| commit | 90f29a66d5bbd58167d84b2011d27c7ffb9a1ee1 | [log] [tgz] |
|---|---|---|
| author | Lei Zhang <antiagainst@gmail.com> | Wed Jun 19 19:38:33 2024 -0700 |
| committer | GitHub <noreply@github.com> | Wed Jun 19 19:38:33 2024 -0700 |
| tree | 3f89493f4ffb9d308cb6396ce502c3c36d82ad47 | |
| parent | 7c410492bdcbfdd4972de827d995bfb8482841a5 [diff] |
Reland "[spirv] Switch to use common target description" (#17699) This relands https://github.com/iree-org/iree/pull/17623. This commit switches SPIR-V side to use the common `#iree_gpu.target` to describe the GPU characteristics. With it we can now remove the ad-hoc Vulkan attributes and dialects and unify how GPU are described across various GPU compiler backends in IREE. SPIR-V has some additional requirements that we need to account for: We have many vendors and APIs to handle there so this commit adds various AMD/ARM/NVIDIA/Qualcomm targets for development purposes so that we can specify them with a shorthand. In order to be extensible, leverage the `feature` field in `#iree_gpu.target` to specify additional capabilities with `cap:` prefix and extensions with `ext:` prefix. We also use the `feature` field to specify what SPIR-V version to target with the `spirv:v1.x` format. Right now the `SPIRVConvertGPUTarget` pass is invoked immediately before configuration. This is to stage the changes. As a next step we need to move it immediately before `ConvertToSPIRV` pass. `--iree-vulkan-target-env` is dropped given now we removed the whole Vulkan dialect and cannot control with a `#vk.target_env` attribute anymore. The default `--iree-vulkan-target-triple` now becomes `vp_android_baseline_2022`, which is a a good lowest common denominator to guarantee the generated SPIR-V is widely accepted. We are not considering SwiftShader now anymore like previously due to testing purposes. The `--iree-vulkan-target-triple` should be renamed given it's not a triple anymore--that will happen later together with other GPU backends (i.e., cuda/hip) to be consistent. In order to support cooperative matrix conversion, we added `WMMA_F16_16x16x16_F16`. For NVIDIA GPUs we are abusing it right now without considering the concrete explicit layout--that is fine given in Vulkan they are opaque anyway. But this need to be fixed if we are targeting WMMA in CUDA. We now contruct a `#iree_gpu.target` to specify the target to drive SPIR-V CodeGen. Progress towards https://github.com/iree-org/iree/issues/16341 ci-extra: test_nvidia_gpu,test_nvidia_a100,test_amd_mi250, build_test_all_macos_arm64,build_and_test_android,test_on_moto-edge-x30 --------- Signed-off-by: Lei Zhang <antiagainst@gmail.com>
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.
Community meeting recordings: IREE YouTube channel
IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.