Han-Chung Wang
Han-Chung Wang
> Speaking of memory footprint btw. I believe there was an issue and/or difference of the dt-fusion pipeline with the current default (CPU) pipeline in means of the weight encodings...
It is not necessary, but I think that it is good to assign it to you because you are working on it.
We are generating bad IRs (19K ops!) for the dispatch, so I'm not going to file an LLVM issue. I identified that there are at least two issues (https://github.com/iree-org/iree/issues/17593 and...
> Maybe right after materializing the layouts you can call the underlying implementation that folds unit dims. IIRC, the unit dims are introduced by tiling. I.e., it happens when we...
The performance issue is come from bad configurations on generic op. It is generating scalar codes. One of potential solutions is to set vector level tile sizes to zeros when...
With recent fixes, I'm able to codegen mmt4d fusion with and without ukernels. The lowering_config is not correct because the lowering_config is not set correctly on generic ops. The consumer...
Prioritize the issue and flesh out more details. The issue depends on https://github.com/iree-org/iree/issues/17718. We will need to support the codegen in the new data-tiling pipeline, because it enables fusion. Today,...
@pashu123 here is an example that you can start with: https://gist.github.com/hanhanW/075f4881664ce095d4af49a29842a6ba#file-z-mlir You can follow https://github.com/iree-org/iree/commit/3d23684f2510264e2793aeb49e53269fd168e4f3 to create a new strategy and add pipeline tests like [this](https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/LLVMGPU/test/ROCDL/pipeline_tile_and_fuse.mlir). In the example, presets...
> Phase 3: (Data Layout Propagation: To push packs and unpacks to the boundaries of the dispatch) The result of phase 3 is not what we're looking for in codegen...
> after materialization within dispatches, you will be propagating the packs and unpacks across dispatches. The materialization happens at codegen level. And the propagation also happens at codegen level. Below...