[英]CUDA C: Kernel outputs bad results
first of all I want to say that this is not a homework assignment and I'm just beginning with CUDA. 首先,我想说这不是一项家庭作业,我只是从CUDA开始。
I'm trying to run the following code to add 2 vectors... The thing is that after every run the result vector (c_device) stays the same and doesn't get the result of the addition of the two vectors. 我正在尝试运行以下代码以添加2个向量...事实是,每次运行后,结果向量(c_device)保持不变,并且没有得到两个向量相加的结果。
I have tried changing the length of the vector and working with ints and unsigned ints and also trying to move between x64 and win32 in visual studio. 我试图更改向量的长度,并使用ints和unsigned int,并且还尝试在Visual Studio中在x64和win32之间移动。
I have attached the code here: 我在这里附加了代码:
This is the .h file 这是.h文件
#ifndef ODINN_CUDA_MAIN_H
#define ODINN_CUDA_MAIN_H
#define ARR_SIZE 100
#define ITER_AMOUNT 1
typedef enum cudaError cudaError_t;
static void HandleError(cudaError_t err, const char *file, int line) {
if (err != CUDA_SUCCESS) {
printf("%s in %s at line %d\n", cudaGetErrorString(err), file, line);
exit(EXIT_FAILURE);
}
}
#define HANDLE_ERROR(err) (HandleError(err, __FILE__, __LINE__))
#define GET_CURRENT_CLOCKS(var) (var = clock())
#define GET_CLOCK_INTERVAL_SEC(start, end, result) (result = ((double)((double)end - (double)start) / (double)CLOCKS_PER_SEC))
__host__ dim3 requestBlockSize(int x, int y=0, int z=0);
__host__ dim3 requestNumBlocks(int x, int y=0, int z=0);
__host__ void allocateVectors(unsigned int **a_host, unsigned int **b_host, unsigned int **c_host, unsigned int **a_device, unsigned int **b_device, unsigned int **c_device);
__global__ void addVectors(unsigned int* a, unsigned int* b, unsigned int* result, int n);
__host__ void cleanUp(unsigned int *a_host, unsigned int *b_host, unsigned int *c_host, unsigned int *a_device, unsigned int *b_device, unsigned int *c_device);
#endif
And this is the .cu file: 这是.cu文件:
#include <cuda.h>
#include <stdio.h>
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <math.h>
#include "main.h"
static cudaDeviceProp prop;
int main(void) {
// Start lazy init now so first cudaMallow will run faster.
cudaSetDevice(0);
cudaFree(0);
unsigned int *a_host, *b_host, *c_host;
unsigned int *a_device, *b_device, *c_device;
double delta_in_sec;
size_t size = sizeof(unsigned int) * ARR_SIZE;
clock_t start_clock, end_clock;
HANDLE_ERROR(cudaGetDeviceProperties(&prop, 0));
dim3 block_size = requestBlockSize(1024);
int blocks_requested = floor((double)(ARR_SIZE / block_size.x));
dim3 n_blocks = requestNumBlocks(blocks_requested > 0 ? blocks_requested : 1);
fprintf(stdout, "Allocating vectors ...\n");
allocateVectors(&a_host, &b_host, &c_host, &a_device, &b_device, &c_device);
fprintf(stdout, "Copying to device ...\n");
HANDLE_ERROR(cudaMemcpy(a_device, a_host, size, cudaMemcpyHostToDevice));
HANDLE_ERROR(cudaMemcpy(b_device, b_host, size, cudaMemcpyHostToDevice));
fprintf(stdout, "Running kernel ...\n");
GET_CURRENT_CLOCKS(start_clock);
for(int i=0; i<ITER_AMOUNT; i++) {\
addVectors<<<n_blocks, block_size>>>(a_device, b_device, c_device, ARR_SIZE);
}
GET_CURRENT_CLOCKS(end_clock);
GET_CLOCK_INTERVAL_SEC(start_clock, end_clock, delta_in_sec);
fprintf(stdout, "Runtime of kernel %d times on arrays in length %d took %f seconds\n"
"Copying results back to host ...\n", ITER_AMOUNT, ARR_SIZE, delta_in_sec);
HANDLE_ERROR(cudaMemcpy(c_host, c_device, size, cudaMemcpyDeviceToHost));;
fprintf(stdout, "%u + %u != %u\n", a_host[0], b_host[0], c_host[0]);
fprintf(stdout, "Cleaning up ...\n");
cleanUp(a_host, b_host, c_host, a_device, b_device, c_device);
fprintf(stdout, "Done!\n");
}
__host__ dim3 requestBlockSize(int x, int y, int z) {
dim3 blocksize(
x <= prop.maxThreadsDim[0] ? x : prop.maxThreadsDim[0],
y <= prop.maxThreadsDim[1] ? y : prop.maxThreadsDim[1],
z <= prop.maxThreadsDim[2] ? z : prop.maxThreadsDim[2]
);
return blocksize;
}
__host__ dim3 requestNumBlocks(int x, int y, int z) {
dim3 numblocks(x, y, z);
return numblocks;
}
__host__ void allocateVectors(unsigned int **a_host, unsigned int **b_host, unsigned int **c_host, unsigned int **a_device, unsigned int **b_device, unsigned int **c_device) {
size_t size = sizeof(unsigned int) * ARR_SIZE;
*a_host = (unsigned int *)malloc(size);
*b_host = (unsigned int *)malloc(size);
*c_host = (unsigned int *)malloc(size);
HANDLE_ERROR(cudaMalloc((void **)a_device, size));
HANDLE_ERROR(cudaMalloc((void **)b_device, size));
HANDLE_ERROR(cudaMalloc((void **)c_device, size));
srand(time(NULL));
for(int i=0; i<ARR_SIZE; i++) {
(*a_host)[i] = rand() % ARR_SIZE;
(*b_host)[i] = rand() % ARR_SIZE;
}
}
__global__ void addVectors(unsigned int* a, unsigned int* b, unsigned int* result, int n) {
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if(idx >= 0 && idx < n)
result[idx] = a[idx] + b[idx];
}
__host__ void cleanUp(unsigned int *a_host, unsigned int *b_host, unsigned int *c_host, unsigned int *a_device, unsigned int *b_device, unsigned int *c_device) {
free(a_host);
free(b_host);
free(c_host);
HANDLE_ERROR(cudaFree(a_device));
HANDLE_ERROR(cudaFree(b_device));
HANDLE_ERROR(cudaFree(c_device));
}
Here is the link to pastebin if you prefer looking at the code there: http://pastebin.com/04jy1CaB 如果您更喜欢查看其中的代码,请访问以下链接到pastebin: http : //pastebin.com/04jy1CaB
I would like to mention that when copying a_device to c_host it works. 我想提到的是,将a_device复制到c_host时可以正常工作。 I have also tried to copy c_host to c_device and see what happens and the same result happened.
我也尝试将c_host复制到c_device,看看会发生什么,并且发生了相同的结果。
Any suggestions? 有什么建议么?
OK, so thanks to talonmies's comment on my question I have realized that I wasn't doing enough error checking and when I did the additional checks I found out that I was passing bad arguments to the kernel call. 好的,因此感谢talonmies对我的问题的评论,我意识到我没有进行足够的错误检查,当我进行其他检查时,发现我正在将错误的参数传递给内核调用。
I was passing invalid dim3 y and z values to the kernel thread amount and block amount parameter. 我将无效的dim3 y和z值传递给内核线程数量和块数量参数。 If you notice my default values were 0 and they should be 1.
如果您注意到我的默认值是0,则应为1。
Long live the debugger :) 调试器万岁:)
Thanks for your help talonmies. 感谢您的帮助。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.