Timing cudaEventRecord() ok for cpu timing?

[shreddn]

I’m writing my first cuda program to compare performance of the device over the host. Here is the code I used to record the host computation time:

    //Create and Record cuda events to time cpu execution
    cudaEvent_t start1, stop1;
    cudaEventCreate(&start1);
    cudaEventRecord(start1,0);
    float cputime;

    doCompOnHost(a_h, b_h, c_h, result_h, N );

    //End Time events for host
    cudaEventCreate(&stop1);
    cudaEventRecord(stop1, 0);
    cudaEventSynchronize(stop1);
    cudaEventElapsedTime(&cputime, start1, stop1);
    cudaEventDestroy(start1);
    cudaEventDestroy(stop1);

Would this return an accurate time for the cpu compuation?

Thanks in advance.

[jph4599]

Short answer, no…events only time GPU.

Long answer from Mister Anderson:

Events precisely measure the actual time taken on the GPU for a kernel (or series of kernels/memcpys) to run. Because the use a high precision timer, you can measure the time of just one kernel call. But, depending on what you are doing, the time measured by the events my not be an accurate measurement of what you want (more in a bit)

Wall clock time measurements are not as precise, you will need to time several hundred or thousand kernel calls in a row to build up a precise average time per call. But, you are accurately measuring the actual time it takes to perform the task including driver overhead and any CPU computations that may be part of your algorithm.

So, the two have different uses. I have a bit of code that runs for a little while on the CPU (5% of the time) and then uses the calculated result from the CPU to launch a GPU kernel which runs for the other 95% of the time.

1. I can use events to measure the time of execution of the kernel on the GPU. Uses: With a precise measurement of the actual kernel execution time, I can compute the effective GFLOP/s and GiB/s rates to see if I am pushing the hardware limits.

But, if what I want is to compare to a CPU code, the time I’m measuring using events is only 95% of the total wall clock time from start to finish!

So: 2) I can use the wall clock time to measure the total time from start to finish of the algorithm to compare to a CPU implementation of the same code.

Of course, you can always use the profiler tool to time kernel calls to give you the same information as events. But events are a little more flexible than the profiler in that they can be inserted in streams and time a series of multiple kernels/memcpys.

[shreddn]

Short answer, no…events only time GPU.

Long answer from Mister Anderson:

Thanks, I just read the kernel calls are asynchronous, so it makes sense why it wouldn’t work for CPU times. Do you know of a comprehensible CPU timing tutorial in C. I just want to time a cpu function. External Media