16
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
__global__ void funct(void){
      printf("Hello from GPU!\n");
}

int main(void){
         funct << < 2, 4 >> >();

for (int i = 0; i < 10; i++){
    cudaDeviceReset();
    //cudaDeviceSynchronize();
    printf("Hello, World from CPU!\n");

    }
return 0;
}

I thought the role of cudaDeviceReset is cudamemcpy. this case we didn't have the result of number. So we were not able to use cudamemcpy. But We used cudaDeviceReset for returning four "Hello from GPU!" result from kernel.

Is that right?

and I replaced cudaDeviceRest() with cudaDeviceSynchronize(). I saw the same result. but I couldn't know the difference between them.

최우석
  • 163
  • 1
  • 1
  • 4

1 Answers1

17

The role of cudaDeviceReset() is documented here

It is used to destroy a CUDA context, which means that all device allocations are removed.

I agree that it appears to have a synchronizing effect. However since the documentation states:

Note that this function will reset the device immediately.

I believe it is unsafe to rely on this behavior. Furthermore, the documentation also calls out the hazard of using this function in a multi-threaded app. Therefore, safe coding would dictate:

  1. Use of device synchronization (e.g. cudaDeviceSynchronize(), or cudaMemcpy(), etc.)

  2. Retrieve whatever data your application would like to preserve that may be in a device allocation, or that a recently running kernel may have updated (in device memory).

  3. Make sure that any host threads that may also have device activity associated with them, are also terminated

  4. Make sure that any C++ objects that may have device activity in their destructors are properly destroyed or out-of-scope

  5. call cudaDeviceReset() as part of application shut-down.

Note that calling cudaDeviceReset() as part of application shut-down should not be considered mandatory. Many applications will work fine without such an idiom.

This answer may also be of interest.

Robert Crovella
  • 143,785
  • 11
  • 213
  • 257
  • 1
    Calling `cudaDeviceReset()` is required for tools like `cuda-memcheck`, so its a good habit to include it at that the end. – Mikhail Apr 13 '20 at 04:31
  • I've reviewed the [documentation](https://docs.nvidia.com/cuda/cuda-memcheck/index.html) just now, and it seems to be called out as a requirement only for leak-checking. – Robert Crovella Apr 13 '20 at 05:05
  • 1
    From my salad days, I recall it also being required for the profiler, here is the comment from Mark Harris: https://devblogs.nvidia.com/pro-tip-clean-up-after-yourself-ensure-correct-profiling/ – Mikhail Apr 13 '20 at 05:18
  • Again, required in specific, atypical cases. – Robert Crovella Apr 13 '20 at 05:48
  • Does it mean that I don't have to use `cudaFree` if I use `cudaDeviceReset`? – Matan Jul 17 '21 at 15:44
  • 1
    It depends on how and where you are using `cudaFree` and for what purpose. We are unlikely to sort this out in the space of comments. – Robert Crovella Jul 17 '21 at 17:13