我有一个CUDA内核，它采用边缘图像并处理它以创建边缘像素的较小的1D阵列。现在这是奇怪的行为。每次我运行内核并计算“d_nlist”中的边缘像素数量时（参见printf附近的代码），每次都得到更大的像素数，即使我使用相同的图像并完全停止程序并重新跑。因此，每次运行它都需要更长的时间才能运行，直到最终，它会抛出一个未捕获的异常。
我的问题是，如何阻止这种情况发生，以便每次运行内核时都能获得一致的结果？
我的设备是Geforce 620。
常量：
THREADS_X = 32
THREADS_Y = 4
PIXELS_PER_THREAD = 4
MAX_QUEUE_LENGTH = THREADS_X * THREADS_Y * PIXELS_PER_THREAD
IMG_WIDTH = 256
IMG_HEIGHT = 256
IMG_SIZE = IMG_WIDTH * IMG_HEIGHT
BLOCKS_X = IMG_WIDTH /（THREADS_X * PIXELS_PER_THREAD）
BLOCKS_Y = IMG_HEIGHT / THREADS_Y
内核如下：
__global__ void convert2DEdgeImageTo1DArray( unsigned char const * const image,
unsigned int* const list, int* const glob_index ) {
unsigned int const x = blockIdx.x * THREADS_X*PIXELS_PER_THREAD + threadIdx.x;
unsigned int const y = blockIdx.y * THREADS_Y + threadIdx.y;
volatile int qindex = -1;
volatile __shared__ int sh_qindex[THREADS_Y];
volatile __shared__ int sh_qstart[THREADS_Y];
sh_qindex[threadIdx.y] = -1;
// Start by making an array
volatile __shared__ unsigned int sh_queue[MAX_QUEUE_LENGTH];
// Fill the queue
for(int i=0; i<PIXELS_PER_THREAD; i++)
{
int const xx = i*THREADS_X + x;
// Read one image pixel from global memory
unsigned char const pixel = image[y*IMG_WIDTH + xx];
unsigned int const queue_val = (y << 16) + xx;
if(pixel)
{
do {
qindex++;
sh_qindex[threadIdx.y] = qindex;
sh_queue[threadIdx.y*THREADS_X*PIXELS_PER_THREAD + qindex] = queue_val;
} while (sh_queue[threadIdx.y*THREADS_X*PIXELS_PER_THREAD + qindex] != queue_val);
}
// Reload index from smem (last thread to write to smem will have updated it)
qindex = sh_qindex[threadIdx.y];
}
// Let thread 0 reserve the space required in the global list
__syncthreads();
if(threadIdx.x == 0 && threadIdx.y == 0)
{
// Find how many items are stored in each list
int total_index = 0;
#pragma unroll
for(int i=0; i<THREADS_Y; i++)
{
sh_qstart[i] = total_index;
total_index += (sh_qindex[i] + 1u);
}
// Calculate the offset in the global list
unsigned int global_offset = atomicAdd(glob_index, total_index);
#pragma unroll
for(int i=0; i<THREADS_Y; i++)
{
sh_qstart[i] += global_offset;
}
}
__syncthreads();
// Copy local queues to global queue
for(int i=0; i<=qindex; i+=THREADS_X)
{
if(i + threadIdx.x > qindex)
break;
unsigned int qvalue = sh_queue[threadIdx.y*THREADS_X*PIXELS_PER_THREAD + i + threadIdx.x];
list[sh_qstart[threadIdx.y] + i + threadIdx.x] = qvalue;
}
}
以下是调用内核的方法：
void call2DTo1DKernel(unsigned char const * const h_image)
{
// Device side allocation
unsigned char *d_image = NULL;
unsigned int *d_list = NULL;
int h_nlist, *d_nlist = NULL;
cudaMalloc((void**)&d_image, sizeof(unsigned char)*IMG_SIZE);
cudaMalloc((void**)&d_list, sizeof(unsigned int)*IMG_SIZE);
cudaMalloc((void**)&d_nlist, sizeof(int));
// Time measurement initialization
cudaEvent_t start, stop, startio, stopio;
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventCreate(&startio);
cudaEventCreate(&stopio);
// Start timer w/ io
cudaEventRecord(startio,0);
// Copy image data to device
cudaMemcpy((void*)d_image, (void*)h_image, sizeof(unsigned char)*IMG_SIZE, cudaMemcpyHostToDevice);
// Start timer
cudaEventRecord(start,0);
// Kernel call
// Phase 1 : Convert 2D binary image to 1D pixel array
dim3 dimBlock1(THREADS_X, THREADS_Y);
dim3 dimGrid1(BLOCKS_X, BLOCKS_Y);
convert2DEdgeImageTo1DArray<<<dimGrid1, dimBlock1>>>(d_image, d_list, d_nlist);
// Stop timer
cudaEventRecord(stop,0);
cudaEventSynchronize(stop);
// Stop timer w/ io
cudaEventRecord(stopio,0);
cudaEventSynchronize(stopio);
// Time measurement
cudaEventElapsedTime(&et,start,stop);
cudaEventElapsedTime(&etio,startio,stopio);
// Time measurement deinitialization
cudaEventDestroy(start);
cudaEventDestroy(stop);
cudaEventDestroy(startio);
cudaEventDestroy(stopio);
// Get list size
cudaMemcpy((void*)&h_nlist, (void*)d_nlist, sizeof(int), cudaMemcpyDeviceToHost);
// Report on console
printf("%d pixels processed...\n", h_nlist);
// Device side dealloc
cudaFree(d_image);
cudaFree(d_space);
cudaFree(d_list);
cudaFree(d_nlist);
}
非常感谢您的帮助。
答案 0 :(得分：1)
作为序言，让我建议一些有用的故障排除步骤：
cuda-memcheck运行您的代码，例如cuda-memcheck ./myapp 如果你执行上述步骤，你会发现你的内核失败了，失败与大小为4的全局写入有关。所以这会把你的注意力集中在内核的最后一段，从评论// Copy local queues to global queue
关于您的代码，您至少有两个问题：
d_nlist变量初始化为零。因此，当您对其进行原子添加时，您将您的值添加到垃圾值，当您重复该过程时，该值往往会增加。以下是一些已删除问题的代码，（我没有尝试整理您的队列复制代码）并添加了错误检查。它为我产生了可重复的结果：
$ cat t216.cu
#include <stdio.h>
#include <stdlib.h>
#define THREADS_X 32
#define THREADS_Y 4
#define PIXELS_PER_THREAD 4
#define MAX_QUEUE_LENGTH (THREADS_X*THREADS_Y*PIXELS_PER_THREAD)
#define IMG_WIDTH 256
#define IMG_HEIGHT 256
#define IMG_SIZE (IMG_WIDTH*IMG_HEIGHT)
#define BLOCKS_X (IMG_WIDTH/(THREADS_X*PIXELS_PER_THREAD))
#define BLOCKS_Y (IMG_HEIGHT/THREADS_Y)
#define cudaCheckErrors(msg) \
do { \
cudaError_t __err = cudaGetLastError(); \
if (__err != cudaSuccess) { \
fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
msg, cudaGetErrorString(__err), \
__FILE__, __LINE__); \
fprintf(stderr, "*** FAILED - ABORTING\n"); \
exit(1); \
} \
} while (0)
__global__ void convert2DEdgeImageTo1DArray( unsigned char const * const image,
unsigned int* const list, int* const glob_index ) {
unsigned int const x = blockIdx.x * THREADS_X*PIXELS_PER_THREAD + threadIdx.x;
unsigned int const y = blockIdx.y * THREADS_Y + threadIdx.y;
volatile int qindex = -1;
volatile __shared__ int sh_qindex[THREADS_Y];
volatile __shared__ int sh_qstart[THREADS_Y];
sh_qindex[threadIdx.y] = -1;
// Start by making an array
volatile __shared__ unsigned int sh_queue[MAX_QUEUE_LENGTH];
// Fill the queue
for(int i=0; i<PIXELS_PER_THREAD; i++)
{
int const xx = i*THREADS_X + x;
// Read one image pixel from global memory
unsigned char const pixel = image[y*IMG_WIDTH + xx];
unsigned int const queue_val = (y << 16) + xx;
if(pixel)
{
do {
qindex++;
sh_qindex[threadIdx.y] = qindex;
sh_queue[threadIdx.y*THREADS_X*PIXELS_PER_THREAD + qindex] = queue_val;
} while (sh_queue[threadIdx.y*THREADS_X*PIXELS_PER_THREAD + qindex] != queue_val);
}
// Reload index from smem (last thread to write to smem will have updated it)
qindex = sh_qindex[threadIdx.y];
}
// Let thread 0 reserve the space required in the global list
__syncthreads();
if(threadIdx.x == 0 && threadIdx.y == 0)
{
// Find how many items are stored in each list
int total_index = 0;
#pragma unroll
for(int i=0; i<THREADS_Y; i++)
{
sh_qstart[i] = total_index;
total_index += (sh_qindex[i] + 1u);
}
// Calculate the offset in the global list
unsigned int global_offset = atomicAdd(glob_index, total_index);
#pragma unroll
for(int i=0; i<THREADS_Y; i++)
{
sh_qstart[i] += global_offset;
}
}
__syncthreads();
// Copy local queues to global queue
/*
for(int i=0; i<=qindex; i+=THREADS_X)
{
if(i + threadIdx.x > qindex)
break;
unsigned int qvalue = sh_queue[threadIdx.y*THREADS_X*PIXELS_PER_THREAD + i + threadIdx.x];
list[sh_qstart[threadIdx.y] + i + threadIdx.x] = qvalue;
}
*/
}
void call2DTo1DKernel(unsigned char const * const h_image)
{
// Device side allocation
unsigned char *d_image = NULL;
unsigned int *d_list = NULL;
int h_nlist=0, *d_nlist = NULL;
cudaMalloc((void**)&d_image, sizeof(unsigned char)*IMG_SIZE);
cudaMalloc((void**)&d_list, sizeof(unsigned int)*IMG_SIZE);
cudaMalloc((void**)&d_nlist, sizeof(int));
cudaCheckErrors("cudamalloc fail");
// Time measurement initialization
cudaEvent_t start, stop, startio, stopio;
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventCreate(&startio);
cudaEventCreate(&stopio);
float et, etio;
// Start timer w/ io
cudaEventRecord(startio,0);
cudaMemcpy(d_nlist, &h_nlist, sizeof(int), cudaMemcpyHostToDevice);
// Copy image data to device
cudaMemcpy((void*)d_image, (void*)h_image, sizeof(unsigned char)*IMG_SIZE, cudaMemcpyHostToDevice);
cudaCheckErrors("cudamemcpy 1");
// Start timer
cudaEventRecord(start,0);
// Kernel call
// Phase 1 : Convert 2D binary image to 1D pixel array
dim3 dimBlock1(THREADS_X, THREADS_Y);
dim3 dimGrid1(BLOCKS_X, BLOCKS_Y);
convert2DEdgeImageTo1DArray<<<dimGrid1, dimBlock1>>>(d_image, d_list, d_nlist);
cudaDeviceSynchronize();
cudaCheckErrors("kernel fail");
// Stop timer
cudaEventRecord(stop,0);
cudaEventSynchronize(stop);
// Stop timer w/ io
cudaEventRecord(stopio,0);
cudaEventSynchronize(stopio);
// Time measurement
cudaEventElapsedTime(&et,start,stop);
cudaEventElapsedTime(&etio,startio,stopio);
// Time measurement deinitialization
cudaEventDestroy(start);
cudaEventDestroy(stop);
cudaEventDestroy(startio);
cudaEventDestroy(stopio);
// Get list size
cudaMemcpy((void*)&h_nlist, (void*)d_nlist, sizeof(int), cudaMemcpyDeviceToHost);
cudaCheckErrors("cudaMemcpy 2");
// Report on console
printf("%d pixels processed...\n", h_nlist);
// Device side dealloc
cudaFree(d_image);
// cudaFree(d_space);
cudaFree(d_list);
cudaFree(d_nlist);
}
int main(){
unsigned char *image;
image = (unsigned char *)malloc(IMG_SIZE * sizeof(unsigned char));
if (image == 0) {printf("malloc fail\n"); return 0;}
for (int i =0 ; i<IMG_SIZE; i++)
image[i] = rand()%2;
call2DTo1DKernel(image);
call2DTo1DKernel(image);
call2DTo1DKernel(image);
call2DTo1DKernel(image);
call2DTo1DKernel(image);
cudaCheckErrors("some error");
return 0;
}
$ nvcc -arch=sm_20 -O3 -o t216 t216.cu
$ ./t216
32617 pixels processed...
32617 pixels processed...
32617 pixels processed...
32617 pixels processed...
32617 pixels processed...
$ ./t216
32617 pixels processed...
32617 pixels processed...
32617 pixels processed...
32617 pixels processed...
32617 pixels processed...
$