Question

I am using Cuda SDK 4.0 and am encountering an issue which has taken me 2 days to whittle down into the following code.

#include <cuda.h>
#include <cuda_runtime.h>
void main (int argc, char ** argv) {

    int* test;
    cudaError_t err;

    err = cudaSetDevice(   1   ); err = cudaMallocHost(&test, 1024*sizeof(int));    
    err = cudaSetDevice(   0   ); err = cudaFreeHost(test);    
} 

This throws the following error when calling cudaFreeHost:

First-chance exception at 0x000007fefd96aa7d in Test.exe: Microsoft C++ exception: cudaError_enum at memory location 0x0022f958..

The err value is cudaErrorInvalidValue

The same error occurs for this variation:

err = cudaSetDevice(   0   ); err = cudaMallocHost(&test, 1024*sizeof(int));    
err = cudaSetDevice(   1   ); err = cudaFreeHost(test);

The following variations dont throw the error:

err = cudaSetDevice(   0   ); err = cudaMallocHost(&test, 1024*sizeof(int));    
err = cudaSetDevice(   0   ); err = cudaFreeHost(test);

and

err = cudaSetDevice(   1   ); err = cudaMallocHost(&test, 1024*sizeof(int));    
err = cudaSetDevice(   1   ); err = cudaFreeHost(test);

I was under the impression you only needed to call cudaSetDevice if you want to allocate memory on a specific GPU. In the above example I am only allocating pinned memory on the CPU.

Is this a bug or did I miss something in the manual?

Answer 1

I found the problem. cudaHostAlloc and cudaMallocHost ARE NOT THE SAME.

For anyone who encounters this problem the solution is to use

cudaHostAlloc(&test, 1024*sizeof(int),cudaHostAllocPortable);

instead of

cudaMallocHost(&test, 1024*sizeof(int));