void b3GpuBatchingPgsSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem inertiaBuf, int numContacts, cl_mem contactBuf, const b3Config& config, int static0Index)
{
B3_PROFILE("solveContacts");
m_data->m_bodyBufferGPU->setFromOpenCLBuffer(bodyBuf,numBodies);
m_data->m_inertiaBufferGPU->setFromOpenCLBuffer(inertiaBuf,numBodies);
m_data->m_pBufContactOutGPU->setFromOpenCLBuffer(contactBuf,numContacts);
if (optionalSortContactsDeterminism)
{
if (gpuSortContactsDeterminism)
{
B3_PROFILE("GPU Sort contact constraints (determinism)");
m_data->m_pBufContactOutGPUCopy->resize(numContacts);
m_data->m_contactKeyValues->resize(numContacts);
m_data->m_pBufContactOutGPU->copyToCL(m_data->m_pBufContactOutGPUCopy->getBufferCL(),numContacts,0,0);
{
b3LauncherCL launcher(m_data->m_queue, m_data->m_setDeterminismSortDataChildShapeBKernel);
launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
launcher.setConst(numContacts);
launcher.launch1D( numContacts, 64 );
}
m_data->m_solverGPU->m_sort32->execute(*m_data->m_contactKeyValues);
{
b3LauncherCL launcher(m_data->m_queue, m_data->m_setDeterminismSortDataChildShapeAKernel);
launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
launcher.setConst(numContacts);
launcher.launch1D( numContacts, 64 );
}
m_data->m_solverGPU->m_sort32->execute(*m_data->m_contactKeyValues);
{
b3LauncherCL launcher(m_data->m_queue, m_data->m_setDeterminismSortDataBodyBKernel);
launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
launcher.setConst(numContacts);
launcher.launch1D( numContacts, 64 );
}
m_data->m_solverGPU->m_sort32->execute(*m_data->m_contactKeyValues);
{
b3LauncherCL launcher(m_data->m_queue, m_data->m_setDeterminismSortDataBodyAKernel);
launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
launcher.setConst(numContacts);
launcher.launch1D( numContacts, 64 );
}
m_data->m_solverGPU->m_sort32->execute(*m_data->m_contactKeyValues);
{
B3_PROFILE("gpu reorderContactKernel (determinism)");
b3Int4 cdata;
cdata.x = numContacts;
//b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_data->m_pBufContactOutGPU->getBufferCL() ), b3BufferInfoCL( m_data->m_solverGPU->m_contactBuffer2->getBufferCL())
// , b3BufferInfoCL( m_data->m_solverGPU->m_sortDataBuffer->getBufferCL()) };
b3LauncherCL launcher(m_data->m_queue,m_data->m_solverGPU->m_reorderContactKernel);
launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
launcher.setBuffer(m_data->m_pBufContactOutGPU->getBufferCL());
launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
launcher.setConst( cdata );
launcher.launch1D( numContacts, 64 );
}
} else
{
B3_PROFILE("CPU Sort contact constraints (determinism)");
b3AlignedObjectArray<b3Contact4> cpuConstraints;
m_data->m_pBufContactOutGPU->copyToHost(cpuConstraints);
bool sort = true;
if (sort)
{
cpuConstraints.quickSort(b3ContactCmp);
for (int i=0;i<cpuConstraints.size();i++)
{
cpuConstraints[i].m_batchIdx = i;
}
}
m_data->m_pBufContactOutGPU->copyFromHost(cpuConstraints);
if (m_debugOutput==100)
{
for (int i=0;i<cpuConstraints.size();i++)
{
printf("c[%d].m_bodyA = %d, m_bodyB = %d, batchId = %d\n",i,cpuConstraints[i].m_bodyAPtrAndSignBit,cpuConstraints[i].m_bodyBPtrAndSignBit, cpuConstraints[i].m_batchIdx);
}
}
m_debugOutput++;
}
}
int nContactOut = m_data->m_pBufContactOutGPU->size();
bool useSolver = true;
if (useSolver)
{
float dt=1./60.;
b3ConstraintCfg csCfg( dt );
csCfg.m_enableParallelSolve = true;
csCfg.m_batchCellSize = 6;
csCfg.m_staticIdx = static0Index;
b3OpenCLArray<b3RigidBodyCL>* bodyBuf = m_data->m_bodyBufferGPU;
void* additionalData = 0;//m_data->m_frictionCGPU;
const b3OpenCLArray<b3InertiaCL>* shapeBuf = m_data->m_inertiaBufferGPU;
b3OpenCLArray<b3GpuConstraint4>* contactConstraintOut = m_data->m_contactCGPU;
int nContacts = nContactOut;
int maxNumBatches = 0;
{
if( m_data->m_solverGPU->m_contactBuffer2)
{
m_data->m_solverGPU->m_contactBuffer2->resize(nContacts);
}
if( m_data->m_solverGPU->m_contactBuffer2 == 0 )
{
m_data->m_solverGPU->m_contactBuffer2 = new b3OpenCLArray<b3Contact4>(m_data->m_context,m_data->m_queue, nContacts );
m_data->m_solverGPU->m_contactBuffer2->resize(nContacts);
}
//clFinish(m_data->m_queue);
{
B3_PROFILE("batching");
//@todo: just reserve it, without copy of original contact (unless we use warmstarting)
const b3OpenCLArray<b3RigidBodyCL>* bodyNative = bodyBuf;
{
//b3OpenCLArray<b3RigidBodyCL>* bodyNative = b3OpenCLArrayUtils::map<adl::TYPE_CL, true>( data->m_device, bodyBuf );
//b3OpenCLArray<b3Contact4>* contactNative = b3OpenCLArrayUtils::map<adl::TYPE_CL, true>( data->m_device, contactsIn );
const int sortAlignment = 512; // todo. get this out of sort
if( csCfg.m_enableParallelSolve )
{
int sortSize = B3NEXTMULTIPLEOF( nContacts, sortAlignment );
b3OpenCLArray<unsigned int>* countsNative = m_data->m_solverGPU->m_numConstraints;
b3OpenCLArray<unsigned int>* offsetsNative = m_data->m_solverGPU->m_offsets;
if (gpuSetSortData)
{ // 2. set cell idx
B3_PROFILE("GPU set cell idx");
struct CB
{
int m_nContacts;
int m_staticIdx;
float m_scale;
b3Int4 m_nSplit;
};
b3Assert( sortSize%64 == 0 );
CB cdata;
cdata.m_nContacts = nContacts;
cdata.m_staticIdx = csCfg.m_staticIdx;
cdata.m_scale = 1.f/csCfg.m_batchCellSize;
cdata.m_nSplit.x = B3_SOLVER_N_SPLIT_X;
cdata.m_nSplit.y = B3_SOLVER_N_SPLIT_Y;
cdata.m_nSplit.z = B3_SOLVER_N_SPLIT_Z;
m_data->m_solverGPU->m_sortDataBuffer->resize(nContacts);
b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_data->m_pBufContactOutGPU->getBufferCL() ), b3BufferInfoCL( bodyBuf->getBufferCL()), b3BufferInfoCL( m_data->m_solverGPU->m_sortDataBuffer->getBufferCL()) };
b3LauncherCL launcher(m_data->m_queue, m_data->m_solverGPU->m_setSortDataKernel );
launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
launcher.setConst( cdata.m_nContacts );
launcher.setConst( cdata.m_scale );
launcher.setConst(cdata.m_nSplit);
launcher.setConst(cdata.m_staticIdx);
launcher.launch1D( sortSize, 64 );
} else
{
m_data->m_solverGPU->m_sortDataBuffer->resize(nContacts);
b3AlignedObjectArray<b3SortData> sortDataCPU;
m_data->m_solverGPU->m_sortDataBuffer->copyToHost(sortDataCPU);
b3AlignedObjectArray<b3Contact4> contactCPU;
m_data->m_pBufContactOutGPU->copyToHost(contactCPU);
b3AlignedObjectArray<b3RigidBodyCL> bodiesCPU;
bodyBuf->copyToHost(bodiesCPU);
float scale = 1.f/csCfg.m_batchCellSize;
b3Int4 nSplit;
nSplit.x = B3_SOLVER_N_SPLIT_X;
nSplit.y = B3_SOLVER_N_SPLIT_Y;
nSplit.z = B3_SOLVER_N_SPLIT_Z;
SetSortDataCPU(&contactCPU[0], &bodiesCPU[0], &sortDataCPU[0], nContacts,scale,nSplit,csCfg.m_staticIdx);
m_data->m_solverGPU->m_sortDataBuffer->copyFromHost(sortDataCPU);
}
if (gpuRadixSort)
{ // 3. sort by cell idx
B3_PROFILE("gpuRadixSort");
//int n = B3_SOLVER_N_SPLIT*B3_SOLVER_N_SPLIT;
//int sortBit = 32;
//if( n <= 0xffff ) sortBit = 16;
//if( n <= 0xff ) sortBit = 8;
//adl::RadixSort<adl::TYPE_CL>::execute( data->m_sort, *data->m_sortDataBuffer, sortSize );
//adl::RadixSort32<adl::TYPE_CL>::execute( data->m_sort32, *data->m_sortDataBuffer, sortSize );
b3OpenCLArray<b3SortData>& keyValuesInOut = *(m_data->m_solverGPU->m_sortDataBuffer);
this->m_data->m_solverGPU->m_sort32->execute(keyValuesInOut);
} else
{
b3OpenCLArray<b3SortData>& keyValuesInOut = *(m_data->m_solverGPU->m_sortDataBuffer);
b3AlignedObjectArray<b3SortData> hostValues;
keyValuesInOut.copyToHost(hostValues);
hostValues.quickSort(sortfnc);
keyValuesInOut.copyFromHost(hostValues);
}
{
// 4. find entries
B3_PROFILE("gpuBoundSearch");
m_data->m_solverGPU->m_search->execute(*m_data->m_solverGPU->m_sortDataBuffer,nContacts,*countsNative,B3_SOLVER_N_CELLS,b3BoundSearchCL::COUNT);
//adl::BoundSearch<adl::TYPE_CL>::execute( data->m_search, *data->m_sortDataBuffer, nContacts, *countsNative,
// B3_SOLVER_N_SPLIT*B3_SOLVER_N_SPLIT, adl::BoundSearchBase::COUNT );
//unsigned int sum;
m_data->m_solverGPU->m_scan->execute(*countsNative,*offsetsNative, B3_SOLVER_N_CELLS);//,&sum );
//printf("sum = %d\n",sum);
}
if (nContacts)
{ // 5. sort constraints by cellIdx
{
B3_PROFILE("gpu m_reorderContactKernel");
b3Int4 cdata;
cdata.x = nContacts;
b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_data->m_pBufContactOutGPU->getBufferCL() ), b3BufferInfoCL( m_data->m_solverGPU->m_contactBuffer2->getBufferCL())
, b3BufferInfoCL( m_data->m_solverGPU->m_sortDataBuffer->getBufferCL()) };
b3LauncherCL launcher(m_data->m_queue,m_data->m_solverGPU->m_reorderContactKernel);
launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
launcher.setConst( cdata );
launcher.launch1D( nContacts, 64 );
}
}
}
}
//clFinish(m_data->m_queue);
// {
// b3AlignedObjectArray<unsigned int> histogram;
// m_data->m_solverGPU->m_numConstraints->copyToHost(histogram);
// printf(",,,\n");
// }
if (nContacts)
{
B3_PROFILE("gpu m_copyConstraintKernel");
b3Int4 cdata; cdata.x = nContacts;
b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_data->m_solverGPU->m_contactBuffer2->getBufferCL() ), b3BufferInfoCL( m_data->m_pBufContactOutGPU->getBufferCL() ) };
b3LauncherCL launcher(m_data->m_queue, m_data->m_solverGPU->m_copyConstraintKernel );
launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
launcher.setConst( cdata );
launcher.launch1D( nContacts, 64 );
//we use the clFinish for proper benchmark/profile
clFinish(m_data->m_queue);
}
bool compareGPU = false;
if (nContacts)
{
if (b3GpuBatchContacts)
{
B3_PROFILE("gpu batchContacts");
maxNumBatches = 150;//250;
m_data->m_solverGPU->batchContacts( m_data->m_pBufContactOutGPU, nContacts, m_data->m_solverGPU->m_numConstraints, m_data->m_solverGPU->m_offsets, csCfg.m_staticIdx );
} else
{
B3_PROFILE("cpu batchContacts");
b3AlignedObjectArray<b3Contact4> cpuContacts;
b3OpenCLArray<b3Contact4>* contactsIn = m_data->m_solverGPU->m_contactBuffer2;
contactsIn->copyToHost(cpuContacts);
b3OpenCLArray<unsigned int>* countsNative = m_data->m_solverGPU->m_numConstraints;
b3OpenCLArray<unsigned int>* offsetsNative = m_data->m_solverGPU->m_offsets;
b3AlignedObjectArray<unsigned int> nNativeHost;
b3AlignedObjectArray<unsigned int> offsetsNativeHost;
{
B3_PROFILE("countsNative/offsetsNative copyToHost");
countsNative->copyToHost(nNativeHost);
offsetsNative->copyToHost(offsetsNativeHost);
}
int numNonzeroGrid=0;
{
B3_PROFILE("batch grid");
for(int i=0; i<B3_SOLVER_N_CELLS; i++)
{
int n = (nNativeHost)[i];
int offset = (offsetsNativeHost)[i];
if( n )
{
numNonzeroGrid++;
//printf("cpu batch cell %d\n",i);
int simdWidth =numBodies+1;//-1;//64;//-1;//32;
//int numBatches = sortConstraintByBatch( &cpuContacts[0]+offset, n, simdWidth,csCfg.m_staticIdx ,numBodies); // on GPU
//int numBatches = sortConstraintByBatch2( &cpuContacts[0]+offset, n, simdWidth,csCfg.m_staticIdx ,numBodies); // on GPU
int numBatches = sortConstraintByBatch3( &cpuContacts[0]+offset, n, simdWidth,csCfg.m_staticIdx ,numBodies); // on GPU
maxNumBatches = b3Max(numBatches,maxNumBatches);
static int globalMaxBatch = 0;
if (maxNumBatches>globalMaxBatch )
{
globalMaxBatch = maxNumBatches;
b3Printf("maxNumBatches = %d\n",maxNumBatches);
}
//we use the clFinish for proper benchmark/profile
clFinish(m_data->m_queue);
}
}
}
{
B3_PROFILE("m_contactBuffer->copyFromHost");
m_data->m_solverGPU->m_contactBuffer2->copyFromHost((b3AlignedObjectArray<b3Contact4>&)cpuContacts);
}
}
}
//printf("maxNumBatches = %d\n", maxNumBatches);
if (nContacts)
{
//B3_PROFILE("gpu convertToConstraints");
m_data->m_solverGPU->convertToConstraints( bodyBuf,
shapeBuf, m_data->m_solverGPU->m_contactBuffer2,
contactConstraintOut,
additionalData, nContacts,
(b3SolverBase::ConstraintCfg&) csCfg );
clFinish(m_data->m_queue);
}
}
}
if (1)
{
int numIter = 4;
m_data->m_solverGPU->m_nIterations = numIter;//10
if (b3GpuSolveConstraint)
{
B3_PROFILE("GPU solveContactConstraint");
/*m_data->m_solverGPU->solveContactConstraint(
m_data->m_bodyBufferGPU,
m_data->m_inertiaBufferGPU,
m_data->m_contactCGPU,0,
nContactOut ,
maxNumBatches);
*/
solveContactConstraint(
m_data->m_bodyBufferGPU,
m_data->m_inertiaBufferGPU,
m_data->m_contactCGPU,0,
nContactOut ,
maxNumBatches,numIter);
}
else
{
B3_PROFILE("Host solveContactConstraint");
m_data->m_solverGPU->solveContactConstraintHost(m_data->m_bodyBufferGPU, m_data->m_inertiaBufferGPU, m_data->m_contactCGPU,0, nContactOut ,maxNumBatches);
}
}
#if 0
if (0)
{
B3_PROFILE("read body velocities back to CPU");
//read body updated linear/angular velocities back to CPU
m_data->m_bodyBufferGPU->read(
m_data->m_bodyBufferCPU->m_ptr,numOfConvexRBodies);
adl::DeviceUtils::waitForCompletion( m_data->m_deviceCL );
}
#endif
}
}
