Ben Vanik
Ben Vanik
Two issues: data tiling as it is plumbed through codegen today is fundamentally incompatible with jit eval, and jit eval today should be stripping all transfer ops and device references....
tests are always good; @MaheshRavishankar, is this something you can rebase or is it stale?
This was/is an issue in CUDA as well and the last time a fix was attempted it ended in sadness. In CUDA changing the device would flush it and also...
Good point - if we use the queue for doing submissions then having that be common would also solve the CUDA bugs we have (or at least fix most of...
Good point - we should pool all of those we can anyway and only submit to the queue (or force a blocking switch) when the pool needs to grow. Async...
Oh, for module loading we really need to fix the compiler to link executables together - ideally we should only have one HAL executable per device (though there may be...
yep! any ops that don't impact data (reshapes, casts, bitcasts, etc) should be able to propagate through or be replicated on either side
nice! thanks for looking at this - tests that only run on certain build systems, on certain targets, are persistently disabled, or run only with specific non-standard configuration is a...
interesting - the compiler `Source` seems to hang on to the buffer provided in `wrap_buffer`, but you could try adding a print to its `close()` - the buffer as provided...
Interesting approach in #20953! I think we can refine this a bit to avoid the need to inject the target registry (which we really want to avoid doing when possible)...