cudaMemset() problem

Accelerated Computing CUDA CUDA Programming and Performance
1

Has anyone else noticed that in the 0.8b package that cudaMemset() does nothing? For me it appears to leave the specified memory buffer on the device unchanged.

For example:

CUDA_SAFE_CALL( cudaMalloc( (void**) &DevOutBuf, 1024));
CUDA_SAFE_CALL( cudaMemset( (void *) DevOutBuf, 0, 1024));

does not set every byte to 0 as it should. If I fill the buffer with a pattern and then try to cudaMemset() it, the original bytes are not overwritten.

2

I have noticed this as well. I’m currently doing a Memcpy with a large buffer of zeroes as the input.

3

For the record I am using the recommended driver that came with the CUDA package (97.73).

I also noticed that cudaMemset() is OK in the emulator but fails on the device.

4

Yeah, I’m having this problem as well, so are others I know writing for CUDA :( Has anyone gotten this to work?

5

I just double checked; cudaMemset(…) and cudaMemset2D(…) both work fine on Linux, using the recommended 97.51 driver.

Is everyone who is having problems using Windows?

6

Same problem! cudamemset not doing anything. The weird part is it doesn’t always happen, but rather randomly… 4 years and nothing solved?

7

Works without a flaw for me. It’s either your code or your driver installation.

Try the following code

int* hbuff1, *hbuff2, *dbuff;

hbuff1=(int*) malloc(1024);

hbuff2=(int*) malloc(1024);

memset(hbuff1, 5, 1025);

cudaMalloc((void**) &dbuff, 1024);

cudaMemcpy((void *)dbuff, (void *)hbuff1,1024,cudaMemcpyHostToDevice);

cudaMemset((void *) dbuff, 0, 1024);

cudaMemcpy((void *)hbuff2, (void *)dbuff ,1024,cudaMemcpyDeviceToHost);

assert(memcmp(hbuff1,hbuff2,1024)!=0);
8

I think this is a bug in cudaMemset, but feel free to point out my error.
The following code yields a string of 16843009s instead of 1s on both Ubuntu 11.04 amd64 and Windows 7x64, latest dev drivers.

[font=“Courier New”]#include <stdio.h>
#include <stdlib.h>
#include <cuda.h>
#define GRID_SIZE 128

int main (int argc , char* argv )
{
int* gridOccupancy = 0;

cudaMalloc((void**)&gridOccupancy,GRID_SIZE*GRID_SIZE*sizeof(int));
cudaMemset(gridOccupancy,1,GRID_SIZE*GRID_SIZE*sizeof(int));

int* hgridOccupancy = (int*) calloc(GRID_SIZE*GRID_SIZE,sizeof(int));
cudaMemcpy(hgridOccupancy,gridOccupancy,GRID_SIZE*GRID_SIZE*sizeof(int),cudaMemcpyDeviceToHost);

for (int m = 0; m < GRID_SIZE*GRID_SIZE; m++) {
	printf ("%d %d\n",m,hgridOccupancy[m]);
}

free(hgridOccupancy);

/* Cleanup */
cudaFree(gridOccupancy);

return EXIT_SUCCESS ;

}[/font]

9

Ok. There is no error. cudaMemset works likes memset - it sets bytes, not words. In your code sample, you are effectively setting each byte of memory to a value of 1, not each word. The results you are seeing are expected.