Daniel Arndt comments

Results 838 comments of


                                            Daniel Arndt

Allow specifying scratch level for ScratchSpace at compile-time

I pushed an alternative implementation that doesn't require duplicating the `ScratchMemorySpace` implementation for `SYCL` but rather allows specifying the pointer types used in `ScratchMemorySpace`.

Allow specifying scratch level for ScratchSpace at compile-time

A more complete version for properly supporting the correct address space can be found at https://github.com/masterleinad/kokkos/tree/specialize_scratch_space_sycl_and_view. That branch adds the required `ViewDataHandle` specializations to actually obtain the correct pointer when...

Allow specifying scratch level for ScratchSpace at compile-time

Talking with @crtrott this morning, we were discussing - adding an integer level template parameter to `ScratchSpace`, possibly instead of defining `ScratchSpaceBase` - making these scratch spaces fulfill the memory...

Define new type trait to identify cuda-type execution space in deep copy

@nmm0 What's the status here?

Ensure kernels submitted by multiple threads to synchronous execution spaces are enqueued correctly

> We need to look at overhead for this, also its not clear to me how fence() would help with the is_running thing here? I am now using an implementation...

Ensure kernels submitted by multiple threads to synchronous execution spaces are enqueued correctly

> @masterleinad - would you like to continue discussing this issue in the Wed. meeting? Yes, we should discuss this, depending on priority tomorrow or later.

Ensure kernels submitted by multiple threads to synchronous execution spaces are enqueued correctly

> We need to look at overhead for this, [...] For an empty lambda using one thread, i.e., no contention, (which should be a worst case for applications using one...

Ensure kernels submitted by multiple threads to synchronous execution spaces are enqueued correctly

Needs #6441.

Ensure kernels submitted by multiple threads to synchronous execution spaces are enqueued correctly

Retest this please.

Ensure kernels submitted by multiple threads to synchronous execution spaces are enqueued correctly

> What exactly are the thread safety guarantees we are getting? E.g. are we also letting two threads submit to the same exec space index? If so we should test...