[Mlir-commits] [mlir] [mlir][spirv] Add vector.interleave to spirv.VectorShuffle conversion (PR #93240)

[Jakub Kuderski]

================
@@ -578,6 +578,42 @@ struct VectorShuffleOpConvert final
   }
 };
 
+struct VectorInterleaveOpConvert final
+    : public OpConversionPattern<vector::InterleaveOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(vector::InterleaveOp interleaveOp, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    // Check the source vector type
+    auto sourceType = interleaveOp.getSourceVectorType();
+    if (sourceType.getRank() != 1 || sourceType.isScalable()) {
+      return rewriter.notifyMatchFailure(interleaveOp,
+                                         "unsupported source vector type");
+    }
+
+    // Check the result vector type
+    auto oldResultType = interleaveOp.getResultVectorType();
+    Type newResultType = getTypeConverter()->convertType(oldResultType);
+    if (!newResultType)
+      return rewriter.notifyMatchFailure(interleaveOp,
+                                         "unsupported result vector type");
+
+    // Interleave the indices
+    int n = sourceType.getNumElements();
+    auto seq = llvm::seq<int64_t>(2 * n);
+    auto indices = llvm::to_vector(
+        llvm::map_range(seq, [n](int i) { return (i % 2 ? n : 0) + i / 2; }));
+
+    // Emit a SPIR-V shuffle.
+    rewriter.replaceOpWithNewOp<spirv::VectorShuffleOp>(
+        interleaveOp, newResultType, adaptor.getLhs(), adaptor.getRhs(),
----------------
kuhar wrote:

In SPIR-V, input vectors of size 1 (`n == 1`) are converted to scalars by the type converter. We should add a test with interleave of `vector<1xf32> -> vector<2xf32>`. In that case, we can't use `VectorShuffleOp` directly when the input types are scalars. When this is the case, we can emit `spirv::CompositeConstructOp`. You can check the vector shuffle pattern, IIRC we do something similar there.

https://github.com/llvm/llvm-project/pull/93240