| commit | 3459230b8838dfd4f425f12984c21ad373ef08b4 | [log] [tgz] |
|---|---|---|
| author | bjacob <benoitjacob@google.com> | Wed Jul 26 17:11:53 2023 -0400 |
| committer | GitHub <noreply@github.com> | Wed Jul 26 17:11:53 2023 -0400 |
| tree | bcb302d049965667bdf08e6411c10c0372da3259 | |
| parent | 749453c2c3d4fa28f4f1bfed0ba00890f6e87b7a [diff] |
Ukernels: mmt4d paths for the arm64 bf16 extension (#14495) This extension is present on recent cores (Arm Cortex-A510/A710/X2) and helps `bf16` matmuls accumulating into `f32` be up to 4x faster than `f32` matmuls thanks to not only the 2x smaller bit width, but also thanks to these new instructions being matrix multiplications. There are two instructions, `bfmmla` and `bfdot`. The `bfmmla` instruction is generally faster, and the Arm Reference Manual says it should always be faster, but on the A510 it's actually slower. As we are not currently into the business of writing code paths tuned for microarchitectures (beyond ISA capabilities), this PR just adds a `bfmmla` kernel, which performs great on A710 and X2 and is still a speedup over f32 on the A510 even though it isn't optimal there. Results from `mmt4d_benchmark` (GFlop/s, single-thread): LHS/RHS/accumulator types | Arm Cortex-A510 | Arm Cortex-A710 | Arm Cortex-X2 --- | --- | --- | --- `f32f32f32` | 12.1 | 38.9 | 91.5 `bf16bf16f32` | 13.3 | 152.4 | 314.4
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.