[Mlir-commits] [mlir] [MLIR][Vector] Add unrolling support for bitcast, interleave, and deinterleave ops (PR #194513)

[Jianhui Li]

================
@@ -1389,6 +1389,222 @@ struct UnrollShapeCastPattern : public OpRewritePattern<vector::ShapeCastOp> {
   vector::UnrollVectorOptions options;
 };
 
+// Unroll vector::BitCastOp into smaller tile-based bitcast operations.
+// Tiles the result vector into target shape chunks and bitcasts corresponding
+// source slices, accounting for element bitwidth ratios.
+// Example: bitcast v8f32 to v16f16 with target shape [4] unrolls into
+// multiple bitcast operations on 4-element tiles.
+struct UnrollBitCastPattern : public OpRewritePattern<vector::BitCastOp> {
+  UnrollBitCastPattern(MLIRContext *context,
+                       const vector::UnrollVectorOptions &options,
+                       PatternBenefit benefit = 1)
+      : OpRewritePattern<vector::BitCastOp>(context, benefit),
+        options(options) {}
+
+  LogicalResult matchAndRewrite(vector::BitCastOp bitCastOp,
+                                PatternRewriter &rewriter) const override {
+    auto targetShape = getTargetShape(options, bitCastOp);
+    if (!targetShape)
+      return failure();
----------------
Jianhui-Li wrote:

replaced. This fails when user doesn't implement the call-back function properly so sometimes it doesn't get the targetShape. 

https://github.com/llvm/llvm-project/pull/194513