如果這類問題已被要求,我表示歉意,請將我鏈接到線程!來自cudaMemcpy2D的錯誤數據
無論如何,我是CUDA新手(我來自OpenCL),並且想嘗試使用它生成圖像。有關CUDA代碼:
__global__
void mandlebrot(uint8_t *pixels, size_t pitch, unsigned long width, unsigned long height) {
unsigned block_size = blockDim.x;
uint2 location = {blockIdx.x*block_size, blockIdx.y*block_size};
ulong2 pixel_location = {threadIdx.x, threadIdx.y};
ulong2 real_location = {location.x + pixel_location.x, location.y + pixel_location.y};
if (real_location.x >= width || real_location.y >= height)
return;
uint8_t *row = (uint8_t *)((char *)pixels + real_location.y * pitch);
row[real_location.x * 4+0] = 0;
row[real_location.x * 4+1] = 255;
row[real_location.x * 4+2] = 0;
row[real_location.x * 4+3] = 255;
}
cudaError_t err = cudaSuccess;
#define CUDA_ERR(e) \
if ((err = e) != cudaSuccess) { \
fprintf(stderr, "Failed to allocate device vector A (error code %s)!\n", cudaGetErrorString(err)); \
exit(-1); \
}
int main(void) {
ulong2 dims = {1000, 1000};
unsigned long block_size = 500;
dim3 threads_per_block(block_size, block_size);
dim3 remainders(dims.x % threads_per_block.x, dims.y % threads_per_block.y);
dim3 blocks(dims.x/threads_per_block.x + (remainders.x == 0 ? 0 : 1), dims.y/threads_per_block.y + (remainders.y == 0 ? 0 : 1));
size_t pitch;
uint8_t *pixels, *h_pixels = NULL;
CUDA_ERR(cudaMallocPitch(&pixels, &pitch, dims.x * 4 * sizeof(uint8_t), dims.y));
mandlebrot<<<blocks, threads_per_block>>>(pixels, pitch, dims.x, dims.y);
h_pixels = (uint8_t *)malloc(dims.x * 4 * sizeof(uint8_t) * dims.y);
memset(h_pixels, 0, dims.x * 4 * sizeof(uint8_t) * dims.y);
CUDA_ERR(cudaMemcpy2D(h_pixels, dims.x * 4 * sizeof(uint8_t), pixels, pitch, dims.x, dims.y, cudaMemcpyDeviceToHost));
save_png("out.png", h_pixels, dims.x, dims.y);
CUDA_ERR(cudaFree(pixels));
free(h_pixels);
CUDA_ERR(cudaDeviceReset());
puts("Success");
return 0;
}
的save_png
功能是我採取數據塊並將其保存爲PNG創造了一個通常的效用函數:
void save_png(const char *filename, uint8_t *buffer, unsigned long width, unsigned long height) {
png_structp png_ptr = png_create_write_struct(PNG_LIBPNG_VER_STRING, NULL, NULL, NULL);
if (!png_ptr) {
std::cerr << "Failed to create png write struct" << std::endl;
return;
}
png_infop info_ptr = png_create_info_struct(png_ptr);
if (!info_ptr) {
std::cerr << "Failed to create info_ptr" << std::endl;
png_destroy_write_struct(&png_ptr, NULL);
return;
}
FILE *fp = fopen(filename, "wb");
if (!fp) {
std::cerr << "Failed to open " << filename << " for writing" << std::endl;
png_destroy_write_struct(&png_ptr, &info_ptr);
return;
}
if (setjmp(png_jmpbuf(png_ptr))) {
png_destroy_write_struct(&png_ptr, &info_ptr);
std::cerr << "Error from libpng!" << std::endl;
return;
}
png_init_io(png_ptr, fp);
png_set_IHDR(png_ptr, info_ptr, width, height, 8, PNG_COLOR_TYPE_RGBA, PNG_INTERLACE_NONE, PNG_COMPRESSION_TYPE_DEFAULT, PNG_FILTER_TYPE_DEFAULT);
png_write_info(png_ptr, info_ptr);
png_byte *row_pnts[height];
size_t i;
for (i = 0; i < height; i++) {
row_pnts[i] = buffer + width * 4 * i;
}
png_write_image(png_ptr, row_pnts);
png_write_end(png_ptr, info_ptr);
png_destroy_write_struct(&png_ptr, &info_ptr);
fclose(fp);
}
反正系統產生的圖像是奇怪的白色條紋,可以看到隨機彩色像素斑點here。
有什麼明顯的我做錯了嗎?我試圖按照CUDA網站上的介紹文檔。否則任何人都可以幫我解決這個問題嗎?在這裏,我只是試圖用綠色像素填充pixels
緩衝區。
我正在使用NVIDIA GeForce GT 650M獨立顯卡的MBP視網膜。如果需要,我可以從cuda示例代碼運行並粘貼輸出到print_devices
。
編輯:請注意,用下面的Makefile編譯過程中沒有錯誤或警告:
all:
nvcc -c mandlebrot.cu -o mandlebrot.cu.o
nvcc mandlebrot.cu.o -o mandlebrot -lpng
,並在運行時沒有錯誤。
啊哈!謝謝 - 我完全忘記了'cudaPeekAtLastError',並感謝提示將整個代碼粘貼到一個塊中。我不知道'cuda-memcheck',謝謝指出。我會解決你指出的問題,並將其標記爲答案,如果一切都解決了! – DanZimm 2014-09-23 19:47:05
我已經在我的答案中粘貼了一個完整的代碼,修復了我找到的所有內容。 – 2014-09-23 20:10:55
非常感謝!沒有意識到我不正確地使用'cudaMemcpy2D'。感謝您採取這種效果來幫助我! – DanZimm 2014-09-23 20:48:59