Kunwar Grover comments

Results 36 comments of


                                            Kunwar Grover

[VectorDistribution] Plumb the VectorDistribute pipeline to support reduction operations (3/4)

This is looking much nicer with the vector.contract distribution. Can you move that to a seperate patch if possible?

[SPIR-V] Shared memory limit exceeded for batch matmul dispatch

AFAIU my patch should also take care of SPIRV

[SPIR-V] Shared memory limit exceeded for batch matmul dispatch

Sure, will do tomorrow

[SPIR-V] Shared memory limit exceeded for batch matmul dispatch

> > AFAIU my patch should also take care of SPIRV > > Checked out the #16927 PR and still see the same error @Groverkss I'll have a look and...

[SPIR-V] Shared memory limit exceeded for batch matmul dispatch

@kuhar I looked into this more. Currently, the cooperative matrix path is reusing the MMA heuristics deduction which my patch adds the shared memory check on. I will send a...

[GPU] Clustered Subgroup Reduction

> I'm thinking of adding this part to the description of `gpu.subgroup_reduce`: > > > If a `cluster_size` is provided, the subgroup is divided into clusters of `cluster_size` lanes each,...

[LLVMGPU] VectorDistribution pipeline for attention

Please only review the last commit. Other commits are the patches this patch depends on.

[LLVMGPU] VectorDistribution pipeline for attention

Closing in favour of https://github.com/iree-org/iree/pull/17773

[VectorDistribution] Add support for multi-subgroup attention

> Is there a way you can avoid maybe decomposing attention operation until vector distribution and handle the layout distribution for attention directly? Probably not… if we do that, we...

[VectorDistribution] Add support for multi-subgroup attention

There are some tests that exceed shared memory, so i'm going to wait for https://github.com/iree-org/iree/pull/18415 to land before i land this.