Running 2 GPUs on CUDA Low Speed Memory Copy Operations Problem

[ErenS]

Hi Everyone,

I am trying to write an application which uses 2 GPUs. I started with “MultiGPU” sample code and build an application.

My problem is the low speed of memory copy operations. For example when i used “cudaMemcpyToSymbol” function, it approximately takes 40 ms when copying 7205763 bytes of data. The same copying operation takes place 4 ms in our single-GPU application. Here is the GPU thread code:

// HOST SIDE

unsigned char* frame_out_host;

…

// DEVICE SIDE

device unsigned char frame_rgb  [7205763];

device unsigned char frame_o    [7205764];

…

static CUT_THREADPROC gpuThread2(int * device)

{

CUDA_SAFE_CALL(cudaSetDevice(*device));

CUT_SAFE_CALL(cutResetTimer(timer));

CUT_SAFE_CALL(cutStartTimer(timer));

// Copy some values to the buffers.

CUDA_SAFE_CALL(cudaMemcpyToSymbol(frame_rgb, &frame[device7205763], (720*576)*3, 0, cudaMemcpyHostToDevice) );

CUT_SAFE_CALL(cutStopTimer(timer));

printf(“Processing time: %5.5f (ms)\n”, cutGetTimerValue(timer));

// Invoke kernel on the device.

kernel_start();

// Get the results back.

CUDA_SAFE_CALL(cudaMemcpyFromSymbol(&frame_out_host[(device)7205764/s_gpuCount], frame_o, (720*576)*3/s_gpuCount, 0, cudaMemcpyDeviceToHost) );

CUT_THREADEND;

}

Additionally, the application outputs correct results in “EmuDebug” mode but fails in “Release” mode. Am I doing something wrong in multi-threading issues?

Thanks…