ptheywood / cuda-concurrency-detection Goto Github PK

Make use of streams optional behind a template arg / extra functional argument (easily)

Cuda even to to record in the default stream before and after the batch of concurrent work (or serial if no streams).

Record time externally.
Repeat each one N times for average time.
Compare using streams and not using streams - streams should be quicker (for the right problem size.)

This again relies on several assumptions, will need flexible time comparison + ignore any outliers?

Need to check how this interacts with elapsed time. Might need to change the order of synchronisation to be global before per-stream?

   // Sync each stop evet
            for(int b = 0; b < batches; b++){
                cudaEvent_t stop = stopEvents.at(b);
                CUDA_CALL(cudaEventSynchronize(stop));
            }
            CUDA_CALL(cudaEventSynchronize(g_stop));

Alternative is to figure out based on s_start + g_stop?

It's noted in the CUDA docs that cudaEventElapsedtime is not reliable when using the non-default stream 😢

Rather than solving a large large saxpy problem, switch to solving multiple saxpy instances? Closer map to target application in that case, though functionally should be the same.

Profile with nvprof - this produces an sqlite database.

Write a script which interrogates the sqlite databse, to check for concurrency ( start times + end times in streams should overlap if any concurrency present)