torch-mlir/docs/Torch-ops-E2E-implementatio...

51 lines
4.2 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters!

This file contains ambiguous Unicode characters that may be confused with others in your current locale. If your use case is intentional and legitimate, you can safely ignore this warning. Use the Escape button to highlight these characters.

## Tutorials
- Linalg.generic op introduction [here](https://www.youtube.com/watch?v=PdQPlPudT90)
- Basic E2E debugging walk-through [here](https://www.youtube.com/watch?v=E5epCJOtrf8&t=1556s)
## Example PR
https://github.com/llvm/torch-mlir/pull/294
## Major steps
### Step 1. Add an end-to-end test to iterate on
Add an end-to-end test to the [end-to-end test suite](https://github.com/llvm/torch-mlir/blob/main/docs/adding_an_e2e_test.md)). Ideally there is an existing file that your op fits into. If not, you can create a new file.
We generally recommend testing by invoking `torch.ops.aten.someop` from Python -- that gives a very precise test for the individual Torch operator you are implementing (calling `torch.ops.aten.someop` from Python always lowers into the MLIR `torch.aten.someop` operation)
The end-to-end test is important to check the correctness of the other steps.
### Step 2. Update ods
Update [torch_ods_gen.py](https://github.com/llvm/torch-mlir/blob/main/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py) with the new op and run [update_torch_ods.sh](https://github.com/llvm/torch-mlir/blob/main/build_tools/update_torch_ods.sh) to generate the ods. Running `update_torch_ods.sh` would dump all the operators with schema into `JITOperatorRegistryDump.txt`. Its convenient to look for ops signatures and operands names in this file.
### Step 3. Propagate types
Its essential to make sure the new op implements shape and dtype inference. See [abstract_interp_lib](https://github.com/llvm/torch-mlir/blob/main/docs/abstract_interp_lib.md) for information on adding shape and dtype inference.
### Step 4. Torch ops lowering
#### Decompose
If your op can be decomposed into other supported ops, then you can add a pattern into [DecomposeComplexOps](https://github.com/llvm/torch-mlir/blob/8d3ca887df5ac5126fa3fc2ec3546c6322a4d066/lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp#L1).
You can find example PRs [here](https://github.com/llvm/torch-mlir/pull/2550) and [here](https://github.com/llvm/torch-mlir/pull/2553).
#### Lower to Linalg
The `Torch` dialect needs to be lowered to [Linalg](https://mlir.llvm.org/docs/Dialects/Linalg/) dialect which can be used as input IR of backends. [Here](https://mlir.llvm.org/docs/Dialects/Linalg/#high-level-description-of-linalg-opsa-namelinalg_opsa) is a high level introduction about Linalg ops and [here](https://www.youtube.com/watch?v=PdQPlPudT90) is a video explaining `linalg.generic` op. The building block is the `linalg.generic` op which consists of indexing maps, iterator types, input/output tensors and a compute payload. You would want to get familiar with the concept of [affine map](https://mlir.llvm.org/docs/Dialects/Affine/#affine-expressions). The `linalg.generic` op anatomy [tutorial](https://www.youtube.com/watch?v=PdQPlPudT90&list=PLHPjgRtRcfTpVGFMrLP2KQyXhvtSQiiai&index=1) covers the basics of `linalg.generic` from a user's perspective.
You can find an [example PR here](https://github.com/llvm/torch-mlir/pull/294).
## Delivering Code
1. The codebase follows the [LLVMs coding conventions](https://llvm.org/docs/CodingStandards.html).The following items might be the most frequently used rules:
- [use-early-exits-and-continue-to-simplify-code](https://llvm.org/docs/CodingStandards.html#use-early-exits-and-continue-to-simplify-code)
- [don-t-use-else-after-a-return](https://llvm.org/docs/CodingStandards.html#don-t-use-else-after-a-return)
- [use-auto-type-deduction-to-make-code-more-readable](https://llvm.org/docs/CodingStandards.html#use-auto-type-deduction-to-make-code-more-readable)
- [anonymous-namespaces](https://llvm.org/docs/CodingStandards.html#anonymous-namespaces)
- [avoid-braces-on-simple-single-statement-bodies-of-if-else-loop-statements](https://llvm.org/docs/CodingStandards.html#don-t-use-braces-on-simple-single-statement-bodies-of-if-else-loop-statements)
2. Try to refactor and reuse existing code/helpers when working on RefineTypes and TorchToLinalg lowering for easier maintenance, testing and better readability. Try not to copy & paste existing code.
3. Squash all the commits into one, including the commits addressing review comments.
4. Use `git clang-format HEAD~1` to automatically format your commit.
5. Rebase on `HEAD` before delivering.