Support decomposition of torch.broadcast_tensors #4253
There are no files selected for viewing
This file contains hidden or 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 file contains hidden or 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 file contains hidden or 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 file contains hidden or 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 file contains hidden or 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 file contains hidden or 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 file contains hidden or 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
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -479,78 +479,117 @@ FailureOr<Value> Torch::unsqueezeTensor(PatternRewriter &rewriter, | |
| return unsqueezed; | ||
| } | ||
|
|
||
| // Checks whether the inputs are broadcast compatible or not. If | ||
| // yes, then computes the final broadcast shape. | ||
| void Torch::computeBroadcastShape(PatternRewriter &rewriter, Location loc, | ||
| SmallVector<Value> inputs, | ||
| SmallVector<int64_t> &resultShape, | ||
| SmallVector<Value> &resultShapeValue) { | ||
|
|
||
| SmallVector<SmallVector<int64_t>> shapes; | ||
| SmallVector<unsigned> ranks; | ||
|
|
||
| for (auto input : inputs) { | ||
| SmallVector<int64_t> shape{ | ||
| cast<BaseTensorType>(input.getType()).getSizes()}; | ||
| shapes.push_back(shape); | ||
| ranks.push_back(shape.size()); | ||
| } | ||
|
|
||
| unsigned maxRank = *std::max_element(ranks.begin(), ranks.end()); | ||
|
|
||
| // Check whether the shapes of the tensors are broadcastable or not. | ||
| // Two tensors are “broadcastable” if the following rules hold: | ||
| // 1.) Each tensor has at least one dimension. | ||
| // 2.) When iterating over the dimension sizes, starting at the trailing | ||
| // dimension, the dimension sizes must either be equal, one of them is 1, or | ||
| // one of them does not exist. | ||
| for (unsigned i = 0; i < maxRank; i++) { | ||
|
|
||
| SmallVector<Value> sizeInputs; | ||
| for (auto [idx, input] : llvm::enumerate(inputs)) { | ||
| int sizeDimIdx = ranks[idx] - i - 1; | ||
| if (sizeDimIdx >= 0) { | ||
| auto sizeDim = rewriter.create<Torch::ConstantIntOp>( | ||
| loc, rewriter.getI64IntegerAttr(sizeDimIdx)); | ||
| sizeInputs.push_back( | ||
| rewriter.createOrFold<AtenSizeIntOp>(loc, input, sizeDim)); | ||
| } | ||
| } | ||
|
|
||
| Value torchCstOne = rewriter.create<Torch::ConstantIntOp>( | ||
vinitdeodhar marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| loc, rewriter.getI64IntegerAttr(1)); | ||
| SmallVector<Value> predicates; | ||
| for (auto sizeVal : sizeInputs) { | ||
| Value cmpSizeEquals = | ||
| rewriter.create<Torch::AtenEqIntOp>(loc, sizeVal, sizeInputs.front()); | ||
vinitdeodhar marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| predicates.push_back(cmpSizeEquals); | ||
| Value cmpSizeEqualsOne = | ||
| rewriter.create<Torch::AtenEqIntOp>(loc, sizeVal, torchCstOne); | ||
| predicates.push_back(cmpSizeEqualsOne); | ||
| } | ||
|
|
||
| Value anyBoolOpList = rewriter.create<PrimListConstructOp>( | ||
| loc, Torch::ListType::get(predicates.front().getType()), predicates); | ||
| Value cmp = rewriter.create<Torch::AtenAnyBoolOp>(loc, anyBoolOpList); | ||
| rewriter.create<Torch::RuntimeAssertOp>( | ||
| loc, cmp, "tensors are not broadcast compatible"); | ||
| } | ||
|
|
||
| // If we reach here then it means both the shapes are broadcast compatible. | ||
vinitdeodhar marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| auto maxRankIdx = | ||
| std::max_element(ranks.begin(), ranks.end()) - ranks.begin(); | ||
vinitdeodhar marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| resultShape = shapes[maxRankIdx]; | ||
| Value shapeTensor = inputs[maxRankIdx]; | ||
|
|
||
| for (unsigned i = 0; i < resultShape.size(); i++) { | ||
| Value sizeDim = rewriter.create<Torch::ConstantIntOp>( | ||
| loc, rewriter.getI64IntegerAttr(i)); | ||
| resultShapeValue.push_back( | ||
| rewriter.createOrFold<AtenSizeIntOp>(loc, shapeTensor, sizeDim)); | ||
| } | ||
| unsigned resultRank = resultShape.size(); | ||
| for (unsigned i = 0; i < maxRank; i++) { | ||
|
|
||
| SmallVector<Value> sizeInputs; | ||
| for (auto [idx, input] : llvm::enumerate(inputs)) { | ||
vinitdeodhar marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| int sizeDimIdx = ranks[idx] - i - 1; | ||
| if (sizeDimIdx >= 0) { | ||
| auto sizeDim = rewriter.create<Torch::ConstantIntOp>( | ||
| loc, rewriter.getI64IntegerAttr(sizeDimIdx)); | ||
| sizeInputs.push_back( | ||
| rewriter.createOrFold<AtenSizeIntOp>(loc, input, sizeDim)); | ||
| } | ||
| } | ||
|
|
||
| // Compute shape value of broadcast result, | ||
| // which is the maximum of dimension sizes across all inputs | ||
| Value maxShapeVal = sizeInputs.front(); | ||
| for (auto sizeInput : sizeInputs) { | ||
| maxShapeVal = rewriter.create<PrimMaxIntOp>(loc, maxShapeVal, sizeInput); | ||
| } | ||
| resultShapeValue[resultRank - i - 1] = maxShapeVal; | ||
|
|
||
| // Compute result shape if all input shapes are known | ||
| bool unknownSize = false; | ||
| for (auto [idx, shape] : llvm::enumerate(shapes)) { | ||
| if (ranks[idx] - i - 1 < shape.size() && | ||
| shape[ranks[idx] - i - 1] == kUnknownSize) { | ||
|
There was a problem hiding this comment. This is a bit confusing. There was a problem hiding this comment. @zjgarvey , this check ensures that out of bounds access to shape is not performed in case input tensors have different ranks. The e2e test BroadcastTensorsModuleList_multiple_ranks covers this case. Let me know if you have a suggestion on making this check easier to read There was a problem hiding this comment. Maybe "confusing" isn't the correct word. I'm rather certain the check is wrong. E.g. For Whereas, |
||
| unknownSize = true; | ||
| } | ||
| } | ||
|
|
||
| if (unknownSize) { | ||
| resultShape[resultRank - i - 1] = kUnknownSize; | ||
| } else { | ||
|
|
||
| int64_t maxShape = 1; | ||
| for (auto [idx, shape] : llvm::enumerate(shapes)) { | ||
| if (ranks[idx] - i - 1 < shape.size()) { | ||
| maxShape = std::max(maxShape, shape[ranks[idx] - i - 1]); | ||
| } | ||
| } | ||
| resultShape[resultRank - i - 1] = maxShape; | ||
vinitdeodhar marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| } | ||
| } | ||
| } | ||
|
|
||
Oops, something went wrong.
Oops, something went wrong.
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.
Uh oh!
There was an error while loading. Please reload this page.