search

mratsim / weave

A state-of-the-art multithreading runtime: message-passing based, fast, scalable, ultra-low overhead
Other
541 stars 21 forks source link

Cuda scheduling #133

Open mratsim opened 4 years ago

mratsim commented 4 years ago

For numerical computing it would be interesting to schedule and keep track of Cuda kernels on Nvidia GPUs with an interface similar to the CPU parallel API.

The focus is on task parallelism and dataflow parallelism (task graphs). Data parallelism (parallelFor) should be handled in the GPU kernel.

From this presentation https://developer.download.nvidia.com/CUDA/training/StreamsAndConcurrencyWebinar.pdf, we can use CudaEvent for synchronizing concurrent kernels: image image (note there seems to be a typo in the code it should be

cudaStreamWaitEvent ( stream, event );       // wait for event in stream1

At first glance an event seems to be fired when the stream is empty.

mratsim commented 4 years ago

Interesting concurrent queue for scheduling tasks on GPU, the broker queue: https://arbook.icg.tugraz.at/schmalstieg/Schmalstieg_353.pdf