2004 lines · plain
1// RUN: mlir-opt %s -transform-interpreter -split-input-file | FileCheck %s2 3///----------------------------------------------------------------------------------------4/// Tests for linalg.dot5///----------------------------------------------------------------------------------------6 7// CHECK-LABEL: contraction_dot8func.func @contraction_dot(%A: memref<1584xf32>, %B: memref<1584xf32>, %C: memref<f32>) {9 10// CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584xf32>11// CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [0] : vector<1584xf32> to f3212 linalg.dot ins(%A, %B: memref<1584xf32>, memref<1584xf32>)13 outs(%C: memref<f32>)14 return15}16 17module attributes {transform.with_named_sequence} {18 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {19 %0 = transform.structured.match ops{["linalg.dot"]} in %arg1 : (!transform.any_op) -> !transform.any_op20 transform.structured.vectorize %0 : !transform.any_op21 transform.yield22 }23}24 25// -----26 27///----------------------------------------------------------------------------------------28/// Tests for linalg.matvec29///----------------------------------------------------------------------------------------30 31// CHECK-LABEL: contraction_matvec32func.func @contraction_matvec(%A: memref<1584x1584xf32>, %B: memref<1584xf32>, %C: memref<1584xf32>) {33 34// CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584x1584xf32>35// CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [1] : vector<1584x1584xf32> to vector<1584xf32>36 linalg.matvec ins(%A, %B: memref<1584x1584xf32>, memref<1584xf32>)37 outs(%C: memref<1584xf32>)38 return39}40 41module attributes {transform.with_named_sequence} {42 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {43 %0 = transform.structured.match ops{["linalg.matvec"]} in %arg1 : (!transform.any_op) -> !transform.any_op44 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op45 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns } : (!transform.any_op) -> !transform.any_op46 transform.yield47 }48}49 50// -----51 52///----------------------------------------------------------------------------------------53/// Tests for linalg.matmul54///----------------------------------------------------------------------------------------55 56// CHECK-LABEL: contraction_matmul57func.func @contraction_matmul(%A: memref<1584x1584xf32>, %B: memref<1584x1584xf32>, %C: memref<1584x1584xf32>) {58// CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584x1584x1584xf32>59// CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [2] : vector<1584x1584x1584xf32> to vector<1584x1584xf32>60 linalg.matmul ins(%A, %B: memref<1584x1584xf32>, memref<1584x1584xf32>)61 outs(%C: memref<1584x1584xf32>)62 return63}64 65module attributes {transform.with_named_sequence} {66 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {67 %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op68 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op69 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns } : (!transform.any_op) -> !transform.any_op70 transform.yield71 }72}73 74// -----75 76// CHECK-LABEL: @float_mixed_precision_matmul77// CHECK-COUNT-3: vector.transfer_read78// CHECK-NOT: arith.extf79// CHECK: vector.contract {{.*}} : vector<1584x1584xbf16>, vector<1584x1584xbf16> into vector<1584x1584xf32>80func.func @float_mixed_precision_matmul(%A: memref<1584x1584xbf16>, %B: memref<1584x1584xbf16>, %C: memref<1584x1584xf32>) {81 linalg.matmul ins(%A, %B: memref<1584x1584xbf16>, memref<1584x1584xbf16>)82 outs(%C: memref<1584x1584xf32>)83 return84}85 86module attributes {transform.with_named_sequence} {87 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {88 %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op89 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op90 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { fold_type_extensions_into_contract } : (!transform.any_op) -> !transform.any_op91 transform.yield92 }93}94 95// -----96 97// CHECK-LABEL: func @vectorization_test_298func.func @vectorization_test_2(%A: memref<8x16xf32>, %B: memref<16x32xf32>,99 %C: memref<8x32xf32>) {100 // CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<8x32x16xf32>101 // CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [2] : vector<8x32x16xf32> to vector<8x32xf32>102 linalg.matmul103 ins(%A, %B: memref<8x16xf32>, memref<16x32xf32>)104 outs(%C: memref<8x32xf32>)105 return106}107 108module attributes {transform.with_named_sequence} {109 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {110 %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op111 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op112 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns } : (!transform.any_op) -> !transform.any_op113 transform.yield114 }115}116 117// -----118 119// CHECK-LABEL: func @matmul_tensors120// CHECK-SAME: (%[[ARG0:.*]]: tensor<8x4xf32>, %[[ARG1:.*]]: tensor<4x12xf32>,121// CHECK-SAME: %[[ARG2:.*]]: tensor<8x12xf32>) -> tensor<8x12xf32>122func.func @matmul_tensors(123 %arg0: tensor<8x4xf32>, %arg1: tensor<4x12xf32>, %arg2: tensor<8x12xf32>)124 -> tensor<8x12xf32> {125 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index126 // CHECK-DAG: %[[V0:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<8x4xf32>, vector<8x12x4xf32>127 // CHECK-DAG: %[[V1:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x12xf32>, vector<8x12x4xf32>128 // CHECK-DAG: %[[V2:.*]] = vector.transfer_read %[[ARG2]][%[[C0]], %[[C0]]], {{.*}} : tensor<8x12xf32>, vector<8x12xf32>129 //130 // linalg matmul lowers gets expanded to a 3D reduction, canonicalization later131 // convert it to a 2D contract.132 // CHECK: %[[MUL:.*]] = arith.mulf %[[V0]], %[[V1]] : vector<8x12x4xf32>133 // CHECK: %[[R:.*]] = vector.multi_reduction <add>, %[[MUL]], %[[V2]] [2] : vector<8x12x4xf32> to vector<8x12xf32>134 // CHECK: %[[W:.*]] = vector.transfer_write %[[R]], %[[ARG2]][%[[C0]], %[[C0]]] {in_bounds = [true, true]} : vector<8x12xf32>, tensor<8x12xf32>135 %0 = linalg.matmul ins(%arg0, %arg1: tensor<8x4xf32>, tensor<4x12xf32>)136 outs(%arg2: tensor<8x12xf32>)137 -> tensor<8x12xf32>138 // CHECK: return %[[W]] : tensor<8x12xf32>139 return %0 : tensor<8x12xf32>140}141 142module attributes {transform.with_named_sequence} {143 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {144 %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op145 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op146 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op147 transform.yield148 }149}150 151// -----152 153///----------------------------------------------------------------------------------------154/// Tests for linalg.batch_matmul155///----------------------------------------------------------------------------------------156 157// CHECK-LABEL: contraction_batch_matmul158func.func @contraction_batch_matmul(%A: memref<1584x1584x1584xf32>, %B: memref<1584x1584x1584xf32>, %C: memref<1584x1584x1584xf32>) {159// CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584x1584x1584x1584xf32>160// CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [3] : vector<1584x1584x1584x1584xf32> to vector<1584x1584x1584xf32>161 linalg.batch_matmul162 ins(%A, %B: memref<1584x1584x1584xf32>, memref<1584x1584x1584xf32>)163 outs(%C: memref<1584x1584x1584xf32>)164 return165}166 167module attributes {transform.with_named_sequence} {168 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {169 %0 = transform.structured.match ops{["linalg.batch_matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op170 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op171 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns } : (!transform.any_op) -> !transform.any_op172 transform.yield173 }174}175 176// -----177 178///----------------------------------------------------------------------------------------179/// Tests for linalg.cantract180///----------------------------------------------------------------------------------------181 182// CHECK-LABEL: @matmul_as_contract183// CHECK-SAME: %[[A:.*]]: tensor<24x12xf32>184// CHECK-SAME: %[[B:.*]]: tensor<12x25xf32>185// CHECK-SAME: %[[C:.*]]: tensor<24x25xf32>186func.func @matmul_as_contract(%A: tensor<24x12xf32>,187 %B: tensor<12x25xf32>,188 %C: tensor<24x25xf32>) -> tensor<24x25xf32> {189 // CHECK: %[[vA:.+]] = vector.transfer_read %[[A]]190 // CHECK: %[[vB:.+]] = vector.transfer_read %[[B]]191 // CHECK: %[[vC:.+]] = vector.transfer_read %[[C]]192 // CHECK: %[[vR:.+]] = vector.contract {{.*}} %[[vA]], %[[vB]], %[[vC]]193 // CHECK: vector.transfer_write %[[vR]], %[[C]]194 %0 = linalg.contract195 indexing_maps = [affine_map<(m, n, k) -> (m, k)>,196 affine_map<(m, n, k) -> (k, n)>,197 affine_map<(m, n, k) -> (m, n)>]198 ins(%A, %B : tensor<24x12xf32>, tensor<12x25xf32>)199 outs(%C : tensor<24x25xf32>) -> tensor<24x25xf32>200 func.return %0 : tensor<24x25xf32>201}202 203module attributes {transform.with_named_sequence} {204 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {205 %0 = transform.structured.match ops{["linalg.contract"]} in %arg1 : (!transform.any_op) -> !transform.any_op206 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op207 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op208 // TODO: also tests the other available vectorization strategies209 transform.yield210 }211}212 213// -----214 215// CHECK-LABEL: @float_mixed_precision_matmul_as_contract216// CHECK-COUNT-3: vector.transfer_read217// CHECK-NOT: arith.extf218// CHECK: vector.contract {{.*}} : vector<24x12xbf16>, vector<12x25xbf16> into vector<24x25xf32>219// CHECK: vector.transfer_write220func.func @float_mixed_precision_matmul_as_contract(%A: tensor<24x12xbf16>,221 %B: tensor<12x25xbf16>,222 %C: tensor<24x25xf32>) -> tensor<24x25xf32> {223 %0 = linalg.contract224 indexing_maps = [affine_map<(m, n, k) -> (m, k)>,225 affine_map<(m, n, k) -> (k, n)>,226 affine_map<(m, n, k) -> (m, n)>]227 ins(%A, %B : tensor<24x12xbf16>, tensor<12x25xbf16>)228 outs(%C : tensor<24x25xf32>) -> tensor<24x25xf32>229 func.return %0 : tensor<24x25xf32>230}231 232module attributes {transform.with_named_sequence} {233 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {234 %0 = transform.structured.match ops{["linalg.contract"]} in %arg1 : (!transform.any_op) -> !transform.any_op235 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op236 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { fold_type_extensions_into_contract } : (!transform.any_op) -> !transform.any_op237 transform.yield238 }239}240 241// -----242 243///----------------------------------------------------------------------------------------244/// Tests for linalg.fill245///----------------------------------------------------------------------------------------246 247// CHECK-LABEL: func @test_vectorize_fill248func.func @test_vectorize_fill(%A : memref<8x16xf32>, %arg0 : f32) {249 // CHECK: %[[V:.*]] = vector.broadcast {{.*}} : f32 to vector<8x16xf32>250 // CHECK: vector.transfer_write %[[V]], {{.*}} : vector<8x16xf32>, memref<8x16xf32>251 linalg.fill ins(%arg0 : f32) outs(%A : memref<8x16xf32>)252 return253}254 255module attributes {transform.with_named_sequence} {256 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {257 %0 = transform.structured.match ops{["linalg.fill"]} in %arg1 : (!transform.any_op) -> !transform.any_op258 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op259 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op260 transform.yield261 }262}263 264// -----265 266// CHECK-LABEL: func @test_vectorize_fill267func.func @test_vectorize_fill_0d(%A : memref<f32>, %arg0 : f32) {268 // CHECK-SAME: (%[[M:.*]]: memref<f32>, %[[val:.*]]: f32)269 // CHECK: %[[VEC:.*]] = vector.broadcast %[[val]] : f32 to vector<f32>270 // CHECK: vector.transfer_write %[[VEC]], %[[M]][] : vector<f32>, memref<f32>271 linalg.fill ins(%arg0 : f32) outs(%A : memref<f32>)272 return273}274 275module attributes {transform.with_named_sequence} {276 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {277 %0 = transform.structured.match ops{["linalg.fill"]} in %arg1 : (!transform.any_op) -> !transform.any_op278 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op279 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op280 transform.yield281 }282}283 284// -----285 286///----------------------------------------------------------------------------------------287/// Tests for linalg.pack288///289/// TODO: Add similar tests for linalg.unpack290///----------------------------------------------------------------------------------------291 292// Note, see a similar test in:293// * vectorization.mlir.294 295func.func @pack_no_padding(%arg0: tensor<32x8x16xf32>, %arg1: tensor<4x1x32x16x2xf32>) -> tensor<4x1x32x16x2xf32> {296 %pack = linalg.pack %arg0 outer_dims_perm = [1, 2, 0] inner_dims_pos = [2, 1] inner_tiles = [16, 2] into %arg1 : tensor<32x8x16xf32> -> tensor<4x1x32x16x2xf32>297 return %pack : tensor<4x1x32x16x2xf32>298}299 300module attributes {transform.with_named_sequence} {301 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {302 %0 = transform.structured.match ops{["linalg.pack"]} in %arg0 : (!transform.any_op) -> !transform.any_op303 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op304 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op305 transform.yield306 }307}308 309// CHECK-LABEL: func.func @pack_no_padding(310// CHECK-SAME: %[[VAL_0:.*]]: tensor<32x8x16xf32>,311// CHECK-SAME: %[[VAL_1:.*]]: tensor<4x1x32x16x2xf32>) -> tensor<4x1x32x16x2xf32> {312// CHECK-DAG: %[[VAL_2:.*]] = ub.poison : f32313// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index314// CHECK: %[[VAL_4:.*]] = vector.transfer_read %[[VAL_0]]{{\[}}%[[VAL_3]], %[[VAL_3]], %[[VAL_3]]], %[[VAL_2]] {in_bounds = [true, true, true]} : tensor<32x8x16xf32>, vector<32x8x16xf32>315// CHECK: %[[VAL_5:.*]] = vector.shape_cast %[[VAL_4]] : vector<32x8x16xf32> to vector<32x4x2x1x16xf32>316// CHECK: %[[VAL_6:.*]] = vector.transpose %[[VAL_5]], [1, 3, 0, 4, 2] : vector<32x4x2x1x16xf32> to vector<4x1x32x16x2xf32>317// CHECK: %[[VAL_8:.*]] = vector.transfer_write %[[VAL_6]], %[[VAL_1]]{{\[}}%[[VAL_3]], %[[VAL_3]], %[[VAL_3]], %[[VAL_3]], %[[VAL_3]]] {in_bounds = [true, true, true, true, true]} : vector<4x1x32x16x2xf32>, tensor<4x1x32x16x2xf32>318// CHECK: return %[[VAL_8]] : tensor<4x1x32x16x2xf32>319 320// -----321 322// Note, see a similar test in:323// * vectorization.mlir.324 325func.func @pack_with_padding(%arg0: tensor<32x7x15xf32>, %arg1: tensor<32x4x1x16x2xf32>) -> tensor<32x4x1x16x2xf32> {326 %pad = arith.constant 0.000000e+00 : f32327 %pack = linalg.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [2, 1] inner_tiles = [16, 2] into %arg1 : tensor<32x7x15xf32> -> tensor<32x4x1x16x2xf32>328 return %pack : tensor<32x4x1x16x2xf32>329}330 331// CHECK-LABEL: func.func @pack_with_padding(332// CHECK-SAME: %[[VAL_0:.*]]: tensor<32x7x15xf32>,333// CHECK-SAME: %[[VAL_1:.*]]: tensor<32x4x1x16x2xf32>) -> tensor<32x4x1x16x2xf32> {334// CHECK: %[[VAL_2:.*]] = arith.constant 0.000000e+00 : f32335// CHECK: %[[VAL_3:.*]] = arith.constant 0 : index336// CHECK: %[[VAL_4:.*]] = vector.transfer_read %[[VAL_0]]{{\[}}%[[VAL_3]], %[[VAL_3]], %[[VAL_3]]], %[[VAL_2]] {in_bounds = [true, false, false]} : tensor<32x7x15xf32>, vector<32x8x16xf32>337// CHECK: %[[VAL_5:.*]] = vector.shape_cast %[[VAL_4]] : vector<32x8x16xf32> to vector<32x4x2x1x16xf32>338// CHECK: %[[VAL_6:.*]] = vector.transpose %[[VAL_5]], [0, 1, 3, 4, 2] : vector<32x4x2x1x16xf32> to vector<32x4x1x16x2xf32>339// CHECK: %[[VAL_8:.*]] = vector.transfer_write %[[VAL_6]], %[[VAL_1]]{{\[}}%[[VAL_3]], %[[VAL_3]], %[[VAL_3]], %[[VAL_3]], %[[VAL_3]]] {in_bounds = [true, true, true, true, true]} : vector<32x4x1x16x2xf32>, tensor<32x4x1x16x2xf32>340// CHECK: return %[[VAL_8]] : tensor<32x4x1x16x2xf32>341 342module attributes {transform.with_named_sequence} {343 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {344 %0 = transform.structured.match ops{["linalg.pack"]} in %arg0 : (!transform.any_op) -> !transform.any_op345 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op346 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op347 transform.yield348 }349}350 351// -----352 353///----------------------------------------------------------------------------------------354/// Tests for linalg.map355///----------------------------------------------------------------------------------------356 357func.func @vectorize_map(%arg0: memref<64xf32>,358 %arg1: memref<64xf32>, %arg2: memref<64xf32>) {359 linalg.map ins(%arg0, %arg1 : memref<64xf32>, memref<64xf32>)360 outs(%arg2 : memref<64xf32>)361 (%in: f32, %in_0: f32, %out: f32) {362 %0 = arith.addf %in, %in_0 : f32363 linalg.yield %0 : f32364 }365 return366}367// CHECK-LABEL: func @vectorize_map368// CHECK: %[[LHS:.*]] = vector.transfer_read369// CHECK-NEXT: %[[RHS:.*]] = vector.transfer_read370// CHECK-NEXT: arith.addf %[[LHS]], %[[RHS]] : vector<64xf32>371 372module attributes {transform.with_named_sequence} {373 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {374 %0 = transform.structured.match ops{["linalg.map"]} in %arg1 : (!transform.any_op) -> !transform.any_op375 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op376 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op377 transform.yield378 }379}380 381// -----382 383///----------------------------------------------------------------------------------------384/// Tests for linalg.transpose385///----------------------------------------------------------------------------------------386 387func.func @vectorize_transpose(%arg0: memref<16x32x64xf32>,388 %arg1: memref<32x64x16xf32>) {389 linalg.transpose ins(%arg0 : memref<16x32x64xf32>)390 outs(%arg1 : memref<32x64x16xf32>) permutation = [1, 2, 0]391 return392}393// CHECK-LABEL: func @vectorize_transpose394// CHECK: vector.transpose395// CHECK-SAME: [1, 2, 0] : vector<16x32x64xf32> to vector<32x64x16xf32>396 397module attributes {transform.with_named_sequence} {398 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {399 %0 = transform.structured.match ops{["linalg.transpose"]} in %arg1 : (!transform.any_op) -> !transform.any_op400 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op401 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op402 transform.yield403 }404}405 406// -----407 408///----------------------------------------------------------------------------------------409/// Tests for linalg.reduce410///----------------------------------------------------------------------------------------411 412func.func @vectorize_reduce(%arg0: memref<16x32x64xf32>,413 %arg1: memref<16x64xf32>) {414 linalg.reduce ins(%arg0 : memref<16x32x64xf32>)415 outs(%arg1 : memref<16x64xf32>) dimensions = [1]416 (%in: f32, %init: f32) {417 %0 = arith.addf %in, %init : f32418 linalg.yield %0 : f32419 }420 return421}422// CHECK-LABEL: func @vectorize_reduce423// CHECK: vector.multi_reduction <add>424// CHECK-SAME: : vector<16x32x64xf32> to vector<16x64xf32>425 426module attributes {transform.with_named_sequence} {427 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {428 %0 = transform.structured.match ops{["linalg.reduce"]} in %arg1 : (!transform.any_op) -> !transform.any_op429 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op430 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op431 transform.yield432 }433}434 435// -----436 437///----------------------------------------------------------------------------------------438/// Tests for linalg.generic439///----------------------------------------------------------------------------------------440 441#matmul_trait = {442 indexing_maps = [443 affine_map<(m, n, k) -> (m, k)>,444 affine_map<(m, n, k) -> (k, n)>,445 affine_map<(m, n, k) -> (m, n)>446 ],447 iterator_types = ["parallel", "parallel", "reduction"]448}449 450// CHECK-LABEL: func @vectorization_test451func.func @vectorization_test(%A: memref<8x16xf32>, %B: memref<16x32xf32>,452 %C: memref<8x32xf32>) {453 // CHECK: vector.transfer_read %{{.*}} : memref<8x16xf32>, vector<8x32x16xf32>454 // CHECK: vector.transfer_read %{{.*}} : memref<16x32xf32>, vector<8x32x16xf32>455 // CHECK: %[[ACC:.*]] = vector.transfer_read %{{.*}} : memref<8x32xf32>, vector<8x32xf32>456 // CHECK: %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} : vector<8x32x16xf32>457 // CHECK: %[[R:.*]] = vector.multi_reduction <add>, %[[MUL]], %[[ACC]] [2] : vector<8x32x16xf32> to vector<8x32xf32>458 // CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xf32>, memref<8x32xf32>459 linalg.generic #matmul_trait460 ins(%A, %B : memref<8x16xf32>, memref<16x32xf32>)461 outs(%C : memref<8x32xf32>) {462 ^bb(%a: f32, %b: f32, %c: f32) :463 %d = arith.mulf %a, %b: f32464 %e = arith.addf %c, %d: f32465 linalg.yield %e : f32466 }467 return468}469 470module attributes {transform.with_named_sequence} {471 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {472 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op473 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op474 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op475 transform.yield476 }477}478 479// -----480 481#map = affine_map<() -> ()>482 483// CHECK-LABEL: func.func @generic_0d(484// CHECK-SAME: %[[ARG_0:.*]]: tensor<f32>, %[[ARG_1:.*]]: tensor<f32>, %[[ARG_2:.*]]: tensor<f32>)485func.func @generic_0d(%arg0: tensor<f32>, %arg1: tensor<f32>,486 %arg2: tensor<f32>) -> tensor<f32> {487// CHECK: %[[PAD:.*]] = ub.poison : f32488// CHECK: %[[READ_0:.*]] = vector.transfer_read %[[ARG_0]][], %[[PAD]] : tensor<f32>, vector<f32>489// CHECK: %[[ARG_0_AS_SCALAR:.*]] = vector.extract %[[READ_0]][] : f32 from vector<f32>490// CHECK: %[[READ_1:.*]] = vector.transfer_read %[[ARG_1]][], %[[PAD]] : tensor<f32>, vector<f32>491// CHECK: %[[ARG_1_AS_SCALAR:.*]] = vector.extract %[[READ_1]][] : f32 from vector<f32>492// CHECK: %[[READ_2:.*]] = vector.transfer_read %[[ARG_2]][], %[[PAD]] : tensor<f32>, vector<f32>493// CHECK: %[[ARG_2_AS_SCALAR:.*]] = vector.extract %[[READ_2]][] : f32 from vector<f32>494// CHECK: %[[MULF:.*]] = arith.mulf %[[ARG_0_AS_SCALAR]], %[[ARG_1_AS_SCALAR]] : f32495// CHECK: %[[ADDF:.*]] = arith.addf %[[ARG_2_AS_SCALAR]], %[[MULF]] : f32496// CHECK: %[[ADDF_BCAST:.*]] = vector.broadcast %[[ADDF]] : f32 to vector<f32>497// CHECK: vector.transfer_write %[[ADDF_BCAST]], %[[ARG_2]][] : vector<f32>, tensor<f32>498 %res = linalg.generic {499 indexing_maps = [#map, #map, #map],500 iterator_types = []501 } ins(%arg0, %arg1 : tensor<f32>, tensor<f32>)502 outs(%arg2 : tensor<f32>) {503 ^bb(%a: f32, %b: f32, %c: f32) :504 %d = arith.mulf %a, %b: f32505 %e = arith.addf %c, %d: f32506 linalg.yield %e : f32507 } -> tensor<f32>508 509 return %res : tensor<f32>510}511 512module attributes {transform.with_named_sequence} {513 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {514 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op515 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op516 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op517 transform.yield518 }519}520 521// -----522 523#matmul_transpose_out_trait = {524 indexing_maps = [525 affine_map<(m, n, k) -> (m, k)>,526 affine_map<(m, n, k) -> (k, n)>,527 affine_map<(m, n, k) -> (n, m)>528 ],529 iterator_types = ["parallel", "parallel", "reduction"]530}531 532// CHECK-LABEL: func @generic_output_transpose533func.func @generic_output_transpose(%A: memref<8x16xf32>, %B: memref<16x32xf32>,534 %C: memref<32x8xf32>) {535 // CHECK: vector.transfer_read %{{.*}} : memref<8x16xf32>, vector<8x32x16xf32>536 // CHECK: vector.transfer_read %{{.*}} : memref<16x32xf32>, vector<8x32x16xf32>537 // CHECK: %[[ACC:.*]] = vector.transfer_read %{{.*}} : memref<32x8xf32>, vector<8x32xf32>538 // CHECK: %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} : vector<8x32x16xf32>539 // CHECK: %[[R:.*]] = vector.multi_reduction <add>, %[[MUL]], %[[ACC]] [2] : vector<8x32x16xf32> to vector<8x32xf32>540 // CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xf32>, memref<32x8xf32>541 linalg.generic #matmul_transpose_out_trait542 ins(%A, %B : memref<8x16xf32>, memref<16x32xf32>)543 outs(%C : memref<32x8xf32>) {544 ^bb(%a: f32, %b: f32, %c: f32) :545 %d = arith.mulf %a, %b: f32546 %e = arith.addf %c, %d: f32547 linalg.yield %e : f32548 }549 return550}551 552module attributes {transform.with_named_sequence} {553 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {554 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op555 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op556 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op557 transform.yield558 }559}560 561// -----562 563#map0 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>564#map1 = affine_map<(d0, d1, d2) -> (d1, d0, d2)>565// CHECK: #[[MAP:.+]] = affine_map<(d0, d1, d2) -> (d1, d0, d2)>566// CHECK: func @generic_interchanged_transpose567func.func @generic_interchanged_transpose(%arg0: tensor<12x128x32xf32>) -> tensor<128x12x32xf32> {568 // CHECK: %[[IN:.+]] = vector.transfer_read569 // CHECK: vector.transfer_write %[[IN]], {{.+}} permutation_map = #[[MAP]]570 %0 = tensor.empty() : tensor<128x12x32xf32>571 %1 = linalg.generic {indexing_maps = [#map0, #map1],572 iterator_types = ["parallel", "parallel", "parallel"]}573 ins(%arg0 : tensor<12x128x32xf32>)574 outs(%0 : tensor<128x12x32xf32>) {575 ^bb0(%arg1: f32, %arg2: f32):576 linalg.yield %arg1 : f32577 } -> tensor<128x12x32xf32>578 return %1 : tensor<128x12x32xf32>579}580 581module attributes {transform.with_named_sequence} {582 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {583 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op584 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op585 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op586 transform.yield587 }588}589 590// -----591 592#matmul_trait = {593 indexing_maps = [594 affine_map<(m, n, k) -> (m, k)>,595 affine_map<(m, n, k) -> (k, n)>,596 affine_map<(m, n, k) -> (m, n)>597 ],598 iterator_types = ["parallel", "parallel", "reduction"]599}600 601// CHECK-LABEL: func @vectorization_test_integer602func.func @vectorization_test_integer(%A: memref<8x16xi32>, %B: memref<16x32xi32>,603 %C: memref<8x32xi32>) {604 // CHECK: vector.transfer_read %{{.*}} : memref<8x16xi32>, vector<8x32x16xi32>605 // CHECK: vector.transfer_read %{{.*}} : memref<16x32xi32>, vector<8x32x16xi32>606 // CHECK: %[[ACC:.*]] = vector.transfer_read %{{.*}} : memref<8x32xi32>, vector<8x32xi32>607 // CHECK: %[[MUL:.*]] = arith.muli %{{.*}}, %{{.*}} : vector<8x32x16xi32>608 // CHECK: vector.multi_reduction <add>, %[[MUL]], %[[ACC]] [2] : vector<8x32x16xi32> to vector<8x32xi32>609 // CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xi32>, memref<8x32xi32>610 linalg.generic #matmul_trait611 ins(%A, %B : memref<8x16xi32>, memref<16x32xi32>)612 outs(%C : memref<8x32xi32>) {613 ^bb(%a: i32, %b: i32, %c: i32) :614 %d = arith.muli %a, %b: i32615 %e = arith.addi %c, %d: i32616 linalg.yield %e : i32617 }618 return619}620 621module attributes {transform.with_named_sequence} {622 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {623 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op624 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op625 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op626 transform.yield627 }628}629 630// -----631 632 633// CHECK-LABEL: func @test_vectorize_scalar_input634func.func @test_vectorize_scalar_input(%A : memref<8x16xf32>, %arg0 : f32) {635 // CHECK: %[[V:.*]] = vector.broadcast {{.*}} : f32 to vector<8x16xf32>636 // CHECK: vector.transfer_write %[[V]], {{.*}} : vector<8x16xf32>, memref<8x16xf32>637 linalg.generic {638 indexing_maps = [affine_map<(m, n) -> ()>, affine_map<(m, n) -> (m, n)>],639 iterator_types = ["parallel", "parallel"]}640 ins(%arg0 : f32)641 outs(%A: memref<8x16xf32>) {642 ^bb(%0: f32, %1: f32) :643 linalg.yield %0 : f32644 }645 return646}647 648module attributes {transform.with_named_sequence} {649 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {650 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op651 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op652 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op653 transform.yield654 }655}656 657// -----658 659// CHECK-LABEL: func @test_do_not_vectorize_unsupported_element_types660func.func @test_do_not_vectorize_unsupported_element_types(%A : memref<8x16xcomplex<f32>>, %arg0 : complex<f32>) {661 // CHECK-NOT: vector.broadcast662 // CHECK-NOT: vector.transfer_write663 linalg.generic {664 indexing_maps = [affine_map<(m, n) -> ()>, affine_map<(m, n) -> (m, n)>],665 iterator_types = ["parallel", "parallel"]}666 ins(%arg0 : complex<f32>)667 outs(%A: memref<8x16xcomplex<f32>>) {668 ^bb(%0: complex<f32>, %1: complex<f32>) :669 linalg.yield %0 : complex<f32>670 }671 return672}673 674module attributes {transform.with_named_sequence} {675 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {676 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op677 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op678 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op679 transform.yield680 }681}682 683// -----684 685#map0 = affine_map<(d0) -> (d0)>686 687func.func @vectorize_affine_apply(%arg0: tensor<5xf32>, %arg3: index) -> tensor<5xi32> {688 %0 = tensor.empty() : tensor<5xi32>689 %1 = linalg.generic {indexing_maps = [#map0, #map0],690 iterator_types = ["parallel"]}691 ins(%arg0 : tensor<5xf32>)692 outs(%0 : tensor<5xi32>) {693 ^bb0(%arg1: f32, %arg2: i32):694 %2 = linalg.index 0 : index695 %11 = affine.apply affine_map<() -> (123)>()696 %12 = affine.apply affine_map<(d0, d1) -> (d0 + d1)>(%2, %11)697 %13 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%12)[%arg3]698 %14 = affine.apply affine_map<(d0) -> (d0 + 1)>(%13)699 %15 = affine.apply affine_map<(d0, d1, d2) -> (d0 + d1 + d2)>(%13, %14, %12)700 %3 = arith.index_cast %15 : index to i32701 linalg.yield %3 : i32702 } -> tensor<5xi32>703 return %1 : tensor<5xi32>704}705 706// CHECK-LABEL: func.func @vectorize_affine_apply707// CHECK-SAME: %arg0: tensor<5xf32>708// CHECK-SAME: %[[ARG1:.*]]: index709// CHECK-DAG: %[[CST:.*]] = arith.constant dense<[123, 124, 125, 126, 127]> : vector<5xindex>710// CHECK-DAG: %[[CST_0:.*]] = arith.constant dense<1> : vector<5xindex>711// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index712// CHECK: %[[EMPTY:.*]] = tensor.empty() : tensor<5xi32>713// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ARG1]] : index to vector<5xindex>714// CHECK: %[[ADDI_1:.*]] = arith.addi %[[BCAST]], %[[CST]] : vector<5xindex>715// CHECK: %[[ADDI_2:.*]] = arith.addi %[[ADDI_1]], %[[CST_0]] : vector<5xindex>716// CHECK: %[[ADDI_3:.*]] = arith.addi %[[ADDI_1]], %[[ADDI_2]] : vector<5xindex>717// CHECK: %[[ADDI_4:.*]] = arith.addi %[[ADDI_3]], %[[CST]] : vector<5xindex>718// CHECK: %[[CAST:.*]] = arith.index_cast %[[ADDI_4]] : vector<5xindex> to vector<5xi32>719// CHECK: vector.transfer_write %[[CAST]], %[[EMPTY]][%[[C0:.*]]] {in_bounds = [true]} : vector<5xi32>, tensor<5xi32>720 721module attributes {transform.with_named_sequence} {722 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {723 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op724 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op725 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { vectorize_nd_extract } : (!transform.any_op) -> !transform.any_op726 transform.yield727 }728}729 730// -----731 732// CHECK-LABEL: func @test_vectorize_trailing_index733 // CHECK-SAME: (%[[ARG0:.*]]: memref<1x2x4x8xindex>)734func.func @test_vectorize_trailing_index(%arg0: memref<1x2x4x8xindex>) {735 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<[0, 1, 2, 3, 4, 5, 6, 7]> : vector<8xindex>736 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index737 linalg.generic {738 indexing_maps = [739 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],740 iterator_types = ["parallel", "parallel", "parallel", "parallel"]}741 outs(%arg0: memref<1x2x4x8xindex>) {742 ^bb0(%arg1: index):743 // CHECK: %[[BCST:.*]] = vector.broadcast %[[CST0]] : vector<8xindex> to vector<1x2x4x8xindex>744 // CHECK: vector.transfer_write %[[BCST]], %[[ARG0]][%[[C0]], %[[C0]], %[[C0]], %[[C0]]] {{.*}} : vector<1x2x4x8xindex>, memref<1x2x4x8xindex>745 %0 = linalg.index 3 : index746 linalg.yield %0 : index747 }748 return749}750 751module attributes {transform.with_named_sequence} {752 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {753 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op754 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op755 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op756 transform.yield757 }758}759 760// -----761 762// CHECK-LABEL: func @test_vectorize_inner_index763 // CHECK-SAME: (%[[ARG0:.*]]: memref<1x2x4x8xindex>)764func.func @test_vectorize_inner_index(%arg0: memref<1x2x4x8xindex>) {765 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<[0, 1]> : vector<2xindex>766 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index767 linalg.generic {768 indexing_maps = [769 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],770 iterator_types = ["parallel", "parallel", "parallel", "parallel"]}771 outs(%arg0: memref<1x2x4x8xindex>) {772 ^bb0(%arg1: index):773 // CHECK: %[[BCST:.*]] = vector.broadcast %[[CST0]] : vector<2xindex> to vector<1x8x4x2xindex>774 // CHECK: %[[TRAN:.*]] = vector.transpose %[[BCST]], [0, 3, 2, 1] : vector<1x8x4x2xindex> to vector<1x2x4x8xindex>775 // CHECK: vector.transfer_write %[[TRAN]], %[[ARG0]][%[[C0]], %[[C0]], %[[C0]], %[[C0]]] {{.*}} : vector<1x2x4x8xindex>, memref<1x2x4x8xindex>776 %0 = linalg.index 1 : index777 linalg.yield %0 : index778 }779 return780}781 782module attributes {transform.with_named_sequence} {783 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {784 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op785 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op786 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op787 transform.yield788 }789}790 791// -----792 793// CHECK-LABEL: func @generic_vectorize794 // CHECK-SAME: (%[[ARG0:.*]]: memref<4x256xf32>, %[[ARG1:.*]]: memref<4x256xf32>,795 // CHECK-SAME: %[[ARG2:.*]]: memref<256xf32>, %[[ARG3:.*]]: f32)796func.func @generic_vectorize(%arg0: memref<4x256xf32>,797 %arg1: memref<4x256xf32>,798 %arg2: memref<256xf32>, %i: f32) {799 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<2.000000e+00> : vector<4x256xf32>800 // CHECK-DAG: %[[CST1:.*]] = arith.constant dense<1.000000e+00> : vector<4x256xf32>801 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index802 %c1_f32 = arith.constant 1.0 : f32803 linalg.generic {804 indexing_maps = [805 affine_map<(d0, d1) -> (d0, d1)>,806 affine_map<(d0, d1) -> (d1)>,807 affine_map<(d0, d1) -> (d0, d1)>,808 affine_map<(d0, d1) -> (d0, d1)>,809 affine_map<(d0, d1) -> (d0, d1)>,810 affine_map<(d0, d1) -> (d0, d1)>,811 affine_map<(d0, d1) -> (d0, d1)>,812 affine_map<(d0, d1) -> (d0, d1)>,813 affine_map<(d0, d1) -> (d0, d1)>,814 affine_map<(d0, d1) -> (d0, d1)>,815 affine_map<(d0, d1) -> (d0, d1)>,816 affine_map<(d0, d1) -> (d0, d1)>],817 iterator_types = ["parallel", "parallel"]}818 ins(%arg1, %arg2: memref<4x256xf32>, memref<256xf32>)819 outs(820 %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0 :821 memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>,822 memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>,823 memref<4x256xf32>, memref<4x256xf32>) {824 ^bb0(%arg3 : f32, %arg4 : f32, %arg5: f32, %arg6: f32, %arg7: f32, %arg8: f32,825 // CHECK: %[[V2:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : memref<4x256xf32>, vector<4x256xf32>826 // CHECK: %[[V0:.*]] = vector.transfer_read %[[ARG2]][%[[C0]]], {{.*}} : memref<256xf32>, vector<4x256xf32>827 // CHECK: %[[V3:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : memref<4x256xf32>, vector<4x256xf32>828 // CHECK: %[[V1:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : memref<4x256xf32>, vector<4x256xf32>829 %arg9 : f32, %arg10 : f32, %arg11 : f32, %arg12 : f32, %arg13 : f32,830 %arg14 : f32):831 // CHECK: %[[ADD:.*]] = arith.addf %[[V0]], %[[V1]] : vector<4x256xf32>832 %6 = arith.addf %arg4, %arg6 : f32833 // CHECK: %[[CMP:.*]] = arith.cmpf ogt, %[[V2]], %[[V1]] : vector<4x256xf32>834 %7 = arith.cmpf ogt, %arg3, %arg6 : f32835 // CHECK: %[[ARG3B:.*]] = vector.broadcast %[[ARG3]] : f32 to vector<4x256xf32>836 %8 = arith.constant 2.0 : f32837 // CHECK: %[[DIV:.*]] = arith.divf %[[V3]], %[[ARG3B]] : vector<4x256xf32>838 %9 = arith.divf %arg5, %i : f32839 // CHECK: %[[EXP:.*]] = math.exp2 %[[V3]] : vector<4x256xf32>840 %10 = math.exp2 %arg5 : f32841 // CHECK: %[[MUL:.*]] = arith.mulf %[[V3]], %[[CST0]] : vector<4x256xf32>842 %11 = arith.mulf %arg5, %8 : f32843 // CHECK: %[[RSQRT:.*]] = math.rsqrt %[[V3]] : vector<4x256xf32>844 %12 = math.rsqrt %arg5 : f32845 // CHECK: %[[SEL:.*]] = arith.select %[[CMP]], %[[V3]], %[[V1]] : vector<4x256xi1>, vector<4x256xf32>846 %13 = arith.select %7, %arg5, %arg6 : f32847 // CHECK: %[[SUB:.*]] = arith.subf %[[V3]], %[[V0]] : vector<4x256xf32>848 %14 = arith.subf %arg5, %arg4 : f32849 // CHECK: %[[TAN:.*]] = math.tanh %[[V3]] : vector<4x256xf32>850 %15 = math.tanh %arg5 : f32851 // CHECK: vector.transfer_write %[[ADD]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>852 // CHECK: vector.transfer_write %[[CST0]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>853 // CHECK: vector.transfer_write %[[CST1]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>854 // CHECK: vector.transfer_write %[[DIV]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>855 // CHECK: vector.transfer_write %[[EXP]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>856 // CHECK: vector.transfer_write %[[MUL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>857 // CHECK: vector.transfer_write %[[RSQRT]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>858 // CHECK: vector.transfer_write %[[SEL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>859 // CHECK: vector.transfer_write %[[SUB]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>860 // CHECK: vector.transfer_write %[[TAN]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>861 linalg.yield %6, %8, %c1_f32, %9, %10, %11, %12, %13, %14, %15 : f32, f32,862 f32, f32, f32, f32, f32, f32, f32, f32863 }864 return865}866 867module attributes {transform.with_named_sequence} {868 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {869 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op870 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op871 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op872 transform.yield873 }874}875 876// -----877 878// CHECK-LABEL: func @generic_vectorize_tensor879// CHECK-SAME: (%[[ARG0:.*]]: tensor<4x256xf32>, %[[ARG1:.*]]: tensor<4x256xf32>,880// CHECK-SAME: %[[ARG2:.*]]: tensor<256xf32>, %[[ARG3:.*]]: f32)881func.func @generic_vectorize_tensor(%arg0: tensor<4x256xf32>,882 %arg1: tensor<4x256xf32>, %arg2: tensor<256xf32>,883 %i: f32) -> (tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,884 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,885 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>) {886 %c1_f32 = arith.constant 1.0 : f32887 %r:10 = linalg.generic {888 indexing_maps = [889 affine_map<(d0, d1) -> (d0, d1)>,890 affine_map<(d0, d1) -> (d1)>,891 affine_map<(d0, d1) -> (d0, d1)>,892 affine_map<(d0, d1) -> (d0, d1)>,893 affine_map<(d0, d1) -> (d0, d1)>,894 affine_map<(d0, d1) -> (d0, d1)>,895 affine_map<(d0, d1) -> (d0, d1)>,896 affine_map<(d0, d1) -> (d0, d1)>,897 affine_map<(d0, d1) -> (d0, d1)>,898 affine_map<(d0, d1) -> (d0, d1)>,899 affine_map<(d0, d1) -> (d0, d1)>,900 affine_map<(d0, d1) -> (d0, d1)>],901 iterator_types = ["parallel", "parallel"]}902 ins(%arg1, %arg2: tensor<4x256xf32>, tensor<256xf32>)903 outs(904 %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0 :905 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,906 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,907 tensor<4x256xf32>, tensor<4x256xf32>) {908 ^bb0(%arg3 : f32, %arg4 : f32, %arg5: f32, %arg6: f32, %arg7: f32, %arg8: f32,909 %arg9 : f32, %arg10 : f32, %arg11 : f32, %arg12 : f32, %arg13 : f32,910 %arg14 : f32):911 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<2.000000e+00> : vector<4x256xf32>912 // CHECK-DAG: %[[CST1:.*]] = arith.constant dense<1.000000e+00> : vector<4x256xf32>913 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index914 // CHECK: %[[V2:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x256xf32>, vector<4x256xf32>915 // CHECK: %[[V0:.*]] = vector.transfer_read %[[ARG2]][%[[C0]]], {{.*}} : tensor<256xf32>, vector<4x256xf32>916 // CHECK: %[[V3:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x256xf32>, vector<4x256xf32>917 // CHECK: %[[V1:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x256xf32>, vector<4x256xf32>918 // CHECK: %[[ADD:.*]] = arith.addf %[[V0]], %[[V1]] : vector<4x256xf32>919 %6 = arith.addf %arg4, %arg6 : f32920 // CHECK: %[[CMP:.*]] = arith.cmpf ogt, %[[V2]], %[[V1]] : vector<4x256xf32>921 %7 = arith.cmpf ogt, %arg3, %arg6 : f32922 // CHECK: %[[ARG3B:.*]] = vector.broadcast %[[ARG3]] : f32 to vector<4x256xf32>923 %8 = arith.constant 2.0 : f32924 // CHECK: %[[DIV:.*]] = arith.divf %[[V3]], %[[ARG3B]] : vector<4x256xf32>925 %9 = arith.divf %arg5, %i : f32926 // CHECK: %[[EXP:.*]] = math.exp2 %[[V3]] : vector<4x256xf32>927 %10 = math.exp2 %arg5 : f32928 // CHECK: %[[MUL:.*]] = arith.mulf %[[V3]], %[[CST0]] : vector<4x256xf32>929 %11 = arith.mulf %arg5, %8 : f32930 // CHECK: %[[RSQRT:.*]] = math.rsqrt %[[V3]] : vector<4x256xf32>931 %12 = math.rsqrt %arg5 : f32932 // CHECK: %[[SEL:.*]] = arith.select %[[CMP]], %[[V3]], %[[V1]] : vector<4x256xi1>, vector<4x256xf32>933 %13 = arith.select %7, %arg5, %arg6 : f32934 // CHECK: %[[SUB:.*]] = arith.subf %[[V3]], %[[V0]] : vector<4x256xf32>935 %14 = arith.subf %arg5, %arg4 : f32936 // CHECK: %[[TAN:.*]] = math.tanh %[[V3]] : vector<4x256xf32>937 %15 = math.tanh %arg5 : f32938 // CHECK: %[[R0:.*]] = vector.transfer_write %[[ADD]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>939 // CHECK: %[[R1:.*]] = vector.transfer_write %[[CST0]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>940 // CHECK: %[[R2:.*]] = vector.transfer_write %[[CST1]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>941 // CHECK: %[[R3:.*]] = vector.transfer_write %[[DIV]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>942 // CHECK: %[[R4:.*]] = vector.transfer_write %[[EXP]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>943 // CHECK: %[[R5:.*]] = vector.transfer_write %[[MUL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>944 // CHECK: %[[R6:.*]] = vector.transfer_write %[[RSQRT]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>945 // CHECK: %[[R7:.*]] = vector.transfer_write %[[SEL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>946 // CHECK: %[[R8:.*]] = vector.transfer_write %[[SUB]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>947 // CHECK: %[[R9:.*]] = vector.transfer_write %[[TAN]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>948 linalg.yield %6, %8, %c1_f32, %9, %10, %11, %12, %13, %14, %15 : f32, f32,949 f32, f32, f32, f32, f32, f32, f32, f32950 } -> (tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,951 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,952 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>)953 // CHECK: return %[[R0]], %[[R1]], %[[R2]], %[[R3]], %[[R4]], %[[R5]], %[[R6]], %[[R7]], %[[R8]], %[[R9]] : tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>954 return %r#0, %r#1, %r#2, %r#3, %r#4, %r#5, %r#6, %r#7, %r#8, %r#9:955 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,956 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,957 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>958}959 960module attributes {transform.with_named_sequence} {961 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {962 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op963 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op964 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op965 transform.yield966 }967}968 969// -----970 971// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, 0, 0, d1)>972// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0) -> (d0, 0, 0, 0)>973// CHECK-DAG: #[[$MAP2:.*]] = affine_map<(d0) -> (0, 0, d0, 0)>974// CHECK-DAG: #[[$MAP3:.*]] = affine_map<(d0, d1) -> (d1, 0, d0, 0)>975// CHECK: func @generic_vectorize_broadcast_transpose976// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index977// CHECK-DAG: %[[PV:.*]] = ub.poison : f32978// CHECK: %[[V0:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %[[PV]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP0]]} : memref<4x4xf32>, vector<4x4x4x4xf32>979// CHECK: %[[V1:.*]] = vector.transfer_read %{{.*}}[%[[C0]]], %[[PV]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP1]]} : memref<4xf32>, vector<4x4x4x4xf32>980// CHECK: %[[V2:.*]] = vector.transfer_read %{{.*}}[%[[C0]]], %[[PV]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP2]]} : memref<4xf32>, vector<4x4x4x4xf32>981// CHECK: %[[V3:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %[[PV]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP3]]} : memref<4x4xf32>, vector<4x4x4x4xf32>982// CHECK: %[[SUB:.*]] = arith.subf %[[V0]], %[[V1]] : vector<4x4x4x4xf32>983// CHECK: %[[ADD0:.*]] = arith.addf %[[V2]], %[[SUB]] : vector<4x4x4x4xf32>984// CHECK: %[[ADD1:.*]] = arith.addf %[[V3]], %[[ADD0]] : vector<4x4x4x4xf32>985// CHECK: vector.transfer_write %[[ADD1]], {{.*}} : vector<4x4x4x4xf32>, memref<4x4x4x4xf32>986func.func @generic_vectorize_broadcast_transpose(987 %A: memref<4xf32>, %B: memref<4x4xf32>, %C: memref<4x4x4x4xf32>) {988 linalg.generic {989 indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d3)>,990 affine_map<(d0, d1, d2, d3) -> (d0)>,991 affine_map<(d0, d1, d2, d3) -> (d2)>,992 affine_map<(d0, d1, d2, d3) -> (d2, d0)>,993 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],994 iterator_types = ["parallel", "parallel", "parallel", "parallel"]}995 ins(%B, %A, %A, %B: memref<4x4xf32>, memref<4xf32>, memref<4xf32>, memref<4x4xf32>)996 outs(%C : memref<4x4x4x4xf32>) {997 ^bb0(%arg0: f32, %arg1: f32, %arg2: f32, %arg3: f32, %arg4: f32):998 %s = arith.subf %arg0, %arg1 : f32999 %a = arith.addf %arg2, %s : f321000 %b = arith.addf %arg3, %a : f321001 linalg.yield %b : f321002 }1003 return1004}1005 1006module attributes {transform.with_named_sequence} {1007 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1008 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1009 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1010 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op1011 transform.yield1012 }1013}1014 1015// -----1016 1017// Test different input maps.1018#matmul_trait = {1019 indexing_maps = [1020 affine_map<(d0, d1, d2, d3) -> (d1, d0)>,1021 affine_map<(d0, d1, d2, d3) -> (d3, d1)>,1022 affine_map<(d0, d1, d2, d3) -> (d3, d1, d0, d2)>,1023 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>1024 ],1025 iterator_types = ["parallel", "parallel", "parallel", "parallel"]1026}1027 1028// CHECK-DAG: #[[MAP0:.*]] = affine_map<(d0, d1) -> (d1, d0, 0, 0)>1029// CHECK-DAG: #[[MAP1:.*]] = affine_map<(d0, d1) -> (0, d1, 0, d0)>1030// CHECK-DAG: #[[MAP2:.*]] = affine_map<(d0, d1, d2, d3) -> (d2, d1, d3, d0)>1031// CHECK: func @vectorization_transpose1032// CHECK: vector.transfer_read {{.*}}{in_bounds = [true, true, true, true], permutation_map = #[[MAP0]]} : memref<14x7xf32>, vector<7x14x8x16xf32>1033// CHECK: vector.transfer_read {{.*}}{in_bounds = [true, true, true, true], permutation_map = #[[MAP1]]} : memref<16x14xf32>, vector<7x14x8x16xf32>1034// CHECK: vector.transfer_read {{.*}}{in_bounds = [true, true, true, true], permutation_map = #[[MAP2]]} : memref<16x14x7x8xf32>, vector<7x14x8x16xf32>1035// CHECK: arith.addf {{.*}} : vector<7x14x8x16xf32>1036// CHECK: arith.addf {{.*}} : vector<7x14x8x16xf32>1037// CHECK: vector.transfer_write {{.*}} : vector<7x14x8x16xf32>, memref<7x14x8x16xf32>1038func.func @vectorization_transpose(%A: memref<14x7xf32>, %B: memref<16x14xf32>,1039 %C: memref<16x14x7x8xf32>, %D: memref<7x14x8x16xf32>) {1040 linalg.generic #matmul_trait1041 ins(%A, %B, %C : memref<14x7xf32>, memref<16x14xf32>, memref<16x14x7x8xf32>)1042 outs(%D : memref<7x14x8x16xf32>) {1043 ^bb(%a: f32, %b: f32, %c: f32, %d: f32) :1044 %e = arith.addf %a, %b: f321045 %f = arith.addf %e, %c: f321046 linalg.yield %f : f321047 }1048 return1049}1050 1051module attributes {transform.with_named_sequence} {1052 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1053 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1054 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1055 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op1056 transform.yield1057 }1058}1059 1060// -----1061 1062// CHECK-LABEL: func @sum_exp1063func.func @sum_exp(%input: tensor<4x16x8xf32>, %output: tensor<4x16xf32>)1064 -> tensor<4x16xf32>1065{1066 // CHECK: vector.transfer_read {{.*}} : tensor<4x16x8xf32>, vector<4x16x8xf32>1067 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<4x16xf32>, vector<4x16xf32>1068 // CHECK: math.exp {{.*}} : vector<4x16x8xf32>1069 // CHECK: vector.multi_reduction <add>, %{{.*}}, %{{.*}} [2] : vector<4x16x8xf32> to vector<4x16xf32>1070 // CHECK: vector.transfer_write {{.*}} : vector<4x16xf32>, tensor<4x16xf32>1071 // CHECK: return {{.*}} : tensor<4x16xf32>1072 %0 = linalg.generic {1073 indexing_maps = [1074 affine_map<(d0, d1, d2) -> (d0, d1, d2)>,1075 affine_map<(d0, d1, d2) -> (d0, d1)>1076 ],1077 iterator_types = ["parallel", "parallel", "reduction"]1078 } ins(%input : tensor<4x16x8xf32>) outs(%output : tensor<4x16xf32>) {1079 ^bb0(%arg0: f32, %arg1: f32):1080 %1 = math.exp %arg0 : f321081 %2 = arith.addf %1, %arg1 : f321082 linalg.yield %2 : f321083 } -> tensor<4x16xf32>1084 return %0 : tensor<4x16xf32>1085}1086 1087module attributes {transform.with_named_sequence} {1088 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1089 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1090 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1091 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1092 transform.yield1093 }1094}1095 1096// -----1097 1098// CHECK-DAG: #[[$M1:.*]] = affine_map<(d0, d1) -> (d1, d0, 0, 0)>1099// CHECK-DAG: #[[$M2:.*]] = affine_map<(d0, d1) -> (0, 0, d1, d0)>1100// CHECK-DAG: #[[$M3:.*]] = affine_map<(d0, d1) -> (d1, d0)>1101 1102// CHECK-LABEL: func @sum_exp_21103func.func @sum_exp_2(%input: tensor<3x2xf32>, %input_2: tensor<5x4xf32>, %output: tensor<5x2xf32>)1104 -> tensor<5x2xf32>1105{1106 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true, true, true], permutation_map = #[[$M1]]} : tensor<3x2xf32>, vector<2x3x4x5xf32>1107 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true, true, true], permutation_map = #[[$M2]]} : tensor<5x4xf32>, vector<2x3x4x5xf32>1108 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true], permutation_map = #[[$M3]]} : tensor<5x2xf32>, vector<2x5xf32>1109 // CHECK: math.exp {{.*}} : vector<2x3x4x5xf32>1110 // CHECK: math.exp {{.*}} : vector<2x3x4x5xf32>1111 // CHECK: addf {{.*}} : vector<2x3x4x5xf32>1112 // CHECK: vector.multi_reduction <add>, {{.*}}, %{{.*}} [1, 2] : vector<2x3x4x5xf32> to vector<2x5xf32>1113 // CHECK: vector.transfer_write {{.*}} {in_bounds = [true, true], permutation_map = #[[$M3]]} : vector<2x5xf32>, tensor<5x2xf32>1114 // CHECK: return {{.*}} : tensor<5x2xf32>1115 %0 = linalg.generic {1116 indexing_maps = [1117 affine_map<(d0, d1, d2, d3) -> (d1, d0)>,1118 affine_map<(d0, d1, d2, d3) -> (d3, d2)>,1119 affine_map<(d0, d1, d2, d3) -> (d3, d0)>1120 ],1121 iterator_types = ["parallel", "reduction", "reduction", "parallel"]1122 } ins(%input, %input_2 : tensor<3x2xf32>, tensor<5x4xf32>) outs(%output : tensor<5x2xf32>) {1123 ^bb0(%arg0: f32, %arg1: f32, %arg2: f32):1124 %1 = math.exp %arg0 : f321125 %2 = math.exp %arg1 : f321126 %3 = arith.addf %1, %2 : f321127 %4 = arith.addf %3, %arg2 : f321128 linalg.yield %4 : f321129 } -> tensor<5x2xf32>1130 return %0 : tensor<5x2xf32>1131}1132 1133 1134module attributes {transform.with_named_sequence} {1135 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1136 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1137 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1138 %5 = transform.structured.vectorize_children_and_apply_patterns %4 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op1139 transform.yield1140 }1141}1142 1143// -----1144 1145// CHECK-LABEL: func @red_maximumf_2d(1146func.func @red_maximumf_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {1147 // CHECK: %[[CMINF:.+]] = arith.constant dense<-3.402820e+38> : vector<4xf32>1148 // CHECK: tensor.empty() : tensor<4xf32>1149 // CHECK: vector.multi_reduction <maximumf>, {{.*}}, %[[CMINF]] [1] : vector<4x4xf32> to vector<4xf32>1150 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>1151 %ident = arith.constant -3.40282e+38 : f321152 %init = tensor.empty() : tensor<4xf32>1153 %fill = linalg.fill ins(%ident : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>1154 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1155 affine_map<(d0, d1) -> (d0)>],1156 iterator_types = ["parallel", "reduction"]}1157 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {1158 ^bb0(%in0: f32, %out0: f32):1159 %max = arith.maximumf %in0, %out0 : f321160 linalg.yield %max : f321161 } -> tensor<4xf32>1162 return %red : tensor<4xf32>1163}1164 1165module attributes {transform.with_named_sequence} {1166 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1167 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1168 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1169 %5 = transform.structured.vectorize_children_and_apply_patterns %4 { vectorize_padding } : (!transform.any_op) -> !transform.any_op1170 transform.yield1171 }1172}1173 1174// -----1175 1176// CHECK-LABEL: func @red_maxnumf_2d(1177func.func @red_maxnumf_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {1178 // CHECK: %[[CMINF:.+]] = arith.constant dense<-3.402820e+38> : vector<4xf32>1179 // CHECK: tensor.empty() : tensor<4xf32>1180 // CHECK: vector.multi_reduction <maxnumf>, {{.*}}, %[[CMINF]] [1] : vector<4x4xf32> to vector<4xf32>1181 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>1182 %ident = arith.constant -3.40282e+38 : f321183 %init = tensor.empty() : tensor<4xf32>1184 %fill = linalg.fill ins(%ident : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>1185 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1186 affine_map<(d0, d1) -> (d0)>],1187 iterator_types = ["parallel", "reduction"]}1188 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {1189 ^bb0(%in0: f32, %out0: f32):1190 %max = arith.maxnumf %in0, %out0 : f321191 linalg.yield %max : f321192 } -> tensor<4xf32>1193 return %red : tensor<4xf32>1194}1195 1196 1197module attributes {transform.with_named_sequence} {1198 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1199 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1200 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1201 %5 = transform.structured.vectorize_children_and_apply_patterns %4 { vectorize_padding } : (!transform.any_op) -> !transform.any_op1202 transform.yield1203 }1204}1205 1206// -----1207 1208// CHECK-LABEL: func @red_minimumf_2d(1209func.func @red_minimumf_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {1210 // CHECK: %[[CMAXF:.+]] = arith.constant dense<3.402820e+38> : vector<4xf32>1211 // CHECK: tensor.empty() : tensor<4xf32>1212 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xf32>, vector<4x4xf32>1213 // CHECK: vector.multi_reduction <minimumf>, {{.*}}, %[[CMAXF]] [1] : vector<4x4xf32> to vector<4xf32>1214 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>1215 %maxf32 = arith.constant 3.40282e+38 : f321216 %init = tensor.empty() : tensor<4xf32>1217 %fill = linalg.fill ins(%maxf32 : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>1218 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1219 affine_map<(d0, d1) -> (d0)>],1220 iterator_types = ["parallel", "reduction"]}1221 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {1222 ^bb0(%in0: f32, %out0: f32):1223 %min = arith.minimumf %out0, %in0 : f321224 linalg.yield %min : f321225 } -> tensor<4xf32>1226 return %red : tensor<4xf32>1227}1228 1229 1230module attributes {transform.with_named_sequence} {1231 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1232 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1233 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1234 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1235 transform.yield1236 }1237}1238 1239// -----1240 1241// CHECK-LABEL: func @red_minnumf_2d(1242func.func @red_minnumf_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {1243 // CHECK: %[[CMAXF:.+]] = arith.constant dense<3.402820e+38> : vector<4xf32>1244 // CHECK: tensor.empty() : tensor<4xf32>1245 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xf32>, vector<4x4xf32>1246 // CHECK: vector.multi_reduction <minnumf>, {{.*}}, %[[CMAXF]] [1] : vector<4x4xf32> to vector<4xf32>1247 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>1248 %maxf32 = arith.constant 3.40282e+38 : f321249 %init = tensor.empty() : tensor<4xf32>1250 %fill = linalg.fill ins(%maxf32 : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>1251 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1252 affine_map<(d0, d1) -> (d0)>],1253 iterator_types = ["parallel", "reduction"]}1254 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {1255 ^bb0(%in0: f32, %out0: f32):1256 %min = arith.minnumf %out0, %in0 : f321257 linalg.yield %min : f321258 } -> tensor<4xf32>1259 return %red : tensor<4xf32>1260}1261 1262 1263module attributes {transform.with_named_sequence} {1264 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1265 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1266 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1267 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1268 transform.yield1269 }1270}1271 1272// -----1273 1274// CHECK-LABEL: func @red_mul_2d(1275func.func @red_mul_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {1276 // CHECK: tensor.empty() : tensor<4xf32>1277 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xf32>, vector<4x4xf32>1278 // CHECK: vector.multi_reduction <mul>, {{.*}}, {{.*}} [1] : vector<4x4xf32> to vector<4xf32>1279 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>1280 %ident = arith.constant 1.0 : f321281 %init = tensor.empty() : tensor<4xf32>1282 %fill = linalg.fill ins(%ident : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>1283 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1284 affine_map<(d0, d1) -> (d0)>],1285 iterator_types = ["parallel", "reduction"]}1286 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {1287 ^bb0(%in0: f32, %out0: f32):1288 %mul = arith.mulf %in0, %out0 : f321289 linalg.yield %mul : f321290 } -> tensor<4xf32>1291 return %red : tensor<4xf32>1292}1293 1294 1295module attributes {transform.with_named_sequence} {1296 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1297 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1298 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1299 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1300 transform.yield1301 }1302}1303 1304// -----1305 1306// CHECK-LABEL: func @red_or_2d(1307func.func @red_or_2d(%arg0: tensor<4x4xi1>) -> tensor<4xi1> {1308 // CHECK: tensor.empty() : tensor<4xi1>1309 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xi1>, vector<4x4xi1>1310 // CHECK: vector.multi_reduction <or>, {{.*}}, {{.*}} [1] : vector<4x4xi1> to vector<4xi1>1311 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1>1312 %ident = arith.constant false1313 %init = tensor.empty() : tensor<4xi1>1314 %fill = linalg.fill ins(%ident : i1) outs(%init : tensor<4xi1>) -> tensor<4xi1>1315 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1316 affine_map<(d0, d1) -> (d0)>],1317 iterator_types = ["parallel", "reduction"]}1318 ins(%arg0 : tensor<4x4xi1>) outs(%fill : tensor<4xi1>) {1319 ^bb0(%in0: i1, %out0: i1):1320 %or = arith.ori %in0, %out0 : i11321 linalg.yield %or : i11322 } -> tensor<4xi1>1323 return %red : tensor<4xi1>1324}1325 1326 1327module attributes {transform.with_named_sequence} {1328 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1329 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1330 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1331 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1332 transform.yield1333 }1334}1335 1336// -----1337 1338// CHECK-LABEL: func @red_and_2d(1339func.func @red_and_2d(%arg0: tensor<4x4xi1>) -> tensor<4xi1> {1340 // CHECK: tensor.empty() : tensor<4xi1>1341 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xi1>, vector<4x4xi1>1342 // CHECK: vector.multi_reduction <and>, {{.*}}, {{.*}} [1] : vector<4x4xi1> to vector<4xi1>1343 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1>1344 %ident = arith.constant true1345 %init = tensor.empty() : tensor<4xi1>1346 %fill = linalg.fill ins(%ident : i1) outs(%init : tensor<4xi1>) -> tensor<4xi1>1347 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1348 affine_map<(d0, d1) -> (d0)>],1349 iterator_types = ["parallel", "reduction"]}1350 ins(%arg0 : tensor<4x4xi1>) outs(%fill : tensor<4xi1>) {1351 ^bb0(%in0: i1, %out0: i1):1352 %and = arith.andi %in0, %out0 : i11353 linalg.yield %and : i11354 } -> tensor<4xi1>1355 return %red : tensor<4xi1>1356}1357 1358 1359module attributes {transform.with_named_sequence} {1360 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1361 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1362 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1363 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1364 transform.yield1365 }1366}1367 1368// -----1369 1370// CHECK-LABEL: func @red_xor_2d(1371func.func @red_xor_2d(%arg0: tensor<4x4xi1>) -> tensor<4xi1> {1372 // CHECK: tensor.empty() : tensor<4xi1>1373 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xi1>, vector<4x4xi1>1374 // CHECK: vector.multi_reduction <xor>, {{.*}}, {{.*}} [1] : vector<4x4xi1> to vector<4xi1>1375 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1>1376 %ident = arith.constant false1377 %init = tensor.empty() : tensor<4xi1>1378 %fill = linalg.fill ins(%ident : i1) outs(%init : tensor<4xi1>) -> tensor<4xi1>1379 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1380 affine_map<(d0, d1) -> (d0)>],1381 iterator_types = ["parallel", "reduction"]}1382 ins(%arg0 : tensor<4x4xi1>) outs(%fill : tensor<4xi1>) {1383 ^bb0(%in0: i1, %out0: i1):1384 %xor = arith.xori %in0, %out0 : i11385 linalg.yield %xor : i11386 } -> tensor<4xi1>1387 return %red : tensor<4xi1>1388}1389 1390 1391module attributes {transform.with_named_sequence} {1392 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1393 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1394 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1395 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1396 transform.yield1397 }1398}1399 1400// -----1401 1402// CHECK-DAG: #[[$M5:.*]] = affine_map<(d0, d1) -> (d0, 0)>1403 1404// CHECK-LABEL: func @explicit_broadcast(1405func.func @explicit_broadcast(%arg0: tensor<4x4xf32>, %arg1: tensor<4x1xf32>) -> tensor<4x4xf32> {1406 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<4x4xf32>, vector<4x4xf32>1407 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true], permutation_map = #[[$M5]]} : tensor<4x1xf32>, vector<4x4xf32>1408 // CHECK: subf {{.*}} : vector<4x4xf32>1409 // CHECK: vector.transfer_write {{.*}} {in_bounds = [true, true]} : vector<4x4xf32>, tensor<4x4xf32>1410 %c0 = arith.constant 0.0 : f321411 %init = tensor.empty() : tensor<4x4xf32>1412 %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<4x4xf32>) -> tensor<4x4xf32>1413 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1414 affine_map<(d0, d1) -> (d0, 0)>,1415 affine_map<(d0, d1) -> (d0, d1)>],1416 iterator_types = ["parallel", "parallel"]}1417 ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x1xf32>)1418 outs(%fill : tensor<4x4xf32>) {1419 ^bb0(%arg7: f32, %arg8: f32, %arg9: f32):1420 %40 = arith.subf %arg7, %arg8 : f321421 linalg.yield %40 : f321422 } -> tensor<4x4xf32>1423 return %red : tensor<4x4xf32>1424}1425 1426 1427module attributes {transform.with_named_sequence} {1428 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1429 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1430 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1431 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1432 transform.yield1433 }1434}1435 1436// -----1437 1438// TODO: Two Linalg Ops in one tests - either split or document "why".1439 1440// CHECK-DAG: #[[$M6:.*]] = affine_map<(d0, d1) -> (d0, 0)>1441 1442// CHECK-LABEL: func @fused_broadcast_red_2d1443func.func @fused_broadcast_red_2d(%arg0: tensor<4x4xf32>, %arg1: tensor<4x1xf32>) -> tensor<4xf32> {1444 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<4x4xf32>, vector<4x4xf32>1445 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true], permutation_map = #[[$M6]]} : tensor<4x1xf32>, vector<4x4xf32>1446 // CHECK: subf {{.*}} : vector<4x4xf32>1447 // CHECK: math.exp {{.*}} : vector<4x4xf32>1448 // CHECK: vector.multi_reduction <add>, {{.*}}, {{.*}} : vector<4x4xf32> to vector<4xf32>1449 // CHECK: vector.transfer_write {{.*}} {in_bounds = [true]} : vector<4xf32>, tensor<4xf32>1450 %c0 = arith.constant 0.0 : f321451 %init = tensor.empty() : tensor<4xf32>1452 %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>1453 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,1454 affine_map<(d0, d1) -> (d0, 0)>,1455 affine_map<(d0, d1) -> (d0)>],1456 iterator_types = ["parallel", "reduction"]}1457 ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x1xf32>)1458 outs(%fill : tensor<4xf32>) {1459 ^bb0(%arg7: f32, %arg8: f32, %arg9: f32):1460 %40 = arith.subf %arg7, %arg8 : f321461 %41 = math.exp %40 : f321462 %42 = arith.addf %41, %arg9 : f321463 linalg.yield %42 : f321464 } -> tensor<4xf32>1465 return %red : tensor<4xf32>1466}1467 1468 1469module attributes {transform.with_named_sequence} {1470 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1471 %0 = transform.structured.match ops{["linalg.fill"]} in %arg1 : (!transform.any_op) -> !transform.any_op1472 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1473 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op1474 1475 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1476 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1477 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1478 transform.yield1479 }1480}1481 1482// -----1483 1484// CHECK-LABEL: func @reduce_to_rank_0(1485// CHECK-SAME: %[[SRC:.*]]: tensor<32xf32>1486func.func @reduce_to_rank_0(%arg0: tensor<32xf32>) -> tensor<f32> {1487 // CHECK-DAG: %[[F0:.*]] = arith.constant 0.000000e+00 : f321488 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index1489 %f0 = arith.constant 0.000000e+00 : f321490 1491 // CHECK: %[[INIT:.*]] = tensor.empty() : tensor<f32>1492 %0 = tensor.empty() : tensor<f32>1493 1494 %1 = linalg.fill ins(%f0 : f32) outs(%0 : tensor<f32>) -> tensor<f32>1495 // CHECK: %[[R:.*]] = vector.transfer_read %[[SRC]][%[[C0]]]1496 // CHECK-SAME: : tensor<32xf32>, vector<32xf32>1497 // CHECK: %[[RED:.*]] = vector.multi_reduction <add>, %[[R]], %[[F0]] [0]1498 // CHECK-SAME: : vector<32xf32> to f321499 // CHECK: %[[RED_V1:.*]] = vector.broadcast %[[RED]] : f32 to vector<f32>1500 // CHECK: %[[RES:.*]] = vector.transfer_write %[[RED_V1]], %[[INIT]][]1501 // CHECK-SAME: : vector<f32>, tensor<f32>1502 %2 = linalg.generic {1503 indexing_maps = [affine_map<(d0) -> (d0)>,1504 affine_map<(d0) -> ()>],1505 iterator_types = ["reduction"]}1506 ins(%arg0 : tensor<32xf32>)1507 outs(%1 : tensor<f32>) {1508 ^bb0(%a: f32, %b: f32):1509 %3 = arith.addf %a, %b : f321510 linalg.yield %3 : f321511 } -> tensor<f32>1512 1513 return %2 : tensor<f32>1514}1515 1516module attributes {transform.with_named_sequence} {1517 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1518 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1519 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1520 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op1521 transform.yield1522 }1523}1524 1525 1526// -----1527 1528// CHECK-LABEL: func @reduce_to_rank_1(1529// CHECK-SAME: %[[SRC:.*]]: tensor<32xf32>1530func.func @reduce_to_rank_1(%arg0: tensor<32xf32>) -> tensor<1xf32> {1531 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index1532 // CHECK-DAG: %[[F0:.*]] = arith.constant dense<0.000000e+00> : vector<1xf32>1533 %f0 = arith.constant 0.000000e+00 : f321534 1535 // CHECK: %[[INIT:.*]] = tensor.empty() : tensor<1xf32>1536 %0 = tensor.empty() : tensor<1xf32>1537 1538 // CHECK: %[[INIT_ZERO:.*]] = vector.transfer_write %[[F0]], %[[INIT]][%[[C0]]]1539 // CHECK-SAME: : vector<1xf32>, tensor<1xf32>1540 %1 = linalg.fill ins(%f0 : f32) outs(%0 : tensor<1xf32>) -> tensor<1xf32>1541 1542 // CHECK: %[[R:.*]] = vector.transfer_read %[[SRC]][%[[C0]]]1543 // CHECK-SAME: : tensor<32xf32>, vector<32xf32>1544 // CHECK: %[[INIT_ZERO_VEC:.*]] = vector.transfer_read %[[INIT_ZERO]][%[[C0]]]1545 // CHECK-SAME: : tensor<1xf32>, vector<f32>1546 // CHECK: %[[INIT_ZERO_SCL:.*]] = vector.extract %[[INIT_ZERO_VEC]][]1547 // CHECK-SAME: : f32 from vector<f32>1548 // CHECK: %[[RED:.*]] = vector.multi_reduction <add>, %[[R]], %[[INIT_ZERO_SCL]] [0]1549 // CHECK-SAME: : vector<32xf32> to f321550 // CHECK: %[[RED_V1:.*]] = vector.broadcast %[[RED]] : f32 to vector<f32>1551 // CHECK: vector.transfer_write %[[RED_V1]], %[[INIT_ZERO]][%[[C0]]]1552 // CHECK-SAME: : vector<f32>, tensor<1xf32>1553 1554 %2 = linalg.generic {1555 indexing_maps = [affine_map<(d0) -> (d0)>,1556 affine_map<(d0) -> (0)>],1557 iterator_types = ["reduction"]}1558 ins(%arg0 : tensor<32xf32>)1559 outs(%1 : tensor<1xf32>) {1560 ^bb0(%a: f32, %b: f32):1561 %3 = arith.addf %a, %b : f321562 linalg.yield %3 : f321563 } -> tensor<1xf32>1564 1565 return %2 : tensor<1xf32>1566}1567 1568module attributes {transform.with_named_sequence} {1569 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1570 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1571 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1572 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op1573 transform.yield1574 }1575}1576 1577 1578// -----1579 1580// This test checks that vectorization does not occur when an input indexing map1581// is not a projected permutation. In the future, this can be converted to a1582// positive test when support is added.1583 1584// CHECK-LABEL: func @not_projected_permutation1585func.func @not_projected_permutation(%arg0: tensor<8x8xf32>) -> tensor<6x6x3x3xf32> {1586 %c0 = arith.constant 0.0 : f321587 %init = tensor.empty() : tensor<6x6x3x3xf32>1588 %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<6x6x3x3xf32>) -> tensor<6x6x3x3xf32>1589 // CHECK: linalg.generic1590 %result = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0 + d2, d1 + d3)>,1591 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],1592 iterator_types = ["parallel", "parallel", "parallel", "parallel"]}1593 ins(%arg0 : tensor<8x8xf32>)1594 outs(%fill : tensor<6x6x3x3xf32>) {1595 ^bb0(%arg7: f32, %arg9: f32):1596 linalg.yield %arg7 : f321597 } -> tensor<6x6x3x3xf32>1598 return %result : tensor<6x6x3x3xf32>1599}1600 1601module attributes {transform.with_named_sequence} {1602 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1603 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1604 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1605 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op1606 transform.yield1607 }1608}1609 1610// -----1611 1612// Check vectorization can handle cases where outputs are a mix of reduced and non-reduced values.1613func.func @mixed_parallel_reduced_results(%arg0 : tensor<2x4x8xf32>,1614 %arg1 : tensor<2x4xf32>, %arg2 : tensor<2x4x8xf32>, %arg3 : tensor<2x4xf32>) ->1615 (tensor<2x4x8xf32>, tensor<2x4xf32>) {1616 %0:2 = linalg.generic {1617 indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>,1618 affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>],1619 iterator_types = ["parallel", "parallel", "reduction"]}1620 ins(%arg0, %arg1 : tensor<2x4x8xf32>, tensor<2x4xf32>)1621 outs(%arg2, %arg3 : tensor<2x4x8xf32>, tensor<2x4xf32>) {1622 ^bb0(%b0 : f32, %b1 : f32, %b2 : f32, %b3 : f32):1623 %1 = arith.mulf %b0, %b1 : f321624 %2 = arith.addf %1, %b3 : f321625 linalg.yield %1, %2 : f32, f321626 } -> (tensor<2x4x8xf32>, tensor<2x4xf32>)1627 return %0#0, %0#1 : tensor<2x4x8xf32>, tensor<2x4xf32>1628}1629// CHECK-LABEL: func @mixed_parallel_reduced_results(1630// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: tensor<2x4x8xf32>1631// CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: tensor<2x4xf32>1632// CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: tensor<2x4x8xf32>1633// CHECK-SAME: %[[ARG3:[a-zA-Z0-9]+]]: tensor<2x4xf32>1634// CHECK-DAG: %[[V0:.+]] = vector.transfer_read %[[ARG0]]1635// CHECK-DAG: %[[V1:.+]] = vector.transfer_read %[[ARG1]]1636// CHECK-DAG: %[[V2:.+]] = vector.transfer_read %[[ARG3]]1637// CHECK-DAG: %[[MUL:.+]] = arith.mulf %[[V0]], %[[V1]]1638// CHECK-DAG: %[[ADD:.+]] = vector.multi_reduction <add>, %[[MUL]], %[[V2]]1639// CHECK-DAG: vector.transfer_write %[[MUL]], %[[ARG2]]1640// CHECK-DAG: vector.transfer_write %[[ADD]], %[[ARG3]]1641 1642module attributes {transform.with_named_sequence} {1643 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1644 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1645 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1646 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op1647 transform.yield1648 }1649}1650 1651// -----1652 1653// This is a regression test. This IR cannot be vectorized, but1654// structured.vectorize_children_and_apply_patterns should nevertheless succeed.1655 1656#map = affine_map<(d0) -> (d0)>1657// CHECK-LABEL: @not_vectorizable1658func.func @not_vectorizable(%arg0: tensor<1x?xf32>, %arg1: index, %arg2: index, %arg3: index) -> tensor<1x128xf32> {1659 %c0 = arith.constant 0 : index1660 %0 = tensor.empty() : tensor<1x128xf32>1661 %1 = scf.for %arg5 = %arg2 to %arg1 step %arg3 iter_args(%arg6 = %0) -> (tensor<1x128xf32>) {1662 %extracted_slice = tensor.extract_slice %arg6[0, 0] [1, %arg1] [1, 1] : tensor<1x128xf32> to tensor<?xf32>1663 %sz0 = tensor.dim %extracted_slice, %c0 : tensor<?xf32>1664 %expanded = tensor.expand_shape %extracted_slice [[0, 1]] output_shape [1, %sz0] : tensor<?xf32> into tensor<1x?xf32>1665 %extracted_slice_0 = tensor.extract_slice %arg0[0, %arg3] [1, %arg2] [1, 1] : tensor<1x?xf32> to tensor<?xf32>1666 %extracted_slice_1 = tensor.extract_slice %expanded[0, %arg3] [1, %arg2] [1, 1] : tensor<1x?xf32> to tensor<?xf32>1667 %2 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%extracted_slice_0 : tensor<?xf32>) outs(%extracted_slice_1 : tensor<?xf32>) {1668 ^bb0(%in: f32, %out: f32):1669 %3 = arith.addf %in, %out : f321670 linalg.yield %3 : f321671 } -> tensor<?xf32>1672 %inserted_slice = tensor.insert_slice %2 into %expanded[0, %arg3] [1, %arg2] [1, 1] : tensor<?xf32> into tensor<1x?xf32>1673 %collapsed = tensor.collapse_shape %inserted_slice [[0, 1]] : tensor<1x?xf32> into tensor<?xf32>1674 %inserted_slice_2 = tensor.insert_slice %collapsed into %arg6[0, 0] [1, %arg1] [1, 1] : tensor<?xf32> into tensor<1x128xf32>1675 scf.yield %inserted_slice_2 : tensor<1x128xf32>1676 }1677 return %1 : tensor<1x128xf32>1678}1679module attributes {transform.with_named_sequence} {1680 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {1681 %0 = transform.structured.match ops{["func.func"]} in %arg0 : (!transform.any_op) -> !transform.any_op1682 %1 = transform.structured.vectorize_children_and_apply_patterns %0 : (!transform.any_op) -> !transform.any_op1683 transform.yield1684 }1685}1686 1687// -----1688 1689// Regression test: %13 was incorrectly detected as a reduction and1690// vectorization failed.1691 1692func.func @wrong_reduction_detection(%input: tensor<120x64xf32>) -> tensor<120x64xf32> {1693 %c0 = arith.constant 0 : index1694 %c4 = arith.constant 4 : index1695 %c64 = arith.constant 64 : index1696 %cst_6 = arith.constant 4.000000e+00 : f321697 %1 = scf.for %arg0 = %c0 to %c64 step %c4 iter_args(%arg1 = %input) -> (tensor<120x64xf32>) {1698 %extracted_slice = tensor.extract_slice %arg1[%c0, %arg0] [1, 4] [1, 1] : tensor<120x64xf32> to tensor<1x4xf32>1699 %10 = linalg.fill {__internal_linalg_transform__ = "1"} ins(%cst_6 : f32) outs(%extracted_slice : tensor<1x4xf32>) -> tensor<1x4xf32>1700 %11 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} outs(%10 : tensor<1x4xf32>) {1701 ^bb0(%out: f32):1702 %12 = linalg.index 0 : index1703 %13 = arith.addi %arg0, %12 : index1704 %18 = arith.index_cast %13 : index to i321705 %20 = arith.uitofp %18 : i32 to f321706 %67 = arith.mulf %out, %20 : f321707 linalg.yield %67 : f321708 } -> tensor<1x4xf32>1709 %inserted_slice = tensor.insert_slice %11 into %arg1[%c0, %arg0] [1, 4] [1, 1] : tensor<1x4xf32> into tensor<120x64xf32>1710 scf.yield %inserted_slice : tensor<120x64xf32>1711 }1712 return %1 : tensor<120x64xf32>1713}1714 1715module attributes {transform.with_named_sequence} {1716 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1717 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1718 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1719 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op1720 transform.yield1721 }1722}1723 1724// CHECK-LABEL: @wrong_reduction_detection1725// CHECK: vector.broadcast1726// CHECK: vector.transfer_write1727 1728// -----1729 1730// Don't vectorize tensor<0xf32> : (!transform.any_op) -> !transform.any_op1731// CHECK-LABEL: @tensor_size01732// CHECK: linalg.generic1733func.func @tensor_size0(%arg0: tensor<0xf32>,1734 %arg1: tensor<f32>) -> tensor<f32> {1735 %0 = linalg.generic1736 {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> ()>],1737 iterator_types = ["reduction"]}1738 ins(%arg0 : tensor<0xf32>) outs(%arg1 : tensor<f32>) {1739 ^bb0(%in: f32, %out: f32):1740 %12 = arith.addf %out, %in : f321741 linalg.yield %12 : f321742 } -> tensor<f32>1743 return %0 : tensor<f32>1744}1745 1746module attributes {transform.with_named_sequence} {1747 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1748 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1749 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1750 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op1751 transform.yield1752 }1753}1754 1755// -----1756 1757func.func @zero_dim_tensor(%input: tensor<f32>, %output: tensor<f32>) -> tensor<f32>1758{1759 %0 = linalg.generic { indexing_maps = [ affine_map<() -> ()>, affine_map<() -> ()> ],1760 iterator_types = [] }1761 ins(%input : tensor<f32>)1762 outs(%output : tensor<f32>) {1763 ^bb0(%arg0: f32, %arg1: f32):1764 %2 = arith.addf %arg0, %arg1 : f321765 linalg.yield %2 : f321766 } -> tensor<f32>1767 return %0 : tensor<f32>1768}1769 1770module attributes {transform.with_named_sequence} {1771 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1772 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1773 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1774 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1775 transform.yield1776 }1777}1778 1779// CHECK-LABEL: func @zero_dim_tensor1780// CHECK: vector.transfer_read {{.*}} : tensor<f32>, vector<f32>1781// CHECK: vector.extract1782// CHECK: vector.transfer_read {{.*}} : tensor<f32>, vector<f32>1783// CHECK: vector.extract1784// CHECK: arith.addf {{.*}} : f321785// CHECK: vector.broadcast %{{.*}} : f32 to vector<f32>1786// CHECK: vector.transfer_write {{.*}} : vector<f32>, tensor<f32>1787 1788// -----1789 1790// Make sure we generate the right transfer writes for multi-output generic ops1791// with different permutation maps.1792 1793func.func @multi_output_generic_different_perm_maps(%in0: tensor<4x1xf32>,1794 %out0: tensor<4x1xf32>,1795 %out1: tensor<1x4xf32>) -> (tensor<4x1xf32>, tensor<1x4xf32>) {1796 %13:2 = linalg.generic {indexing_maps = [ affine_map<(d0, d1) -> (d1, d0)>,1797 affine_map<(d0, d1) -> (d1, d0)>,1798 affine_map<(d0, d1) -> (d0, d1)> ],1799 iterator_types = ["parallel", "parallel"]}1800 ins(%in0 : tensor<4x1xf32>)1801 outs(%out0, %out1 : tensor<4x1xf32>, tensor<1x4xf32>) {1802 ^bb0(%in: f32, %out: f32, %out_2: f32):1803 %16 = arith.addf %in, %in : f321804 linalg.yield %16, %16 : f32, f321805 } -> (tensor<4x1xf32>, tensor<1x4xf32>)1806 return %13#0, %13#1 : tensor<4x1xf32>, tensor<1x4xf32>1807}1808 1809module attributes {transform.with_named_sequence} {1810 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1811 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1812 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1813 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op1814 transform.yield1815 }1816}1817 1818// CHECK-LABEL: func @multi_output_generic_different_perm_maps1819// CHECK: %[[VAL_5:.*]] = vector.transfer_read %{{.*}} {in_bounds = [true, true]} : tensor<4x1xf32>, vector<4x1xf32>1820// CHECK: %[[VAL_6:.*]] = arith.addf %[[VAL_5]], %[[VAL_5]] : vector<4x1xf32>1821// CHECK: %[[VAL_7:.*]] = vector.transpose %[[VAL_6]], [1, 0] : vector<4x1xf32> to vector<1x4xf32>1822// CHECK: %[[VAL_8:.*]] = vector.transpose %[[VAL_7]], [1, 0] : vector<1x4xf32> to vector<4x1xf32>1823// CHECK: vector.transfer_write %[[VAL_8]], %{{.*}} {in_bounds = [true, true]} : vector<4x1xf32>, tensor<4x1xf32>1824// CHECK: vector.transfer_write %[[VAL_7]], %{{.*}} {in_bounds = [true, true]} : vector<1x4xf32>, tensor<1x4xf32>1825 1826// -----1827 1828// Extracted from: https://github.com/llvm/llvm-project/issues/972471829 1830#map = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>1831#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>1832 1833func.func @generic_with_reduction_and_broadcast(%arg0: tensor<1x12x197x197xf32>) -> (tensor<1x12x197x1xf32>) {1834 %0 = tensor.empty() : tensor<1x12x197x1xf32>1835 %1 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%arg0 : tensor<1x12x197x197xf32>) outs(%0 : tensor<1x12x197x1xf32>) {1836 ^bb0(%in: f32, %out: f32):1837 %818 = arith.addf %in, %out : f321838 linalg.yield %818 : f321839 } -> tensor<1x12x197x1xf32>1840 return %1 : tensor<1x12x197x1xf32>1841}1842module attributes {transform.with_named_sequence} {1843 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {1844 %0 = transform.structured.match ops{["linalg.generic"]} in %arg0 : (!transform.any_op) -> !transform.any_op1845 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1846 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op1847 transform.yield1848 }1849}1850 1851// CHECK: #[[$ATTR_32:.+]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2)>1852 1853// CHECK-LABEL: func.func @generic_with_reduction_and_broadcast(1854// CHECK-SAME: %[[VAL_0:.*]]: tensor<1x12x197x197xf32>) -> tensor<1x12x197x1xf32> {1855// CHECK: %[[VAL_1:.*]] = ub.poison : f321856// CHECK: %[[VAL_2:.*]] = arith.constant 0 : index1857// CHECK: %[[VAL_3:.*]] = tensor.empty() : tensor<1x12x197x1xf32>1858// CHECK: %[[VAL_4:.*]] = vector.transfer_read %[[VAL_0]]{{\[}}%[[VAL_2]], %[[VAL_2]], %[[VAL_2]], %[[VAL_2]]], %[[VAL_1]] {in_bounds = [true, true, true, true]} : tensor<1x12x197x197xf32>, vector<1x12x197x197xf32>1859// CHECK: %[[VAL_5:.*]] = vector.transfer_read %[[VAL_3]]{{\[}}%[[VAL_2]], %[[VAL_2]], %[[VAL_2]], %[[VAL_2]]], %[[VAL_1]] {in_bounds = [true, true, true], permutation_map = #[[$ATTR_32]]} : tensor<1x12x197x1xf32>, vector<1x12x197xf32>1860// CHECK: %[[VAL_6:.*]] = vector.multi_reduction <add>, %[[VAL_4]], %[[VAL_5]] [3] : vector<1x12x197x197xf32> to vector<1x12x197xf32>1861// CHECK: %[[VAL_7:.*]] = vector.broadcast %[[VAL_6]] : vector<1x12x197xf32> to vector<1x1x12x197xf32>1862// CHECK: %[[VAL_8:.*]] = vector.transpose %[[VAL_7]], [1, 2, 3, 0] : vector<1x1x12x197xf32> to vector<1x12x197x1xf32>1863// CHECK: %[[VAL_9:.*]] = vector.transfer_write %[[VAL_8]], %[[VAL_3]]{{\[}}%[[VAL_2]], %[[VAL_2]], %[[VAL_2]], %[[VAL_2]]] {in_bounds = [true, true, true, true]} : vector<1x12x197x1xf32>, tensor<1x12x197x1xf32>1864// CHECK: return %[[VAL_9]] : tensor<1x12x197x1xf32>1865 1866// -----1867 1868// CHECK-LABEL: func @float_mixed_precision_matmul_as_generic1869// CHECK-COUNT-3: vector.transfer_read1870// CHECK-NOT: arith.extf1871// CHECK: vector.contract {{.*}} : vector<8x16xbf16>, vector<16x32xbf16> into vector<8x32xf32>1872// CHECK: vector.transfer_write1873func.func @float_mixed_precision_matmul_as_generic(%A: memref<8x16xbf16>, %B: memref<16x32xbf16>,1874 %C: memref<8x32xf32>) {1875 linalg.generic {1876 indexing_maps = [1877 affine_map<(m, n, k) -> (m, k)>,1878 affine_map<(m, n, k) -> (k, n)>,1879 affine_map<(m, n, k) -> (m, n)>1880 ],1881 iterator_types = ["parallel", "parallel", "reduction"]1882 }1883 ins(%A, %B : memref<8x16xbf16>, memref<16x32xbf16>)1884 outs(%C : memref<8x32xf32>) {1885 ^bb(%in: bf16, %in_0: bf16, %c: f32) :1886 %a = arith.extf %in : bf16 to f321887 %b = arith.extf %in_0 : bf16 to f321888 %d = arith.mulf %a, %b: f321889 %e = arith.addf %c, %d: f321890 linalg.yield %e : f321891 }1892 return1893}1894 1895module attributes {transform.with_named_sequence} {1896 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1897 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1898 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1899 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { fold_type_extensions_into_contract } : (!transform.any_op) -> !transform.any_op1900 transform.yield1901 }1902}1903 1904// -----1905 1906// CHECK-LABEL: func @integer_mixed_precision_matmul_as_generic1907// CHECK-COUNT-3: vector.transfer_read1908// CHECK-NOT: arith.extsi1909// CHECK: vector.contract {{.*}} : vector<8x16xi8>, vector<16x32xi8> into vector<8x32xi32>1910// CHECK: vector.transfer_write1911func.func @integer_mixed_precision_matmul_as_generic(%A: memref<8x16xi8>, %B: memref<16x32xi8>,1912 %C: memref<8x32xi32>) {1913 linalg.generic {1914 indexing_maps = [1915 affine_map<(m, n, k) -> (m, k)>,1916 affine_map<(m, n, k) -> (k, n)>,1917 affine_map<(m, n, k) -> (m, n)>1918 ],1919 iterator_types = ["parallel", "parallel", "reduction"]1920 }1921 ins(%A, %B : memref<8x16xi8>, memref<16x32xi8>)1922 outs(%C : memref<8x32xi32>) {1923 ^bb(%in: i8, %in_0: i8, %c: i32) :1924 %a = arith.extsi %in : i8 to i321925 %b = arith.extsi %in_0 : i8 to i321926 %d = arith.muli %a, %b: i321927 %e = arith.addi %c, %d: i321928 linalg.yield %e : i321929 }1930 return1931}1932 1933module attributes {transform.with_named_sequence} {1934 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1935 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op1936 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1937 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { fold_type_extensions_into_contract } : (!transform.any_op) -> !transform.any_op1938 transform.yield1939 }1940}1941 1942// -----1943 1944///----------------------------------------------------------------------------------------1945/// Tests for memref.copy1946///----------------------------------------------------------------------------------------1947 1948// CHECK-LABEL: func @test_vectorize_copy1949func.func @test_vectorize_copy(%A : memref<8x16xf32>, %B : memref<8x16xf32>) {1950 // CHECK: %[[V:.*]] = vector.transfer_read {{.*}} : memref<8x16xf32>, vector<8x16xf32>1951 // CHECK: vector.transfer_write %[[V]], {{.*}} : vector<8x16xf32>, memref<8x16xf32>1952 memref.copy %A, %B : memref<8x16xf32> to memref<8x16xf32>1953 return1954}1955 1956module attributes {transform.with_named_sequence} {1957 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1958 %0 = transform.structured.match ops{["memref.copy"]} in %arg1 : (!transform.any_op) -> !transform.any_op1959 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1960 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op1961 transform.yield1962 }1963}1964 1965// -----1966 1967// CHECK-LABEL: func @test_vectorize_copy_0d1968func.func @test_vectorize_copy_0d(%A : memref<f32>, %B : memref<f32>) {1969 // CHECK-SAME: (%[[A:.*]]: memref<f32>, %[[B:.*]]: memref<f32>)1970 // CHECK: %[[V:.*]] = vector.transfer_read %[[A]][]{{.*}} : memref<f32>, vector<f32>1971 // CHECK: %[[val:.*]] = vector.extract %[[V]][] : f32 from vector<f32>1972 // CHECK: %[[VV:.*]] = vector.broadcast %[[val]] : f32 to vector<f32>1973 // CHECK: vector.transfer_write %[[VV]], %[[B]][] : vector<f32>, memref<f32>1974 memref.copy %A, %B : memref<f32> to memref<f32>1975 return1976}1977 1978module attributes {transform.with_named_sequence} {1979 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1980 %0 = transform.structured.match ops{["memref.copy"]} in %arg1 : (!transform.any_op) -> !transform.any_op1981 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op1982 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op1983 transform.yield1984 }1985}1986 1987// -----1988 1989// CHECK-LABEL: func @test_vectorize_copy_complex1990// CHECK-NOT: vector<1991func.func @test_vectorize_copy_complex(%A : memref<8x16xcomplex<f32>>, %B : memref<8x16xcomplex<f32>>) {1992 memref.copy %A, %B : memref<8x16xcomplex<f32>> to memref<8x16xcomplex<f32>>1993 return1994}1995 1996module attributes {transform.with_named_sequence} {1997 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {1998 %0 = transform.structured.match ops{["memref.copy"]} in %arg1 : (!transform.any_op) -> !transform.any_op1999 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op2000 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op2001 transform.yield2002 }2003}2004