16x8_t ab1357 = vcombine_s16(vshrn_n_s32(a, 16), vshrn_n_s32(b, 16)); + // Saturated add + return vreinterpretq_m128i_s16(vqaddq_s16(ab0246, ab1357)); +#endif +} + +// Horizontally add adjacent pairs of signed 16-bit integers in a and b using +// saturation, and pack the signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadds_pi16 +FORCE_INLINE __m64 _mm_hadds_pi16(__m64 _a, __m64 _b) +{ + int16x4_t a = vreinterpret_s16_m64(_a); + int16x4_t b = vreinterpret_s16_m64(_b); +#if defined(__aarch64__) + return vreinterpret_s64_s16(vqadd_s16(vuzp1_s16(a, b), vuzp2_s16(a, b))); +#else + int16x4x2_t res = vuzp_s16(a, b); + return vreinterpret_s64_s16(vqadd_s16(res.val[0], res.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of 16-bit integers in a and b, and pack +// the signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_epi16 +FORCE_INLINE __m128i _mm_hsub_epi16(__m128i _a, __m128i _b) +{ + int16x8_t a = vreinterpretq_s16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); +#if defined(__aarch64__) + return vreinterpretq_m128i_s16( + vsubq_s16(vuzp1q_s16(a, b), vuzp2q_s16(a, b))); +#else + int16x8x2_t c = vuzpq_s16(a, b); + return vreinterpretq_m128i_s16(vsubq_s16(c.val[0], c.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of 32-bit integers in a and b, and pack +// the signed 32-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_epi32 +FORCE_INLINE __m128i _mm_hsub_epi32(__m128i _a, __m128i _b) +{ + int32x4_t a = vreinterpretq_s32_m128i(_a); + int32x4_t b = vreinterpretq_s32_m128i(_b); +#if defined(__aarch64__) + return vreinterpretq_m128i_s32( + vsubq_s32(vuzp1q_s32(a, b), vuzp2q_s32(a, b))); +#else + int32x4x2_t c = vuzpq_s32(a, b); + return vreinterpretq_m128i_s32(vsubq_s32(c.val[0], c.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of 16-bit integers in a and b, and pack +// the signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_pi16 +FORCE_INLINE __m64 _mm_hsub_pi16(__m64 _a, __m64 _b) +{ + int16x4_t a = vreinterpret_s16_m64(_a); + int16x4_t b = vreinterpret_s16_m64(_b); +#if defined(__aarch64__) + return vreinterpret_m64_s16(vsub_s16(vuzp1_s16(a, b), vuzp2_s16(a, b))); +#else + int16x4x2_t c = vuzp_s16(a, b); + return vreinterpret_m64_s16(vsub_s16(c.val[0], c.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of 32-bit integers in a and b, and pack +// the signed 32-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_hsub_pi32 +FORCE_INLINE __m64 _mm_hsub_pi32(__m64 _a, __m64 _b) +{ + int32x2_t a = vreinterpret_s32_m64(_a); + int32x2_t b = vreinterpret_s32_m64(_b); +#if defined(__aarch64__) + return vreinterpret_m64_s32(vsub_s32(vuzp1_s32(a, b), vuzp2_s32(a, b))); +#else + int32x2x2_t c = vuzp_s32(a, b); + return vreinterpret_m64_s32(vsub_s32(c.val[0], c.val[1])); +#endif +} + +// Computes saturated pairwise difference of each argument as a 16-bit signed +// integer values a and b. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsubs_epi16 +FORCE_INLINE __m128i _mm_hsubs_epi16(__m128i _a, __m128i _b) +{ + int16x8_t a = vreinterpretq_s16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); +#if defined(__aarch64__) + return vreinterpretq_m128i_s16( + vqsubq_s16(vuzp1q_s16(a, b), vuzp2q_s16(a, b))); +#else + int16x8x2_t c = vuzpq_s16(a, b); + return vreinterpretq_m128i_s16(vqsubq_s16(c.val[0], c.val[1])); +#endif +} + +// Horizontally subtract adjacent pairs of signed 16-bit integers in a and b +// using saturation, and pack the signed 16-bit results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsubs_pi16 +FORCE_INLINE __m64 _mm_hsubs_pi16(__m64 _a, __m64 _b) +{ + int16x4_t a = vreinterpret_s16_m64(_a); + int16x4_t b = vreinterpret_s16_m64(_b); +#if defined(__aarch64__) + return vreinterpret_m64_s16(vqsub_s16(vuzp1_s16(a, b), vuzp2_s16(a, b))); +#else + int16x4x2_t c = vuzp_s16(a, b); + return vreinterpret_m64_s16(vqsub_s16(c.val[0], c.val[1])); +#endif +} + +// Vertically multiply each unsigned 8-bit integer from a with the corresponding +// signed 8-bit integer from b, producing intermediate signed 16-bit integers. +// Horizontally add adjacent pairs of intermediate signed 16-bit integers, +// and pack the saturated results in dst. +// +// FOR j := 0 to 7 +// i := j*16 +// dst[i+15:i] := Saturate_To_Int16( a[i+15:i+8]*b[i+15:i+8] + +// a[i+7:i]*b[i+7:i] ) +// ENDFOR +FORCE_INLINE __m128i _mm_maddubs_epi16(__m128i _a, __m128i _b) +{ +#if defined(__aarch64__) + uint8x16_t a = vreinterpretq_u8_m128i(_a); + int8x16_t b = vreinterpretq_s8_m128i(_b); + int16x8_t tl = vmulq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(a))), + vmovl_s8(vget_low_s8(b))); + int16x8_t th = vmulq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(a))), + vmovl_s8(vget_high_s8(b))); + return vreinterpretq_m128i_s16( + vqaddq_s16(vuzp1q_s16(tl, th), vuzp2q_s16(tl, th))); +#else + // This would be much simpler if x86 would choose to zero extend OR sign + // extend, not both. This could probably be optimized better. + uint16x8_t a = vreinterpretq_u16_m128i(_a); + int16x8_t b = vreinterpretq_s16_m128i(_b); + + // Zero extend a + int16x8_t a_odd = vreinterpretq_s16_u16(vshrq_n_u16(a, 8)); + int16x8_t a_even = vreinterpretq_s16_u16(vbicq_u16(a, vdupq_n_u16(0xff00))); + + // Sign extend by shifting left then shifting right. + int16x8_t b_even = vshrq_n_s16(vshlq_n_s16(b, 8), 8); + int16x8_t b_odd = vshrq_n_s16(b, 8); + + // multiply + int16x8_t prod1 = vmulq_s16(a_even, b_even); + int16x8_t prod2 = vmulq_s16(a_odd, b_odd); + + // saturated add + return vreinterpretq_m128i_s16(vqaddq_s16(prod1, prod2)); +#endif +} + +// Vertically multiply each unsigned 8-bit integer from a with the corresponding +// signed 8-bit integer from b, producing intermediate signed 16-bit integers. +// Horizontally add adjacent pairs of intermediate signed 16-bit integers, and +// pack the saturated results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_maddubs_pi16 +FORCE_INLINE __m64 _mm_maddubs_pi16(__m64 _a, __m64 _b) +{ + uint16x4_t a = vreinterpret_u16_m64(_a); + int16x4_t b = vreinterpret_s16_m64(_b); + + // Zero extend a + int16x4_t a_odd = vreinterpret_s16_u16(vshr_n_u16(a, 8)); + int16x4_t a_even = vreinterpret_s16_u16(vand_u16(a, vdup_n_u16(0xff))); + + // Sign extend by shifting left then shifting right. + int16x4_t b_even = vshr_n_s16(vshl_n_s16(b, 8), 8); + int16x4_t b_odd = vshr_n_s16(b, 8); + + // multiply + int16x4_t prod1 = vmul_s16(a_even, b_even); + int16x4_t prod2 = vmul_s16(a_odd, b_odd); + + // saturated add + return vreinterpret_m64_s16(vqadd_s16(prod1, prod2)); +} + +// Multiply packed signed 16-bit integers in a and b, producing intermediate +// signed 32-bit integers. Shift right by 15 bits while rounding up, and store +// the packed 16-bit integers in dst. +// +// r0 := Round(((int32_t)a0 * (int32_t)b0) >> 15) +// r1 := Round(((int32_t)a1 * (int32_t)b1) >> 15) +// r2 := Round(((int32_t)a2 * (int32_t)b2) >> 15) +// ... +// r7 := Round(((int32_t)a7 * (int32_t)b7) >> 15) +FORCE_INLINE __m128i _mm_mulhrs_epi16(__m128i a, __m128i b) +{ + // Has issues due to saturation + // return vreinterpretq_m128i_s16(vqrdmulhq_s16(a, b)); + + // Multiply + int32x4_t mul_lo = vmull_s16(vget_low_s16(vreinterpretq_s16_m128i(a)), + vget_low_s16(vreinterpretq_s16_m128i(b))); + int32x4_t mul_hi = vmull_s16(vget_high_s16(vreinterpretq_s16_m128i(a)), + vget_high_s16(vreinterpretq_s16_m128i(b))); + + // Rounding narrowing shift right + // narrow = (int16_t)((mul + 16384) >> 15); + int16x4_t narrow_lo = vrshrn_n_s32(mul_lo, 15); + int16x4_t narrow_hi = vrshrn_n_s32(mul_hi, 15); + + // Join together + return vreinterpretq_m128i_s16(vcombine_s16(narrow_lo, narrow_hi)); +} + +// Multiply packed signed 16-bit integers in a and b, producing intermediate +// signed 32-bit integers. Truncate each intermediate integer to the 18 most +// significant bits, round by adding 1, and store bits [16:1] to dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mulhrs_pi16 +FORCE_INLINE __m64 _mm_mulhrs_pi16(__m64 a, __m64 b) +{ + int32x4_t mul_extend = + vmull_s16((vreinterpret_s16_m64(a)), (vreinterpret_s16_m64(b))); + + // Rounding narrowing shift right + return vreinterpret_m64_s16(vrshrn_n_s32(mul_extend, 15)); +} + +// Shuffle packed 8-bit integers in a according to shuffle control mask in the +// corresponding 8-bit element of b, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_epi8 +FORCE_INLINE __m128i _mm_shuffle_epi8(__m128i a, __m128i b) +{ + int8x16_t tbl = vreinterpretq_s8_m128i(a); // input a + uint8x16_t idx = vreinterpretq_u8_m128i(b); // input b + uint8x16_t idx_masked = + vandq_u8(idx, vdupq_n_u8(0x8F)); // avoid using meaningless bits +#if defined(__aarch64__) + return vreinterpretq_m128i_s8(vqtbl1q_s8(tbl, idx_masked)); +#elif defined(__GNUC__) + int8x16_t ret; + // %e and %f represent the even and odd D registers + // respectively. + __asm__ __volatile__( + "vtbl.8 %e[ret], {%e[tbl], %f[tbl]}, %e[idx]\n" + "vtbl.8 %f[ret], {%e[tbl], %f[tbl]}, %f[idx]\n" + : [ret] "=&w"(ret) + : [tbl] "w"(tbl), [idx] "w"(idx_masked)); + return vreinterpretq_m128i_s8(ret); +#else + // use this line if testing on aarch64 + int8x8x2_t a_split = {vget_low_s8(tbl), vget_high_s8(tbl)}; + return vreinterpretq_m128i_s8( + vcombine_s8(vtbl2_s8(a_split, vget_low_u8(idx_masked)), + vtbl2_s8(a_split, vget_high_u8(idx_masked)))); +#endif +} + +// Shuffle packed 8-bit integers in a according to shuffle control mask in the +// corresponding 8-bit element of b, and store the results in dst. +// +// FOR j := 0 to 7 +// i := j*8 +// IF b[i+7] == 1 +// dst[i+7:i] := 0 +// ELSE +// index[2:0] := b[i+2:i] +// dst[i+7:i] := a[index*8+7:index*8] +// FI +// ENDFOR +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_shuffle_pi8 +FORCE_INLINE __m64 _mm_shuffle_pi8(__m64 a, __m64 b) +{ + const int8x8_t controlMask = + vand_s8(vreinterpret_s8_m64(b), vdup_n_s8((int8_t) (0x1 << 7 | 0x07))); + int8x8_t res = vtbl1_s8(vreinterpret_s8_m64(a), controlMask); + return vreinterpret_m64_s8(res); } // Negate packed 16-bit integers in a when the corresponding signed @@ -3212,6 +7166,45 @@ FORCE_INLINE __m128i _mm_sign_epi32(__m128i _a, __m128i _b) return vreinterpretq_m128i_s32(res); } +// Negate packed 8-bit integers in a when the corresponding signed +// 8-bit integer in b is negative, and store the results in dst. +// Element in dst are zeroed out when the corresponding element +// in b is zero. +// +// for i in 0..15 +// if b[i] < 0 +// r[i] := -a[i] +// else if b[i] == 0 +// r[i] := 0 +// else +// r[i] := a[i] +// fi +// done +FORCE_INLINE __m128i _mm_sign_epi8(__m128i _a, __m128i _b) +{ + int8x16_t a = vreinterpretq_s8_m128i(_a); + int8x16_t b = vreinterpretq_s8_m128i(_b); + + // signed shift right: faster than vclt + // (b < 0) ? 0xFF : 0 + uint8x16_t ltMask = vreinterpretq_u8_s8(vshrq_n_s8(b, 7)); + + // (b == 0) ? 0xFF : 0 +#if defined(__aarch64__) + int8x16_t zeroMask = vreinterpretq_s8_u8(vceqzq_s8(b)); +#else + int8x16_t zeroMask = vreinterpretq_s8_u8(vceqq_s8(b, vdupq_n_s8(0))); +#endif + + // bitwise select either a or negative 'a' (vnegq_s8(a) return negative 'a') + // based on ltMask + int8x16_t masked = vbslq_s8(ltMask, vnegq_s8(a), a); + // res = masked & (~zeroMask) + int8x16_t res = vbicq_s8(masked, zeroMask); + + return vreinterpretq_m128i_s8(res); +} + // Negate packed 16-bit integers in a when the corresponding signed 16-bit // integer in b is negative, and store the results in dst. Element in dst are // zeroed out when the corresponding element in b is zero. @@ -3244,7 +7237,7 @@ FORCE_INLINE __m64 _mm_sign_pi16(__m64 _a, __m64 _b) int16x4_t zeroMask = vreinterpret_s16_u16(vceq_s16(b, vdup_n_s16(0))); #endif - // bitwise select either a or nagative 'a' (vneg_s16(a) return nagative 'a') + // bitwise select either a or negative 'a' (vneg_s16(a) return negative 'a') // based on ltMask int16x4_t masked = vbsl_s16(ltMask, vneg_s16(a), a); // res = masked & (~zeroMask) @@ -3285,7 +7278,7 @@ FORCE_INLINE __m64 _mm_sign_pi32(__m64 _a, __m64 _b) int32x2_t zeroMask = vreinterpret_s32_u32(vceq_s32(b, vdup_n_s32(0))); #endif - // bitwise select either a or nagative 'a' (vneg_s32(a) return nagative 'a') + // bitwise select either a or negative 'a' (vneg_s32(a) return negative 'a') // based on ltMask int32x2_t masked = vbsl_s32(ltMask, vneg_s32(a), a); // res = masked & (~zeroMask) @@ -3326,7 +7319,7 @@ FORCE_INLINE __m64 _mm_sign_pi8(__m64 _a, __m64 _b) int8x8_t zeroMask = vreinterpret_s8_u8(vceq_s8(b, vdup_n_s8(0))); #endif - // bitwise select either a or nagative 'a' (vneg_s8(a) return nagative 'a') + // bitwise select either a or negative 'a' (vneg_s8(a) return negative 'a') // based on ltMask int8x8_t masked = vbsl_s8(ltMask, vneg_s8(a), a); // res = masked & (~zeroMask) @@ -3335,1423 +7328,346 @@ FORCE_INLINE __m64 _mm_sign_pi8(__m64 _a, __m64 _b) return vreinterpret_m64_s8(res); } -// Average packed unsigned 16-bit integers in a and b, and store the results in -// dst. -// -// FOR j := 0 to 3 -// i := j*16 -// dst[i+15:i] := (a[i+15:i] + b[i+15:i] + 1) >> 1 -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_avg_pu16 -FORCE_INLINE __m64 _mm_avg_pu16(__m64 a, __m64 b) -{ - return vreinterpret_m64_u16( - vrhadd_u16(vreinterpret_u16_m64(a), vreinterpret_u16_m64(b))); -} +/* SSE4.1 */ -// Average packed unsigned 8-bit integers in a and b, and store the results in -// dst. +// Blend packed 16-bit integers from a and b using control mask imm8, and store +// the results in dst. // // FOR j := 0 to 7 -// i := j*8 -// dst[i+7:i] := (a[i+7:i] + b[i+7:i] + 1) >> 1 +// i := j*16 +// IF imm8[j] +// dst[i+15:i] := b[i+15:i] +// ELSE +// dst[i+15:i] := a[i+15:i] +// FI // ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_avg_pu8 -FORCE_INLINE __m64 _mm_avg_pu8(__m64 a, __m64 b) +// FORCE_INLINE __m128i _mm_blend_epi16(__m128i a, __m128i b, +// __constrange(0,255) int imm) +#define _mm_blend_epi16(a, b, imm) \ + __extension__({ \ + const uint16_t _mask[8] = {((imm) & (1 << 0)) ? (uint16_t) -1 : 0x0, \ + ((imm) & (1 << 1)) ? (uint16_t) -1 : 0x0, \ + ((imm) & (1 << 2)) ? (uint16_t) -1 : 0x0, \ + ((imm) & (1 << 3)) ? (uint16_t) -1 : 0x0, \ + ((imm) & (1 << 4)) ? (uint16_t) -1 : 0x0, \ + ((imm) & (1 << 5)) ? (uint16_t) -1 : 0x0, \ + ((imm) & (1 << 6)) ? (uint16_t) -1 : 0x0, \ + ((imm) & (1 << 7)) ? (uint16_t) -1 : 0x0}; \ + uint16x8_t _mask_vec = vld1q_u16(_mask); \ + uint16x8_t _a = vreinterpretq_u16_m128i(a); \ + uint16x8_t _b = vreinterpretq_u16_m128i(b); \ + vreinterpretq_m128i_u16(vbslq_u16(_mask_vec, _b, _a)); \ + }) + +// Blend packed double-precision (64-bit) floating-point elements from a and b +// using control mask imm8, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blend_pd +#define _mm_blend_pd(a, b, imm) \ + __extension__({ \ + const uint64_t _mask[2] = { \ + ((imm) & (1 << 0)) ? ~UINT64_C(0) : UINT64_C(0), \ + ((imm) & (1 << 1)) ? ~UINT64_C(0) : UINT64_C(0)}; \ + uint64x2_t _mask_vec = vld1q_u64(_mask); \ + uint64x2_t _a = vreinterpretq_u64_m128d(a); \ + uint64x2_t _b = vreinterpretq_u64_m128d(b); \ + vreinterpretq_m128d_u64(vbslq_u64(_mask_vec, _b, _a)); \ + }) + +// Blend packed single-precision (32-bit) floating-point elements from a and b +// using mask, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blend_ps +FORCE_INLINE __m128 _mm_blend_ps(__m128 _a, __m128 _b, const char imm8) { - return vreinterpret_m64_u8( - vrhadd_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); + const uint32_t ALIGN_STRUCT(16) + data[4] = {((imm8) & (1 << 0)) ? UINT32_MAX : 0, + ((imm8) & (1 << 1)) ? UINT32_MAX : 0, + ((imm8) & (1 << 2)) ? UINT32_MAX : 0, + ((imm8) & (1 << 3)) ? UINT32_MAX : 0}; + uint32x4_t mask = vld1q_u32(data); + float32x4_t a = vreinterpretq_f32_m128(_a); + float32x4_t b = vreinterpretq_f32_m128(_b); + return vreinterpretq_m128_f32(vbslq_f32(mask, b, a)); } -// Average packed unsigned 8-bit integers in a and b, and store the results in +// Blend packed 8-bit integers from a and b using mask, and store the results in // dst. // -// FOR j := 0 to 7 -// i := j*8 -// dst[i+7:i] := (a[i+7:i] + b[i+7:i] + 1) >> 1 -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pavgb -#define _m_pavgb(a, b) _mm_avg_pu8(a, b) - -// Average packed unsigned 16-bit integers in a and b, and store the results in -// dst. -// -// FOR j := 0 to 3 -// i := j*16 -// dst[i+15:i] := (a[i+15:i] + b[i+15:i] + 1) >> 1 -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pavgw -#define _m_pavgw(a, b) _mm_avg_pu16(a, b) - -// Extract a 16-bit integer from a, selected with imm8, and store the result in -// the lower element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pextrw -#define _m_pextrw(a, imm) _mm_extract_pi16(a, imm) - -// Copy a to dst, and insert the 16-bit integer i into dst at the location -// specified by imm8. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=m_pinsrw -#define _m_pinsrw(a, i, imm) _mm_insert_pi16(a, i, imm) - -// Compare packed signed 16-bit integers in a and b, and store packed maximum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmaxsw -#define _m_pmaxsw(a, b) _mm_max_pi16(a, b) - -// Compare packed unsigned 8-bit integers in a and b, and store packed maximum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmaxub -#define _m_pmaxub(a, b) _mm_max_pu8(a, b) - -// Compare packed signed 16-bit integers in a and b, and store packed minimum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pminsw -#define _m_pminsw(a, b) _mm_min_pi16(a, b) - -// Compare packed unsigned 8-bit integers in a and b, and store packed minimum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pminub -#define _m_pminub(a, b) _mm_min_pu8(a, b) - -// Create mask from the most significant bit of each 8-bit element in a, and -// store the result in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmovmskb -#define _m_pmovmskb(a) _mm_movemask_pi8(a) - -// Multiply the packed unsigned 16-bit integers in a and b, producing -// intermediate 32-bit integers, and store the high 16 bits of the intermediate -// integers in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmulhuw -#define _m_pmulhuw(a, b) _mm_mulhi_pu16(a, b) - -// Compute the absolute differences of packed unsigned 8-bit integers in a and -// b, then horizontally sum each consecutive 8 differences to produce four -// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low -// 16 bits of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=m_psadbw -#define _m_psadbw(a, b) _mm_sad_pu8(a, b) - -// Computes the average of the 16 unsigned 8-bit integers in a and the 16 -// unsigned 8-bit integers in b and rounds. -// -// r0 := (a0 + b0) / 2 -// r1 := (a1 + b1) / 2 -// ... -// r15 := (a15 + b15) / 2 -// -// https://msdn.microsoft.com/en-us/library/vstudio/8zwh554a(v%3dvs.90).aspx -FORCE_INLINE __m128i _mm_avg_epu8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u8( - vrhaddq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); -} - -// Shift a left by imm8 bytes while shifting in zeros, and store the results in -// dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_bslli_si128 -#define _mm_bslli_si128(a, imm) _mm_slli_si128(a, imm) - -// Shift a right by imm8 bytes while shifting in zeros, and store the results in -// dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_bsrli_si128 -#define _mm_bsrli_si128(a, imm) _mm_srli_si128(a, imm) - -// Computes the average of the 8 unsigned 16-bit integers in a and the 8 -// unsigned 16-bit integers in b and rounds. -// -// r0 := (a0 + b0) / 2 -// r1 := (a1 + b1) / 2 -// ... -// r7 := (a7 + b7) / 2 -// -// https://msdn.microsoft.com/en-us/library/vstudio/y13ca3c8(v=vs.90).aspx -FORCE_INLINE __m128i _mm_avg_epu16(__m128i a, __m128i b) -{ - return (__m128i) vrhaddq_u16(vreinterpretq_u16_m128i(a), - vreinterpretq_u16_m128i(b)); -} - -// Adds the four single-precision, floating-point values of a and b. -// -// r0 := a0 + b0 -// r1 := a1 + b1 -// r2 := a2 + b2 -// r3 := a3 + b3 -// -// https://msdn.microsoft.com/en-us/library/vstudio/c9848chc(v=vs.100).aspx -FORCE_INLINE __m128 _mm_add_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_f32( - vaddq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -} - -// Add packed double-precision (64-bit) floating-point elements in a and b, and -// store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_add_pd -FORCE_INLINE __m128d _mm_add_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vaddq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - double *da = (double *) &a; - double *db = (double *) &b; - double c[2]; - c[0] = da[0] + db[0]; - c[1] = da[1] + db[1]; - return vld1q_f32((float32_t *) c); -#endif -} - -// Add the lower double-precision (64-bit) floating-point element in a and b, -// store the result in the lower element of dst, and copy the upper element from -// a to the upper element of dst. -// -// dst[63:0] := a[63:0] + b[63:0] -// dst[127:64] := a[127:64] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_add_sd -FORCE_INLINE __m128d _mm_add_sd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return _mm_move_sd(a, _mm_add_pd(a, b)); -#else - double *da = (double *) &a; - double *db = (double *) &b; - double c[2]; - c[0] = da[0] + db[0]; - c[1] = da[1]; - return vld1q_f32((float32_t *) c); -#endif -} - -// Add 64-bit integers a and b, and store the result in dst. -// -// dst[63:0] := a[63:0] + b[63:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_add_si64 -FORCE_INLINE __m64 _mm_add_si64(__m64 a, __m64 b) -{ - return vreinterpret_m64_s64( - vadd_s64(vreinterpret_s64_m64(a), vreinterpret_s64_m64(b))); -} - -// adds the scalar single-precision floating point values of a and b. -// https://msdn.microsoft.com/en-us/library/be94x2y6(v=vs.100).aspx -FORCE_INLINE __m128 _mm_add_ss(__m128 a, __m128 b) -{ - float32_t b0 = vgetq_lane_f32(vreinterpretq_f32_m128(b), 0); - float32x4_t value = vsetq_lane_f32(b0, vdupq_n_f32(0), 0); - // the upper values in the result must be the remnants of . - return vreinterpretq_m128_f32(vaddq_f32(a, value)); -} - -// Adds the 4 signed or unsigned 64-bit integers in a to the 4 signed or -// unsigned 32-bit integers in b. -// https://msdn.microsoft.com/en-us/library/vstudio/09xs4fkk(v=vs.100).aspx -FORCE_INLINE __m128i _mm_add_epi64(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s64( - vaddq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b))); -} - -// Adds the 4 signed or unsigned 32-bit integers in a to the 4 signed or -// unsigned 32-bit integers in b. -// -// r0 := a0 + b0 -// r1 := a1 + b1 -// r2 := a2 + b2 -// r3 := a3 + b3 -// -// https://msdn.microsoft.com/en-us/library/vstudio/09xs4fkk(v=vs.100).aspx -FORCE_INLINE __m128i _mm_add_epi32(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s32( - vaddq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -} - -// Adds the 8 signed or unsigned 16-bit integers in a to the 8 signed or -// unsigned 16-bit integers in b. -// https://msdn.microsoft.com/en-us/library/fceha5k4(v=vs.100).aspx -FORCE_INLINE __m128i _mm_add_epi16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s16( - vaddq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -} - -// Adds the 16 signed or unsigned 8-bit integers in a to the 16 signed or -// unsigned 8-bit integers in b. -// https://technet.microsoft.com/en-us/subscriptions/yc7tcyzs(v=vs.90) -FORCE_INLINE __m128i _mm_add_epi8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s8( - vaddq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -} - -// Adds the 8 signed 16-bit integers in a to the 8 signed 16-bit integers in b -// and saturates. -// -// r0 := SignedSaturate(a0 + b0) -// r1 := SignedSaturate(a1 + b1) -// ... -// r7 := SignedSaturate(a7 + b7) -// -// https://msdn.microsoft.com/en-us/library/1a306ef8(v=vs.100).aspx -FORCE_INLINE __m128i _mm_adds_epi16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s16( - vqaddq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -} - -// Add packed signed 8-bit integers in a and b using saturation, and store the -// results in dst. -// // FOR j := 0 to 15 -// i := j*8 -// dst[i+7:i] := Saturate8( a[i+7:i] + b[i+7:i] ) +// i := j*8 +// IF mask[i+7] +// dst[i+7:i] := b[i+7:i] +// ELSE +// dst[i+7:i] := a[i+7:i] +// FI // ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_adds_epi8 -FORCE_INLINE __m128i _mm_adds_epi8(__m128i a, __m128i b) +FORCE_INLINE __m128i _mm_blendv_epi8(__m128i _a, __m128i _b, __m128i _mask) { - return vreinterpretq_m128i_s8( - vqaddq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -} - -// Adds the 16 unsigned 8-bit integers in a to the 16 unsigned 8-bit integers in -// b and saturates.. -// https://msdn.microsoft.com/en-us/library/9hahyddy(v=vs.100).aspx -FORCE_INLINE __m128i _mm_adds_epu8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u8( - vqaddq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); -} - -// Multiplies the 8 signed or unsigned 16-bit integers from a by the 8 signed or -// unsigned 16-bit integers from b. -// -// r0 := (a0 * b0)[15:0] -// r1 := (a1 * b1)[15:0] -// ... -// r7 := (a7 * b7)[15:0] -// -// https://msdn.microsoft.com/en-us/library/vstudio/9ks1472s(v=vs.100).aspx -FORCE_INLINE __m128i _mm_mullo_epi16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s16( - vmulq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -} - -// Multiplies the 4 signed or unsigned 32-bit integers from a by the 4 signed or -// unsigned 32-bit integers from b. -// https://msdn.microsoft.com/en-us/library/vstudio/bb531409(v=vs.100).aspx -FORCE_INLINE __m128i _mm_mullo_epi32(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s32( - vmulq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -} - -// Multiply the packed unsigned 16-bit integers in a and b, producing -// intermediate 32-bit integers, and store the high 16 bits of the intermediate -// integers in dst. -// -// FOR j := 0 to 3 -// i := j*16 -// tmp[31:0] := a[i+15:i] * b[i+15:i] -// dst[i+15:i] := tmp[31:16] -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_pmulhuw -#define _m_pmulhuw(a, b) _mm_mulhi_pu16(a, b) - -// Multiplies the four single-precision, floating-point values of a and b. -// -// r0 := a0 * b0 -// r1 := a1 * b1 -// r2 := a2 * b2 -// r3 := a3 * b3 -// -// https://msdn.microsoft.com/en-us/library/vstudio/22kbk6t9(v=vs.100).aspx -FORCE_INLINE __m128 _mm_mul_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_f32( - vmulq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -} - -// Multiply packed double-precision (64-bit) floating-point elements in a and b, -// and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mul_pd -FORCE_INLINE __m128d _mm_mul_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vmulq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - double *da = (double *) &a; - double *db = (double *) &b; - double c[2]; - c[0] = da[0] * db[0]; - c[1] = da[1] * db[1]; - return vld1q_f32((float32_t *) c); -#endif -} - -// Multiply the lower double-precision (64-bit) floating-point element in a and -// b, store the result in the lower element of dst, and copy the upper element -// from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_mul_sd -FORCE_INLINE __m128d _mm_mul_sd(__m128d a, __m128d b) -{ - return _mm_move_sd(a, _mm_mul_pd(a, b)); -} - -// Multiply the lower single-precision (32-bit) floating-point element in a and -// b, store the result in the lower element of dst, and copy the upper 3 packed -// elements from a to the upper elements of dst. -// -// dst[31:0] := a[31:0] * b[31:0] -// dst[127:32] := a[127:32] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mul_ss -FORCE_INLINE __m128 _mm_mul_ss(__m128 a, __m128 b) -{ - return _mm_move_ss(a, _mm_mul_ps(a, b)); -} - -// Multiply the low unsigned 32-bit integers from each packed 64-bit element in -// a and b, and store the unsigned 64-bit results in dst. -// -// r0 := (a0 & 0xFFFFFFFF) * (b0 & 0xFFFFFFFF) -// r1 := (a2 & 0xFFFFFFFF) * (b2 & 0xFFFFFFFF) -FORCE_INLINE __m128i _mm_mul_epu32(__m128i a, __m128i b) -{ - // vmull_u32 upcasts instead of masking, so we downcast. - uint32x2_t a_lo = vmovn_u64(vreinterpretq_u64_m128i(a)); - uint32x2_t b_lo = vmovn_u64(vreinterpretq_u64_m128i(b)); - return vreinterpretq_m128i_u64(vmull_u32(a_lo, b_lo)); -} - -// Multiply the low unsigned 32-bit integers from a and b, and store the -// unsigned 64-bit result in dst. -// -// dst[63:0] := a[31:0] * b[31:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mul_su32 -FORCE_INLINE __m64 _mm_mul_su32(__m64 a, __m64 b) -{ - return vreinterpret_m64_u64(vget_low_u64( - vmull_u32(vreinterpret_u32_m64(a), vreinterpret_u32_m64(b)))); -} - -// Multiply the low signed 32-bit integers from each packed 64-bit element in -// a and b, and store the signed 64-bit results in dst. -// -// r0 := (int64_t)(int32_t)a0 * (int64_t)(int32_t)b0 -// r1 := (int64_t)(int32_t)a2 * (int64_t)(int32_t)b2 -FORCE_INLINE __m128i _mm_mul_epi32(__m128i a, __m128i b) -{ - // vmull_s32 upcasts instead of masking, so we downcast. - int32x2_t a_lo = vmovn_s64(vreinterpretq_s64_m128i(a)); - int32x2_t b_lo = vmovn_s64(vreinterpretq_s64_m128i(b)); - return vreinterpretq_m128i_s64(vmull_s32(a_lo, b_lo)); -} - -// Multiplies the 8 signed 16-bit integers from a by the 8 signed 16-bit -// integers from b. -// -// r0 := (a0 * b0) + (a1 * b1) -// r1 := (a2 * b2) + (a3 * b3) -// r2 := (a4 * b4) + (a5 * b5) -// r3 := (a6 * b6) + (a7 * b7) -// https://msdn.microsoft.com/en-us/library/yht36sa6(v=vs.90).aspx -FORCE_INLINE __m128i _mm_madd_epi16(__m128i a, __m128i b) -{ - int32x4_t low = vmull_s16(vget_low_s16(vreinterpretq_s16_m128i(a)), - vget_low_s16(vreinterpretq_s16_m128i(b))); - int32x4_t high = vmull_s16(vget_high_s16(vreinterpretq_s16_m128i(a)), - vget_high_s16(vreinterpretq_s16_m128i(b))); - - int32x2_t low_sum = vpadd_s32(vget_low_s32(low), vget_high_s32(low)); - int32x2_t high_sum = vpadd_s32(vget_low_s32(high), vget_high_s32(high)); - - return vreinterpretq_m128i_s32(vcombine_s32(low_sum, high_sum)); -} - -// Conditionally store 8-bit integer elements from a into memory using mask -// (elements are not stored when the highest bit is not set in the corresponding -// element) and a non-temporal memory hint. mem_addr does not need to be aligned -// on any particular boundary. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_maskmoveu_si128 -FORCE_INLINE void _mm_maskmoveu_si128(__m128i a, __m128i mask, char *mem_addr) -{ - int8x16_t shr_mask = vshrq_n_s8(vreinterpretq_s8_m128i(mask), 7); - __m128 b = _mm_load_ps((const float *) mem_addr); - int8x16_t masked = - vbslq_s8(vreinterpretq_u8_s8(shr_mask), vreinterpretq_s8_m128i(a), - vreinterpretq_s8_m128(b)); - vst1q_s8((int8_t *) mem_addr, masked); -} - -// Multiply packed signed 16-bit integers in a and b, producing intermediate -// signed 32-bit integers. Shift right by 15 bits while rounding up, and store -// the packed 16-bit integers in dst. -// -// r0 := Round(((int32_t)a0 * (int32_t)b0) >> 15) -// r1 := Round(((int32_t)a1 * (int32_t)b1) >> 15) -// r2 := Round(((int32_t)a2 * (int32_t)b2) >> 15) -// ... -// r7 := Round(((int32_t)a7 * (int32_t)b7) >> 15) -FORCE_INLINE __m128i _mm_mulhrs_epi16(__m128i a, __m128i b) -{ - // Has issues due to saturation - // return vreinterpretq_m128i_s16(vqrdmulhq_s16(a, b)); - - // Multiply - int32x4_t mul_lo = vmull_s16(vget_low_s16(vreinterpretq_s16_m128i(a)), - vget_low_s16(vreinterpretq_s16_m128i(b))); - int32x4_t mul_hi = vmull_s16(vget_high_s16(vreinterpretq_s16_m128i(a)), - vget_high_s16(vreinterpretq_s16_m128i(b))); - - // Rounding narrowing shift right - // narrow = (int16_t)((mul + 16384) >> 15); - int16x4_t narrow_lo = vrshrn_n_s32(mul_lo, 15); - int16x4_t narrow_hi = vrshrn_n_s32(mul_hi, 15); - - // Join together - return vreinterpretq_m128i_s16(vcombine_s16(narrow_lo, narrow_hi)); -} - -// Vertically multiply each unsigned 8-bit integer from a with the corresponding -// signed 8-bit integer from b, producing intermediate signed 16-bit integers. -// Horizontally add adjacent pairs of intermediate signed 16-bit integers, -// and pack the saturated results in dst. -// -// FOR j := 0 to 7 -// i := j*16 -// dst[i+15:i] := Saturate_To_Int16( a[i+15:i+8]*b[i+15:i+8] + -// a[i+7:i]*b[i+7:i] ) -// ENDFOR -FORCE_INLINE __m128i _mm_maddubs_epi16(__m128i _a, __m128i _b) -{ -#if defined(__aarch64__) + // Use a signed shift right to create a mask with the sign bit + uint8x16_t mask = + vreinterpretq_u8_s8(vshrq_n_s8(vreinterpretq_s8_m128i(_mask), 7)); uint8x16_t a = vreinterpretq_u8_m128i(_a); - int8x16_t b = vreinterpretq_s8_m128i(_b); - int16x8_t tl = vmulq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(a))), - vmovl_s8(vget_low_s8(b))); - int16x8_t th = vmulq_s16(vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(a))), - vmovl_s8(vget_high_s8(b))); - return vreinterpretq_m128i_s16( - vqaddq_s16(vuzp1q_s16(tl, th), vuzp2q_s16(tl, th))); + uint8x16_t b = vreinterpretq_u8_m128i(_b); + return vreinterpretq_m128i_u8(vbslq_u8(mask, b, a)); +} + +// Blend packed double-precision (64-bit) floating-point elements from a and b +// using mask, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blendv_pd +FORCE_INLINE __m128d _mm_blendv_pd(__m128d _a, __m128d _b, __m128d _mask) +{ + uint64x2_t mask = + vreinterpretq_u64_s64(vshrq_n_s64(vreinterpretq_s64_m128d(_mask), 63)); +#if defined(__aarch64__) + float64x2_t a = vreinterpretq_f64_m128d(_a); + float64x2_t b = vreinterpretq_f64_m128d(_b); + return vreinterpretq_m128d_f64(vbslq_f64(mask, b, a)); #else - // This would be much simpler if x86 would choose to zero extend OR sign - // extend, not both. This could probably be optimized better. - uint16x8_t a = vreinterpretq_u16_m128i(_a); - int16x8_t b = vreinterpretq_s16_m128i(_b); - - // Zero extend a - int16x8_t a_odd = vreinterpretq_s16_u16(vshrq_n_u16(a, 8)); - int16x8_t a_even = vreinterpretq_s16_u16(vbicq_u16(a, vdupq_n_u16(0xff00))); - - // Sign extend by shifting left then shifting right. - int16x8_t b_even = vshrq_n_s16(vshlq_n_s16(b, 8), 8); - int16x8_t b_odd = vshrq_n_s16(b, 8); - - // multiply - int16x8_t prod1 = vmulq_s16(a_even, b_even); - int16x8_t prod2 = vmulq_s16(a_odd, b_odd); - - // saturated add - return vreinterpretq_m128i_s16(vqaddq_s16(prod1, prod2)); + uint64x2_t a = vreinterpretq_u64_m128d(_a); + uint64x2_t b = vreinterpretq_u64_m128d(_b); + return vreinterpretq_m128d_u64(vbslq_u64(mask, b, a)); #endif } -// Computes the fused multiple add product of 32-bit floating point numbers. -// -// Return Value -// Multiplies A and B, and adds C to the temporary result before returning it. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_fmadd -FORCE_INLINE __m128 _mm_fmadd_ps(__m128 a, __m128 b, __m128 c) +// Blend packed single-precision (32-bit) floating-point elements from a and b +// using mask, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blendv_ps +FORCE_INLINE __m128 _mm_blendv_ps(__m128 _a, __m128 _b, __m128 _mask) +{ + // Use a signed shift right to create a mask with the sign bit + uint32x4_t mask = + vreinterpretq_u32_s32(vshrq_n_s32(vreinterpretq_s32_m128(_mask), 31)); + float32x4_t a = vreinterpretq_f32_m128(_a); + float32x4_t b = vreinterpretq_f32_m128(_b); + return vreinterpretq_m128_f32(vbslq_f32(mask, b, a)); +} + +// Round the packed double-precision (64-bit) floating-point elements in a up +// to an integer value, and store the results as packed double-precision +// floating-point elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_pd +FORCE_INLINE __m128d _mm_ceil_pd(__m128d a) { #if defined(__aarch64__) - return vreinterpretq_m128_f32(vfmaq_f32(vreinterpretq_f32_m128(c), - vreinterpretq_f32_m128(b), - vreinterpretq_f32_m128(a))); + return vreinterpretq_m128d_f64(vrndpq_f64(vreinterpretq_f64_m128d(a))); #else - return _mm_add_ps(_mm_mul_ps(a, b), c); + double *f = (double *) &a; + return _mm_set_pd(ceil(f[1]), ceil(f[0])); #endif } -// Alternatively add and subtract packed double-precision (64-bit) -// floating-point elements in a to/from packed elements in b, and store the -// results in dst. -// -// FOR j := 0 to 1 -// i := j*64 -// IF ((j & 1) == 0) -// dst[i+63:i] := a[i+63:i] - b[i+63:i] -// ELSE -// dst[i+63:i] := a[i+63:i] + b[i+63:i] -// FI -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_addsub_pd -FORCE_INLINE __m128d _mm_addsub_pd(__m128d a, __m128d b) +// Round the packed single-precision (32-bit) floating-point elements in a up to +// an integer value, and store the results as packed single-precision +// floating-point elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_ps +FORCE_INLINE __m128 _mm_ceil_ps(__m128 a) { - __m128d mask = _mm_set_pd(1.0f, -1.0f); -#if defined(__aarch64__) - return vreinterpretq_m128d_f64(vfmaq_f64(vreinterpretq_f64_m128d(a), - vreinterpretq_f64_m128d(b), - vreinterpretq_f64_m128d(mask))); +#if defined(__aarch64__) || defined(__ARM_FEATURE_DIRECTED_ROUNDING) + return vreinterpretq_m128_f32(vrndpq_f32(vreinterpretq_f32_m128(a))); #else - return _mm_add_pd(_mm_mul_pd(b, mask), a); + float *f = (float *) &a; + return _mm_set_ps(ceilf(f[3]), ceilf(f[2]), ceilf(f[1]), ceilf(f[0])); #endif } -// Alternatively add and subtract packed single-precision (32-bit) -// floating-point elements in a to/from packed elements in b, and store the -// results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=addsub_ps -FORCE_INLINE __m128 _mm_addsub_ps(__m128 a, __m128 b) +// Round the lower double-precision (64-bit) floating-point element in b up to +// an integer value, store the result as a double-precision floating-point +// element in the lower element of dst, and copy the upper element from a to the +// upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_sd +FORCE_INLINE __m128d _mm_ceil_sd(__m128d a, __m128d b) { - __m128 mask = {-1.0f, 1.0f, -1.0f, 1.0f}; - return _mm_fmadd_ps(b, mask, a); + return _mm_move_sd(a, _mm_ceil_pd(b)); } -// Horizontally add adjacent pairs of double-precision (64-bit) floating-point -// elements in a and b, and pack the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadd_pd -FORCE_INLINE __m128d _mm_hadd_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vpaddq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - double *da = (double *) &a; - double *db = (double *) &b; - double c[] = {da[0] + da[1], db[0] + db[1]}; - return vreinterpretq_m128d_u64(vld1q_u64((uint64_t *) c)); -#endif -} - -// Compute the absolute differences of packed unsigned 8-bit integers in a and -// b, then horizontally sum each consecutive 8 differences to produce two -// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low -// 16 bits of 64-bit elements in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sad_epu8 -FORCE_INLINE __m128i _mm_sad_epu8(__m128i a, __m128i b) -{ - uint16x8_t t = vpaddlq_u8(vabdq_u8((uint8x16_t) a, (uint8x16_t) b)); - uint16_t r0 = t[0] + t[1] + t[2] + t[3]; - uint16_t r4 = t[4] + t[5] + t[6] + t[7]; - uint16x8_t r = vsetq_lane_u16(r0, vdupq_n_u16(0), 0); - return (__m128i) vsetq_lane_u16(r4, r, 4); -} - -// Compute the absolute differences of packed unsigned 8-bit integers in a and -// b, then horizontally sum each consecutive 8 differences to produce four -// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low -// 16 bits of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_sad_pu8 -FORCE_INLINE __m64 _mm_sad_pu8(__m64 a, __m64 b) -{ - uint16x4_t t = - vpaddl_u8(vabd_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); - uint16_t r0 = t[0] + t[1] + t[2] + t[3]; - return vreinterpret_m64_u16(vset_lane_u16(r0, vdup_n_u16(0), 0)); -} - -// Compute the absolute differences of packed unsigned 8-bit integers in a and -// b, then horizontally sum each consecutive 8 differences to produce four -// unsigned 16-bit integers, and pack these unsigned 16-bit integers in the low -// 16 bits of dst. +// Round the lower single-precision (32-bit) floating-point element in b up to +// an integer value, store the result as a single-precision floating-point +// element in the lower element of dst, and copy the upper 3 packed elements +// from a to the upper elements of dst. // -// FOR j := 0 to 7 -// i := j*8 -// tmp[i+7:i] := ABS(a[i+7:i] - b[i+7:i]) -// ENDFOR -// dst[15:0] := tmp[7:0] + tmp[15:8] + tmp[23:16] + tmp[31:24] + tmp[39:32] + -// tmp[47:40] + tmp[55:48] + tmp[63:56] dst[63:16] := 0 -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_psadbw -#define _m_psadbw(a, b) _mm_sad_pu8(a, b) - -// Divides the four single-precision, floating-point values of a and b. -// -// r0 := a0 / b0 -// r1 := a1 / b1 -// r2 := a2 / b2 -// r3 := a3 / b3 -// -// https://msdn.microsoft.com/en-us/library/edaw8147(v=vs.100).aspx -FORCE_INLINE __m128 _mm_div_ps(__m128 a, __m128 b) -{ -#if defined(__aarch64__) && !SSE2NEON_PRECISE_DIV - return vreinterpretq_m128_f32( - vdivq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -#else - float32x4_t recip = vrecpeq_f32(vreinterpretq_f32_m128(b)); - recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(b))); -#if SSE2NEON_PRECISE_DIV - // Additional Netwon-Raphson iteration for accuracy - recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(b))); -#endif - return vreinterpretq_m128_f32(vmulq_f32(vreinterpretq_f32_m128(a), recip)); -#endif -} - -// Divides the scalar single-precision floating point value of a by b. -// https://msdn.microsoft.com/en-us/library/4y73xa49(v=vs.100).aspx -FORCE_INLINE __m128 _mm_div_ss(__m128 a, __m128 b) -{ - float32_t value = - vgetq_lane_f32(vreinterpretq_f32_m128(_mm_div_ps(a, b)), 0); - return vreinterpretq_m128_f32( - vsetq_lane_f32(value, vreinterpretq_f32_m128(a), 0)); -} - -// Divide packed double-precision (64-bit) floating-point elements in a by -// packed elements in b, and store the results in dst. -// -// FOR j := 0 to 1 -// i := 64*j -// dst[i+63:i] := a[i+63:i] / b[i+63:i] -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_div_pd -FORCE_INLINE __m128d _mm_div_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vdivq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - double *da = (double *) &a; - double *db = (double *) &b; - double c[2]; - c[0] = da[0] / db[0]; - c[1] = da[1] / db[1]; - return vld1q_f32((float32_t *) c); -#endif -} - -// Divide the lower double-precision (64-bit) floating-point element in a by the -// lower double-precision (64-bit) floating-point element in b, store the result -// in the lower element of dst, and copy the upper element from a to the upper -// element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_div_sd -FORCE_INLINE __m128d _mm_div_sd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - float64x2_t tmp = - vdivq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b)); - return vreinterpretq_m128d_f64( - vsetq_lane_f64(vgetq_lane_f64(vreinterpretq_f64_m128d(a), 1), tmp, 1)); -#else - return _mm_move_sd(a, _mm_div_pd(a, b)); -#endif -} - -// Compute the approximate reciprocal of packed single-precision (32-bit) -// floating-point elements in a, and store the results in dst. The maximum -// relative error for this approximation is less than 1.5*2^-12. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_rcp_ps -FORCE_INLINE __m128 _mm_rcp_ps(__m128 in) -{ - float32x4_t recip = vrecpeq_f32(vreinterpretq_f32_m128(in)); - recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(in))); -#if SSE2NEON_PRECISE_DIV - // Additional Netwon-Raphson iteration for accuracy - recip = vmulq_f32(recip, vrecpsq_f32(recip, vreinterpretq_f32_m128(in))); -#endif - return vreinterpretq_m128_f32(recip); -} - -// Compute the approximate reciprocal of the lower single-precision (32-bit) -// floating-point element in a, store the result in the lower element of dst, -// and copy the upper 3 packed elements from a to the upper elements of dst. The -// maximum relative error for this approximation is less than 1.5*2^-12. -// -// dst[31:0] := (1.0 / a[31:0]) +// dst[31:0] := CEIL(b[31:0]) // dst[127:32] := a[127:32] // -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_rcp_ss -FORCE_INLINE __m128 _mm_rcp_ss(__m128 a) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_ss +FORCE_INLINE __m128 _mm_ceil_ss(__m128 a, __m128 b) { - return _mm_move_ss(a, _mm_rcp_ps(a)); + return _mm_move_ss(a, _mm_ceil_ps(b)); } -// Computes the approximations of square roots of the four single-precision, -// floating-point values of a. First computes reciprocal square roots and then -// reciprocals of the four values. -// -// r0 := sqrt(a0) -// r1 := sqrt(a1) -// r2 := sqrt(a2) -// r3 := sqrt(a3) -// -// https://msdn.microsoft.com/en-us/library/vstudio/8z67bwwk(v=vs.100).aspx -FORCE_INLINE __m128 _mm_sqrt_ps(__m128 in) -{ -#if SSE2NEON_PRECISE_SQRT - float32x4_t recip = vrsqrteq_f32(vreinterpretq_f32_m128(in)); - - // Test for vrsqrteq_f32(0) -> positive infinity case. - // Change to zero, so that s * 1/sqrt(s) result is zero too. - const uint32x4_t pos_inf = vdupq_n_u32(0x7F800000); - const uint32x4_t div_by_zero = - vceqq_u32(pos_inf, vreinterpretq_u32_f32(recip)); - recip = vreinterpretq_f32_u32( - vandq_u32(vmvnq_u32(div_by_zero), vreinterpretq_u32_f32(recip))); - - // Additional Netwon-Raphson iteration for accuracy - recip = vmulq_f32( - vrsqrtsq_f32(vmulq_f32(recip, recip), vreinterpretq_f32_m128(in)), - recip); - recip = vmulq_f32( - vrsqrtsq_f32(vmulq_f32(recip, recip), vreinterpretq_f32_m128(in)), - recip); - - // sqrt(s) = s * 1/sqrt(s) - return vreinterpretq_m128_f32(vmulq_f32(vreinterpretq_f32_m128(in), recip)); -#elif defined(__aarch64__) - return vreinterpretq_m128_f32(vsqrtq_f32(vreinterpretq_f32_m128(in))); -#else - float32x4_t recipsq = vrsqrteq_f32(vreinterpretq_f32_m128(in)); - float32x4_t sq = vrecpeq_f32(recipsq); - return vreinterpretq_m128_f32(sq); -#endif -} - -// Computes the approximation of the square root of the scalar single-precision -// floating point value of in. -// https://msdn.microsoft.com/en-us/library/ahfsc22d(v=vs.100).aspx -FORCE_INLINE __m128 _mm_sqrt_ss(__m128 in) -{ - float32_t value = - vgetq_lane_f32(vreinterpretq_f32_m128(_mm_sqrt_ps(in)), 0); - return vreinterpretq_m128_f32( - vsetq_lane_f32(value, vreinterpretq_f32_m128(in), 0)); -} - -// Computes the approximations of the reciprocal square roots of the four -// single-precision floating point values of in. -// The current precision is 1% error. -// https://msdn.microsoft.com/en-us/library/22hfsh53(v=vs.100).aspx -FORCE_INLINE __m128 _mm_rsqrt_ps(__m128 in) -{ - float32x4_t out = vrsqrteq_f32(vreinterpretq_f32_m128(in)); -#if SSE2NEON_PRECISE_SQRT - // Additional Netwon-Raphson iteration for accuracy - out = vmulq_f32( - out, vrsqrtsq_f32(vmulq_f32(vreinterpretq_f32_m128(in), out), out)); - out = vmulq_f32( - out, vrsqrtsq_f32(vmulq_f32(vreinterpretq_f32_m128(in), out), out)); -#endif - return vreinterpretq_m128_f32(out); -} - -// Compute the approximate reciprocal square root of the lower single-precision -// (32-bit) floating-point element in a, store the result in the lower element -// of dst, and copy the upper 3 packed elements from a to the upper elements of -// dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_rsqrt_ss -FORCE_INLINE __m128 _mm_rsqrt_ss(__m128 in) -{ - return vsetq_lane_f32(vgetq_lane_f32(_mm_rsqrt_ps(in), 0), in, 0); -} - -// Compare packed signed 16-bit integers in a and b, and store packed maximum -// values in dst. -// -// FOR j := 0 to 3 -// i := j*16 -// dst[i+15:i] := MAX(a[i+15:i], b[i+15:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pi16 -FORCE_INLINE __m64 _mm_max_pi16(__m64 a, __m64 b) -{ - return vreinterpret_m64_s16( - vmax_s16(vreinterpret_s16_m64(a), vreinterpret_s16_m64(b))); -} - -// Compare packed signed 16-bit integers in a and b, and store packed maximum -// values in dst. -// -// FOR j := 0 to 3 -// i := j*16 -// dst[i+15:i] := MAX(a[i+15:i], b[i+15:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pi16 -#define _m_pmaxsw(a, b) _mm_max_pi16(a, b) - -// Computes the maximums of the four single-precision, floating-point values of -// a and b. -// https://msdn.microsoft.com/en-us/library/vstudio/ff5d607a(v=vs.100).aspx -FORCE_INLINE __m128 _mm_max_ps(__m128 a, __m128 b) -{ -#if SSE2NEON_PRECISE_MINMAX - float32x4_t _a = vreinterpretq_f32_m128(a); - float32x4_t _b = vreinterpretq_f32_m128(b); - return vbslq_f32(vcltq_f32(_b, _a), _a, _b); -#else - return vreinterpretq_m128_f32( - vmaxq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -#endif -} - -// Compare packed unsigned 8-bit integers in a and b, and store packed maximum -// values in dst. -// -// FOR j := 0 to 7 -// i := j*8 -// dst[i+7:i] := MAX(a[i+7:i], b[i+7:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pu8 -FORCE_INLINE __m64 _mm_max_pu8(__m64 a, __m64 b) -{ - return vreinterpret_m64_u8( - vmax_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); -} - -// Compare packed unsigned 8-bit integers in a and b, and store packed maximum -// values in dst. -// -// FOR j := 0 to 7 -// i := j*8 -// dst[i+7:i] := MAX(a[i+7:i], b[i+7:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pu8 -#define _m_pmaxub(a, b) _mm_max_pu8(a, b) - -// Compare packed signed 16-bit integers in a and b, and store packed minimum -// values in dst. -// -// FOR j := 0 to 3 -// i := j*16 -// dst[i+15:i] := MIN(a[i+15:i], b[i+15:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pi16 -FORCE_INLINE __m64 _mm_min_pi16(__m64 a, __m64 b) -{ - return vreinterpret_m64_s16( - vmin_s16(vreinterpret_s16_m64(a), vreinterpret_s16_m64(b))); -} - -// Compare packed signed 16-bit integers in a and b, and store packed minimum -// values in dst. -// -// FOR j := 0 to 3 -// i := j*16 -// dst[i+15:i] := MIN(a[i+15:i], b[i+15:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pi16 -#define _m_pminsw(a, b) _mm_min_pi16(a, b) - -// Computes the minima of the four single-precision, floating-point values of a -// and b. -// https://msdn.microsoft.com/en-us/library/vstudio/wh13kadz(v=vs.100).aspx -FORCE_INLINE __m128 _mm_min_ps(__m128 a, __m128 b) -{ -#if SSE2NEON_PRECISE_MINMAX - float32x4_t _a = vreinterpretq_f32_m128(a); - float32x4_t _b = vreinterpretq_f32_m128(b); - return vbslq_f32(vcltq_f32(_a, _b), _a, _b); -#else - return vreinterpretq_m128_f32( - vminq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -#endif -} - -// Compare packed unsigned 8-bit integers in a and b, and store packed minimum -// values in dst. -// -// FOR j := 0 to 7 -// i := j*8 -// dst[i+7:i] := MIN(a[i+7:i], b[i+7:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pu8 -FORCE_INLINE __m64 _mm_min_pu8(__m64 a, __m64 b) -{ - return vreinterpret_m64_u8( - vmin_u8(vreinterpret_u8_m64(a), vreinterpret_u8_m64(b))); -} - -// Compare packed unsigned 8-bit integers in a and b, and store packed minimum -// values in dst. -// -// FOR j := 0 to 7 -// i := j*8 -// dst[i+7:i] := MIN(a[i+7:i], b[i+7:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pu8 -#define _m_pminub(a, b) _mm_min_pu8(a, b) - -// Computes the maximum of the two lower scalar single-precision floating point -// values of a and b. -// https://msdn.microsoft.com/en-us/library/s6db5esz(v=vs.100).aspx -FORCE_INLINE __m128 _mm_max_ss(__m128 a, __m128 b) -{ - float32_t value = vgetq_lane_f32(_mm_max_ps(a, b), 0); - return vreinterpretq_m128_f32( - vsetq_lane_f32(value, vreinterpretq_f32_m128(a), 0)); -} - -// Computes the minimum of the two lower scalar single-precision floating point -// values of a and b. -// https://msdn.microsoft.com/en-us/library/0a9y7xaa(v=vs.100).aspx -FORCE_INLINE __m128 _mm_min_ss(__m128 a, __m128 b) -{ - float32_t value = vgetq_lane_f32(_mm_min_ps(a, b), 0); - return vreinterpretq_m128_f32( - vsetq_lane_f32(value, vreinterpretq_f32_m128(a), 0)); -} - -// Computes the pairwise maxima of the 16 unsigned 8-bit integers from a and the -// 16 unsigned 8-bit integers from b. -// https://msdn.microsoft.com/en-us/library/st6634za(v=vs.100).aspx -FORCE_INLINE __m128i _mm_max_epu8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u8( - vmaxq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); -} - -// Compare packed double-precision (64-bit) floating-point elements in a and b, -// and store packed maximum values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_pd -FORCE_INLINE __m128d _mm_max_pd(__m128d a, __m128d b) +// Compare packed 64-bit integers in a and b for equality, and store the results +// in dst +FORCE_INLINE __m128i _mm_cmpeq_epi64(__m128i a, __m128i b) { #if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vmaxq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); + return vreinterpretq_m128i_u64( + vceqq_u64(vreinterpretq_u64_m128i(a), vreinterpretq_u64_m128i(b))); #else - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) > (*(double *) &b0) ? a0 : b0; - d[1] = (*(double *) &a1) > (*(double *) &b1) ? a1 : b1; - - return vreinterpretq_m128d_u64(vld1q_u64(d)); + // ARMv7 lacks vceqq_u64 + // (a == b) -> (a_lo == b_lo) && (a_hi == b_hi) + uint32x4_t cmp = + vceqq_u32(vreinterpretq_u32_m128i(a), vreinterpretq_u32_m128i(b)); + uint32x4_t swapped = vrev64q_u32(cmp); + return vreinterpretq_m128i_u32(vandq_u32(cmp, swapped)); #endif } -// Compare the lower double-precision (64-bit) floating-point elements in a and -// b, store the maximum value in the lower element of dst, and copy the upper -// element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_sd -FORCE_INLINE __m128d _mm_max_sd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return _mm_move_sd(a, _mm_max_pd(a, b)); -#else - double *da = (double *) &a; - double *db = (double *) &b; - double c[2] = {fmax(da[0], db[0]), da[1]}; - return vld1q_f32((float32_t *) c); -#endif -} - -// Computes the pairwise minima of the 16 unsigned 8-bit integers from a and the -// 16 unsigned 8-bit integers from b. -// https://msdn.microsoft.com/ko-kr/library/17k8cf58(v=vs.100).aspxx -FORCE_INLINE __m128i _mm_min_epu8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u8( - vminq_u8(vreinterpretq_u8_m128i(a), vreinterpretq_u8_m128i(b))); -} - -// Compare packed double-precision (64-bit) floating-point elements in a and b, -// and store packed minimum values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_pd -FORCE_INLINE __m128d _mm_min_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vminq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) < (*(double *) &b0) ? a0 : b0; - d[1] = (*(double *) &a1) < (*(double *) &b1) ? a1 : b1; - return vreinterpretq_m128d_u64(vld1q_u64(d)); -#endif -} - -// Compare the lower double-precision (64-bit) floating-point elements in a and -// b, store the minimum value in the lower element of dst, and copy the upper -// element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_sd -FORCE_INLINE __m128d _mm_min_sd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return _mm_move_sd(a, _mm_min_pd(a, b)); -#else - double *da = (double *) &a; - double *db = (double *) &b; - double c[2] = {fmin(da[0], db[0]), da[1]}; - return vld1q_f32((float32_t *) c); -#endif -} - -// Computes the pairwise minima of the 8 signed 16-bit integers from a and the 8 -// signed 16-bit integers from b. -// https://msdn.microsoft.com/en-us/library/vstudio/6te997ew(v=vs.100).aspx -FORCE_INLINE __m128i _mm_min_epi16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s16( - vminq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -} - -// Compare packed signed 8-bit integers in a and b, and store packed maximum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epi8 -FORCE_INLINE __m128i _mm_max_epi8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s8( - vmaxq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -} - -// Compare packed unsigned 16-bit integers in a and b, and store packed maximum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epu16 -FORCE_INLINE __m128i _mm_max_epu16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u16( - vmaxq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); -} - -// Compare packed signed 8-bit integers in a and b, and store packed minimum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_epi8 -FORCE_INLINE __m128i _mm_min_epi8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s8( - vminq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -} - -// Compare packed unsigned 16-bit integers in a and b, and store packed minimum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_epu16 -FORCE_INLINE __m128i _mm_min_epu16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u16( - vminq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); -} - -// Computes the pairwise maxima of the 8 signed 16-bit integers from a and the 8 -// signed 16-bit integers from b. -// https://msdn.microsoft.com/en-us/LIBRary/3x060h7c(v=vs.100).aspx -FORCE_INLINE __m128i _mm_max_epi16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s16( - vmaxq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -} - -// epi versions of min/max -// Computes the pariwise maximums of the four signed 32-bit integer values of a -// and b. -// -// A 128-bit parameter that can be defined with the following equations: -// r0 := (a0 > b0) ? a0 : b0 -// r1 := (a1 > b1) ? a1 : b1 -// r2 := (a2 > b2) ? a2 : b2 -// r3 := (a3 > b3) ? a3 : b3 -// -// https://msdn.microsoft.com/en-us/library/vstudio/bb514055(v=vs.100).aspx -FORCE_INLINE __m128i _mm_max_epi32(__m128i a, __m128i b) +// Converts the four signed 16-bit integers in the lower 64 bits to four signed +// 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepi16_epi32(__m128i a) { return vreinterpretq_m128i_s32( - vmaxq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); + vmovl_s16(vget_low_s16(vreinterpretq_s16_m128i(a)))); } -// Computes the pariwise minima of the four signed 32-bit integer values of a -// and b. -// -// A 128-bit parameter that can be defined with the following equations: -// r0 := (a0 < b0) ? a0 : b0 -// r1 := (a1 < b1) ? a1 : b1 -// r2 := (a2 < b2) ? a2 : b2 -// r3 := (a3 < b3) ? a3 : b3 -// -// https://msdn.microsoft.com/en-us/library/vstudio/bb531476(v=vs.100).aspx -FORCE_INLINE __m128i _mm_min_epi32(__m128i a, __m128i b) +// Converts the two signed 16-bit integers in the lower 32 bits two signed +// 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepi16_epi64(__m128i a) { - return vreinterpretq_m128i_s32( - vminq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); + int16x8_t s16x8 = vreinterpretq_s16_m128i(a); /* xxxx xxxx xxxx 0B0A */ + int32x4_t s32x4 = vmovl_s16(vget_low_s16(s16x8)); /* 000x 000x 000B 000A */ + int64x2_t s64x2 = vmovl_s32(vget_low_s32(s32x4)); /* 0000 000B 0000 000A */ + return vreinterpretq_m128i_s64(s64x2); } -// Compare packed unsigned 32-bit integers in a and b, and store packed maximum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epu32 -FORCE_INLINE __m128i _mm_max_epu32(__m128i a, __m128i b) +// Converts the two signed 32-bit integers in the lower 64 bits to two signed +// 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepi32_epi64(__m128i a) +{ + return vreinterpretq_m128i_s64( + vmovl_s32(vget_low_s32(vreinterpretq_s32_m128i(a)))); +} + +// Converts the four unsigned 8-bit integers in the lower 16 bits to four +// unsigned 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepi8_epi16(__m128i a) +{ + int8x16_t s8x16 = vreinterpretq_s8_m128i(a); /* xxxx xxxx xxxx DCBA */ + int16x8_t s16x8 = vmovl_s8(vget_low_s8(s8x16)); /* 0x0x 0x0x 0D0C 0B0A */ + return vreinterpretq_m128i_s16(s16x8); +} + +// Converts the four unsigned 8-bit integers in the lower 32 bits to four +// unsigned 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepi8_epi32(__m128i a) +{ + int8x16_t s8x16 = vreinterpretq_s8_m128i(a); /* xxxx xxxx xxxx DCBA */ + int16x8_t s16x8 = vmovl_s8(vget_low_s8(s8x16)); /* 0x0x 0x0x 0D0C 0B0A */ + int32x4_t s32x4 = vmovl_s16(vget_low_s16(s16x8)); /* 000D 000C 000B 000A */ + return vreinterpretq_m128i_s32(s32x4); +} + +// Converts the two signed 8-bit integers in the lower 32 bits to four +// signed 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepi8_epi64(__m128i a) +{ + int8x16_t s8x16 = vreinterpretq_s8_m128i(a); /* xxxx xxxx xxxx xxBA */ + int16x8_t s16x8 = vmovl_s8(vget_low_s8(s8x16)); /* 0x0x 0x0x 0x0x 0B0A */ + int32x4_t s32x4 = vmovl_s16(vget_low_s16(s16x8)); /* 000x 000x 000B 000A */ + int64x2_t s64x2 = vmovl_s32(vget_low_s32(s32x4)); /* 0000 000B 0000 000A */ + return vreinterpretq_m128i_s64(s64x2); +} + +// Converts the four unsigned 16-bit integers in the lower 64 bits to four +// unsigned 32-bit integers. +FORCE_INLINE __m128i _mm_cvtepu16_epi32(__m128i a) { return vreinterpretq_m128i_u32( - vmaxq_u32(vreinterpretq_u32_m128i(a), vreinterpretq_u32_m128i(b))); + vmovl_u16(vget_low_u16(vreinterpretq_u16_m128i(a)))); } -// Compare packed unsigned 32-bit integers in a and b, and store packed minimum -// values in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epu32 -FORCE_INLINE __m128i _mm_min_epu32(__m128i a, __m128i b) +// Converts the two unsigned 16-bit integers in the lower 32 bits to two +// unsigned 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepu16_epi64(__m128i a) { - return vreinterpretq_m128i_u32( - vminq_u32(vreinterpretq_u32_m128i(a), vreinterpretq_u32_m128i(b))); + uint16x8_t u16x8 = vreinterpretq_u16_m128i(a); /* xxxx xxxx xxxx 0B0A */ + uint32x4_t u32x4 = vmovl_u16(vget_low_u16(u16x8)); /* 000x 000x 000B 000A */ + uint64x2_t u64x2 = vmovl_u32(vget_low_u32(u32x4)); /* 0000 000B 0000 000A */ + return vreinterpretq_m128i_u64(u64x2); } -// Multiply the packed unsigned 16-bit integers in a and b, producing -// intermediate 32-bit integers, and store the high 16 bits of the intermediate -// integers in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mulhi_pu16 -FORCE_INLINE __m64 _mm_mulhi_pu16(__m64 a, __m64 b) +// Converts the two unsigned 32-bit integers in the lower 64 bits to two +// unsigned 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepu32_epi64(__m128i a) { - return vreinterpret_m64_u16(vshrn_n_u32( - vmull_u16(vreinterpret_u16_m64(a), vreinterpret_u16_m64(b)), 16)); + return vreinterpretq_m128i_u64( + vmovl_u32(vget_low_u32(vreinterpretq_u32_m128i(a)))); } -// Multiplies the 8 signed 16-bit integers from a by the 8 signed 16-bit -// integers from b. -// -// r0 := (a0 * b0)[31:16] -// r1 := (a1 * b1)[31:16] -// ... -// r7 := (a7 * b7)[31:16] -// -// https://msdn.microsoft.com/en-us/library/vstudio/59hddw1d(v=vs.100).aspx -FORCE_INLINE __m128i _mm_mulhi_epi16(__m128i a, __m128i b) +// Zero extend packed unsigned 8-bit integers in a to packed 16-bit integers, +// and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtepu8_epi16 +FORCE_INLINE __m128i _mm_cvtepu8_epi16(__m128i a) { - /* FIXME: issue with large values because of result saturation */ - // int16x8_t ret = vqdmulhq_s16(vreinterpretq_s16_m128i(a), - // vreinterpretq_s16_m128i(b)); /* =2*a*b */ return - // vreinterpretq_m128i_s16(vshrq_n_s16(ret, 1)); - int16x4_t a3210 = vget_low_s16(vreinterpretq_s16_m128i(a)); - int16x4_t b3210 = vget_low_s16(vreinterpretq_s16_m128i(b)); - int32x4_t ab3210 = vmull_s16(a3210, b3210); /* 3333222211110000 */ - int16x4_t a7654 = vget_high_s16(vreinterpretq_s16_m128i(a)); - int16x4_t b7654 = vget_high_s16(vreinterpretq_s16_m128i(b)); - int32x4_t ab7654 = vmull_s16(a7654, b7654); /* 7777666655554444 */ - uint16x8x2_t r = - vuzpq_u16(vreinterpretq_u16_s32(ab3210), vreinterpretq_u16_s32(ab7654)); - return vreinterpretq_m128i_u16(r.val[1]); + uint8x16_t u8x16 = vreinterpretq_u8_m128i(a); /* xxxx xxxx HGFE DCBA */ + uint16x8_t u16x8 = vmovl_u8(vget_low_u8(u8x16)); /* 0H0G 0F0E 0D0C 0B0A */ + return vreinterpretq_m128i_u16(u16x8); } -// Multiply the packed unsigned 16-bit integers in a and b, producing -// intermediate 32-bit integers, and store the high 16 bits of the intermediate -// integers in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mulhi_epu16 -FORCE_INLINE __m128i _mm_mulhi_epu16(__m128i a, __m128i b) +// Converts the four unsigned 8-bit integers in the lower 32 bits to four +// unsigned 32-bit integers. +// https://msdn.microsoft.com/en-us/library/bb531467%28v=vs.100%29.aspx +FORCE_INLINE __m128i _mm_cvtepu8_epi32(__m128i a) { - uint16x4_t a3210 = vget_low_u16(vreinterpretq_u16_m128i(a)); - uint16x4_t b3210 = vget_low_u16(vreinterpretq_u16_m128i(b)); - uint32x4_t ab3210 = vmull_u16(a3210, b3210); -#if defined(__aarch64__) - uint32x4_t ab7654 = - vmull_high_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b)); - uint16x8_t r = vuzp2q_u16(vreinterpretq_u16_u32(ab3210), - vreinterpretq_u16_u32(ab7654)); - return vreinterpretq_m128i_u16(r); -#else - uint16x4_t a7654 = vget_high_u16(vreinterpretq_u16_m128i(a)); - uint16x4_t b7654 = vget_high_u16(vreinterpretq_u16_m128i(b)); - uint32x4_t ab7654 = vmull_u16(a7654, b7654); - uint16x8x2_t r = - vuzpq_u16(vreinterpretq_u16_u32(ab3210), vreinterpretq_u16_u32(ab7654)); - return vreinterpretq_m128i_u16(r.val[1]); + uint8x16_t u8x16 = vreinterpretq_u8_m128i(a); /* xxxx xxxx xxxx DCBA */ + uint16x8_t u16x8 = vmovl_u8(vget_low_u8(u8x16)); /* 0x0x 0x0x 0D0C 0B0A */ + uint32x4_t u32x4 = vmovl_u16(vget_low_u16(u16x8)); /* 000D 000C 000B 000A */ + return vreinterpretq_m128i_u32(u32x4); +} + +// Converts the two unsigned 8-bit integers in the lower 16 bits to two +// unsigned 64-bit integers. +FORCE_INLINE __m128i _mm_cvtepu8_epi64(__m128i a) +{ + uint8x16_t u8x16 = vreinterpretq_u8_m128i(a); /* xxxx xxxx xxxx xxBA */ + uint16x8_t u16x8 = vmovl_u8(vget_low_u8(u8x16)); /* 0x0x 0x0x 0x0x 0B0A */ + uint32x4_t u32x4 = vmovl_u16(vget_low_u16(u16x8)); /* 000x 000x 000B 000A */ + uint64x2_t u64x2 = vmovl_u32(vget_low_u32(u32x4)); /* 0000 000B 0000 000A */ + return vreinterpretq_m128i_u64(u64x2); +} + +// Conditionally multiply the packed double-precision (64-bit) floating-point +// elements in a and b using the high 4 bits in imm8, sum the four products, and +// conditionally store the sum in dst using the low 4 bits of imm8. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_dp_pd +FORCE_INLINE __m128d _mm_dp_pd(__m128d a, __m128d b, const int imm) +{ + // Generate mask value from constant immediate bit value + const int64_t bit0Mask = imm & 0x01 ? UINT64_MAX : 0; + const int64_t bit1Mask = imm & 0x02 ? UINT64_MAX : 0; +#if !SSE2NEON_PRECISE_DP + const int64_t bit4Mask = imm & 0x10 ? UINT64_MAX : 0; + const int64_t bit5Mask = imm & 0x20 ? UINT64_MAX : 0; #endif -} - -// Computes pairwise add of each argument as single-precision, floating-point -// values a and b. -// https://msdn.microsoft.com/en-us/library/yd9wecaa.aspx -FORCE_INLINE __m128 _mm_hadd_ps(__m128 a, __m128 b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128_f32( - vpaddq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); + // Conditional multiplication +#if !SSE2NEON_PRECISE_DP + __m128d mul = _mm_mul_pd(a, b); + const __m128d mulMask = + _mm_castsi128_pd(_mm_set_epi64x(bit5Mask, bit4Mask)); + __m128d tmp = _mm_and_pd(mul, mulMask); #else - float32x2_t a10 = vget_low_f32(vreinterpretq_f32_m128(a)); - float32x2_t a32 = vget_high_f32(vreinterpretq_f32_m128(a)); - float32x2_t b10 = vget_low_f32(vreinterpretq_f32_m128(b)); - float32x2_t b32 = vget_high_f32(vreinterpretq_f32_m128(b)); - return vreinterpretq_m128_f32( - vcombine_f32(vpadd_f32(a10, a32), vpadd_f32(b10, b32))); -#endif -} - -// Computes pairwise add of each argument as a 16-bit signed or unsigned integer -// values a and b. -FORCE_INLINE __m128i _mm_hadd_epi16(__m128i _a, __m128i _b) -{ - int16x8_t a = vreinterpretq_s16_m128i(_a); - int16x8_t b = vreinterpretq_s16_m128i(_b); #if defined(__aarch64__) - return vreinterpretq_m128i_s16(vpaddq_s16(a, b)); + double d0 = (imm & 0x10) ? vgetq_lane_f64(vreinterpretq_f64_m128d(a), 0) * + vgetq_lane_f64(vreinterpretq_f64_m128d(b), 0) + : 0; + double d1 = (imm & 0x20) ? vgetq_lane_f64(vreinterpretq_f64_m128d(a), 1) * + vgetq_lane_f64(vreinterpretq_f64_m128d(b), 1) + : 0; #else - return vreinterpretq_m128i_s16( - vcombine_s16(vpadd_s16(vget_low_s16(a), vget_high_s16(a)), - vpadd_s16(vget_low_s16(b), vget_high_s16(b)))); + double d0 = (imm & 0x10) ? ((double *) &a)[0] * ((double *) &b)[0] : 0; + double d1 = (imm & 0x20) ? ((double *) &a)[1] * ((double *) &b)[1] : 0; #endif -} - -// Horizontally subtract adjacent pairs of double-precision (64-bit) -// floating-point elements in a and b, and pack the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_pd -FORCE_INLINE __m128d _mm_hsub_pd(__m128d _a, __m128d _b) -{ + __m128d tmp = _mm_set_pd(d1, d0); +#endif + // Sum the products #if defined(__aarch64__) - return vreinterpretq_m128d_f64(vsubq_f64( - vuzp1q_f64(vreinterpretq_f64_m128d(_a), vreinterpretq_f64_m128d(_b)), - vuzp2q_f64(vreinterpretq_f64_m128d(_a), vreinterpretq_f64_m128d(_b)))); + double sum = vpaddd_f64(vreinterpretq_f64_m128d(tmp)); #else - double *da = (double *) &_a; - double *db = (double *) &_b; - double c[] = {da[0] - da[1], db[0] - db[1]}; - return vreinterpretq_m128d_u64(vld1q_u64((uint64_t *) c)); + double sum = *((double *) &tmp) + *(((double *) &tmp) + 1); #endif -} - -// Horizontally substract adjacent pairs of single-precision (32-bit) -// floating-point elements in a and b, and pack the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsub_ps -FORCE_INLINE __m128 _mm_hsub_ps(__m128 _a, __m128 _b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128_f32(vsubq_f32( - vuzp1q_f32(vreinterpretq_f32_m128(_a), vreinterpretq_f32_m128(_b)), - vuzp2q_f32(vreinterpretq_f32_m128(_a), vreinterpretq_f32_m128(_b)))); -#else - float32x4x2_t c = - vuzpq_f32(vreinterpretq_f32_m128(_a), vreinterpretq_f32_m128(_b)); - return vreinterpretq_m128_f32(vsubq_f32(c.val[0], c.val[1])); -#endif -} - -// Horizontally add adjacent pairs of 16-bit integers in a and b, and pack the -// signed 16-bit results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadd_pi16 -FORCE_INLINE __m64 _mm_hadd_pi16(__m64 a, __m64 b) -{ - return vreinterpret_m64_s16( - vpadd_s16(vreinterpret_s16_m64(a), vreinterpret_s16_m64(b))); -} - -// Horizontally add adjacent pairs of 32-bit integers in a and b, and pack the -// signed 32-bit results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hadd_pi32 -FORCE_INLINE __m64 _mm_hadd_pi32(__m64 a, __m64 b) -{ - return vreinterpret_m64_s32( - vpadd_s32(vreinterpret_s32_m64(a), vreinterpret_s32_m64(b))); -} - -// Computes pairwise difference of each argument as a 16-bit signed or unsigned -// integer values a and b. -FORCE_INLINE __m128i _mm_hsub_epi16(__m128i _a, __m128i _b) -{ - int32x4_t a = vreinterpretq_s32_m128i(_a); - int32x4_t b = vreinterpretq_s32_m128i(_b); - // Interleave using vshrn/vmovn - // [a0|a2|a4|a6|b0|b2|b4|b6] - // [a1|a3|a5|a7|b1|b3|b5|b7] - int16x8_t ab0246 = vcombine_s16(vmovn_s32(a), vmovn_s32(b)); - int16x8_t ab1357 = vcombine_s16(vshrn_n_s32(a, 16), vshrn_n_s32(b, 16)); - // Subtract - return vreinterpretq_m128i_s16(vsubq_s16(ab0246, ab1357)); -} - -// Computes saturated pairwise sub of each argument as a 16-bit signed -// integer values a and b. -FORCE_INLINE __m128i _mm_hadds_epi16(__m128i _a, __m128i _b) -{ -#if defined(__aarch64__) - int16x8_t a = vreinterpretq_s16_m128i(_a); - int16x8_t b = vreinterpretq_s16_m128i(_b); - return vreinterpretq_s64_s16( - vqaddq_s16(vuzp1q_s16(a, b), vuzp2q_s16(a, b))); -#else - int32x4_t a = vreinterpretq_s32_m128i(_a); - int32x4_t b = vreinterpretq_s32_m128i(_b); - // Interleave using vshrn/vmovn - // [a0|a2|a4|a6|b0|b2|b4|b6] - // [a1|a3|a5|a7|b1|b3|b5|b7] - int16x8_t ab0246 = vcombine_s16(vmovn_s32(a), vmovn_s32(b)); - int16x8_t ab1357 = vcombine_s16(vshrn_n_s32(a, 16), vshrn_n_s32(b, 16)); - // Saturated add - return vreinterpretq_m128i_s16(vqaddq_s16(ab0246, ab1357)); -#endif -} - -// Computes saturated pairwise difference of each argument as a 16-bit signed -// integer values a and b. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_hsubs_epi16 -FORCE_INLINE __m128i _mm_hsubs_epi16(__m128i _a, __m128i _b) -{ -#if defined(__aarch64__) - int16x8_t a = vreinterpretq_s16_m128i(_a); - int16x8_t b = vreinterpretq_s16_m128i(_b); - return vreinterpretq_s64_s16( - vqsubq_s16(vuzp1q_s16(a, b), vuzp2q_s16(a, b))); -#else - int32x4_t a = vreinterpretq_s32_m128i(_a); - int32x4_t b = vreinterpretq_s32_m128i(_b); - // Interleave using vshrn/vmovn - // [a0|a2|a4|a6|b0|b2|b4|b6] - // [a1|a3|a5|a7|b1|b3|b5|b7] - int16x8_t ab0246 = vcombine_s16(vmovn_s32(a), vmovn_s32(b)); - int16x8_t ab1357 = vcombine_s16(vshrn_n_s32(a, 16), vshrn_n_s32(b, 16)); - // Saturated subtract - return vreinterpretq_m128i_s16(vqsubq_s16(ab0246, ab1357)); -#endif -} - -// Computes pairwise add of each argument as a 32-bit signed or unsigned integer -// values a and b. -FORCE_INLINE __m128i _mm_hadd_epi32(__m128i _a, __m128i _b) -{ - int32x4_t a = vreinterpretq_s32_m128i(_a); - int32x4_t b = vreinterpretq_s32_m128i(_b); - return vreinterpretq_m128i_s32( - vcombine_s32(vpadd_s32(vget_low_s32(a), vget_high_s32(a)), - vpadd_s32(vget_low_s32(b), vget_high_s32(b)))); -} - -// Computes pairwise difference of each argument as a 32-bit signed or unsigned -// integer values a and b. -FORCE_INLINE __m128i _mm_hsub_epi32(__m128i _a, __m128i _b) -{ - int64x2_t a = vreinterpretq_s64_m128i(_a); - int64x2_t b = vreinterpretq_s64_m128i(_b); - // Interleave using vshrn/vmovn - // [a0|a2|b0|b2] - // [a1|a2|b1|b3] - int32x4_t ab02 = vcombine_s32(vmovn_s64(a), vmovn_s64(b)); - int32x4_t ab13 = vcombine_s32(vshrn_n_s64(a, 32), vshrn_n_s64(b, 32)); - // Subtract - return vreinterpretq_m128i_s32(vsubq_s32(ab02, ab13)); -} - -// Kahan summation for accurate summation of floating-point numbers. -// http://blog.zachbjornson.com/2019/08/11/fast-float-summation.html -FORCE_INLINE void _sse2neon_kadd_f32(float *sum, float *c, float y) -{ - y -= *c; - float t = *sum + y; - *c = (t - *sum) - y; - *sum = t; + // Conditionally store the sum + const __m128d sumMask = + _mm_castsi128_pd(_mm_set_epi64x(bit1Mask, bit0Mask)); + __m128d res = _mm_and_pd(_mm_set_pd1(sum), sumMask); + return res; } // Conditionally multiply the packed single-precision (32-bit) floating-point @@ -4798,1611 +7714,64 @@ FORCE_INLINE __m128 _mm_dp_ps(__m128 a, __m128 b, const int imm) return vreinterpretq_m128_f32(res); } -/* Compare operations */ - -// Compares for less than -// https://msdn.microsoft.com/en-us/library/vstudio/f330yhc8(v=vs.100).aspx -FORCE_INLINE __m128 _mm_cmplt_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_u32( - vcltq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -} - -// Compares for less than -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/fy94wye7(v=vs.100) -FORCE_INLINE __m128 _mm_cmplt_ss(__m128 a, __m128 b) -{ - return _mm_move_ss(a, _mm_cmplt_ps(a, b)); -} - -// Compares for greater than. -// -// r0 := (a0 > b0) ? 0xffffffff : 0x0 -// r1 := (a1 > b1) ? 0xffffffff : 0x0 -// r2 := (a2 > b2) ? 0xffffffff : 0x0 -// r3 := (a3 > b3) ? 0xffffffff : 0x0 -// -// https://msdn.microsoft.com/en-us/library/vstudio/11dy102s(v=vs.100).aspx -FORCE_INLINE __m128 _mm_cmpgt_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_u32( - vcgtq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -} - -// Compares for greater than. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/1xyyyy9e(v=vs.100) -FORCE_INLINE __m128 _mm_cmpgt_ss(__m128 a, __m128 b) -{ - return _mm_move_ss(a, _mm_cmpgt_ps(a, b)); -} - -// Compares for greater than or equal. -// https://msdn.microsoft.com/en-us/library/vstudio/fs813y2t(v=vs.100).aspx -FORCE_INLINE __m128 _mm_cmpge_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_u32( - vcgeq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -} - -// Compares for greater than or equal. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/kesh3ddc(v=vs.100) -FORCE_INLINE __m128 _mm_cmpge_ss(__m128 a, __m128 b) -{ - return _mm_move_ss(a, _mm_cmpge_ps(a, b)); -} - -// Compares for less than or equal. -// -// r0 := (a0 <= b0) ? 0xffffffff : 0x0 -// r1 := (a1 <= b1) ? 0xffffffff : 0x0 -// r2 := (a2 <= b2) ? 0xffffffff : 0x0 -// r3 := (a3 <= b3) ? 0xffffffff : 0x0 -// -// https://msdn.microsoft.com/en-us/library/vstudio/1s75w83z(v=vs.100).aspx -FORCE_INLINE __m128 _mm_cmple_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_u32( - vcleq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -} - -// Compares for less than or equal. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/a7x0hbhw(v=vs.100) -FORCE_INLINE __m128 _mm_cmple_ss(__m128 a, __m128 b) -{ - return _mm_move_ss(a, _mm_cmple_ps(a, b)); -} - -// Compares for equality. -// https://msdn.microsoft.com/en-us/library/vstudio/36aectz5(v=vs.100).aspx -FORCE_INLINE __m128 _mm_cmpeq_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_u32( - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -} - -// Compares for equality. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/k423z28e(v=vs.100) -FORCE_INLINE __m128 _mm_cmpeq_ss(__m128 a, __m128 b) -{ - return _mm_move_ss(a, _mm_cmpeq_ps(a, b)); -} - -// Compares for inequality. -// https://msdn.microsoft.com/en-us/library/sf44thbx(v=vs.100).aspx -FORCE_INLINE __m128 _mm_cmpneq_ps(__m128 a, __m128 b) -{ - return vreinterpretq_m128_u32(vmvnq_u32( - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)))); -} - -// Compares for inequality. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/ekya8fh4(v=vs.100) -FORCE_INLINE __m128 _mm_cmpneq_ss(__m128 a, __m128 b) -{ - return _mm_move_ss(a, _mm_cmpneq_ps(a, b)); -} - -// Compares for not greater than or equal. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/wsexys62(v=vs.100) -FORCE_INLINE __m128 _mm_cmpnge_ps(__m128 a, __m128 b) -{ - return _mm_cmplt_ps(a, b); -} - -// Compares for not greater than or equal. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/fk2y80s8(v=vs.100) -FORCE_INLINE __m128 _mm_cmpnge_ss(__m128 a, __m128 b) -{ - return _mm_cmplt_ss(a, b); -} - -// Compares for not greater than. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/d0xh7w0s(v=vs.100) -FORCE_INLINE __m128 _mm_cmpngt_ps(__m128 a, __m128 b) -{ - return _mm_cmple_ps(a, b); -} - -// Compares for not greater than. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/z7x9ydwh(v=vs.100) -FORCE_INLINE __m128 _mm_cmpngt_ss(__m128 a, __m128 b) -{ - return _mm_cmple_ss(a, b); -} - -// Compares for not less than or equal. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/6a330kxw(v=vs.100) -FORCE_INLINE __m128 _mm_cmpnle_ps(__m128 a, __m128 b) -{ - return _mm_cmpgt_ps(a, b); -} - -// Compares for not less than or equal. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/z7x9ydwh(v=vs.100) -FORCE_INLINE __m128 _mm_cmpnle_ss(__m128 a, __m128 b) -{ - return _mm_cmpgt_ss(a, b); -} - -// Compares for not less than. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/4686bbdw(v=vs.100) -FORCE_INLINE __m128 _mm_cmpnlt_ps(__m128 a, __m128 b) -{ - return _mm_cmpge_ps(a, b); -} - -// Compares for not less than. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/56b9z2wf(v=vs.100) -FORCE_INLINE __m128 _mm_cmpnlt_ss(__m128 a, __m128 b) -{ - return _mm_cmpge_ss(a, b); -} - -// Compares the 16 signed or unsigned 8-bit integers in a and the 16 signed or -// unsigned 8-bit integers in b for equality. -// https://msdn.microsoft.com/en-us/library/windows/desktop/bz5xk21a(v=vs.90).aspx -FORCE_INLINE __m128i _mm_cmpeq_epi8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u8( - vceqq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -} - -// Compare packed double-precision (64-bit) floating-point elements in a and b -// for equality, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpeq_pd -FORCE_INLINE __m128d _mm_cmpeq_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_u64( - vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - // (a == b) -> (a_lo == b_lo) && (a_hi == b_hi) - uint32x4_t cmp = - vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(b)); - uint32x4_t swapped = vrev64q_u32(cmp); - return vreinterpretq_m128d_u32(vandq_u32(cmp, swapped)); -#endif -} - -// Compare the lower double-precision (64-bit) floating-point elements in a and -// b for equality, store the result in the lower element of dst, and copy the -// upper element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpeq_sd -FORCE_INLINE __m128d _mm_cmpeq_sd(__m128d a, __m128d b) -{ - return _mm_move_sd(a, _mm_cmpeq_pd(a, b)); -} - -// Compare packed double-precision (64-bit) floating-point elements in a and b -// for greater-than-or-equal, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpge_pd -FORCE_INLINE __m128d _mm_cmpge_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_u64( - vcgeq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) >= (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); - d[1] = (*(double *) &a1) >= (*(double *) &b1) ? ~UINT64_C(0) : UINT64_C(0); - - return vreinterpretq_m128d_u64(vld1q_u64(d)); -#endif -} - -// Compare the lower double-precision (64-bit) floating-point elements in a and -// b for greater-than-or-equal, store the result in the lower element of dst, -// and copy the upper element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpge_sd -FORCE_INLINE __m128d _mm_cmpge_sd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return _mm_move_sd(a, _mm_cmpge_pd(a, b)); -#else - // expand "_mm_cmpge_pd()" to reduce unnecessary operations - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) >= (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); - d[1] = a1; - - return vreinterpretq_m128d_u64(vld1q_u64(d)); -#endif -} - -// Compares the 8 signed or unsigned 16-bit integers in a and the 8 signed or -// unsigned 16-bit integers in b for equality. -// https://msdn.microsoft.com/en-us/library/2ay060te(v=vs.100).aspx -FORCE_INLINE __m128i _mm_cmpeq_epi16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u16( - vceqq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -} - -// Compare packed 32-bit integers in a and b for equality, and store the results -// in dst -FORCE_INLINE __m128i _mm_cmpeq_epi32(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u32( - vceqq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -} - -// Compare packed 64-bit integers in a and b for equality, and store the results -// in dst -FORCE_INLINE __m128i _mm_cmpeq_epi64(__m128i a, __m128i b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128i_u64( - vceqq_u64(vreinterpretq_u64_m128i(a), vreinterpretq_u64_m128i(b))); -#else - // ARMv7 lacks vceqq_u64 - // (a == b) -> (a_lo == b_lo) && (a_hi == b_hi) - uint32x4_t cmp = - vceqq_u32(vreinterpretq_u32_m128i(a), vreinterpretq_u32_m128i(b)); - uint32x4_t swapped = vrev64q_u32(cmp); - return vreinterpretq_m128i_u32(vandq_u32(cmp, swapped)); -#endif -} - -// Compares the 16 signed 8-bit integers in a and the 16 signed 8-bit integers -// in b for lesser than. -// https://msdn.microsoft.com/en-us/library/windows/desktop/9s46csht(v=vs.90).aspx -FORCE_INLINE __m128i _mm_cmplt_epi8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u8( - vcltq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -} - -// Compare packed double-precision (64-bit) floating-point elements in a and b -// for less-than, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmplt_pd -FORCE_INLINE __m128d _mm_cmplt_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_u64( - vcltq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) < (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); - d[1] = (*(double *) &a1) < (*(double *) &b1) ? ~UINT64_C(0) : UINT64_C(0); - - return vreinterpretq_m128d_u64(vld1q_u64(d)); -#endif -} - -// Compare the lower double-precision (64-bit) floating-point elements in a and -// b for less-than, store the result in the lower element of dst, and copy the -// upper element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmplt_sd -FORCE_INLINE __m128d _mm_cmplt_sd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return _mm_move_sd(a, _mm_cmplt_pd(a, b)); -#else - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) < (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); - d[1] = a1; - - return vreinterpretq_m128d_u64(vld1q_u64(d)); -#endif -} - -// Compare packed double-precision (64-bit) floating-point elements in a and b -// for not-equal, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpneq_pd -FORCE_INLINE __m128d _mm_cmpneq_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_s32(vmvnq_s32(vreinterpretq_s32_u64( - vceqq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))))); -#else - // (a == b) -> (a_lo == b_lo) && (a_hi == b_hi) - uint32x4_t cmp = - vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(b)); - uint32x4_t swapped = vrev64q_u32(cmp); - return vreinterpretq_m128d_u32(vmvnq_u32(vandq_u32(cmp, swapped))); -#endif -} - -// Compare the lower double-precision (64-bit) floating-point elements in a and -// b for not-equal, store the result in the lower element of dst, and copy the -// upper element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpneq_sd -FORCE_INLINE __m128d _mm_cmpneq_sd(__m128d a, __m128d b) -{ - return _mm_move_sd(a, _mm_cmpneq_pd(a, b)); -} - -// Compare packed double-precision (64-bit) floating-point elements in a and b -// for not-greater-than-or-equal, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnge_pd -FORCE_INLINE __m128d _mm_cmpnge_pd(__m128d a, __m128d b) -{ - return _mm_cmplt_pd(a, b); -} - -// Compare the lower double-precision (64-bit) floating-point elements in a and -// b for not-greater-than-or-equal, store the result in the lower element of -// dst, and copy the upper element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpnge_sd -FORCE_INLINE __m128d _mm_cmpnge_sd(__m128d a, __m128d b) -{ - return _mm_cmplt_sd(a, b); -} - -// Compare the lower double-precision (64-bit) floating-point element in a and b -// for equality, and return the boolean result (0 or 1). -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_comieq_sd -FORCE_INLINE int _mm_comieq_sd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return !!vgetq_lane_u64(vceqq_f64(a, b), 0); -#else - uint32x4_t a_not_nan = - vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(a)); - uint32x4_t b_not_nan = - vceqq_u32(vreinterpretq_u32_m128d(b), vreinterpretq_u32_m128d(b)); - uint32x4_t a_and_b_not_nan = vandq_u32(a_not_nan, b_not_nan); - uint32x4_t a_eq_b = - vceqq_u32(vreinterpretq_u32_m128d(a), vreinterpretq_u32_m128d(b)); - uint64x2_t and_results = vandq_u64(vreinterpretq_u64_u32(a_and_b_not_nan), - vreinterpretq_u64_u32(a_eq_b)); - return !!vgetq_lane_u64(and_results, 0); -#endif -} - -// Compares the 16 signed 8-bit integers in a and the 16 signed 8-bit integers -// in b for greater than. -// -// r0 := (a0 > b0) ? 0xff : 0x0 -// r1 := (a1 > b1) ? 0xff : 0x0 -// ... -// r15 := (a15 > b15) ? 0xff : 0x0 -// -// https://msdn.microsoft.com/zh-tw/library/wf45zt2b(v=vs.100).aspx -FORCE_INLINE __m128i _mm_cmpgt_epi8(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u8( - vcgtq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -} - -// Compare packed double-precision (64-bit) floating-point elements in a and b -// for greater-than, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpgt_pd -FORCE_INLINE __m128d _mm_cmpgt_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_u64( - vcgtq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) > (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); - d[1] = (*(double *) &a1) > (*(double *) &b1) ? ~UINT64_C(0) : UINT64_C(0); - - return vreinterpretq_m128d_u64(vld1q_u64(d)); -#endif -} - -// Compare the lower double-precision (64-bit) floating-point elements in a and -// b for greater-than, store the result in the lower element of dst, and copy -// the upper element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmpgt_sd -FORCE_INLINE __m128d _mm_cmpgt_sd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return _mm_move_sd(a, _mm_cmpgt_pd(a, b)); -#else - // expand "_mm_cmpge_pd()" to reduce unnecessary operations - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) > (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); - d[1] = a1; - - return vreinterpretq_m128d_u64(vld1q_u64(d)); -#endif -} - -// Compare packed double-precision (64-bit) floating-point elements in a and b -// for less-than-or-equal, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmple_pd -FORCE_INLINE __m128d _mm_cmple_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_u64( - vcleq_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t b1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) <= (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); - d[1] = (*(double *) &a1) <= (*(double *) &b1) ? ~UINT64_C(0) : UINT64_C(0); - - return vreinterpretq_m128d_u64(vld1q_u64(d)); -#endif -} - -// Compare the lower double-precision (64-bit) floating-point elements in a and -// b for less-than-or-equal, store the result in the lower element of dst, and -// copy the upper element from a to the upper element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cmple_sd -FORCE_INLINE __m128d _mm_cmple_sd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return _mm_move_sd(a, _mm_cmple_pd(a, b)); -#else - // expand "_mm_cmpge_pd()" to reduce unnecessary operations - uint64_t a0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(a)); - uint64_t a1 = (uint64_t) vget_high_u64(vreinterpretq_u64_m128d(a)); - uint64_t b0 = (uint64_t) vget_low_u64(vreinterpretq_u64_m128d(b)); - uint64_t d[2]; - d[0] = (*(double *) &a0) <= (*(double *) &b0) ? ~UINT64_C(0) : UINT64_C(0); - d[1] = a1; - - return vreinterpretq_m128d_u64(vld1q_u64(d)); -#endif -} - -// Compares the 8 signed 16-bit integers in a and the 8 signed 16-bit integers -// in b for less than. -// -// r0 := (a0 < b0) ? 0xffff : 0x0 -// r1 := (a1 < b1) ? 0xffff : 0x0 -// ... -// r7 := (a7 < b7) ? 0xffff : 0x0 -// -// https://technet.microsoft.com/en-us/library/t863edb2(v=vs.100).aspx -FORCE_INLINE __m128i _mm_cmplt_epi16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u16( - vcltq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -} - -// Compares the 8 signed 16-bit integers in a and the 8 signed 16-bit integers -// in b for greater than. -// -// r0 := (a0 > b0) ? 0xffff : 0x0 -// r1 := (a1 > b1) ? 0xffff : 0x0 -// ... -// r7 := (a7 > b7) ? 0xffff : 0x0 -// -// https://technet.microsoft.com/en-us/library/xd43yfsa(v=vs.100).aspx -FORCE_INLINE __m128i _mm_cmpgt_epi16(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u16( - vcgtq_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -} - - -// Compares the 4 signed 32-bit integers in a and the 4 signed 32-bit integers -// in b for less than. -// https://msdn.microsoft.com/en-us/library/vstudio/4ak0bf5d(v=vs.100).aspx -FORCE_INLINE __m128i _mm_cmplt_epi32(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u32( - vcltq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -} - -// Compares the 4 signed 32-bit integers in a and the 4 signed 32-bit integers -// in b for greater than. -// https://msdn.microsoft.com/en-us/library/vstudio/1s9f2z0y(v=vs.100).aspx -FORCE_INLINE __m128i _mm_cmpgt_epi32(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_u32( - vcgtq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -} - -// Compares the 2 signed 64-bit integers in a and the 2 signed 64-bit integers -// in b for greater than. -FORCE_INLINE __m128i _mm_cmpgt_epi64(__m128i a, __m128i b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128i_u64( - vcgtq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b))); -#else - return vreinterpretq_m128i_s64(vshrq_n_s64( - vqsubq_s64(vreinterpretq_s64_m128i(b), vreinterpretq_s64_m128i(a)), - 63)); -#endif -} - -// Compares the four 32-bit floats in a and b to check if any values are NaN. -// Ordered compare between each value returns true for "orderable" and false for -// "not orderable" (NaN). -// https://msdn.microsoft.com/en-us/library/vstudio/0h9w00fx(v=vs.100).aspx see -// also: -// http://stackoverflow.com/questions/8627331/what-does-ordered-unordered-comparison-mean -// http://stackoverflow.com/questions/29349621/neon-isnanval-intrinsics -FORCE_INLINE __m128 _mm_cmpord_ps(__m128 a, __m128 b) -{ - // Note: NEON does not have ordered compare builtin - // Need to compare a eq a and b eq b to check for NaN - // Do AND of results to get final - uint32x4_t ceqaa = - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); - uint32x4_t ceqbb = - vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); - return vreinterpretq_m128_u32(vandq_u32(ceqaa, ceqbb)); -} - -// Compares for ordered. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/343t62da(v=vs.100) -FORCE_INLINE __m128 _mm_cmpord_ss(__m128 a, __m128 b) -{ - return _mm_move_ss(a, _mm_cmpord_ps(a, b)); -} - -// Compares for unordered. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/khy6fk1t(v=vs.100) -FORCE_INLINE __m128 _mm_cmpunord_ps(__m128 a, __m128 b) -{ - uint32x4_t f32a = - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); - uint32x4_t f32b = - vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); - return vreinterpretq_m128_u32(vmvnq_u32(vandq_u32(f32a, f32b))); -} - -// Compares for unordered. -// https://docs.microsoft.com/en-us/previous-versions/visualstudio/visual-studio-2010/2as2387b(v=vs.100) -FORCE_INLINE __m128 _mm_cmpunord_ss(__m128 a, __m128 b) -{ - return _mm_move_ss(a, _mm_cmpunord_ps(a, b)); -} - -// Compares the lower single-precision floating point scalar values of a and b -// using a less than operation. : -// https://msdn.microsoft.com/en-us/library/2kwe606b(v=vs.90).aspx Important -// note!! The documentation on MSDN is incorrect! If either of the values is a -// NAN the docs say you will get a one, but in fact, it will return a zero!! -FORCE_INLINE int _mm_comilt_ss(__m128 a, __m128 b) -{ - uint32x4_t a_not_nan = - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); - uint32x4_t b_not_nan = - vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); - uint32x4_t a_and_b_not_nan = vandq_u32(a_not_nan, b_not_nan); - uint32x4_t a_lt_b = - vcltq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); - return (vgetq_lane_u32(vandq_u32(a_and_b_not_nan, a_lt_b), 0) != 0) ? 1 : 0; -} - -// Compares the lower single-precision floating point scalar values of a and b -// using a greater than operation. : -// https://msdn.microsoft.com/en-us/library/b0738e0t(v=vs.100).aspx -FORCE_INLINE int _mm_comigt_ss(__m128 a, __m128 b) -{ - // return vgetq_lane_u32(vcgtq_f32(vreinterpretq_f32_m128(a), - // vreinterpretq_f32_m128(b)), 0); - uint32x4_t a_not_nan = - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); - uint32x4_t b_not_nan = - vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); - uint32x4_t a_and_b_not_nan = vandq_u32(a_not_nan, b_not_nan); - uint32x4_t a_gt_b = - vcgtq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); - return (vgetq_lane_u32(vandq_u32(a_and_b_not_nan, a_gt_b), 0) != 0) ? 1 : 0; -} - -// Compares the lower single-precision floating point scalar values of a and b -// using a less than or equal operation. : -// https://msdn.microsoft.com/en-us/library/1w4t7c57(v=vs.90).aspx -FORCE_INLINE int _mm_comile_ss(__m128 a, __m128 b) -{ - // return vgetq_lane_u32(vcleq_f32(vreinterpretq_f32_m128(a), - // vreinterpretq_f32_m128(b)), 0); - uint32x4_t a_not_nan = - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); - uint32x4_t b_not_nan = - vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); - uint32x4_t a_and_b_not_nan = vandq_u32(a_not_nan, b_not_nan); - uint32x4_t a_le_b = - vcleq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); - return (vgetq_lane_u32(vandq_u32(a_and_b_not_nan, a_le_b), 0) != 0) ? 1 : 0; -} - -// Compares the lower single-precision floating point scalar values of a and b -// using a greater than or equal operation. : -// https://msdn.microsoft.com/en-us/library/8t80des6(v=vs.100).aspx -FORCE_INLINE int _mm_comige_ss(__m128 a, __m128 b) -{ - // return vgetq_lane_u32(vcgeq_f32(vreinterpretq_f32_m128(a), - // vreinterpretq_f32_m128(b)), 0); - uint32x4_t a_not_nan = - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); - uint32x4_t b_not_nan = - vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); - uint32x4_t a_and_b_not_nan = vandq_u32(a_not_nan, b_not_nan); - uint32x4_t a_ge_b = - vcgeq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); - return (vgetq_lane_u32(vandq_u32(a_and_b_not_nan, a_ge_b), 0) != 0) ? 1 : 0; -} - -// Compares the lower single-precision floating point scalar values of a and b -// using an equality operation. : -// https://msdn.microsoft.com/en-us/library/93yx2h2b(v=vs.100).aspx -FORCE_INLINE int _mm_comieq_ss(__m128 a, __m128 b) -{ - // return vgetq_lane_u32(vceqq_f32(vreinterpretq_f32_m128(a), - // vreinterpretq_f32_m128(b)), 0); - uint32x4_t a_not_nan = - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); - uint32x4_t b_not_nan = - vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); - uint32x4_t a_and_b_not_nan = vandq_u32(a_not_nan, b_not_nan); - uint32x4_t a_eq_b = - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b)); - return (vgetq_lane_u32(vandq_u32(a_and_b_not_nan, a_eq_b), 0) != 0) ? 1 : 0; -} - -// Compares the lower single-precision floating point scalar values of a and b -// using an inequality operation. : -// https://msdn.microsoft.com/en-us/library/bafh5e0a(v=vs.90).aspx -FORCE_INLINE int _mm_comineq_ss(__m128 a, __m128 b) -{ - // return !vgetq_lane_u32(vceqq_f32(vreinterpretq_f32_m128(a), - // vreinterpretq_f32_m128(b)), 0); - uint32x4_t a_not_nan = - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(a)); - uint32x4_t b_not_nan = - vceqq_f32(vreinterpretq_f32_m128(b), vreinterpretq_f32_m128(b)); - uint32x4_t a_or_b_nan = vmvnq_u32(vandq_u32(a_not_nan, b_not_nan)); - uint32x4_t a_neq_b = vmvnq_u32( - vceqq_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); - return (vgetq_lane_u32(vorrq_u32(a_or_b_nan, a_neq_b), 0) != 0) ? 1 : 0; -} - -// according to the documentation, these intrinsics behave the same as the -// non-'u' versions. We'll just alias them here. -#define _mm_ucomieq_ss _mm_comieq_ss -#define _mm_ucomige_ss _mm_comige_ss -#define _mm_ucomigt_ss _mm_comigt_ss -#define _mm_ucomile_ss _mm_comile_ss -#define _mm_ucomilt_ss _mm_comilt_ss -#define _mm_ucomineq_ss _mm_comineq_ss - -/* Conversions */ - -// Convert packed signed 32-bit integers in b to packed single-precision -// (32-bit) floating-point elements, store the results in the lower 2 elements -// of dst, and copy the upper 2 packed elements from a to the upper elements of -// dst. -// -// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) -// dst[63:32] := Convert_Int32_To_FP32(b[63:32]) -// dst[95:64] := a[95:64] -// dst[127:96] := a[127:96] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_pi2ps -FORCE_INLINE __m128 _mm_cvt_pi2ps(__m128 a, __m64 b) -{ - return vreinterpretq_m128_f32( - vcombine_f32(vcvt_f32_s32(vreinterpret_s32_m64(b)), - vget_high_f32(vreinterpretq_f32_m128(a)))); -} - -// Convert the signed 32-bit integer b to a single-precision (32-bit) -// floating-point element, store the result in the lower element of dst, and -// copy the upper 3 packed elements from a to the upper elements of dst. -// -// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) -// dst[127:32] := a[127:32] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_si2ss -FORCE_INLINE __m128 _mm_cvt_si2ss(__m128 a, int b) -{ - return vreinterpretq_m128_f32( - vsetq_lane_f32((float) b, vreinterpretq_f32_m128(a), 0)); -} - -// Convert the signed 32-bit integer b to a single-precision (32-bit) -// floating-point element, store the result in the lower element of dst, and -// copy the upper 3 packed elements from a to the upper elements of dst. -// -// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) -// dst[127:32] := a[127:32] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi32_ss -#define _mm_cvtsi32_ss(a, b) _mm_cvt_si2ss(a, b) - -// Convert the signed 64-bit integer b to a single-precision (32-bit) -// floating-point element, store the result in the lower element of dst, and -// copy the upper 3 packed elements from a to the upper elements of dst. -// -// dst[31:0] := Convert_Int64_To_FP32(b[63:0]) -// dst[127:32] := a[127:32] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi64_ss -FORCE_INLINE __m128 _mm_cvtsi64_ss(__m128 a, int64_t b) -{ - return vreinterpretq_m128_f32( - vsetq_lane_f32((float) b, vreinterpretq_f32_m128(a), 0)); -} - -// Convert the lower single-precision (32-bit) floating-point element in a to a -// 32-bit integer, and store the result in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_ss2si -FORCE_INLINE int _mm_cvt_ss2si(__m128 a) -{ -#if defined(__aarch64__) - return vgetq_lane_s32(vcvtnq_s32_f32(vreinterpretq_f32_m128(a)), 0); -#else - float32_t data = vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); - float32_t diff = data - floor(data); - if (diff > 0.5) - return (int32_t) ceil(data); - if (unlikely(diff == 0.5)) { - int32_t f = (int32_t) floor(data); - int32_t c = (int32_t) ceil(data); - return c & 1 ? f : c; - } - return (int32_t) floor(data); -#endif -} - -// Convert packed 16-bit integers in a to packed single-precision (32-bit) -// floating-point elements, and store the results in dst. -// -// FOR j := 0 to 3 -// i := j*16 -// m := j*32 -// dst[m+31:m] := Convert_Int16_To_FP32(a[i+15:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi16_ps -FORCE_INLINE __m128 _mm_cvtpi16_ps(__m64 a) -{ - return vreinterpretq_m128_f32( - vcvtq_f32_s32(vmovl_s16(vreinterpret_s16_m64(a)))); -} - -// Convert packed 32-bit integers in b to packed single-precision (32-bit) -// floating-point elements, store the results in the lower 2 elements of dst, -// and copy the upper 2 packed elements from a to the upper elements of dst. -// -// dst[31:0] := Convert_Int32_To_FP32(b[31:0]) -// dst[63:32] := Convert_Int32_To_FP32(b[63:32]) -// dst[95:64] := a[95:64] -// dst[127:96] := a[127:96] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi32_ps -FORCE_INLINE __m128 _mm_cvtpi32_ps(__m128 a, __m64 b) -{ - return vreinterpretq_m128_f32( - vcombine_f32(vcvt_f32_s32(vreinterpret_s32_m64(b)), - vget_high_f32(vreinterpretq_f32_m128(a)))); -} - -// Convert packed signed 32-bit integers in a to packed single-precision -// (32-bit) floating-point elements, store the results in the lower 2 elements -// of dst, then covert the packed signed 32-bit integers in b to -// single-precision (32-bit) floating-point element, and store the results in -// the upper 2 elements of dst. -// -// dst[31:0] := Convert_Int32_To_FP32(a[31:0]) -// dst[63:32] := Convert_Int32_To_FP32(a[63:32]) -// dst[95:64] := Convert_Int32_To_FP32(b[31:0]) -// dst[127:96] := Convert_Int32_To_FP32(b[63:32]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi32x2_ps -FORCE_INLINE __m128 _mm_cvtpi32x2_ps(__m64 a, __m64 b) -{ - return vreinterpretq_m128_f32(vcvtq_f32_s32( - vcombine_s32(vreinterpret_s32_m64(a), vreinterpret_s32_m64(b)))); -} - -// Convert the lower packed 8-bit integers in a to packed single-precision -// (32-bit) floating-point elements, and store the results in dst. -// -// FOR j := 0 to 3 -// i := j*8 -// m := j*32 -// dst[m+31:m] := Convert_Int8_To_FP32(a[i+7:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi8_ps -FORCE_INLINE __m128 _mm_cvtpi8_ps(__m64 a) -{ - return vreinterpretq_m128_f32(vcvtq_f32_s32( - vmovl_s16(vget_low_s16(vmovl_s8(vreinterpret_s8_m64(a)))))); -} - -// Convert packed unsigned 16-bit integers in a to packed single-precision -// (32-bit) floating-point elements, and store the results in dst. -// -// FOR j := 0 to 3 -// i := j*16 -// m := j*32 -// dst[m+31:m] := Convert_UInt16_To_FP32(a[i+15:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpu16_ps -FORCE_INLINE __m128 _mm_cvtpu16_ps(__m64 a) -{ - return vreinterpretq_m128_f32( - vcvtq_f32_u32(vmovl_u16(vreinterpret_u16_m64(a)))); -} - -// Convert the lower packed unsigned 8-bit integers in a to packed -// single-precision (32-bit) floating-point elements, and store the results in -// dst. -// -// FOR j := 0 to 3 -// i := j*8 -// m := j*32 -// dst[m+31:m] := Convert_UInt8_To_FP32(a[i+7:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpu8_ps -FORCE_INLINE __m128 _mm_cvtpu8_ps(__m64 a) -{ - return vreinterpretq_m128_f32(vcvtq_f32_u32( - vmovl_u16(vget_low_u16(vmovl_u8(vreinterpret_u8_m64(a)))))); -} - -// Converts the four single-precision, floating-point values of a to signed -// 32-bit integer values using truncate. -// https://msdn.microsoft.com/en-us/library/vstudio/1h005y6x(v=vs.100).aspx -FORCE_INLINE __m128i _mm_cvttps_epi32(__m128 a) -{ - return vreinterpretq_m128i_s32(vcvtq_s32_f32(vreinterpretq_f32_m128(a))); -} - -// Convert the lower double-precision (64-bit) floating-point element in a to a -// 64-bit integer with truncation, and store the result in dst. -// -// dst[63:0] := Convert_FP64_To_Int64_Truncate(a[63:0]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttsd_si64 -FORCE_INLINE int64_t _mm_cvttsd_si64(__m128d a) -{ -#if defined(__aarch64__) - return vgetq_lane_s64(vcvtq_s64_f64(vreinterpretq_f64_m128d(a)), 0); -#else - double ret = *((double *) &a); - return (int64_t) ret; -#endif -} - -// Convert the lower double-precision (64-bit) floating-point element in a to a -// 64-bit integer with truncation, and store the result in dst. -// -// dst[63:0] := Convert_FP64_To_Int64_Truncate(a[63:0]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvttsd_si64x -#define _mm_cvttsd_si64x(a) _mm_cvttsd_si64(a) - -// Converts the four signed 32-bit integer values of a to single-precision, -// floating-point values -// https://msdn.microsoft.com/en-us/library/vstudio/36bwxcx5(v=vs.100).aspx -FORCE_INLINE __m128 _mm_cvtepi32_ps(__m128i a) -{ - return vreinterpretq_m128_f32(vcvtq_f32_s32(vreinterpretq_s32_m128i(a))); -} - -// Convert packed signed 32-bit integers in a to packed double-precision -// (64-bit) floating-point elements, and store the results in dst. -// -// FOR j := 0 to 1 -// i := j*32 -// m := j*64 -// dst[m+63:m] := Convert_Int32_To_FP64(a[i+31:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtepi32_pd -FORCE_INLINE __m128d _mm_cvtepi32_pd(__m128i a) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vcvtq_f64_s64(vmovl_s32(vget_low_s32(vreinterpretq_s32_m128i(a))))); -#else - double a0 = (double) vgetq_lane_s32(vreinterpretq_s32_m128i(a), 0); - double a1 = (double) vgetq_lane_s32(vreinterpretq_s32_m128i(a), 1); - return _mm_set_pd(a1, a0); -#endif -} - -// Convert packed signed 32-bit integers in a to packed double-precision -// (64-bit) floating-point elements, and store the results in dst. -// -// FOR j := 0 to 1 -// i := j*32 -// m := j*64 -// dst[m+63:m] := Convert_Int32_To_FP64(a[i+31:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpi32_pd -FORCE_INLINE __m128d _mm_cvtpi32_pd(__m64 a) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vcvtq_f64_s64(vmovl_s32(vreinterpret_s32_m64(a)))); -#else - double a0 = (double) vget_lane_s32(vreinterpret_s32_m64(a), 0); - double a1 = (double) vget_lane_s32(vreinterpret_s32_m64(a), 1); - return _mm_set_pd(a1, a0); -#endif -} - -// Converts the four unsigned 8-bit integers in the lower 16 bits to four -// unsigned 32-bit integers. -FORCE_INLINE __m128i _mm_cvtepu8_epi16(__m128i a) -{ - uint8x16_t u8x16 = vreinterpretq_u8_m128i(a); /* xxxx xxxx xxxx DCBA */ - uint16x8_t u16x8 = vmovl_u8(vget_low_u8(u8x16)); /* 0x0x 0x0x 0D0C 0B0A */ - return vreinterpretq_m128i_u16(u16x8); -} - -// Converts the four unsigned 8-bit integers in the lower 32 bits to four -// unsigned 32-bit integers. -// https://msdn.microsoft.com/en-us/library/bb531467%28v=vs.100%29.aspx -FORCE_INLINE __m128i _mm_cvtepu8_epi32(__m128i a) -{ - uint8x16_t u8x16 = vreinterpretq_u8_m128i(a); /* xxxx xxxx xxxx DCBA */ - uint16x8_t u16x8 = vmovl_u8(vget_low_u8(u8x16)); /* 0x0x 0x0x 0D0C 0B0A */ - uint32x4_t u32x4 = vmovl_u16(vget_low_u16(u16x8)); /* 000D 000C 000B 000A */ - return vreinterpretq_m128i_u32(u32x4); -} - -// Converts the two unsigned 8-bit integers in the lower 16 bits to two -// unsigned 64-bit integers. -FORCE_INLINE __m128i _mm_cvtepu8_epi64(__m128i a) -{ - uint8x16_t u8x16 = vreinterpretq_u8_m128i(a); /* xxxx xxxx xxxx xxBA */ - uint16x8_t u16x8 = vmovl_u8(vget_low_u8(u8x16)); /* 0x0x 0x0x 0x0x 0B0A */ - uint32x4_t u32x4 = vmovl_u16(vget_low_u16(u16x8)); /* 000x 000x 000B 000A */ - uint64x2_t u64x2 = vmovl_u32(vget_low_u32(u32x4)); /* 0000 000B 0000 000A */ - return vreinterpretq_m128i_u64(u64x2); -} - -// Converts the four unsigned 8-bit integers in the lower 16 bits to four -// unsigned 32-bit integers. -FORCE_INLINE __m128i _mm_cvtepi8_epi16(__m128i a) -{ - int8x16_t s8x16 = vreinterpretq_s8_m128i(a); /* xxxx xxxx xxxx DCBA */ - int16x8_t s16x8 = vmovl_s8(vget_low_s8(s8x16)); /* 0x0x 0x0x 0D0C 0B0A */ - return vreinterpretq_m128i_s16(s16x8); -} - -// Converts the four unsigned 8-bit integers in the lower 32 bits to four -// unsigned 32-bit integers. -FORCE_INLINE __m128i _mm_cvtepi8_epi32(__m128i a) -{ - int8x16_t s8x16 = vreinterpretq_s8_m128i(a); /* xxxx xxxx xxxx DCBA */ - int16x8_t s16x8 = vmovl_s8(vget_low_s8(s8x16)); /* 0x0x 0x0x 0D0C 0B0A */ - int32x4_t s32x4 = vmovl_s16(vget_low_s16(s16x8)); /* 000D 000C 000B 000A */ - return vreinterpretq_m128i_s32(s32x4); -} - -// Converts the two signed 8-bit integers in the lower 32 bits to four -// signed 64-bit integers. -FORCE_INLINE __m128i _mm_cvtepi8_epi64(__m128i a) -{ - int8x16_t s8x16 = vreinterpretq_s8_m128i(a); /* xxxx xxxx xxxx xxBA */ - int16x8_t s16x8 = vmovl_s8(vget_low_s8(s8x16)); /* 0x0x 0x0x 0x0x 0B0A */ - int32x4_t s32x4 = vmovl_s16(vget_low_s16(s16x8)); /* 000x 000x 000B 000A */ - int64x2_t s64x2 = vmovl_s32(vget_low_s32(s32x4)); /* 0000 000B 0000 000A */ - return vreinterpretq_m128i_s64(s64x2); -} - -// Converts the four signed 16-bit integers in the lower 64 bits to four signed -// 32-bit integers. -FORCE_INLINE __m128i _mm_cvtepi16_epi32(__m128i a) -{ - return vreinterpretq_m128i_s32( - vmovl_s16(vget_low_s16(vreinterpretq_s16_m128i(a)))); -} - -// Converts the two signed 16-bit integers in the lower 32 bits two signed -// 32-bit integers. -FORCE_INLINE __m128i _mm_cvtepi16_epi64(__m128i a) -{ - int16x8_t s16x8 = vreinterpretq_s16_m128i(a); /* xxxx xxxx xxxx 0B0A */ - int32x4_t s32x4 = vmovl_s16(vget_low_s16(s16x8)); /* 000x 000x 000B 000A */ - int64x2_t s64x2 = vmovl_s32(vget_low_s32(s32x4)); /* 0000 000B 0000 000A */ - return vreinterpretq_m128i_s64(s64x2); -} - -// Converts the four unsigned 16-bit integers in the lower 64 bits to four -// unsigned 32-bit integers. -FORCE_INLINE __m128i _mm_cvtepu16_epi32(__m128i a) -{ - return vreinterpretq_m128i_u32( - vmovl_u16(vget_low_u16(vreinterpretq_u16_m128i(a)))); -} - -// Converts the two unsigned 16-bit integers in the lower 32 bits to two -// unsigned 64-bit integers. -FORCE_INLINE __m128i _mm_cvtepu16_epi64(__m128i a) -{ - uint16x8_t u16x8 = vreinterpretq_u16_m128i(a); /* xxxx xxxx xxxx 0B0A */ - uint32x4_t u32x4 = vmovl_u16(vget_low_u16(u16x8)); /* 000x 000x 000B 000A */ - uint64x2_t u64x2 = vmovl_u32(vget_low_u32(u32x4)); /* 0000 000B 0000 000A */ - return vreinterpretq_m128i_u64(u64x2); -} - -// Converts the two unsigned 32-bit integers in the lower 64 bits to two -// unsigned 64-bit integers. -FORCE_INLINE __m128i _mm_cvtepu32_epi64(__m128i a) -{ - return vreinterpretq_m128i_u64( - vmovl_u32(vget_low_u32(vreinterpretq_u32_m128i(a)))); -} - -// Converts the two signed 32-bit integers in the lower 64 bits to two signed -// 64-bit integers. -FORCE_INLINE __m128i _mm_cvtepi32_epi64(__m128i a) -{ - return vreinterpretq_m128i_s64( - vmovl_s32(vget_low_s32(vreinterpretq_s32_m128i(a)))); -} - -// Converts the four single-precision, floating-point values of a to signed -// 32-bit integer values. -// -// r0 := (int) a0 -// r1 := (int) a1 -// r2 := (int) a2 -// r3 := (int) a3 -// -// https://msdn.microsoft.com/en-us/library/vstudio/xdc42k5e(v=vs.100).aspx -// *NOTE*. The default rounding mode on SSE is 'round to even', which ARMv7-A -// does not support! It is supported on ARMv8-A however. -FORCE_INLINE __m128i _mm_cvtps_epi32(__m128 a) -{ -#if defined(__aarch64__) - return vreinterpretq_m128i_s32(vcvtnq_s32_f32(a)); -#else - uint32x4_t signmask = vdupq_n_u32(0x80000000); - float32x4_t half = vbslq_f32(signmask, vreinterpretq_f32_m128(a), - vdupq_n_f32(0.5f)); /* +/- 0.5 */ - int32x4_t r_normal = vcvtq_s32_f32(vaddq_f32( - vreinterpretq_f32_m128(a), half)); /* round to integer: [a + 0.5]*/ - int32x4_t r_trunc = - vcvtq_s32_f32(vreinterpretq_f32_m128(a)); /* truncate to integer: [a] */ - int32x4_t plusone = vreinterpretq_s32_u32(vshrq_n_u32( - vreinterpretq_u32_s32(vnegq_s32(r_trunc)), 31)); /* 1 or 0 */ - int32x4_t r_even = vbicq_s32(vaddq_s32(r_trunc, plusone), - vdupq_n_s32(1)); /* ([a] + {0,1}) & ~1 */ - float32x4_t delta = vsubq_f32( - vreinterpretq_f32_m128(a), - vcvtq_f32_s32(r_trunc)); /* compute delta: delta = (a - [a]) */ - uint32x4_t is_delta_half = vceqq_f32(delta, half); /* delta == +/- 0.5 */ - return vreinterpretq_m128i_s32(vbslq_s32(is_delta_half, r_even, r_normal)); -#endif -} - -// Convert packed single-precision (32-bit) floating-point elements in a to -// packed 16-bit integers, and store the results in dst. Note: this intrinsic -// will generate 0x7FFF, rather than 0x8000, for input values between 0x7FFF and -// 0x7FFFFFFF. -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pi16 -FORCE_INLINE __m64 _mm_cvtps_pi16(__m128 a) -{ - return vreinterpret_m64_s16( - vmovn_s32(vreinterpretq_s32_m128i(_mm_cvtps_epi32(a)))); -} - -// Copy the lower 32-bit integer in a to dst. -// -// dst[31:0] := a[31:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si32 -FORCE_INLINE int _mm_cvtsi128_si32(__m128i a) -{ - return vgetq_lane_s32(vreinterpretq_s32_m128i(a), 0); -} - -// Copy the lower 64-bit integer in a to dst. -// -// dst[63:0] := a[63:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si64 -FORCE_INLINE int64_t _mm_cvtsi128_si64(__m128i a) -{ - return vgetq_lane_s64(vreinterpretq_s64_m128i(a), 0); -} - -// Copy the lower 64-bit integer in a to dst. -// -// dst[63:0] := a[63:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsi128_si64x -#define _mm_cvtsi128_si64x(a) _mm_cvtsi128_si64(a) - -// Moves 32-bit integer a to the least significant 32 bits of an __m128 object, -// zero extending the upper bits. -// -// r0 := a -// r1 := 0x0 -// r2 := 0x0 -// r3 := 0x0 -// -// https://msdn.microsoft.com/en-us/library/ct3539ha%28v=vs.90%29.aspx -FORCE_INLINE __m128i _mm_cvtsi32_si128(int a) -{ - return vreinterpretq_m128i_s32(vsetq_lane_s32(a, vdupq_n_s32(0), 0)); -} - -// Moves 64-bit integer a to the least significant 64 bits of an __m128 object, -// zero extending the upper bits. -// -// r0 := a -// r1 := 0x0 -FORCE_INLINE __m128i _mm_cvtsi64_si128(int64_t a) -{ - return vreinterpretq_m128i_s64(vsetq_lane_s64(a, vdupq_n_s64(0), 0)); -} - -// Cast vector of type __m128 to type __m128d. This intrinsic is only used for -// compilation and does not generate any instructions, thus it has zero latency. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castps_pd -FORCE_INLINE __m128d _mm_castps_pd(__m128 a) -{ - return vreinterpretq_m128d_s32(vreinterpretq_s32_m128(a)); -} - -// Applies a type cast to reinterpret four 32-bit floating point values passed -// in as a 128-bit parameter as packed 32-bit integers. -// https://msdn.microsoft.com/en-us/library/bb514099.aspx -FORCE_INLINE __m128i _mm_castps_si128(__m128 a) -{ - return vreinterpretq_m128i_s32(vreinterpretq_s32_m128(a)); -} - -// Cast vector of type __m128i to type __m128d. This intrinsic is only used for -// compilation and does not generate any instructions, thus it has zero latency. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castsi128_pd -FORCE_INLINE __m128d _mm_castsi128_pd(__m128i a) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64(vreinterpretq_f64_m128i(a)); -#else - return vreinterpretq_m128d_f32(vreinterpretq_f32_m128i(a)); -#endif -} - -// Applies a type cast to reinterpret four 32-bit integers passed in as a -// 128-bit parameter as packed 32-bit floating point values. -// https://msdn.microsoft.com/en-us/library/bb514029.aspx -FORCE_INLINE __m128 _mm_castsi128_ps(__m128i a) -{ - return vreinterpretq_m128_s32(vreinterpretq_s32_m128i(a)); -} - -// Loads 128-bit value. : -// https://msdn.microsoft.com/en-us/library/atzzad1h(v=vs.80).aspx -FORCE_INLINE __m128i _mm_load_si128(const __m128i *p) -{ - return vreinterpretq_m128i_s32(vld1q_s32((const int32_t *) p)); -} - -// Load a double-precision (64-bit) floating-point element from memory into both -// elements of dst. -// -// dst[63:0] := MEM[mem_addr+63:mem_addr] -// dst[127:64] := MEM[mem_addr+63:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load1_pd -FORCE_INLINE __m128d _mm_load1_pd(const double *p) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64(vld1q_dup_f64(p)); -#else - return vreinterpretq_m128d_s64(vdupq_n_s64(*(const int64_t *) p)); -#endif -} - -// Load a double-precision (64-bit) floating-point element from memory into both -// elements of dst. -// -// dst[63:0] := MEM[mem_addr+63:mem_addr] -// dst[127:64] := MEM[mem_addr+63:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_pd1 -#define _mm_load_pd1 _mm_load1_pd - -// Load a double-precision (64-bit) floating-point element from memory into both -// elements of dst. -// -// dst[63:0] := MEM[mem_addr+63:mem_addr] -// dst[127:64] := MEM[mem_addr+63:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loaddup_pd -#define _mm_loaddup_pd _mm_load1_pd - -// Load a double-precision (64-bit) floating-point element from memory into the -// upper element of dst, and copy the lower element from a to dst. mem_addr does -// not need to be aligned on any particular boundary. -// -// dst[63:0] := a[63:0] -// dst[127:64] := MEM[mem_addr+63:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadh_pd -FORCE_INLINE __m128d _mm_loadh_pd(__m128d a, const double *p) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vcombine_f64(vget_low_f64(vreinterpretq_f64_m128d(a)), vld1_f64(p))); -#else - return vreinterpretq_m128d_f32(vcombine_f32( - vget_low_f32(vreinterpretq_f32_m128d(a)), vld1_f32((const float *) p))); -#endif -} - -// Load a double-precision (64-bit) floating-point element from memory into both -// elements of dst. -// -// dst[63:0] := MEM[mem_addr+63:mem_addr] -// dst[127:64] := MEM[mem_addr+63:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_load_pd1 -#define _mm_load_pd1 _mm_load1_pd - -// Load a double-precision (64-bit) floating-point element from memory into both -// elements of dst. -// -// dst[63:0] := MEM[mem_addr+63:mem_addr] -// dst[127:64] := MEM[mem_addr+63:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loaddup_pd -#define _mm_loaddup_pd _mm_load1_pd - -// Loads 128-bit value. : -// https://msdn.microsoft.com/zh-cn/library/f4k12ae8(v=vs.90).aspx -FORCE_INLINE __m128i _mm_loadu_si128(const __m128i *p) -{ - return vreinterpretq_m128i_s32(vld1q_s32((const int32_t *) p)); -} - -// Load unaligned 32-bit integer from memory into the first element of dst. -// -// dst[31:0] := MEM[mem_addr+31:mem_addr] -// dst[MAX:32] := 0 -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_loadu_si32 -FORCE_INLINE __m128i _mm_loadu_si32(const void *p) -{ - return vreinterpretq_m128i_s32( - vsetq_lane_s32(*(const int32_t *) p, vdupq_n_s32(0), 0)); -} - -// Convert packed double-precision (64-bit) floating-point elements in a to -// packed single-precision (32-bit) floating-point elements, and store the -// results in dst. -// -// FOR j := 0 to 1 -// i := 32*j -// k := 64*j -// dst[i+31:i] := Convert_FP64_To_FP32(a[k+64:k]) -// ENDFOR -// dst[127:64] := 0 -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtpd_ps -FORCE_INLINE __m128 _mm_cvtpd_ps(__m128d a) -{ -#if defined(__aarch64__) - float32x2_t tmp = vcvt_f32_f64(vreinterpretq_f64_m128d(a)); - return vreinterpretq_m128_f32(vcombine_f32(tmp, vdup_n_f32(0))); -#else - float a0 = (float) ((double *) &a)[0]; - float a1 = (float) ((double *) &a)[1]; - return _mm_set_ps(0, 0, a1, a0); -#endif -} - -// Copy the lower double-precision (64-bit) floating-point element of a to dst. -// -// dst[63:0] := a[63:0] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtsd_f64 -FORCE_INLINE double _mm_cvtsd_f64(__m128d a) -{ -#if defined(__aarch64__) - return (double) vgetq_lane_f64(vreinterpretq_f64_m128d(a), 0); -#else - return ((double *) &a)[0]; -#endif -} - -// Convert packed single-precision (32-bit) floating-point elements in a to -// packed double-precision (64-bit) floating-point elements, and store the -// results in dst. -// -// FOR j := 0 to 1 -// i := 64*j -// k := 32*j -// dst[i+63:i] := Convert_FP32_To_FP64(a[k+31:k]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pd -FORCE_INLINE __m128d _mm_cvtps_pd(__m128 a) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vcvt_f64_f32(vget_low_f32(vreinterpretq_f32_m128(a)))); -#else - double a0 = (double) vgetq_lane_f32(vreinterpretq_f32_m128(a), 0); - double a1 = (double) vgetq_lane_f32(vreinterpretq_f32_m128(a), 1); - return _mm_set_pd(a1, a0); -#endif -} - -// Cast vector of type __m128d to type __m128i. This intrinsic is only used for -// compilation and does not generate any instructions, thus it has zero latency. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castpd_si128 -FORCE_INLINE __m128i _mm_castpd_si128(__m128d a) -{ - return vreinterpretq_m128i_s64(vreinterpretq_s64_m128d(a)); -} - -// Cast vector of type __m128d to type __m128. This intrinsic is only used for -// compilation and does not generate any instructions, thus it has zero latency. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_castpd_ps -FORCE_INLINE __m128 _mm_castpd_ps(__m128d a) -{ - return vreinterpretq_m128_s64(vreinterpretq_s64_m128d(a)); -} - -// Blend packed single-precision (32-bit) floating-point elements from a and b -// using mask, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blendv_ps -FORCE_INLINE __m128 _mm_blendv_ps(__m128 _a, __m128 _b, __m128 _mask) -{ - // Use a signed shift right to create a mask with the sign bit - uint32x4_t mask = - vreinterpretq_u32_s32(vshrq_n_s32(vreinterpretq_s32_m128(_mask), 31)); - float32x4_t a = vreinterpretq_f32_m128(_a); - float32x4_t b = vreinterpretq_f32_m128(_b); - return vreinterpretq_m128_f32(vbslq_f32(mask, b, a)); -} - -// Blend packed single-precision (32-bit) floating-point elements from a and b -// using mask, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blend_ps -FORCE_INLINE __m128 _mm_blend_ps(__m128 _a, __m128 _b, const char imm8) -{ - const uint32_t ALIGN_STRUCT(16) - data[4] = {((imm8) & (1 << 0)) ? UINT32_MAX : 0, - ((imm8) & (1 << 1)) ? UINT32_MAX : 0, - ((imm8) & (1 << 2)) ? UINT32_MAX : 0, - ((imm8) & (1 << 3)) ? UINT32_MAX : 0}; - uint32x4_t mask = vld1q_u32(data); - float32x4_t a = vreinterpretq_f32_m128(_a); - float32x4_t b = vreinterpretq_f32_m128(_b); - return vreinterpretq_m128_f32(vbslq_f32(mask, b, a)); -} - -// Blend packed double-precision (64-bit) floating-point elements from a and b -// using mask, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_blendv_pd -FORCE_INLINE __m128d _mm_blendv_pd(__m128d _a, __m128d _b, __m128d _mask) -{ - uint64x2_t mask = - vreinterpretq_u64_s64(vshrq_n_s64(vreinterpretq_s64_m128d(_mask), 63)); -#if defined(__aarch64__) - float64x2_t a = vreinterpretq_f64_m128d(_a); - float64x2_t b = vreinterpretq_f64_m128d(_b); - return vreinterpretq_m128d_f64(vbslq_f64(mask, b, a)); -#else - uint64x2_t a = vreinterpretq_u64_m128d(_a); - uint64x2_t b = vreinterpretq_u64_m128d(_b); - return vreinterpretq_m128d_u64(vbslq_u64(mask, b, a)); -#endif -} - -typedef struct { - uint16_t res0; - uint8_t res1 : 6; - uint8_t bit22 : 1; - uint8_t bit23 : 1; - uint8_t res2; -#if defined(__aarch64__) - uint32_t res3; -#endif -} fpcr_bitfield; - -// Macro: Set the rounding mode bits of the MXCSR control and status register to -// the value in unsigned 32-bit integer a. The rounding mode may contain any of -// the following flags: _MM_ROUND_NEAREST, _MM_ROUND_DOWN, _MM_ROUND_UP, -// _MM_ROUND_TOWARD_ZERO -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_MM_SET_ROUNDING_MODE -FORCE_INLINE void _MM_SET_ROUNDING_MODE(int rounding) -{ - union { - fpcr_bitfield field; -#if defined(__aarch64__) - uint64_t value; -#else - uint32_t value; -#endif - } r; - -#if defined(__aarch64__) - asm volatile("mrs %0, FPCR" : "=r"(r.value)); /* read */ -#else - asm volatile("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ -#endif - - switch (rounding) { - case _MM_ROUND_TOWARD_ZERO: - r.field.bit22 = 1; - r.field.bit23 = 1; - break; - case _MM_ROUND_DOWN: - r.field.bit22 = 0; - r.field.bit23 = 1; - break; - case _MM_ROUND_UP: - r.field.bit22 = 1; - r.field.bit23 = 0; - break; - default: //_MM_ROUND_NEAREST - r.field.bit22 = 0; - r.field.bit23 = 0; - } - -#if defined(__aarch64__) - asm volatile("msr FPCR, %0" ::"r"(r)); /* write */ -#else - asm volatile("vmsr FPSCR, %0" ::"r"(r)); /* write */ -#endif -} - -FORCE_INLINE void _mm_setcsr(unsigned int a) -{ - _MM_SET_ROUNDING_MODE(a); -} - -// Round the packed single-precision (32-bit) floating-point elements in a using -// the rounding parameter, and store the results as packed single-precision +// Extracts the selected signed or unsigned 32-bit integer from a and zero +// extends. +// FORCE_INLINE int _mm_extract_epi32(__m128i a, __constrange(0,4) int imm) +#define _mm_extract_epi32(a, imm) \ + vgetq_lane_s32(vreinterpretq_s32_m128i(a), (imm)) + +// Extracts the selected signed or unsigned 64-bit integer from a and zero +// extends. +// FORCE_INLINE __int64 _mm_extract_epi64(__m128i a, __constrange(0,2) int imm) +#define _mm_extract_epi64(a, imm) \ + vgetq_lane_s64(vreinterpretq_s64_m128i(a), (imm)) + +// Extracts the selected signed or unsigned 8-bit integer from a and zero +// extends. +// FORCE_INLINE int _mm_extract_epi8(__m128i a, __constrange(0,16) int imm) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_extract_epi8 +#define _mm_extract_epi8(a, imm) vgetq_lane_u8(vreinterpretq_u8_m128i(a), (imm)) + +// Extracts the selected single-precision (32-bit) floating-point from a. +// FORCE_INLINE int _mm_extract_ps(__m128 a, __constrange(0,4) int imm) +#define _mm_extract_ps(a, imm) vgetq_lane_s32(vreinterpretq_s32_m128(a), (imm)) + +// Round the packed double-precision (64-bit) floating-point elements in a down +// to an integer value, and store the results as packed double-precision // floating-point elements in dst. -// software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_round_ps -FORCE_INLINE __m128 _mm_round_ps(__m128 a, int rounding) +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_floor_pd +FORCE_INLINE __m128d _mm_floor_pd(__m128d a) { #if defined(__aarch64__) - switch (rounding) { - case (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC): - return vreinterpretq_m128_f32(vrndnq_f32(vreinterpretq_f32_m128(a))); - case (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC): - return vreinterpretq_m128_f32(vrndmq_f32(vreinterpretq_f32_m128(a))); - case (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC): - return vreinterpretq_m128_f32(vrndpq_f32(vreinterpretq_f32_m128(a))); - case (_MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC): - return vreinterpretq_m128_f32(vrndq_f32(vreinterpretq_f32_m128(a))); - default: //_MM_FROUND_CUR_DIRECTION - return vreinterpretq_m128_f32(vrndiq_f32(vreinterpretq_f32_m128(a))); - } + return vreinterpretq_m128d_f64(vrndmq_f64(vreinterpretq_f64_m128d(a))); #else - float *v_float = (float *) &a; - __m128 zero, neg_inf, pos_inf; - - switch (rounding) { - case (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC): - return _mm_cvtepi32_ps(_mm_cvtps_epi32(a)); - case (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC): - return (__m128){floorf(v_float[0]), floorf(v_float[1]), - floorf(v_float[2]), floorf(v_float[3])}; - case (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC): - return (__m128){ceilf(v_float[0]), ceilf(v_float[1]), ceilf(v_float[2]), - ceilf(v_float[3])}; - case (_MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC): - zero = _mm_set_ps(0.0f, 0.0f, 0.0f, 0.0f); - neg_inf = _mm_set_ps(floorf(v_float[0]), floorf(v_float[1]), - floorf(v_float[2]), floorf(v_float[3])); - pos_inf = _mm_set_ps(ceilf(v_float[0]), ceilf(v_float[1]), - ceilf(v_float[2]), ceilf(v_float[3])); - return _mm_blendv_ps(pos_inf, neg_inf, _mm_cmple_ps(a, zero)); - default: //_MM_FROUND_CUR_DIRECTION - return (__m128){roundf(v_float[0]), roundf(v_float[1]), - roundf(v_float[2]), roundf(v_float[3])}; - } + double *f = (double *) &a; + return _mm_set_pd(floor(f[1]), floor(f[0])); #endif } -// Convert packed single-precision (32-bit) floating-point elements in a to -// packed 32-bit integers, and store the results in dst. -// -// FOR j := 0 to 1 -// i := 32*j -// dst[i+31:i] := Convert_FP32_To_Int32(a[i+31:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvt_ps2pi -FORCE_INLINE __m64 _mm_cvt_ps2pi(__m128 a) -{ -#if defined(__aarch64__) - return vreinterpret_m64_s32( - vget_low_s32(vcvtnq_s32_f32(vreinterpretq_f32_m128(a)))); -#else - return vreinterpret_m64_s32( - vcvt_s32_f32(vget_low_f32(vreinterpretq_f32_m128( - _mm_round_ps(a, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))))); -#endif -} - -// Convert packed single-precision (32-bit) floating-point elements in a to -// packed 32-bit integers, and store the results in dst. -// -// FOR j := 0 to 1 -// i := 32*j -// dst[i+31:i] := Convert_FP32_To_Int32(a[i+31:i]) -// ENDFOR -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_cvtps_pi32 -#define _mm_cvtps_pi32(a) _mm_cvt_ps2pi(a) - -// Round the packed single-precision (32-bit) floating-point elements in a up to -// an integer value, and store the results as packed single-precision -// floating-point elements in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_ps -FORCE_INLINE __m128 _mm_ceil_ps(__m128 a) -{ - return _mm_round_ps(a, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC); -} - -// Round the lower single-precision (32-bit) floating-point element in b up to -// an integer value, store the result as a single-precision floating-point -// element in the lower element of dst, and copy the upper 3 packed elements -// from a to the upper elements of dst. -// -// dst[31:0] := CEIL(b[31:0]) -// dst[127:32] := a[127:32] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_ceil_ss -FORCE_INLINE __m128 _mm_ceil_ss(__m128 a, __m128 b) -{ - return _mm_move_ss( - a, _mm_round_ps(b, _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC)); -} - // Round the packed single-precision (32-bit) floating-point elements in a down // to an integer value, and store the results as packed single-precision // floating-point elements in dst. // https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_floor_ps FORCE_INLINE __m128 _mm_floor_ps(__m128 a) { - return _mm_round_ps(a, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC); +#if defined(__aarch64__) || defined(__ARM_FEATURE_DIRECTED_ROUNDING) + return vreinterpretq_m128_f32(vrndmq_f32(vreinterpretq_f32_m128(a))); +#else + float *f = (float *) &a; + return _mm_set_ps(floorf(f[3]), floorf(f[2]), floorf(f[1]), floorf(f[0])); +#endif +} + +// Round the lower double-precision (64-bit) floating-point element in b down to +// an integer value, store the result as a double-precision floating-point +// element in the lower element of dst, and copy the upper element from a to the +// upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_floor_sd +FORCE_INLINE __m128d _mm_floor_sd(__m128d a, __m128d b) +{ + return _mm_move_sd(a, _mm_floor_pd(b)); } // Round the lower single-precision (32-bit) floating-point element in b down to @@ -6416,372 +7785,147 @@ FORCE_INLINE __m128 _mm_floor_ps(__m128 a) // https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_floor_ss FORCE_INLINE __m128 _mm_floor_ss(__m128 a, __m128 b) { - return _mm_move_ss( - a, _mm_round_ps(b, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC)); + return _mm_move_ss(a, _mm_floor_ps(b)); } -// Load 128-bits of integer data from unaligned memory into dst. This intrinsic -// may perform better than _mm_loadu_si128 when the data crosses a cache line -// boundary. -// -// dst[127:0] := MEM[mem_addr+127:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_lddqu_si128 -#define _mm_lddqu_si128 _mm_loadu_si128 +// Inserts the least significant 32 bits of b into the selected 32-bit integer +// of a. +// FORCE_INLINE __m128i _mm_insert_epi32(__m128i a, int b, +// __constrange(0,4) int imm) +#define _mm_insert_epi32(a, b, imm) \ + __extension__({ \ + vreinterpretq_m128i_s32( \ + vsetq_lane_s32((b), vreinterpretq_s32_m128i(a), (imm))); \ + }) -/* Miscellaneous Operations */ +// Inserts the least significant 64 bits of b into the selected 64-bit integer +// of a. +// FORCE_INLINE __m128i _mm_insert_epi64(__m128i a, __int64 b, +// __constrange(0,2) int imm) +#define _mm_insert_epi64(a, b, imm) \ + __extension__({ \ + vreinterpretq_m128i_s64( \ + vsetq_lane_s64((b), vreinterpretq_s64_m128i(a), (imm))); \ + }) -// Shifts the 8 signed 16-bit integers in a right by count bits while shifting -// in the sign bit. +// Inserts the least significant 8 bits of b into the selected 8-bit integer +// of a. +// FORCE_INLINE __m128i _mm_insert_epi8(__m128i a, int b, +// __constrange(0,16) int imm) +#define _mm_insert_epi8(a, b, imm) \ + __extension__({ \ + vreinterpretq_m128i_s8( \ + vsetq_lane_s8((b), vreinterpretq_s8_m128i(a), (imm))); \ + }) + +// Copy a to tmp, then insert a single-precision (32-bit) floating-point +// element from b into tmp using the control in imm8. Store tmp to dst using +// the mask in imm8 (elements are zeroed out when the corresponding bit is set). +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=insert_ps +#define _mm_insert_ps(a, b, imm8) \ + __extension__({ \ + float32x4_t tmp1 = \ + vsetq_lane_f32(vgetq_lane_f32(b, (imm8 >> 6) & 0x3), \ + vreinterpretq_f32_m128(a), 0); \ + float32x4_t tmp2 = \ + vsetq_lane_f32(vgetq_lane_f32(tmp1, 0), vreinterpretq_f32_m128(a), \ + ((imm8 >> 4) & 0x3)); \ + const uint32_t data[4] = {((imm8) & (1 << 0)) ? UINT32_MAX : 0, \ + ((imm8) & (1 << 1)) ? UINT32_MAX : 0, \ + ((imm8) & (1 << 2)) ? UINT32_MAX : 0, \ + ((imm8) & (1 << 3)) ? UINT32_MAX : 0}; \ + uint32x4_t mask = vld1q_u32(data); \ + float32x4_t all_zeros = vdupq_n_f32(0); \ + \ + vreinterpretq_m128_f32( \ + vbslq_f32(mask, all_zeros, vreinterpretq_f32_m128(tmp2))); \ + }) + +// epi versions of min/max +// Computes the pariwise maximums of the four signed 32-bit integer values of a +// and b. // -// r0 := a0 >> count -// r1 := a1 >> count -// ... -// r7 := a7 >> count +// A 128-bit parameter that can be defined with the following equations: +// r0 := (a0 > b0) ? a0 : b0 +// r1 := (a1 > b1) ? a1 : b1 +// r2 := (a2 > b2) ? a2 : b2 +// r3 := (a3 > b3) ? a3 : b3 // -// https://msdn.microsoft.com/en-us/library/3c9997dk(v%3dvs.90).aspx -FORCE_INLINE __m128i _mm_sra_epi16(__m128i a, __m128i count) +// https://msdn.microsoft.com/en-us/library/vstudio/bb514055(v=vs.100).aspx +FORCE_INLINE __m128i _mm_max_epi32(__m128i a, __m128i b) { - int64_t c = (int64_t) vget_low_s64((int64x2_t) count); - if (unlikely(c > 15)) - return _mm_cmplt_epi16(a, _mm_setzero_si128()); - return vreinterpretq_m128i_s16(vshlq_s16((int16x8_t) a, vdupq_n_s16(-c))); + return vreinterpretq_m128i_s32( + vmaxq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); } -// Shifts the 4 signed 32-bit integers in a right by count bits while shifting -// in the sign bit. -// -// r0 := a0 >> count -// r1 := a1 >> count -// r2 := a2 >> count -// r3 := a3 >> count -// -// https://msdn.microsoft.com/en-us/library/ce40009e(v%3dvs.100).aspx -FORCE_INLINE __m128i _mm_sra_epi32(__m128i a, __m128i count) -{ - int64_t c = (int64_t) vget_low_s64((int64x2_t) count); - if (unlikely(c > 31)) - return _mm_cmplt_epi32(a, _mm_setzero_si128()); - return vreinterpretq_m128i_s32(vshlq_s32((int32x4_t) a, vdupq_n_s32(-c))); -} - -// Packs the 16 signed 16-bit integers from a and b into 8-bit integers and -// saturates. -// https://msdn.microsoft.com/en-us/library/k4y4f7w5%28v=vs.90%29.aspx -FORCE_INLINE __m128i _mm_packs_epi16(__m128i a, __m128i b) +// Compare packed signed 8-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epi8 +FORCE_INLINE __m128i _mm_max_epi8(__m128i a, __m128i b) { return vreinterpretq_m128i_s8( - vcombine_s8(vqmovn_s16(vreinterpretq_s16_m128i(a)), - vqmovn_s16(vreinterpretq_s16_m128i(b)))); + vmaxq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); } -// Packs the 16 signed 16 - bit integers from a and b into 8 - bit unsigned -// integers and saturates. -// -// r0 := UnsignedSaturate(a0) -// r1 := UnsignedSaturate(a1) -// ... -// r7 := UnsignedSaturate(a7) -// r8 := UnsignedSaturate(b0) -// r9 := UnsignedSaturate(b1) -// ... -// r15 := UnsignedSaturate(b7) -// -// https://msdn.microsoft.com/en-us/library/07ad1wx4(v=vs.100).aspx -FORCE_INLINE __m128i _mm_packus_epi16(const __m128i a, const __m128i b) -{ - return vreinterpretq_m128i_u8( - vcombine_u8(vqmovun_s16(vreinterpretq_s16_m128i(a)), - vqmovun_s16(vreinterpretq_s16_m128i(b)))); -} - -// Packs the 8 signed 32-bit integers from a and b into signed 16-bit integers -// and saturates. -// -// r0 := SignedSaturate(a0) -// r1 := SignedSaturate(a1) -// r2 := SignedSaturate(a2) -// r3 := SignedSaturate(a3) -// r4 := SignedSaturate(b0) -// r5 := SignedSaturate(b1) -// r6 := SignedSaturate(b2) -// r7 := SignedSaturate(b3) -// -// https://msdn.microsoft.com/en-us/library/393t56f9%28v=vs.90%29.aspx -FORCE_INLINE __m128i _mm_packs_epi32(__m128i a, __m128i b) -{ - return vreinterpretq_m128i_s16( - vcombine_s16(vqmovn_s32(vreinterpretq_s32_m128i(a)), - vqmovn_s32(vreinterpretq_s32_m128i(b)))); -} - -// Packs the 8 unsigned 32-bit integers from a and b into unsigned 16-bit -// integers and saturates. -// -// r0 := UnsignedSaturate(a0) -// r1 := UnsignedSaturate(a1) -// r2 := UnsignedSaturate(a2) -// r3 := UnsignedSaturate(a3) -// r4 := UnsignedSaturate(b0) -// r5 := UnsignedSaturate(b1) -// r6 := UnsignedSaturate(b2) -// r7 := UnsignedSaturate(b3) -FORCE_INLINE __m128i _mm_packus_epi32(__m128i a, __m128i b) +// Compare packed unsigned 16-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epu16 +FORCE_INLINE __m128i _mm_max_epu16(__m128i a, __m128i b) { return vreinterpretq_m128i_u16( - vcombine_u16(vqmovun_s32(vreinterpretq_s32_m128i(a)), - vqmovun_s32(vreinterpretq_s32_m128i(b)))); + vmaxq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); } -// Interleaves the lower 8 signed or unsigned 8-bit integers in a with the lower -// 8 signed or unsigned 8-bit integers in b. -// -// r0 := a0 -// r1 := b0 -// r2 := a1 -// r3 := b1 -// ... -// r14 := a7 -// r15 := b7 -// -// https://msdn.microsoft.com/en-us/library/xf7k860c%28v=vs.90%29.aspx -FORCE_INLINE __m128i _mm_unpacklo_epi8(__m128i a, __m128i b) +// Compare packed unsigned 32-bit integers in a and b, and store packed maximum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epu32 +FORCE_INLINE __m128i _mm_max_epu32(__m128i a, __m128i b) { -#if defined(__aarch64__) - return vreinterpretq_m128i_s8( - vzip1q_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -#else - int8x8_t a1 = vreinterpret_s8_s16(vget_low_s16(vreinterpretq_s16_m128i(a))); - int8x8_t b1 = vreinterpret_s8_s16(vget_low_s16(vreinterpretq_s16_m128i(b))); - int8x8x2_t result = vzip_s8(a1, b1); - return vreinterpretq_m128i_s8(vcombine_s8(result.val[0], result.val[1])); -#endif + return vreinterpretq_m128i_u32( + vmaxq_u32(vreinterpretq_u32_m128i(a), vreinterpretq_u32_m128i(b))); } -// Interleaves the lower 4 signed or unsigned 16-bit integers in a with the -// lower 4 signed or unsigned 16-bit integers in b. +// Computes the pariwise minima of the four signed 32-bit integer values of a +// and b. // -// r0 := a0 -// r1 := b0 -// r2 := a1 -// r3 := b1 -// r4 := a2 -// r5 := b2 -// r6 := a3 -// r7 := b3 +// A 128-bit parameter that can be defined with the following equations: +// r0 := (a0 < b0) ? a0 : b0 +// r1 := (a1 < b1) ? a1 : b1 +// r2 := (a2 < b2) ? a2 : b2 +// r3 := (a3 < b3) ? a3 : b3 // -// https://msdn.microsoft.com/en-us/library/btxb17bw%28v=vs.90%29.aspx -FORCE_INLINE __m128i _mm_unpacklo_epi16(__m128i a, __m128i b) +// https://msdn.microsoft.com/en-us/library/vstudio/bb531476(v=vs.100).aspx +FORCE_INLINE __m128i _mm_min_epi32(__m128i a, __m128i b) { -#if defined(__aarch64__) - return vreinterpretq_m128i_s16( - vzip1q_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -#else - int16x4_t a1 = vget_low_s16(vreinterpretq_s16_m128i(a)); - int16x4_t b1 = vget_low_s16(vreinterpretq_s16_m128i(b)); - int16x4x2_t result = vzip_s16(a1, b1); - return vreinterpretq_m128i_s16(vcombine_s16(result.val[0], result.val[1])); -#endif -} - -// Interleaves the lower 2 signed or unsigned 32 - bit integers in a with the -// lower 2 signed or unsigned 32 - bit integers in b. -// -// r0 := a0 -// r1 := b0 -// r2 := a1 -// r3 := b1 -// -// https://msdn.microsoft.com/en-us/library/x8atst9d(v=vs.100).aspx -FORCE_INLINE __m128i _mm_unpacklo_epi32(__m128i a, __m128i b) -{ -#if defined(__aarch64__) return vreinterpretq_m128i_s32( - vzip1q_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -#else - int32x2_t a1 = vget_low_s32(vreinterpretq_s32_m128i(a)); - int32x2_t b1 = vget_low_s32(vreinterpretq_s32_m128i(b)); - int32x2x2_t result = vzip_s32(a1, b1); - return vreinterpretq_m128i_s32(vcombine_s32(result.val[0], result.val[1])); -#endif + vminq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); } -FORCE_INLINE __m128i _mm_unpacklo_epi64(__m128i a, __m128i b) +// Compare packed signed 8-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_epi8 +FORCE_INLINE __m128i _mm_min_epi8(__m128i a, __m128i b) { - int64x1_t a_l = vget_low_s64(vreinterpretq_s64_m128i(a)); - int64x1_t b_l = vget_low_s64(vreinterpretq_s64_m128i(b)); - return vreinterpretq_m128i_s64(vcombine_s64(a_l, b_l)); -} - -// Selects and interleaves the lower two single-precision, floating-point values -// from a and b. -// -// r0 := a0 -// r1 := b0 -// r2 := a1 -// r3 := b1 -// -// https://msdn.microsoft.com/en-us/library/25st103b%28v=vs.90%29.aspx -FORCE_INLINE __m128 _mm_unpacklo_ps(__m128 a, __m128 b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128_f32( - vzip1q_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -#else - float32x2_t a1 = vget_low_f32(vreinterpretq_f32_m128(a)); - float32x2_t b1 = vget_low_f32(vreinterpretq_f32_m128(b)); - float32x2x2_t result = vzip_f32(a1, b1); - return vreinterpretq_m128_f32(vcombine_f32(result.val[0], result.val[1])); -#endif -} - -// Unpack and interleave double-precision (64-bit) floating-point elements from -// the low half of a and b, and store the results in dst. -// -// DEFINE INTERLEAVE_QWORDS(src1[127:0], src2[127:0]) { -// dst[63:0] := src1[63:0] -// dst[127:64] := src2[63:0] -// RETURN dst[127:0] -// } -// dst[127:0] := INTERLEAVE_QWORDS(a[127:0], b[127:0]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_unpacklo_pd -FORCE_INLINE __m128d _mm_unpacklo_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vzip1q_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - return vreinterpretq_m128d_s64( - vcombine_s64(vget_low_s64(vreinterpretq_s64_m128d(a)), - vget_low_s64(vreinterpretq_s64_m128d(b)))); -#endif -} - -// Unpack and interleave double-precision (64-bit) floating-point elements from -// the high half of a and b, and store the results in dst. -// -// DEFINE INTERLEAVE_HIGH_QWORDS(src1[127:0], src2[127:0]) { -// dst[63:0] := src1[127:64] -// dst[127:64] := src2[127:64] -// RETURN dst[127:0] -// } -// dst[127:0] := INTERLEAVE_HIGH_QWORDS(a[127:0], b[127:0]) -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_unpackhi_pd -FORCE_INLINE __m128d _mm_unpackhi_pd(__m128d a, __m128d b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128d_f64( - vzip2q_f64(vreinterpretq_f64_m128d(a), vreinterpretq_f64_m128d(b))); -#else - return vreinterpretq_m128d_s64( - vcombine_s64(vget_high_s64(vreinterpretq_s64_m128d(a)), - vget_high_s64(vreinterpretq_s64_m128d(b)))); -#endif -} - -// Selects and interleaves the upper two single-precision, floating-point values -// from a and b. -// -// r0 := a2 -// r1 := b2 -// r2 := a3 -// r3 := b3 -// -// https://msdn.microsoft.com/en-us/library/skccxx7d%28v=vs.90%29.aspx -FORCE_INLINE __m128 _mm_unpackhi_ps(__m128 a, __m128 b) -{ -#if defined(__aarch64__) - return vreinterpretq_m128_f32( - vzip2q_f32(vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))); -#else - float32x2_t a1 = vget_high_f32(vreinterpretq_f32_m128(a)); - float32x2_t b1 = vget_high_f32(vreinterpretq_f32_m128(b)); - float32x2x2_t result = vzip_f32(a1, b1); - return vreinterpretq_m128_f32(vcombine_f32(result.val[0], result.val[1])); -#endif -} - -// Interleaves the upper 8 signed or unsigned 8-bit integers in a with the upper -// 8 signed or unsigned 8-bit integers in b. -// -// r0 := a8 -// r1 := b8 -// r2 := a9 -// r3 := b9 -// ... -// r14 := a15 -// r15 := b15 -// -// https://msdn.microsoft.com/en-us/library/t5h7783k(v=vs.100).aspx -FORCE_INLINE __m128i _mm_unpackhi_epi8(__m128i a, __m128i b) -{ -#if defined(__aarch64__) return vreinterpretq_m128i_s8( - vzip2q_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); -#else - int8x8_t a1 = - vreinterpret_s8_s16(vget_high_s16(vreinterpretq_s16_m128i(a))); - int8x8_t b1 = - vreinterpret_s8_s16(vget_high_s16(vreinterpretq_s16_m128i(b))); - int8x8x2_t result = vzip_s8(a1, b1); - return vreinterpretq_m128i_s8(vcombine_s8(result.val[0], result.val[1])); -#endif + vminq_s8(vreinterpretq_s8_m128i(a), vreinterpretq_s8_m128i(b))); } -// Interleaves the upper 4 signed or unsigned 16-bit integers in a with the -// upper 4 signed or unsigned 16-bit integers in b. -// -// r0 := a4 -// r1 := b4 -// r2 := a5 -// r3 := b5 -// r4 := a6 -// r5 := b6 -// r6 := a7 -// r7 := b7 -// -// https://msdn.microsoft.com/en-us/library/03196cz7(v=vs.100).aspx -FORCE_INLINE __m128i _mm_unpackhi_epi16(__m128i a, __m128i b) +// Compare packed unsigned 16-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_min_epu16 +FORCE_INLINE __m128i _mm_min_epu16(__m128i a, __m128i b) { -#if defined(__aarch64__) - return vreinterpretq_m128i_s16( - vzip2q_s16(vreinterpretq_s16_m128i(a), vreinterpretq_s16_m128i(b))); -#else - int16x4_t a1 = vget_high_s16(vreinterpretq_s16_m128i(a)); - int16x4_t b1 = vget_high_s16(vreinterpretq_s16_m128i(b)); - int16x4x2_t result = vzip_s16(a1, b1); - return vreinterpretq_m128i_s16(vcombine_s16(result.val[0], result.val[1])); -#endif + return vreinterpretq_m128i_u16( + vminq_u16(vreinterpretq_u16_m128i(a), vreinterpretq_u16_m128i(b))); } -// Interleaves the upper 2 signed or unsigned 32-bit integers in a with the -// upper 2 signed or unsigned 32-bit integers in b. -// https://msdn.microsoft.com/en-us/library/65sa7cbs(v=vs.100).aspx -FORCE_INLINE __m128i _mm_unpackhi_epi32(__m128i a, __m128i b) +// Compare packed unsigned 32-bit integers in a and b, and store packed minimum +// values in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_max_epu32 +FORCE_INLINE __m128i _mm_min_epu32(__m128i a, __m128i b) { -#if defined(__aarch64__) - return vreinterpretq_m128i_s32( - vzip2q_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); -#else - int32x2_t a1 = vget_high_s32(vreinterpretq_s32_m128i(a)); - int32x2_t b1 = vget_high_s32(vreinterpretq_s32_m128i(b)); - int32x2x2_t result = vzip_s32(a1, b1); - return vreinterpretq_m128i_s32(vcombine_s32(result.val[0], result.val[1])); -#endif -} - -// Interleaves the upper signed or unsigned 64-bit integer in a with the -// upper signed or unsigned 64-bit integer in b. -// -// r0 := a1 -// r1 := b1 -FORCE_INLINE __m128i _mm_unpackhi_epi64(__m128i a, __m128i b) -{ - int64x1_t a_h = vget_high_s64(vreinterpretq_s64_m128i(a)); - int64x1_t b_h = vget_high_s64(vreinterpretq_s64_m128i(b)); - return vreinterpretq_m128i_s64(vcombine_s64(a_h, b_h)); + return vreinterpretq_m128i_u32( + vminq_u32(vreinterpretq_u32_m128i(a), vreinterpretq_u32_m128i(b))); } // Horizontally compute the minimum amongst the packed unsigned 16-bit integers @@ -6837,6 +7981,339 @@ FORCE_INLINE __m128i _mm_minpos_epu16(__m128i a) return dst; } +// Compute the sum of absolute differences (SADs) of quadruplets of unsigned +// 8-bit integers in a compared to those in b, and store the 16-bit results in +// dst. Eight SADs are performed using one quadruplet from b and eight +// quadruplets from a. One quadruplet is selected from b starting at on the +// offset specified in imm8. Eight quadruplets are formed from sequential 8-bit +// integers selected from a starting at the offset specified in imm8. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_mpsadbw_epu8 +FORCE_INLINE __m128i _mm_mpsadbw_epu8(__m128i a, __m128i b, const int imm) +{ + uint8x16_t _a, _b; + + switch (imm & 0x4) { + case 0: + // do nothing + _a = vreinterpretq_u8_m128i(a); + break; + case 4: + _a = vreinterpretq_u8_u32(vextq_u32(vreinterpretq_u32_m128i(a), + vreinterpretq_u32_m128i(a), 1)); + break; + default: +#if defined(__GNUC__) || defined(__clang__) + __builtin_unreachable(); +#endif + break; + } + + switch (imm & 0x3) { + case 0: + _b = vreinterpretq_u8_u32( + vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 0))); + break; + case 1: + _b = vreinterpretq_u8_u32( + vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 1))); + break; + case 2: + _b = vreinterpretq_u8_u32( + vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 2))); + break; + case 3: + _b = vreinterpretq_u8_u32( + vdupq_n_u32(vgetq_lane_u32(vreinterpretq_u32_m128i(b), 3))); + break; + default: +#if defined(__GNUC__) || defined(__clang__) + __builtin_unreachable(); +#endif + break; + } + + int16x8_t c04, c15, c26, c37; + uint8x8_t low_b = vget_low_u8(_b); + c04 = vabsq_s16(vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(_a), low_b))); + _a = vextq_u8(_a, _a, 1); + c15 = vabsq_s16(vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(_a), low_b))); + _a = vextq_u8(_a, _a, 1); + c26 = vabsq_s16(vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(_a), low_b))); + _a = vextq_u8(_a, _a, 1); + c37 = vabsq_s16(vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(_a), low_b))); +#if defined(__aarch64__) + // |0|4|2|6| + c04 = vpaddq_s16(c04, c26); + // |1|5|3|7| + c15 = vpaddq_s16(c15, c37); + + int32x4_t trn1_c = + vtrn1q_s32(vreinterpretq_s32_s16(c04), vreinterpretq_s32_s16(c15)); + int32x4_t trn2_c = + vtrn2q_s32(vreinterpretq_s32_s16(c04), vreinterpretq_s32_s16(c15)); + return vreinterpretq_m128i_s16(vpaddq_s16(vreinterpretq_s16_s32(trn1_c), + vreinterpretq_s16_s32(trn2_c))); +#else + int16x4_t c01, c23, c45, c67; + c01 = vpadd_s16(vget_low_s16(c04), vget_low_s16(c15)); + c23 = vpadd_s16(vget_low_s16(c26), vget_low_s16(c37)); + c45 = vpadd_s16(vget_high_s16(c04), vget_high_s16(c15)); + c67 = vpadd_s16(vget_high_s16(c26), vget_high_s16(c37)); + + return vreinterpretq_m128i_s16( + vcombine_s16(vpadd_s16(c01, c23), vpadd_s16(c45, c67))); +#endif +} + +// Multiply the low signed 32-bit integers from each packed 64-bit element in +// a and b, and store the signed 64-bit results in dst. +// +// r0 := (int64_t)(int32_t)a0 * (int64_t)(int32_t)b0 +// r1 := (int64_t)(int32_t)a2 * (int64_t)(int32_t)b2 +FORCE_INLINE __m128i _mm_mul_epi32(__m128i a, __m128i b) +{ + // vmull_s32 upcasts instead of masking, so we downcast. + int32x2_t a_lo = vmovn_s64(vreinterpretq_s64_m128i(a)); + int32x2_t b_lo = vmovn_s64(vreinterpretq_s64_m128i(b)); + return vreinterpretq_m128i_s64(vmull_s32(a_lo, b_lo)); +} + +// Multiplies the 4 signed or unsigned 32-bit integers from a by the 4 signed or +// unsigned 32-bit integers from b. +// https://msdn.microsoft.com/en-us/library/vstudio/bb531409(v=vs.100).aspx +FORCE_INLINE __m128i _mm_mullo_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_s32( + vmulq_s32(vreinterpretq_s32_m128i(a), vreinterpretq_s32_m128i(b))); +} + +// Packs the 8 unsigned 32-bit integers from a and b into unsigned 16-bit +// integers and saturates. +// +// r0 := UnsignedSaturate(a0) +// r1 := UnsignedSaturate(a1) +// r2 := UnsignedSaturate(a2) +// r3 := UnsignedSaturate(a3) +// r4 := UnsignedSaturate(b0) +// r5 := UnsignedSaturate(b1) +// r6 := UnsignedSaturate(b2) +// r7 := UnsignedSaturate(b3) +FORCE_INLINE __m128i _mm_packus_epi32(__m128i a, __m128i b) +{ + return vreinterpretq_m128i_u16( + vcombine_u16(vqmovun_s32(vreinterpretq_s32_m128i(a)), + vqmovun_s32(vreinterpretq_s32_m128i(b)))); +} + +// Round the packed double-precision (64-bit) floating-point elements in a using +// the rounding parameter, and store the results as packed double-precision +// floating-point elements in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_round_pd +FORCE_INLINE __m128d _mm_round_pd(__m128d a, int rounding) +{ +#if defined(__aarch64__) + switch (rounding) { + case (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC): + return vreinterpretq_m128d_f64(vrndnq_f64(vreinterpretq_f64_m128d(a))); + case (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC): + return _mm_floor_pd(a); + case (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC): + return _mm_ceil_pd(a); + case (_MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC): + return vreinterpretq_m128d_f64(vrndq_f64(vreinterpretq_f64_m128d(a))); + default: //_MM_FROUND_CUR_DIRECTION + return vreinterpretq_m128d_f64(vrndiq_f64(vreinterpretq_f64_m128d(a))); + } +#else + double *v_double = (double *) &a; + + if (rounding == (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_NEAREST)) { + double res[2], tmp; + for (int i = 0; i < 2; i++) { + tmp = (v_double[i] < 0) ? -v_double[i] : v_double[i]; + double roundDown = floor(tmp); // Round down value + double roundUp = ceil(tmp); // Round up value + double diffDown = tmp - roundDown; + double diffUp = roundUp - tmp; + if (diffDown < diffUp) { + /* If it's closer to the round down value, then use it */ + res[i] = roundDown; + } else if (diffDown > diffUp) { + /* If it's closer to the round up value, then use it */ + res[i] = roundUp; + } else { + /* If it's equidistant between round up and round down value, + * pick the one which is an even number */ + double half = roundDown / 2; + if (half != floor(half)) { + /* If the round down value is odd, return the round up value + */ + res[i] = roundUp; + } else { + /* If the round up value is odd, return the round down value + */ + res[i] = roundDown; + } + } + res[i] = (v_double[i] < 0) ? -res[i] : res[i]; + } + return _mm_set_pd(res[1], res[0]); + } else if (rounding == (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_DOWN)) { + return _mm_floor_pd(a); + } else if (rounding == (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_UP)) { + return _mm_ceil_pd(a); + } + return _mm_set_pd(v_double[1] > 0 ? floor(v_double[1]) : ceil(v_double[1]), + v_double[0] > 0 ? floor(v_double[0]) : ceil(v_double[0])); +#endif +} + +// Round the packed single-precision (32-bit) floating-point elements in a using +// the rounding parameter, and store the results as packed single-precision +// floating-point elements in dst. +// software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_round_ps +FORCE_INLINE __m128 _mm_round_ps(__m128 a, int rounding) +{ +#if defined(__aarch64__) || defined(__ARM_FEATURE_DIRECTED_ROUNDING) + switch (rounding) { + case (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC): + return vreinterpretq_m128_f32(vrndnq_f32(vreinterpretq_f32_m128(a))); + case (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC): + return _mm_floor_ps(a); + case (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC): + return _mm_ceil_ps(a); + case (_MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC): + return vreinterpretq_m128_f32(vrndq_f32(vreinterpretq_f32_m128(a))); + default: //_MM_FROUND_CUR_DIRECTION + return vreinterpretq_m128_f32(vrndiq_f32(vreinterpretq_f32_m128(a))); + } +#else + float *v_float = (float *) &a; + + if (rounding == (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_NEAREST)) { + uint32x4_t signmask = vdupq_n_u32(0x80000000); + float32x4_t half = vbslq_f32(signmask, vreinterpretq_f32_m128(a), + vdupq_n_f32(0.5f)); /* +/- 0.5 */ + int32x4_t r_normal = vcvtq_s32_f32(vaddq_f32( + vreinterpretq_f32_m128(a), half)); /* round to integer: [a + 0.5]*/ + int32x4_t r_trunc = vcvtq_s32_f32( + vreinterpretq_f32_m128(a)); /* truncate to integer: [a] */ + int32x4_t plusone = vreinterpretq_s32_u32(vshrq_n_u32( + vreinterpretq_u32_s32(vnegq_s32(r_trunc)), 31)); /* 1 or 0 */ + int32x4_t r_even = vbicq_s32(vaddq_s32(r_trunc, plusone), + vdupq_n_s32(1)); /* ([a] + {0,1}) & ~1 */ + float32x4_t delta = vsubq_f32( + vreinterpretq_f32_m128(a), + vcvtq_f32_s32(r_trunc)); /* compute delta: delta = (a - [a]) */ + uint32x4_t is_delta_half = + vceqq_f32(delta, half); /* delta == +/- 0.5 */ + return vreinterpretq_m128_f32( + vcvtq_f32_s32(vbslq_s32(is_delta_half, r_even, r_normal))); + } else if (rounding == (_MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_DOWN)) { + return _mm_floor_ps(a); + } else if (rounding == (_MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC) || + (rounding == _MM_FROUND_CUR_DIRECTION && + _MM_GET_ROUNDING_MODE() == _MM_ROUND_UP)) { + return _mm_ceil_ps(a); + } + return _mm_set_ps(v_float[3] > 0 ? floorf(v_float[3]) : ceilf(v_float[3]), + v_float[2] > 0 ? floorf(v_float[2]) : ceilf(v_float[2]), + v_float[1] > 0 ? floorf(v_float[1]) : ceilf(v_float[1]), + v_float[0] > 0 ? floorf(v_float[0]) : ceilf(v_float[0])); +#endif +} + +// Round the lower double-precision (64-bit) floating-point element in b using +// the rounding parameter, store the result as a double-precision floating-point +// element in the lower element of dst, and copy the upper element from a to the +// upper element of dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_round_sd +FORCE_INLINE __m128d _mm_round_sd(__m128d a, __m128d b, int rounding) +{ + return _mm_move_sd(a, _mm_round_pd(b, rounding)); +} + +// Round the lower single-precision (32-bit) floating-point element in b using +// the rounding parameter, store the result as a single-precision floating-point +// element in the lower element of dst, and copy the upper 3 packed elements +// from a to the upper elements of dst. Rounding is done according to the +// rounding[3:0] parameter, which can be one of: +// (_MM_FROUND_TO_NEAREST_INT |_MM_FROUND_NO_EXC) // round to nearest, and +// suppress exceptions +// (_MM_FROUND_TO_NEG_INF |_MM_FROUND_NO_EXC) // round down, and +// suppress exceptions +// (_MM_FROUND_TO_POS_INF |_MM_FROUND_NO_EXC) // round up, and suppress +// exceptions +// (_MM_FROUND_TO_ZERO |_MM_FROUND_NO_EXC) // truncate, and suppress +// exceptions _MM_FROUND_CUR_DIRECTION // use MXCSR.RC; see +// _MM_SET_ROUNDING_MODE +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_round_ss +FORCE_INLINE __m128 _mm_round_ss(__m128 a, __m128 b, int rounding) +{ + return _mm_move_ss(a, _mm_round_ps(b, rounding)); +} + +// Load 128-bits of integer data from memory into dst using a non-temporal +// memory hint. mem_addr must be aligned on a 16-byte boundary or a +// general-protection exception may be generated. +// +// dst[127:0] := MEM[mem_addr+127:mem_addr] +// +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_load_si128 +FORCE_INLINE __m128i _mm_stream_load_si128(__m128i *p) +{ +#if __has_builtin(__builtin_nontemporal_store) + return __builtin_nontemporal_load(p); +#else + return vreinterpretq_m128i_s64(vld1q_s64((int64_t *) p)); +#endif +} + +// Compute the bitwise NOT of a and then AND with a 128-bit vector containing +// all 1's, and return 1 if the result is zero, otherwise return 0. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_test_all_ones +FORCE_INLINE int _mm_test_all_ones(__m128i a) +{ + return (uint64_t) (vgetq_lane_s64(a, 0) & vgetq_lane_s64(a, 1)) == + ~(uint64_t) 0; +} + +// Compute the bitwise AND of 128 bits (representing integer data) in a and +// mask, and return 1 if the result is zero, otherwise return 0. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_test_all_zeros +FORCE_INLINE int _mm_test_all_zeros(__m128i a, __m128i mask) +{ + int64x2_t a_and_mask = + vandq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(mask)); + return !(vgetq_lane_s64(a_and_mask, 0) | vgetq_lane_s64(a_and_mask, 1)); +} + +// Compute the bitwise AND of 128 bits (representing integer data) in a and +// mask, and set ZF to 1 if the result is zero, otherwise set ZF to 0. Compute +// the bitwise NOT of a and then AND with mask, and set CF to 1 if the result is +// zero, otherwise set CF to 0. Return 1 if both the ZF and CF values are zero, +// otherwise return 0. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=mm_test_mix_ones_zero +FORCE_INLINE int _mm_test_mix_ones_zeros(__m128i a, __m128i mask) +{ + uint64x2_t zf = + vandq_u64(vreinterpretq_u64_m128i(mask), vreinterpretq_u64_m128i(a)); + uint64x2_t cf = + vbicq_u64(vreinterpretq_u64_m128i(mask), vreinterpretq_u64_m128i(a)); + uint64x2_t result = vandq_u64(zf, cf); + return !(vgetq_lane_u64(result, 0) | vgetq_lane_u64(result, 1)); +} + // Compute the bitwise AND of 128 bits (representing integer data) in a and b, // and set ZF to 1 if the result is zero, otherwise set ZF to 0. Compute the // bitwise NOT of a and then AND with b, and set CF to 1 if the result is zero, @@ -6850,6 +8327,14 @@ FORCE_INLINE int _mm_testc_si128(__m128i a, __m128i b) return !(vgetq_lane_s64(s64, 0) | vgetq_lane_s64(s64, 1)); } +// Compute the bitwise AND of 128 bits (representing integer data) in a and b, +// and set ZF to 1 if the result is zero, otherwise set ZF to 0. Compute the +// bitwise NOT of a and then AND with b, and set CF to 1 if the result is zero, +// otherwise set CF to 0. Return 1 if both the ZF and CF values are zero, +// otherwise return 0. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_testnzc_si128 +#define _mm_testnzc_si128(a, b) _mm_test_mix_ones_zeros(a, b) + // Compute the bitwise AND of 128 bits (representing integer data) in a and b, // and set ZF to 1 if the result is zero, otherwise set ZF to 0. Compute the // bitwise NOT of a and then AND with b, and set CF to 1 if the result is zero, @@ -6862,305 +8347,99 @@ FORCE_INLINE int _mm_testz_si128(__m128i a, __m128i b) return !(vgetq_lane_s64(s64, 0) | vgetq_lane_s64(s64, 1)); } -// Extracts the selected signed or unsigned 8-bit integer from a and zero -// extends. -// FORCE_INLINE int _mm_extract_epi8(__m128i a, __constrange(0,16) int imm) -#define _mm_extract_epi8(a, imm) vgetq_lane_u8(vreinterpretq_u8_m128i(a), (imm)) +/* SSE4.2 */ -// Inserts the least significant 8 bits of b into the selected 8-bit integer -// of a. -// FORCE_INLINE __m128i _mm_insert_epi8(__m128i a, int b, -// __constrange(0,16) int imm) -#define _mm_insert_epi8(a, b, imm) \ - __extension__({ \ - vreinterpretq_m128i_s8( \ - vsetq_lane_s8((b), vreinterpretq_s8_m128i(a), (imm))); \ - }) - -// Extracts the selected signed or unsigned 16-bit integer from a and zero -// extends. -// https://msdn.microsoft.com/en-us/library/6dceta0c(v=vs.100).aspx -// FORCE_INLINE int _mm_extract_epi16(__m128i a, __constrange(0,8) int imm) -#define _mm_extract_epi16(a, imm) \ - vgetq_lane_u16(vreinterpretq_u16_m128i(a), (imm)) - -// Extract a 16-bit integer from a, selected with imm8, and store the result in -// the lower element of dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_extract_pi16 -#define _mm_extract_pi16(a, imm) \ - (int32_t) vget_lane_u16(vreinterpret_u16_m64(a), (imm)) - -// Inserts the least significant 16 bits of b into the selected 16-bit integer -// of a. -// https://msdn.microsoft.com/en-us/library/kaze8hz1%28v=vs.100%29.aspx -// FORCE_INLINE __m128i _mm_insert_epi16(__m128i a, int b, -// __constrange(0,8) int imm) -#define _mm_insert_epi16(a, b, imm) \ - __extension__({ \ - vreinterpretq_m128i_s16( \ - vsetq_lane_s16((b), vreinterpretq_s16_m128i(a), (imm))); \ - }) - -// Copy a to dst, and insert the 16-bit integer i into dst at the location -// specified by imm8. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_insert_pi16 -#define _mm_insert_pi16(a, b, imm) \ - __extension__({ \ - vreinterpret_m64_s16( \ - vset_lane_s16((b), vreinterpret_s16_m64(a), (imm))); \ - }) - -// Extracts the selected signed or unsigned 32-bit integer from a and zero -// extends. -// FORCE_INLINE int _mm_extract_epi32(__m128i a, __constrange(0,4) int imm) -#define _mm_extract_epi32(a, imm) \ - vgetq_lane_s32(vreinterpretq_s32_m128i(a), (imm)) - -// Extracts the selected single-precision (32-bit) floating-point from a. -// FORCE_INLINE int _mm_extract_ps(__m128 a, __constrange(0,4) int imm) -#define _mm_extract_ps(a, imm) vgetq_lane_s32(vreinterpretq_s32_m128(a), (imm)) - -// Inserts the least significant 32 bits of b into the selected 32-bit integer -// of a. -// FORCE_INLINE __m128i _mm_insert_epi32(__m128i a, int b, -// __constrange(0,4) int imm) -#define _mm_insert_epi32(a, b, imm) \ - __extension__({ \ - vreinterpretq_m128i_s32( \ - vsetq_lane_s32((b), vreinterpretq_s32_m128i(a), (imm))); \ - }) - -// Extracts the selected signed or unsigned 64-bit integer from a and zero -// extends. -// FORCE_INLINE __int64 _mm_extract_epi64(__m128i a, __constrange(0,2) int imm) -#define _mm_extract_epi64(a, imm) \ - vgetq_lane_s64(vreinterpretq_s64_m128i(a), (imm)) - -// Inserts the least significant 64 bits of b into the selected 64-bit integer -// of a. -// FORCE_INLINE __m128i _mm_insert_epi64(__m128i a, __int64 b, -// __constrange(0,2) int imm) -#define _mm_insert_epi64(a, b, imm) \ - __extension__({ \ - vreinterpretq_m128i_s64( \ - vsetq_lane_s64((b), vreinterpretq_s64_m128i(a), (imm))); \ - }) - -// Count the number of bits set to 1 in unsigned 32-bit integer a, and -// return that count in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_popcnt_u32 -FORCE_INLINE int _mm_popcnt_u32(unsigned int a) +// Compares the 2 signed 64-bit integers in a and the 2 signed 64-bit integers +// in b for greater than. +FORCE_INLINE __m128i _mm_cmpgt_epi64(__m128i a, __m128i b) { #if defined(__aarch64__) -#if __has_builtin(__builtin_popcount) - return __builtin_popcount(a); + return vreinterpretq_m128i_u64( + vcgtq_s64(vreinterpretq_s64_m128i(a), vreinterpretq_s64_m128i(b))); #else - return (int) vaddlv_u8(vcnt_u8(vcreate_u8((uint64_t) a))); -#endif -#else - uint32_t count = 0; - uint8x8_t input_val, count8x8_val; - uint16x4_t count16x4_val; - uint32x2_t count32x2_val; - - input_val = vld1_u8((uint8_t *) &a); - count8x8_val = vcnt_u8(input_val); - count16x4_val = vpaddl_u8(count8x8_val); - count32x2_val = vpaddl_u16(count16x4_val); - - vst1_u32(&count, count32x2_val); - return count; + return vreinterpretq_m128i_s64(vshrq_n_s64( + vqsubq_s64(vreinterpretq_s64_m128i(b), vreinterpretq_s64_m128i(a)), + 63)); #endif } -// Count the number of bits set to 1 in unsigned 64-bit integer a, and -// return that count in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_popcnt_u64 -FORCE_INLINE int64_t _mm_popcnt_u64(uint64_t a) +// Starting with the initial value in crc, accumulates a CRC32 value for +// unsigned 16-bit integer v. +// https://msdn.microsoft.com/en-us/library/bb531411(v=vs.100) +FORCE_INLINE uint32_t _mm_crc32_u16(uint32_t crc, uint16_t v) { -#if defined(__aarch64__) -#if __has_builtin(__builtin_popcountll) - return __builtin_popcountll(a); +#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) + __asm__ __volatile__("crc32ch %w[c], %w[c], %w[v]\n\t" + : [c] "+r"(crc) + : [v] "r"(v)); +#elif (__ARM_ARCH == 8) && defined(__ARM_FEATURE_CRC32) + crc = __crc32ch(crc, v); #else - return (int64_t) vaddlv_u8(vcnt_u8(vcreate_u8(a))); -#endif -#else - uint64_t count = 0; - uint8x8_t input_val, count8x8_val; - uint16x4_t count16x4_val; - uint32x2_t count32x2_val; - uint64x1_t count64x1_val; - - input_val = vld1_u8((uint8_t *) &a); - count8x8_val = vcnt_u8(input_val); - count16x4_val = vpaddl_u8(count8x8_val); - count32x2_val = vpaddl_u16(count16x4_val); - count64x1_val = vpaddl_u32(count32x2_val); - vst1_u64(&count, count64x1_val); - return count; + crc = _mm_crc32_u8(crc, v & 0xff); + crc = _mm_crc32_u8(crc, (v >> 8) & 0xff); #endif + return crc; } -// Macro: Transpose the 4x4 matrix formed by the 4 rows of single-precision -// (32-bit) floating-point elements in row0, row1, row2, and row3, and store the -// transposed matrix in these vectors (row0 now contains column 0, etc.). -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=MM_TRANSPOSE4_PS -#define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \ - do { \ - float32x4x2_t ROW01 = vtrnq_f32(row0, row1); \ - float32x4x2_t ROW23 = vtrnq_f32(row2, row3); \ - row0 = vcombine_f32(vget_low_f32(ROW01.val[0]), \ - vget_low_f32(ROW23.val[0])); \ - row1 = vcombine_f32(vget_low_f32(ROW01.val[1]), \ - vget_low_f32(ROW23.val[1])); \ - row2 = vcombine_f32(vget_high_f32(ROW01.val[0]), \ - vget_high_f32(ROW23.val[0])); \ - row3 = vcombine_f32(vget_high_f32(ROW01.val[1]), \ - vget_high_f32(ROW23.val[1])); \ - } while (0) - -/* Crypto Extensions */ - -#if defined(__ARM_FEATURE_CRYPTO) -// Wraps vmull_p64 -FORCE_INLINE uint64x2_t _sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b) +// Starting with the initial value in crc, accumulates a CRC32 value for +// unsigned 32-bit integer v. +// https://msdn.microsoft.com/en-us/library/bb531394(v=vs.100) +FORCE_INLINE uint32_t _mm_crc32_u32(uint32_t crc, uint32_t v) { - poly64_t a = vget_lane_p64(vreinterpret_p64_u64(_a), 0); - poly64_t b = vget_lane_p64(vreinterpret_p64_u64(_b), 0); - return vreinterpretq_u64_p128(vmull_p64(a, b)); -} -#else // ARMv7 polyfill -// ARMv7/some A64 lacks vmull_p64, but it has vmull_p8. -// -// vmull_p8 calculates 8 8-bit->16-bit polynomial multiplies, but we need a -// 64-bit->128-bit polynomial multiply. -// -// It needs some work and is somewhat slow, but it is still faster than all -// known scalar methods. -// -// Algorithm adapted to C from -// https://www.workofard.com/2017/07/ghash-for-low-end-cores/, which is adapted -// from "Fast Software Polynomial Multiplication on ARM Processors Using the -// NEON Engine" by Danilo Camara, Conrado Gouvea, Julio Lopez and Ricardo Dahab -// (https://hal.inria.fr/hal-01506572) -static uint64x2_t _sse2neon_vmull_p64(uint64x1_t _a, uint64x1_t _b) -{ - poly8x8_t a = vreinterpret_p8_u64(_a); - poly8x8_t b = vreinterpret_p8_u64(_b); - - // Masks - uint8x16_t k48_32 = vcombine_u8(vcreate_u8(0x0000ffffffffffff), - vcreate_u8(0x00000000ffffffff)); - uint8x16_t k16_00 = vcombine_u8(vcreate_u8(0x000000000000ffff), - vcreate_u8(0x0000000000000000)); - - // Do the multiplies, rotating with vext to get all combinations - uint8x16_t d = vreinterpretq_u8_p16(vmull_p8(a, b)); // D = A0 * B0 - uint8x16_t e = - vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 1))); // E = A0 * B1 - uint8x16_t f = - vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 1), b)); // F = A1 * B0 - uint8x16_t g = - vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 2))); // G = A0 * B2 - uint8x16_t h = - vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 2), b)); // H = A2 * B0 - uint8x16_t i = - vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 3))); // I = A0 * B3 - uint8x16_t j = - vreinterpretq_u8_p16(vmull_p8(vext_p8(a, a, 3), b)); // J = A3 * B0 - uint8x16_t k = - vreinterpretq_u8_p16(vmull_p8(a, vext_p8(b, b, 4))); // L = A0 * B4 - - // Add cross products - uint8x16_t l = veorq_u8(e, f); // L = E + F - uint8x16_t m = veorq_u8(g, h); // M = G + H - uint8x16_t n = veorq_u8(i, j); // N = I + J - - // Interleave. Using vzip1 and vzip2 prevents Clang from emitting TBL - // instructions. -#if defined(__aarch64__) - uint8x16_t lm_p0 = vreinterpretq_u8_u64( - vzip1q_u64(vreinterpretq_u64_u8(l), vreinterpretq_u64_u8(m))); - uint8x16_t lm_p1 = vreinterpretq_u8_u64( - vzip2q_u64(vreinterpretq_u64_u8(l), vreinterpretq_u64_u8(m))); - uint8x16_t nk_p0 = vreinterpretq_u8_u64( - vzip1q_u64(vreinterpretq_u64_u8(n), vreinterpretq_u64_u8(k))); - uint8x16_t nk_p1 = vreinterpretq_u8_u64( - vzip2q_u64(vreinterpretq_u64_u8(n), vreinterpretq_u64_u8(k))); +#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) + __asm__ __volatile__("crc32cw %w[c], %w[c], %w[v]\n\t" + : [c] "+r"(crc) + : [v] "r"(v)); +#elif (__ARM_ARCH == 8) && defined(__ARM_FEATURE_CRC32) + crc = __crc32cw(crc, v); #else - uint8x16_t lm_p0 = vcombine_u8(vget_low_u8(l), vget_low_u8(m)); - uint8x16_t lm_p1 = vcombine_u8(vget_high_u8(l), vget_high_u8(m)); - uint8x16_t nk_p0 = vcombine_u8(vget_low_u8(n), vget_low_u8(k)); - uint8x16_t nk_p1 = vcombine_u8(vget_high_u8(n), vget_high_u8(k)); + crc = _mm_crc32_u16(crc, v & 0xffff); + crc = _mm_crc32_u16(crc, (v >> 16) & 0xffff); #endif - // t0 = (L) (P0 + P1) << 8 - // t1 = (M) (P2 + P3) << 16 - uint8x16_t t0t1_tmp = veorq_u8(lm_p0, lm_p1); - uint8x16_t t0t1_h = vandq_u8(lm_p1, k48_32); - uint8x16_t t0t1_l = veorq_u8(t0t1_tmp, t0t1_h); - - // t2 = (N) (P4 + P5) << 24 - // t3 = (K) (P6 + P7) << 32 - uint8x16_t t2t3_tmp = veorq_u8(nk_p0, nk_p1); - uint8x16_t t2t3_h = vandq_u8(nk_p1, k16_00); - uint8x16_t t2t3_l = veorq_u8(t2t3_tmp, t2t3_h); - - // De-interleave -#if defined(__aarch64__) - uint8x16_t t0 = vreinterpretq_u8_u64( - vuzp1q_u64(vreinterpretq_u64_u8(t0t1_l), vreinterpretq_u64_u8(t0t1_h))); - uint8x16_t t1 = vreinterpretq_u8_u64( - vuzp2q_u64(vreinterpretq_u64_u8(t0t1_l), vreinterpretq_u64_u8(t0t1_h))); - uint8x16_t t2 = vreinterpretq_u8_u64( - vuzp1q_u64(vreinterpretq_u64_u8(t2t3_l), vreinterpretq_u64_u8(t2t3_h))); - uint8x16_t t3 = vreinterpretq_u8_u64( - vuzp2q_u64(vreinterpretq_u64_u8(t2t3_l), vreinterpretq_u64_u8(t2t3_h))); -#else - uint8x16_t t1 = vcombine_u8(vget_high_u8(t0t1_l), vget_high_u8(t0t1_h)); - uint8x16_t t0 = vcombine_u8(vget_low_u8(t0t1_l), vget_low_u8(t0t1_h)); - uint8x16_t t3 = vcombine_u8(vget_high_u8(t2t3_l), vget_high_u8(t2t3_h)); - uint8x16_t t2 = vcombine_u8(vget_low_u8(t2t3_l), vget_low_u8(t2t3_h)); -#endif - // Shift the cross products - uint8x16_t t0_shift = vextq_u8(t0, t0, 15); // t0 << 8 - uint8x16_t t1_shift = vextq_u8(t1, t1, 14); // t1 << 16 - uint8x16_t t2_shift = vextq_u8(t2, t2, 13); // t2 << 24 - uint8x16_t t3_shift = vextq_u8(t3, t3, 12); // t3 << 32 - - // Accumulate the products - uint8x16_t cross1 = veorq_u8(t0_shift, t1_shift); - uint8x16_t cross2 = veorq_u8(t2_shift, t3_shift); - uint8x16_t mix = veorq_u8(d, cross1); - uint8x16_t r = veorq_u8(mix, cross2); - return vreinterpretq_u64_u8(r); + return crc; } -#endif // ARMv7 polyfill -// Perform a carry-less multiplication of two 64-bit integers, selected from a -// and b according to imm8, and store the results in dst. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_clmulepi64_si128 -FORCE_INLINE __m128i _mm_clmulepi64_si128(__m128i _a, __m128i _b, const int imm) +// Starting with the initial value in crc, accumulates a CRC32 value for +// unsigned 64-bit integer v. +// https://msdn.microsoft.com/en-us/library/bb514033(v=vs.100) +FORCE_INLINE uint64_t _mm_crc32_u64(uint64_t crc, uint64_t v) { - uint64x2_t a = vreinterpretq_u64_m128i(_a); - uint64x2_t b = vreinterpretq_u64_m128i(_b); - switch (imm & 0x11) { - case 0x00: - return vreinterpretq_m128i_u64( - _sse2neon_vmull_p64(vget_low_u64(a), vget_low_u64(b))); - case 0x01: - return vreinterpretq_m128i_u64( - _sse2neon_vmull_p64(vget_high_u64(a), vget_low_u64(b))); - case 0x10: - return vreinterpretq_m128i_u64( - _sse2neon_vmull_p64(vget_low_u64(a), vget_high_u64(b))); - case 0x11: - return vreinterpretq_m128i_u64( - _sse2neon_vmull_p64(vget_high_u64(a), vget_high_u64(b))); - default: - abort(); +#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) + __asm__ __volatile__("crc32cx %w[c], %w[c], %x[v]\n\t" + : [c] "+r"(crc) + : [v] "r"(v)); +#else + crc = _mm_crc32_u32((uint32_t) (crc), v & 0xffffffff); + crc = _mm_crc32_u32((uint32_t) (crc), (v >> 32) & 0xffffffff); +#endif + return crc; +} + +// Starting with the initial value in crc, accumulates a CRC32 value for +// unsigned 8-bit integer v. +// https://msdn.microsoft.com/en-us/library/bb514036(v=vs.100) +FORCE_INLINE uint32_t _mm_crc32_u8(uint32_t crc, uint8_t v) +{ +#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) + __asm__ __volatile__("crc32cb %w[c], %w[c], %w[v]\n\t" + : [c] "+r"(crc) + : [v] "r"(v)); +#elif (__ARM_ARCH == 8) && defined(__ARM_FEATURE_CRC32) + crc = __crc32cb(crc, v); +#else + crc ^= v; + for (int bit = 0; bit < 8; bit++) { + if (crc & 1) + crc = (crc >> 1) ^ UINT32_C(0x82f63b78); + else + crc = (crc >> 1); } +#endif + return crc; } +/* AES */ + #if !defined(__ARM_FEATURE_CRYPTO) /* clang-format off */ #define SSE2NEON_AES_DATA(w) \ @@ -7238,7 +8517,7 @@ FORCE_INLINE __m128i _mm_aesenc_si128(__m128i EncBlock, __m128i RoundKey) v = vqtbx4q_u8(v, _sse2neon_vld1q_u8_x4(SSE2NEON_sbox + 0xc0), w - 0xc0); // mix columns - w = (v << 1) ^ (uint8x16_t)(((int8x16_t) v >> 7) & 0x1b); + w = (v << 1) ^ (uint8x16_t) (((int8x16_t) v >> 7) & 0x1b); w ^= (uint8x16_t) vrev32q_u16((uint16x8_t) v); w ^= vqtbl1q_u8(v ^ w, vld1q_u8(ror32by8)); @@ -7246,9 +8525,9 @@ FORCE_INLINE __m128i _mm_aesenc_si128(__m128i EncBlock, __m128i RoundKey) return vreinterpretq_m128i_u8(w) ^ RoundKey; #else /* ARMv7-A NEON implementation */ -#define SSE2NEON_AES_B2W(b0, b1, b2, b3) \ - (((uint32_t)(b3) << 24) | ((uint32_t)(b2) << 16) | ((uint32_t)(b1) << 8) | \ - (b0)) +#define SSE2NEON_AES_B2W(b0, b1, b2, b3) \ + (((uint32_t) (b3) << 24) | ((uint32_t) (b2) << 16) | \ + ((uint32_t) (b1) << 8) | (uint32_t) (b0)) #define SSE2NEON_AES_F2(x) ((x << 1) ^ (((x >> 7) & 1) * 0x011b /* WPOLY */)) #define SSE2NEON_AES_F3(x) (SSE2NEON_AES_F2(x) ^ x) #define SSE2NEON_AES_U0(p) \ @@ -7299,22 +8578,22 @@ FORCE_INLINE __m128i _mm_aesenclast_si128(__m128i a, __m128i RoundKey) { /* FIXME: optimized for NEON */ uint8_t v[4][4] = { - [0] = {SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 0)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 5)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 10)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 15)]}, - [1] = {SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 4)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 9)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 14)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 3)]}, - [2] = {SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 8)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 13)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 2)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 7)]}, - [3] = {SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 12)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 1)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 6)], - SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 11)]}, + {SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 0)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 5)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 10)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 15)]}, + {SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 4)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 9)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 14)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 3)]}, + {SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 8)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 13)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 2)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 7)]}, + {SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 12)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 1)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 6)], + SSE2NEON_sbox[vreinterpretq_nth_u8_m128i(a, 11)]}, }; for (int i = 0; i < 16; i++) vreinterpretq_nth_u8_m128i(a, i) = @@ -7380,211 +8659,135 @@ FORCE_INLINE __m128i _mm_aeskeygenassist_si128(__m128i a, const int rcon) } #endif -/* Streaming Extensions */ +/* Others */ -// Guarantees that every preceding store is globally visible before any -// subsequent store. -// https://msdn.microsoft.com/en-us/library/5h2w73d1%28v=vs.90%29.aspx -FORCE_INLINE void _mm_sfence(void) +// Perform a carry-less multiplication of two 64-bit integers, selected from a +// and b according to imm8, and store the results in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_clmulepi64_si128 +FORCE_INLINE __m128i _mm_clmulepi64_si128(__m128i _a, __m128i _b, const int imm) { - __sync_synchronize(); -} - -// Store 64-bits of integer data from a into memory using a non-temporal memory -// hint. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_pi -FORCE_INLINE void _mm_stream_pi(__m64 *p, __m64 a) -{ - vst1_s64((int64_t *) p, vreinterpret_s64_m64(a)); -} - -// Store 128-bits (composed of 4 packed single-precision (32-bit) floating- -// point elements) from a into memory using a non-temporal memory hint. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_ps -FORCE_INLINE void _mm_stream_ps(float *p, __m128 a) -{ -#if __has_builtin(__builtin_nontemporal_store) - __builtin_nontemporal_store(a, (float32x4_t *) p); -#else - vst1q_f32(p, vreinterpretq_f32_m128(a)); -#endif -} - -// Store 128-bits (composed of 2 packed double-precision (64-bit) floating-point -// elements) from a into memory using a non-temporal memory hint. mem_addr must -// be aligned on a 16-byte boundary or a general-protection exception may be -// generated. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_pd -FORCE_INLINE void _mm_stream_pd(double *p, __m128d a) -{ -#if __has_builtin(__builtin_nontemporal_store) - __builtin_nontemporal_store(a, (float32x4_t *) p); -#elif defined(__aarch64__) - vst1q_f64(p, vreinterpretq_f64_m128d(a)); -#else - vst1q_s64((int64_t *) p, vreinterpretq_s64_m128d(a)); -#endif -} - -// Stores the data in a to the address p without polluting the caches. If the -// cache line containing address p is already in the cache, the cache will be -// updated. -// https://msdn.microsoft.com/en-us/library/ba08y07y%28v=vs.90%29.aspx -FORCE_INLINE void _mm_stream_si128(__m128i *p, __m128i a) -{ -#if __has_builtin(__builtin_nontemporal_store) - __builtin_nontemporal_store(a, p); -#else - vst1q_s64((int64_t *) p, vreinterpretq_s64_m128i(a)); -#endif -} - -// Store 32-bit integer a into memory using a non-temporal hint to minimize -// cache pollution. If the cache line containing address mem_addr is already in -// the cache, the cache will be updated. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_si32 -FORCE_INLINE void _mm_stream_si32(int *p, int a) -{ - vst1q_lane_s32((int32_t *) p, vdupq_n_s32(a), 0); -} - -// Load 128-bits of integer data from memory into dst using a non-temporal -// memory hint. mem_addr must be aligned on a 16-byte boundary or a -// general-protection exception may be generated. -// -// dst[127:0] := MEM[mem_addr+127:mem_addr] -// -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_stream_load_si128 -FORCE_INLINE __m128i _mm_stream_load_si128(__m128i *p) -{ -#if __has_builtin(__builtin_nontemporal_store) - return __builtin_nontemporal_load(p); -#else - return vreinterpretq_m128i_s64(vld1q_s64((int64_t *) p)); -#endif -} - -// Cache line containing p is flushed and invalidated from all caches in the -// coherency domain. : -// https://msdn.microsoft.com/en-us/library/ba08y07y(v=vs.100).aspx -FORCE_INLINE void _mm_clflush(void const *p) -{ - (void) p; - // no corollary for Neon? -} - -// Allocate aligned blocks of memory. -// https://software.intel.com/en-us/ -// cpp-compiler-developer-guide-and-reference-allocating-and-freeing-aligned-memory-blocks -FORCE_INLINE void *_mm_malloc(size_t size, size_t align) -{ - void *ptr; - if (align == 1) - return malloc(size); - if (align == 2 || (sizeof(void *) == 8 && align == 4)) - align = sizeof(void *); - if (!posix_memalign(&ptr, align, size)) - return ptr; - return NULL; -} - -// Conditionally store 8-bit integer elements from a into memory using mask -// (elements are not stored when the highest bit is not set in the corresponding -// element) and a non-temporal memory hint. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_maskmove_si64 -FORCE_INLINE void _mm_maskmove_si64(__m64 a, __m64 mask, char *mem_addr) -{ - int8x8_t shr_mask = vshr_n_s8(vreinterpret_s8_m64(mask), 7); - __m128 b = _mm_load_ps((const float *) mem_addr); - int8x8_t masked = - vbsl_s8(vreinterpret_u8_s8(shr_mask), vreinterpret_s8_m64(a), - vreinterpret_s8_u64(vget_low_u64(vreinterpretq_u64_m128(b)))); - vst1_s8((int8_t *) mem_addr, masked); -} - -// Conditionally store 8-bit integer elements from a into memory using mask -// (elements are not stored when the highest bit is not set in the corresponding -// element) and a non-temporal memory hint. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_m_maskmovq -#define _m_maskmovq(a, mask, mem_addr) _mm_maskmove_si64(a, mask, mem_addr) - -// Free aligned memory that was allocated with _mm_malloc. -// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_free -FORCE_INLINE void _mm_free(void *addr) -{ - free(addr); -} - -// Starting with the initial value in crc, accumulates a CRC32 value for -// unsigned 8-bit integer v. -// https://msdn.microsoft.com/en-us/library/bb514036(v=vs.100) -FORCE_INLINE uint32_t _mm_crc32_u8(uint32_t crc, uint8_t v) -{ -#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) - __asm__ __volatile__("crc32cb %w[c], %w[c], %w[v]\n\t" - : [c] "+r"(crc) - : [v] "r"(v)); -#else - crc ^= v; - for (int bit = 0; bit < 8; bit++) { - if (crc & 1) - crc = (crc >> 1) ^ UINT32_C(0x82f63b78); - else - crc = (crc >> 1); + uint64x2_t a = vreinterpretq_u64_m128i(_a); + uint64x2_t b = vreinterpretq_u64_m128i(_b); + switch (imm & 0x11) { + case 0x00: + return vreinterpretq_m128i_u64( + _sse2neon_vmull_p64(vget_low_u64(a), vget_low_u64(b))); + case 0x01: + return vreinterpretq_m128i_u64( + _sse2neon_vmull_p64(vget_high_u64(a), vget_low_u64(b))); + case 0x10: + return vreinterpretq_m128i_u64( + _sse2neon_vmull_p64(vget_low_u64(a), vget_high_u64(b))); + case 0x11: + return vreinterpretq_m128i_u64( + _sse2neon_vmull_p64(vget_high_u64(a), vget_high_u64(b))); + default: + abort(); } -#endif - return crc; } -// Starting with the initial value in crc, accumulates a CRC32 value for -// unsigned 16-bit integer v. -// https://msdn.microsoft.com/en-us/library/bb531411(v=vs.100) -FORCE_INLINE uint32_t _mm_crc32_u16(uint32_t crc, uint16_t v) +FORCE_INLINE unsigned int _sse2neon_mm_get_denormals_zero_mode() { -#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) - __asm__ __volatile__("crc32ch %w[c], %w[c], %w[v]\n\t" - : [c] "+r"(crc) - : [v] "r"(v)); + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; #else - crc = _mm_crc32_u8(crc, v & 0xff); - crc = _mm_crc32_u8(crc, (v >> 8) & 0xff); + uint32_t value; #endif - return crc; + } r; + +#if defined(__aarch64__) + __asm__ __volatile__("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + __asm__ __volatile__("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + return r.field.bit24 ? _MM_DENORMALS_ZERO_ON : _MM_DENORMALS_ZERO_OFF; } -// Starting with the initial value in crc, accumulates a CRC32 value for -// unsigned 32-bit integer v. -// https://msdn.microsoft.com/en-us/library/bb531394(v=vs.100) -FORCE_INLINE uint32_t _mm_crc32_u32(uint32_t crc, uint32_t v) +// Count the number of bits set to 1 in unsigned 32-bit integer a, and +// return that count in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_popcnt_u32 +FORCE_INLINE int _mm_popcnt_u32(unsigned int a) { -#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) - __asm__ __volatile__("crc32cw %w[c], %w[c], %w[v]\n\t" - : [c] "+r"(crc) - : [v] "r"(v)); +#if defined(__aarch64__) +#if __has_builtin(__builtin_popcount) + return __builtin_popcount(a); #else - crc = _mm_crc32_u16(crc, v & 0xffff); - crc = _mm_crc32_u16(crc, (v >> 16) & 0xffff); + return (int) vaddlv_u8(vcnt_u8(vcreate_u8((uint64_t) a))); +#endif +#else + uint32_t count = 0; + uint8x8_t input_val, count8x8_val; + uint16x4_t count16x4_val; + uint32x2_t count32x2_val; + + input_val = vld1_u8((uint8_t *) &a); + count8x8_val = vcnt_u8(input_val); + count16x4_val = vpaddl_u8(count8x8_val); + count32x2_val = vpaddl_u16(count16x4_val); + + vst1_u32(&count, count32x2_val); + return count; #endif - return crc; } -// Starting with the initial value in crc, accumulates a CRC32 value for -// unsigned 64-bit integer v. -// https://msdn.microsoft.com/en-us/library/bb514033(v=vs.100) -FORCE_INLINE uint64_t _mm_crc32_u64(uint64_t crc, uint64_t v) +// Count the number of bits set to 1 in unsigned 64-bit integer a, and +// return that count in dst. +// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm_popcnt_u64 +FORCE_INLINE int64_t _mm_popcnt_u64(uint64_t a) { -#if defined(__aarch64__) && defined(__ARM_FEATURE_CRC32) - __asm__ __volatile__("crc32cx %w[c], %w[c], %x[v]\n\t" - : [c] "+r"(crc) - : [v] "r"(v)); +#if defined(__aarch64__) +#if __has_builtin(__builtin_popcountll) + return __builtin_popcountll(a); #else - crc = _mm_crc32_u32((uint32_t)(crc), v & 0xffffffff); - crc = _mm_crc32_u32((uint32_t)(crc), (v >> 32) & 0xffffffff); + return (int64_t) vaddlv_u8(vcnt_u8(vcreate_u8(a))); +#endif +#else + uint64_t count = 0; + uint8x8_t input_val, count8x8_val; + uint16x4_t count16x4_val; + uint32x2_t count32x2_val; + uint64x1_t count64x1_val; + + input_val = vld1_u8((uint8_t *) &a); + count8x8_val = vcnt_u8(input_val); + count16x4_val = vpaddl_u8(count8x8_val); + count32x2_val = vpaddl_u16(count16x4_val); + count64x1_val = vpaddl_u32(count32x2_val); + vst1_u64(&count, count64x1_val); + return count; #endif - return crc; } -FORCE_INLINE void _mm_empty (void) { } +FORCE_INLINE void _sse2neon_mm_set_denormals_zero_mode(unsigned int flag) +{ + // AArch32 Advanced SIMD arithmetic always uses the Flush-to-zero setting, + // regardless of the value of the FZ bit. + union { + fpcr_bitfield field; +#if defined(__aarch64__) + uint64_t value; +#else + uint32_t value; +#endif + } r; + +#if defined(__aarch64__) + __asm__ __volatile__("mrs %0, FPCR" : "=r"(r.value)); /* read */ +#else + __asm__ __volatile__("vmrs %0, FPSCR" : "=r"(r.value)); /* read */ +#endif + + r.field.bit24 = (flag & _MM_DENORMALS_ZERO_MASK) == _MM_DENORMALS_ZERO_ON; + +#if defined(__aarch64__) + __asm__ __volatile__("msr FPCR, %0" ::"r"(r)); /* write */ +#else + __asm__ __volatile__("vmsr FPSCR, %0" ::"r"(r)); /* write */ +#endif +} #if defined(__GNUC__) || defined(__clang__) #pragma pop_macro("ALIGN_STRUCT") diff --git a/engine/OcclusionSystem.cpp b/engine/OcclusionSystem.cpp index 86491631..1c3fad1c 100644 --- a/engine/OcclusionSystem.cpp +++ b/engine/OcclusionSystem.cpp @@ -1268,7 +1268,9 @@ void CEdgeList::CullSmallOccluders() // Sort the surfaces by screen area, in descending order int nSurfCount = m_Surfaces.Count(); s_pSortSurfaces = m_Surfaces.Base(); - qsort( m_SurfaceSort.Base(), nSurfCount, sizeof(int), SurfCompare ); + + if( m_SurfaceSort.Base() ) + qsort( m_SurfaceSort.Base(), nSurfCount, sizeof(int), SurfCompare ); // We're going to keep the greater of r_occludermin + All surfaces with a screen area >= r_occluderarea int nMinSurfaces = r_occludermincount.GetInt(); @@ -1282,7 +1284,7 @@ void CEdgeList::CullSmallOccluders() bool *bUseSurface = (bool*)stackalloc( nSurfCount * sizeof(bool) ); memset( bUseSurface, 0, nSurfCount * sizeof(bool) ); - + int i; for ( i = 0; i < nSurfCount; ++i ) { diff --git a/engine/gl_rsurf.cpp b/engine/gl_rsurf.cpp index 82ed85ac..836a3443 100644 --- a/engine/gl_rsurf.cpp +++ b/engine/gl_rsurf.cpp @@ -4099,7 +4099,9 @@ CBrushBatchRender::brushrender_t *CBrushBatchRender::FindOrCreateRenderBatch( mo surfaceList.Sort( SurfaceCmp ); renderT.pPlanes = new cplane_t *[planeList.Count()]; renderT.planeCount = planeList.Count(); - memcpy( renderT.pPlanes, planeList.Base(), sizeof(cplane_t *)*planeList.Count() ); + + if( planeList.Base() ) + memcpy( renderT.pPlanes, planeList.Base(), sizeof(cplane_t *)*planeList.Count() ); renderT.pSurfaces = new brushrendersurface_t[surfaceList.Count()]; renderT.surfaceCount = surfaceList.Count(); diff --git a/game/client/c_rumble.cpp b/game/client/c_rumble.cpp index a22ef884..478cb853 100644 --- a/game/client/c_rumble.cpp +++ b/game/client/c_rumble.cpp @@ -166,7 +166,7 @@ void GenerateSquareWaveEffect( RumbleWaveform_t *pWaveform, const WaveGenParams_ while( i < NUM_WAVE_SAMPLES ) { - for( j = 0 ; j < steps ; j++ ) + for( j = 0 ; j < steps && i < NUM_WAVE_SAMPLES; j++ ) { if( params.leftChannel ) { @@ -177,7 +177,7 @@ void GenerateSquareWaveEffect( RumbleWaveform_t *pWaveform, const WaveGenParams_ pWaveform->amplitude_right[i++] = params.minAmplitude; } } - for( j = 0 ; j < steps ; j++ ) + for( j = 0 ; j < steps && i < NUM_WAVE_SAMPLES; j++ ) { if( params.leftChannel ) { diff --git a/game/client/c_vote_controller.cpp b/game/client/c_vote_controller.cpp index c900c27e..52184061 100644 --- a/game/client/c_vote_controller.cpp +++ b/game/client/c_vote_controller.cpp @@ -33,7 +33,7 @@ void C_VoteController::RecvProxy_VoteType( const CRecvProxyData *pData, void *pS if( pMe->m_iActiveIssueIndex == pData->m_Value.m_Int ) return; - pMe->m_iActiveIssueIndex = pData->m_Value.m_Int; + memcpy( &pMe->m_iActiveIssueIndex, &pData->m_Value.m_Int, sizeof(pData->m_Value.m_Int) ); pMe->m_bTypeDirty = true; // Since the contents of a new vote are in three parts, we can't directly send an event to the Hud @@ -186,4 +186,4 @@ void C_VoteController::FireGameEvent( IGameEvent *event ) } } } -} \ No newline at end of file +} diff --git a/game/client/client_hl1mp.vpc b/game/client/client_hl1mp.vpc new file mode 100644 index 00000000..ef5eddf4 --- /dev/null +++ b/game/client/client_hl1mp.vpc @@ -0,0 +1,109 @@ +//----------------------------------------------------------------------------- +// CLIENT_HL1MP.VPC +// +// Project Script +//----------------------------------------------------------------------------- + +$Macro SRCDIR "..\.." +$Macro GAMENAME "hl1mp" + +$Include "$SRCDIR\game\client\client_base.vpc" + +$Configuration +{ + $Compiler + { + $AdditionalIncludeDirectories "$BASE;.\hl1,.\hl2,.\hl2\elements,$SRCDIR\game\shared\hl1,$SRCDIR\game\shared\hl2" + $PreprocessorDefinitions "$BASE;HL1_CLIENT_DLL;HL1MP_CLIENT_DLL" + } +} + +$Project "Client (HL1MP)" +{ + $Folder "Source Files" + { + -$File "geiger.cpp" + -$File "history_resource.cpp" + -$File "train.cpp" + + $File "c_team_objectiveresource.cpp" + $File "c_team_objectiveresource.h" + $File "hud_chat.cpp" + $File "$SRCDIR\game\shared\predicted_viewmodel.cpp" + $File "$SRCDIR\game\shared\predicted_viewmodel.h" + + $Folder "HL2 DLL" + { + $File "hl2\c_antlion_dust.cpp" + $File "hl2\c_basehelicopter.cpp" + $File "hl2\c_basehelicopter.h" + $File "hl2\c_basehlcombatweapon.h" + $File "hl2\c_corpse.cpp" + $File "hl2\c_corpse.h" + $File "hl2\c_hl2_playerlocaldata.h" + $File "hl2\c_rotorwash.cpp" + $File "$SRCDIR\game\shared\hl2\citadel_effects_shared.h" + $File "$SRCDIR\game\shared\hl2\env_headcrabcanister_shared.h" + $File "hl2\fx_bugbait.cpp" + $File "$SRCDIR\game\shared\hl2\hl2_shareddefs.h" + $File "hl2\hl_in_main.cpp" + $File "hl2\hl_prediction.cpp" + $File "hl2\vgui_rootpanel_hl2.cpp" + } + + $Folder "HL1 DLL" + { + $File "hl1\hl1_c_legacytempents.cpp" + $File "hl1\hl1_c_player.cpp" + $File "hl1\hl1_c_player.h" + $File "hl1\hl1_c_rpg_rocket.cpp" + $File "hl1\hl1_c_weapon__stubs.cpp" + $File "hl1\hl1_clientmode.cpp" + $File "hl1\hl1_clientmode.h" + $File "hl1\hl1_clientscoreboard.cpp" + $File "hl1\hl1_hud_deathnotice.cpp" + $File "hl1\hl1_fx_gauss.cpp" + $File "hl1\hl1_fx_gibs.cpp" + $File "hl1\hl1_fx_impacts.cpp" + $File "hl1\hl1_fx_shelleject.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_gamemovement.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_gamemovement.h" + $File "hl1\hl1_hud_ammo.cpp" + $File "hl1\hl1_hud_battery.cpp" + $File "hl1\hl1_hud_damageindicator.cpp" + $File "hl1\hl1_hud_damagetiles.cpp" + $File "hl1\hl1_hud_flashlight.cpp" + $File "hl1\hl1_hud_geiger.cpp" + $File "hl1\hl1_hud_health.cpp" + $File "hl1\hl1_hud_history_resource.cpp" + $File "hl1\hl1_hud_numbers.cpp" + $File "hl1\hl1_hud_numbers.h" + $File "hl1\hl1_hud_train.cpp" + $File "hl1\hl1_hud_weaponselection.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_basecombatweapon_shared.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_gamerules.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_player_shared.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_player_shared.h" + $File "$SRCDIR\game\shared\hl1\hl1_usermessages.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_357.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_crossbow.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_egon.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_gauss.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_glock.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_handgrenade.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_hornetgun.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_mp5.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_rpg.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_sachel.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_shotgun.cpp" + $File "$SRCDIR\game\server\hl1\hl1_weapon_crowbar.cpp" + } + + $Folder "HL1MP DLL" + { + $File "hl1\c_hl1mp_player.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_basecombatweapon_shared.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_gamerules.cpp" + } + } +} diff --git a/game/client/hud_vote.h b/game/client/hud_vote.h index 91ba1e9b..33df877d 100644 --- a/game/client/hud_vote.h +++ b/game/client/hud_vote.h @@ -119,6 +119,9 @@ class CHudVote : public vgui::EditablePanel, public CHudElement { DECLARE_CLASS_SIMPLE( CHudVote, vgui::EditablePanel ); +public: + DECLARE_MULTIPLY_INHERITED(); + CHudVote( const char *pElementName ); virtual void LevelInit( void ); diff --git a/game/client/viewrender.cpp b/game/client/viewrender.cpp index 0240c238..f6193ef4 100644 --- a/game/client/viewrender.cpp +++ b/game/client/viewrender.cpp @@ -4010,7 +4010,7 @@ void CRendering3dView::DrawOpaqueRenderables( ERenderDepthMode DepthMode ) } } - if ( 0 && r_threaded_renderables.GetBool() ) + if ( r_threaded_renderables.GetBool() ) { ParallelProcess( "BoneSetupNpcsLast", arrBoneSetupNpcsLast.Base() + numOpaqueEnts - numNpcs, numNpcs, &SetupBonesOnBaseAnimating ); ParallelProcess( "BoneSetupNpcsLast NonNPCs", arrBoneSetupNpcsLast.Base(), numNonNpcsAnimating, &SetupBonesOnBaseAnimating ); diff --git a/game/client/wscript b/game/client/wscript index 56373a8a..3732c8b4 100755 --- a/game/client/wscript +++ b/game/client/wscript @@ -17,7 +17,7 @@ games = { 'hl1': ['client_base.vpc', 'client_hl1.vpc'], 'episodic': ['client_base.vpc', 'client_episodic.vpc'], 'portal': ['client_base.vpc', 'client_portal.vpc'], - 'hl1mp': ['client_base.vpc', 'client_hl1.vpc'], + 'hl1mp': ['client_base.vpc', 'client_hl1mp.vpc'], 'cstrike': ['client_base.vpc', 'client_cstrike.vpc'], 'dod': ['client_base.vpc', 'client_dod.vpc'] } diff --git a/game/server/AI_Criteria.h b/game/server/AI_Criteria.h index 72c54080..e9a2c13f 100644 --- a/game/server/AI_Criteria.h +++ b/game/server/AI_Criteria.h @@ -93,7 +93,7 @@ private: CUtlRBTree< CritEntry_t, short > m_Lookup; }; -#pragma pack(1) +//#pragma pack(1) template struct response_interval_t { @@ -150,7 +150,7 @@ struct AI_ResponseParams responseparams_interval_t predelay; //21 }; -#pragma pack() +//#pragma pack() //----------------------------------------------------------------------------- // Purpose: Generic container for a response to a match to a criteria set diff --git a/game/server/AI_ResponseSystem.cpp b/game/server/AI_ResponseSystem.cpp index 510a1d4e..7b739679 100644 --- a/game/server/AI_ResponseSystem.cpp +++ b/game/server/AI_ResponseSystem.cpp @@ -44,7 +44,6 @@ inline static char *CopyString( const char *in ) return out; } -#pragma pack(1) class Matcher { public: @@ -542,7 +541,6 @@ struct Rule bool m_bMatchOnce : 1; bool m_bEnabled : 1; }; -#pragma pack() //----------------------------------------------------------------------------- // Purpose: diff --git a/game/server/server_hl1mp.vpc b/game/server/server_hl1mp.vpc new file mode 100644 index 00000000..5692cbda --- /dev/null +++ b/game/server/server_hl1mp.vpc @@ -0,0 +1,179 @@ +//----------------------------------------------------------------------------- +// SERVER_HL1MP.VPC +// +// Project Script +//----------------------------------------------------------------------------- + +$Macro SRCDIR "..\.." +$Macro GAMENAME "hl1mp" + +$Include "$SRCDIR\game\server\server_base.vpc" + +$Configuration +{ + $Compiler + { + $AdditionalIncludeDirectories "$BASE;$SRCDIR\game\shared\hl1,$SRCDIR\game\shared\hl2,.\hl1,.\hl2" + $PreprocessorDefinitions "$BASE;HL1_DLL;HL1MP_DLL" + } +} + +$Project "Server (HL1MP)" +{ + $Folder "Source Files" + { + $File "hl1\hl1mp_gameinterface.cpp" + $File "basegrenade_concussion.cpp" + $File "basegrenade_contact.cpp" + $File "basegrenade_timed.cpp" + $File "hl2\Func_Monitor.cpp" + $File "GrenadeThrown.cpp" + $File "GrenadeThrown.h" + $File "h_cycler.cpp" + $File "$SRCDIR\game\shared\predicted_viewmodel.cpp" + $File "$SRCDIR\game\shared\predicted_viewmodel.h" + $File "$SRCDIR\game\shared\hl2\survival_gamerules.cpp" + $File "team_spawnpoint.cpp" + $File "team_spawnpoint.h" + $File "$SRCDIR\game\shared\weapon_parse_default.cpp" + + $Folder "HL2 DLL" + { + $File "hl2\ai_behavior_police.h" + $File "hl2\ai_goal_police.h" + $File "hl2\ai_interactions.h" + $File "hl2\antlion_maker.h" + $File "hl2\CBaseSpriteProjectile.cpp" + $File "hl2\CBaseSpriteProjectile.h" + $File "$SRCDIR\game\shared\hl2\citadel_effects_shared.h" + $File "hl2\energy_wave.h" + $File "$SRCDIR\game\shared\hl2\env_alyxemp_shared.h" + $File "$SRCDIR\game\shared\hl2\hl2_shareddefs.h" + $File "$SRCDIR\game\shared\hl2\hl_movedata.h" + $File "hl2\look_door.cpp" + $File "hl2\monster_dummy.cpp" + $File "hl2\npc_metropolice.h" + $File "hl2\npc_playercompanion.h" + $File "npc_Talker.cpp" + $File "npc_Talker.h" + $File "hl2\prop_combine_ball.h" + $File "hl2\script_intro.h" + $File "hl2\vehicle_crane.h" + $File "hl2\weapon_crowbar.h" + $File "hl2\weapon_physcannon.h" + $File "hl2\weapon_stunstick.h" + + $Folder "unused" + { + $File "hl2\grenade_beam.cpp" + $File "hl2\grenade_beam.h" + $File "hl2\grenade_homer.cpp" + $File "hl2\grenade_homer.h" + } + } + + $Folder "HL1 DLL" + { + $File "actanimating.cpp" + $File "actanimating.h" + $File "hl1\hl1_ai_basenpc.cpp" + $File "hl1\hl1_ai_basenpc.h" + $File "hl1\hl1_basecombatweapon.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_basecombatweapon_shared.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_basecombatweapon_shared.h" + $File "hl1\hl1_basegrenade.cpp" + $File "hl1\hl1_basegrenade.h" + $File "hl1_CBaseHelicopter.h" + $File "hl1\hl1_client.cpp" + $File "hl1\hl1_ents.cpp" + $File "hl1\hl1_env_speaker.cpp" + $File "hl1\hl1_eventlog.cpp" + $File "hl1\hl1_func_recharge.cpp" + $File "hl1\hl1_func_tank.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_gamemovement.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_gamemovement.h" + $File "$SRCDIR\game\shared\hl1\hl1_gamerules.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_gamerules.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_gamerules.h" + $File "hl1\hl1_grenade_mp5.cpp" + $File "hl1\hl1_grenade_mp5.h" + $File "hl1\hl1_grenade_spit.cpp" + $File "hl1\hl1_grenade_spit.h" + $File "hl1\hl1_item_ammo.cpp" + $File "hl1\hl1_item_battery.cpp" + $File "hl1\hl1_item_healthkit.cpp" + $File "hl1\hl1_item_longjump.cpp" + $File "hl1\hl1_item_suit.cpp" + $File "hl1\hl1_items.cpp" + $File "hl1\hl1_items.h" + $File "hl1\hl1_monstermaker.cpp" + $File "hl1\hl1_monstermaker.h" + $File "hl1\hl1_npc_aflock.cpp" + $File "hl1\hl1_npc_agrunt.cpp" + $File "hl1\hl1_npc_apache.cpp" + $File "hl1\hl1_npc_barnacle.cpp" + $File "hl1\hl1_npc_barnacle.h" + $File "hl1\hl1_npc_barney.cpp" + $File "hl1\hl1_npc_barney.h" + $File "hl1\hl1_npc_bigmomma.cpp" + $File "hl1\hl1_npc_bloater.cpp" + $File "hl1\hl1_npc_bullsquid.cpp" + $File "hl1\hl1_npc_bullsquid.h" + $File "hl1\hl1_npc_controller.cpp" + $File "hl1\hl1_npc_gargantua.cpp" + $File "hl1\hl1_npc_gargantua.h" + $File "hl1\hl1_npc_gman.cpp" + $File "hl1\hl1_npc_hassassin.cpp" + $File "hl1\hl1_npc_headcrab.cpp" + $File "hl1\hl1_npc_headcrab.h" + $File "hl1\hl1_npc_hgrunt.cpp" + $File "hl1\hl1_npc_hgrunt.h" + $File "hl1\hl1_npc_hornet.cpp" + $File "hl1\hl1_npc_hornet.h" + $File "hl1\hl1_npc_houndeye.cpp" + $File "hl1\hl1_npc_houndeye.h" + $File "hl1\hl1_npc_ichthyosaur.cpp" + $File "hl1\hl1_npc_ichthyosaur.h" + $File "hl1\hl1_npc_leech.cpp" + $File "hl1\hl1_npc_nihilanth.cpp" + $File "hl1\hl1_npc_osprey.cpp" + $File "hl1\hl1_npc_roach.cpp" + $File "hl1\hl1_npc_scientist.cpp" + $File "hl1\hl1_npc_scientist.h" + $File "hl1\hl1_npc_snark.cpp" + $File "hl1\hl1_npc_snark.h" + $File "hl1\hl1_npc_talker.cpp" + $File "hl1\hl1_npc_talker.h" + $File "hl1\hl1_npc_tentacle.cpp" + $File "hl1\hl1_npc_turret.cpp" + $File "hl1\hl1_npc_vortigaunt.cpp" + $File "hl1\hl1_npc_vortigaunt.h" + $File "hl1\hl1_npc_zombie.cpp" + $File "hl1\hl1_npc_zombie.h" + $File "hl1\hl1_player.cpp" + $File "hl1\hl1_player.h" + $File "$SRCDIR\game\shared\hl1\hl1_player_shared.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_player_shared.h" + $File "hl1\hl1_playermove.cpp" + $File "$SRCDIR\game\shared\hl1\hl1_usermessages.cpp" + $File "hl1\hl1_weapon_snark.cpp" + $File "hl1\hl1_weapon_tripmine.cpp" + $File "hl1\hl1_weaponbox.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_basecombatweapon_shared.cpp" + $File "hl1\hl1mp_bot_temp.cpp" + $File "hl1\hl1mp_player.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_357.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_crossbow.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_egon.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_gauss.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_glock.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_handgrenade.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_hornetgun.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_mp5.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_rpg.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_sachel.cpp" + $File "$SRCDIR\game\shared\hl1\hl1mp_weapon_shotgun.cpp" + $File "$SRCDIR\game\server\hl1\hl1_weapon_crowbar.cpp" + } + } +} diff --git a/game/server/wscript b/game/server/wscript index 848f9c40..21e4e4ee 100755 --- a/game/server/wscript +++ b/game/server/wscript @@ -14,7 +14,7 @@ games = { 'hl2mp': ['server_base.vpc', 'server_hl2mp.vpc'], 'portal': ['server_base.vpc', 'server_portal.vpc'], 'hl1': ['server_base.vpc', 'server_hl1.vpc'], - 'hl1mp': ['server_base.vpc', 'server_hl1.vpc'], + 'hl1mp': ['server_base.vpc', 'server_hl1mp.vpc'], 'cstrike': ['server_base.vpc', 'server_cstrike.vpc', 'nav_mesh.vpc'], 'dod': ['server_base.vpc', 'server_dod.vpc'], 'tf': [ diff --git a/game/shared/saverestore.cpp b/game/shared/saverestore.cpp index bd1a7bef..387aefc2 100644 --- a/game/shared/saverestore.cpp +++ b/game/shared/saverestore.cpp @@ -203,6 +203,7 @@ CSave::CSave( CSaveRestoreData *pdata ) inline int CSave::DataEmpty( const char *pdata, int size ) { + static int void_data = 0; if ( size != 4 ) { const char *pLimit = pdata + size; @@ -214,7 +215,7 @@ inline int CSave::DataEmpty( const char *pdata, int size ) return 1; } - return ( *((int *)pdata) == 0 ); + return memcmp(pdata, &void_data, sizeof(int)) == 0; } //----------------------------------------------------------------------------- diff --git a/gameui/BaseSaveGameDialog.cpp b/gameui/BaseSaveGameDialog.cpp index 99f7fc1c..c59d18ce 100644 --- a/gameui/BaseSaveGameDialog.cpp +++ b/gameui/BaseSaveGameDialog.cpp @@ -560,7 +560,7 @@ int SaveReadNameAndComment( FileHandle_t f, OUT_Z_CAP(nameSize) char *name, int int nNumberOfFields; char *pData; - int nFieldSize; + short nFieldSize; pData = pSaveData; @@ -580,9 +580,12 @@ int SaveReadNameAndComment( FileHandle_t f, OUT_Z_CAP(nameSize) char *name, int pTokenList = NULL; // short, short (size, index of field name) - nFieldSize = *(short *)pData; + memcpy( &nFieldSize, pData, sizeof(short) ); + pData += sizeof(short); - pFieldName = pTokenList[ *(short *)pData ]; + short index; + memcpy( &index, pData, sizeof(short) ); + pFieldName = pTokenList[index]; if (stricmp(pFieldName, "GameHeader")) { @@ -592,7 +595,7 @@ int SaveReadNameAndComment( FileHandle_t f, OUT_Z_CAP(nameSize) char *name, int // int (fieldcount) pData += sizeof(short); - nNumberOfFields = *(int*)pData; + memcpy( &nNumberOfFields, pData, sizeof(int) ); pData += nFieldSize; // Each field is a short (size), short (index of name), binary string of "size" bytes (data) @@ -603,10 +606,12 @@ int SaveReadNameAndComment( FileHandle_t f, OUT_Z_CAP(nameSize) char *name, int // szName // Actual Data - nFieldSize = *(short *)pData; + memcpy( &nFieldSize, pData, sizeof(short) ); pData += sizeof(short); - pFieldName = pTokenList[ *(short *)pData ]; + short index; + memcpy( &index, pData, sizeof(short)); + pFieldName = pTokenList[index]; pData += sizeof(short); if (!stricmp(pFieldName, "comment")) diff --git a/public/bone_setup.cpp b/public/bone_setup.cpp index 35d72d5e..564f9ea6 100644 --- a/public/bone_setup.cpp +++ b/public/bone_setup.cpp @@ -21,6 +21,8 @@ #include "convar.h" #include "tier0/tslist.h" #include "vphysics_interface.h" +#include "mathlib/compressed_vector.h" + #ifdef CLIENT_DLL #include "posedebugger.h" #endif @@ -378,14 +380,18 @@ void CalcBoneQuaternion( int frame, float s, { if ( panim->flags & STUDIO_ANIM_RAWROT ) { - q = *(panim->pQuat48()); + Quaternion48 tmp; + memcpy( &tmp, panim->pQuat48(), sizeof(Quaternion48) ); + q = tmp; Assert( q.IsValid() ); return; - } - + } + if ( panim->flags & STUDIO_ANIM_RAWROT2 ) { - q = *(panim->pQuat64()); + Quaternion64 tmp; + memcpy( &tmp, panim->pQuat64(), sizeof(Quaternion64) ); + q = tmp; Assert( q.IsValid() ); return; } diff --git a/public/dt_send.cpp b/public/dt_send.cpp index caad4191..27580f8f 100644 --- a/public/dt_send.cpp +++ b/public/dt_send.cpp @@ -265,7 +265,7 @@ void SendProxy_UInt16ToInt32( const SendProp *pProp, const void *pStruct, const void SendProxy_UInt32ToInt32( const SendProp *pProp, const void *pStruct, const void *pData, DVariant *pOut, int iElement, int objectID) { - *((unsigned long*)&pOut->m_Int) = *((unsigned long*)pData); + memcpy( &pOut->m_Int, pData, sizeof(unsigned long) ); } #ifdef SUPPORTS_INT64 void SendProxy_UInt64ToInt64( const SendProp *pProp, const void *pStruct, const void *pData, DVariant *pOut, int iElement, int objectID) diff --git a/public/mathlib/compressed_vector.h b/public/mathlib/compressed_vector.h index 6a495229..796a50bc 100644 --- a/public/mathlib/compressed_vector.h +++ b/public/mathlib/compressed_vector.h @@ -149,7 +149,7 @@ class Quaternion64 { public: // Construction/destruction: - Quaternion64(void); + Quaternion64(void) {}; Quaternion64(vec_t X, vec_t Y, vec_t Z); // assignment @@ -197,7 +197,7 @@ class Quaternion48 { public: // Construction/destruction: - Quaternion48(void); + Quaternion48(void) {}; Quaternion48(vec_t X, vec_t Y, vec_t Z); // assignment diff --git a/public/mathlib/lightdesc.h b/public/mathlib/lightdesc.h index 1096d623..3f0e1656 100644 --- a/public/mathlib/lightdesc.h +++ b/public/mathlib/lightdesc.h @@ -34,7 +34,7 @@ enum LightType_OptimizationFlags_t struct LightDesc_t { LightType_t m_Type; //< MATERIAL_LIGHT_xxx - Vector m_Color; //< color+intensity + Vector m_Color; //< color+intensity Vector m_Position; //< light source center position Vector m_Direction; //< for SPOT, direction it is pointing float m_Range; //< distance range for light.0=infinite @@ -60,6 +60,7 @@ public: LightDesc_t(void) { + m_Type = MATERIAL_LIGHT_DISABLE; } // constructors for various useful subtypes diff --git a/public/mathlib/vector4d.h b/public/mathlib/vector4d.h index d63cf52b..89fcce01 100644 --- a/public/mathlib/vector4d.h +++ b/public/mathlib/vector4d.h @@ -23,6 +23,10 @@ #include "tier0/dbg.h" #include "mathlib/math_pfns.h" +#ifdef __arm__ +#include "sse2neon.h" +#endif + // forward declarations class Vector; class Vector2D; @@ -141,10 +145,8 @@ public: inline void Set( vec_t X, vec_t Y, vec_t Z, vec_t W ); inline void InitZero( void ); -#ifndef __arm__ inline __m128 &AsM128() { return *(__m128*)&x; } inline const __m128 &AsM128() const { return *(const __m128*)&x; } -#endif private: // No copy constructors allowed if we're in optimal mode @@ -616,9 +618,7 @@ inline void Vector4DAligned::Set( vec_t X, vec_t Y, vec_t Z, vec_t W ) inline void Vector4DAligned::InitZero( void ) { -#if defined (__arm__) - x = y = z = w = 0; -#elif !defined( _X360 ) +#if !defined( _X360 ) this->AsM128() = _mm_set1_ps( 0.0f ); #else this->AsM128() = __vspltisw( 0 ); @@ -629,7 +629,7 @@ inline void Vector4DAligned::InitZero( void ) inline void Vector4DMultiplyAligned( Vector4DAligned const& a, Vector4DAligned const& b, Vector4DAligned& c ) { Assert( a.IsValid() && b.IsValid() ); -#if !defined( _X360 ) || defined (__arm__) +#if !defined( _X360 ) c.x = a.x * b.x; c.y = a.y * b.y; c.z = a.z * b.z; @@ -643,7 +643,7 @@ inline void Vector4DWeightMAD( vec_t w, Vector4DAligned const& vInA, Vector4DAli { Assert( vInA.IsValid() && vInB.IsValid() && IsFinite(w) ); -#if !defined( _X360 ) || defined (__arm__) +#if !defined( _X360 ) vOutA.x += vInA.x * w; vOutA.y += vInA.y * w; vOutA.z += vInA.z * w; @@ -664,7 +664,6 @@ inline void Vector4DWeightMAD( vec_t w, Vector4DAligned const& vInA, Vector4DAli #endif } -#ifndef __arm__ inline void Vector4DWeightMADSSE( vec_t w, Vector4DAligned const& vInA, Vector4DAligned& vOutA, Vector4DAligned const& vInB, Vector4DAligned& vOutB ) { Assert( vInA.IsValid() && vInB.IsValid() && IsFinite(w) ); @@ -686,7 +685,6 @@ inline void Vector4DWeightMADSSE( vec_t w, Vector4DAligned const& vInA, Vector4D vOutB.AsM128() = __vmaddfp( vInB.AsM128(), temp, vOutB.AsM128() ); #endif } -#endif #endif // VECTOR4D_H diff --git a/public/mathlib/vmatrix.h b/public/mathlib/vmatrix.h index e49a888d..48738df0 100644 --- a/public/mathlib/vmatrix.h +++ b/public/mathlib/vmatrix.h @@ -423,6 +423,12 @@ void MatrixInverseTranspose( const VMatrix& src, VMatrix& dst ); //----------------------------------------------------------------------------- inline VMatrix::VMatrix() { + Init( + 0.f, 0.f, 0.f, 0.f, + 0.f, 0.f, 0.f, 0.f, + 0.f, 0.f, 0.f, 0.f, + 0.f, 0.f, 0.f, 0.f + ); } inline VMatrix::VMatrix( diff --git a/public/studio.h b/public/studio.h index ae20a045..a30875be 100644 --- a/public/studio.h +++ b/public/studio.h @@ -639,6 +639,7 @@ struct mstudioanim_t byte bone; byte flags; // weighing options + // valid for animating data only inline byte *pData( void ) const { return (((byte *)this) + sizeof( struct mstudioanim_t )); }; inline mstudioanim_valueptr_t *pRotV( void ) const { return (mstudioanim_valueptr_t *)(pData()); }; @@ -650,8 +651,9 @@ struct mstudioanim_t inline Vector48 *pPos( void ) const { return (Vector48 *)(pData() + ((flags & STUDIO_ANIM_RAWROT) != 0) * sizeof( *pQuat48() ) + ((flags & STUDIO_ANIM_RAWROT2) != 0) * sizeof( *pQuat64() ) ); }; short nextoffset; + inline mstudioanim_t *pNext( void ) const { if (nextoffset != 0) return (mstudioanim_t *)(((byte *)this) + nextoffset); else return NULL; }; -}; +} ALIGN16; struct mstudiomovement_t { diff --git a/public/togles/linuxwin/dxabstract_types.h b/public/togles/linuxwin/dxabstract_types.h index 37818c2d..49861e86 100644 --- a/public/togles/linuxwin/dxabstract_types.h +++ b/public/togles/linuxwin/dxabstract_types.h @@ -1195,7 +1195,7 @@ typedef enum _D3DVERTEXBLENDFLAGS D3DVBF_3WEIGHTS = 3, // 4 matrix blending D3DVBF_TWEENING = 255, // blending using D3DRS_TWEENFACTOR D3DVBF_0WEIGHTS = 256, // one matrix is used with weight 1.0 - D3DVBF_FORCE_DWORD = 0x7fffffff, // force 32-bit size enum + D3DVBF_FORCE_DWORD = 0xffffffff, // force 32-bit size enum } D3DVERTEXBLENDFLAGS; typedef struct _D3DINDEXBUFFER_DESC @@ -1533,7 +1533,7 @@ typedef enum _D3DTRANSFORMSTATETYPE D3DTS_VIEW = 2, D3DTS_PROJECTION = 3, D3DTS_TEXTURE0 = 16, - D3DTS_FORCE_DWORD = 0x7fffffff, /* force 32-bit size enum */ + D3DTS_FORCE_DWORD = 0xffffffff, /* force 32-bit size enum */ } D3DTRANSFORMSTATETYPE; // **** FIXED FUNCTION STUFF - None of this stuff needs support in GL. diff --git a/public/togles/linuxwin/glentrypoints.h b/public/togles/linuxwin/glentrypoints.h index fa5a4cf8..310932bc 100644 --- a/public/togles/linuxwin/glentrypoints.h +++ b/public/togles/linuxwin/glentrypoints.h @@ -38,18 +38,8 @@ #include "interface.h" #include "togles/rendermechanism.h" -#ifdef LINUX -#include -#endif - void *VoidFnPtrLookup_GlMgr(const char *fn, bool &okay, const bool bRequired, void *fallback=NULL); -/* -#define GL_USE_EXECUTE_HELPER_FOR_ALL_API_CALLS 1 -#define GL_TRACK_API_TIME 1 -#define GL_DUMP_ALL_API_CALLS 1 -*/ - #if GL_USE_EXECUTE_HELPER_FOR_ALL_API_CALLS class CGLExecuteHelperBase { @@ -57,7 +47,7 @@ public: inline void StartCall(const char *pName); inline void StopCall(const char *pName); #if GL_TRACK_API_TIME - uint64 m_nStartTime; + TmU64 m_nStartTime; #endif }; @@ -313,32 +303,30 @@ public: int m_nOpenGLVersionMinor; // if GL_VERSION is 2.1.0, this will be set to 1. int m_nOpenGLVersionPatch; // if GL_VERSION is 2.1.0, this will be set to 0. bool m_bHave_OpenGL; - + char *m_pGLDriverStrings[cGLTotalDriverStrings]; - GLDriverProvider_t m_nDriverProvider; + GLDriverProvider_t m_nDriverProvider; + +#ifdef LOAD_HARDFP +#define _APIENTRY __attribute__((pcs("aapcs"))) APIENTRY +#else +#define _APIENTRY APIENTRY +#endif #ifdef OSX #define GL_EXT(x,glmajor,glminor) bool m_bHave_##x; #define GL_FUNC(ext,req,ret,fn,arg,call) CDynamicFunctionOpenGL< req, ret (*) arg, ret > fn; #define GL_FUNC_VOID(ext,req,fn,arg,call) CDynamicFunctionOpenGL< req, void (*) arg, void > fn; #else - -#ifdef LOAD_HARDFP -#define _APIENTRY __attribute__((pcs("aapcs"))) APIENTRY #define GL_EXT(x,glmajor,glminor) bool m_bHave_##x; #define GL_FUNC(ext,req,ret,fn,arg,call) CDynamicFunctionOpenGL< req, ret (_APIENTRY *) arg, ret > fn; #define GL_FUNC_VOID(ext,req,fn,arg,call) CDynamicFunctionOpenGL< req, void (_APIENTRY *) arg, void > fn; -#else -#define GL_EXT(x,glmajor,glminor) bool m_bHave_##x; -#define GL_FUNC(ext,req,ret,fn,arg,call) CDynamicFunctionOpenGL< req, ret (APIENTRY *) arg, ret > fn; -#define GL_FUNC_VOID(ext,req,fn,arg,call) CDynamicFunctionOpenGL< req, void (APIENTRY *) arg, void > fn; #endif -#endif - #include "togles/glfuncs.inl" - #undef GL_FUNC_VOID - #undef GL_FUNC - #undef GL_EXT +#include "togles/glfuncs.inl" +#undef GL_FUNC_VOID +#undef GL_FUNC +#undef GL_EXT bool HasSwapTearExtension() const { @@ -366,30 +354,54 @@ typedef void * (*GL_GetProcAddressCallbackFunc_t)(const char *, bool &, const bo DLL_IMPORT void ClearOpenGLEntryPoints(); #endif -inline uint64 get_nsecs() -{ - struct timespec time={0,0}; - clock_gettime(CLOCK_MONOTONIC, &time); - return time.tv_nsec; -} - #if GL_USE_EXECUTE_HELPER_FOR_ALL_API_CALLS inline void CGLExecuteHelperBase::StartCall(const char *pName) { (void)pName; - m_nStartTime = get_nsecs(); + +#if GL_TELEMETRY_ZONES + tmEnter( TELEMETRY_LEVEL3, TMZF_NONE, pName ); +#endif + +#if GL_TRACK_API_TIME + m_nStartTime = tmFastTime(); +#endif + +#if GL_DUMP_ALL_API_CALLS + static bool s_bDumpCalls; + if ( s_bDumpCalls ) + { + char buf[128]; + buf[0] = 'G'; + buf[1] = 'L'; + buf[2] = ':'; + size_t l = strlen( pName ); + memcpy( buf + 3, pName, l ); + buf[3 + l] = '\n'; + buf[4 + l] = '\0'; + Plat_DebugString( buf ); + } +#endif } inline void CGLExecuteHelperBase::StopCall(const char *pName) -{ - if( gGL ) - { - uint64 time = get_nsecs() - m_nStartTime; - printf("Function %s finished in %llu\n", pName, time); - - if( strcmp(pName, "glBufferSubData") == 0 && time > 1000000 ) - DebuggerBreak(); - } +{ +#if GL_TRACK_API_TIME + uint64 nTotalCycles = tmFastTime() - m_nStartTime; +#endif + +#if GL_TELEMETRY_ZONES + tmLeave( TELEMETRY_LEVEL3 ); +#endif + +#if GL_TRACK_API_TIME + //double flMilliseconds = g_Telemetry.flRDTSCToMilliSeconds * nTotalCycles; + if (gGL) + { + gGL->m_nTotalGLCycles += nTotalCycles; + gGL->m_nTotalGLCalls++; + } +#endif } #endif diff --git a/serverbrowser/ServerBrowserDialog.cpp b/serverbrowser/ServerBrowserDialog.cpp index 0e2f3f43..3faf2d14 100644 --- a/serverbrowser/ServerBrowserDialog.cpp +++ b/serverbrowser/ServerBrowserDialog.cpp @@ -154,9 +154,10 @@ CServerBrowserDialog::~CServerBrowserDialog() SaveUserData(); if (m_pSavedData) - { m_pSavedData->deleteThis(); - } + + if( m_pFilterData ) + m_pFilterData->deleteThis(); } @@ -813,4 +814,4 @@ void CServerBrowserDialog::OnKeyCodePressed( vgui::KeyCode code ) } BaseClass::OnKeyCodePressed( code ); -} \ No newline at end of file +} diff --git a/studiorender/studiorendercontext.cpp b/studiorender/studiorendercontext.cpp index c29a27b6..62a7c8d3 100644 --- a/studiorender/studiorendercontext.cpp +++ b/studiorender/studiorendercontext.cpp @@ -763,7 +763,9 @@ void CStudioRenderContext::R_StudioBuildMeshGroup( const char *pModelName, bool for (i = 0; i < pStripGroup->numIndices; ++i) { - meshBuilder.Index( *pStripGroup->pIndex(i) ); + unsigned short index; + memcpy( &index, pStripGroup->pIndex(i), sizeof(index) ); + meshBuilder.Index( index ); meshBuilder.AdvanceIndex(); } diff --git a/tier0/cpu.cpp b/tier0/cpu.cpp index d5bf6c9b..90ca43ac 100644 --- a/tier0/cpu.cpp +++ b/tier0/cpu.cpp @@ -142,9 +142,9 @@ static bool IsWin98OrOlder() static bool CheckSSETechnology(void) { -#if defined(__SANITIZE_ADDRESS__) +#if defined(__SANITIZE_ADDRESS__) || defined (__arm__) return false; -#elif defined( _X360 ) || defined( _PS3 ) || defined (__arm__) +#elif defined( _X360 ) || defined( _PS3 ) return true; #else if ( IsWin98OrOlder() ) { @@ -162,10 +162,8 @@ static bool CheckSSETechnology(void) static bool CheckSSE2Technology(void) { -#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) +#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) || defined (__arm__) return false; -#elif defined (__arm__) - return true; #else unsigned long eax,ebx,edx,unused; if ( !cpuid(1,eax,ebx,unused,edx) ) @@ -177,10 +175,8 @@ static bool CheckSSE2Technology(void) bool CheckSSE3Technology(void) { -#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) +#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) || defined (__arm__) return false; -#elif defined (__arm__) - return true; #else unsigned long eax,ebx,edx,ecx; if( !cpuid(1,eax,ebx,ecx,edx) ) @@ -192,10 +188,8 @@ bool CheckSSE3Technology(void) bool CheckSSSE3Technology(void) { -#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) +#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) || defined (__arm__) return false; -#elif defined (__arm__) - return true; #else // SSSE 3 is implemented by both Intel and AMD // detection is done the same way for both vendors @@ -209,10 +203,8 @@ bool CheckSSSE3Technology(void) bool CheckSSE41Technology(void) { -#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) +#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) || defined (__arm__) return false; -#elif defined (__arm__) - return true; #else // SSE 4.1 is implemented by both Intel and AMD // detection is done the same way for both vendors @@ -227,10 +219,8 @@ bool CheckSSE41Technology(void) bool CheckSSE42Technology(void) { -#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) +#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) || defined (__arm__) return false; -#elif defined (__arm__) - return true; #else // SSE4.2 is an Intel-only feature @@ -249,10 +239,8 @@ bool CheckSSE42Technology(void) bool CheckSSE4aTechnology( void ) { -#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) +#if defined( _X360 ) || defined( _PS3 ) || defined(__SANITIZE_ADDRESS__) || defined (__arm__) return false; -#elif defined (__arm__) - return true; #else // SSE 4a is an AMD-only feature diff --git a/tier1/checksum_crc.cpp b/tier1/checksum_crc.cpp index b9dacbb6..29093d1e 100644 --- a/tier1/checksum_crc.cpp +++ b/tier1/checksum_crc.cpp @@ -105,6 +105,8 @@ void CRC32_ProcessBuffer(CRC32_t *pulCRC, const void *pBuffer, int nBuffer) unsigned int nFront; int nMain; + CRC32_t tmp; + JustAfew: switch (nBuffer) @@ -119,7 +121,8 @@ JustAfew: ulCrc = pulCRCTable[*pb++ ^ (unsigned char)ulCrc] ^ (ulCrc >> 8); case 4: - ulCrc ^= LittleLong( *(CRC32_t *)pb ); + memcpy( &tmp, pb, sizeof(CRC32_t) ); + ulCrc ^= LittleLong( tmp ); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); @@ -162,12 +165,15 @@ JustAfew: nMain = nBuffer >> 3; while (nMain--) { - ulCrc ^= LittleLong( *(CRC32_t *)pb ); + memcpy( &tmp, pb, sizeof(CRC32_t) ); + ulCrc ^= LittleLong( tmp ); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); - ulCrc ^= LittleLong( *(CRC32_t *)(pb + 4) ); + + memcpy( &tmp, pb+4, sizeof(CRC32_t) ); + ulCrc ^= LittleLong( tmp ); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); ulCrc = pulCRCTable[(unsigned char)ulCrc] ^ (ulCrc >> 8); diff --git a/tier1/processor_detect_linux.cpp b/tier1/processor_detect_linux.cpp index 11248ab2..9e2490bd 100644 --- a/tier1/processor_detect_linux.cpp +++ b/tier1/processor_detect_linux.cpp @@ -12,9 +12,9 @@ bool CheckSSETechnology(void) { return false; } bool CheckSSE2Technology(void) { return false; } bool Check3DNowTechnology(void) { return false; } #elif defined (__arm__) -bool CheckMMXTechnology(void) { return true; } -bool CheckSSETechnology(void) { return true; } -bool CheckSSE2Technology(void) { return true; } +bool CheckMMXTechnology(void) { return false; } +bool CheckSSETechnology(void) { return false; } +bool CheckSSE2Technology(void) { return false; } bool Check3DNowTechnology(void) { return false; } #else diff --git a/tier1/snappy-stubs-internal.h b/tier1/snappy-stubs-internal.h index 1413825a..ab6a5a5f 100644 --- a/tier1/snappy-stubs-internal.h +++ b/tier1/snappy-stubs-internal.h @@ -100,7 +100,8 @@ static const int64 kint64max = static_cast(0x7FFFFFFFFFFFFFFFLL); // x86 and PowerPC can simply do these loads and stores native. -#if defined(__i386__) || defined(__x86_64__) || defined(__powerpc__) +// fuck this shit +#if 0 // defined(__i386__) || defined(__x86_64__) || defined(__powerpc__) #define UNALIGNED_LOAD16(_p) (*reinterpret_cast(_p)) #define UNALIGNED_LOAD32(_p) (*reinterpret_cast(_p)) diff --git a/togles/linuxwin/cglmbuffer.cpp b/togles/linuxwin/cglmbuffer.cpp index 1fb77b62..6d01e001 100644 --- a/togles/linuxwin/cglmbuffer.cpp +++ b/togles/linuxwin/cglmbuffer.cpp @@ -472,10 +472,7 @@ CGLMBuffer::CGLMBuffer( GLMContext *pCtx, EGLMBufferType type, uint size, uint o m_bPseudo = true; #endif - const char *szRenderer = (const char*)gGL->glGetString(GL_VENDOR); -// Msg("GL_VENDOR: %s\n", szRenderer); - - if( strcmp(szRenderer, "ARM") == 0 ) + if( strcmp(gGL->m_pGLDriverStrings[cGLVendorString], "ARM") == 0 ) g_bUsePseudoBufs = true; // works faster with Mali gpu #if GL_ENABLE_INDEX_VERIFICATION diff --git a/togles/linuxwin/cglmprogram.cpp b/togles/linuxwin/cglmprogram.cpp index 05a9cf7f..28d24f91 100644 --- a/togles/linuxwin/cglmprogram.cpp +++ b/togles/linuxwin/cglmprogram.cpp @@ -350,11 +350,10 @@ void CGLMProgram::Compile( EGLMProgramLang lang ) // compile gGL->glCompileShader( glslDesc->m_object.glsl ); - - + GLint isCompiled = 0; gGL->glGetShaderiv(glslDesc->m_object.glsl, GL_COMPILE_STATUS, &isCompiled); - + if(isCompiled == GL_FALSE) { GLint maxLength = 0; diff --git a/togles/linuxwin/cglmtex.cpp b/togles/linuxwin/cglmtex.cpp index 5230ad11..962fce7c 100644 --- a/togles/linuxwin/cglmtex.cpp +++ b/togles/linuxwin/cglmtex.cpp @@ -3649,9 +3649,9 @@ void CGLMTex::WriteTexels( GLMTexLockDesc *desc, bool writeWholeSlice, bool noDa Assert( writeWholeSlice ); //subimage not implemented in this path yet // compressed path // http://www.opengl.org/sdk/docs/man/xhtml/glCompressedTexImage2D.xml - if( gGL->m_bHave_GL_EXT_texture_compression_dxt1 ) +/* if( gGL->m_bHave_GL_EXT_texture_compression_dxt1 ) gGL->glCompressedTexImage2D( target, desc->m_req.m_mip, intformat, slice->m_xSize, slice->m_ySize, 0, slice->m_storageSize, sliceAddress ); - else + else*/ CompressedTexImage2D( target, desc->m_req.m_mip, intformat, slice->m_xSize, slice->m_ySize, 0, slice->m_storageSize, sliceAddress ); } else diff --git a/togles/linuxwin/decompress.o b/togles/linuxwin/decompress.o deleted file mode 100644 index 3e88c1046fa3c29e782b4c3fc597b8552ee8b93d..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 35548 zcmb@t2{={V_W*n^*E}2{bH-cT$P{HNG88funM0<8LWaykS90T)At?$)L`5XCxaN6` z%wvXY%9J_2eR$vB|M&lf=X;*-`EH%F_g-tSwbx#I?X}k4r-Ib9E+dgh#J@izf)w#T zEd(O+(Gf=Kun(uroYR;RbGnFiweC!4ep7ZGckt!S#XyJFWO1EH8{DN4DX% zL3dEYY~wu8v@lgW;Wrs>A7D0dbue4dgxUI|nIp{6%>QBVFMGHf;0=2H%bqaFfLWoM z#}fWC1_3+H1I7R&uz_KxDKTx>ItIXrNQPXA0Kzb1&tk_&v0EfiXb9kn-6HjH#%>{TeI7RmV{A04XjJH+ zheGh~fd@61vr;`NyeK$SJhlpnC6HiuP@aX@S`?{2P*@!b`0k19R;aD4zX9~*U8f?*lvALBhvA))}D`TNJL ztSmh9C0xhmG_D)Z?1Jm)oZp-ViUC;K&N-^?2m>@5o>>|0026f15hf_9Ku>jC2R!f1 z>CQP^C*B|c5#fwh#dQEJ9d?9iJ7L;-pex5MYY>gQJ z*dzw=vLCkF;kW z6})rcQ31Lr9kH7K`j1}!DWYx|;eNt9CS;)uj06=hLIn?yU=|`!feVSng`z)! zLZT8XxPk)4O^c}lH8QZfP!e63I#4Ia)Z$gq&OiZ_xcZ}Yg$lG{kx&6D7Rmv$gL2xC z4KRoV%4tP5{9_+G44!@P+$6g2zy<(19;S%@Ulj5GQpCa*1;|;D}m{&0^1P* z!tx@w@yry!q!oa8XkhC*om-CvJ>7vouzD!*uhAdi87$7TNmRA+5y4Auzr2up%Mgw(()!_6L9 z^O3@%2;(Ti1d1?*0%9>pGy)}o=>a*wUPNo*Xc@3oDC`dJ{XY_m|Jszqfy4&R87+i2 zK%3*XVUcmm|1Odjy!MoME_kwR5mLH%R4}p+w?>#mA^T*3xj|1AfPsJ-1N1Zx)Ic|0 zV8SApI#5I5)(+i>c*K25(4@pGc;TgnfyfBVNQA073b`NQi^3~ZAop?gq>_Xg002`5 zJY`^>9xzWGm?v->-%&6X>_Jdee}J%DxGRT>jquQ=jPRw!iv-~nY7aTrBY#IYQsSLI zBapvwe@J~%geesAw=7s?=xHVB0cwCR3(VIU&0E0$KwuG3m?gZ(AD98oof5RUFn_3a z;i-e&$rWKr8{w}@i|0xNzIL9}l#(!wLaxc0qiF+8z#FiG{T=XgyroK@g4v{jiNGRu zPhts_ROV=>#(NIE_=$fCm)Hp&IEc++XB5Lp zp@PthN&tv2fZyRNU^_A&3x_um3LUObf~yZV01*Z-uTvpKK!rbX^>7C`o&joo6jUH} z5h~chEr+X+0ScS~8au`V70@Davvn#Iz`4u<-&_GZ>+q}Oz>m)0mZ1Xh7YuOdN>SnZ zlz8ULFs5@ZT%QUm0KVtHXKIgEu<0~BtFFhz-*$E@M?8F1Z% zc`_vY!j0gX?FsWVpa)2J264l1r#Z|3w(8$bxM2|S=gA0*WY|eX+yDq$06Wfut>T5b zkR!WsO#lXJm?n7ArSMw8)SYnt(HWsf3l+lAj!1?BQ4Q$Ho-hkRg_X!I>;ySg6<8_| zW!-@4C&7C5!toTu(rJdxNuVAE5KDUj4eBU{f*x?tgdq$-URpTWVE_=^GQ4xCW~n-G ztH2k+jthiZ2C6N%cL#u=00r040>h(p!X^cKr=zG1K!*{V0ncoRYl9^L3>=^Wcqu@v zAWX6W8nBdyW5f2v3I=BYLmJ>Zf%xKhsc*|*OA*G|09**b^`IPpB=rE!TMI3Pm-=$3 z0U!r2-f|C5YlhvXg`RD~h6Iw$?j7s|#iK4Lv<(2@SRVQpR6t-uv&v|pS)j0e&@83| zFH9$t1A;7e4hiLeVUe9Bv1pALT*KxZvWc+GL)ak!E6_ujV1vz_ma}^ncbK(d$4RL= zg13*-2L`zMZ~D+d1FScA8>A0aIAUm{=Xqe*R`rL8H4c>o@&?&m=#dtf8kmX@-VcS{ zWq=;Zf;P5B=D+tx@b-a68=$p4qMfh$$8*72hYFn0EJVluaycMAfi=hOl0l(l01O&v zED_{7Kn=w8BmZT9mV)V6iISj$eH3jCO<$(18u@gYy{V1^kcBBGuw6n#(TY_+fHmfEJPH0iM;l0-idV#)3BCZE2XaqqJfhgpF&c&*@EQRJU2#FG+R zqj0!(M`7SFpT1#qpxyqO|~!1d6M~Rw5%Hypa-M7zr5Sub^Zp zp-1EZi>*Y-c6hD;d0l~4oq;3C0?rZkj8K6q2>{dLqm;nL#TO|7QK146Q;_@Vf}LuL zBuvo|#>l8<0q{Oyl7^}lga}v zaAiaW!YmtZ4k`dSYa>_|N^s1vhjTI90Ep;R4gY|kf-cw*02uJ>hc()*em(U{7z+W6E^#K&d zIIa)aS|s)_sb>yY8a==p&cVRw5(hvycvt{dg2AAKa}ih!1AuyY1cH+`7z9t{^-4AB1z0ZbeTwhNTg1m$!f4bVzZ zP9+iuf^3xs(bh8`$VfW~B(R6>wS#9EJU|WL?Ti+;hpd}ndjc9m^aq0%0zhi*A!owS zHe^-_SscP3p#~rud}tB?p7Z>Xui2^Jeb!~q28Z3M=`b$_%t z@Hrl(zyyQ81GgCj6)1x(Ne#y%1}LW%KBsAdLsvVpLAEZC9Bdj}U{KWoaLo#2kyR*p z5a~P{!QQX}j>{G*Uva@eRM>`)>* z2wRl6DsYzH0vQ(#@~|8uL&D?$3utwapn_LmfC^aPBq{L!hP}@JXZD8@*aQ9e+skfw zoW{RKw*WDr!Ui~{?VBxL2azli4ND8O?ur|j{Rwgc(A^nc1YmsL0E55>!`=&Q;c&Pg z2rPJ1ISiNB0(T_P<8ZKUVE=gl3~wnM@r&SB3H&-dM?eJjIZ66$I z+C66@J{>Y7?4ZB_1NKT=D8c~>pLpaXJ#GL`Nvi_9pbC;OO+y$bgT3MkRNxVSpjwH@ zKcvOAg2cuU2^34!0cwFrkT+G^gRSN7Nn(=6BF%r#e|7ff?}Gz*Fn=f%3kbpGgEtb)%Uize(CI^=_kkwtfy?x_2WP`TSX0C~ z5R8ZWi7Mda0&5LS#rQwY1cW^Nb2NN-#uo_zJH!XWY5~g(2Weh>8G7ibY5tGS&B4iD z9Gs=W?gG3Fu(L{FbmMvy1iB%B7Qwmh{vtTbMw|v+7yu9WxD5AUz)LAX1rTWQ;{Cyi zbf2&f4!o4$oN^vu^)M^AWgHEzHA0gzoC59+cLetzppOQgb`j=|AtJZ`IVEqCQsWiG zK^Hy3HpuX4A>1amk^$2muD*`*Lq%xP9u5q+01m)(HNfgC5ynY@S5$&|08j-e!30S) zQGf$-6XpSr3Kay=V%oyhx4_vK@X!OS@Ct{AVaz7jVZ1=!3=DW$+#$6dsU{^%y$`7A z;lcNCe&Bd~=%o&|BU=xJz6Qnz^4S4gGjSEH13=RPX`I0VSizG5r8!)Jxp?4{VQ-GF zfZG6W{7@xt*vH~Itl+%oXZ|~qs zIdR~s#rxo75dln89Lkx;5pf_U&0>fEYf#Js5d9BfbYEj%9$dnNfXJY)=?FeV0RG(^ zZLkK30Q;3~f^<728bO+*-U+6xhwB0TL2zpZxeFK?(+P8h$qzaAHTMzzA88q=U|J9Z z8-NlpO~r~}aNuhXsr#D#p9rCx{X^EkwNmxqcEA#VNo`?v0OiqKkA@UK6j%Je#XTg$ z{eZXdg1!(?B?m%7#vO6O0#*s%RuKodjrtjzvuwMpY;7=;<0~MBE1)G3)Z2Qpv zxPrWs6t72ttyBSQ#c>Vu^s@bsR`XFb6NmFB4iMw!ad18b;+weVQbgHdk_YpI%^@7Z z5U}+mxYqD12O!t<2QEwj4;5eJiB zNmaPs%>lk0!ZgVt1a?*c_7mB)IVu3gq*}#n$<~6)O|T1)9h-f)DsT+@Kn^{J0TLbE zV^m-+_)DaKD+6|m0(#~F!at4xMu!LWjKChHwUk5_5CCzz_zL)b$R2WmZ_MEGD_r8X z@J76#QTh)s6)qtR*!=*Q_z&=q0>S{_^)DGHe?k~rIHz;5J?vWrYE8g_+C%#jFa~6H z00Rj2C+wkHHTKY*762Iobpoi5!@U8n7d)*;cozV>1z>~K01s$fIzjmZzyRGLyxAVQ z4QO2&?4i9%d&sSmKRAFZ+aRiV;0K+B*gL|P}N@V|G%CQY1BrOyk$pLeNyO3iD_>f@04K%hX zkqsbi1BW&8@K6qT2yihUvcmwndL|s+YBL}}9qb!k7zxan7Z@C^#2{2qCIbrzAmC6H zNeb7Cj=tRJZccR0VZfEq<_M&|3cjy3Z1w?%foE1K;(0WGZF7P!M}}Ml`;s95T;PN4 zLkrv-LwMg4*kEv>_ux$I@Y4vd=Ud=AcwtY%i{0Y&3<2kykiW1z&?9T0u>UvtV7>h- zA6)@?FZq=IFY>`-*9XITgMY+8kaq$T0se?s18#Z&7EXygh#f0nDwPtA z@TUXf@xumD0tEWQqyN!&(IXD<8)SIkF9{NbAVI-jSC4)`B3La5pp^ghgRw9i#>0JK z=l9p@s7xavcgSQpNAlQbCWTt1SFbo6Q|8Z=d{M zeNjV_>X%fdTg)BaNrpST=<^441?B`- ze9LfQTGl8HWE>M_3)E?kS1PfQe8o`C`ir+xJ}PZfX+M;UwR?fGp}vDVwf{CR`HQA! z=nNB1b;@@aZbfn`y?=axn`H1QC3IH#w7S5LSciCuJW{K(8te@aBBO?A{v~5vo_H`?1F?3mDJg#d>!Qen%e>9BfLa}27}n= zb+Qgx3eLI7$QU-^W7T@1*@f5M=g3H0t{zhstg}Nfv11JxZIFH67*o&N44smYye8-P zvPev&pRxY0eQ@Gz60yhyReE*k2IN&`SzEzXdq&Pe&Wcm>{PqI9H_uo~`%-W+;`>@Z zJ|i{r^!t^NIPTJwXf|hpgx8`P!iT}-z5cG%mk^1YCtS5o=X{eJ6Vr)|7XRRBZB8Qo zRKoEMn>)eOrbKQmt>(kQ6^i^+=jh4r)K({l=TB_PSDvfs%PCDeW4v-!&!gtq#3HZR zH4h%N2SeJKhD#PtR&zY|76X^QeOMHKGHOS5D*WMwUyRGGWGcpxqKn9tE?uMQbn2|{ zpRAbG1uBoDnJyWUe&Q`={3RMZ8rwqd!)!&-_=oomqLjD)ql}&GFTp*p#NQ}l#BdR$ z`XHE=oK1aIyg@zKH0BNOB%&^6fxMVx?0yU>?@xr5pe%JBWeWb6d( z^u_gNGsK%*^LE|}m8|9fP_B3{qDVlQiD*RI@sc5AktXezoS-4y49zI_l#7X8)k@^Z z*H1;ak*RW_C?enJeuG#;4oobtoFO$PS>bjw&1e6lV)~<+itXb?72;Zn z>BI6zRTXlCrzBin58oiRo)a#*KYTy2n84+7rZ$hGOfoD@=nrke@zna|Z``T%zSEck zUB&mcvk?mJsLZc|RjvWXku9^mUeG6z$%vNOHm?PK*=KKW1kGgM*V<*RU+tlNG%u@A zW{^=4=UlOV5mCfI$41?084TP&;xCG*UE~sa<*kVyLQX@O_X?6;W%cYOt?#jUh<*Cj z3@Hp5d)Qfh%kU-AJo>R0_q>uW3AeZADmJf~3bwzTWm#8K7*&o$&M#cA68Y(b;xkct3WsbT|8+G-*K!aZJ zntJa;E#ynBRf?!$#2IwrrFV#C4Q?_bYRp7av3;vgHjvkc4oz8gLF(Xcl(_uTiKH8E zH70vEkm@Iukd3Uokl-!Km_8R-lN2(A(sOJUieKVYDS5OP6q-?8KIv9dQQyQJ+6v2?Lt+ose$nvbT|{2}9(que>S z04aHpH@F2?qicCb43=`$^UgO|hniNet@rbib=Y58rFbZnw{Z!f@|ImWIhxl9p~b%~ z7WEBLs3iH0*9k&!`?K!58J+~QHGji z)n7!9pu8{i=u9wIQVHT-N1lzLtdeEOjeR+VzJ}Cp4}Vi56g&Xl*S#ESbgT`?W5?E7 z(Dzi_q^t#V*;522hCG{n`sv2xKJk~Os%5RPye)Co9oN{%2gkD~MdL3u( z)P%R;w?-$7iyBqg{r7AE-jk=tw1sOwoYWcne7d&8mQSFH`2^2b zi83PI53P#PrGKrH1-!GzlyNdP=}ljh6V6VP)|uXYoz<3`(y#FK+MjPL<|+5Qnr#@$ z^xXqAo8`adE8I$q@K!Zymj6?fEoa${OZ_Bb@zapyS?ZLZ^BzC9-`&^Pd;J+IWs{{9 zKWpwhJg;0~NilVatxymdb$RZ6-Zcy$gD|`7*JaXE)WolYZRKxbnh1 zu3kMW+W*)6*4PB|hSE&;1n~sndhg?1UHE z*hFLJmN0h5Z_@d%6!5or@6B_B4|;TPhYx<@yVkgbnawEuEHnM>xMM1vSkdm?cJ~&6LSoP$?CG&mmnfI8r%dKcr|3;1A`6{pZ6Db^4q3RH>g!qX$Y9#GIsI}y zi^;if=dz3hL_**9%Q$6c{oFW*;bD38GP1->gVgAN$?~_W%8j*;k)6@1y%ARoxS7rm z1!Il*uk_D;n3z>~`qz&F?G`&*r}u!;pT~3PRpe6*+SgD1UZh=d4d6_)d$)E%U-PDf zC)D8VNL~?HBJOFX`hWrS9RA>^=i%usZ>UnG;K^$3LwTl@3x_I3(r z?q>vF&YJGFkAJ;7Pmn7rolp#8A7skBS}!boY(loF*QGt>p2C*8`uJnsQa8kEt`pwK z_Qvqjhh|#=IZ?^ttnVh?8Wj|rwh3Ce&=vlBFu=%qoNZS{zA_m*CBQK2s}ip^tUETU z_p|R^9dc1axFUiJVgFB`%HW^!|qAKGlVzeHJuj>ASG`IZ{EsWu)NIXzb88Mat1}*PNIs)6lg%)9P4vQOlc7 zO|q-Sui2flsQz|QToaqzI(gtwc!K5fNT(Eyz$*uwlE}$PLlaG{ZSZOg6ggpYGq1*G z^ph~Z$Ufz_J)L}+ALrd=;!{mdSBb@PUR+4PH&1DQCCL;Evz1?6XqlhHqqFXPuK^_-ml^JmS-X zcPr7JsAf-^;r zJET~lyZ4QEDbMlbaYzg`w)=T=zWs7>iPAk~JYML4|EJd$a!g*X!OiuV2PV1fGY6g9 z#8|6JfKP!6zfz$>f}MEu=KI901+`~A>F5}Gta{QLbv~s)@Fr7gS11keINNi=kiVJV zYqw3sEhwXeBY50LM>zxicK36rmkhnlwTpFbMa)*$lVYQ$GZ=l6UUf)4KQBlB6=|=b zoaJzC^IUqIS!8CvIp4=A_tW!R)(v-SeD5alcJs*h_TL61rw4+^^!sM}OxC_kq_0~o=9OyN*EY;S}OH+TiZLvvOL>o*}P${Ws%85ETY90 zA!176n*s;)MK4oj`o%3%J`fk+(0OA^{rlF+j~}o0ewe=}J;VNsy*eR|+sdaR*tK10 zb;~A#`@`S0nzZcY(?#nj?(f0hRXiMI8_d4=W6x$T@Jc2G-?VoO^AB^qLYm}n8Ofc< zAbsIlvY>F5f=@ZkfPqEi_Xe^@rxj1QX5Y|VCG3&x4L@~Tq*Giyz5lwD6+hbZkgELh z$hy&O8?5TH#TfcsmTy6HjmkNTF*@#3)zmRU7YeV+iE=9-sPZ-~V!oe!7otimh(Gz8 zQd8f7knPt`EaH`4vdb8JxKLp9T4&9_UT1jmb|TH#D~c7iJ-S;DqDqp+rhcq=rOUtX zLXO4#IPrd|$GG4w*Ou|zmV0#InHP#TbdbPfY5!O}HmtS$@q?eb2j5(8aL&UmQ$I|( zQ9>SBJ~r!A)DceZ&sR{B{qqqU+7Gwt#A$*Z*;)@+t-el%_@r)U-M z5Oi7HG*t?F`Z;YMy8P`&jqZG<`ri6#Y&psnCEr(LQrL1T zd3S67Pw~6VJeWbejYME&WU-6?w66R59HEIH5vo_!T;UlSHvfTb`4(Eh1fq!{BDv9guX`yUP^KP7m=&8LR z)~}Zj%wk_5K7X#TzkW6LVERKl&EHkcm3xLdqE}DGkERAxEH>F?{#upY8qYR=p&>fy zXLa||gh}nFOxYZjp@)+Yk{TO#Mrq~Sg|<&Wtll@Ln0Bf3XB$P#bC--TmsGk1KvTc> zKhC>-cuy7?y|D)h99 z&t_iG9W~1jJI@;*ikRrGoqpBlzcS&P_HngeQHG0{^|inxgXLzAZihk2Emg;w%L+ns z2gqcPLGYmxzJK!smx6HV{n{>e@cN)QL)d|*p1QC&s!--M9Zu$beMhigFiKgUe&x*B zS1;nnkT37ZcgH*OO?-hT&*%R$AFQCA*+npirRio@>joO){r~+fPmRO zcM?fVXGP?jOI)k_D6L*zlg&yo*8U={-`m3Qcw+uFF;nC~jLc?%@$*ly{g*rjUsNi% zMwU4DNi5j9)0UE*{UTV!zO?2qQ+p}7iP`sCM#+HkdY<@-FsI#H-_yG0Kb&su`c8&_ z%yMBpPBx-V6#QGj6rf1P6B1}*zEum|OiYLrANao2k&%1VeM zI=MoQZx!((Bn=?`MbW7zayxB3?qvlQK@2SmZS8~pp@oA^v0n1*d zeR%vU#UAC}kIvWToPsbRoGjZArq{=`_|v`*kB#*8O^Jh#TaePr00LNNc&%@X{JZ8o zrha+`w_j8|%MlyaJ}>pb%3G=fFHS;b&d9xV@2;f&TK$hD1H5&lnY6}@EU*2(&rLj( z$6I^RDHEw&lc^M1KV;6GlA%6DD{)cj#ASh7*1Zyv5~?(l0``4~PTue0nth|&-~Dz+ z$XlxivNETjCD|GInFrjXyCeIXBpUbCwVv!@G(&A8^xNJG@!{TF%@n(?a#Q-#_gO0m zDtZ>4Gt&<@Bih^e3H>JLZ@Up%O>E4?&p8nlqm*LEeiT0+mpb|D)_m7|#Hy6oFbHGmZ*}L?3J$3J7;^PZDcr_nb7PT z2hBNTark}>!>4EK#NPgz*nj)JOwg&BDf4AIAI`P00Gp8wWj@g|1Im^6ReS{KxP@R~ zaH;>@>38Jx1=3DGfUH#iQhSQbK$08MmMRH0XqmkBebhb@0`EdE2h*`O8aJ zPLgk~wRFgJ znSK3M7r=blPqy&l&$JxYcQp^68s06)JU3_{dj|7jbt!n7orfof?(1;KN8~oax2(utJ%{`3=d99|(2oK6eXzyE%M z^70pRW#xHjzK(qAHdz(xpE1F%v$d97eO79sjodWM&RILIT$~dXyLA^*99aXs^pmhPk{z&Dzg$}9s=1JT$U9H!WYS-Cf?F5uJ zOm_529!k=`hce_prm0;wUB-!X>Qva@SET#uMDz^&__?iO=+{1`Wq<4dn_RYA3W8Ge-lTLZ2Hp`^? zTrIz$%BjU$p46#fkZbnj+q_P4=97#{Ue|0x$o2fl?>o*YNh{k$&J^?}^`w?Ij;}a>Dp_fFU@;%bj_1D4l_u2jI`**bp-TvPY;@t| z{a@I~$VGZpMSkH}ev;mE6kpGwc!9cBezKzwck)KqbW_S|gm|)6jDJvd+_kU{bE;z} z)eHjkJTJDrVY(e^sK_Uk-$QYyRv^(gP3s<=q}*MHbE`3I^QXK%h`sHy74DBGxUEWZ zJGVY{9Ah!(`D}UFbZWMj=6N?+eo1DJgN>x=j*mg4a`=Zo6pM}y#K-DsBRpKW??+rp zPEcJgRU4!X<^OXm^j?~dt;bde*G#5H!P)-Sg}CCq9}+7Dk_KCGy!_Z}A1*_+w8qnS5+3DG2b=d4ZhB?4!D)L78K8Og{F1w zy4wZ`CWWkO;lZ~2>KimEkG=fr=JBZ01QB7xx_-XX8DTcx(DvJUr99(r5&6ecs>P^E znARsaE>+O@mKR^bj*_22KP_P_0E4KAWDm=@|e1U#H==4m;T>%_}1N$a>>}(8Ktu>(1c61jULxiF!55L?5!qM>3wf zyIRNo{rk;pTgp$*bAMm?5-lpIW}%uu)M2=NeTycfY%7pY=8*hym$Zs_<2%Y%8skhg zzMP>CluN3977B9^o1*`&doA^8(LT?8<+S?6G`rEMq1t)K!aIv$D6{J9Bmd5b-t6Fw z5(T!;0*llR2XXZKK_?gc3qJlR%pCuImZA7mJ`4A$&RhA*vO=+`#}y@?U5at*n^d0t zywyp3<$rD{sb8tv)^9wpClh}LgaZ-P%B1dHr;g1N^&7pE~vD4D>a_-LQrY7u{ zzg%6NRSvWAA$@!Qq3cb)Pf^I3oGz5Bv--prw0P`d`^jn=F~YJcpP-N${})-#pRQ1= zIRCb`F)maf+C$GAOA(RH!k73_O!Q-Jrik~pTQrJ=S zT{qFmH}+l+t-BNgiF(Y5d)iv1 zh_R31oLei6*sCc6X44JYdWXAwMJ(moMywxh^`<`vb`3Lm{5_RTVnOIK?KGvtyL{(S zehL?&pT|hyiRVOr+X3SipL_D7UoKSzJBeof-1em)^z)Z!H+?eV>BdXzp@dAZg&L!o zg^l6d^T^Sh<~?gZN%odkO2`IU#*8#I?+bo!pU16~+1}ZAjik*(9+W@t*3^GuKVfom zDy&~8pkm~&)vZyp9GTtU@hyS0R06M97ey`ik{Y$T&c^t3Cbd8+1%hjDnyxjjcU$>A z!FQ%U{48c+*}}v5q&zO>ri;qT(}{x}MfL}y_F?k#<<&#i$_lp*#*0QH*he?4RB9=A ze@h6Z`5oL^>CcH*Q2$%=8sihiRkiRg5EnLmvfT1E&3jsoJj>Z7wZ0bGM`G(vzFqu8 zZnw~W>Ax>7{=IlDtpChk`ZFC5*4>6x2x`3bf68m2ml)4&tes;yud(HTo@VJSl{-$D z>lG#?MUmWJ>+Dxl#C{f=cbRG?>NpXLulxFGxm4bnv@QB^Q?%Ipp$gl;z|~rx4rUfA zz5>>v{7@}VQTnf5jBUrBc)s_1?_ptkA(+MJYS@XG8r#w86V!fhsdwF#)m?r!b2NX} z_I2yuk(4yGEO_^tM<`JL>2!azYO}>mYfk*lGu6d&qWx{s1%~&UhhnsPRqQ76QDlL4 zl-A~h*RQ?psZ9Wh|8Kimw@(Y@lFm85fe0;IMF+|nqC}^CUY8PN@q#qkL~0g**o{*L0ewlT@e@I43B}PpmtsDPj)N`0{sPX70rpGXQ1(+h5|hsv{~M9 z0Ln?K{W^>yJ3IIQiLv&vzF#2NbYS?@`M~hI9gCfP_n40Q##W89SllS)O2GpmWYQvT zm1!#W0vnxzS#+qEVd~eMbBtK7%huUG{hX@{v(luRSEM5Q1O}4Wc7hq#aIu!Tif&Kk z-rhWBpW5cj%Gy&Zqomd+WzK=sdO>1H_vYJ0$FVIRLnn{xe>mPG%eVz`d3u$GFM8cw z{_wWDD91P38yEic*OPNv(X>u{p9$CFX)lb}{K6a^2gDV14mKpwN&D-=$YMUtQ z`QP{P-oG=gH)S8HUo)llklW-nYvb~|tr5z5E!UNUoo1Wn0L!Q<4d+`{I6L`62<&KsX<{4(+_ zq=vI-{4Ld+@+MqW`}Cuka_Kqqsdw@9+1jEIU&ncpr+JM%+H}KQ@0*oF?ws|!`St5? zroq*P;5C`tdL-qG%c<#VO;3>x$`70zDDQ2!1|?qj%&qmru$@Y5M%TODn~=O>F|N>G zUt?S#_LJJ|SM{0Xhf~MJ>{I_l*LMt6NvkCFkulOo*PRjl)lW;Hc@eyK`A(EgsgiGp z;<;JVxyuB(;edOl{KLs5-O9gzI@s->PLJCny`i&sEhy!>uZXT@u*f=`Qvl$_d$0MA@)}mI{VBMpODS`IE8^~ zMBld=&zVeXglgl0jv+6KT@!M!n5OdSDjHH=aqyzc%6Yp|RW5CM;qiqJO@2vGWB21E zi)YW~#h#12vCtj8@ORTEOrgQIVJ%Lu=S9pbQsRo#_tl|Q|0^cZ9vjAy9XA z1erih$`-x(f!sc9`vqF^7spWikZ~OOaOygnXApC9#hD=HK~9>K5#t2+u~C%(8Mo%{ zxkuI+@uzsaCFY-U#vdG)x0FH+tsx1#kf z++ftuy#M+cr=!IO6aIYx>TIff%QV$gc7<{)_DD0%oY15=W_DZmGSk+E=Mgq4?dOEj zZ5~I}Mz>6>ea%ex&=;FHDMl@Xy_|3?lgUaVC2^8ylhH(9&+|0n*)Q@(`q5Rs6Mhp! zhg*92tGN*?d^}>T88W`18@vv5>uOa}LQT{u=OYtK!KqkNQoY zcc!?Mi@IF8={$DJ!Rp_sG!HKKluQoNt!lWxeztp=QC#5)<3(MA({F~_E_vzb-!iHi zQH`7{$EA%GFb};emLK0>j_e=QuZvwPdLwJ`dU7A<-(e&vpmgA^mBF;b^h3|@bLIA_ zTYs2MKcp`xJ)Jh&uOVbgwPtYiEm8F??OGQK?w`Cy_4m2bzTnCnvzw~)s2aUaxkB8| z^y&|znAbgfHZs+QvH6X^xMaxCi}~L?(RjftGF3dPww;eT2E)90$8AXhUaQNzVYsbh z!+ncQ)-pZ4nM#0aa zf4Fon=$gKuU9)R+itQm2Y8kK{x&5wZX=WmS>iPKn^#=+~snPU0d1tw`k1d6M>1B*D z5=B>}e<3AHa1agbadk7j$w-$Ycbcza^W4t&W%jeBoev0t_UCTvVBPY%zc*i;WFhO< z;5l`=wk)>~!6t zJa5wlR_pT$C8_6ZwjV|kgD%lymNaRc)XTF6yc2xmZJTy(8#d^dF#mXWmHqmS9~7q^ z+*p3lbC+Y~g@vQ~C3i|+`UWgD#p7RcQjcmZ!jsFgtoAChJTR{cOjYA*&0ULrbMf}2 z+e_W_Qe{Vm^nx_)nDCP*H0^T9-+@`FeACw``$mLgKHb%mil#$ zOH6(X3ghzRcu^lWa1MX5#bk%?3aioVeg7$?_<*S*Vt8iunYKFQi(wD{MOQ{9-|IA# zgAGn`KTizPk3Epz;Y#Xg&nJtK4bV8FX}z-@JhzIRQ+By{s{I+aT`0xHcOn_N-S3tR zaf=2AEy+QK&!V?p?qF^!-e_m?vp6xf5d6gXwr=Rc4RQx7dLgHtSvE5S%Ng19B>pn} zVEjFqoWRR;$ppf$Pmz`N$Dey~xyiF=-uS>-ORphFrdB(B%haHDljF^(-*U%rLds)3 zrSZ@YJmO?Jsp(bA`5#214Ic2J=N{8e9gvwKzNT`4Yc}|wVjlgY%fkwSKXG}*VKrsuDx8Lu(Re(tgEQpOiHY~5!3NZGu7?-jRC4R(#BV@ibBJmhMhCt zRP&@B8)#+!8FOdha=0x|Y3UY4)@lV-M)es8{x-DRij3^)Jv zIIuQ@_gxES%cA%r?|WzAz8k(bn6A_2ZI~%-I9^uY#C%H$%lW2Z!d@RR#utc;F{U^dMXnk3H6x#$9r^$n ze9B_gekpxi8znIJ^2z*M+LH#2yt_?T*NPL2R^okmGg=d-wY?n&VwO}}wdky+8#x}1 z3=DoN(&x-l`LI;$hiV*nD`%b=>s#<{_e1;y$%%Bfx0jwU&|TB&pnIa7@J_w^#1oM- zdJpMJG@{?#d2%A*y-#q*FOQp*wv?KT_8E>;Yza!ZCeG@yFrWIz zI7_!D)mU*T7#DV*lyLlF%gRZfnXMP?_f>r&Ctsbsd#NtZDf)MHDwo9fmgupMTuus& zR4;rU+Qj9!p||Plh4&L|%q!kba$fi(@b$@y&BRtsgz-vW>kofGPOC4*gq7+lhVBuoi zO_Yz{$Wxo$;L4LLdRA(Cr@n|Lg+I{fr_Y&XeaY}3&7{DOY)9?A0#>kt-+K^IYSY{6-MG6STl3rukYNEV4Bq>S0 zTDK&NGINoO&M~iWd_ns>a_r8kW%q&GGS{WtKSjsO@4h`3I`CqU(s00NAh&0~to+IR z)90yAj82?oe8zQ!p4KbU%V6cAZug67yyUD3S-%_2$#(PuQz zZpz~Jm#oe!59gFliTz}nCzL0BytgVw`@|&*>_d1jT66w;Cthnz=WO=Vb9%C;I8F5$ z|F*hD8hgUY0 z1)t(d?+HINBo5Rql7FiuPa}QJ+o)&NL+znBJBJz$Mw*q-kv^f@{k|rora6ElBi9NZ zDc>COYrn9WRPlRxW=G%{E??~;`2VY1_1o8sV(`_`jO0@Rvq6*xx)TTIR}Sd*uDl^Q zJXkos&aXJu6*N^&w?yHzrDi$1%- ziovRL<^^Bf^ZrWr;}?kaMWPqJFHqDn8b>8ooo+qdFyil0qv)i z?ONNCOOHN@bI+VwC)r4l)cBJov#?^6nGvlm{pN83W6|fnM|xMPr!QJk+bcwsPhZrP zv@c87ck;>8>+6CD9-NO5B~|+9=#QT&i96KN z-!g^Xn2#3{bP=QenCmoF`mxA!AzYMK)Xsfx`LgnBDa!k{ zddqB`dkpR4#SNs8=BtT?T$cZ%u{!~Wst@}Ieng7Us-lf3q=YDyWKU!X?KIif>`E)C zr;EV(3Bi=lF-J!SkdNq|ii{Oy`_ph%G(s-F(`$XTbr^~-3J{wQt zEf_m?RpPV__rh0A4Yg})vUA2ahoyI`fUw7F!N;-S$}2&VN9z{c7dIXZq$& z=%kaqw0Lmv>W)*6x2-F*l?>n6asNSsCsX%a2pY2`)u&{Y>f6RoomSrYt4_sm^Fg(6 zqpGxPYVT@G8V7x@@_o87)9~%&dmFMMFTQ9re(^bZ*lbnfL3WLCm-?h!JgV;7&*aMB zqODWygVU$UoYBY(JE2~c;^b6e-}B;}Nv+0i$=JDR%9!LT$Bt+7U1X2%GcWmTPQo|U z;x#UvWTtf5rBG{8p72P=vbgutWoP@u%as51#khKK%&U*BH=TWFr_i%-Sx}{iMV;Q9 z(Mib;^_pezmDj#bzrI;x;%2MPejoB2>X+ZVn|iG>-`3uFZHcklqtw8-*fT4<8~(gUo%5nfm9kzx+f_T%=hW-$TK%wEYUY}GwaV$vCYQ#ZS=+v>;-q@sC9fVuyNpWJ zWjDQWx|(z;qrza3V!l?+`oJ2!zH-I)Z(mRwVbdn&&8z|wHT};zeTPZ}980s4n(Tg| zvA@d1GjqyE?|ii{sq5$B^g*6`6Ytg@E2$|zmcDS_rL((N9}yzI&ex5tK0MnZ`$PP@ zk^gqq+giA3v6SoX*)~er1#x4)KC{UZ0<<=GB&Nr7sJpdxk*Px8m-7x^T{bR1dUKu6 zY?lVzqS!zg@A?Isek>Vdo8;Wzw60QcP0V<4D8I^HWA3%Jo27@2I^o(TO6u|=4U-c~ z@;|+b8?vn-q2IP?vfgqF9KXpAPyeCkvg~|X4?At!_v#X99_zJd*N?E* z_HK2zwtOCNH^Kb%aoPK#J83HxSWDlYRqEkCW_?OdsX@{5A4VRE`o%pC_o$g{;d`NE z<6=XzskZqyu03|CZ&au()%d7&HD^xmyR#SB_OY&%9{NS?O4oY#i*zEh?BKE=C$2r; z-m|w=UB~L4rrUQMUpcpO$A#L(w&l|OC1XeSd?>YU(up>O(mh{Zvwc>xRCRQgO`xe# zslnsW-g&pO-~T zZP=%~TlvBhcf;Xt3Nlhv5`4PP7Ue}gj!$gNs+c+O_S3V5`$J<78tokDwWIUIYqmCZ6zhv)8#YOWWW@^`q}8U$p8XGx|C7qsq{dg<>PnS@h& zPkB%1G^t7Fv)IRo;mGhY;#h)jeb`)-sQY*=5}cH)GThEI@)thdF9hqhIxh^*U z?@6N$*ajZlYTf&bPk3#9TkBpHaanWv&Cit=?FyPc!ne-E%Wbe<7p0G1qJ~@6n@Ab- zDZIE+yQdHs7|o#TQic4JZtYOxQT$>C}?feM*zNoD0A6e6y!> z^eW9mpWA2p$plrL@3%F5vDNAO*N3=W?U8(5bJWkviq%FQ5rIKcAA;-s&*d8Gw+#`B z?MqJi*!Icn>OCwlcb8zI@n%4`!B$h<0~S45ZMS~;fDR9*emXP1!Q*3x_xp=ob??+Q zJl8zoA6)v;e@nNKH;qe{?&!W&VXTVupxJ@XCVR`ioU))aV6x>)nXM;#JA{?`eb!lS z7n5$R($Ov5_xW2xC3(A55~3T1-sWaH#>z{ooOc=7M=1ne^$PfM(j+YH_4KvV>rXfy z)4hJsG_ZS?;b7~nZ)EzUlx~wXzSqGd#A&>Lagk!miFS!9qbALW2(R|ue`iOAoZ0+q zM>;s{{8ztyP?AGKj^km+e`~Up_RF2BTK~3Z_31*t9peIHKczUXS=w<+{`Vq>gppG; ze#EGT{=Md4<91n@PCmtDZ`LjKS?YD`s`iS>(nZ}iR%j*`YDuX@=FC~Ac+dJ$M$Y+; z<`1?_d{A)MU_+&{yn5Y$pph5_@@*FNnxte{JT#~7zH(mq(Jup!7TbS+5=SaZc@iL|MrGJzm4s`F-ZfqA{PP z4@{RAdOymUE0q^ywQJOHw~ad=NKbxIpw-tp#34{iVq)j1-?AorUXhajspg>VsXaYQ zruO`LFVta==8tZfVIQ}SoHHvX@aYP5N82w8H@`ns6P8z9u_rNYo^1B0u`~Bgo4L9% zO*$_?y7!a0dGqC7dK}OiD|L9$zOtahP62mgzDk(IUEX^))@?@0=AFyG*WWl^P^BIz zZM{&-f7SZj9{sx)wUMo`6n}@{7L41PHKjO)kL#WdpsI@Q1kWWLC$66Qa+usOv{3HWPV%cGxvGH+6h_< z-IK3g)~>It>{#dPeej^hVdoz=^kWat>(^@gz!Ih7i6*q+zMMON_*r6ctS=-@S>hZVcJRF%(Y_E7ff5%|$Bc<2h$;V7BNe#Vl-+B3hTPO8b zshpIU*D9@#s?E~(KBtvbc5T|YA+_{owsV!=@cYTbtM|4)(eYiq`Swfidxzh8^XlR4 zErO=o$O9eToiMb`wD*%-n|$cbSSE%;t`mu6F$5r)V6OA5=E=bjcaq#k!)x!~@s zv8OG2^v~RRYTK;~-_8V8zU#fl!STY4qy1maYJ6io`K6}AzWRjP85_b=H3t^G{CwK* zyz(f!u{%Ebbw9MU0G#f9>CS@9LE`_J7$b3VIQJ!*pQ4>$%FE9qT7u+1JnV z;;wzFh6&SVr-~kY&1ju^E8}cs?T`S^D($`d_iL!hH@<4S-8R0@s{THMPi`7_P~X>G z(qrRZi7daL!+q7e?MlfV{arrv$4`;$HQ~``M~lx@mvRkPo1A|4AZ^(A4?iO>d8{gX zYuWFG(ND=|*Ud}$vve|D7=n%2675~8z;dR+S_)l(g_&$$m+QX(}$Y0B!9>j}Ycvnr;W zjGSLFa?qka3N4YHvzRIQV4{TbzamfVP zM_1apRvazNDfTh*e_}LsS~t^KtJ69K9{4n^N8QmbzAj61N;jyz>(V|wdylo&{3VSO z=1wZid-LyDrHftb6NkSH>Yej^k->7?vpYVXOqJ{+IY@H-K$(qR25dWeP-&p1$D=-?;oF>-TVBH z?%DtRV{H{a!Oh;f!gpv^T=v>e>kaN|n;k4q{Ce)p)}-pvo#BScql-7kyMbSgqxonVT9nY8jqdkoqDzwaC48pI2;L|5c4CD*fl^cl*%oN!H{G zq6^>ab;QRtUnQ>T_};quj_hJxIjbKTL%Uiw42ut#cqICmf!4|Mq6rPpif0TS_NjNW z-Ha(BsRXY~=PXm5w~L)xXQqGFSYos$YFTRD>540Poevyzv$^EiCN?@Mr#9l^x%-=k?C#rzZNi2I_7%ohSd3879OE!SOGQslM^8sr{BQre({qxZngZqNnupjZ!6J5PD*Myb>a&M@kwo-RQUn}W(sZpYGtM%w@lJZ zp~F%sGledTTU#j@Np6vlRp?@tuh7Blih`WkRfV=@mldSOw3BrHQ{J|nDV2(%n=aMN z|DSa&e$};6xGf<$gbP^wr|?)_XCu0uq?(`n_)A8cw`!HVP`OKK<`C<9+gh zZYp#!1P7AGE^gJzCSO5HFs6{Nj}b;~N&j=*es8md!Z^v?R)5%Ln|9-9yc4evU0C5O z(PF*C<1Ki-EEJqtomFtPRG2(Q!Oc=(#yABhr$1CPqYX8s66st|p1xe*i#z<@&T+r? zm9;`%E6M0T?QAX~z9vK^*Pw|Xf1bG1=Rf4+n!dsK)lb~H#T|x-Gg3m`&lXK|4T|RK zO<75SF3ZtklB5Sq0Wlv>t*HM1ikpuEj!n58k15Aa6vj~*6;gs4+(}c4O%Vr@Ds%>NljN_x*OZhLuf^{Ul0wJd z0nJGYivNkb{wL;V!tY5^P!jI`4rxwSctH2RA5~f53B`O3k)}yjsH7P>)G^5lZ)lG9 zP17VRyr-Gw*B7qWK=JQ(8Q~knDs*d_=;9~zIQ*Vugw_<#_?3_mWGQxP5x1vUw?*8A zV$&8ezcz%mhE_U z+>Yly$45BEbC=^A|6@Egc>8#T&J2t2iHy|o5-fiO=E0#}L6(zU^~Q!o`9y?x1PhT- zo+6RGN0gVpMQCtnM5MS!Q`TyhM{rXTaZH#;h__+OV5r~~8WI^L@(2#|_s|(41cU@c z>F780a1>WX_oiTIQ_J{hkB}&5b5{?~7V#JkXY0|uj-C(}5$YXDe`|EcSq$epBfirU z$9j6=d42Ky5b?c%_})-_&v)WBap$2NiW|f|#T|zUB2i?NhgXou%Rflu>k$wvqP;@9 zLL`bY8mi;x6D11s5=Hq(hXm{+Vw1^$H6^qn2& z<)A$bh1=j>_z{}W8yOyF3)jO-EbX^OlJ_g^W6%55Rzr}W z*A3cqL4wB6CW!kDgLL5|acm36!^uztyC$?^D)MDW@-IY5UWESmc|&M-4b~lI37-bB9`j-BMyeqa1M-zOCi6f<@Hzx zx4?9`2WG?L@HFK2>pcD{ybmA37qAM}!g|<~jyWE$0=3{EI24Y67A$S26Jl5B4n1HX z41u%ZJopz&b&?TpfZO3txF6=g(=3g0LmQUHJFql< zI`W<{2!_HamijF~ya=vjsZKK5cObtD9)w3A|IY|q=K|u(@HR_zO3?ls`Iqn?_%Hm# zQXT&6!}~)Dc48^t2eBID_YGWMAMNJITSG_a0^Q&Y7|2rHNW`;Y99#ld!L=|IZiBnw zK6n(Kgje7VSPV*Dl8(|yz1I_!V9qaOg%s5{`!azKZLOgA-WV zFP?~fVJM7*^I;rZ!P2;mh|^#u{2LyE$KW}55f;E(@F7c&^HY}U*CJmJzeB!mByWc- zOZ}8t8mEF-7wrac6f}dj&;h!zRL2eR3>d^x-Eg!oLjEtf8YaUuxE*G)R5u&(A$XRh zx);%Y7x@zS6js0*mg@XqX`G~NQ{0)Q{m>1nKy^3>>cOEb)ftW00y?l%*A?wEkoSf& zVK|%%7sACX)m@2r4cyG~cEVRI_4^4s(kTI6e?{0E_J!JTFf@Y3&Srv`By{0bXk8#)2V+uaU!VX2-n9DuwI90o^18$40T;p)cmWo{hwvq=gFoPKS{ARz3U~?Lf{$P|tcO2gJ37z7c*u+;Ac;+wD(K7y6-6|84zIbT>BFGHvJ zc|GM}SEvm8u+&cz@jz$@heLB{4ac**KUf;?hkOu>f-!Ir{0pvTX?zMx`z) z?Vt}#h9_VZ)TDLg@sr?mmewl)@n)FC(wIH)0P^SHP51Ns zs}F}EZ;sd++95v?@l+^6J`iyTj6{AR;&`|W`DDZ!U>fqf5buSjU@k0#x8YM*0sn;! zur-|n=k1e$-C<8S0O~+vXbP=Zs^x^(6;46k8?ir}$;vUyFD>OhrBm z@g8^p`BR8<;YH+&5Z{4iEG=K4?@HMIa2yO~X?wz1em=0Ye^$XXxE*G))c+vjBk&~h z`G~K=LgdTfE7$;~RGR8_Wa+(3SC-bp81ZDZPe*iM~zS+ZrmcR8J4O!r5>=Ol4_# zSuE|xJ@5eXrx54Di^vxtz5~nHR)X-1rRBay{sU~I)>OY8RAi}t50?6CBR?3LLu=^7 zQvZo8_4h?S5Jtk;Fb*z(tKeFg#!|fucnJAp@ERX&;pFc~tTo&YP)Fz7@qW@GSfc`S%pu z?g}#@f3Mey-m`ya={+<5mZBBC=M;1z=gf?v9|-6~MnyE~-#i!L6RZ>E6B8xqczZ;7 z2s)mTk%A8Wen1>FspAzHrANO7_#M$N0eFngAHMVVjQPLY==T8s=Q#RWUWeDk+s8B7 zPZSX18%k>=p7-$djPRM&vX#Et4-SZo;uZZh!Xv^X#LtIPO@ksr{Qg&$-~jrmC)#QA zn`b#qKWU(ten2qOCxmuINNALg4*fcBkocDli2XvMb?BEBVLlO2bDEay<1g}!@R&(O z{USo6!)QnU|4sb+75b?BYcilYI%Uj%)&xDe(V-~5{kI<C=O-H`6{t&cJvwZ~6Zhj>^VD3=h!Wh7{=|x>HPmkn<{91&zKNfQ zWOF|idM?vT!LO-#TU;o^^`x<2BkI?@Ebh#0JcZJ&Wj}sD^{0M`|LMo?xBk>`ANq-R z4$WIqzlC%zqj_o`vzrnvACF33`fUh3Nc?!vDJ^kr-oGd4CSF$#Xu3oPQ(g1QQ`}td zI?cE2cX9}QR{X8H&D;B(5-peOM$6^bR$eZ> zO#WJ~92Mhwy4Zi+>4o}f^TCLI+}`a!{XVX^LB#?wYrE66!Ut`UC3MvpJxy&HY@cU(5Y+kM@fb zwex=Ig?_v%c=;`c3m-&+FoPZs70M5y8`tCK8`o;_FbS?+^qz(XV``q|W2a?5^(A!f LxVe$qnrHtX@qk%C diff --git a/togles/linuxwin/dx9asmtogl2.cpp b/togles/linuxwin/dx9asmtogl2.cpp index 969b6a91..41d08a94 100644 --- a/togles/linuxwin/dx9asmtogl2.cpp +++ b/togles/linuxwin/dx9asmtogl2.cpp @@ -2148,23 +2148,32 @@ static uint PrintDoubleInt( char *pBuf, uint nBufSize, double f, uint nMinChars if ( bAnyDigitsLeft ) { - uint n = remainder % 100U; remainder /= 100U; *reinterpret_cast(pDst - 1) = reinterpret_cast(pDigits)[n]; - n = remainder % 100U; remainder /= 100U; *reinterpret_cast(pDst - 1 - 2) = reinterpret_cast(pDigits)[n]; + uint n = remainder % 100U; remainder /= 100U; + memcpy( reinterpret_cast(pDst - 1), &(reinterpret_cast(pDigits)[n]), sizeof(uint16) ); + n = remainder % 100U; remainder /= 100U; + memcpy( reinterpret_cast(pDst - 3), &(reinterpret_cast(pDigits)[n]), sizeof(uint16) ); Assert( remainder < 100U ); - *reinterpret_cast(pDst - 1 - 4) = reinterpret_cast(pDigits)[remainder]; + memcpy( reinterpret_cast(pDst - 5), &(reinterpret_cast(pDigits)[remainder]), sizeof(uint16) ); pDst -= 6; } else { - uint n = remainder % 100U; remainder /= 100U; *reinterpret_cast(pDst - 1) = reinterpret_cast(pDigits)[n]; --pDst; if ( ( n >= 10 ) || ( remainder ) ) --pDst; + uint n = remainder % 100U; remainder /= 100U; + memcpy( reinterpret_cast(pDst - 1), &(reinterpret_cast(pDigits)[n]), sizeof(uint16) ); + --pDst; if ( ( n >= 10 ) || ( remainder ) ) --pDst; + if ( remainder ) { - n = remainder % 100U; remainder /= 100U; *reinterpret_cast(pDst - 1) = reinterpret_cast(pDigits)[n]; --pDst; if ( ( n >= 10 ) || ( remainder ) ) --pDst; + n = remainder % 100U; remainder /= 100U; + memcpy( reinterpret_cast(pDst - 1), &(reinterpret_cast(pDigits)[n]), sizeof(uint16) ); + + --pDst; if ( ( n >= 10 ) || ( remainder ) ) --pDst; if ( remainder ) { Assert( remainder < 100U ); - *reinterpret_cast(pDst - 1) = reinterpret_cast(pDigits)[remainder]; --pDst; if ( remainder >= 10 ) --pDst; + memcpy( reinterpret_cast(pDst - 1), &(reinterpret_cast(pDigits)[remainder]), sizeof(uint16) ); + --pDst; if ( remainder >= 10 ) --pDst; } } } diff --git a/togles/linuxwin/glentrypoints.cpp b/togles/linuxwin/glentrypoints.cpp index e4f78713..2d48c251 100644 --- a/togles/linuxwin/glentrypoints.cpp +++ b/togles/linuxwin/glentrypoints.cpp @@ -206,7 +206,7 @@ void ToGLDisconnectLibraries() static void GetOpenGLVersion(int *major, int *minor, int *patch) { *major = *minor = *patch = 0; - static CDynamicFunctionOpenGL< true, const GLubyte *( APIENTRY *)(GLenum name), const GLubyte * > glGetString("glGetString"); + static CDynamicFunctionOpenGL< true, const GLubyte *( _APIENTRY *)(GLenum name), const GLubyte * > glGetString("glGetString"); if (glGetString) { const char *version = (const char *) glGetString(GL_VERSION); @@ -271,7 +271,7 @@ static bool CheckOpenGLExtension_internal(const char *ext, const int coremajor, } // okay, see if the GL_EXTENSIONS string reports it. - static CDynamicFunctionOpenGL< true, const GLubyte *( APIENTRY *)(GLenum name), const GLubyte * > glGetString("glGetString"); + static CDynamicFunctionOpenGL< true, const GLubyte *( _APIENTRY *)(GLenum name), const GLubyte * > glGetString("glGetString"); if (!glGetString) return false; @@ -284,7 +284,7 @@ static bool CheckOpenGLExtension_internal(const char *ext, const int coremajor, #if _WIN32 if (!ptr) { - static CDynamicFunctionOpenGL< true, const char *( APIENTRY *)( ), const char * > wglGetExtensionsStringEXT("wglGetExtensionsStringEXT"); + static CDynamicFunctionOpenGL< true, const char *( _APIENTRY *)( ), const char * > wglGetExtensionsStringEXT("wglGetExtensionsStringEXT"); if (wglGetExtensionsStringEXT) { extensions = wglGetExtensionsStringEXT(); diff --git a/togles/linuxwin/glmgr.cpp b/togles/linuxwin/glmgr.cpp index 7fb01e38..01a2bf72 100644 --- a/togles/linuxwin/glmgr.cpp +++ b/togles/linuxwin/glmgr.cpp @@ -221,7 +221,7 @@ void APIENTRY GL_Debug_Output_Callback(GLenum source, GLenum type, GLuint id, GL return; } - if ( gl_debug_output.GetBool() || type == GL_DEBUG_TYPE_ERROR_ARB ) + if ( gl_debug_output.GetBool() || type == GL_DEBUG_TYPE_ERROR_ARB || type == GL_DEBUG_SEVERITY_MEDIUM_ARB ) { Msg( "GL: [%s][%s][%s][%d]: %s\n", sSource, sType, sSeverity, id, message ); } diff --git a/vphysics/physics_collide.cpp b/vphysics/physics_collide.cpp index 6a6d185b..1c2609f6 100644 --- a/vphysics/physics_collide.cpp +++ b/vphysics/physics_collide.cpp @@ -1641,8 +1641,13 @@ void CPhysicsCollision::VCollideLoad( vcollide_t *pOutput, int solidCount, const memcpy( &size, pBuffer + position, sizeof(int) ); position += sizeof(int); - pOutput->solids[i] = CPhysCollide::UnserializeFromBuffer( pBuffer + position, size, i, swap ); + char *tmpbuf = new char[size]; + memcpy(tmpbuf, pBuffer + position, size); + + pOutput->solids[i] = CPhysCollide::UnserializeFromBuffer( tmpbuf, size, i, swap ); position += size; + + delete[] tmpbuf; } END_IVP_ALLOCATION();