Skip to content

[MLIR] Enable scalable vectorization for linalg.batch_matmul #172333

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

Open
momchil-velikov wants to merge 1 commit into
base: main
Choose a base branch
from
+85 −0
Open

[MLIR] Enable scalable vectorization for linalg.batch_matmul #172333

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
1 change: 1 addition & 0 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 @@ -2640,6 +2640,7 @@ vectorizeScalableVectorPrecondition(Operation *op,
// Cond 4: Only the following ops are supported in the
// presence of scalable vectors
return success(isElementwise(linalgOp) || isa<linalg::MatmulOp>(op) ||
isa<linalg::BatchMatmulOp>(op) ||
isa<linalg::DepthwiseConv1DNwcWcOp>(op) ||
isa<linalg::MatvecOp>(op) || isa<linalg::Mmt4DOp>(op) ||
isa<linalg::BatchMmt4DOp>(op) ||
Expand Down
84 changes: 84 additions & 0 deletions mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1725,3 +1725,87 @@ module attributes {transform.with_named_sequence} {
transform.yield
}
}

// -----

func.func @batch_matmul(%A: memref<?x?x?xf32>, %B: memref<?x?x?xf32>, %C: memref<?x?x?xf32>) {
linalg.batch_matmul ins(%A, %B: memref<?x?x?xf32>, memref<?x?x?xf32>)
outs(%C: memref<?x?x?xf32>)
return
}

// CHECK-LABEL: func.func @batch_matmul(
// CHECK-SAME: %[[A:.*]]: memref<?x?x?xf32>, %[[B:.*]]: memref<?x?x?xf32>, %[[C:.*]]: memref<?x?x?xf32>
// CHECK: %[[c0:.*]] = arith.constant 0 : index
Copy link
Contributor

@banach-space banach-space Dec 16, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Let's use CAPS for all LIT variables

Suggested change
// CHECK: %[[c0:.*]] = arith.constant 0 : index
// CHECK: %[[C0:.*]] = arith.constant 0 : index

// CHECK: %[[BATCH_DIM:.*]] = memref.dim %[[A]], %[[c0]] : memref<?x?x?xf32>
// CHECK: %[[c1:.*]] = arith.constant 1 : index
// CHECK: %[[M:.*]] = memref.dim %[[A]], %[[c1]] : memref<?x?x?xf32>
// CHECK: %[[c2:.*]] = arith.constant 2 : index
// CHECK: %[[N:.*]] = memref.dim %[[B]], %[[c2]] : memref<?x?x?xf32>
// CHECK: %[[c2_2:.*]] = arith.constant 2 : index
// CHECK: %[[K:.*]] = memref.dim %[[A]], %[[c2_2]] : memref<?x?x?xf32>
// CHECK: %[[c0_4:.*]] = arith.constant 0 : index
// CHECK: %[[P0:.*]] = ub.poison : f32
// CHECK: %[[MA:.*]] = vector.create_mask %[[BATCH_DIM]], %[[M]], %[[K]] : vector<4x8x4xi1>
// CHECK: %[[VA:.*]] = vector.mask %[[MA]] { vector.transfer_read %[[A]][%[[c0_4]], %[[c0_4]], %[[c0_4]]], %[[P0]] {in_bounds = [true, true, true, true], permutation_map = #{{.*}}} : memref<?x?x?xf32>, vector<4x8x16x4xf32> } : vector<4x8x4xi1> -> vector<4x8x16x4xf32>
// CHECK: %[[P1:.*]] = ub.poison : f32
// CHECK: %[[MB:.*]] = vector.create_mask %[[BATCH_DIM]], %[[K]], %[[N]] : vector<4x4x16xi1>
// CHECK: %[[VB:.*]] = vector.mask %[[MB]] { vector.transfer_read %[[B]][%[[c0_4]], %[[c0_4]], %[[c0_4]]], %[[P1]] {in_bounds = [true, true, true, true], permutation_map = #{{.*}}} : memref<?x?x?xf32>, vector<4x8x16x4xf32> } : vector<4x4x16xi1> -> vector<4x8x16x4xf32>
// CHECK: %[[P2:.*]] = ub.poison : f32
// CHECK: %[[MC:.*]] = vector.create_mask %[[BATCH_DIM]], %[[M]], %[[N]] : vector<4x8x16xi1>
// CHECK: %[[VC:.*]] = vector.mask %[[MC]] { vector.transfer_read %[[C]][%[[c0_4]], %[[c0_4]], %[[c0_4]]], %[[P2]] {in_bounds = [true, true, true]} : memref<?x?x?xf32>, vector<4x8x16xf32> } : vector<4x8x16xi1> -> vector<4x8x16xf32>
// CHECK: %[[MUL:.*]] = arith.mulf %[[VA]], %[[VB]] : vector<4x8x16x4xf32>
// CHECK: %[[MRED:.*]] = vector.create_mask %[[BATCH_DIM]], %[[M]], %[[N]], %[[K]] : vector<4x8x16x4xi1>
// CHECK: %[[RED:.*]] = vector.mask %[[MRED]] { vector.multi_reduction <add>, %[[MUL]], %[[VC]] [3] : vector<4x8x16x4xf32> to vector<4x8x16xf32> } : vector<4x8x16x4xi1> -> vector<4x8x16xf32>
// CHECK: %[[c0_5:.*]] = arith.constant 0 : index
// CHECK: vector.mask %[[MC]] { vector.transfer_write %[[RED]], %[[C]][%[[c0_5]], %[[c0_5]], %[[c0_5]]] {in_bounds = [true, true, true]} : vector<4x8x16xf32>, memref<?x?x?xf32> } : vector<4x8x16xi1>

module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%matmul = transform.structured.match ops{["linalg.batch_matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
transform.structured.vectorize %matmul vector_sizes [4, 8, 16, 4] : !transform.any_op
transform.yield
}
}

// -----

func.func @batch_matmul_scalable(%A: memref<?x?x?xf32>, %B: memref<?x?x?xf32>, %C: memref<?x?x?xf32>) {
linalg.batch_matmul ins(%A, %B: memref<?x?x?xf32>, memref<?x?x?xf32>)
outs(%C: memref<?x?x?xf32>)
return
}

// CHECK-LABEL: func.func @batch_matmul_scalable
// CHECK-SAME: (%[[A:.*]]: memref<?x?x?xf32>, %[[B:.*]]: memref<?x?x?xf32>, %[[C:.*]]: memref<?x?x?xf32>) {
// CHECK: %[[c0:.*]] = arith.constant 0 : index
// CHECK: %[[BATCH_DIM:.*]] = memref.dim %[[A]], %[[c0]] : memref<?x?x?xf32>
// CHECK: %[[c1:.*]] = arith.constant 1 : index
// CHECK: %[[M:.*]] = memref.dim %[[A]], %[[c1]] : memref<?x?x?xf32>
// CHECK: %[[c2:.*]] = arith.constant 2 : index
// CHECK: %[[N:.*]] = memref.dim %[[B]], %[[c2]] : memref<?x?x?xf32>
// CHECK: %[[c2_2:.*]] = arith.constant 2 : index
// CHECK: %[[K:.*]] = memref.dim %[[A]], %[[c2_2]] : memref<?x?x?xf32>
// CHECK: %[[c0_4:.*]] = arith.constant 0 : index
// CHECK: %[[P0:.*]] = ub.poison : f32
// CHECK: %[[MA:.*]] = vector.create_mask %[[BATCH_DIM]], %[[M]], %[[K]] : vector<4x8x4xi1>
// CHECK: %[[VA:.*]] = vector.mask %[[MA]] { vector.transfer_read %[[A]][%[[c0_4]], %[[c0_4]], %[[c0_4]]], %[[P0]] {in_bounds = [true, true, true, true], permutation_map = #{{.*}}} : memref<?x?x?xf32>, vector<4x8x[16]x4xf32> } : vector<4x8x4xi1> -> vector<4x8x[16]x4xf32>
// CHECK: %[[P1:.*]] = ub.poison : f32
// CHECK: %[[MB:.*]] = vector.create_mask %[[BATCH_DIM]], %[[K]], %[[N]] : vector<4x4x[16]xi1>
// CHECK: %[[VB:.*]] = vector.mask %[[MB]] { vector.transfer_read %[[B]][%[[c0_4]], %[[c0_4]], %[[c0_4]]], %[[P1]] {in_bounds = [true, true, true, true], permutation_map = #{{.*}}} : memref<?x?x?xf32>, vector<4x8x[16]x4xf32> } : vector<4x4x[16]xi1> -> vector<4x8x[16]x4xf32>
// CHECK: %[[P2:.*]] = ub.poison : f32
// CHECK: %[[MC:.*]] = vector.create_mask %[[BATCH_DIM]], %[[M]], %[[N]] : vector<4x8x[16]xi1>
// CHECK: %[[VC:.*]] = vector.mask %[[MC]] { vector.transfer_read %[[C]][%[[c0_4]], %[[c0_4]], %[[c0_4]]], %[[P2]] {in_bounds = [true, true, true]} : memref<?x?x?xf32>, vector<4x8x[16]xf32> } : vector<4x8x[16]xi1> -> vector<4x8x[16]xf32>
// CHECK: %[[MUL:.*]] = arith.mulf %[[VA]], %[[VB]] : vector<4x8x[16]x4xf32>
// CHECK: %[[MRED:.*]] = vector.create_mask %[[BATCH_DIM]], %[[M]], %[[N]], %[[K]] : vector<4x8x[16]x4xi1>
// CHECK: %[[RED:.*]] = vector.mask %[[MRED]] { vector.multi_reduction <add>, %[[MUL]], %[[VC]] [3] : vector<4x8x[16]x4xf32> to vector<4x8x[16]xf32> } : vector<4x8x[16]x4xi1> -> vector<4x8x[16]xf32>
// CHECK: %[[c0_5:.*]] = arith.constant 0 : index
// CHECK: vector.mask %[[MC]] { vector.transfer_write %[[RED]], %[[C]][%[[c0_5]], %[[c0_5]], %[[c0_5]]] {in_bounds = [true, true, true]} : vector<4x8x[16]xf32>, memref<?x?x?xf32> } : vector<4x8x[16]xi1>

module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%matmul = transform.structured.match ops{["linalg.batch_matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
transform.structured.vectorize %matmul vector_sizes [4, 8, [16], 4] : !transform.any_op
transform.yield
}
}