AMD clEnqueueMapBuffer和固定內存

Question

我目前使用AMD GPU。

/*device memory*/ 
pattern_obj = clCreateBuffer(context, CL_MEM_READ_ONLY, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_char), NULL, &ret); 
text_objA = clCreateBuffer(context, CL_MEM_READ_ONLY, fileLen * sizeof(char)/2, NULL, &ret); 
text_objB = clCreateBuffer(context, CL_MEM_READ_ONLY, fileLen * sizeof(char)/2, NULL, &ret); 
failure_obj = clCreateBuffer(context, CL_MEM_READ_ONLY, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_int), NULL, &ret); 
ret_objA = clCreateBuffer(context, CL_MEM_WRITE_ONLY, MAX_PATTERN_NUM * sizeof(cl_int), NULL, &ret); 
ret_objB = clCreateBuffer(context, CL_MEM_WRITE_ONLY, MAX_PATTERN_NUM * sizeof(cl_int), NULL, &ret); 

/*pinned memory*/ 
mPattern_obj = clCreateBuffer(context, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_char), NULL, &ret); 
mText_obj = clCreateBuffer(context, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, fileLen * sizeof(char), NULL, &ret); 
mFailure_obj = clCreateBuffer(context, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_int), NULL, &ret); 
mRet_obj = clCreateBuffer(context, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, MAX_PATTERN_NUM * sizeof(cl_int) * 2, NULL, &ret); 

/*mapped pointer for pinned memory*/ 
pattern = (cl_char *)clEnqueueMapBuffer(command_queue[0], mPattern_obj, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_char), 0, NULL, NULL, &ret); 
strings = (char *)clEnqueueMapBuffer(command_queue[0], mText_obj, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, fileLen * sizeof(char), 0, NULL, NULL, &ret); 
failure = (cl_int *)clEnqueueMapBuffer(command_queue[0], mFailure_obj, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_int), 0, NULL, NULL, &ret); 
matched = (cl_int *)clEnqueueMapBuffer(command_queue[0], mRet_obj, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, MAX_PATTERN_NUM * sizeof(cl_int) * 2, 0, NULL, NULL, &ret); 

/*Initialize the mapped pointer*/ 

ret = clEnqueueWriteBuffer(command_queue[0], pattern_obj, CL_FALSE, 0, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_char), (void *)&pattern[0], 0, NULL, NULL); 
ret = clEnqueueWriteBuffer(command_queue[0], text_objA, CL_FALSE, 0, halfSize, (void *)&strings[0], 0, NULL, NULL); 
ret = clEnqueueWriteBuffer(command_queue[0], failure_obj, CL_FALSE, 0, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_int), (void *)&failure[0], 0, NULL, NULL); 

ret = clEnqueueNDRangeKernel(command_queue[0], kernel[0], 1, NULL, &global_item_size, &local_item_size, 0, NULL, &kmp_event); 

clWaitForEvents(1, &kmp_event); 

ret = clEnqueueWriteBuffer(command_queue[1], pattern_obj, CL_FALSE, 0, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_char), (void *)&pattern[0], 0, NULL, NULL); 
ret = clEnqueueWriteBuffer(command_queue[1], text_objB, CL_FALSE, 0, halfSize, (void *)&strings[halfOffset], 0, NULL, NULL); 
ret = clEnqueueWriteBuffer(command_queue[1], failure_obj, CL_FALSE, 0, MAX_PATTERN_SIZE * MAX_PATTERN_NUM * sizeof(cl_int), (void *)&failure[0], 0, NULL, NULL); 

ret = clEnqueueNDRangeKernel(command_queue[1], kernel[1], 1, NULL, &global_item_size, &local_item_size, 0, NULL, &kmp_event); 

clWaitForEvents(1, &kmp_event); 

ret = clEnqueueReadBuffer(command_queue[0], ret_objA, CL_FALSE, 0, MAX_PATTERN_NUM * sizeof(cl_int), (void *)&matched[0], 0, NULL, NULL); 

ret = clEnqueueReadBuffer(command_queue[1], ret_objB, CL_FALSE, 0, MAX_PATTERN_NUM * sizeof(cl_int), (void *)&matched[MAX_PATTERN_NUM], 0, NULL, NULL); 

這是我固定內存的代碼。

我使用MapBuffer和固定內存來重疊數據傳輸和執行。

我知道我的代碼是用於Nvidia GPU，我認爲Nvidia和AMD之間存在不匹配。

AMD GPU沒有關於重疊傳輸和計算的示例代碼。

所以，我不知道我必須做什麼。

我應該更改哪些代碼才能使我的代碼在AMD GPU中工作？

我在A & B中劃分了text_obj，因爲我想通過將文本拆分爲正面和背面來傳輸文本。

如）文字： 「ababcccc」 text_objA => ABAB和text_objB => CCCC

Answer 1

是，NVIDIA和AMD都不同固定內存。 NVIDIA有兩個緩衝區（一個設備，一個主機）的奇怪要求，它奇蹟般地檢測到複製操作的主機指針是「特殊的」，並且使得複製速度更快。請注意，在較新的驅動程序中，當您映射/取消映射設備緩衝區時，您可能會獲得類似的性能。 AMD也是這麼做的（不需要兩個緩衝區）。閱讀AMD的開發人員指南;他們有一整段內存傳輸。底線：使用文檔的指導，但嘗試一切和配置您的設備，以找到真正的工作。