我对cuda的尝试很少，并试图编写一个盒式过滤器。我读到盒式滤镜是一种滤镜，其中结果图像中的每个像素的值等于输入图像中其相邻像素的平均值。 我有这个文件 http://www.nvidia.com/content/nvision2008/tech_presentations/Game_Developer_Track/NVISION08-Image_Processing_and_Video_with_CUDA.pdf 并改变了一点代码。这是我的功能。
#define TILE_W 16
#define TILE_H 16
#define R 2 // filter radius
#define D (R*2+1) // filter diameter
#define S (D*D) // filter size
#define BLOCK_W (TILE_W+(2*R))
#define BLOCK_H (TILE_H+(2*R))
__global__ void d_filter(unsigned char *g_idata, unsigned char *g_odata, unsigned int width, unsigned int height)
{
__shared__ unsigned char smem[BLOCK_W*BLOCK_H];
int x = blockIdx.x*TILE_W + threadIdx.x - R;
int y = blockIdx.y*TILE_H + threadIdx.y - R;
// clamp to edge of image
x = max(0, x);
x = min(x, width-1);
y = max(y, 0);
y = min(y, height-1);
unsigned int index = y*width + x;
unsigned int bindex = threadIdx.y*blockDim.y+threadIdx.x;
// each thread copies its pixel of the block to shared memory
smem[bindex] = g_idata[index];
__syncthreads();
// only threads inside the apron will write results
if ((threadIdx.x >= R) && (threadIdx.x < (BLOCK_W-R)) && (threadIdx.y >= R) && (threadIdx.y < (BLOCK_H-R))) {
float sum = 0;
for(int dy=-R; dy<=R; dy++) {
for(int dx=-R; dx<=R; dx++) {
float i = smem[bindex + (dy*blockDim.x) + dx];
sum += i;
}
}
g_odata[index] = sum / S;
}
}
编辑：这是一个有效的新版本。问题出现在内核发布中。
#include <fstream>
#include <stdio.h>
#include <stdlib.h>
#include <cuda_runtime.h>
#include <assert.h>
#define PGMHeaderSize 0x40
inline bool loadPPM(const char *file, unsigned char **data, unsigned int *w, unsigned int *h, unsigned int *channels)
{
FILE *fp = NULL;
fp = fopen(file, "rb");
if (!fp) {
fprintf(stderr, "__LoadPPM() : unable to open file\n" );
return false;
}
// check header
char header[PGMHeaderSize];
if (fgets(header, PGMHeaderSize, fp) == NULL)
{
fprintf(stderr,"__LoadPPM() : reading PGM header returned NULL\n" );
return false;
}
if (strncmp(header, "P5", 2) == 0)
{
*channels = 1;
}
else if (strncmp(header, "P6", 2) == 0)
{
*channels = 3;
}
else
{
fprintf(stderr,"__LoadPPM() : File is not a PPM or PGM image\n" );
*channels = 0;
return false;
}
// parse header, read maxval, width and height
unsigned int width = 0;
unsigned int height = 0;
unsigned int maxval = 0;
unsigned int i = 0;
while (i < 3)
{
if (fgets(header, PGMHeaderSize, fp) == NULL)
{
fprintf(stderr,"__LoadPPM() : reading PGM header returned NULL\n" );
return false;
}
if (header[0] == '#')
{
continue;
}
if (i == 0)
{
i += sscanf(header, "%u %u %u", &width, &height, &maxval);
}
else if (i == 1)
{
i += sscanf(header, "%u %u", &height, &maxval);
}
else if (i == 2)
{
i += sscanf(header, "%u", &maxval);
}
}
// check if given handle for the data is initialized
if (NULL != *data)
{
if (*w != width || *h != height)
{
fprintf(stderr, "__LoadPPM() : Invalid image dimensions.\n" );
}
}
else
{
*data = (unsigned char *) malloc(sizeof(unsigned char) * width * height * *channels);
if (!data) {
fprintf(stderr, "Unable to allocate hostmemory\n");
return false;
}
*w = width;
*h = height;
}
// read and close file
if (fread(*data, sizeof(unsigned char), width * height * *channels, fp) == 0)
{
fprintf(stderr, "__LoadPPM() : read data returned error.\n" );
fclose(fp);
return false;
}
fclose(fp);
return true;
}
inline bool savePPM(const char *file, unsigned char *data, unsigned int w, unsigned int h, unsigned int channels)
{
assert(NULL != data);
assert(w > 0);
assert(h > 0);
std::fstream fh(file, std::fstream::out | std::fstream::binary);
if (fh.bad())
{
fprintf(stderr, "__savePPM() : Opening file failed.\n" );
return false;
}
if (channels == 1)
{
fh << "P5\n";
}
else if (channels == 3)
{
fh << "P6\n";
}
else
{
fprintf(stderr, "__savePPM() : Invalid number of channels.\n" );
return false;
}
fh << w << "\n" << h << "\n" << 0xff << std::endl;
for (unsigned int i = 0; (i < (w*h*channels)) && fh.good(); ++i)
{
fh << data[i];
}
fh.flush();
if (fh.bad())
{
fprintf(stderr,"__savePPM() : Writing data failed.\n" );
return false;
}
fh.close();
return true;
}
#define TILE_W 16
#define TILE_H 16
#define Rx 2 // filter radius in x direction
#define Ry 2 // filter radius in y direction
#define FILTER_W (Rx*2+1) // filter diameter in x direction
#define FILTER_H (Ry*2+1) // filter diameter in y direction
#define S (FILTER_W*FILTER_H) // filter size
#define BLOCK_W (TILE_W+(2*Rx))
#define BLOCK_H (TILE_H+(2*Ry))
__global__ void box_filter(const unsigned char *in, unsigned char *out, const unsigned int w, const unsigned int h){
//Indexes
const int x = blockIdx.x * TILE_W + threadIdx.x - Rx; // x image index
const int y = blockIdx.y * TILE_H + threadIdx.y - Ry; // y image index
const int d = y * w + x; // data index
//shared mem
__shared__ float shMem[BLOCK_W][BLOCK_H];
if(x<0 || y<0 || x>=w || y>=h) { // Threads which are not in the picture just write 0 to the shared mem
shMem[threadIdx.x][threadIdx.y] = 0;
return;
}
shMem[threadIdx.x][threadIdx.y] = in[d];
__syncthreads();
// box filter (only for threads inside the tile)
if ((threadIdx.x >= Rx) && (threadIdx.x < (BLOCK_W-Rx)) && (threadIdx.y >= Ry) && (threadIdx.y < (BLOCK_H-Ry))) {
float sum = 0;
for(int dx=-Rx; dx<=Rx; dx++) {
for(int dy=-Ry; dy<=Ry; dy++) {
sum += shMem[threadIdx.x+dx][threadIdx.y+dy];
}
}
out[d] = sum / S;
}
}
#define checkCudaErrors(err) __checkCudaErrors (err, __FILE__, __LINE__)
inline void __checkCudaErrors(cudaError err, const char *file, const int line)
{
if (cudaSuccess != err)
{
fprintf(stderr, "%s(%i) : CUDA Runtime API error %d: %s.\n",
file, line, (int)err, cudaGetErrorString(err));
exit(EXIT_FAILURE);
}
}
int main(){
unsigned char *data=NULL, *d_idata=NULL, *d_odata=NULL;
unsigned int w,h,channels;
if(! loadPPM("../../data/lena_bw.pgm", &data, &w, &h, &channels)){
fprintf(stderr, "Failed to open File\n");
exit(EXIT_FAILURE);
}
printf("Loaded file with w:%d h:%d channels:%d \n",w,h,channels);
unsigned int numElements = w*h*channels;
size_t datasize = numElements * sizeof(unsigned char);
// Allocate the Device Memory
printf("Allocate Devicememory for data\n");
checkCudaErrors(cudaMalloc((void **)&d_idata, datasize));
checkCudaErrors(cudaMalloc((void **)&d_odata, datasize));
// Copy to device
printf("Copy idata from the host memory to the CUDA device\n");
checkCudaErrors(cudaMemcpy(d_idata, data, datasize, cudaMemcpyHostToDevice));
// Launch Kernel
int GRID_W = w/TILE_W +1;
int GRID_H = h/TILE_H +1;
dim3 threadsPerBlock(BLOCK_W, BLOCK_H);
dim3 blocksPerGrid(GRID_W,GRID_H);
printf("CUDA kernel launch with [%d %d] blocks of [%d %d] threads\n", blocksPerGrid.x, blocksPerGrid.y, threadsPerBlock.x, threadsPerBlock.y);
box_filter<<<blocksPerGrid, threadsPerBlock>>>(d_idata, d_odata, w,h);
checkCudaErrors(cudaGetLastError());
// Copy data from device to host
printf("Copy odata from the CUDA device to the host memory\n");
checkCudaErrors(cudaMemcpy(data, d_odata, datasize, cudaMemcpyDeviceToHost));
// Free Device memory
printf("Free Device memory\n");
checkCudaErrors(cudaFree(d_idata));
checkCudaErrors(cudaFree(d_odata));
// Save Picture
printf("Save Picture\n");
bool saved = false;
if (channels==1)
saved = savePPM("output.pgm", data, w, h, channels);
else if (channels==3)
saved = savePPM("output.ppm", data, w, h, channels);
else fprintf(stderr, "ERROR: Unable to save file - wrong channel!\n");
// Free Host memory
printf("Free Host memory\n");
free(data);
if (!saved){
fprintf(stderr, "Failed to save File\n");
exit(EXIT_FAILURE);
}
printf("Done\n");
}
过滤功能有问题。 loadPPM和savePPM（cuda样本的一部分）正在使用其他内核函数，但是使用这个过滤器函数我会得到一个黑色图像。
所以问题是：我错了什么？
其他一些理解问题： 这里https://www.nvidia.com/docs/IO/116711/sc11-cuda-c-basics.pdf我读到线程只能在块内共享（共享内存，syncthreads，..）。所以在我的函数中，图像被分割成矩形块，图像处理幻灯片的第9页上的图片大约是一个块？块边缘的像素怎么样？它们没变吗？
感谢您的回答。
答案 0 :(得分：2)
您的代码中的一个问题是您的内核需要2D网格和2D线程块：
int x = blockIdx.x*TILE_W + threadIdx.x - R;
int y = blockIdx.y*TILE_H + threadIdx.y - R;
^^^^^^^^^^ ^^^^^^^^^^^
2D grid 2D threadblock
但您正在启动具有1D网格和线程块定义的内核：
int threadsPerBlock = 256; // creates 1D threadblock
int blocksPerGrid =(numElements + threadsPerBlock - 1) / threadsPerBlock; //1D grid
....
d_filter<<<blocksPerGrid, threadsPerBlock>>>(d_idata, d_odata, w,h);
因此，当您启动该内核时，threadIdx.y将始终为零blockIdx.y
当我制作一个不依赖于PPM图像加载/存储的代码的修改版本（因此，使用合成数据），并进行必要的更改以启动2D网格和线程块，以与内核保持一致，代码似乎正确地为我运行并产生似乎可能是过滤输出的输出，而不是零：
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#define TILE_W 16
#define TILE_H 16
#define R 2 // filter radius
#define D (R*2+1) // filter diameter
#define S (D*D) // filter size
#define BLOCK_W (TILE_W+(2*R))
#define BLOCK_H (TILE_H+(2*R))
__global__ void d_filter(unsigned char *g_idata, unsigned char *g_odata, unsigned int width, unsigned int height)
{
__shared__ unsigned char smem[BLOCK_W*BLOCK_H];
int x = blockIdx.x*TILE_W + threadIdx.x - R;
int y = blockIdx.y*TILE_H + threadIdx.y - R;
// clamp to edge of image
x = max(0, x);
x = min(x, width-1);
y = max(y, 0);
y = min(y, height-1);
unsigned int index = y*width + x;
unsigned int bindex = threadIdx.y*blockDim.y+threadIdx.x;
// each thread copies its pixel of the block to shared memory
smem[bindex] = g_idata[index];
__syncthreads();
// only threads inside the apron will write results
if ((threadIdx.x >= R) && (threadIdx.x < (BLOCK_W-R)) && (threadIdx.y >= R) && (threadIdx.y < (BLOCK_H-R))) {
float sum = 0;
for(int dy=-R; dy<=R; dy++) {
for(int dx=-R; dx<=R; dx++) {
float i = smem[bindex + (dy*blockDim.x) + dx];
sum += i;
}
}
g_odata[index] = sum / S;
}
}
const unsigned int imgw = 512;
const unsigned int imgh = 256;
void loadImg(unsigned char **data, unsigned int *w, unsigned int *h, unsigned int *ch){
*w = imgw;
*h = imgh;
*ch = 1;
*data = (unsigned char *)malloc(imgw*imgh*sizeof(unsigned char));
for (int i = 0; i < imgw*imgh; i++) (*data)[i] = i%8;
}
int main(){
unsigned char *data=NULL, *d_idata=NULL, *d_odata;
unsigned int w,h,channels;
loadImg(&data, &w, &h, &channels);
printf("Loaded file with w:%d h:%d channels:%d \n",w,h,channels);
unsigned int numElements = w*h*channels;
size_t datasize = numElements * sizeof(unsigned char);
cudaError_t err = cudaSuccess;
// Allocate the Device Memory
printf("Allocate Devicememory for data\n");
err = cudaMalloc((void **)&d_idata, datasize);
if ( err != cudaSuccess)
{
fprintf(stderr, "Failed to allocate device memory for idata (error code %s)!\n", cudaGetErrorString(err));
exit(EXIT_FAILURE);
}
err = cudaMalloc((void **)&d_odata, datasize);
if ( err != cudaSuccess)
{
fprintf(stderr, "Failed to allocate device memory for odata (error code %s)!\n", cudaGetErrorString(err));
exit(EXIT_FAILURE);
}
// Copy to device
printf("Copy idata from the host memory to the CUDA device\n");
err =cudaMemcpy(d_idata, data, datasize, cudaMemcpyHostToDevice);
if (err != cudaSuccess)
{
fprintf(stderr, "Failed to copy idata from host to device (error code %s)!\n", cudaGetErrorString(err));
exit(EXIT_FAILURE);
}
// Launch Kernel
dim3 threadsPerBlock(BLOCK_W, BLOCK_H);
dim3 blocksPerGrid((w+threadsPerBlock.x-1)/threadsPerBlock.x, (h+threadsPerBlock.y-1)/threadsPerBlock.y);
printf("CUDA kernel launch with %d,%d blocks of %d,%d threads\n", blocksPerGrid.x, blocksPerGrid.y, threadsPerBlock.x, threadsPerBlock.y);
d_filter<<<blocksPerGrid, threadsPerBlock>>>(d_idata, d_odata, w,h);
err=cudaGetLastError();
if (err != cudaSuccess)
{
fprintf(stderr, "Failed to launch kernel (error code %s)!\n", cudaGetErrorString(err));
exit(EXIT_FAILURE);
}
// Copy data from device to host
printf("Copy odata from the CUDA device to the host memory\n");
err=cudaMemcpy(data, d_odata, datasize, cudaMemcpyDeviceToHost);
if (err != cudaSuccess)
{
fprintf(stderr, "Failed to copy odata from device to host (error code %s)!\n", cudaGetErrorString(err));
exit(EXIT_FAILURE);
}
// Free Device memory
printf("Free Device memory\n");
err=cudaFree(d_idata);
if (err != cudaSuccess)
{
fprintf(stderr, "Failed to free device idata (error code %s)!\n", cudaGetErrorString(err));
exit(EXIT_FAILURE);
}
err=cudaFree(d_odata);
if (err != cudaSuccess)
{
fprintf(stderr, "Failed to free device odata (error code %s)!\n", cudaGetErrorString(err));
exit(EXIT_FAILURE);
}
printf("results:\n");
for (int i = 0; i < 16; i++){
for (int j = 0; j < 16; j++) printf("%d ", data[i*w+j]);
printf("\n");}
// Free Host memory
printf("Free Host memory\n");
free(data);
printf("\nDone\n");
}
当我使用cuda-memcheck运行上述代码时，我得到了这个：
C:\ProgramData\NVIDIA Corporation\CUDA Samples\v5.0\bin\win32\Debug>cuda-memcheck test
========= CUDA-MEMCHECK
Loaded file with w:512 h:256 channels:1
Allocate Devicememory for data
Copy idata from the host memory to the CUDA device
CUDA kernel launch with 26,13 blocks of 20,20 threads
Copy odata from the CUDA device to the host memory
Free Device memory
results:
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
0 1 2 3 4 5 4 3 3 2 2 3 4 5 4 3
Free Host memory
Done
========= ERROR SUMMARY: 0 errors
C:\ProgramData\NVIDIA Corporation\CUDA Samples\v5.0\bin\win32\Debug>