在CUDA中分配设备变量时出现问题 [英] Having problem assigning a device variable in CUDA

问题描述

我无法尝试为设备变量分配值，然后将其复制到主机变量。

I'm having trouble trying to assign a value to a device variable and then copying this to a host variable.

我从d_test和h_test = 0.0开始。我有一个简单的内核将设备变量d_test设置为1.0。然后我将它复制到宿主变量h_test并打印。问题是，当我打印我得到h_test = 0.0。我究竟做错了什么？以下是代码：

I start with d_test and h_test = 0.0. I have a simple kernel to set the device variable, d_test, to 1.0. I then copy this to the host variable h_test and print. The problem is that when I print I get h_test = 0.0. What am I doing wrong? Here's the code:

// -*- mode: C -*-
#include <stdio.h>
#include <stdlib.h>
#include <cuda_runtime.h>

// device variable and kernel
__device__ float d_test;
__global__ void kernel1(float d_test) { d_test = 1.0; }


int main() {

  // initialise variables
  float h_test = 0.0;
  cudaMemset(&d_test,0,sizeof(float));

  // invoke kernel
  kernel1 <<<1,1>>> (d_test);

  // Copy device variable to host and print
  cudaMemcpy(&h_test,&d_test,sizeof(float),cudaMemcpyDeviceToHost);
  printf("%f\n",h_test);  

}

推荐答案

1. 正如pezcode所说， kernel1 的参数 d_test 会影响你的全局变量，因此当它分配给 d_test 时，的参数，而不是你想要的全局变量。 kernel1 不需要在此示例中使用参数。

1. As pezcode correctly notes, kernel1's parameter d_test shadows your global variable, so when it assigns to d_test, it is actually changing the value of its parameter, instead of the global variable as you intend. kernel1 need not take an argument in this example.

而不是 cudaMemcpy 时，使用 cudaMemcpyFromSymbol li>


Instead of cudaMemcpy, use cudaMemcpyFromSymbol when copying from a global __device__ variable.

这是完整的解决方案：

// -*- mode: C -*-
#include <stdio.h>
#include <stdlib.h>
#include <cuda_runtime.h>

// device variable and kernel
__device__ float d_test;
__global__ void kernel1() { d_test = 1.0; }

int main() {

  // initialise variables
  float h_test = 0.0;
  cudaMemset(&d_test,0,sizeof(float));

  // invoke kernel
  kernel1 <<<1,1>>> ();

  // Copy device variable to host and print
  cudaMemcpyFromSymbol(&h_test, "d_test", sizeof(float), 0, cudaMemcpyDeviceToHost);
  printf("%f\n",h_test);  
}

输出：

$ nvcc test.cu -run
1.000000

这篇关于在CUDA中分配设备变量时出现问题的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！