// // Copyright (c) 2017 The Khronos Group Inc. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. // #include "../../test_common/harness/compat.h" #include #include #include #include #include #include #include "../../test_common/harness/conversions.h" #include "procs.h" static const char *async_global_to_local_kernel2D = R"OpenCLC( #pragma OPENCL EXTENSION cl_khr_extended_async_copies : enable #define STRUCT_SIZE %d typedef struct __attribute__((packed)) { uchar byte[STRUCT_SIZE]; } VarSizeStruct __attribute__((aligned(1))); __kernel void test_fn(const __global VarSizeStruct *src, __global VarSizeStruct *dst, __local VarSizeStruct *localBuffer, int numElementsPerLine, int lineCopiesPerWorkgroup, int lineCopiesPerWorkItem, int srcStride, int dstStride) { // Zero the local storage first for (int i = 0; i < lineCopiesPerWorkItem; i++) { for (int j = 0; j < numElementsPerLine; j++) { const int index = (get_local_id(0) * lineCopiesPerWorkItem + i) * dstStride + j; for (int k = 0; k < STRUCT_SIZE; k++) { localBuffer[index].byte[k] = 0; } } } // Do this to verify all kernels are done zeroing the local buffer before we // try the copy barrier( CLK_LOCAL_MEM_FENCE ); event_t event = async_work_group_copy_2D2D(localBuffer, 0, src, lineCopiesPerWorkgroup * get_group_id(0) * srcStride, sizeof(VarSizeStruct), (size_t)numElementsPerLine, (size_t)lineCopiesPerWorkgroup, srcStride, dstStride, 0); // Wait for the copy to complete, then verify by manually copying to the dest wait_group_events(1, &event); for (int i = 0; i < lineCopiesPerWorkItem; i++) { for (int j = 0; j < numElementsPerLine; j++) { const int local_index = (get_local_id(0) * lineCopiesPerWorkItem + i) * dstStride + j; const int global_index = (get_global_id(0) * lineCopiesPerWorkItem + i) * dstStride + j; dst[global_index] = localBuffer[local_index]; } } } )OpenCLC"; static const char *async_local_to_global_kernel2D = R"OpenCLC( #pragma OPENCL EXTENSION cl_khr_extended_async_copies : enable #define STRUCT_SIZE %d typedef struct __attribute__((packed)) { uchar byte[STRUCT_SIZE]; } VarSizeStruct __attribute__((aligned(1))); __kernel void test_fn(const __global VarSizeStruct *src, __global VarSizeStruct *dst, __local VarSizeStruct *localBuffer, int numElementsPerLine, int lineCopiesPerWorkgroup, int lineCopiesPerWorkItem, int srcStride, int dstStride) { // Zero the local storage first for (int i = 0; i < lineCopiesPerWorkItem; i++) { for (int j = 0; j < numElementsPerLine; j++) { const int index = (get_local_id(0) * lineCopiesPerWorkItem + i) * srcStride + j; for (int k = 0; k < STRUCT_SIZE; k++) { localBuffer[index].byte[k] = 0; } } } // Do this to verify all kernels are done zeroing the local buffer before we try the copy barrier(CLK_LOCAL_MEM_FENCE); for (int i = 0; i < lineCopiesPerWorkItem; i++) { for (int j = 0; j < numElementsPerLine; j++) { const int local_index = (get_local_id(0) * lineCopiesPerWorkItem + i) * srcStride + j; const int global_index = (get_global_id(0)*lineCopiesPerWorkItem + i) * srcStride + j; localBuffer[local_index] = src[global_index]; } } // Do this to verify all kernels are done copying to the local buffer before we try the copy barrier(CLK_LOCAL_MEM_FENCE); event_t event = async_work_group_copy_2D2D(dst, lineCopiesPerWorkgroup * get_group_id(0) * dstStride, localBuffer, 0, sizeof(VarSizeStruct), (size_t)numElementsPerLine, (size_t)lineCopiesPerWorkgroup, srcStride, dstStride, 0 ); wait_group_events(1, &event); }; )OpenCLC"; int test_copy2D(const cl_device_id deviceID, const cl_context context, const cl_command_queue queue, const char *const kernelCode, const size_t elementSize, const int srcMargin, const int dstMargin, const bool localIsDst) { int error; log_info("Testing %d byte element with srcMargin = %d, dstMargin = %d\n", elementSize, srcMargin, dstMargin); cl_long max_local_mem_size; error = clGetDeviceInfo(deviceID, CL_DEVICE_LOCAL_MEM_SIZE, sizeof(max_local_mem_size), &max_local_mem_size, NULL); test_error(error, "clGetDeviceInfo for CL_DEVICE_LOCAL_MEM_SIZE failed."); cl_long max_global_mem_size; error = clGetDeviceInfo(deviceID, CL_DEVICE_GLOBAL_MEM_SIZE, sizeof(max_global_mem_size), &max_global_mem_size, NULL); test_error(error, "clGetDeviceInfo for CL_DEVICE_GLOBAL_MEM_SIZE failed."); cl_long max_alloc_size; error = clGetDeviceInfo(deviceID, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(max_alloc_size), &max_alloc_size, NULL); test_error(error, "clGetDeviceInfo for CL_DEVICE_MAX_MEM_ALLOC_SIZE failed."); cl_long max_work_group_size; error = clGetDeviceInfo(deviceID, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(max_work_group_size), &max_work_group_size, NULL); test_error(error, "clGetDeviceInfo for CL_DEVICE_MAX_WORK_GROUP_SIZE failed."); if (max_alloc_size > max_global_mem_size / 2) max_alloc_size = max_global_mem_size / 2; unsigned int num_of_compute_devices; error = clGetDeviceInfo(deviceID, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(num_of_compute_devices), &num_of_compute_devices, NULL); test_error(error, "clGetDeviceInfo for CL_DEVICE_MAX_COMPUTE_UNITS failed."); char programSource[4096] = { 0 }; const char *programPtr = programSource; sprintf(programSource, kernelCode, elementSize); // log_info("program: %s\n", programSource); clProgramWrapper program; clKernelWrapper kernel; error = create_single_kernel_helper(context, &program, &kernel, 1, &programPtr, "test_fn"); test_error(error, "Unable to create testing kernel"); size_t max_workgroup_size; error = clGetKernelWorkGroupInfo( kernel, deviceID, CL_KERNEL_WORK_GROUP_SIZE, sizeof(max_workgroup_size), &max_workgroup_size, NULL); test_error( error, "clGetKernelWorkGroupInfo failed for CL_KERNEL_WORK_GROUP_SIZE."); size_t max_local_workgroup_size[3]; error = clGetDeviceInfo(deviceID, CL_DEVICE_MAX_WORK_ITEM_SIZES, sizeof(max_local_workgroup_size), max_local_workgroup_size, NULL); test_error(error, "clGetDeviceInfo failed for CL_DEVICE_MAX_WORK_ITEM_SIZES"); // Pick the minimum of the device and the kernel if (max_workgroup_size > max_local_workgroup_size[0]) max_workgroup_size = max_local_workgroup_size[0]; const size_t numElementsPerLine = 10; const cl_int dstStride = numElementsPerLine + dstMargin; const cl_int srcStride = numElementsPerLine + srcMargin; const size_t lineCopiesPerWorkItem = 13; const size_t localStorageSpacePerWorkitem = lineCopiesPerWorkItem * elementSize * (localIsDst ? dstStride : srcStride); size_t maxLocalWorkgroupSize = (((int)max_local_mem_size / 2) / localStorageSpacePerWorkitem); // Calculation can return 0 on embedded devices due to 1KB local mem limit if (maxLocalWorkgroupSize == 0) { maxLocalWorkgroupSize = 1; } size_t localWorkgroupSize = maxLocalWorkgroupSize; if (maxLocalWorkgroupSize > max_workgroup_size) localWorkgroupSize = max_workgroup_size; const size_t maxTotalLinesIn = (max_alloc_size / elementSize + srcMargin) / srcStride; const size_t maxTotalLinesOut = (max_alloc_size / elementSize + dstMargin) / dstStride; const size_t maxTotalLines = std::min(maxTotalLinesIn, maxTotalLinesOut); const size_t maxLocalWorkgroups = maxTotalLines / (localWorkgroupSize * lineCopiesPerWorkItem); const size_t localBufferSize = localWorkgroupSize * localStorageSpacePerWorkitem - (localIsDst ? dstMargin : srcMargin); const size_t numberOfLocalWorkgroups = std::min(1111, (int)maxLocalWorkgroups); const size_t totalLines = numberOfLocalWorkgroups * localWorkgroupSize * lineCopiesPerWorkItem; const size_t inBufferSize = elementSize * (totalLines * numElementsPerLine + (totalLines - 1) * srcMargin); const size_t outBufferSize = elementSize * (totalLines * numElementsPerLine + (totalLines - 1) * dstMargin); const size_t globalWorkgroupSize = numberOfLocalWorkgroups * localWorkgroupSize; if ((localBufferSize / 4) > max_work_group_size) { log_info("Skipping due to resource requirements local:%db " "max_work_group_size:%d\n", localBufferSize, max_work_group_size); return 0; } void *const inBuffer = (void *)malloc(inBufferSize); void *const outBuffer = (void *)malloc(outBufferSize); void *const outBufferCopy = (void *)malloc(outBufferSize); const cl_int lineCopiesPerWorkItemInt = static_cast(lineCopiesPerWorkItem); const cl_int numElementsPerLineInt = static_cast(numElementsPerLine); const cl_int lineCopiesPerWorkgroup = static_cast(lineCopiesPerWorkItem * localWorkgroupSize); log_info( "Global: %d, local %d, local buffer %db, global in buffer %db, " "global out buffer %db, each work group will copy %d lines and each " "work item item will copy %d lines.\n", (int)globalWorkgroupSize, (int)localWorkgroupSize, (int)localBufferSize, (int)inBufferSize, (int)outBufferSize, lineCopiesPerWorkgroup, lineCopiesPerWorkItemInt); size_t threads[1], localThreads[1]; threads[0] = globalWorkgroupSize; localThreads[0] = localWorkgroupSize; MTdata d = init_genrand(gRandomSeed); generate_random_data(kChar, inBufferSize, d, inBuffer); generate_random_data(kChar, outBufferSize, d, outBuffer); free_mtdata(d); d = NULL; memcpy(outBufferCopy, outBuffer, outBufferSize); clMemWrapper streams[2]; streams[0] = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR, inBufferSize, inBuffer, &error); test_error(error, "Unable to create input buffer"); streams[1] = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR, outBufferSize, outBuffer, &error); test_error(error, "Unable to create output buffer"); error = clSetKernelArg(kernel, 0, sizeof(streams[0]), &streams[0]); test_error(error, "Unable to set kernel argument"); error = clSetKernelArg(kernel, 1, sizeof(streams[1]), &streams[1]); test_error(error, "Unable to set kernel argument"); error = clSetKernelArg(kernel, 2, localBufferSize, NULL); test_error(error, "Unable to set kernel argument"); error = clSetKernelArg(kernel, 3, sizeof(numElementsPerLineInt), &numElementsPerLineInt); test_error(error, "Unable to set kernel argument"); error = clSetKernelArg(kernel, 4, sizeof(lineCopiesPerWorkgroup), &lineCopiesPerWorkgroup); test_error(error, "Unable to set kernel argument"); error = clSetKernelArg(kernel, 5, sizeof(lineCopiesPerWorkItemInt), &lineCopiesPerWorkItemInt); test_error(error, "Unable to set kernel argument"); error = clSetKernelArg(kernel, 6, sizeof(srcStride), &srcStride); test_error(error, "Unable to set kernel argument"); error = clSetKernelArg(kernel, 7, sizeof(dstStride), &dstStride); test_error(error, "Unable to set kernel argument"); // Enqueue error = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, threads, localThreads, 0, NULL, NULL); test_error(error, "Unable to queue kernel"); // Read error = clEnqueueReadBuffer(queue, streams[1], CL_TRUE, 0, outBufferSize, outBuffer, 0, NULL, NULL); test_error(error, "Unable to read results"); // Verify int failuresPrinted = 0; for (int i = 0; i < (int)globalWorkgroupSize * lineCopiesPerWorkItem * elementSize; i += elementSize) { for (int j = 0; j < (int)numElementsPerLine * elementSize; j += elementSize) { int inIdx = i * srcStride + j; int outIdx = i * dstStride + j; if (memcmp(((char *)inBuffer) + inIdx, ((char *)outBuffer) + outIdx, elementSize) != 0) { unsigned char *inchar = (unsigned char *)inBuffer + inIdx; unsigned char *outchar = (unsigned char *)outBuffer + outIdx; char values[4096] = { 0 }; if (failuresPrinted == 0) { // Print first failure message log_error("ERROR: Results of copy did not validate!\n"); } sprintf(values + strlen(values), "%d -> [", inIdx); for (int k = 0; k < (int)elementSize; k++) sprintf(values + strlen(values), "%2x ", inchar[k]); sprintf(values + strlen(values), "] != ["); for (int k = 0; k < (int)elementSize; k++) sprintf(values + strlen(values), "%2x ", outchar[k]); sprintf(values + strlen(values), "]"); log_error("%s\n", values); failuresPrinted++; } if (failuresPrinted > 5) { log_error("Not printing further failures...\n"); return -1; } } if (i < (int)(globalWorkgroupSize * lineCopiesPerWorkItem - 1) * elementSize) { int outIdx = i * dstStride + numElementsPerLine * elementSize; if (memcmp(((char *)outBuffer) + outIdx, ((char *)outBufferCopy) + outIdx, dstMargin * elementSize) != 0) { if (failuresPrinted == 0) { // Print first failure message log_error("ERROR: Results of copy did not validate!\n"); } log_error( "2D copy corrupted data in output buffer in the stride " "offset of line %d\n", i); failuresPrinted++; } if (failuresPrinted > 5) { log_error("Not printing further failures...\n"); return -1; } } } free(inBuffer); free(outBuffer); free(outBufferCopy); return failuresPrinted ? -1 : 0; } int test_copy2D_all_types(cl_device_id deviceID, cl_context context, cl_command_queue queue, const char *kernelCode, bool localIsDst) { const unsigned int elemSizes[] = { 1, 2, 3, 4, 5, 6, 7, 8, 13, 16, 32, 47, 64 }; // The margins below represent the number of elements between the end of // one line and the start of the next. The strides are equivalent to the // length of the line plus the chosen margin. // These have to be multipliers, because the margin must be a multiple of // element size. const unsigned int marginMultipliers[] = { 0, 10, 100 }; int errors = 0; if (!is_extension_available(deviceID, "cl_khr_extended_async_copies")) { log_info( "Device does not support extended async copies. Skipping test.\n"); } else { for (const unsigned int elemSize : elemSizes) { for (const unsigned int srcMarginMultiplier : marginMultipliers) { for (const unsigned int dstMarginMultiplier : marginMultipliers) { if (test_copy2D(deviceID, context, queue, kernelCode, elemSize, srcMarginMultiplier * elemSize, dstMarginMultiplier * elemSize, localIsDst)) { errors++; } } } } } return errors ? -1 : 0; } int test_async_copy_global_to_local2D(cl_device_id deviceID, cl_context context, cl_command_queue queue, int num_elements) { return test_copy2D_all_types(deviceID, context, queue, async_global_to_local_kernel2D, true); } int test_async_copy_local_to_global2D(cl_device_id deviceID, cl_context context, cl_command_queue queue, int num_elements) { return test_copy2D_all_types(deviceID, context, queue, async_local_to_global_kernel2D, false); }