[mlir][xegpu] Patch dynamic descriptor creation

[akroviakov]

This brief PR fixes the bug in XeGPU's CreateNdDescOp tensor creation with dynamic offset and strides.

[llvmbot]

@llvm/pr-subscribers-mlir

Author: Artem Kroviakov (akroviakov)

Changes

This brief PR fixes the bug in XeGPU's CreateNdDescOp tensor creation with dynamic offset and strides.

---

Full diff: https://github.com/llvm/llvm-project/pull/93580.diff

1 Files Affected:

• (modified) mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp (+1-1)

diff --git a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
index 22959224d56c2..37904f55cc23a 100644
--- a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
+++ b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
@@ -110,7 +110,7 @@ void CreateNdDescOp::build(OpBuilder &builder, OperationState &state,
 
   dispatchIndexOpFoldResults(offsets, dynamicOffsets, staticOffsets);
   dispatchIndexOpFoldResults(shape, dynamicShape, staticShape);
-  dispatchIndexOpFoldResults(strides, dynamicStrides, staticOffsets);
+  dispatchIndexOpFoldResults(strides, dynamicStrides, dynamicOffsets);
 
   auto staticOffsetsAttr = builder.getDenseI64ArrayAttr(staticOffsets);
   auto staticShapeAttr = builder.getDenseI64ArrayAttr(staticShape);

[llvmbot]

@llvm/pr-subscribers-mlir-gpu

Author: Artem Kroviakov (akroviakov)

Changes

This brief PR fixes the bug in XeGPU's CreateNdDescOp tensor creation with dynamic offset and strides.

---

Full diff: https://github.com/llvm/llvm-project/pull/93580.diff

1 Files Affected:

• (modified) mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp (+1-1)

diff --git a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
index 22959224d56c2..37904f55cc23a 100644
--- a/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
+++ b/mlir/lib/Dialect/XeGPU/IR/XeGPUOps.cpp
@@ -110,7 +110,7 @@ void CreateNdDescOp::build(OpBuilder &builder, OperationState &state,
 
   dispatchIndexOpFoldResults(offsets, dynamicOffsets, staticOffsets);
   dispatchIndexOpFoldResults(shape, dynamicShape, staticShape);
-  dispatchIndexOpFoldResults(strides, dynamicStrides, staticOffsets);
+  dispatchIndexOpFoldResults(strides, dynamicStrides, dynamicOffsets);
 
   auto staticOffsetsAttr = builder.getDenseI64ArrayAttr(staticOffsets);
   auto staticShapeAttr = builder.getDenseI64ArrayAttr(staticShape);

[Garra1980]

cc @chencha3

[adam-smnk]

Thanks a lot for the fix!

I guess XeGPU op verifier is too simple.
@chencha3 we should probably look into that in a separate PR

[chencha3]

Thanks a lot for the fix!

I guess XeGPU op verifier is too simple. @chencha3 we should probably look into that in a separate PR

Yes, you are right. Currently some parameters of XeGPU ops are architecture specific, and we left them checked by an arch class in downstream. We need to rethink this part.

[chencha3]

LGTM