Is CUDA warp scheduling deterministic? -


I am thinking that the warranty scheduling order of the CUDA application is deterministic

especially thinking I'm sticking to the sequence of execution execution on the same device with multiple runs in the same kernel as the same input data. If not, is there anything that can stress the order of execution of execution (if an order dependent is debugging algorithms then say in case)?

Accurate behavior of scheduling CUDA taana has not been defined. So you can not depend on it. Specifically, if several warps are ready to be executed in a given issue slot, then there is no description that the taan will be selected by the warp scheduler.

There is no external method for precise control, warrant execution orders.

It is certainly possible that the code can be prepared which determines taane ID, something like this:

  #include  & Gt; & Gt; (); CudaDeviceSynchronize (); CudaMemcpyFromSymbol (h_warp_order, warp_order, N_WARPS * sizeof (int)); For (int i = 0; i  

To execute at one time, it will be very disabled by giving only a taunt, of course.

In general, the best parallel algorithm has little or no order dependency.


Comments

Popular posts from this blog

java - org.apache.http.ProtocolException: Target host is not specified -

java - Gradle dependencies: compile project by relative path -

ruby on rails - Object doesn't support #inspect when used with .include -