[Mlir-commits] [mlir] [mlir] Add direct vectorization lowering for `tensor.pack` ops (PR #78660)

[Diego Caballero]

================
@@ -1393,6 +1400,182 @@ vectorizeAsLinalgGeneric(RewriterBase &rewriter, VectorizationState &state,
   return success();
 }
 
+/// Given a tensor::PackOp, return the permutation from the "tiled"
+/// shape to the "packed" shape, defined as the following:
+/// The "packed" shape is the same as the `dest` shape of the pack op.
+/// The "tiled" shape is a permutation of the `dest` shape such that
+/// each outer dimension is in the original `source` order, and the
+/// inner_tile dimensions immediately follow their corresponding outer
+/// dimension.
+/// i.e. for the following tensor.pack:
+/// ```mlir
+/// %pack = tensor.pack %0 padding_value(%1)
+///   outer_dims_perm = [0, 2, 1]
+///   inner_dims_pos = [2, 1]
+///   inner_tiles = [16, 2]
+///   into %2 : tensor<32x8x16> -> tensor<32x1x4x16x2>
+/// ```
+/// The "packed" shape is `32x1x4x16x2`
+/// The "tiled" shape is `32x(4x2)x(1x16)`
+static SmallVector<int64_t>
+getTiledShapeToPackedShapePerm(tensor::PackOp packOp) {
+  auto innerTiles = packOp.getInnerTiles();
+  int64_t srcRank = packOp.getSourceRank();
+  auto innerDimsPos = packOp.getInnerDimsPos();
+  if (innerDimsPos.empty())
+    innerDimsPos = to_vector(llvm::seq<int64_t>(innerTiles.size()));
+  auto outerDimsPerm = packOp.getOuterDimsPerm();
+  if (outerDimsPerm.empty())
+    outerDimsPerm = to_vector(llvm::seq<int64_t>(srcRank));
+  auto packedIdxToTiledIdx = [&](int64_t idx) -> int64_t {
+    int64_t srcIdx;
+    if (idx >= srcRank)
+      srcIdx = innerDimsPos[idx - srcRank];
+    else
+      srcIdx = outerDimsPerm[idx];
+    int64_t tiledIdx = srcIdx;
+    for (int64_t pos : innerDimsPos)
+      if (pos < srcIdx)
+        tiledIdx++;
+    if (idx >= srcRank)
+      tiledIdx++;
+    return tiledIdx;
+  };
+  SmallVector<int64_t> perm;
+  for (size_t i = 0; i < packOp.getDestRank(); i++)
+    perm.push_back(packedIdxToTiledIdx(i));
+  return perm;
+}
+
+/// Given a tensor::PackOp, return the "tiled" `dest` shape as described
+/// above in `getTiledShapeToPackedShapePerm`.
+static SmallVector<int64_t> getTiledPackShape(tensor::PackOp packOp,
+                                              ArrayRef<int64_t> destShape) {
+  auto perm = getTiledShapeToPackedShapePerm(packOp);
+  return applyPermutation(destShape, invertPermutationVector(perm));
+}
+
+/// Create a masked TransferReadOp from `source` with shape `readShape`.
----------------
dcaballe wrote:

Can we refactor/generalize createTransferRead and createTransferWrite so that we can use it in other place where we are doing the same?

https://github.com/llvm/llvm-project/pull/78660