1// RUN: mlir-opt -split-input-file -test-fold-arith-extf-into-vector-contract-patterns %s | FileCheck %s 2 3 4// CHECK-DAG: #[[$map0:.*]] = affine_map<(d0, d1, d2) -> (d0, d2)> 5// CHECK-DAG: #[[$map1:.*]] = affine_map<(d0, d1, d2) -> (d2, d1)> 6// CHECK-DAG: #[[$map2:.*]] = affine_map<(d0, d1, d2) -> (d0, d1)> 7// CHECK-LABEL: func.func @fold_arith_extf_into_contract 8// CHECK-SAME: (%[[ARG0:.*]]: vector<64x64xf16>, %[[ARG1:.*]]: vector<64x64xf16>, %[[ARG2:.*]]: vector<64x64xf32>) 9// CHECK-NEXT: %[[R:.+]] = vector.contract {indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]], 10// CHECK-SAME: iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} 11// CHECK-SAME: %[[ARG0]], %[[ARG1]], %[[ARG2]] : vector<64x64xf16>, vector<64x64xf16> into vector<64x64xf32> 12// CHECK-NEXT: return %[[R]] : vector<64x64xf32> 13func.func @fold_arith_extf_into_contract( 14 %arg0: vector<64x64xf16>, 15 %arg1: vector<64x64xf16>, 16 %arg2: vector<64x64xf32>) -> vector<64x64xf32> { 17 %lhs_f32 = arith.extf %arg0 : vector<64x64xf16> to vector<64x64xf32> 18 %rhs_f32 = arith.extf %arg1 : vector<64x64xf16> to vector<64x64xf32> 19 %result = vector.contract { 20 indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], 21 iterator_types = ["parallel", "parallel", "reduction"], 22 kind = #vector.kind<add>} 23 %lhs_f32, %rhs_f32, %arg2 : vector<64x64xf32>, vector<64x64xf32> into vector<64x64xf32> 24 return %result : vector<64x64xf32> 25} 26 27// ----- 28 29// CHECK-DAG: #[[$map0:.*]] = affine_map<(d0, d1, d2) -> (d0, d2)> 30// CHECK-DAG: #[[$map1:.*]] = affine_map<(d0, d1, d2) -> (d2, d1)> 31// CHECK-DAG: #[[$map2:.*]] = affine_map<(d0, d1, d2) -> (d0, d1)> 32// CHECK-LABEL: func.func @fold_arith_extf_into_contract_scalable 33// CHECK-SAME: (%[[ARG0:.*]]: vector<[64]x64xf16>, %[[ARG1:.*]]: vector<64x64xf16>, %[[ARG2:.*]]: vector<[64]x64xf32>) 34// CHECK-NEXT: %[[R:.+]] = vector.contract {indexing_maps = [#[[$map0]], #[[$map1]], #[[$map2]]], 35// CHECK-SAME: iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} 36// CHECK-SAME: %[[ARG0]], %[[ARG1]], %[[ARG2]] : vector<[64]x64xf16>, vector<64x64xf16> into vector<[64]x64xf32> 37// CHECK-NEXT: return %[[R]] : vector<[64]x64xf32> 38func.func @fold_arith_extf_into_contract_scalable( 39 %arg0: vector<[64]x64xf16>, 40 %arg1: vector<64x64xf16>, 41 %arg2: vector<[64]x64xf32>) -> vector<[64]x64xf32> { 42 %lhs_f32 = arith.extf %arg0 : vector<[64]x64xf16> to vector<[64]x64xf32> 43 %rhs_f32 = arith.extf %arg1 : vector<64x64xf16> to vector<64x64xf32> 44 %result = vector.contract { 45 indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], 46 iterator_types = ["parallel", "parallel", "reduction"], 47 kind = #vector.kind<add>} 48 %lhs_f32, %rhs_f32, %arg2 : vector<[64]x64xf32>, vector<64x64xf32> into vector<[64]x64xf32> 49 return %result : vector<[64]x64xf32> 50} 51 52// ----- 53 54// CHECK-LABEL: func.func @fold_arith_extsi_into_contract 55// CHECK-SAME: (%[[ARG0:.*]]: vector<64x64xi8>, %[[ARG1:.*]]: vector<64x64xi8>, %[[ARG2:.*]]: vector<64x64xi32>) 56// CHECK-NEXT: %[[R:.+]] = vector.contract 57// CHECK-SAME: iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} 58// CHECK-SAME: %[[ARG0]], %[[ARG1]], %[[ARG2]] : vector<64x64xi8>, vector<64x64xi8> into vector<64x64xi32> 59// CHECK-NEXT: return %[[R]] : vector<64x64xi32> 60func.func @fold_arith_extsi_into_contract( 61 %arg0: vector<64x64xi8>, 62 %arg1: vector<64x64xi8>, 63 %arg2: vector<64x64xi32>) -> vector<64x64xi32> { 64 %lhs_i32 = arith.extsi %arg0 : vector<64x64xi8> to vector<64x64xi32> 65 %rhs_i32 = arith.extsi %arg1 : vector<64x64xi8> to vector<64x64xi32> 66 %result = vector.contract { 67 indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], 68 iterator_types = ["parallel", "parallel", "reduction"], 69 kind = #vector.kind<add>} 70 %lhs_i32, %rhs_i32, %arg2 : vector<64x64xi32>, vector<64x64xi32> into vector<64x64xi32> 71 return %result : vector<64x64xi32> 72} 73