Support qdq decomposition of DotGeneralOp and ConvolutionOp #2461
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
This was referenced Jul 24, 2024
GleasonK
approved these changes
Jul 24, 2024
3963618 to
8ba3db8
Compare
sdasgup3
commented
Jul 25, 2024
abhigunj
approved these changes
Jul 25, 2024
sdasgup3
added a commit
that referenced
this pull request
Jul 26, 2024
[ParentPR](#2459) Previously while creating the QDQ pattern we use `create` API without using the result type and hence reply on the type inference to derive the result type. That works for most of the element-wise operations, however, for dot_general and convolution the result type can might be infeasible to infer in the presense of input quantize types. The PR fixes that. Note to the reviewers: To may just focus on the very last commit of the chain. The rest is coming from parent PR. [ChildPR](#2461)
8ba3db8 to
8d39480
Compare
sdasgup3
added a commit
that referenced
this pull request
Jul 26, 2024
…2462) [ParentPR](#2461) `quant-to-math` pass [assumes](https://github.com/openxla/stablehlo/blob/eba821aa1c54a21d70331d7926dfc8b929f988f3/stablehlo/transforms/StablehloLegalizeQuantToInt.cpp#L984) that the return type of quantized dot_general or convolution is always having `i32` as storage type. With that the following program with the result storage type of `i8`fail to materialize all the intermediate converted values. ``` func.func @dot_general_per_tensor_quantization(%arg0: tensor<2x3x4x!quant.uniform<i8:f32, 1.0:17>>, %arg1: tensor<2x3x5x!quant.uniform<i8:f32, 1.0:0>>) -> tensor<2x4x5x!quant.uniform<i8:f32, 1.0:17>> { // expected-error@+1 {{failed to legalize operation 'stablehlo.dot_general' that was explicitly marked illegal}} %0 = "stablehlo.dot_general"(%arg0, %arg1) { dot_dimension_numbers = #stablehlo.dot< lhs_batching_dimensions = [0], rhs_batching_dimensions = [0], lhs_contracting_dimensions = [1], rhs_contracting_dimensions = [1] > } : (tensor<2x3x4x!quant.uniform<i8:f32, 1.0:17>>, tensor<2x3x5x!quant.uniform<i8:f32, 1.0:0>>) -> tensor<2x4x5x!quant.uniform<i8:f32, 1.0:17>> func.return %0 : tensor<2x4x5x!quant.uniform<i8:f32, 1.0:17>> } ``` One option to fix this to provode source/target materialization [link](https://mlir.llvm.org/docs/DialectConversion/#type-converter), but we found that for other Ops [e.g.](https://github.com/openxla/stablehlo/blob/eba821aa1c54a21d70331d7926dfc8b929f988f3/stablehlo/transforms/StablehloLegalizeQuantToInt.cpp#L510), there is a precedent on how to convert the math computed in `i32` is converted back to result type. The PR implements the missing conversion. Note to the reviewers: To may just focus on the very last commit of the chain. The rest is coming from parent PR.
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.
ParentPR
The PR
DotGeneralOpandConvolutionOp.Note to the reviewers: To may just focus on the very last commit of the chain. The rest is coming from parent PR.
childPR