Accidently left some very slow copyToHost in the batching code, removing it makes it faster :-)

src/Bullet3OpenCL/RigidBody/b3GpuPgsContactSolver.cpp

@@ -895,7 +895,7 @@ void b3GpuPgsContactSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem
 								}
 								m_data->m_solverGPU->m_contactBuffer2->copyFromHost(outContacts);
 
-/*								"void ReorderContactKernel(__global struct b3Contact4Data* in, __global struct b3Contact4Data* out, __global int2* sortData, int4 cb )\n"
+								/*								"void ReorderContactKernel(__global struct b3Contact4Data* in, __global struct b3Contact4Data* out, __global int2* sortData, int4 cb )\n"
 								"{\n"
 								"	int nContacts = cb.x;\n"
 								"	int gIdx = GET_GLOBAL_IDX;\n"
@@ -934,11 +934,11 @@ void b3GpuPgsContactSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem
 
 				//clFinish(m_data->m_queue);
 
-//				{
-//				b3AlignedObjectArray<unsigned int> histogram;
-//				m_data->m_solverGPU->m_numConstraints->copyToHost(histogram);
-//				printf(",,,\n");
-//				}
+				//				{
+				//				b3AlignedObjectArray<unsigned int> histogram;
+				//				m_data->m_solverGPU->m_numConstraints->copyToHost(histogram);
+				//				printf(",,,\n");
+				//				}
 
 
 				if (nContacts)
@@ -949,7 +949,7 @@ void b3GpuPgsContactSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem
 						for (int i=0;i<nContacts;i++)
 						{
 							m_data->m_pBufContactOutGPU->copyFromOpenCLArray(*m_data->m_solverGPU->m_contactBuffer2);
-//							m_data->m_solverGPU->m_contactBuffer2->getBufferCL(); 
+							//							m_data->m_solverGPU->m_contactBuffer2->getBufferCL(); 
 							//						m_data->m_pBufContactOutGPU->getBufferCL() 
 						}
 
@@ -980,6 +980,7 @@ void b3GpuPgsContactSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem
 						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 );
+						clFinish(m_data->m_queue);
 					} else
 					{
 						B3_PROFILE("cpu batchContacts");
@@ -1003,25 +1004,16 @@ void b3GpuPgsContactSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem
 						int numNonzeroGrid=0;
 
 						{
-							B3_PROFILE("batch grid");
+							B3_PROFILE("cpu 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 )
@@ -1029,13 +1021,11 @@ void b3GpuPgsContactSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem
 										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");
@@ -1051,7 +1041,7 @@ void b3GpuPgsContactSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem
 
 				if (nContacts)
 				{
-                    //B3_PROFILE("gpu convertToConstraints");
+					B3_PROFILE("gpu convertToConstraints");
 					m_data->m_solverGPU->convertToConstraints( bodyBuf, 
 						shapeBuf, m_data->m_solverGPU->m_contactBuffer2,
 						contactConstraintOut, 

src/Bullet3OpenCL/RigidBody/b3Solver.cpp

@@ -978,6 +978,9 @@ void b3Solver::convertToConstraints( const b3OpenCLArray<b3RigidBodyCL>* bodyBuf
 		cdata.m_positionDrift = cfg.m_positionDrift;
 		cdata.m_positionConstraintCoeff = cfg.m_positionConstraintCoeff;
 
+		
+		if (convertConstraintOnCpu)
+		{
 			b3AlignedObjectArray<b3RigidBodyCL> gBodies;
 		bodyBuf->copyToHost(gBodies);
 
@@ -990,8 +993,6 @@ void b3Solver::convertToConstraints( const b3OpenCLArray<b3RigidBodyCL>* bodyBuf
 		b3AlignedObjectArray<b3GpuConstraint4> gConstraintOut;
 		gConstraintOut.resize(nContacts);
 		
-		if (convertConstraintOnCpu)
-		{
 			B3_PROFILE("cpu contactToConstraintKernel");
 			for (int gIdx=0;gIdx<nContacts;gIdx++)
 			{