/* * Copyright (c) Advanced Micro Devices, Inc., or its affiliates. * * SPDX-License-Identifier: MIT */ #include "OCLPerfMatrixTranspose.h" #include #include #include #include "CL/cl.h" #include "Timer.h" static const unsigned int NUM_BLOCK_SIZES = 2; static const unsigned int blockSizes[NUM_BLOCK_SIZES] = {8, 16}; static const unsigned int NUM_MATRIX_DIMS = 2; static const unsigned int matrixDims[NUM_MATRIX_DIMS] = {1024, 1920}; static const char* matrixtranspose_kernel = "kernel void matrixTranspose(global uint *restrict inBuf, global uint " "*restrict outBuf, local uint *localBuf, uint blockSize, uint width, uint " "height)\n" "{\n" " uint globalIdx = get_global_id(0);\n" " uint globalIdy = get_global_id(1);\n" " uint localIdx = get_local_id(0);\n" " uint localIdy = get_local_id(1);\n" " /* copy from input to local memory */\n" " /* Note that we transpose the x and y coordinates when storing */\n" " localBuf[localIdx*blockSize + localIdy] = inBuf[globalIdy*width + " "globalIdx];\n" " /* wait until the whole block is filled */\n" " barrier(CLK_LOCAL_MEM_FENCE);\n" " uint groupIdx = get_group_id(0);\n" " uint groupIdy = get_group_id(1);\n" " /* calculate the corresponding target location for transpose by " "inverting x and y values*/\n" " /* Here we don't swap localIdx and localIdy, this is to get larger " "bursts when threads write to memory. */\n" " /* To make this work, we've swapped the coordinates when we write to " "local memory. */\n" " uint targetGlobalIdx = groupIdy*blockSize + localIdx;\n" " uint targetGlobalIdy = groupIdx*blockSize + localIdy;\n" " /* calculate the corresponding raster indices of source and target " "*/\n" " uint targetIndex = targetGlobalIdy*height + targetGlobalIdx;\n" " uint sourceIndex = localIdy * blockSize + localIdx;\n" " outBuf[targetIndex] = localBuf[sourceIndex];\n" "}\n"; OCLPerfMatrixTranspose::OCLPerfMatrixTranspose() { _numSubTests = NUM_BLOCK_SIZES * NUM_MATRIX_DIMS; } OCLPerfMatrixTranspose::~OCLPerfMatrixTranspose() {} void OCLPerfMatrixTranspose::setData(cl_mem buffer) { unsigned int* data = (unsigned int*)_wrapper->clEnqueueMapBuffer( cmd_queue_, buffer, true, CL_MAP_WRITE, 0, bufSize_, 0, NULL, NULL, &error_); for (unsigned int i = 0; i < height_; i++) { for (unsigned int j = 0; j < width_; j++) { *(data + i * width_ + j) = i * width_ + j; } } error_ = _wrapper->clEnqueueUnmapMemObject(cmd_queue_, buffer, data, 0, NULL, NULL); } void OCLPerfMatrixTranspose::fillData(cl_mem buffer, unsigned int val) { unsigned int* data = (unsigned int*)_wrapper->clEnqueueMapBuffer( cmd_queue_, buffer, true, CL_MAP_WRITE, 0, bufSize_, 0, NULL, NULL, &error_); for (unsigned int i = 0; i < width_ * height_; i++) { data[i] = val; } error_ = _wrapper->clEnqueueUnmapMemObject(cmd_queue_, buffer, data, 0, NULL, NULL); } void OCLPerfMatrixTranspose::checkData(cl_mem buffer) { unsigned int* data = (unsigned int*)_wrapper->clEnqueueMapBuffer( cmd_queue_, buffer, true, CL_MAP_READ, 0, bufSize_, 0, NULL, NULL, &error_); bool err = false; for (unsigned int i = 0; (i < width_) && !err; i++) { for (unsigned int j = 0; (j < height_) && !err; j++) { if (*(data + i * height_ + j) != (j * width_ + i)) { printf("Data mismatch at (%d, %d)! Got %d, expected %d\n", j, i, *(data + i * height_ + j), j * width_ + i); err = true; break; } } break; } error_ = _wrapper->clEnqueueUnmapMemObject(cmd_queue_, buffer, data, 0, NULL, NULL); } static void CL_CALLBACK notify_callback(const char* errinfo, const void* private_info, size_t cb, void* user_data) {} void OCLPerfMatrixTranspose::open(unsigned int test, char* units, double& conversion, unsigned int deviceId) { cl_uint numPlatforms; cl_platform_id platform = NULL; cl_uint num_devices = 0; cl_device_id* devices = NULL; cl_device_id device = NULL; _crcword = 0; conversion = 1.0f; _deviceId = deviceId; _openTest = test; context_ = 0; cmd_queue_ = 0; program_ = 0; kernel_ = 0; inBuffer_ = 0; outBuffer_ = 0; blockSize_ = blockSizes[_openTest % NUM_BLOCK_SIZES]; width_ = matrixDims[_openTest / NUM_BLOCK_SIZES]; height_ = width_; // We compute a square domain bufSize_ = width_ * height_ * sizeof(cl_uint); error_ = _wrapper->clGetPlatformIDs(0, NULL, &numPlatforms); CHECK_RESULT(error_ != CL_SUCCESS, "clGetPlatformIDs failed"); if (0 < numPlatforms) { cl_platform_id* platforms = new cl_platform_id[numPlatforms]; error_ = _wrapper->clGetPlatformIDs(numPlatforms, platforms, NULL); CHECK_RESULT(error_ != CL_SUCCESS, "clGetPlatformIDs failed"); #if 0 // Get last for default platform = platforms[numPlatforms-1]; for (unsigned i = 0; i < numPlatforms; ++i) { #endif platform = platforms[_platformIndex]; char pbuf[100]; error_ = _wrapper->clGetPlatformInfo(platforms[_platformIndex], CL_PLATFORM_VENDOR, sizeof(pbuf), pbuf, NULL); num_devices = 0; /* Get the number of requested devices */ error_ = _wrapper->clGetDeviceIDs(platforms[_platformIndex], type_, 0, NULL, &num_devices); // Runtime returns an error when no GPU devices are present instead of just // returning 0 devices // CHECK_RESULT(error_ != CL_SUCCESS, "clGetDeviceIDs failed"); // Choose platform with GPU devices if (num_devices > 0) { // platform = platforms[_platformIndex]; // break; } #if 0 } #endif delete platforms; } /* * If we could find our platform, use it. If not, die as we need the AMD * platform for these extensions. */ CHECK_RESULT(platform == 0, "Couldn't find platform with GPU devices, cannot proceed"); devices = (cl_device_id*)malloc(num_devices * sizeof(cl_device_id)); CHECK_RESULT(devices == 0, "no devices"); /* Get the requested device */ error_ = _wrapper->clGetDeviceIDs(platform, type_, num_devices, devices, NULL); CHECK_RESULT(error_ != CL_SUCCESS, "clGetDeviceIDs failed"); CHECK_RESULT(_deviceId >= num_devices, "Requested deviceID not available"); device = devices[_deviceId]; context_ = _wrapper->clCreateContext(NULL, 1, &device, notify_callback, NULL, &error_); CHECK_RESULT(context_ == 0, "clCreateContext failed"); char charbuf[1024]; size_t retsize; error_ = _wrapper->clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 1024, charbuf, &retsize); CHECK_RESULT(error_ != CL_SUCCESS, "clGetDeviceInfo failed"); cmd_queue_ = _wrapper->clCreateCommandQueue(context_, device, 0, NULL); CHECK_RESULT(cmd_queue_ == 0, "clCreateCommandQueue failed"); inBuffer_ = _wrapper->clCreateBuffer(context_, CL_MEM_READ_ONLY, bufSize_, NULL, &error_); CHECK_RESULT(inBuffer_ == 0, "clCreateBuffer(inBuffer) failed"); setData(inBuffer_); outBuffer_ = _wrapper->clCreateBuffer(context_, CL_MEM_WRITE_ONLY, bufSize_, NULL, &error_); CHECK_RESULT(outBuffer_ == 0, "clCreateBuffer(outBuffer) failed"); fillData(outBuffer_, 0xdeadbeef); program_ = _wrapper->clCreateProgramWithSource(context_, 1, (const char**)&matrixtranspose_kernel, NULL, &error_); CHECK_RESULT(program_ == 0, "clCreateProgramWithSource failed"); char* buildOps = NULL; error_ = _wrapper->clBuildProgram(program_, 1, &device, buildOps, NULL, NULL); if (error_ != CL_SUCCESS) { cl_int intError; char log[16384]; intError = _wrapper->clGetProgramBuildInfo(program_, device, CL_PROGRAM_BUILD_LOG, 16384 * sizeof(char), log, NULL); printf("Build error -> %s\n", log); CHECK_RESULT(0, "clBuildProgram failed"); } kernel_ = _wrapper->clCreateKernel(program_, "matrixTranspose", &error_); CHECK_RESULT(kernel_ == 0, "clCreateKernel failed"); error_ = _wrapper->clSetKernelArg(kernel_, 0, sizeof(cl_mem), (void*)&inBuffer_); error_ = _wrapper->clSetKernelArg(kernel_, 1, sizeof(cl_mem), (void*)&outBuffer_); error_ = _wrapper->clSetKernelArg(kernel_, 2, sizeof(cl_uint) * blockSize_ * blockSize_, NULL); error_ = _wrapper->clSetKernelArg(kernel_, 3, sizeof(cl_uint), (void*)&blockSize_); error_ = _wrapper->clSetKernelArg(kernel_, 4, sizeof(cl_uint), (void*)&width_); error_ = _wrapper->clSetKernelArg(kernel_, 5, sizeof(cl_uint), (void*)&height_); } void OCLPerfMatrixTranspose::run(void) { size_t global_work_size[2] = {width_, height_}; size_t local_work_size[2] = {blockSize_, blockSize_}; CPerfCounter timer; timer.Reset(); timer.Start(); for (unsigned int i = 0; i < MAX_ITERATIONS; i++) { error_ = _wrapper->clEnqueueNDRangeKernel(cmd_queue_, kernel_, 2, NULL, (const size_t*)global_work_size, (const size_t*)local_work_size, 0, NULL, NULL); } CHECK_RESULT(error_, "clEnqueueNDRangeKernel failed"); _wrapper->clFinish(cmd_queue_); timer.Stop(); double sec = timer.GetElapsedTime(); checkData(outBuffer_); // Compute GB/s double perf = ((double)bufSize_ * (double)MAX_ITERATIONS * (double)(1e-09)) / sec; _perfInfo = (float)perf; testDescString = ""; char str[90]; snprintf(str, sizeof(str), "(%d,%d) matrix with (%2d,%2d) block size %fms (GB/s) ", width_, height_, blockSize_, blockSize_, (sec / (double)MAX_ITERATIONS) * 1000.); testDescString += str; } unsigned int OCLPerfMatrixTranspose::close(void) { _wrapper->clFinish(cmd_queue_); if (inBuffer_) { error_ = _wrapper->clReleaseMemObject(inBuffer_); CHECK_RESULT_NO_RETURN(error_ != CL_SUCCESS, "clReleaseMemObject(inBuffer_) failed"); } if (outBuffer_) { error_ = _wrapper->clReleaseMemObject(outBuffer_); CHECK_RESULT_NO_RETURN(error_ != CL_SUCCESS, "clReleaseMemObject(outBuffer_) failed"); } if (kernel_) { error_ = _wrapper->clReleaseKernel(kernel_); CHECK_RESULT_NO_RETURN(error_ != CL_SUCCESS, "clReleaseKernel failed"); } if (program_) { error_ = _wrapper->clReleaseProgram(program_); CHECK_RESULT_NO_RETURN(error_ != CL_SUCCESS, "clReleaseProgram failed"); } if (cmd_queue_) { error_ = _wrapper->clReleaseCommandQueue(cmd_queue_); CHECK_RESULT_NO_RETURN(error_ != CL_SUCCESS, "clReleaseCommandQueue failed"); } if (context_) { error_ = _wrapper->clReleaseContext(context_); CHECK_RESULT_NO_RETURN(error_ != CL_SUCCESS, "clReleaseContext failed"); } return _crcword; }