1
0
mirror of https://github.com/alliedmodders/hl2sdk.git synced 2024-12-22 09:38:56 +08:00
hl2sdk/mathlib/ssenoise.cpp
2014-10-30 12:30:57 -04:00

110 lines
4.2 KiB
C++

//========= Copyright Valve Corporation, All rights reserved. ============//
//
// Purpose: Fast low quality noise suitable for real time use
//
//=====================================================================================//
#include <math.h>
#include <float.h> // Needed for FLT_EPSILON
#include "basetypes.h"
#include <memory.h>
#include "tier0/dbg.h"
#include "mathlib/mathlib.h"
#include "mathlib/vector.h"
#include "mathlib/ssemath.h"
// memdbgon must be the last include file in a .cpp file!!!
#include "tier0/memdbgon.h"
#include "noisedata.h"
#define MAGIC_NUMBER (1<<15) // gives 8 bits of fraction
static fltx4 Four_MagicNumbers = { MAGIC_NUMBER, MAGIC_NUMBER, MAGIC_NUMBER, MAGIC_NUMBER };
static ALIGN16 int32 idx_mask[4]= {0xffff, 0xffff, 0xffff, 0xffff};
#define MASK255 (*((fltx4 *)(& idx_mask )))
// returns 0..1
static inline float GetLatticePointValue( int idx_x, int idx_y, int idx_z )
{
NOTE_UNUSED(perm_d);
NOTE_UNUSED(impulse_ycoords);
NOTE_UNUSED(impulse_zcoords);
int ret_idx = perm_a[idx_x & 0xff];
ret_idx = perm_b[( idx_y + ret_idx ) & 0xff];
ret_idx = perm_c[( idx_z + ret_idx ) & 0xff];
return impulse_xcoords[ret_idx];
}
fltx4 NoiseSIMD( const fltx4 & x, const fltx4 & y, const fltx4 & z )
{
// use magic to convert to integer index
fltx4 x_idx = AndSIMD( MASK255, AddSIMD( x, Four_MagicNumbers ) );
fltx4 y_idx = AndSIMD( MASK255, AddSIMD( y, Four_MagicNumbers ) );
fltx4 z_idx = AndSIMD( MASK255, AddSIMD( z, Four_MagicNumbers ) );
fltx4 lattice000 = Four_Zeros, lattice001 = Four_Zeros, lattice010 = Four_Zeros, lattice011 = Four_Zeros;
fltx4 lattice100 = Four_Zeros, lattice101 = Four_Zeros, lattice110 = Four_Zeros, lattice111 = Four_Zeros;
// FIXME: Converting the input vectors to int indices will cause load-hit-stores (48 bytes)
// Converting the indexed noise values back to vectors will cause more (128 bytes)
// The noise table could store vectors if we chunked it into 2x2x2 blocks.
fltx4 xfrac = Four_Zeros, yfrac = Four_Zeros, zfrac = Four_Zeros;
#define DOPASS(i) \
{ unsigned int xi = SubInt( x_idx, i ); \
unsigned int yi = SubInt( y_idx, i ); \
unsigned int zi = SubInt( z_idx, i ); \
SubFloat( xfrac, i ) = (xi & 0xff)*(1.0/256.0); \
SubFloat( yfrac, i ) = (yi & 0xff)*(1.0/256.0); \
SubFloat( zfrac, i ) = (zi & 0xff)*(1.0/256.0); \
xi>>=8; \
yi>>=8; \
zi>>=8; \
\
SubFloat( lattice000, i ) = GetLatticePointValue( xi,yi,zi ); \
SubFloat( lattice001, i ) = GetLatticePointValue( xi,yi,zi+1 ); \
SubFloat( lattice010, i ) = GetLatticePointValue( xi,yi+1,zi ); \
SubFloat( lattice011, i ) = GetLatticePointValue( xi,yi+1,zi+1 ); \
SubFloat( lattice100, i ) = GetLatticePointValue( xi+1,yi,zi ); \
SubFloat( lattice101, i ) = GetLatticePointValue( xi+1,yi,zi+1 ); \
SubFloat( lattice110, i ) = GetLatticePointValue( xi+1,yi+1,zi ); \
SubFloat( lattice111, i ) = GetLatticePointValue( xi+1,yi+1,zi+1 ); \
}
DOPASS( 0 );
DOPASS( 1 );
DOPASS( 2 );
DOPASS( 3 );
// now, we have 8 lattice values for each of four points as m128s, and interpolant values for
// each axis in m128 form in [xyz]frac. Perfom the trilinear interpolation as SIMD ops
// first, do x interpolation
fltx4 l2d00 = AddSIMD( lattice000, MulSIMD( xfrac, SubSIMD( lattice100, lattice000 ) ) );
fltx4 l2d01 = AddSIMD( lattice001, MulSIMD( xfrac, SubSIMD( lattice101, lattice001 ) ) );
fltx4 l2d10 = AddSIMD( lattice010, MulSIMD( xfrac, SubSIMD( lattice110, lattice010 ) ) );
fltx4 l2d11 = AddSIMD( lattice011, MulSIMD( xfrac, SubSIMD( lattice111, lattice011 ) ) );
// now, do y interpolation
fltx4 l1d0 = AddSIMD( l2d00, MulSIMD( yfrac, SubSIMD( l2d10, l2d00 ) ) );
fltx4 l1d1 = AddSIMD( l2d01, MulSIMD( yfrac, SubSIMD( l2d11, l2d01 ) ) );
// final z interpolation
fltx4 rslt = AddSIMD( l1d0, MulSIMD( zfrac, SubSIMD( l1d1, l1d0 ) ) );
// map to 0..1
return MulSIMD( Four_Twos, SubSIMD( rslt, Four_PointFives ) );
}
fltx4 NoiseSIMD( FourVectors const &pos )
{
return NoiseSIMD( pos.x, pos.y, pos.z );
}