Rob Suderman
Rob Suderman
So what is weirder is the `i64` values seem to be constructed later in the lowering process. Working to understand why at the moment.
Alright, figured it out. The problem is the conversion for `tosa.apply_scale` generates the `i64` operations to handle the rescaling behavior. This creates `i64` types that the gpu doesn't like. This...
This should handle the fix once it lands and integrates: https://reviews.llvm.org/D122948
@antiagainst Good news: we have the i32 only version of the rescale lowering, which is pretty trivial to enable. There is an optional boolean for whether to include the 32-bit-only...
It would be awesome to see this land. ONNX has done an admirable amount of work in defragmenting the ML space and having it target a hardware / codegen motivated...
`tosa.resize` is landing a change relatively soon so once the new version is checked in we can find the no-op canonicalizer. We should try to identify both the cases for...
This should be fixed now.
Verified that it removes the following case: ``` %233 = "tosa.resize"(%232) {border = [0, 0], mode = "BILINEAR", offset = [0, 0], scale = [2, 2, 2, 2]} : (tensor)...
There should be some performance improvements landed at IREE head including the broadcast case and the pass-through canonicalization. I have a patch that handles single broadcasting in flight but it...
Ah okay, I see the failure. Technically we only support gathers on i32 values at most. This is a TOSA limitation. There are actually other failures as well (e.g. topk...