csgo-2018-source/common/ps3/vjobchain.cpp

383 lines
13 KiB
C++
Raw Normal View History

2021-07-25 12:11:47 +08:00
//========== Copyright <20> Valve Corporation, All rights reserved. ========
#include "vjobchain.h"
#include "vjobutils.h"
#include "vjobs/root.h"
#include "tier1/strtools.h"
#include "tier0/miniprofiler.h"
CMiniProfiler g_mpRun, g_mpJoin, g_mpPush, g_mpPush2;
int VjobChain::Init( VJobsRoot * pRoot, uint nMaxContention, const char* pFormatName, ... )
{
Assert( !( uintp( &m_guard ) & 0x7F ) );
{
va_list args
va_start( args, pFormatName );
V_vsnprintf( m_name, sizeof( m_name ), pFormatName, args );
}
cell::Spurs::JobChainAttribute attr;
uint8_t nVjobChainPriority[8] = {0,12,12,12,12,12,0,0}; // priority lower than the main job queue, in order to yield
attr.initialize( &attr, m_headBlock, 128, 1, nVjobChainPriority, nMaxContention, true, 0,1, false, 256, 1 );
attr.setName( m_name );
m_pLastBlock = NULL; // NOT running by default
CELL_MUST_SUCCEED( JobChain::createWithAttribute( &pRoot->m_spurs, &m_spursJobChain, &attr ) );
V_memset( &m_notifyArea, 0, sizeof( m_notifyArea ) );
V_memset( &m_jobNotify, 0, sizeof( m_jobNotify ) );
m_jobNotify.header = *( pRoot->m_pJobNotify );
m_jobNotify.header.useInOutBuffer = 1;
Assert( !( uint( &m_jobNotify ) & 63 ) ); // should be 64-byte aligned
AddInputDma( &m_jobNotify, sizeof( m_notifyArea ), &m_notifyArea );
m_notifyArea.m_nCopyFrom = 1;
m_notifyArea.m_nCopyTo = 0; // SPU will mark copyTo = 1, PPU will mark it back to 0; at this time, we may actually mark the notify as completed
m_jobNotify.workArea.userData[1] = 0; // function: default
CELL_MUST_SUCCEED( m_guard.initialize( &m_spursJobChain, &m_guard, 1 /*notifyCount*/, 1 /*requestSpuCount(ignored)*/, 1 /*autoReset*/ ) );
m_headBlock[0] = CELL_SPURS_JOB_COMMAND_SYNC; // wait for all previous list commands to finish
m_headBlock[1] = CELL_SPURS_JOB_COMMAND_JOB( &m_jobNotify );
m_headBlock[2] = CELL_SPURS_JOB_COMMAND_GUARD( &m_guard );
m_headBlock[BLOCK_COMMANDS] = ( uint64 )-1ll;
CELL_MUST_SUCCEED( m_spursJobChain.run() );
m_nSpinWaitNotify = 0;
return CELL_OK;
}
int VjobChain::Run()
{
if( !IsRunning() )
{
CMiniProfilerGuard mpg( &g_mpRun );
Assert( m_notifyArea.m_nCopyTo ); // the jobchain must be joined if its not in Running state
m_notifyArea.m_nCopyTo = 0;
m_pLastBlock = m_headBlock;
m_nCurrentBlockCommands = 3; // right after the SYNC-JOB(notify)-GUARD prefix
m_headBlock[m_nCurrentBlockCommands] = CELL_SPURS_JOB_COMMAND_JTS;
// __lwsync(); // make sure we complete sync reset and write JTS before notify - probably not necessary because the guard should have a barrier for sure
m_guard.notify(); // let the jobchain go
return CELL_OK;
}
else
{
return CELL_OK; // it's valid to try to run a running chain in our interface...
}
}
int VjobChain::End( )
{
if( IsRunning() )
{
Assert( m_pLastBlock[m_nCurrentBlockCommands] == CELL_SPURS_JOB_COMMAND_JTS );
Assert( m_notifyArea.m_nCopyTo == 0 ); // make sure we reset sync correctly
m_pLastBlock[m_nCurrentBlockCommands] = CELL_SPURS_JOB_COMMAND_RESET_PC( m_headBlock );
return 0;
}
else
return -1; // you should not end non-running instance
}
void VjobChain::Shutdown()
{
if( IsRunning() )
{
m_pLastBlock[m_nCurrentBlockCommands] = CELL_SPURS_JOB_COMMAND_END;
}
m_spursJobChain.shutdown();
m_spursJobChain.join();
}
int VjobChain::Join()
{
if( IsRunning() )
{
#ifdef _DEBUG
CellSpursJobChainInfo info;
m_spursJobChain.getInfo( &info );
#endif
CMiniProfilerGuard mpg( &g_mpJoin );
// wait for reset sync notification to come through
volatile job_notify::NotifyArea_t *pNotify = &m_notifyArea;
Assert( pNotify->m_nCopyFrom );
while( !pNotify->m_nCopyTo )
{
++m_nSpinWaitNotify;
}
if( m_nSpinWaitNotify )
{
// <HACK> Sergiy : I'm taking this out for now because jobchain double-buffering is effectively temporarily hosed
// Warning( "VjobChain %s: stall in join, %d spins\n", m_name, m_nSpinWaitNotify );
m_nSpinWaitNotify = 0;
}
// free up the memory of the jobs that are now known to have dispatched
if( m_headBlock != m_pLastBlock )
{
uint64 *pBlock = m_headBlock;
do
{
Assert( pBlock[BLOCK_COMMANDS] == (uint64)-1ll );
uint64 eaNext = pBlock[BLOCK_COMMANDS - 1];
Assert( ( eaNext & 0xFFFFFFFF00000007ull ) == 3 );
pBlock = ( uint64 * )( uintp( eaNext ) & ~7 );
Assert( pBlock[BLOCK_COMMANDS] == (uint64)-1ll );
delete[]pBlock;
}
while( pBlock != m_pLastBlock );
}
m_pLastBlock = NULL; // idle state
return CELL_OK;
}
else
return 0; // valid to join twice;
}
void VjobChain::Push( uint64 nCommand )
{
Assert( IsRunning() );
uint64 * pNextCommand = &m_pLastBlock[m_nCurrentBlockCommands++]; // JTS to patch
if( m_nCurrentBlockCommands < BLOCK_COMMANDS )
{
CMiniProfilerGuard mpg( &g_mpPush );
pNextCommand[1] = CELL_SPURS_JOB_COMMAND_JTS;
__lwsync(); // ordering: create JobHeader, insert next JTS --> patch current JTS with JOB command
*pNextCommand = nCommand;
}
else
{
CMiniProfilerGuard mpg( &g_mpPush2 );
m_pLastBlock = new uint64[BLOCK_COMMANDS+1];
m_pLastBlock[BLOCK_COMMANDS] = ( uint64 )-1ll; // marker
m_pLastBlock[0] = nCommand;
m_pLastBlock[m_nCurrentBlockCommands = 1] = CELL_SPURS_JOB_COMMAND_JTS;
__lwsync(); // ordering: create JobHeader, allocate & reset new segment with JOB command in it --> patch JTS in old segment with NEXT command
*pNextCommand = CELL_SPURS_JOB_COMMAND_NEXT( m_pLastBlock );
}
}
void VjobChain::Push( const uint64 * nCommands, uint nCommandCount )
{
// todo: make it more optimal by removing extra lwsync's
for( uint i = 0; i < nCommandCount; ++i )
{
Push( nCommands[i] );
}
}
int VjobChain2::Init( VJobsRoot * pRoot, uint nMaxContention, const char* pName )
{
m_vjobChainRing = ( VjobChain * )MemAlloc_AllocAligned( VJOB_CHAINS * sizeof( VjobChain ), 128 );
for( uint i = 0; i < 2; ++i )
{
int nError = m_vjobChainRing[i].Init( pRoot, nMaxContention, "%s%d", pName, i );
if( nError )
return nError;
}
m_nCurrentChain = 0;
return 0;
}
void VjobChain2::Begin()
{
m_vjobChainRing[( m_nCurrentChain + 1 ) % VJOB_CHAINS].Join();
VjobChain & jobchain = Jobchain();
jobchain.Join(); // join the jobchain that we'll be using now
jobchain.Run();
}
void VjobChain2::End()
{
VjobChain & jobchain = Jobchain();
jobchain.End();
m_nCurrentChain = ( m_nCurrentChain + 1 ) % VJOB_CHAINS; // swap the job chain
}
void VjobChain2::Shutdown()
{
for( uint i = 0; i < 2; ++i )
{
m_vjobChainRing[i].Shutdown();
}
MemAlloc_FreeAligned( m_vjobChainRing );
}
int VjobChain3::Init( VJobsRoot * pRoot, uint nMaxContention, uint nMinCommandsPerBuffer, uint8_t nVjobChainPriority[8], const char* pName, uint nDmaTags )
{
m_pName = pName;
const uint nSizeOfJobDescriptor = 128, nMaxGrabbedJob = 4;
// we need at least 4 commands
uint nBufferSize = sizeof( VjobBufferHeader_t ) + sizeof( uint64 ) * MAX( nMinCommandsPerBuffer, VjobBuffer_t::VERBATIM_COMMAND_COUNT + 2 ); // +2 is for user's command and JTN
nBufferSize = AlignValue( nBufferSize, 128 );
m_nMaxCommandsPerBuffer = ( nBufferSize - sizeof( VjobBufferHeader_t ) ) / sizeof( uint64 );
uint nAllocationSize = sizeof( cell::Spurs::JobChain ) + nBufferSize * BUFFER_COUNT;
m_pSpursJobChain = ( cell::Spurs::JobChain* )MemAlloc_AllocAligned( nAllocationSize, 128 );
V_memset( m_pSpursJobChain, 0, nAllocationSize );
m_pBuffers[0] = ( VjobBuffer_t * )( m_pSpursJobChain + 1 );
m_nFrontBuffer = 0;
m_pFrontBuffer = m_pBuffers[0];
for( int i = 1; i < BUFFER_COUNT; ++i )
{
m_pBuffers[i] = ( VjobBuffer_t * )( uintp( m_pBuffers[ i - 1 ] ) + nBufferSize );
}
cell::Spurs::JobChainAttribute attr;
attr.initialize( &attr, m_pFrontBuffer->m_spursCommands, nSizeOfJobDescriptor, nMaxGrabbedJob, nVjobChainPriority, nMaxContention, true, nDmaTags, nDmaTags + 1, false, 256, 1 );
attr.setName( pName );
CELL_MUST_SUCCEED( JobChain::createWithAttribute( &pRoot->m_spurs, m_pSpursJobChain, &attr ) );
for( int i = 0; i < BUFFER_COUNT; ++i )
{
Assert( !( uintp( m_pBuffers[i] ) & 0x7F ) );
m_pBuffers[i]->Init( pRoot, m_pSpursJobChain );
}
StartCommandBuffer( 0, CELL_SPURS_JOB_COMMAND_NOP );
CELL_MUST_SUCCEED( m_pSpursJobChain->run() );
#ifdef _DEBUG
sys_timer_usleep( 100 );
#endif
m_nSpinWaitNotify = 0;
return CELL_OK;
}
void VjobBuffer_t::Init( VJobsRoot * pRoot, cell::Spurs::JobChain * pSpursJobChain )
{
m_jobNotify.header = *( pRoot->m_pJobNotify );
m_jobNotify.header.useInOutBuffer = 1;
AddInputDma( &m_jobNotify, sizeof( m_notifyArea ), &m_notifyArea );
m_notifyArea.m_nCopyFrom = 1;
// SPU will mark copyTo = 1, PPU will mark it back to 0; at this time, we may actually mark the notify as completed; 1 means "previous buffer is free"
// Then we'll start command buffer, which will reset the ready flag. But then we run the jobchain, which will run job_notify and set the flag back again, thus starting the ring
m_notifyArea.m_nCopyTo = 1;
m_jobNotify.workArea.userData[1] = 0; // function: default
uint nCommands = 0;
m_spursCommands[nCommands++] = CELL_SPURS_JOB_COMMAND_SYNC; // wait for all previous list commands to finish
m_spursCommands[nCommands++] = CELL_SPURS_JOB_COMMAND_JOB( &m_jobNotify );
#ifdef VJOBCHAIN3_GUARD
Assert( !( uintp( &m_guard ) & -128 ) );
CELL_MUST_SUCCEED( m_guard.initialize( pSpursJobChain, &m_guard, 1 /*notifyCount*/, 1 /*requestSpuCount(ignored)*/, 1 /*autoReset*/ ) );
m_spursCommands[nCommands++] = CELL_SPURS_JOB_COMMAND_GUARD( &m_guard );
#endif
#ifdef _DEBUG
m_jobNotify2.header = *( pRoot->m_pJobNotify );
m_jobNotify2.header.useInOutBuffer = 1;
AddInputDma( &m_jobNotify2, sizeof( m_notifyArea2 ), &m_notifyArea2 );
m_jobNotify2.workArea.userData[1] = 0; // function: default
m_notifyArea2.m_nCopyFrom = 1;
m_notifyArea2.m_nCopyTo = 0; // just for debugging, to see when this job gets executed
m_spursCommands[nCommands++] = CELL_SPURS_JOB_COMMAND_JOB( &m_jobNotify2 );
#endif
Assert( nCommands == VjobBuffer_t::VERBATIM_COMMAND_COUNT );
m_spursCommands[VjobBuffer_t::VERBATIM_COMMAND_COUNT] = CELL_SPURS_JOB_COMMAND_JTS;
}
void VjobChain3::WaitForEntryNotify( VjobBuffer_t * pBuffer )
{
volatile job_notify::NotifyArea_t *pNotify = &pBuffer->m_notifyArea;
Assert( pNotify->m_nCopyFrom );
while( !pNotify->m_nCopyTo )
{
++m_nSpinWaitNotify;
sys_timer_usleep( 30 );
}
if( m_nSpinWaitNotify )
{
Warning( "VjobChain %s: stall in WaitForEntryNotify, %d spins\n", m_pName, m_nSpinWaitNotify );
m_nSpinWaitNotify = 0;
}
}
uint64* VjobChain3::SwapCommandBuffer( uint64 nInsertCommand )
{
uint64 * pSpursIsSpinningHere = &m_pFrontBuffer->m_spursCommands[ m_nFrontBufferCommandCount ];
Assert( m_nFrontBufferCommandCount < m_nMaxCommandsPerBuffer );
uint nNext1Buffer = ( m_nFrontBuffer + 1 ) % BUFFER_COUNT, nNext2Buffer = ( m_nFrontBuffer + 2 ) % BUFFER_COUNT;
VjobBuffer_t * pNext1Buffer = m_pBuffers[ nNext1Buffer ], * pNext2Buffer = m_pBuffers[ nNext2Buffer ];
// before we can declare the next1 buffer "front", we need to make sure it's fully ready to accept commands, i.e. that it was fully read by SPURS
// for that, we check the next2 buffer notification area
WaitForEntryNotify( pNext1Buffer );
WaitForEntryNotify( pNext2Buffer );
// if next2 buffer has been notified, next1 must have been notified long ago
Assert( pNext1Buffer->m_notifyArea.m_nCopyTo );
uint64* pInsertionPoint = StartCommandBuffer( nNext1Buffer, nInsertCommand );
Assert( pNext1Buffer == m_pFrontBuffer );
// implicit lwsync is here
*pSpursIsSpinningHere = CELL_SPURS_JOB_COMMAND_NEXT( pNext1Buffer->m_spursCommands ); // jump to the next buffer
return pInsertionPoint;
}
uint64* VjobChain3::StartCommandBuffer( uint nNext1Buffer, uint64 nInsertCommand )
{
m_nFrontBuffer = nNext1Buffer;
m_pFrontBuffer = m_pBuffers[ nNext1Buffer ];
// the ready marker is presumed to be present; SPURS must have gone through this buffer in the previous ring, otherwise we can't use it
Assert( m_pFrontBuffer->m_notifyArea.m_nCopyTo == 1 );
// reset the ready marker; SPURS didn't get to this buffer yet (we're about to reuse them and we didn't jump to it yet)
m_pFrontBuffer->m_notifyArea.m_nCopyTo = 0;
#ifdef _DEBUG
m_pFrontBuffer->m_notifyArea2.m_nCopyFrom++;
m_pFrontBuffer->m_notifyArea2.m_nCopyTo = 0;
#endif
uint64 * pCommand = &m_pFrontBuffer->m_spursCommands[ VjobBuffer_t::VERBATIM_COMMAND_COUNT ];
*pCommand = nInsertCommand;
m_pFrontBuffer->m_spursCommands[ VjobBuffer_t::VERBATIM_COMMAND_COUNT + 1 ] = CELL_SPURS_JOB_COMMAND_JTS;
m_nFrontBufferCommandCount = VjobBuffer_t::VERBATIM_COMMAND_COUNT + 1;
#ifdef VJOBCHAIN3_GUARD
m_pFrontBuffer->m_guard.notify(); // let the jobchain go through
// implicit lwsync is here
#else
__lwsync();
#endif
return pCommand;
}
void VjobChain3::End()
{
Assert( m_nFrontBufferCommandCount < m_nMaxCommandsPerBuffer );
m_pFrontBuffer->m_spursCommands[ m_nFrontBufferCommandCount ] = CELL_SPURS_JOB_COMMAND_END;
m_pSpursJobChain->shutdown();
}
void VjobChain3::Join()
{
Assert( m_pFrontBuffer->m_spursCommands[ m_nFrontBufferCommandCount ] == CELL_SPURS_JOB_COMMAND_END );
m_pSpursJobChain->join();
MemAlloc_FreeAligned( m_pSpursJobChain );
m_pSpursJobChain = NULL;
}