Jerry Wu
Jerry Wu
Make it optional and off by default on PR runs SGTM
I'm actually more than happy to try to fix this : ) Also assign to myself
It looks like there is a size regression with drop unit dims on vector transfer + convert trivial shape_cast to no-op https://github.com/openxla/iree/pull/14220#issuecomment-1606270765 I'll investigate it further
The problem is at `VectorReductionToGPU`, the patterns in `mlir::vector::populatePropagateWarpVectorDistributionPatterns` can't handle the `vector.shape_cast`, which results in bad warp distribution: ```mlir ----- After VectorReduceToGPU ----- func.func @main_dispatch_84_generic_2x256_i8xi32() { %c0 = arith.constant...
The patch trying to fix vector distribution is sent out for review https://reviews.llvm.org/D154870
Unassigned myself as I'm not working on this now
Sure, I'll take a look this week
So my understanding of the TODO tasks are: 1. Currently we don't support DT for `[f32, i8, f32]`. But we should experiment with DT + codegen to see if a...
I have a prototype to promote matmul for available ukernel: #15873 I think one issue is currently we don't fuse `arith.sitofp` into ukernel dispatch. AFAIK there are two potential ways:...
Unassigned myself as I'm not actively working on this