| commit | 7f5857d84220d2e8cc56ce1980de1337739ea931 | [log] [tgz] |
|---|---|---|
| author | bjacob <benoitjacob@google.com> | Tue Jul 25 18:15:41 2023 -0400 |
| committer | GitHub <noreply@github.com> | Tue Jul 25 22:15:41 2023 +0000 |
| tree | 47c9b91369970d40098408f124ad2e4bfcbbbae9 | |
| parent | a5a6b67f9553d7f949206ccbb560fe37f8100f22 [diff] |
Ukernels: mmt4d paths for arm64 fp16 extensions (#14490) On the arm64 architecture, widely-available `fp16` and `fp16fml` extensions allow greatly speeding up the `f16f16f16` and `f16f16f32` cases respectively. Along the way, this renames `fullfp16` -> `fp16` --- we follow LLVM's nomenclature for ISA extensions and there is some confusion there as both strings exist in LLVM, but from actually exercising this, it appears that `fp16` is what we need here, and that conveniently allows dropping those 4 characters, `full`. Also along the way, reported and worked around https://github.com/llvm/llvm-project/issues/64104 Results from `mmt4d_benchmark`, Gflop/s, single thread. Results with these new code paths are **in bold** and are to be compared to the results immediately above them (same element types, not using the new code path). benchmark | Arm Cortex-X2 | Arm Cortex-A510 --- | --- | --- BM_mmt4d_f16f16f32_tile_8x8x1 | 47.1 | 9.5 BM_mmt4d_f16f16f32_tile_8x8x1_fp16fml | **92.0** | **12.1** BM_mmt4d_f16f16f16_tile_8x8x1 | 46.8 | 9.49 BM_mmt4d_f16f16f16_tile_8x8x1_fp16 | **178.5** | **21.3**
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.