Cherry pick: Allow creation of TCP groups where an op has multiple uses #76
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
As titled. Here's the corresponding torch-mlir PR: https://github.robot.car/cruise/mla-robocomp-torch-mlir/pull/156. Upstream cleanup was done via these PRs: - [x] [cruise-automation/pull/12](cruise-automation#12) - [x] [cruise-automation/pull/13](cruise-automation#13) - [x] [cruise-automation/pull/14](cruise-automation#14)
…uise-automation#3) This is useful for ML ranker C7: ``` ERROR: [RoboCompiler] - RoboCompiler step 'RoboCompiler Optimize' failed with the following error: cruise/mlp/robotorch/project/trajectory_ranking/model_architectures/per_npc_ranker.py:712:0: error: failed to legalize operation 'torch.aten.index.Tensor_hacked_twin' that was explicitly marked illegal ``` TorchToTosa doesn't match the AtenIndexTensorHackedTwinOp due to [this](https://sourcegraph.robot.car/github.robot.car/cruise/mla-robocomp-torch-mlir/-/blob/lib/Conversion/TorchToTosa/TorchToTosa.cpp?L3788) constraint: ``` // Right now only support multiple indexes with same shape // TODO for different shape multiple indexes, add broadcast_to for small // shape for (auto indexShapeOneDim : indexesShape) { if (!llvm::equal(indexesShape[0], indexShapeOneDim)) { return rewriter.notifyMatchFailure( op, "unimplemented: Only support multi indexes with same shape"); } } ``` where we see multiple index tensors of different shapes: ``` %278 = torch.prim.ListConstruct %277, %273, %270, %205 : (!torch.vtensor<[1,1,1,1],si64>, !torch.vtensor<[24,1,1],si64>, !torch.vtensor<[2,1],si64>, !torch.vtensor<[40],si64>) -> !torch.list<vtensor> %279 = torch.aten.index.Tensor_hacked_twin %arg9, %278 : !torch.vtensor<[1,24,2,40],f32>, !torch.list<vtensor> -> !torch.vtensor<[1,24,2,40],f32> ``` We can dig into supporting this case in TorchToTosa eventually, however this is to unblock the immediate case for C7 where we plan to map this to a data movement kernel (like gather).
Merge cruise-automation#21 from upstream.
Move bottom up fuser declaration to header file No testing required since it's a minor restructuring --------- Co-authored-by: Muhammad Abubakar <muhammad.abubakar@getcruise.com>
…-automation#7) Move bottom up fuser declaration to header file No testing required since it's a minor restructuring Cherry-pick from cruise-automation@c4c94fb --------- Co-authored-by: Muhammad Abubakar <muhammad.abubakar@getcruise.com>
Bumps TCP to `c4c94fb25f2c3470839796025f40f0487b3b69b6`.
…ion#25) As titled, lower torch transposed convolution to a custom TCP op to avoid a mis-compilation in `TorchToTosa`. --------- Co-authored-by: Srinath Avadhanula <srinath.avadhanula@getcruise.com>
…ruise-automation#25) (cruise-automation#10) As titled, lower torch transposed convolution to a custom TCP op to avoid a mis-compilation in `TorchToTosa`. Cherry-pick from upstream: cruise-automation#25 --------- Co-authored-by: Srinath Avadhanula <srinath.avadhanula@getcruise.com>
Add a utility to aid in converting torch ops to `tcp.custom_op` --------- Co-authored-by: Srinath Avadhanula <srinath.avadhanula@getcruise.com>
…ruise-automation#12) Add a utility to aid in converting torch ops to `tcp.custom_op` Cherry picking cruise-automation#33 --------- Co-authored-by: Srinath Avadhanula <srinath.avadhanula@getcruise.com>
…cruise-automation#34) Conv 1D fails legalization in TorchToTosa, which I believe [only supports](https://sourcegraph.com/github.com/llvm/torch-mlir@faa4517e83d82348259165412d0744ba776360b3/-/blob/lib/Conversion/TorchToTosa/TorchToTosa.cpp?L1871) converting aten.convolution for the 2D case: ``` if (inputTy.getRank() != 4) return rewriter.notifyMatchFailure( op, "Unimplemented: only 2D convolutions supported"); ``` This makes it so we dynamically make just conv2d legal, so the remaining variants can convert to `tcp.custom_op`.
Updates mlir-tcp to `90768ec2801ed9959144e8a1ca800e34ee2c7f54` resolves some minor merge conflicts around testing changes in [this PR](cruise-automation#28).
[LLVM](https://github.com/llvm/llvm-project/tree/70eb0e37a86747f9266e4c8380baa89746f5e23b) (Jan 29) [StableHlo](https://github.com/openxla/stablehlo/tree/fd52182f76cadb82f2064fe5fc49a4fb4347a826) (Jan 31) [Torch-MLIR](https://github.com/llvm/torch-mlir/tree/c7d7d7f00494b588c31ac617e91354b12709009d) (Jan 31) [TCP](https://github.com/cruise-automation/mlir-tcp/tree/b0d474e30f04362c54cf29b9bdeb5bd697e7b2dc) (Feb 1)
) Cherry-pick from upstream PR: cruise-automation#39
[LLVM](https://github.com/llvm/llvm-project/tree/e5ed7b6e2fd368b722b6359556cd0125881e7638) (Feb 27) [StableHlo](https://github.com/openxla/stablehlo/tree/d214e2e2bd568a80a9f25515907b64dd7b98c538) (Mar 5) [Torch-MLIR](https://github.com/llvm/torch-mlir/tree/a86e89ecb5c7929a39a38743fb7cacadf1ff41bb) (Mar 4) [TCP](https://github.com/cruise-automation/mlir-tcp/tree/8d5d10e91e1afc9a5bb88ac647de4929fa31d735) (Mar 5) Disable the python related tests that use the hermetic python deps
Replace the tcp custom op for `torch.tensorrt.execute_engine_variadic` to `tensorrt.execute_engine` for generic usage purpose. To test, from c/c repo: `bazel test @mlir-tcp//test/...`
[LLVM](https://github.com/llvm/llvm-project/tree/a952c123880eb1168f1021b116485e27170d48ca) (April 11) [Torch-MLIR](https://github.com/llvm/torch-mlir/tree/45eaeaaf3640f1f7f7b4743cf5844c3b9d519fdf) (April 12) [StableHlo](https://github.com/openxla/stablehlo/tree/271e8634de184fbfafd677d3876170feb6d08c97) (March 28) [TCP](https://github.com/cruise-automation/mlir-tcp/tree/04f327c698dd6f588ce09d8d9c99533c1e9855f9) (April 12)
Support lowering of the index_from_tensor op Part of: https://github.robot.car/cruise/cruise/pull/240509
Picks the dynamic legality fixes to TorchToTcp amidst other changes.
Add missing file from previous PR
Create tensor from index array handle
Remove `LRDPrefilterPrediction` related content Test from c/c: `bazel test @mlir-tcp//test/...`
…tomation#74) We previously only allowed an op to have a single use during the creation a group for it. This PR relaxes that to allow multiple uses as long as all the uses belong to the same region. --------- Co-authored-by: Srinath Avadhanula <srinath.avadhanula@getcruise.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Cherry picking #74
We previously only allowed an op to have a single use during the creation a group for it. This PR relaxes that to allow multiple uses as long as all the uses belong to the same region.