Skip to content

[mlir][linalg] Fix scalable vectorisation of tensor.extract #100325

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 1 commit into from
Jul 25, 2024
Merged

[mlir][linalg] Fix scalable vectorisation of tensor.extract #100325

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
9 changes: 5 additions & 4 deletions mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1069,19 +1069,20 @@ vectorizeTensorExtract(RewriterBase &rewriter, VectorizationState &state,
// * for vector indices (e.g. `vector<1x1x4xindex>`) - extract the bottom
// (0th) element and use that.
SmallVector<Value> transferReadIdxs;
auto resTrailingDim = resultType.getShape().back();
auto zero = rewriter.create<arith::ConstantOp>(
loc, rewriter.getI32Type(), rewriter.getZeroAttr(rewriter.getI32Type()));
for (size_t i = 0; i < extractOp.getIndices().size(); i++) {
auto idx = bvm.lookup(extractOp.getIndices()[i]);
Value idx = bvm.lookup(extractOp.getIndices()[i]);
if (idx.getType().isIndex()) {
transferReadIdxs.push_back(idx);
continue;
}

auto indexAs1dVector = rewriter.create<vector::ShapeCastOp>(
loc, VectorType::get({resTrailingDim}, rewriter.getIndexType()),
bvm.lookup(extractOp.getIndices()[i]));
loc,
VectorType::get(resultType.getShape().back(), rewriter.getIndexType(),
resultType.getScalableDims().back()),
idx);
transferReadIdxs.push_back(
rewriter.create<vector::ExtractElementOp>(loc, indexAs1dVector, zero));
}
Expand Down
111 changes: 98 additions & 13 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
@@ -1,29 +1,52 @@
// RUN: mlir-opt %s -transform-interpreter -split-input-file | FileCheck %s

func.func @masked_static_vectorize_nd_tensor_extract_with_affine_apply_contiguous(%6: tensor<80x16xf32>, %arg0: index, %extracted_slice : tensor<1x3xf32>) -> tensor<1x3xf32> {
func.func @masked_static_vectorize_nd_tensor_extract_with_affine_apply_contiguous(
%src: tensor<80x16xf32>,
%output : tensor<1x3xf32>,
%idx: index) -> tensor<1x3xf32> {

%c79 = arith.constant 79 : index
%1 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
} outs(%extracted_slice : tensor<1x3xf32>) {
} outs(%output : tensor<1x3xf32>) {
^bb0(%out: f32):
%2 = linalg.index 1 : index
%3 = affine.apply affine_map<(d0, d1) -> (d0 + d1)>(%2, %arg0)
%extracted = tensor.extract %6[%c79, %3] : tensor<80x16xf32>
%3 = affine.apply affine_map<(d0, d1) -> (d0 + d1)>(%2, %idx)
%extracted = tensor.extract %src[%c79, %3] : tensor<80x16xf32>
linalg.yield %extracted : f32
} -> tensor<1x3xf32>
return %1 : tensor<1x3xf32>
}

// CHECK-LABEL: func.func @masked_static_vectorize_nd_tensor_extract_with_affine_apply_contiguous
// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 3 : index
// CHECK: %[[VAL_8:.*]] = vector.create_mask %[[VAL_4]], %[[VAL_5]] : vector<1x4xi1>
// CHECK: %[[VAL_9:.*]] = vector.mask %[[VAL_8]] { vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<1x3xf32>, vector<1x4xf32> } : vector<1x4xi1> -> vector<1x4xf32>
// CHECK: %[[VAL_11:.*]] = vector.broadcast {{.*}} : index to vector<4xindex>
// CHECK: %[[VAL_12:.*]] = arith.addi {{.*}} : vector<4xindex>
// CHECK: %[[VAL_20:.*]] = vector.mask %[[VAL_8]] { vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<80x16xf32>, vector<1x4xf32> } : vector<1x4xi1> -> vector<1x4xf32>
// CHECK: %[[VAL_22:.*]] = vector.mask %[[VAL_8]] { vector.transfer_write {{.*}} {in_bounds = [true, true]} : vector<1x4xf32>, tensor<1x3xf32> } : vector<1x4xi1> -> tensor<1x3xf32>
// CHECK-SAME: %[[SRC:.*]]: tensor<80x16xf32>,
// CHECK-SAME: %[[OUTPUT:.*]]: tensor<1x3xf32>,
// CHECK-SAME: %[[IDX_IN:.*]]: index) -> tensor<1x3xf32> {

/// Create the mask
// CHECK-DAG: %[[DIM_0:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[DIM_1:.*]] = arith.constant 3 : index
// CHECK-DAG: %[[C79:.*]] = arith.constant 79 : index
// CHECK: %[[MASK:.*]] = vector.create_mask %[[DIM_0]], %[[DIM_1]] : vector<1x4xi1>

/// TODO: This transfer_read is redundant - remove
// CHECK: vector.mask %[[MASK]] { vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<1x3xf32>, vector<1x4xf32> } : vector<1x4xi1> -> vector<1x4xf32>

/// Caluclate the index vector
// CHECK: %[[STEP:.*]] = vector.step : vector<4xindex>
// CHECK: %[[IDX_BC:.*]] = vector.broadcast %[[IDX_IN]] : index to vector<4xindex>
// CHECK: %[[IDX_VEC:.*]] = arith.addi %[[STEP]], %[[IDX_BC]] : vector<4xindex>
// CHECK: %[[C0:.*]] = arith.constant 0 : i32
// CHECK: %[[SC:.*]] = vector.shape_cast %[[IDX_VEC]] : vector<4xindex> to vector<4xindex>

/// Extract the starting point from the index vector
// CHECK: %[[IDX_START:.*]] = vector.extractelement %[[SC]]{{\[}}%[[C0]] : i32] : vector<4xindex>

// Final read and write
// CHECK: %[[READ:.*]] = vector.mask %[[MASK]] { vector.transfer_read %[[SRC]]{{\[}}%[[C79]], %[[IDX_START]]], {{.*}} {in_bounds = [true, true]} : tensor<80x16xf32>, vector<1x4xf32> } : vector<1x4xi1> -> vector<1x4xf32>
// CHECK: %[[C0_1:.*]] = arith.constant 0 : index
// CHECK: vector.mask %[[MASK]] { vector.transfer_write %[[READ]], %[[OUTPUT]]{{\[}}%[[C0_1]], %[[C0_1]]] {in_bounds = [true, true]} : vector<1x4xf32>, tensor<1x3xf32> } : vector<1x4xi1> -> tensor<1x3xf32>

module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
Expand All @@ -33,7 +56,69 @@ module attributes {transform.with_named_sequence} {
}
}

// -----
// -----

// Identical to the above, but with scalable vectors.

func.func @masked_static_vectorize_nd_tensor_extract_with_affine_apply_contiguous_scalable(
%src: tensor<80x16xf32>,
%output : tensor<1x3xf32>,
%idx: index) -> tensor<1x3xf32> {

%c79 = arith.constant 79 : index
%1 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
} outs(%output : tensor<1x3xf32>) {
^bb0(%out: f32):
%2 = linalg.index 1 : index
%3 = affine.apply affine_map<(d0, d1) -> (d0 + d1)>(%2, %idx)
%extracted = tensor.extract %src[%c79, %3] : tensor<80x16xf32>
linalg.yield %extracted : f32
} -> tensor<1x3xf32>

return %1 : tensor<1x3xf32>
}

// CHECK-LABEL: func.func @masked_static_vectorize_nd_tensor_extract_with_affine_apply_contiguous_scalable
// CHECK-SAME: %[[SRC:.*]]: tensor<80x16xf32>,
// CHECK-SAME: %[[OUTPUT:.*]]: tensor<1x3xf32>,
// CHECK-SAME: %[[IDX_IN:.*]]: index) -> tensor<1x3xf32> {

/// Create the mask
// CHECK-DAG: %[[DIM_0:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[DIM_1:.*]] = arith.constant 3 : index
// CHECK-DAG: %[[C79:.*]] = arith.constant 79 : index
// CHECK: %[[MASK:.*]] = vector.create_mask %[[DIM_0]], %[[DIM_1]] : vector<1x[4]xi1>

/// TODO: This transfer_read is redundant - remove
// CHECK: vector.mask %[[MASK]] { vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<1x3xf32>, vector<1x[4]xf32> } : vector<1x[4]xi1> -> vector<1x[4]xf32>

/// Caluclate the index vector
// CHECK: %[[STEP:.*]] = vector.step : vector<[4]xindex>
// CHECK: %[[IDX_BC:.*]] = vector.broadcast %[[IDX_IN]] : index to vector<[4]xindex>
// CHECK: %[[IDX_VEC:.*]] = arith.addi %[[STEP]], %[[IDX_BC]] : vector<[4]xindex>
// CHECK: %[[C0:.*]] = arith.constant 0 : i32
// CHECK: %[[SC:.*]] = vector.shape_cast %[[IDX_VEC]] : vector<[4]xindex> to vector<[4]xindex>

/// Extract the starting point from the index vector
// CHECK: %[[IDX_START:.*]] = vector.extractelement %[[SC]]{{\[}}%[[C0]] : i32] : vector<[4]xindex>

// Final read and write
// CHECK: %[[READ:.*]] = vector.mask %[[MASK]] { vector.transfer_read %[[SRC]]{{\[}}%[[C79]], %[[IDX_START]]], {{.*}} {in_bounds = [true, true]} : tensor<80x16xf32>, vector<1x[4]xf32> } : vector<1x[4]xi1> -> vector<1x[4]xf32>
// CHECK: %[[C0_1:.*]] = arith.constant 0 : index
// CHECK: vector.mask %[[MASK]] { vector.transfer_write %[[READ]], %[[OUTPUT]]{{\[}}%[[C0_1]], %[[C0_1]]] {in_bounds = [true, true]} : vector<1x[4]xf32>, tensor<1x3xf32> } : vector<1x[4]xi1> -> tensor<1x3xf32>


module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
transform.structured.vectorize %0 vector_sizes [1, [4]] {vectorize_nd_extract} : !transform.any_op
transform.yield
}
}

// -----

func.func @masked_dynamic_vectorize_nd_tensor_extract_with_affine_apply_contiguous(%6: tensor<?x?xf32>, %arg0: index, %extracted_slice : tensor<?x?xf32>) -> tensor<?x?xf32> {
%c79 = arith.constant 79 : index
Expand Down
Loading