Skip to content

[mlir][linalg] Fix vectorization of tensor.extract #118105

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Dec 5, 2024
Merged

[mlir][linalg] Fix vectorization of tensor.extract #118105

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
47376c5
[mlir][linalg] Fix vectorization of tensor.extract
banach-space Nov 29, 2024
aa35313
fixup! [mlir][linalg] Fix vectorization of tensor.extract
banach-space Dec 4, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 11 additions & 1 deletion mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1165,8 +1165,18 @@ vectorizeTensorExtract(RewriterBase &rewriter, VectorizationState &state,
loc, resultType, extractOp.getTensor(), transferReadIdxs,
permutationMap, inBounds);

// Mask this broadcasting xfer_read here rather than relying on the generic
// path (the generic path assumes identity masking map, which wouldn't be
// valid here).
SmallVector<int64_t> readMaskShape = {1};
auto readMaskType = VectorType::get(readMaskShape, rewriter.getI1Type());
auto allTrue = rewriter.create<vector::ConstantMaskOp>(
loc, readMaskType, vector::ConstantMaskKind::AllTrue);
auto *maskedReadOp =
mlir::vector::maskOperation(rewriter, transferReadOp, allTrue);

LDBG("Vectorised as scalar broadcast load: " << extractOp << "\n");
return VectorizationResult{VectorizationStatus::NewOp, transferReadOp};
return VectorizationResult{VectorizationStatus::NewOp, maskedReadOp};
}

// 2b. Handle contiguous access.
Expand Down
52 changes: 52 additions & 0 deletions mlir/test/Dialect/Linalg/vectorize-tensor-extract-masked.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -425,3 +425,55 @@ module attributes {transform.with_named_sequence} {
transform.yield
}
}

// -----

#map = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
func.func @scalar_broadcast(%init : tensor<1x1x3xi32>, %src: tensor<1x3x2x4xi32>, %idx :index) -> tensor<1x1x3xi32> {

%c0 = arith.constant 0 :index

%res = linalg.generic {
indexing_maps = [#map],
iterator_types = ["parallel", "parallel", "parallel"]}
outs(%init : tensor<1x1x3xi32>) {
^bb0(%out: i32):
%val = tensor.extract %src[%idx, %idx, %idx, %idx] : tensor<1x3x2x4xi32>
linalg.yield %val : i32
} -> tensor<1x1x3xi32>

return %res : tensor<1x1x3xi32>
}

// CHECK: #[[$MAP:.+]] = affine_map<(d0, d1, d2, d3) -> (0, 0, 0)>
// CHECK-LABEL: func.func @scalar_broadcast(
// CHECK-SAME: %[[INIT:.*]]: tensor<1x1x3xi32>,
// CHECK-SAME: %[[SRC:.*]]: tensor<1x3x2x4xi32>,
// CHECK-SAME: %[[IDX:.*]]: index) -> tensor<1x1x3xi32> {

/// Compute the mask for saving the final result
// CHECK: %[[C1:.*]] = arith.constant 1 : index
// CHECK: %[[C1_2:.*]] = arith.constant 1 : index
// CHECK: %[[C3:.*]] = arith.constant 3 : index
// CHECK: %[[MASK_RES:.*]] = vector.create_mask %[[C1]], %[[C1_2]], %[[C3]] : vector<1x1x4xi1>

/// Read and broadcast the scalar
// CHECK: %[[PAD:.*]] = arith.constant 0 : i32
// CHECK: %[[MASK_READ:.*]] = vector.constant_mask [1] : vector<1xi1>
// CHECK: %[[READ:.*]] = vector.mask %[[MASK_READ]] {
// CHECK-SAME: vector.transfer_read %[[SRC]]{{\[}}%[[IDX]], %[[IDX]], %[[IDX]], %[[IDX]]], %[[PAD]]
// CHECK-SAME: {in_bounds = [true, true, true], permutation_map = #[[$MAP]]} : tensor<1x3x2x4xi32>, vector<1x1x4xi32>
// CHECK-SAME: } : vector<1xi1> -> vector<1x1x4xi32>

/// Save the result in the output tensor
// CHECK: vector.mask %[[MASK_RES]] {
// CHECK-SAME: vector.transfer_write %[[READ]], %[[INIT]]{{.*}} {in_bounds = [true, true, true]} : vector<1x1x4xi32>, tensor<1x1x3xi32>
// CHECK-SAME: } : vector<1x1x4xi1> -> tensor<1x1x3xi32>

module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%module: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["linalg.generic"]} in %module : (!transform.any_op) -> !transform.any_op
transform.structured.vectorize %0 vector_sizes [1, 1, 4] {vectorize_nd_extract} : !transform.any_op
transform.yield
}
}
28 changes: 15 additions & 13 deletions mlir/test/Dialect/Linalg/vectorize-tensor-extract.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,7 +66,7 @@ module attributes {transform.with_named_sequence} {
// -----

#map = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
func.func @vectorize_nd_tensor_extract_constant_idx(%arg0: tensor<3x3xf32>, %arg2: tensor<1x1x3xf32>) -> tensor<1x1x3xf32> {
func.func @vectorize_nd_tensor_extract_scalar_broadcast(%arg0: tensor<3x3xf32>, %arg2: tensor<1x1x3xf32>) -> tensor<1x1x3xf32> {
%c0 = arith.constant 1 : index
%c1 = arith.constant 2 : index
%2 = linalg.generic {
Expand All @@ -80,17 +80,17 @@ func.func @vectorize_nd_tensor_extract_constant_idx(%arg0: tensor<3x3xf32>, %arg
return %2 : tensor<1x1x3xf32>
}

// CHECK: #[[$MAP:.*]] = affine_map<(d0, d1) -> (0, 0, 0)>
// CHECK-LABEL: func.func @vectorize_nd_tensor_extract_constant_idx(
// CHECK: #[[$MAP:.+]] = affine_map<(d0, d1) -> (0, 0, 0)>
// CHECK-LABEL: func.func @vectorize_nd_tensor_extract_scalar_broadcast(
// CHECK-SAME: %[[ARG_0:.*]]: tensor<3x3xf32>,
// CHECK-SAME: %[[ARG_1:.*]]: tensor<1x1x3xf32>) -> tensor<1x1x3xf32> {
// CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[C2:.*]] = arith.constant 2 : index
// CHECK-DAG: %[[C0_f32_2:.*]] = arith.constant 0.000000e+00 : f32
// CHECK-DAG: %[[C0_f32:.*]] = arith.constant 0.000000e+00 : f32
// CHECK: %[[READ:.*]] = vector.transfer_read %[[ARG_0]][%[[C1]], %[[C2]]], %[[C0_f32]] {in_bounds = [true, true, true], permutation_map = #[[$MAP]]} : tensor<3x3xf32>, vector<1x1x3xf32>
// CHECK: %[[C0_4:.*]] = arith.constant 0 : index
// CHECK: vector.transfer_write %[[READ]], %[[ARG_1]][%[[C0_4]], %[[C0_4]], %[[C0_4]]] : vector<1x1x3xf32>, tensor<1x1x3xf32>
// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[MASK:.*]] = vector.constant_mask [1] : vector<1xi1>
// CHECK: %[[READ:.*]] = vector.mask %[[MASK]] { vector.transfer_read %[[ARG_0]][%[[C1]], %[[C2]]], {{.*}} {in_bounds = [true, true, true], permutation_map = #[[$MAP]]} : tensor<3x3xf32>, vector<1x1x3xf32> } : vector<1xi1> -> vector<1x1x3xf32>
// CHECK: %[[C0_2:.*]] = arith.constant 0 : index
// CHECK: vector.transfer_write %[[READ]], %[[ARG_1]]{{\[}}%[[C0_2]], %[[C0_2]], %[[C0_2]]] : vector<1x1x3xf32>, tensor<1x1x3xf32>

module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
Expand Down Expand Up @@ -823,7 +823,7 @@ func.func @vectorize_scalar_broadcast_column_tensor(%in: tensor<1x1x4xi32>) -> t
return %out:tensor<1x1x4xi32>
}

// CHECK: #[[$ATTR_1:.+]] = affine_map<(d0, d1) -> (0, 0, 0)>
// CHECK: #[[$MAP:.+]] = affine_map<(d0, d1) -> (0, 0, 0)>
// CHECK-LABEL: func.func @vectorize_scalar_broadcast_column_tensor(
// CHECK-SAME: %[[VAL_0:.*]]: tensor<1x1x4xi32>) -> tensor<1x1x4xi32> {
// CHECK: %[[VAL_1:.*]] = arith.constant 4 : index
Expand All @@ -844,12 +844,14 @@ func.func @vectorize_scalar_broadcast_column_tensor(%in: tensor<1x1x4xi32>) -> t
// CHECK: %[[VAL_16:.*]] = arith.constant dense<true> : vector<1x1x4xi1>
// CHECK: %[[VAL_17:.*]] = arith.constant dense<0> : vector<1x1x4xi32>
// CHECK: %[[VAL_18:.*]] = arith.constant 0 : index
// CHECK: %[[VAL_20:.*]] = vector.shape_cast %[[VAL_15]] : vector<1x1x4xindex> to vector<4xindex>
// CHECK: %[[VAL_21:.*]] = vector.extract %[[VAL_20]][0] : index from vector<4xindex>
// CHECK: %[[VAL_22:.*]] = arith.constant 0 : i32
// CHECK: %[[VAL_23:.*]] = vector.transfer_read %[[VAL_3]]{{\[}}%[[VAL_21]], %[[VAL_2]]], %[[VAL_22]] {in_bounds = [true, true, true], permutation_map = #[[$ATTR_1]]} : tensor<15x1xi32>, vector<1x1x4xi32>
// CHECK: %[[VAL_19:.*]] = vector.shape_cast %[[VAL_15]] : vector<1x1x4xindex> to vector<4xindex>
// CHECK: %[[VAL_20:.*]] = vector.extract %[[VAL_19]][0] : index from vector<4xindex>
// CHECK: %[[VAL_21:.*]] = arith.constant 0 : i32
// CHECK: %[[VAL_22:.*]] = vector.constant_mask [1] : vector<1xi1>
// CHECK: %[[VAL_23:.*]] = vector.mask %[[VAL_22]] { vector.transfer_read %[[VAL_3]]{{\[}}%[[VAL_20]], %[[VAL_2]]], %[[VAL_21]] {in_bounds = [true, true, true], permutation_map = #[[$MAP]]} : tensor<15x1xi32>, vector<1x1x4xi32> } : vector<1xi1> -> vector<1x1x4xi32>
// CHECK: %[[VAL_24:.*]] = arith.constant 0 : index
// CHECK: %[[VAL_25:.*]] = vector.transfer_write %[[VAL_23]], %[[VAL_0]]{{\[}}%[[VAL_24]], %[[VAL_24]], %[[VAL_24]]] : vector<1x1x4xi32>, tensor<1x1x4xi32>
// CHECK: return %[[VAL_25]] : tensor<1x1x4xi32>

module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
Expand Down
Loading