Skip to content

[MLIR][AArch64] Add an extra test for Neon I8MM (NFC) #135777

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 3 commits into from
Jun 3, 2025
Merged

[MLIR][AArch64] Add an extra test for Neon I8MM (NFC) #135777

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9cc437b
[MLIR][AArch64] Add an extra test for Neon I8MM
momchil-velikov Apr 15, 2025
2ff98bb
[fixup] Add a comment explaining the added test
momchil-velikov May 14, 2025
e4e0ba9
[fixup] Some more comments
momchil-velikov May 15, 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
106 changes: 106 additions & 0 deletions mlir/test/Dialect/ArmNeon/lower-to-arm-neon.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -355,6 +355,112 @@ func.func @vector_arm_neon_k_unroll_vecmat(%lhs: vector<1x32xi8>, %rhs: vector<2
return %res : vector<1x2xi32>
}

// -----

// Test with more than one iteration across all of the M, N and K dimensions.
// Shows multiple independent accumulators as well as accumulator reuse.
// For each iterarion [I, J, K] sub-tiles are extracted from offsets as follows:
// LHS: [2*I, 8*K]
// RHS: [2*J, 8*K]
// ACC: [2*I, 2*J]
// Sub-tile insert offsets for the result are as like ACC (there are redundant
// inserts).

// CHECK-LABEL: @vector_arm_neon_mnk_unroll
// CHECK-SAME: %[[LHS:arg0]]: vector<4x16xi8>
// CHECK-SAME: %[[RHS:arg1]]: vector<4x16xi8>
// CHECK-SAME: %[[ACC:arg2]]: vector<4x4xi32>
// CHECK-SAME: -> vector<4x4xi32> {
// CHECK-NEXT: %cst = arith.constant dense<0> : vector<4x4xi32>

// Iteration: [0, 0, 0]
// CHECK-NEXT: %[[VAL_0:[0-9]+]] = vector.extract_strided_slice %[[LHS]] {offsets = [0, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_1:[0-9]+]] = vector.extract_strided_slice %[[RHS]] {offsets = [0, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_2:[0-9]+]] = vector.extract_strided_slice %[[ACC]] {offsets = [0, 0], sizes = [2, 2], strides = [1, 1]} : vector<4x4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_3:[0-9]+]] = vector.shape_cast %[[VAL_0]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_4:[0-9]+]] = vector.shape_cast %[[VAL_1]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_5:[0-9]+]] = vector.shape_cast %[[VAL_2]] : vector<2x2xi32> to vector<4xi32>
// CHECK-NEXT: %[[KACC_0_v0:[0-9]+]] = arm_neon.intr.smmla %[[VAL_5]], %[[VAL_3]], %[[VAL_4]] : vector<16xi8> to vector<4xi32>
// CHECK-NEXT: %[[VAL_7:[0-9]+]] = vector.shape_cast %[[KACC_0_v0]] : vector<4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_8:[0-9]+]] = vector.insert_strided_slice %[[VAL_7]], %cst {offsets = [0, 0], strides = [1, 1]} : vector<2x2xi32> into vector<4x4xi32>

// Iteration: [0, 0, 1]
// CHECK-NEXT: %[[VAL_9:[0-9]+]] = vector.extract_strided_slice %[[LHS]] {offsets = [0, 8], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_10:[0-9]+]] = vector.extract_strided_slice %[[RHS]] {offsets = [0, 8], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_11:[0-9]+]] = vector.shape_cast %[[VAL_9]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_12:[0-9]+]] = vector.shape_cast %[[VAL_10]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[KACC_0_v1:[0-9]+]] = arm_neon.intr.smmla %[[KACC_0_v0]], %[[VAL_11]], %[[VAL_12]] : vector<16xi8> to vector<4xi32>
// CHECK-NEXT: %[[VAL_14:[0-9]+]] = vector.shape_cast %[[KACC_0_v1]] : vector<4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_15:[0-9]+]] = vector.insert_strided_slice %[[VAL_14]], %[[VAL_8]] {offsets = [0, 0], strides = [1, 1]} : vector<2x2xi32> into vector<4x4xi32>

// Iteration: [0, 1, 0]
// CHECK-NEXT: %[[VAL_16:[0-9]+]] = vector.extract_strided_slice %[[LHS]] {offsets = [0, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_17:[0-9]+]] = vector.extract_strided_slice %[[RHS]] {offsets = [2, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_18:[0-9]+]] = vector.extract_strided_slice %[[ACC]] {offsets = [0, 2], sizes = [2, 2], strides = [1, 1]} : vector<4x4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_19:[0-9]+]] = vector.shape_cast %[[VAL_16]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_20:[0-9]+]] = vector.shape_cast %[[VAL_17]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_21:[0-9]+]] = vector.shape_cast %[[VAL_18]] : vector<2x2xi32> to vector<4xi32>
// CHECK-NEXT: %[[KACC_1_v0:[0-9]+]] = arm_neon.intr.smmla %[[VAL_21]], %[[VAL_19]], %[[VAL_20]] : vector<16xi8> to vector<4xi32>
// CHECK-NEXT: %[[VAL_23:[0-9]+]] = vector.shape_cast %[[KACC_1_v0]] : vector<4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_24:[0-9]+]] = vector.insert_strided_slice %[[VAL_23]], %[[VAL_15]] {offsets = [0, 2], strides = [1, 1]} : vector<2x2xi32> into vector<4x4xi32>

// Iteration: [0, 1, 1]
// CHECK-NEXT: %[[VAL_25:[0-9]+]] = vector.extract_strided_slice %[[LHS]] {offsets = [0, 8], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_26:[0-9]+]] = vector.extract_strided_slice %[[RHS]] {offsets = [2, 8], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_27:[0-9]+]] = vector.shape_cast %[[VAL_25]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_28:[0-9]+]] = vector.shape_cast %[[VAL_26]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[KACC_1_v1:[0-9]+]] = arm_neon.intr.smmla %[[KACC_1_v0]], %[[VAL_27]], %[[VAL_28]] : vector<16xi8> to vector<4xi32>
// CHECK-NEXT: %[[VAL_30:[0-9]+]] = vector.shape_cast %[[KACC_1_v1]] : vector<4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_31:[0-9]+]] = vector.insert_strided_slice %[[VAL_30]], %[[VAL_24]] {offsets = [0, 2], strides = [1, 1]} : vector<2x2xi32> into vector<4x4xi32>

// Iteration: [1, 0, 0]
// CHECK-NEXT: %[[VAL_32:[0-9]+]] = vector.extract_strided_slice %[[LHS]] {offsets = [2, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_33:[0-9]+]] = vector.extract_strided_slice %[[RHS]] {offsets = [0, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_34:[0-9]+]] = vector.extract_strided_slice %[[ACC]] {offsets = [2, 0], sizes = [2, 2], strides = [1, 1]} : vector<4x4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_35:[0-9]+]] = vector.shape_cast %[[VAL_32]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_36:[0-9]+]] = vector.shape_cast %[[VAL_33]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_37:[0-9]+]] = vector.shape_cast %[[VAL_34]] : vector<2x2xi32> to vector<4xi32>
// CHECK-NEXT: %[[KACC_2_v0:[0-9]+]] = arm_neon.intr.smmla %[[VAL_37]], %[[VAL_35]], %[[VAL_36]] : vector<16xi8> to vector<4xi32>
// CHECK-NEXT: %[[VAL_39:[0-9]+]] = vector.shape_cast %[[KACC_2_v0]] : vector<4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_40:[0-9]+]] = vector.insert_strided_slice %[[VAL_39]], %[[VAL_31]] {offsets = [2, 0], strides = [1, 1]} : vector<2x2xi32> into vector<4x4xi32>

// Iteration: [1, 0, 1]
// CHECK-NEXT: %[[VAL_41:[0-9]+]] = vector.extract_strided_slice %[[LHS]] {offsets = [2, 8], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_42:[0-9]+]] = vector.extract_strided_slice %[[RHS]] {offsets = [0, 8], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_43:[0-9]+]] = vector.shape_cast %[[VAL_41]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_44:[0-9]+]] = vector.shape_cast %[[VAL_42]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[KACC_2_v1:[0-9]+]] = arm_neon.intr.smmla %[[KACC_2_v0]], %[[VAL_43]], %[[VAL_44]] : vector<16xi8> to vector<4xi32>
// CHECK-NEXT: %[[VAL_46:[0-9]+]] = vector.shape_cast %[[KACC_2_v1]] : vector<4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_47:[0-9]+]] = vector.insert_strided_slice %[[VAL_46]], %[[VAL_40]] {offsets = [2, 0], strides = [1, 1]} : vector<2x2xi32> into vector<4x4xi32>

// Iteration: [1, 1, 0]
// CHECK-NEXT: %[[VAL_48:[0-9]+]] = vector.extract_strided_slice %[[LHS]] {offsets = [2, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_49:[0-9]+]] = vector.extract_strided_slice %[[RHS]] {offsets = [2, 0], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_50:[0-9]+]] = vector.extract_strided_slice %[[ACC]] {offsets = [2, 2], sizes = [2, 2], strides = [1, 1]} : vector<4x4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_51:[0-9]+]] = vector.shape_cast %[[VAL_48]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_52:[0-9]+]] = vector.shape_cast %[[VAL_49]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_53:[0-9]+]] = vector.shape_cast %[[VAL_50]] : vector<2x2xi32> to vector<4xi32>
// CHECK-NEXT: %[[KACC_3_v0:[0-9]+]] = arm_neon.intr.smmla %[[VAL_53]], %[[VAL_51]], %[[VAL_52]] : vector<16xi8> to vector<4xi32>
// CHECK-NEXT: %[[VAL_55:[0-9]+]] = vector.shape_cast %[[KACC_3_v0]] : vector<4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_56:[0-9]+]] = vector.insert_strided_slice %[[VAL_55]], %[[VAL_47]] {offsets = [2, 2], strides = [1, 1]} : vector<2x2xi32> into vector<4x4xi32>

// Iteration: [1, 1, 1]
// CHECK-NEXT: %[[VAL_57:[0-9]+]] = vector.extract_strided_slice %[[LHS]] {offsets = [2, 8], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_58:[0-9]+]] = vector.extract_strided_slice %[[RHS]] {offsets = [2, 8], sizes = [2, 8], strides = [1, 1]} : vector<4x16xi8> to vector<2x8xi8>
// CHECK-NEXT: %[[VAL_59:[0-9]+]] = vector.shape_cast %[[VAL_57]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[VAL_60:[0-9]+]] = vector.shape_cast %[[VAL_58]] : vector<2x8xi8> to vector<16xi8>
// CHECK-NEXT: %[[KACC_3_v1:[0-9]+]] = arm_neon.intr.smmla %[[KACC_3_v0]], %[[VAL_59]], %[[VAL_60]] : vector<16xi8> to vector<4xi32>
// CHECK-NEXT: %[[VAL_62:[0-9]+]] = vector.shape_cast %[[KACC_3_v1]] : vector<4xi32> to vector<2x2xi32>
// CHECK-NEXT: %[[VAL_63:[0-9]+]] = vector.insert_strided_slice %[[VAL_62]], %[[VAL_56]] {offsets = [2, 2], strides = [1, 1]} : vector<2x2xi32> into vector<4x4xi32>

// CHECK-NEXT: return %[[VAL_63]] : vector<4x4xi32>
func.func @vector_arm_neon_mnk_unroll(%lhs: vector<4x16xi8>, %rhs: vector<4x16xi8>, %acc : vector<4x4xi32>) -> vector<4x4xi32> {
%lhs_extsi = arith.extsi %lhs : vector<4x16xi8> to vector<4x16xi32>
%rhs_extsi = arith.extsi %rhs : vector<4x16xi8> to vector<4x16xi32>
%res = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %lhs_extsi, %rhs_extsi, %acc : vector<4x16xi32>, vector<4x16xi32> into vector<4x4xi32>
return %res : vector<4x4xi32>
}

module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%module: !transform.any_op {transform.readonly}) {
%func = transform.structured.match ops{["func.func"]} in %module : (!transform.any_op) -> !transform.op<"func.func">
Expand Down
Loading