The CUDA SDK 3.0 oclDotProduct is just an example. I attached the code adaption of oclDoProduct.cpp.
// -------------------------------------------------------- // Core sequence... copy input data to GPU, compute, copy results back /* Profiling test */ cl_event evts[3]; // Asynchronous write of data to GPU device shrLog("clEnqueueWriteBuffer (SrcA and SrcB)...\n"); ciErrNum = clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcA, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize * 4, srcA, 0, NULL, &evts[0]); //clFlush(cqCommandQueue); ciErrNum |= clEnqueueWriteBuffer(cqCommandQueue, cmDevSrcB, CL_FALSE, 0, sizeof(cl_float) * szGlobalWorkSize * 4, srcB, 0, NULL, &evts[1]); shrCheckErrorEX (ciErrNum, CL_SUCCESS, pCleanup); //clFlush(cqCommandQueue); // Launch kernel shrLog("clEnqueueNDRangeKernel (DotProduct)...\n"); ciErrNum = clEnqueueNDRangeKernel(cqCommandQueue, ckKernel, 1, NULL, &szGlobalWorkSize, &szLocalWorkSize, 0, NULL, &evts[2]); shrCheckErrorEX (ciErrNum, CL_SUCCESS, pCleanup); // Read back results and check accumulated errors shrLog("clEnqueueReadBuffer (Dst)...\n\n"); ciErrNum = clEnqueueReadBuffer(cqCommandQueue, cmDevDst, CL_TRUE, 0, sizeof(cl_float) * szGlobalWorkSize, dst, 0, NULL, NULL); shrCheckErrorEX (ciErrNum, CL_SUCCESS, pCleanup); /* Profiling test */ for(int i = 0; i < 3; i++){ cl_ulong ocl_strttime, ocl_stoptime, ocl_submittime, ocl_queuedTime; clGetEventProfilingInfo(evts, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &ocl_strttime, NULL); clGetEventProfilingInfo(evts, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &ocl_stoptime, NULL); clGetEventProfilingInfo(evts, CL_PROFILING_COMMAND_SUBMIT, sizeof(cl_ulong), &ocl_submittime, NULL); clGetEventProfilingInfo(evts, CL_PROFILING_COMMAND_QUEUED, sizeof(cl_ulong), &ocl_queuedTime, NULL); printf("CL_strtTime: %llu; CL_stopTime: %llu; CL_submitTime: %llu; CL_queuedTime: %llu\n", ocl_strttime, ocl_stoptime, ocl_submittime, ocl_queuedTime); }
