[mlir][vector] Determine vector sizes from the result shape in the case of tensor pack

When the vector sizes are not passed as inputs to the vector transform operation,
the vector sizes are queried from the static result shape in the case of tensor.pack op.

Author: pashu123

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

@@ -1412,10 +1412,11 @@ static SmallVector<int64_t> getTiledPackShape(tensor::PackOp packOp,
 
 /// Create a TransferReadOp from `source` with static shape `readShape`. If the
 /// vector type for the read is not the same as the type of `source`, then a
-/// mask is created on the read.
+/// mask is created on the read.  If `doMasking` parameter is set to false we
+/// update the `inBounds` attribute instead of masking.
 static Value createReadOrMaskedRead(OpBuilder &builder, Location loc,
                                     Value source, ArrayRef<int64_t> readShape,
-                                    Value padValue) {
+                                    Value padValue, bool doMasking = true) {
   assert(llvm::none_of(readShape,
                        [](int64_t s) { return s == ShapedType::kDynamic; }));
   auto sourceShape = dyn_cast<ShapedType>(source.getType()).getShape();
@@ -1424,14 +1425,21 @@ static Value createReadOrMaskedRead(OpBuilder &builder, Location loc,
   auto vectorType = VectorType::get(readShape, padValue.getType());
   int64_t readRank = readShape.size();
   auto zero = builder.create<arith::ConstantIndexOp>(loc, 0);
+  SmallVector<bool> inBoundsVal(readRank, true);
+  if (!doMasking) {
+    // Update the inBounds attribute.
+    for (unsigned i = 0; i < readRank; i++)
+      inBoundsVal[i] = sourceShape[i] == readShape[i];
+  }
   auto transferReadOp = builder.create<vector::TransferReadOp>(
       loc,
       /*vectorType=*/vectorType,
       /*source=*/source,
       /*indices=*/SmallVector<Value>(readRank, zero),
       /*padding=*/padValue,
-      /*inBounds=*/SmallVector<bool>(readRank, true));
-  if (llvm::equal(readShape, sourceShape)) {
+      /*inBounds=*/inBoundsVal);
+
+  if (llvm::equal(readShape, sourceShape) || !doMasking) {
     return transferReadOp;
   }
   SmallVector<OpFoldResult> mixedSourceDims =
@@ -1482,11 +1490,10 @@ static Operation *createWriteOrMaskedWrite(OpBuilder &builder, Location loc,
   return write;
 }
 
-/// Vectorize tensor::PackOp with (1) static innerTiles and (2) constant
-/// padding value into:
+/// Vectorize tensor::PackOp with (1) static innerTiles (2) constant
+/// padding value and (3) input vector sizes into:
 /// masked_transfer_read->shape_cast->transpose->transfer_write_in_bounds
 /// As in the following example:
-///
 /// %pack = tensor.pack %src inner_dims_pos = [2, 1] inner_tiles = [16, 2]
 ///     into %dst : tensor<32x8x16xf32> -> tensor<32x4x1x16x2xf32>
 ///
@@ -1505,6 +1512,10 @@ static Operation *createWriteOrMaskedWrite(OpBuilder &builder, Location loc,
 ///     %empty[%c0_0, %c0_0, %c0_0, %c0_0, %c0_0]
 ///     {in_bounds = [true, true, true, true, true]}
 ///     : vector<32x4x1x16x2xf32>, tensor<32x4x1x16x2xf32>
+///
+/// If the (3) input vector sizes are not provided, the vector sizes are
+/// determined by the result tensor shape. Also, we update the inBounds
+/// attribute instead of masking.
 static LogicalResult
 vectorizeAsTensorPackOp(RewriterBase &rewriter, tensor::PackOp packOp,
                         ArrayRef<int64_t> inputVectorSizes,
@@ -1525,6 +1536,16 @@ vectorizeAsTensorPackOp(RewriterBase &rewriter, tensor::PackOp packOp,
   (void)status; // prevent unused variable warning on non-assert builds.
   assert(succeeded(status) && "failed to reify result shapes");
 
+  // If the input vector sizes are not provided, then the vector sizes are
+  // determined by the result tensor shape. In case the vector sizes aren't
+  // provided, we update the inBounds attribute instead of masking.
+  bool doMasking = true;
+  if (inputVectorSizes.empty()) {
+    ArrayRef<int64_t> resultTensorShape = packOp.getDestType().getShape();
+    inputVectorSizes = resultTensorShape.take_front(packOp.getSourceRank());
+    doMasking = false;
+  }
+
   // Create masked TransferReadOp.
   SmallVector<int64_t> inputShape(inputVectorSizes);
   auto innerTiles = packOp.getStaticInnerTiles();
@@ -1536,7 +1557,7 @@ vectorizeAsTensorPackOp(RewriterBase &rewriter, tensor::PackOp packOp,
   for (auto [idx, size] : enumerate(innerTiles))
     inputShape[innerDimsPos[idx]] *= size;
   auto maskedRead = createReadOrMaskedRead(rewriter, loc, packOp.getSource(),
-                                           inputShape, padValue);
+                                           inputShape, padValue, doMasking);
 
   // Create ShapeCastOp.
   SmallVector<int64_t> destShape(inputVectorSizes);
@@ -1763,7 +1784,7 @@ vectorizeDynamicLinalgOpPrecondition(linalg::LinalgOp op,
 /// Returns success if `inputVectorSizes` is a valid masking configuraion for
 /// given `shape`, i.e., it meets:
 ///   1. The numbers of elements in both array are equal.
-///   2. `inputVectorSizes` does nos have dynamic dimensions.
+///   2. `inputVectorSizes` does not have dynamic dimensions.
 ///   3. All the values in `inputVectorSizes` are greater than or equal to
 ///      static sizes in `shape`.
 static LogicalResult
@@ -1881,18 +1902,25 @@ static LogicalResult vectorizeLinalgOpPrecondition(
   return success();
 }
 
-/// TODO: Use a matcher to check for a constant padding value.
 static LogicalResult
 vectorizePackOpPrecondition(tensor::PackOp packOp,
                             ArrayRef<int64_t> inputVectorSizes) {
   auto padValue = packOp.getPaddingValue();
-  if (padValue && !padValue.getDefiningOp<arith::ConstantOp>()) {
+  Attribute cstAttr;
+  if (padValue && !matchPattern(padValue, m_Constant(&cstAttr))) {
     LDBG("pad value is not constant: " << packOp << "\n");
     return failure();
   }
-
   ArrayRef<int64_t> resultTensorShape = packOp.getDestType().getShape();
-  if (failed(isValidMaskedInputVector(
+  bool satisfyEmptyCond = true;
+  if (inputVectorSizes.empty()) {
+    if (!packOp.getDestType().hasStaticShape() ||
+        !packOp.getSourceType().hasStaticShape())
+      satisfyEmptyCond = false;
+  }
+
+  if (!satisfyEmptyCond &&
+      failed(isValidMaskedInputVector(
           resultTensorShape.take_front(packOp.getSourceRank()),
           inputVectorSizes)))
     return failure();

mlir/test/Dialect/Linalg/vectorization-unsupported.mlir

@@ -109,3 +109,20 @@ module attributes {transform.with_named_sequence} {
     transform.yield
   }
 }
+
+  // -----
+
+func.func @test_pack_no_vectorize_dynamic_shape(%arg0: tensor<?xf32>, %arg1: tensor<4x16xf32>) -> tensor<4x16xf32> {
+  %pad = arith.constant 0.000000e+00 : f32
+  // expected-error @+1 {{Attempted to vectorize, but failed}}
+  %pack = tensor.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [0] inner_tiles = [16] into %arg1 : tensor<?xf32> -> tensor<4x16xf32>
+  return %pack : tensor<4x16xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["tensor.pack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    transform.structured.vectorize %0 : !transform.any_op
+    transform.yield
+  }
+}

mlir/test/Dialect/Linalg/vectorization.mlir

@@ -812,3 +812,58 @@ func.func @test_vectorize_unpack_no_masks(%source: tensor<8x8x32x16xf32>, %dest:
     transform.yield
   } 
 }
+
+  // -----
+
+// CHECK-LABEL: test_vectorize_pack_no_vector_sizes
+func.func @test_vectorize_pack_no_vector_sizes(%arg0: tensor<64x4xf32>, %arg1: tensor<2x4x16x2xf32>) -> tensor<2x4x16x2xf32> {
+  %pack = tensor.pack %arg0 outer_dims_perm = [1, 0] inner_dims_pos = [0, 1] inner_tiles = [16, 2] into %arg1 : tensor<64x4xf32> -> tensor<2x4x16x2xf32>
+  return %pack : tensor<2x4x16x2xf32>
+}
+//  CHECK-DAG: %[[cst:.*]] = arith.constant 0.000000e+00 : f32
+//  CHECK-DAG: %[[c0:.*]] = arith.constant 0 : index
+//      CHECK: %[[read:.*]] = vector.transfer_read %{{.*}}[%[[c0]], %[[c0]]], %[[cst]]
+// CHECK-SAME:    {in_bounds = [true, true]} : tensor<64x4xf32>, vector<64x4xf32>
+//      CHECK: %[[shape_cast:.*]] = vector.shape_cast %[[read]] : vector<64x4xf32> to vector<4x16x2x2xf32>
+//      CHECK: %[[transpose:.*]] = vector.transpose %[[shape_cast]], [2, 0, 1, 3] : vector<4x16x2x2xf32> to vector<2x4x16x2xf32>
+//  CHECK-DAG: %[[c0_1:.*]] = arith.constant 0 : index
+//  CHECK-DAG: %[[empty:.*]] = tensor.empty() : tensor<2x4x16x2xf32>
+//      CHECK: %[[write:.*]] = vector.transfer_write %[[transpose]], %[[empty]][%[[c0_1]], %[[c0_1]], %[[c0_1]], %[[c0_1]]]
+// CHECK-SAME:   {in_bounds = [true, true, true, true]} : vector<2x4x16x2xf32>, tensor<2x4x16x2xf32>
+//      CHECK: return %[[write]] : tensor<2x4x16x2xf32>
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["tensor.pack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    transform.structured.vectorize %0 : !transform.any_op
+    transform.yield
+  }
+}
+
+  // -----
+
+// CHECK-LABEL: test_vectorize_padded_pack_no_vector_sizes
+func.func @test_vectorize_padded_pack_no_vector_sizes(%arg0: tensor<32x7x15xf32>, %arg1: tensor<32x4x1x16x2xf32>) -> tensor<32x4x1x16x2xf32> {
+  %pad = arith.constant 0.000000e+00 : f32
+  %pack = tensor.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [2, 1] inner_tiles = [16, 2] into %arg1 : tensor<32x7x15xf32> -> tensor<32x4x1x16x2xf32>
+  return %pack : tensor<32x4x1x16x2xf32>
+}
+//  CHECK-DAG: %[[cst:.*]] = arith.constant 0.000000e+00 : f32
+//  CHECK-DAG: %[[c0:.*]] = arith.constant 0 : index
+//      CHECK: %[[transfer_read:.*]] =  vector.transfer_read %{{.*}}[%[[c0]], %[[c0]], %[[c0]]], %[[cst]]
+// CHECK-SAME:   {in_bounds = [true, false, false]} : tensor<32x7x15xf32>, vector<32x8x16xf32> 
+//      CHECK: %[[shape_cast:.*]] = vector.shape_cast %[[transfer_read]] : vector<32x8x16xf32> to vector<32x4x2x1x16xf32>
+//      CHECK: %[[transpose:.*]] = vector.transpose %[[shape_cast]], [0, 1, 3, 4, 2] : vector<32x4x2x1x16xf32> to vector<32x4x1x16x2xf32>
+//  CHECK-DAG: %[[c0_1:.*]] = arith.constant 0 : index
+//  CHECK-DAG: %[[empty:.*]] = tensor.empty() : tensor<32x4x1x16x2xf32>
+//      CHECK: %[[write:.*]] = vector.transfer_write %[[transpose]], %[[empty]][%[[c0_1]], %[[c0_1]], %[[c0_1]], %[[c0_1]], %[[c0_1]]]
+// CHECK-SAME:   {in_bounds = [true, true, true, true, true]} : vector<32x4x1x16x2xf32>, tensor<32x4x1x16x2xf32>
+//      CHECK: return %[[write]] : tensor<32x4x1x16x2xf32>
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["tensor.pack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    transform.structured.vectorize %0 : !transform.any_op
+    transform.yield
+  }
+}