void FindCliff( size_t nOps, FILE* plot )
{
size_t nGroups = 32;
size_t nThreadsPerGroup=60;
std::vector<GEN::Instruction> ops;
ConstructShader(ops,nOps,nThreadsPerGroup);
GEN::Encoder enc;
HAXWell::Blob isa;
isa.SetLength( enc.GetBufferSize(ops.size()) );
isa.SetLength( enc.Encode( isa.GetBytes(), ops.data(), ops.size() ) );
//PrintISA(stdout,isa);
int CURBE[HAXWell::MAX_DISPATCH_COUNT][8];
for( size_t i=0; i<HAXWell::MAX_DISPATCH_COUNT; i++ )
for( size_t j=0; j<8; j++ )
CURBE[i][j] = 2*i;
HAXWell::ShaderArgs args;
args.nDispatchThreadCount = nThreadsPerGroup;
args.nSIMDMode = 8;
args.nCURBEAllocsPerThread = 1;
args.pCURBE = CURBE;
args.nIsaLength = isa.GetLength();
args.pIsa = isa.GetBytes();
HAXWell::ShaderHandle hShader = HAXWell::CreateShader(args);
HAXWell::BufferHandle hBuffer = HAXWell::CreateBuffer( 0, 32*nGroups*args.nDispatchThreadCount );
HAXWell::TimerHandle h = HAXWell::BeginTimer();
HAXWell::DispatchShader( hShader, &hBuffer,1,nGroups );
HAXWell::EndTimer( h );
HAXWell::Finish();
HAXWell::timer_t nTime = HAXWell::ReadTimer(h);
size_t nThreads = nGroups*args.nDispatchThreadCount;
unsigned int* pBuff = (unsigned int*)HAXWell::MapBuffer(hBuffer);
size_t nAvg=0;
for( size_t i=0; i<nThreadsPerGroup*nGroups; i++ )
{
unsigned int startlo = pBuff[8*i+1];
unsigned int endlo = pBuff[8*i+3];
nAvg += endlo-startlo;
}
HAXWell::UnmapBuffer(hBuffer);
HAXWell::ReleaseShader(hShader);
HAXWell::ReleaseBuffer(hBuffer);
double latency = ((double)nAvg) / (nThreadsPerGroup*nGroups);
fprintf(plot, "%u, %f, %u\n", 16*ops.size(), latency, ops.size() );
printf("%u, %f, %u\n", 16*ops.size(), latency, ops.size() );
}
HAXWell::timer_t InstructionIssueTest( size_t nRegs, size_t simd )
{
size_t nGroups = 128;
size_t nThreadsPerGroup=60;
std::vector<GEN::Instruction> ops;
ConstructShader(ops,nRegs,nThreadsPerGroup, simd);
GEN::Encoder enc;
HAXWell::Blob isa;
isa.SetLength( enc.GetBufferSize(ops.size()) );
isa.SetLength( enc.Encode( isa.GetBytes(), ops.data(), ops.size() ) );
// PrintISA(stdout,isa);
int CURBE[HAXWell::MAX_DISPATCH_COUNT][8];
for( size_t i=0; i<HAXWell::MAX_DISPATCH_COUNT; i++ )
for( size_t j=0; j<8; j++ )
CURBE[i][j] = 2*i;
HAXWell::ShaderArgs args;
args.nDispatchThreadCount = nThreadsPerGroup;
args.nSIMDMode = 16;//(simd <= 8) ? 8 : 16; // dispatch mode really doesn't seem to matter
args.nCURBEAllocsPerThread = 1;
args.pCURBE = CURBE;
args.nIsaLength = isa.GetLength();
args.pIsa = isa.GetBytes();
HAXWell::ShaderHandle hShader = HAXWell::CreateShader(args);
HAXWell::BufferHandle hBuffer = HAXWell::CreateBuffer( 0, 32*nGroups*args.nDispatchThreadCount );
HAXWell::TimerHandle h = HAXWell::BeginTimer();
HAXWell::DispatchShader( hShader, &hBuffer,1,nGroups );
HAXWell::EndTimer( h );
HAXWell::Finish();
HAXWell::ReleaseBuffer(hBuffer);
HAXWell::ReleaseShader(hShader);
return HAXWell::ReadTimer(h);
}
double ScatterReadTest( size_t nDivergent )
{
size_t nGroups = 32;
size_t nThreadsPerGroup=60;
std::vector<GEN::Instruction> ops;
ConstructShader(ops,nThreadsPerGroup);
GEN::Encoder enc;
HAXWell::Blob isa;
isa.SetLength( enc.GetBufferSize(ops.size()) );
isa.SetLength( enc.Encode( isa.GetBytes(), ops.data(), ops.size() ) );
//PrintISA(stdout,isa);
int CURBE[HAXWell::MAX_DISPATCH_COUNT][2][8];
for( size_t i=0; i<HAXWell::MAX_DISPATCH_COUNT; i++ )
{
// first CURBE reg is offset into timings buffer
for( size_t j=0; j<8; j++ )
CURBE[i][0][j] = 2*i;
// second CURBE reg is per-channel read offsets for scatter read
for( size_t j=0; j<8; j++ )
CURBE[i][1][j] = 8*i + j;
// send a certain number of them to completely different cache lines
for( size_t j=0; j<nDivergent; j++ )
CURBE[i][1][j] += 16*j;
#ifdef SAME_ADDRESS // nah, use zero instead....
for( size_t j=0; j<8; j++ )
CURBE[i][1][j] = 0;
#endif
}
HAXWell::ShaderArgs args;
args.nDispatchThreadCount = nThreadsPerGroup;
args.nSIMDMode = 16;
args.nCURBEAllocsPerThread = 2;
args.pCURBE = CURBE;
args.nIsaLength = isa.GetLength();
args.pIsa = isa.GetBytes();
HAXWell::ShaderHandle hShader = HAXWell::CreateShader(args);
HAXWell::BufferHandle hBuffers[] =
{
HAXWell::CreateBuffer( 0, 32*nGroups*args.nDispatchThreadCount ),
HAXWell::CreateBuffer( 0, 32*nGroups*args.nDispatchThreadCount*1024 ) ,
};
HAXWell::DispatchShader( hShader, hBuffers,2,nGroups );
HAXWell::Finish();
unsigned int* pBuff = (unsigned int*)HAXWell::MapBuffer(hBuffers[0]);
size_t nAvg=0;
for( size_t i=0; i<nThreadsPerGroup*nGroups; i++ )
{
unsigned int startlo = pBuff[8*i];
unsigned int endlo = pBuff[8*i+2];
nAvg += endlo-startlo;
}
HAXWell::UnmapBuffer(hBuffers[0]);
HAXWell::ReleaseBuffer(hBuffers[0]);
HAXWell::ReleaseBuffer(hBuffers[1]);
HAXWell::ReleaseShader(hShader);
return (nAvg)/(double)(nGroups*nThreadsPerGroup);
}