| commit | 2231682b5c2ad2d61e2991ae7327b0f55a0bcf54 | [log] [tgz] |
|---|---|---|
| author | Manish Gupta <manigupta@google.com> | Sun Jan 22 23:26:25 2023 -0800 |
| committer | GitHub <noreply@github.com> | Sun Jan 22 23:26:25 2023 -0800 |
| tree | e08b00e2e149e437475c3f207f76b05a9b280edc | |
| parent | 6c0bfc0226e15f75121783dc754a78a1e100cfaf [diff] |
Adds Native Tensor Core (F16) Support [mma.sync.16816.f16.f16 and ldmatrix] (#11817) Adds native Tensor Core (F16) support for NVIDIA GPUs. Specifically, adds support to target `mma.sync.aligned.m16n8k16.row.col.f16.f16.f16.f16` and `ldmatrix` instructions on NVIDIA A100 GPUs. For now, these instructions are not enabled by default and requires using (`--iree-codegen-llvmgpu-use-mma-sync=true`). IREE runs row-row matmul and the native Tensor Core support requires `ldmatrix` for operandA and `ldmatrix.trans` for operandB. `ldmatrix.trans` transposes the data while loading to use Tensor Core `mma*row.col*` instruction. There are LLVM/MLIR upstream pathes and previous IREE PRs that enable targeting native Tensor Core F16 support. A few LLVM/MLIR upstream patches: - [[mlir][NVGPU] Adding Support for cp_async_zfill via Inline Asm](https://reviews.llvm.org/D132269) - [[mlir][NVGPU] Handle Native mma.sync sizes and ldmatrix(x4) for matrixB](https://reviews.llvm.org/D135749) - [[mlir][NVGPU] Fix affine maps computing indices for LdMatrixOp srcMemref](https://reviews.llvm.org/D138978) - [[mlir][NVGPU] Fix NVGPU Bazel Layering](https://reviews.llvm.org/D141982) Previous IREE PRs: - [Support GEMM Pipelining *without* Epilogue Peeling](https://github.com/iree-org/iree/pull/10388) - [Add support for GEMM e2e Test For CUDA backend on F16 input](https://github.com/iree-org/iree/pull/10842)
IREE (Intermediate Representation Execution Environment, pronounced as “eerie”) is an MLIR-based end-to-end compiler and runtime that lowers Machine Learning (ML) models to a unified IR that scales up to meet the needs of the datacenter and down to satisfy the constraints and special considerations of mobile and edge deployments.
See our website for project details, user guides, and instructions on building from source.
IREE is still in its early phase. We have settled down on the overarching infrastructure and are actively improving various software components as well as project logistics. It is still quite far from ready for everyday use and is made available without any support at the moment. With that said, we welcome any kind of feedback on any communication channels!
See our website for more information.
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.