[Mlir-commits] [mlir] [mlir][linalg] Allow extra DPS inputs in convolution dim inference (PR #198462)

[llvmlistbot at llvm.org]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir

Author: Ahmad Tameem (Tameem-10xE)

<details>
<summary>Changes</summary>

As pointed out by @<!-- -->hanhanW in https://github.com/iree-org/iree/pull/24016#issuecomment-4444591355, this PR fixes a mismatch between `inferConvolutionDims` and the `LinalgConvolutionOpInterface` verifier. This is needed for quantized convolution ops, which can have extra scalar zero-point operands in addition to the input and filter operands.

---
Full diff: https://github.com/llvm/llvm-project/pull/198462.diff


1 Files Affected:

- (modified) mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp (+1-1) 


``````````diff
diff --git a/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp b/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp
index 2ba77cea8f16e..b761782541ff3 100644
--- a/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp
+++ b/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp
@@ -884,7 +884,7 @@ inferConvolutionDimsImpl(LinalgOp linalgOp,
 /// Returns a failure if `output_image` (and implicitly `filter_loop`) is empty.
 FailureOr<ConvolutionDimensions>
 mlir::linalg::inferConvolutionDims(LinalgOp linalgOp) {
-  if (linalgOp.getNumDpsInits() != 1 || linalgOp.getNumDpsInputs() != 2)
+  if (linalgOp.getNumDpsInits() != 1 || linalgOp.getNumDpsInputs() < 2)
     return failure();
 
   auto indexingMaps = linalgOp.getIndexingMapsArray();

``````````

</details>


https://github.com/llvm/llvm-project/pull/198462