| commit | b71aa530ffbf7a00a0cac65dbbed4c230444e582 | [log] [tgz] |
|---|---|---|
| author | Ben Vanik <ben.vanik@gmail.com> | Wed Nov 30 15:50:36 2022 -0800 |
| committer | Ben Vanik <ben.vanik@gmail.com> | Mon Dec 05 22:01:17 2022 -0800 |
| tree | 47246e46e3e6f399455a66878742b16319ec9de0 | |
| parent | aca8cc69cf0d0195622afede602165fe67305224 [diff] |
Adds support for HAL executable object linkage. Objects are passed through flow/stream and handled by the HAL infrastructure during interface materialization (where we create the variants based on target configuration). Each backend then gets the objects specified for it and can use those in backend-dependent ways. For now the LLVM-CPU backend only supports external function calls useful for microkernels and such. This allows for a majority of IREE's features when defining flow/stream executables that call out to externs (binding/operand packing/optimization, inlining, linking, and automatic multi-targeting). In the future support can be added for generating the boilerplate for external device functions called all the way from (annotated) source inputs. The GPU backends (CUDA/Vulkan SPIR-V) currently only support entire top-level function definition (CUDA kernels or SPIR-V compute shaders). In the future support can be added for linking (PTX linking or spirv-link) to enable the microkernel-style substitution of ops that supports fusion and interface optimization. Objects can be embedded as data or referenced by file path allowing for both JIT and precompilation approaches (a codegen backend could take some input IR, produce via an external tool some objects, and then rewrite the IR to reference those objects). Because paths are hard the `--iree-hal-executable-object-search-path=` flag can be used (repeatedly) to add search paths. When coming from frontends it's probably best to rely on embedding.
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.