Ben Vanik
Ben Vanik
Agreed! Related to the training wheels, #9295 shows a good example of the kind of code we want to be able to entirely group into dispatch regions. Specifically as it...
The `IREE_SIZE_OPTIMIZED` is saying "build IREE in a size-optimized configuration" - since IREE is a library we want to be able to control this independently of global build configuration -...
(sorry, incomplete thought) enabling IREE_SIZE_OPTIMIZED if in MinSizeRel would make sense if it didn't also aggressively disable things like status messages and such - IREE_SIZE_OPTIMIZED is like an extreme MinSizeRel...
I'm not sure - `llvm-strip --strip-debug` can strip debug info from bitcode if a reproducer is needed or `strip --strip-debug` can be used for object files. An LLVM engineer wanting...
Yep! You could ask on discord/discourse and see if anyone wants to take it on but otherwise it should be fairly simple to get started and landed - a first...
Gotcha, thanks for explaining as the intended workflow was not clear to me! Sounds like you think this is a good workflow for MLIR codegen targeting LLVM: it's really useful...
(where possible the best solution here is still to fuse this into producers - when we are doing that we should only need to pad/swizzle in very rare cases)
(still valid?)
To clarify, this just means we should be fusing those dispatches together such that they are accessing the source data once for all 3 reductions, correct? ("pattern matching ML layers"...
Cool! This will be a really great improvement!