[mlir][linalg] Enable scalable vectorization of linalg.unpack (WIP)

This patch updates `vectorizeAsTensorUnpackOp` to support scalable
vectorization by requiring user-specified vector sizes for both the
_read_ and _write_ operations involved in `linalg.unpack`. Detailed
rationale and an example are provided below.

Conceptually, `linalg.unpack` consists of the following high-level steps:
  1. _Read_ from the source tensor.
  2. Transpose the value read in step (1).
  3. _Write_ the value from step (2) into the destination tensor.

Currently, when vectorizing with user-provided vector sizes, only the
sizes for the _write_ operation (step 3) are required. Sizes for the
_read_ operation (step 1) are inferred from static shapes and inner tile
sizes.

This logic breaks when the input shapes or tile sizes are dynamic
(indeed, `vectorizeUnPackOpPrecondition` rejects such cases ATM and the
vectorization fails). This patch addresses the issue by requiring
explicit vector sizes for both the read and write sides, enabling
scalable vectorization in such cases.

Example:

```mlir
func.func @unpack(%in: tensor<1x1x8x?xf32>, %out: tensor<8x?xf32>) -> tensor<8x?xf32> {
  %vs = vector.vscale
  %c8 = arith.constant 8 : index
  %tile_size = arith.muli %vs, %c8 : index

  %unpack = linalg.unpack  %in
    inner_dims_pos = [0, 1]
    inner_tiles = [8, %tile_size]
    into %out : tensor<1x1x8x?xf32> -> tensor<8x?xf32>
  return %unpack : tensor<8x?xf32>
}

module attributes {transform.with_named_sequence} {
  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
    %0 = transform.structured.match ops{["linalg.unpack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
    transform.structured.vectorize %0 vector_sizes [1, 1, 8, [8],  8, [8]] : !transform.any_op
    //                                              \         /    \    /
    //                                              read-sizes   write-sizes
    transform.yield
  }
}
```

Finally, this patch also extends `createReadOrMaskedRead` and
`createWriteOrMaskedWrite` to take scalable flags.

Author: banach-space

mlir/include/mlir/Dialect/Vector/Utils/VectorUtils.h

@@ -228,7 +228,7 @@ bool isLinearizableVector(VectorType type);
 Value createReadOrMaskedRead(OpBuilder &builder, Location loc, Value source,
                              ArrayRef<int64_t> inputVectorSizes, Value padValue,
                              bool useInBoundsInsteadOfMasking = false,
-                             ArrayRef<bool> scalableDims = {});
+                             ArrayRef<bool> inputScalableVecDims = {});
 
 /// Returns success if `inputVectorSizes` is a valid masking configuraion for
 /// given `shape`, i.e., it meets:

mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp

@@ -1805,7 +1805,8 @@ vectorizeAsTensorPackOp(RewriterBase &rewriter, linalg::PackOp packOp,
     inputShape[innerDimsPos[idx]] *= size;
   auto maskedRead = vector::createReadOrMaskedRead(
       rewriter, loc, packOp.getSource(), inputShape, padValue,
-      useInBoundsInsteadOfMasking);
+      useInBoundsInsteadOfMasking,
+      /*inputScalableVecSizes=*/{});
 
   // Create ShapeCastOp.
   SmallVector<int64_t> destShape(inputVectorSizes);
@@ -1840,6 +1841,10 @@ vectorizeAsTensorPackOp(RewriterBase &rewriter, linalg::PackOp packOp,
 ///
 /// When collapsing scalable flags, conservatively avoids cases with two
 /// scalable dims. We could re-visit this in the future.
+///
+///  If the vector sizes are not provided:
+///   * the vector sizes are determined by the input operand and attributes,
+///   * update the inBounds attribute instead of masking.
 static VectorType getCollapsedVecType(VectorType type,
                                       ArrayRef<AffineMap> reassociation) {
   assert(type.getNumScalableDims() < 2 &&
@@ -1878,11 +1883,19 @@ static VectorType getCollapsedVecType(VectorType type,
 ///   vector::TransferWriteOp. - Write the result vector back to the destination
 ///   tensor.
 ///   If the vector sizes are not provided:
-///   * the vector sizes are determined by the input operand and attributes,
-///   * update the inBounds attribute instead of masking.
+/// Vectorize `linalg.unpack %src into %dest` as:
+///   // Reads a vector from the source tensor
+///   %read = vector.transfer_read  %src
+///   // Transpose %read as specified in `outer_dims_perm` attribute
+///   %tr = vector.transpose %read
+///   // Reshape the data based on the target
+///   %sc = vector.shape_cast %tr
+///   // Write the result vector to the destination tensor.
+///   vector.transfer_write %sc into %dest
 static LogicalResult
 vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
                           ArrayRef<int64_t> inputVectorSizes,
+                          ArrayRef<bool> inputScalableVecDims,
                           SmallVectorImpl<Value> &newResults) {
 
   // TODO: Introduce a parent class that will handle the insertion point update.
@@ -1899,25 +1912,54 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
 
   auto destSize = unpackOp.getDestRank();
 
-  if (!inputVectorSizes.empty())
-    assert(inputVectorSizes.size() == destSize &&
+  if (!inputVectorSizes.empty()) {
+    assert(inputVectorSizes.size() == destSize + sourceShape.size() &&
            "Incorrect number of input vector sizes");
+  }
+
+  SmallVector<bool> readScalableVectorFlags;
+  SmallVector<bool> writeScalableVectorFlags;
+  SmallVector<int64_t> readVectorSizes;
+  SmallVector<int64_t> writeVectorSizes;
 
-  // vectorSizes is the shape of the vector that will be used to do final
+  // Split input-vector-sizes into vector sizes for the read and write
+  // operations.
+  if (!inputVectorSizes.empty()) {
+    readVectorSizes.append(inputVectorSizes.begin(),
+                           inputVectorSizes.begin() + sourceShape.size());
+    writeVectorSizes.append(inputVectorSizes.begin() + sourceShape.size(),
+                            inputVectorSizes.end());
+  }
+  if (!inputScalableVecDims.empty()) {
+    readScalableVectorFlags.append(inputScalableVecDims.begin(),
+                                   inputScalableVecDims.begin() +
+                                       sourceShape.size());
+    writeScalableVectorFlags.append(inputScalableVecDims.begin() +
+                                        sourceShape.size(),
+                                    inputScalableVecDims.end());
+  } else {
+    readScalableVectorFlags = SmallVector<bool>(sourceShape.size(), false);
+    writeScalableVectorFlags = SmallVector<bool>(destSize, false);
+  }
+
+  // writeVectorSizes is the shape of the vector that will be used to do final
   // write on the destination tensor. It is set like this: Let's say the
   // source tensor is rank 'M' and the dest tensor rank 'N', where N <= M.
   // Thus:
-  // 1. vectorSizes = sourceShape.take_front(N)
-  // 2. if outer_dims_perms is present: do that permutation on vectorSizes.
+  // 1. writeVectorSizes = sourceShape.take_front(N)
+  // 2. if outer_dims_perms is present: do that permutation on writeVectorSizes.
   // 3. multiply all the locations in vectorSize pointed by innerDimPos by the
   //    innerTiles attribute value.
-  SmallVector<int64_t> vectorSizes(inputVectorSizes);
-  if (vectorSizes.empty()) {
-    llvm::append_range(vectorSizes, sourceShape.take_front(destSize));
+  // SmallVector<int64_t> writeVectorSizes(inputVectorSizes);
+  if (writeVectorSizes.empty()) {
+    if (ShapedType::isDynamicShape(sourceShape))
+      return failure();
+
+    llvm::append_range(writeVectorSizes, sourceShape.take_front(destSize));
     if (!outerDimsPerm.empty())
-      applyPermutationToVector(vectorSizes, outerDimsPerm);
+      applyPermutationToVector(writeVectorSizes, outerDimsPerm);
     for (auto [i, pos] : llvm::enumerate(innerDimPos))
-      vectorSizes[pos] *= innerTiles[i];
+      writeVectorSizes[pos] *= innerTiles[i];
 
     useInBoundsInsteadOfMasking = true;
   }
@@ -1941,17 +1983,20 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
   //   After applying outer_dims_perm: [8, 16]
   //   After appending the rest of the sourceShape: [8, 16, 32, 16]
 
-  SmallVector<int64_t> readVectorSizes(vectorSizes.begin(), vectorSizes.end());
-
-  for (auto [index, size] : enumerate(innerTiles)) {
-    readVectorSizes[innerDimPos[index]] =
-        llvm::divideCeil(readVectorSizes[innerDimPos[index]], size);
-  }
-  if (!outerDimsPerm.empty()) {
-    applyPermutationToVector(readVectorSizes, outerDimsPerm);
+  if (readVectorSizes.empty()) {
+    // Compute read-vector-sizes based on the write-vector-sizes and inner tile
+    // sizes. Note, this will only work when all sizes are static.
+    readVectorSizes = writeVectorSizes;
+    for (auto [index, size] : enumerate(innerTiles)) {
+      readVectorSizes[innerDimPos[index]] =
+          llvm::divideCeil(readVectorSizes[innerDimPos[index]], size);
+    }
+    if (!outerDimsPerm.empty()) {
+      applyPermutationToVector(readVectorSizes, outerDimsPerm);
+    }
+    readVectorSizes.append(sourceShape.begin() + writeVectorSizes.size(),
+                           sourceShape.end());
   }
-  readVectorSizes.append(sourceShape.begin() + vectorSizes.size(),
-                         sourceShape.end());
 
   Location loc = unpackOp->getLoc();
 
@@ -1963,7 +2008,7 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
   // to shape of source, then a mask is necessary.
   Value readResult = vector::createReadOrMaskedRead(
       rewriter, loc, unpackOp.getSource(), readVectorSizes, padValue,
-      /*useInBoundsInsteadOfMasking=*/false);
+      /*useInBoundsInsteadOfMasking=*/false, readScalableVectorFlags);
 
   PackingMetadata packMetadata;
   SmallVector<int64_t> lastDimToInsertPosPerm =
@@ -2009,7 +2054,7 @@ vectorizeAsTensorPadOp(RewriterBase &rewriter, tensor::PadOp padOp,
   assert(succeeded(status) && "failed to reify result shapes");
   auto maskedRead = vector::createReadOrMaskedRead(
       rewriter, loc, padOp.getSource(), inputVectorSizes, padValue,
-      /*useInBoundsInsteadOfMasking=*/false);
+      /*useInBoundsInsteadOfMasking=*/false, /*inputScalableVecSizes=*/{});
 
   // Create Xfer write Op
   Value dest = tensor::EmptyOp::create(rewriter, loc, reifiedReturnShapes[0],
@@ -2093,6 +2138,9 @@ static LogicalResult
 vectorizeUnPackOpPrecondition(linalg::UnPackOp unpackOp,
                               ArrayRef<int64_t> inputVectorSizes) {
 
+  // FIXME!!!
+  return success();
+
   if (llvm::any_of(unpackOp.getInnerTiles(), [](OpFoldResult res) {
         return !getConstantIntValue(res).has_value();
       })) {
@@ -2429,6 +2477,7 @@ vectorizePackOpPrecondition(linalg::PackOp packOp,
     LDBG() << "pad value is not constant: " << packOp;
     return failure();
   }
+
   ArrayRef<int64_t> resultTensorShape = packOp.getDestType().getShape();
   bool satisfyEmptyCond = true;
   if (inputVectorSizes.empty()) {
@@ -2507,12 +2556,14 @@ vectorizeScalableVectorPrecondition(Operation *op,
   if (numOfScalableDims == 0)
     return success();
 
+  // TODO: Check the following!
   auto linalgOp = dyn_cast<LinalgOp>(op);
 
-  // Cond 1: There's been no need for scalable vectorisation of
-  // non-linalg Ops so far
-  if (!linalgOp)
-    return failure();
+  // Cond 1: Reject Ops that don't implement the LinalgOp interface, with the
+  // exception of UnpackOp for which there is a dedicated hook.
+  if (!linalgOp) {
+    return isa<linalg::UnPackOp>(op) ? success() : failure();
+  }
 
   // Cond 2: There's been no need for more than 2 scalable dims so far
   if (numOfScalableDims > 2)
@@ -2610,7 +2661,7 @@ vectorizeScalableVectorPrecondition(Operation *op,
                  isa<linalg::MatmulTransposeAOp>(op) ||
                  isa<linalg::DepthwiseConv1DNwcWcOp>(op) ||
                  isa<linalg::MatvecOp>(op) || isa<linalg::Mmt4DOp>(op) ||
-                 hasReductionIterator(linalgOp));
+                 isa<linalg::UnPackOp>(op) || hasReductionIterator(linalgOp));
 }
 
 LogicalResult mlir::linalg::vectorizeOpPrecondition(
@@ -2743,7 +2794,8 @@ FailureOr<VectorizationResult> mlir::linalg::vectorize(
           })
           .Case<linalg::UnPackOp>([&](auto unpackOp) {
             return vectorizeAsTensorUnpackOp(rewriter, unpackOp,
-                                             inputVectorSizes, results);
+                                             inputVectorSizes,
+                                             inputScalableVecDims, results);
           })
           .Case<tensor::InsertSliceOp>([&](auto sliceOp) {
             return vectorizeAsInsertSliceOp(rewriter, sliceOp, inputVectorSizes,
@@ -3135,7 +3187,8 @@ vectorizeAsInsertSliceOp(RewriterBase &rewriter, tensor::InsertSliceOp sliceOp,
       vecType.getRank(), arith::ConstantIndexOp::create(rewriter, loc, 0));
   Value read = mlir::vector::createReadOrMaskedRead(
       rewriter, loc, source, vecType.getShape(), padValue,
-      /*useInBoundsInsteadOfMasking=*/inputVectorSizes.empty());
+      /*useInBoundsInsteadOfMasking=*/inputVectorSizes.empty(),
+      /*inputScalableVecSizes=*/{});
 
   // Create write
   auto writeIndices =

mlir/lib/Dialect/Vector/Utils/VectorUtils.cpp

@@ -279,14 +279,16 @@ vector::createUnrollIterator(VectorType vType, int64_t targetRank) {
   // Attempt to unroll until targetRank or the first scalable dimension (which
   // cannot be unrolled).
   auto shapeToUnroll = vType.getShape().drop_back(targetRank);
-  auto scalableDimsToUnroll = vType.getScalableDims().drop_back(targetRank);
-  auto it = llvm::find(scalableDimsToUnroll, true);
-  auto firstScalableDim = it - scalableDimsToUnroll.begin();
+  auto inputScalableVecDimsToUnroll =
+      vType.getScalableDims().drop_back(targetRank);
+  auto it = llvm::find(inputScalableVecDimsToUnroll, true);
+  auto firstScalableDim = it - inputScalableVecDimsToUnroll.begin();
   if (firstScalableDim == 0)
     return {};
   // All scalable dimensions should be removed now.
-  scalableDimsToUnroll = scalableDimsToUnroll.slice(0, firstScalableDim);
-  assert(!llvm::is_contained(scalableDimsToUnroll, true) &&
+  inputScalableVecDimsToUnroll =
+      inputScalableVecDimsToUnroll.slice(0, firstScalableDim);
+  assert(!llvm::is_contained(inputScalableVecDimsToUnroll, true) &&
          "unexpected leading scalable dimension");
   // Create an unroll iterator for leading dimensions.
   shapeToUnroll = shapeToUnroll.slice(0, firstScalableDim);
@@ -319,15 +321,15 @@ Value vector::createReadOrMaskedRead(OpBuilder &builder, Location loc,
                                      ArrayRef<int64_t> inputVectorSizes,
                                      Value padValue,
                                      bool useInBoundsInsteadOfMasking,
-                                     ArrayRef<bool> scalableDims) {
+                                     ArrayRef<bool> inputScalableVecDims) {
   assert(!llvm::is_contained(inputVectorSizes, ShapedType::kDynamic) &&
          "invalid input vector sizes");
   auto sourceShapedType = cast<ShapedType>(source.getType());
   auto sourceShape = sourceShapedType.getShape();
   assert(sourceShape.size() == inputVectorSizes.size() &&
          "expected same ranks.");
-  auto vectorType =
-      VectorType::get(inputVectorSizes, padValue.getType(), scalableDims);
+  auto vectorType = VectorType::get(inputVectorSizes, padValue.getType(),
+                                    inputScalableVecDims);
   assert(padValue.getType() == sourceShapedType.getElementType() &&
          "expected same pad element type to match source element type");
   int64_t readRank = inputVectorSizes.size();
@@ -356,8 +358,8 @@ Value vector::createReadOrMaskedRead(OpBuilder &builder, Location loc,
           ? memref::getMixedSizes(builder, loc, source)
           : tensor::getMixedSizes(builder, loc, source);
 
-  auto maskType =
-      VectorType::get(inputVectorSizes, builder.getI1Type(), scalableDims);
+  auto maskType = VectorType::get(inputVectorSizes, builder.getI1Type(),
+                                  inputScalableVecDims);
   Value mask =
       vector::CreateMaskOp::create(builder, loc, maskType, mixedSourceDims);
   return mlir::vector::maskOperation(builder, transferReadOp, mask)