[Mlir-commits] [mlir] [MLIR][XeGPU] Support Layout propagation for interleave and deintereleave op (PR #194966)

[Artem Kroviakov]

================
@@ -458,6 +458,85 @@ xegpu::inferBitCastSourceLayout(xegpu::DistributeLayoutAttr resLayout,
   return finalSrcLayout;
 }
 
+/// Infers the source layout attribute for an interleave operation given the
+/// result layout attribute. Interleave doubles the size of the innermost
+/// dimension, so the layout inference is similar to bitcast where the source
+/// element type is larger than the result element type (ratio = 2).
+xegpu::DistributeLayoutAttr
+xegpu::inferInterleaveSourceLayout(xegpu::DistributeLayoutAttr resLayout) {
+
+  SmallVector<int64_t> sgData = resLayout.getEffectiveSgDataAsInt();
+  SmallVector<int64_t> instData = resLayout.getEffectiveInstDataAsInt();
+  SmallVector<int64_t> laneData = resLayout.getEffectiveLaneDataAsInt();
+  size_t sgDataSize = sgData.size();
+  size_t instDataSize = instData.size();
+  size_t laneDataSize = laneData.size();
+  int64_t sgDataValue = -1;
+  int64_t instDataValue = -1;
+  int64_t laneDataValue = -1;
+  int64_t dim = resLayout.getRank() - 1;
+
+  // Interleave doubles the innermost dimension, so we need to halve the
+  // layout values (similar to bitcast with ratio = 2)
+  int ratio = 2;
----------------
akroviakov wrote:

```suggestion
  constexpr int ratio = 2;
```

https://github.com/llvm/llvm-project/pull/194966