Open benvanik opened 4 months ago
This seems to be already fixed.
on latest build following works fine.
./iree-compile --compile-mode=hal-executable --iree-hal-target-backends=cuda ~/dev/iree/tools/test/iree-benchmark-executable.mlir
./iree-compile --compile-mode=hal-executable --iree-hal-target-backends=vulkan-spirv ~/dev/iree/tools/test/iree-benchmark-executable.mlir
Today it does not seem to lower to
WorkgroupSize
(SPIR-V) orblockDim
(CUDA):Supporting this is needed to write portable inputs that lower to all backends (we can't use gpu dialect ops).
This input with
iree-compile --iree-hal-target-backends=vulkan-spirv
or=cuda
fails in the same way, passes onllvm-cpu
: