Why these measurements so different?

Accelerated Computing CUDA CUDA Programming and Performance
1

In the host code, I measured the total time of these six kernel launchs with CUDA timer. In the windows 7, the result is about 1 ms and in Linux , the result is 0.17ms. But in the output of Visual Profiler, the total consumed time of these kernel is 0.13ms! why there is such big diferences between these results? Thanks for your help!

cutCreateTimer(&timer);

    cutStartTimer(timer);

    cufftExecD2Z(Plan,cubfft,fft_result);

    cudaThreadSynchronize();

    // first kernel

tri_transpose<<<grid,threads>>>(fft_result,cu_result, argu);

    cudaThreadSynchronize();

    //second kernel

cudaFuncSetCacheConfig(tri_solver, cudaFuncCachePreferShared);

tri_solver<<<grid2,threads2>>>(cu_result,cu_act_trace,cu_tri_main);

    cudaThreadSynchronize();

    // third kernel

tri_transpose2<<<grid,threads>>>(cu_result, cubfft);

    cudaThreadSynchronize();

    // fourth kernel

cufftExecD2Z(Plan,cubfft,fft_result);

    cudaThreadSynchronize();

    // fifth kernel

post_process<<<grid2,threads2>>>(fft_result,cu_result,argu);

   // sixth kernel

cutStopTimer(timer);
2

I believe cutil timers only have millisecond precision on Windows. Use a higher precision timer or time multiple iterations before concluding that the code is slower on windows (which still might be the case though - kernel invocations are slower on windows).

3

The windows WDDM drivers also have substantially more latency and overhead in calling kernels than the linux drivers.