Skip to content

[mlir] Vectorize tensor.pad with low padding for unit dims #133808

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
nirvedhmeshram merged 4 commits into from
Apr 2, 2025
+87 −4
Merged

[mlir] Vectorize tensor.pad with low padding for unit dims #133808

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
8b74b2b
[mlir] Vectorize tenosr.pad with low padding for unit dims
nirvedhmeshram Mar 31, 2025
81a9fc9
address reviwer comments
nirvedhmeshram Apr 1, 2025
1465643
typos
nirvedhmeshram Apr 1, 2025
20551dd
reviwer comments 2
nirvedhmeshram Apr 2, 2025
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
11 changes: 7 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 @@ -2178,11 +2178,14 @@ vectorizePadOpPrecondition(tensor::PadOp padOp,
inputVectorSizes)))
return failure();

if (llvm::any_of(padOp.getLow(), [](Value v) {
std::optional<int64_t> res = getConstantIntValue(v);
return !res.has_value() || res.value() != 0;
if (llvm::any_of(llvm::enumerate(padOp.getLow()), [&](const auto &en) {
Value padValue = en.value();
unsigned pos = en.index();
std::optional<int64_t> res = getConstantIntValue(padValue);
return (!res.has_value() || res.value() != 0) &&
resultTensorShape[pos] != 1;
})) {
LDBG("low pad must all be zero: " << padOp << "\n");
LDBG("low pad must all be zero for all non unit dims: " << padOp << "\n");
return failure();
}

Expand Down
24 changes: 24 additions & 0 deletions mlir/test/Dialect/Linalg/vectorization-unsupported.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -305,6 +305,30 @@ module attributes {transform.with_named_sequence} {

// -----

func.func @test_masked_vectorize_lowpad(
%0 : tensor<?x?xf32>, %h0 : index, %h1 : index, %l0 : index)
-> tensor<2x4xf32> {
// expected-error @+3 {{Attempted to vectorize, but failed}}
%cst = arith.constant 42.43 : f32
%c0 = arith.constant 0 : index
%1 = tensor.pad %0 low[%l0, %c0] high[%h0, %h1] {
^bb0(%hh1: index, %hh2: index):
tensor.yield %cst : f32
} : tensor<?x?xf32> to tensor<2x4xf32>
return %1: tensor<2x4xf32>
}

module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["tensor.pad"]} in %arg1
: (!transform.any_op) -> !transform.any_op
transform.structured.vectorize %0 vector_sizes [2, 4] : !transform.any_op
transform.yield
}
}

// -----

// With dynamically shaped source, the vectorizer infers the vector size for
// xfer Ops from the destination tensor and, conservatively, assumes
// out-of-bounds accesses. Out-of-bounds accesses require a pad value, but
Expand Down
40 changes: 40 additions & 0 deletions mlir/test/Dialect/Linalg/vectorization.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -666,6 +666,46 @@ module attributes {transform.with_named_sequence} {

// -----

// CHECK-LABEL: func @test_masked_vectorize_unit_lowpad
func.func @test_masked_vectorize_unit_lowpad(
%0 : tensor<?x?xf32>, %h0 : index, %h1 : index, %l0 : index)
-> tensor<1x4xf32>
{
// CHECK-DAG: %[[c42:.*]] = arith.constant 4.243000e+01 : f32
// CHECK-DAG: %[[c0:.*]] = arith.constant 0 : index
// CHECK: %[[c0_1:.*]] = arith.constant 0 : index
// CHECK-DAG: %[[d0:.*]] = tensor.dim {{.*}} : tensor<?x?xf32>
// CHECK-DAG: %[[d1:.*]] = tensor.dim {{.*}} : tensor<?x?xf32>
// CHECK: %[[mask:.*]] = vector.create_mask %[[d0]], %[[d1]] : vector<1x4xi1>
// CHECK: %[[masked_read:.*]] = vector.mask %[[mask]] {
// CHECK-SAME: vector.transfer_read %{{.*}}[%[[c0_1]], %[[c0_1]]], %[[c42]]
// CHECK-SAME: {in_bounds = [true, true]} : tensor<?x?xf32>, vector<1x4xf32>
// CHECK-SAME: } : vector<1x4xi1> -> vector<1x4xf32>
// CHECK-DAG: %[[empty:.*]] = tensor.empty() : tensor<1x4xf32>
// CHECK-DAG: %[[c0_2:.*]] = arith.constant 0 : index
// CHECK: %[[masked_write:.*]] = vector.transfer_write %[[masked_read]], %[[empty]][%[[c0_2]], %[[c0_2]]]
// CHECK-SAME: {in_bounds = [true, true]} : vector<1x4xf32>, tensor<1x4xf32>
// CHECK: return %[[masked_write]] : tensor<1x4xf32>
%cst = arith.constant 42.43 : f32
%c0 = arith.constant 0 : index
%1 = tensor.pad %0 low[%l0, %c0] high[%h0, %h1] {
^bb0(%hh1: index, %hh2: index):
tensor.yield %cst : f32
} : tensor<?x?xf32> to tensor<1x4xf32>
return %1: tensor<1x4xf32>
}

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

// -----

// Input identical as the test in vectorization-with-patterns.mlir. Output is
// different - vector sizes are inferred (rather than user-specified) and hence
// masking was used.
Expand Down