All geek questions in one place

CUDA initialization error after fork

I get an "initialization error" after calling fork (). If I run the same program without a plug, everything works fine.

if (fork() == 0) {
    ...
    cudaMalloc(....);
    ...
}

What can cause this?

The following is a complete example. If I comment on the call to cudaGetDeviceCount, it works fine.

#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <cuda_runtime.h>

#define PERR(call) \
  if (call) {\
   fprintf(stderr, "%s:%d Error [%s] on "#call"\n", __FILE__, __LINE__,\
      cudaGetErrorString(cudaGetLastError()));\
   exit(1);\
  }

int
main(int argc, char **argv)
{
  float *v_d;
  int gpucount;

  cudaGetDeviceCount(&gpucount);

  if (fork() == 0) {
    cudaSetDevice(0);
    PERR(cudaMalloc(&v_d, 1000*sizeof(float)));
  }
  wait(NULL);
  return 0;
}

Simple Makefile:

PROGS = fork
CUDA_PATH = /usr/local/cuda
CXXFLAGS = -g -O0 -Wall
CXXINCLUDES = -I$(CUDA_PATH)/include
NVCC := $(CUDA_PATH)/bin/nvcc -ccbin $(CXX) -Xcompiler "$(CXXFLAGS)"

fork: fork.cxx
        $(NVCC) $^ -o $@ $(LIBS)

clean:
        (rm $(PROGS) *.o)

In this case, I am trying to get only the number of devices available in the parent process. This approach does this:

  if (fork() == 0) {
    PERR(cudaGetDeviceCount(&gpucount));
    return(gpucount);
  }
  wait(&gpucount);
  gpucount =  WEXITSTATUS(gpucount);
+4
source share
1 answer

fork()creates a child process. Processes have their own address spaces. A CUDA context cannot be shared between two different processes for many reasons, one of which is that different pointers will be meaningless in a different address space.

CUDA fork(), . cudaSetDevice(0); CUDA, , cudaGetDeviceCount();

, , - CUDA , . , ( CUDA simpleIPC ). ( CUDA .)

/ .

( 2 CUDA), :

$ cat t345.cu
#include <unistd.h>     /* Symbolic Constants */
#include <sys/types.h>  /* Primitive System Data Types */
#include <errno.h>      /* Errors */
#include <stdio.h>      /* Input/Output */
#include <sys/wait.h>   /* Wait for Process Termination */
#include <stdlib.h>     /* General Utilities */


#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 addkernel(int *data){
  *data += 1;
}

int main()
{
    pid_t childpid; /* variable to store the child pid */
    int retval;     /* child process: user-provided return code */
    int status;     /* parent process: child exit status */

    /* only 1 int variable is needed because each process would have its
       own instance of the variable
       here, 2 int variables are used for clarity */

    /* now create new process */
    childpid = fork();

    if (childpid >= 0) /* fork succeeded */
    {
        if (childpid == 0) /* fork() returns 0 to the child process */
        {
            printf("CHILD: I am the child process!\n");
            printf("CHILD: Here my PID: %d\n", getpid());
            printf("CHILD: My parent PID is: %d\n", getppid());
            printf("CHILD: The value of my copy of childpid is: %d\n", childpid);
            int *h_a, *d_a;
            h_a = (int *)malloc(sizeof(int));
            cudaSetDevice(0);
            cudaCheckErrors("CHILD cudaSetDevice fail");
            cudaMalloc(&d_a, sizeof(int));
            cudaCheckErrors("cudaMalloc fail");
            *h_a = 1;
            cudaMemcpy(d_a, h_a, sizeof(int), cudaMemcpyHostToDevice);
            cudaCheckErrors("cudaMemcpy H2D fail");
            addkernel<<<1,1>>>(d_a);
            cudaDeviceSynchronize();
            cudaCheckErrors("kernel fail");
            cudaMemcpy(h_a, d_a, sizeof(int), cudaMemcpyDeviceToHost);
            cudaCheckErrors("cudaMemcpy D2H fail");
            printf("CHILD: result: %d\n", *h_a);

            printf("CHILD: Sleeping for 1 second...\n");
            sleep(1); /* sleep for 1 second */
            cudaDeviceReset();
            printf("CHILD: Enter an exit value (0 to 255): ");
            scanf(" %d", &retval);
            printf("CHILD: Goodbye!\n");
            exit(retval); /* child exits with user-provided return code */
        }
        else /* fork() returns new pid to the parent process */
        {
            printf("PARENT: I am the parent process!\n");
            printf("PARENT: Here my PID: %d\n", getpid());
            printf("PARENT: The value of my copy of childpid is %d\n", childpid);
            printf("PARENT: I will now wait for my child to exit.\n");
            int *h_a, *d_a;
            h_a = (int *)malloc(sizeof(int));
            cudaSetDevice(1);
            cudaCheckErrors("PARENT cudaSetDevice fail");
            cudaMalloc(&d_a, sizeof(int));
            cudaCheckErrors("cudaMalloc fail");
            *h_a = 2;
            cudaMemcpy(d_a, h_a, sizeof(int), cudaMemcpyHostToDevice);
            cudaCheckErrors("cudaMemcpy H2D fail");
            addkernel<<<1,1>>>(d_a);
            cudaDeviceSynchronize();
            cudaCheckErrors("kernel fail");
            cudaMemcpy(h_a, d_a, sizeof(int), cudaMemcpyDeviceToHost);
            cudaCheckErrors("cudaMemcpy D2H fail");
            printf("PARENT: result: %d\n", *h_a);
            wait(&status); /* wait for child to exit, and store its status */
            printf("PARENT: Child exit code is: %d\n", WEXITSTATUS(status));
            cudaSetDevice(0);
            cudaCheckErrors("PARENT cudaSetDevice  2 fail");
            int *h_a2, *d_a2;
            cudaMalloc(&d_a2, sizeof(int));
            cudaCheckErrors("cudaMalloc fail");
            h_a2 = (int *)malloc(sizeof(int));
            *h_a2 = 5;
            cudaMemcpy(d_a2, h_a2, sizeof(int), cudaMemcpyHostToDevice);
            cudaCheckErrors("cudaMemcpy H2D fail");
            addkernel<<<1,1>>>(d_a2);
            cudaDeviceSynchronize();
            cudaCheckErrors("kernel fail");
            cudaMemcpy(h_a2, d_a2, sizeof(int), cudaMemcpyDeviceToHost);
            cudaCheckErrors("cudaMemcpy D2H fail");
            printf("PARENT: result2: %d\n", *h_a2);
            printf("PARENT: Goodbye!\n");
            exit(0);  /* parent exits */
        }
    }
    else /* fork returns -1 on failure */
    {
        perror("fork"); /* display error message */
        exit(0);
    }
}
$ nvcc -arch=sm_20 -o t345 t345.cu
$ ./t345
CHILD: I am the child process!
CHILD: Here my PID: 23603
CHILD: My parent PID is: 23602
CHILD: The value of my copy of childpid is: 0
PARENT: I am the parent process!
PARENT: Here my PID: 23602
PARENT: The value of my copy of childpid is 23603
PARENT: I will now wait for my child to exit.
CHILD: result: 2
CHILD: Sleeping for 1 second...
PARENT: result: 3
CHILD: Enter an exit value (0 to 255): 10
CHILD: Goodbye!
PARENT: Child exit code is: 10
PARENT: result2: 6
PARENT: Goodbye!
$

( )

+4

Source: https://habr.com/ru/post/1535648/

More articles:


All Articles