texture is not thread-safe ?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I tried to lauch the kernel of the sample simpleTexture in a host thread, the thread code is

float angle = 0.5f;    // angle to rotate image by (in radians)

float* d_data = NULL;

unsigned int width, height;

cudaArray* cu_array;

DWORD WINAPI Thread_Proc(LPVOID para)

{

    CUDA_SAFE_CALL(cudaSetDevice(0));

    cudaBindTexture( tex, cu_array);

    dim3 dimBlock(8, 8, 1);

    dim3 dimGrid(width / dimBlock.x, height / dimBlock.y, 1);

    transformKernel<<< dimGrid, dimBlock, 0 >>>( d_data, width, height, angle);

    CUT_CHECK_ERROR("Kernel execution failed");

    cudaUnbindTexture( tex);

}

The global variables and memory is initialized by the main thread as done in the sample.

CUT_CHECK_ERROR report a “unspecified driver error” in the debug mode. The EmuDebug mode work well.

Any idea for this?

2

I also tried to put the commands

cudaBindTexture( tex, cu_array);

and

cudaUnbindTexture( tex);

into the main thread. Then the command

transformKernel<<< dimGrid, dimBlock, 0 >>>( d_data, width, height, angle);

never return.