[mlir][vector] Add n-d deinterleave lowering

This patch implements the lowering for vector
deinterleave for vector of n-dimensions. Process
involves unrolling the n-d vector to a series
of one-dimensional vectors. The deinterleave
operation is then used on these vectors.

From:
```
%0, %1 = vector.deinterleave %a : vector<2x[4]xi8> -> vector<2x[2]xi8>
```

To:
```
%2 = llvm.extractvalue %0[0] : !llvm.array<2 x vector<8xf32>>
%3 = llvm.mlir.poison : vector<8xf32>
%4 = llvm.shufflevector %2, %3 [0, 2, 4, 6] : vector<8xf32>
%5 = llvm.shufflevector %2, %3 [1, 3, 5, 7] : vector<8xf32>
%6 = llvm.insertvalue %4, %1[0] : !llvm.array<2 x vector<4xf32>>
%7 = llvm.insertvalue %5, %1[0] : !llvm.array<2 x vector<4xf32>>
%8 = llvm.extractvalue %0[1] : !llvm.array<2 x vector<8xf32>>
...etc.
```

Author: mub-at-arm

mlir/lib/Dialect/Vector/Transforms/LowerVectorInterleave.cpp

@@ -79,6 +79,67 @@ class UnrollInterleaveOp final : public OpRewritePattern<vector::InterleaveOp> {
   int64_t targetRank = 1;
 };
 
+/// A one-shot unrolling of vector.deinterleave to the `targetRank`.
+///
+/// Example:
+///
+/// ```mlir
+/// vector.deinterleave %a : vector<1x2x3x8xi64> -> vector<1x2x3x4xi64>
+/// ```
+/// Would be unrolled to:
+/// ```mlir
+/// %result = arith.constant dense<0> : vector<1x2x3x8xi64>
+/// %0 = vector.extract %a[0, 0, 0]                 ─┐
+///        : vector<4xi64> from vector<1x2x3x4xi64>  |  | - Repeated 6x for
+/// %1, %2 = vector.deinterleave %0 :                |      all leading
+/// positions
+///        : vector<8xi64> -> vector<4xi64>          |
+/// %3 = vector.insert %1, %result [0, 0, 0]         |
+///        : vector<4xi64> into vector<1x2x3x4xi64>  |
+/// %3 = vector.insert %2, %result [0, 0, 0]         |
+///        : vector<4xi64> into vector<1x2x3x4xi64>  ┘
+/// ```
+///
+/// Note: If any leading dimension before the `targetRank` is scalable the
+/// unrolling will stop before the scalable dimension.
+
+class UnrollDeinterleaveOp final
+    : public OpRewritePattern<vector::DeinterleaveOp> {
+public:
+  UnrollDeinterleaveOp(int64_t targetRank, MLIRContext *context,
+                       PatternBenefit benefit = 1)
+      : OpRewritePattern(context, benefit), targetRank(targetRank) {};
+
+  LogicalResult matchAndRewrite(vector::DeinterleaveOp op,
+                                PatternRewriter &rewriter) const override {
+    VectorType resultType = op.getResultVectorType();
+    auto unrollIterator = vector::createUnrollIterator(resultType, targetRank);
+    if (!unrollIterator)
+      return failure();
+
+    auto loc = op.getLoc();
+    Value evenResult = rewriter.create<arith::ConstantOp>(
+        loc, resultType, rewriter.getZeroAttr(resultType));
+    Value oddResult = rewriter.create<arith::ConstantOp>(
+        loc, resultType, rewriter.getZeroAttr(resultType));
+
+    for (auto position : *unrollIterator) {
+      auto extractSrc =
+          rewriter.create<vector::ExtractOp>(loc, op.getSource(), position);
+      auto deinterleave =
+          rewriter.create<vector::DeinterleaveOp>(loc, extractSrc);
+      evenResult = rewriter.create<vector::InsertOp>(
+          loc, deinterleave.getRes1(), evenResult, position);
+      oddResult = rewriter.create<vector::InsertOp>(loc, deinterleave.getRes2(),
+                                                    oddResult, position);
+    }
+    rewriter.replaceOp(op, ValueRange{evenResult, oddResult});
+    return success();
+  }
+
+private:
+  int64_t targetRank = 1;
+};
 /// Rewrite vector.interleave op into an equivalent vector.shuffle op, when
 /// applicable: `sourceType` must be 1D and non-scalable.
 ///
@@ -116,7 +177,8 @@ struct InterleaveToShuffle final : OpRewritePattern<vector::InterleaveOp> {
 
 void mlir::vector::populateVectorInterleaveLoweringPatterns(
     RewritePatternSet &patterns, int64_t targetRank, PatternBenefit benefit) {
-  patterns.add<UnrollInterleaveOp>(targetRank, patterns.getContext(), benefit);
+  patterns.add<UnrollInterleaveOp, UnrollDeinterleaveOp>(
+      targetRank, patterns.getContext(), benefit);
 }
 
 void mlir::vector::populateVectorInterleaveToShufflePatterns(

mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir

@@ -2565,6 +2565,22 @@ func.func @vector_deinterleave_1d_scalable(%a: vector<[4]xi32>) -> (vector<[2]xi
     return %0, %1 : vector<[2]xi32>, vector<[2]xi32>
 }
 
+// CHECK-LABEL: @vector_deinterleave_2d
+// CHECK-SAME: %[[SRC:.*]]: vector<2x8xf32>) -> (vector<2x4xf32>, vector<2x4xf32>)
+func.func @vector_deinterleave_2d(%a: vector<2x8xf32>) -> (vector<2x4xf32>, vector<2x4xf32>) {
+  // CHECK: llvm.shufflevector
+  // CHECK-NOT: vector.deinterleave %{{.*}} : vector<2x8xf32>
+  %0, %1 = vector.deinterleave %a : vector<2x8xf32> -> vector<2x4xf32>
+  return %0, %1 : vector<2x4xf32>, vector<2x4xf32>
+}
+
+func.func @vector_deinterleave_2d_scalable(%a: vector<2x[8]xf32>) -> (vector<2x[4]xf32>, vector<2x[4]xf32>) {
+    // CHECK: llvm.intr.vector.deinterleave2
+    // CHECK-NOT: vector.deinterleave %{{.*}} : vector<2x[8]xf32>
+    %0, %1 = vector.deinterleave %a : vector<2x[8]xf32> -> vector<2x[4]xf32>
+    return %0, %1 : vector<2x[4]xf32>, vector<2x[4]xf32>
+}
+
 // -----
 
 // CHECK-LABEL: func.func @vector_bitcast_2d