CUDA C:内核输出结果不好

时间:2013-06-16 04:30:15

标签: c++ c cuda nvidia nsight

首先,我想说这不是家庭作业,我只是从CUDA开始。

我正在尝试运行以下代码来添加2​​个向量...事情是每次运行后结果向量(c_device)保持不变并且不会得到两个向量的加法结果。

我尝试更改向量的长度并使用整数和无符号整数,并尝试在visual studio中的x64和win32之间移动。

我在这里附上了代码:

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

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));
}

如果您希望查看代码,请输入以下代码:http://pastebin.com/04jy1CaB

我想提一下,当将a_device复制到c_host时,它可以工作。 我还尝试将c_host复制到c_device,看看会发生什么,并发生相同的结果。

有什么建议吗?

1 个答案:

答案 0 :(得分:1)

好的,所以感谢talonmies对我的问题的评论,我意识到我没有做足够的错误检查,当我做了额外的检查时,我发现我将错误的参数传递给内核调用。

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。 如果您发现我的默认值为0,则它​​们应为1.

调试器万岁:)

感谢你帮助talonmies。