1// RUN: mlir-opt %s -test-linalg-transform-patterns=test-linalg-to-vector-patterns -split-input-file | FileCheck %s 2 3// ----- 4 5// CHECK-LABEL: contraction_dot 6func.func @contraction_dot(%A: memref<1584xf32>, %B: memref<1584xf32>, %C: memref<f32>) { 7 8// CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584xf32> 9// CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [0] : vector<1584xf32> to f32 10 linalg.dot ins(%A, %B: memref<1584xf32>, memref<1584xf32>) 11 outs(%C: memref<f32>) 12 return 13} 14 15// ----- 16 17// CHECK-LABEL: contraction_matvec 18func.func @contraction_matvec(%A: memref<1584x1584xf32>, %B: memref<1584xf32>, %C: memref<1584xf32>) { 19 20// CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584x1584xf32> 21// CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [1] : vector<1584x1584xf32> to vector<1584xf32> 22 linalg.matvec ins(%A, %B: memref<1584x1584xf32>, memref<1584xf32>) 23 outs(%C: memref<1584xf32>) 24 return 25} 26 27// ----- 28 29// CHECK-LABEL: contraction_matmul 30func.func @contraction_matmul(%A: memref<1584x1584xf32>, %B: memref<1584x1584xf32>, %C: memref<1584x1584xf32>) { 31// CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584x1584x1584xf32> 32// CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [2] : vector<1584x1584x1584xf32> to vector<1584x1584xf32> 33 linalg.matmul ins(%A, %B: memref<1584x1584xf32>, memref<1584x1584xf32>) 34 outs(%C: memref<1584x1584xf32>) 35 return 36} 37 38// ----- 39 40// CHECK-LABEL: contraction_batch_matmul 41func.func @contraction_batch_matmul(%A: memref<1584x1584x1584xf32>, %B: memref<1584x1584x1584xf32>, %C: memref<1584x1584x1584xf32>) { 42// CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584x1584x1584x1584xf32> 43// CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [3] : vector<1584x1584x1584x1584xf32> to vector<1584x1584x1584xf32> 44 linalg.batch_matmul 45 ins(%A, %B: memref<1584x1584x1584xf32>, memref<1584x1584x1584xf32>) 46 outs(%C: memref<1584x1584x1584xf32>) 47 return 48} 49 50// ----- 51 52#matmul_trait = { 53 args_in = 2, 54 args_out = 1, 55 indexing_maps = [ 56 affine_map<(m, n, k) -> (m, k)>, 57 affine_map<(m, n, k) -> (k, n)>, 58 affine_map<(m, n, k) -> (m, n)> 59 ], 60 iterator_types = ["parallel", "parallel", "reduction"] 61} 62 63// CHECK-LABEL: func @vectorization_test 64func.func @vectorization_test(%A: memref<8x16xf32>, %B: memref<16x32xf32>, 65 %C: memref<8x32xf32>) { 66 // CHECK: vector.transfer_read %{{.*}} : memref<8x16xf32>, vector<8x32x16xf32> 67 // CHECK: vector.transfer_read %{{.*}} : memref<16x32xf32>, vector<8x32x16xf32> 68 // CHECK: %[[ACC:.*]] = vector.transfer_read %{{.*}} : memref<8x32xf32>, vector<8x32xf32> 69 // CHECK: %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} : vector<8x32x16xf32> 70 // CHECK: %[[R:.*]] = vector.multi_reduction <add>, %[[MUL]], %[[ACC]] [2] : vector<8x32x16xf32> to vector<8x32xf32> 71 // CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xf32>, memref<8x32xf32> 72 linalg.generic #matmul_trait 73 ins(%A, %B : memref<8x16xf32>, memref<16x32xf32>) 74 outs(%C : memref<8x32xf32>) { 75 ^bb(%a: f32, %b: f32, %c: f32) : 76 %d = arith.mulf %a, %b: f32 77 %e = arith.addf %c, %d: f32 78 linalg.yield %e : f32 79 } 80 return 81} 82 83// ----- 84 85#matmul_transpose_out_trait = { 86 args_in = 2, 87 args_out = 1, 88 indexing_maps = [ 89 affine_map<(m, n, k) -> (m, k)>, 90 affine_map<(m, n, k) -> (k, n)>, 91 affine_map<(m, n, k) -> (n, m)> 92 ], 93 iterator_types = ["parallel", "parallel", "reduction"] 94} 95 96// CHECK-LABEL: func @generic_output_transpose 97func.func @generic_output_transpose(%A: memref<8x16xf32>, %B: memref<16x32xf32>, 98 %C: memref<32x8xf32>) { 99 // CHECK: vector.transfer_read %{{.*}} : memref<8x16xf32>, vector<8x32x16xf32> 100 // CHECK: vector.transfer_read %{{.*}} : memref<16x32xf32>, vector<8x32x16xf32> 101 // CHECK: %[[ACC:.*]] = vector.transfer_read %{{.*}} : memref<32x8xf32>, vector<8x32xf32> 102 // CHECK: %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} : vector<8x32x16xf32> 103 // CHECK: %[[R:.*]] = vector.multi_reduction <add>, %[[MUL]], %[[ACC]] [2] : vector<8x32x16xf32> to vector<8x32xf32> 104 // CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xf32>, memref<32x8xf32> 105 linalg.generic #matmul_transpose_out_trait 106 ins(%A, %B : memref<8x16xf32>, memref<16x32xf32>) 107 outs(%C : memref<32x8xf32>) { 108 ^bb(%a: f32, %b: f32, %c: f32) : 109 %d = arith.mulf %a, %b: f32 110 %e = arith.addf %c, %d: f32 111 linalg.yield %e : f32 112 } 113 return 114} 115 116// ----- 117 118#map0 = affine_map<(d0, d1, d2) -> (d0, d1, d2)> 119#map1 = affine_map<(d0, d1, d2) -> (d1, d0, d2)> 120// CHECK: #[[MAP:.+]] = affine_map<(d0, d1, d2) -> (d1, d0, d2)> 121// CHECK: func @generic_interchanged_transpose 122func.func @generic_interchanged_transpose(%arg0: tensor<12x128x32xf32>) -> tensor<128x12x32xf32> { 123 // CHECK: %[[IN:.+]] = vector.transfer_read 124 // CHECK: vector.transfer_write %[[IN]], {{.+}} permutation_map = #[[MAP]] 125 %0 = linalg.init_tensor [128, 12, 32] : tensor<128x12x32xf32> 126 %1 = linalg.generic {indexing_maps = [#map0, #map1], 127 iterator_types = ["parallel", "parallel", "parallel"]} 128 ins(%arg0 : tensor<12x128x32xf32>) 129 outs(%0 : tensor<128x12x32xf32>) { 130 ^bb0(%arg1: f32, %arg2: f32): 131 linalg.yield %arg1 : f32 132 } -> tensor<128x12x32xf32> 133 return %1 : tensor<128x12x32xf32> 134} 135 136// ----- 137 138#matmul_trait = { 139 args_in = 2, 140 args_out = 1, 141 indexing_maps = [ 142 affine_map<(m, n, k) -> (m, k)>, 143 affine_map<(m, n, k) -> (k, n)>, 144 affine_map<(m, n, k) -> (m, n)> 145 ], 146 iterator_types = ["parallel", "parallel", "reduction"] 147} 148 149// CHECK-LABEL: func @vectorization_test_integer 150func.func @vectorization_test_integer(%A: memref<8x16xi32>, %B: memref<16x32xi32>, 151 %C: memref<8x32xi32>) { 152 // CHECK: vector.transfer_read %{{.*}} : memref<8x16xi32>, vector<8x32x16xi32> 153 // CHECK: vector.transfer_read %{{.*}} : memref<16x32xi32>, vector<8x32x16xi32> 154 // CHECK: %[[ACC:.*]] = vector.transfer_read %{{.*}} : memref<8x32xi32>, vector<8x32xi32> 155 // CHECK: %[[MUL:.*]] = arith.muli %{{.*}}, %{{.*}} : vector<8x32x16xi32> 156 // CHECK: vector.multi_reduction <add>, %[[MUL]], %[[ACC]] [2] : vector<8x32x16xi32> to vector<8x32xi32> 157 // CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xi32>, memref<8x32xi32> 158 linalg.generic #matmul_trait 159 ins(%A, %B : memref<8x16xi32>, memref<16x32xi32>) 160 outs(%C : memref<8x32xi32>) { 161 ^bb(%a: i32, %b: i32, %c: i32) : 162 %d = arith.muli %a, %b: i32 163 %e = arith.addi %c, %d: i32 164 linalg.yield %e : i32 165 } 166 return 167} 168 169// ----- 170 171// CHECK-LABEL: func @vectorization_test_2 172func.func @vectorization_test_2(%A: memref<8x16xf32>, %B: memref<16x32xf32>, 173 %C: memref<8x32xf32>) { 174 // CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<8x32x16xf32> 175 // CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [2] : vector<8x32x16xf32> to vector<8x32xf32> 176 linalg.matmul 177 ins(%A, %B: memref<8x16xf32>, memref<16x32xf32>) 178 outs(%C: memref<8x32xf32>) 179 return 180} 181 182// ----- 183 184// CHECK-LABEL: func @test_vectorize_scalar_input 185func.func @test_vectorize_scalar_input(%A : memref<8x16xf32>, %arg0 : f32) { 186 // CHECK: %[[V:.*]] = vector.broadcast {{.*}} : f32 to vector<8x16xf32> 187 // CHECK: vector.transfer_write %[[V]], {{.*}} : vector<8x16xf32>, memref<8x16xf32> 188 linalg.generic { 189 indexing_maps = [affine_map<(m, n) -> ()>, affine_map<(m, n) -> (m, n)>], 190 iterator_types = ["parallel", "parallel"]} 191 ins(%arg0 : f32) 192 outs(%A: memref<8x16xf32>) { 193 ^bb(%0: f32, %1: f32) : 194 linalg.yield %0 : f32 195 } 196 return 197} 198 199// ----- 200 201// CHECK-LABEL: func @test_do_not_vectorize_unsupported_element_types 202func.func @test_do_not_vectorize_unsupported_element_types(%A : memref<8x16xcomplex<f32>>, %arg0 : complex<f32>) { 203 // CHECK-NOT: vector.broadcast 204 // CHECK-NOT: vector.transfer_write 205 linalg.generic { 206 indexing_maps = [affine_map<(m, n) -> ()>, affine_map<(m, n) -> (m, n)>], 207 iterator_types = ["parallel", "parallel"]} 208 ins(%arg0 : complex<f32>) 209 outs(%A: memref<8x16xcomplex<f32>>) { 210 ^bb(%0: complex<f32>, %1: complex<f32>) : 211 linalg.yield %0 : complex<f32> 212 } 213 return 214} 215 216// ----- 217 218// CHECK-LABEL: func @test_vectorize_fill 219func.func @test_vectorize_fill(%A : memref<8x16xf32>, %arg0 : f32) { 220 // CHECK: %[[V:.*]] = vector.broadcast {{.*}} : f32 to vector<8x16xf32> 221 // CHECK: vector.transfer_write %[[V]], {{.*}} : vector<8x16xf32>, memref<8x16xf32> 222 linalg.fill ins(%arg0 : f32) outs(%A : memref<8x16xf32>) 223 return 224} 225 226// ----- 227 228// CHECK-LABEL: func @test_vectorize_fill 229func.func @test_vectorize_fill_scalar(%A : memref<f32>, %arg0 : f32) { 230 // CHECK-SAME: (%[[M:.*]]: memref<f32>, %[[val:.*]]: f32) 231 // CHECK: %[[VEC:.*]] = vector.broadcast %[[val]] : f32 to vector<f32> 232 // CHECK: vector.transfer_write %[[VEC]], %[[M]][] : vector<f32>, memref<f32> 233 linalg.fill ins(%arg0 : f32) outs(%A : memref<f32>) 234 return 235} 236 237// ----- 238 239// CHECK-LABEL: func @test_vectorize_copy 240func.func @test_vectorize_copy(%A : memref<8x16xf32>, %B : memref<8x16xf32>) { 241 // CHECK: %[[V:.*]] = vector.transfer_read {{.*}} : memref<8x16xf32>, vector<8x16xf32> 242 // CHECK: vector.transfer_write %[[V]], {{.*}} : vector<8x16xf32>, memref<8x16xf32> 243 memref.copy %A, %B : memref<8x16xf32> to memref<8x16xf32> 244 return 245} 246 247// ----- 248 249// CHECK-LABEL: func @test_vectorize_copy_scalar 250func.func @test_vectorize_copy_scalar(%A : memref<f32>, %B : memref<f32>) { 251 // CHECK-SAME: (%[[A:.*]]: memref<f32>, %[[B:.*]]: memref<f32>) 252 // CHECK: %[[V:.*]] = vector.transfer_read %[[A]][]{{.*}} : memref<f32>, vector<f32> 253 // CHECK: %[[val:.*]] = vector.extractelement %[[V]][] : vector<f32> 254 // CHECK: %[[VV:.*]] = vector.broadcast %[[val]] : f32 to vector<f32> 255 // CHECK: vector.transfer_write %[[VV]], %[[B]][] : vector<f32>, memref<f32> 256 memref.copy %A, %B : memref<f32> to memref<f32> 257 return 258} 259 260// ----- 261 262// CHECK-LABEL: func @test_vectorize_trailing_index 263 // CHECK-SAME: (%[[ARG0:.*]]: memref<1x2x4x8xindex>) 264func.func @test_vectorize_trailing_index(%arg0: memref<1x2x4x8xindex>) { 265 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<[0, 1, 2, 3, 4, 5, 6, 7]> : vector<8xindex> 266 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 267 linalg.generic { 268 indexing_maps = [ 269 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], 270 iterator_types = ["parallel", "parallel", "parallel", "parallel"]} 271 outs(%arg0: memref<1x2x4x8xindex>) { 272 ^bb0(%arg1: index): 273 // CHECK: %[[BCST:.*]] = vector.broadcast %[[CST0]] : vector<8xindex> to vector<1x2x4x8xindex> 274 // CHECK: vector.transfer_write %[[BCST]], %[[ARG0]][%[[C0]], %[[C0]], %[[C0]], %[[C0]]] {{.*}} : vector<1x2x4x8xindex>, memref<1x2x4x8xindex> 275 %0 = linalg.index 3 : index 276 linalg.yield %0 : index 277 } 278 return 279} 280 281// ----- 282 283// CHECK-LABEL: func @test_vectorize_inner_index 284 // CHECK-SAME: (%[[ARG0:.*]]: memref<1x2x4x8xindex>) 285func.func @test_vectorize_inner_index(%arg0: memref<1x2x4x8xindex>) { 286 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<[0, 1]> : vector<2xindex> 287 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 288 linalg.generic { 289 indexing_maps = [ 290 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], 291 iterator_types = ["parallel", "parallel", "parallel", "parallel"]} 292 outs(%arg0: memref<1x2x4x8xindex>) { 293 ^bb0(%arg1: index): 294 // CHECK: %[[BCST:.*]] = vector.broadcast %[[CST0]] : vector<2xindex> to vector<1x8x4x2xindex> 295 // CHECK: %[[TRAN:.*]] = vector.transpose %[[BCST]], [0, 3, 2, 1] : vector<1x8x4x2xindex> to vector<1x2x4x8xindex> 296 // CHECK: vector.transfer_write %[[TRAN]], %[[ARG0]][%[[C0]], %[[C0]], %[[C0]], %[[C0]]] {{.*}} : vector<1x2x4x8xindex>, memref<1x2x4x8xindex> 297 %0 = linalg.index 1 : index 298 linalg.yield %0 : index 299 } 300 return 301} 302 303// ----- 304 305// CHECK-LABEL: func @generic_vectorize 306 // CHECK-SAME: (%[[ARG0:.*]]: memref<4x256xf32>, %[[ARG1:.*]]: memref<4x256xf32>, 307 // CHECK-SAME: %[[ARG2:.*]]: memref<256xf32>, %[[ARG3:.*]]: f32) 308func.func @generic_vectorize(%arg0: memref<4x256xf32>, 309 %arg1: memref<4x256xf32>, 310 %arg2: memref<256xf32>, %i: f32) { 311 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<2.000000e+00> : vector<4x256xf32> 312 // CHECK-DAG: %[[CST1:.*]] = arith.constant dense<1.000000e+00> : vector<4x256xf32> 313 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 314 %c1_f32 = arith.constant 1.0 : f32 315 linalg.generic { 316 args_in = 0 : i64, 317 args_out = 10 : i64, 318 indexing_maps = [ 319 affine_map<(d0, d1) -> (d0, d1)>, 320 affine_map<(d0, d1) -> (d1)>, 321 affine_map<(d0, d1) -> (d0, d1)>, 322 affine_map<(d0, d1) -> (d0, d1)>, 323 affine_map<(d0, d1) -> (d0, d1)>, 324 affine_map<(d0, d1) -> (d0, d1)>, 325 affine_map<(d0, d1) -> (d0, d1)>, 326 affine_map<(d0, d1) -> (d0, d1)>, 327 affine_map<(d0, d1) -> (d0, d1)>, 328 affine_map<(d0, d1) -> (d0, d1)>, 329 affine_map<(d0, d1) -> (d0, d1)>, 330 affine_map<(d0, d1) -> (d0, d1)>], 331 iterator_types = ["parallel", "parallel"]} 332 ins(%arg1, %arg2: memref<4x256xf32>, memref<256xf32>) 333 outs( 334 %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0 : 335 memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>, 336 memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>, 337 memref<4x256xf32>, memref<4x256xf32>) { 338 ^bb0(%arg3 : f32, %arg4 : f32, %arg5: f32, %arg6: f32, %arg7: f32, %arg8: f32, 339 // CHECK: %[[V2:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : memref<4x256xf32>, vector<4x256xf32> 340 // CHECK: %[[V0:.*]] = vector.transfer_read %[[ARG2]][%[[C0]]], {{.*}} : memref<256xf32>, vector<4x256xf32> 341 // CHECK: %[[V3:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : memref<4x256xf32>, vector<4x256xf32> 342 // CHECK: %[[V1:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : memref<4x256xf32>, vector<4x256xf32> 343 %arg9 : f32, %arg10 : f32, %arg11 : f32, %arg12 : f32, %arg13 : f32, 344 %arg14 : f32): 345 // CHECK: %[[ADD:.*]] = arith.addf %[[V0]], %[[V1]] : vector<4x256xf32> 346 %6 = arith.addf %arg4, %arg6 : f32 347 // CHECK: %[[CMP:.*]] = arith.cmpf ogt, %[[V2]], %[[V1]] : vector<4x256xf32> 348 %7 = arith.cmpf ogt, %arg3, %arg6 : f32 349 // CHECK: %[[ARG3B:.*]] = vector.broadcast %[[ARG3]] : f32 to vector<4x256xf32> 350 %8 = arith.constant 2.0 : f32 351 // CHECK: %[[DIV:.*]] = arith.divf %[[V3]], %[[ARG3B]] : vector<4x256xf32> 352 %9 = arith.divf %arg5, %i : f32 353 // CHECK: %[[EXP:.*]] = math.exp2 %[[V3]] : vector<4x256xf32> 354 %10 = math.exp2 %arg5 : f32 355 // CHECK: %[[MUL:.*]] = arith.mulf %[[V3]], %[[CST0]] : vector<4x256xf32> 356 %11 = arith.mulf %arg5, %8 : f32 357 // CHECK: %[[RSQRT:.*]] = math.rsqrt %[[V3]] : vector<4x256xf32> 358 %12 = math.rsqrt %arg5 : f32 359 // CHECK: %[[SEL:.*]] = arith.select %[[CMP]], %[[V3]], %[[V1]] : vector<4x256xi1>, vector<4x256xf32> 360 %13 = arith.select %7, %arg5, %arg6 : f32 361 // CHECK: %[[SUB:.*]] = arith.subf %[[V3]], %[[V0]] : vector<4x256xf32> 362 %14 = arith.subf %arg5, %arg4 : f32 363 // CHECK: %[[TAN:.*]] = math.tanh %[[V3]] : vector<4x256xf32> 364 %15 = math.tanh %arg5 : f32 365 // CHECK: vector.transfer_write %[[ADD]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 366 // CHECK: vector.transfer_write %[[CST0]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 367 // CHECK: vector.transfer_write %[[CST1]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 368 // CHECK: vector.transfer_write %[[DIV]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 369 // CHECK: vector.transfer_write %[[EXP]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 370 // CHECK: vector.transfer_write %[[MUL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 371 // CHECK: vector.transfer_write %[[RSQRT]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 372 // CHECK: vector.transfer_write %[[SEL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 373 // CHECK: vector.transfer_write %[[SUB]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 374 // CHECK: vector.transfer_write %[[TAN]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32> 375 linalg.yield %6, %8, %c1_f32, %9, %10, %11, %12, %13, %14, %15 : f32, f32, 376 f32, f32, f32, f32, f32, f32, f32, f32 377 } 378 return 379} 380 381// ----- 382 383// CHECK-LABEL: func @generic_vectorize_tensor 384// CHECK-SAME: (%[[ARG0:.*]]: tensor<4x256xf32>, %[[ARG1:.*]]: tensor<4x256xf32>, 385// CHECK-SAME: %[[ARG2:.*]]: tensor<256xf32>, %[[ARG3:.*]]: f32) 386func.func @generic_vectorize_tensor(%arg0: tensor<4x256xf32>, 387 %arg1: tensor<4x256xf32>, %arg2: tensor<256xf32>, 388 %i: f32) -> (tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, 389 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, 390 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>) { 391 %c1_f32 = arith.constant 1.0 : f32 392 %r:10 = linalg.generic { 393 indexing_maps = [ 394 affine_map<(d0, d1) -> (d0, d1)>, 395 affine_map<(d0, d1) -> (d1)>, 396 affine_map<(d0, d1) -> (d0, d1)>, 397 affine_map<(d0, d1) -> (d0, d1)>, 398 affine_map<(d0, d1) -> (d0, d1)>, 399 affine_map<(d0, d1) -> (d0, d1)>, 400 affine_map<(d0, d1) -> (d0, d1)>, 401 affine_map<(d0, d1) -> (d0, d1)>, 402 affine_map<(d0, d1) -> (d0, d1)>, 403 affine_map<(d0, d1) -> (d0, d1)>, 404 affine_map<(d0, d1) -> (d0, d1)>, 405 affine_map<(d0, d1) -> (d0, d1)>], 406 iterator_types = ["parallel", "parallel"]} 407 ins(%arg1, %arg2: tensor<4x256xf32>, tensor<256xf32>) 408 outs( 409 %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0 : 410 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, 411 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, 412 tensor<4x256xf32>, tensor<4x256xf32>) { 413 ^bb0(%arg3 : f32, %arg4 : f32, %arg5: f32, %arg6: f32, %arg7: f32, %arg8: f32, 414 %arg9 : f32, %arg10 : f32, %arg11 : f32, %arg12 : f32, %arg13 : f32, 415 %arg14 : f32): 416 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<2.000000e+00> : vector<4x256xf32> 417 // CHECK-DAG: %[[CST1:.*]] = arith.constant dense<1.000000e+00> : vector<4x256xf32> 418 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 419 // CHECK: %[[V2:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x256xf32>, vector<4x256xf32> 420 // CHECK: %[[V0:.*]] = vector.transfer_read %[[ARG2]][%[[C0]]], {{.*}} : tensor<256xf32>, vector<4x256xf32> 421 // CHECK: %[[V3:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x256xf32>, vector<4x256xf32> 422 // CHECK: %[[V1:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x256xf32>, vector<4x256xf32> 423 // CHECK: %[[ADD:.*]] = arith.addf %[[V0]], %[[V1]] : vector<4x256xf32> 424 %6 = arith.addf %arg4, %arg6 : f32 425 // CHECK: %[[CMP:.*]] = arith.cmpf ogt, %[[V2]], %[[V1]] : vector<4x256xf32> 426 %7 = arith.cmpf ogt, %arg3, %arg6 : f32 427 // CHECK: %[[ARG3B:.*]] = vector.broadcast %[[ARG3]] : f32 to vector<4x256xf32> 428 %8 = arith.constant 2.0 : f32 429 // CHECK: %[[DIV:.*]] = arith.divf %[[V3]], %[[ARG3B]] : vector<4x256xf32> 430 %9 = arith.divf %arg5, %i : f32 431 // CHECK: %[[EXP:.*]] = math.exp2 %[[V3]] : vector<4x256xf32> 432 %10 = math.exp2 %arg5 : f32 433 // CHECK: %[[MUL:.*]] = arith.mulf %[[V3]], %[[CST0]] : vector<4x256xf32> 434 %11 = arith.mulf %arg5, %8 : f32 435 // CHECK: %[[RSQRT:.*]] = math.rsqrt %[[V3]] : vector<4x256xf32> 436 %12 = math.rsqrt %arg5 : f32 437 // CHECK: %[[SEL:.*]] = arith.select %[[CMP]], %[[V3]], %[[V1]] : vector<4x256xi1>, vector<4x256xf32> 438 %13 = arith.select %7, %arg5, %arg6 : f32 439 // CHECK: %[[SUB:.*]] = arith.subf %[[V3]], %[[V0]] : vector<4x256xf32> 440 %14 = arith.subf %arg5, %arg4 : f32 441 // CHECK: %[[TAN:.*]] = math.tanh %[[V3]] : vector<4x256xf32> 442 %15 = math.tanh %arg5 : f32 443 // CHECK: %[[R0:.*]] = vector.transfer_write %[[ADD]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 444 // CHECK: %[[R1:.*]] = vector.transfer_write %[[CST0]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 445 // CHECK: %[[R2:.*]] = vector.transfer_write %[[CST1]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 446 // CHECK: %[[R3:.*]] = vector.transfer_write %[[DIV]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 447 // CHECK: %[[R4:.*]] = vector.transfer_write %[[EXP]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 448 // CHECK: %[[R5:.*]] = vector.transfer_write %[[MUL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 449 // CHECK: %[[R6:.*]] = vector.transfer_write %[[RSQRT]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 450 // CHECK: %[[R7:.*]] = vector.transfer_write %[[SEL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 451 // CHECK: %[[R8:.*]] = vector.transfer_write %[[SUB]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 452 // CHECK: %[[R9:.*]] = vector.transfer_write %[[TAN]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32> 453 linalg.yield %6, %8, %c1_f32, %9, %10, %11, %12, %13, %14, %15 : f32, f32, 454 f32, f32, f32, f32, f32, f32, f32, f32 455 } -> (tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, 456 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, 457 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>) 458 // 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> 459 return %r#0, %r#1, %r#2, %r#3, %r#4, %r#5, %r#6, %r#7, %r#8, %r#9: 460 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, 461 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, 462 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32> 463} 464 465// ----- 466 467// CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, 0, 0, d1)> 468// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0) -> (d0, 0, 0, 0)> 469// CHECK-DAG: #[[$MAP2:.*]] = affine_map<(d0) -> (0, 0, d0, 0)> 470// CHECK-DAG: #[[$MAP3:.*]] = affine_map<(d0, d1) -> (d1, 0, d0, 0)> 471// CHECK: func @generic_vectorize_broadcast_transpose 472// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 473// CHECK-DAG: %[[CF:.*]] = arith.constant 0.000000e+00 : f32 474// CHECK: %[[V0:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %[[CF]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP0]]} : memref<4x4xf32>, vector<4x4x4x4xf32> 475// CHECK: %[[V1:.*]] = vector.transfer_read %{{.*}}[%[[C0]]], %[[CF]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP1]]} : memref<4xf32>, vector<4x4x4x4xf32> 476// CHECK: %[[V2:.*]] = vector.transfer_read %{{.*}}[%[[C0]]], %[[CF]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP2]]} : memref<4xf32>, vector<4x4x4x4xf32> 477// CHECK: %[[V3:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %[[CF]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP3]]} : memref<4x4xf32>, vector<4x4x4x4xf32> 478// CHECK: %[[SUB:.*]] = arith.subf %[[V0]], %[[V1]] : vector<4x4x4x4xf32> 479// CHECK: %[[ADD0:.*]] = arith.addf %[[V2]], %[[SUB]] : vector<4x4x4x4xf32> 480// CHECK: %[[ADD1:.*]] = arith.addf %[[V3]], %[[ADD0]] : vector<4x4x4x4xf32> 481// CHECK: vector.transfer_write %[[ADD1]], {{.*}} : vector<4x4x4x4xf32>, memref<4x4x4x4xf32> 482func.func @generic_vectorize_broadcast_transpose( 483 %A: memref<4xf32>, %B: memref<4x4xf32>, %C: memref<4x4x4x4xf32>) { 484 linalg.generic { 485 indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d3)>, 486 affine_map<(d0, d1, d2, d3) -> (d0)>, 487 affine_map<(d0, d1, d2, d3) -> (d2)>, 488 affine_map<(d0, d1, d2, d3) -> (d2, d0)>, 489 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], 490 iterator_types = ["parallel", "parallel", "parallel", "parallel"]} 491 ins(%B, %A, %A, %B: memref<4x4xf32>, memref<4xf32>, memref<4xf32>, memref<4x4xf32>) 492 outs(%C : memref<4x4x4x4xf32>) { 493 ^bb0(%arg0: f32, %arg1: f32, %arg2: f32, %arg3: f32, %arg4: f32): 494 %s = arith.subf %arg0, %arg1 : f32 495 %a = arith.addf %arg2, %s : f32 496 %b = arith.addf %arg3, %a : f32 497 linalg.yield %b : f32 498 } 499 return 500} 501 502// ----- 503 504// Test different input maps. 505#matmul_trait = { 506 indexing_maps = [ 507 affine_map<(d0, d1, d2, d3) -> (d1, d0)>, 508 affine_map<(d0, d1, d2, d3) -> (d3, d1)>, 509 affine_map<(d0, d1, d2, d3) -> (d3, d1, d0, d2)>, 510 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)> 511 ], 512 iterator_types = ["parallel", "parallel", "parallel", "parallel"] 513} 514 515// CHECK-DAG: #[[MAP0:.*]] = affine_map<(d0, d1) -> (d1, d0, 0, 0)> 516// CHECK-DAG: #[[MAP1:.*]] = affine_map<(d0, d1) -> (0, d1, 0, d0)> 517// CHECK-DAG: #[[MAP2:.*]] = affine_map<(d0, d1, d2, d3) -> (d2, d1, d3, d0)> 518// CHECK: func @vectorization_transpose 519// CHECK: vector.transfer_read {{.*}}{in_bounds = [true, true, true, true], permutation_map = #[[MAP0]]} : memref<14x7xf32>, vector<7x14x8x16xf32> 520// CHECK: vector.transfer_read {{.*}}{in_bounds = [true, true, true, true], permutation_map = #[[MAP1]]} : memref<16x14xf32>, vector<7x14x8x16xf32> 521// CHECK: vector.transfer_read {{.*}}{in_bounds = [true, true, true, true], permutation_map = #[[MAP2]]} : memref<16x14x7x8xf32>, vector<7x14x8x16xf32> 522// CHECK: arith.addf {{.*}} : vector<7x14x8x16xf32> 523// CHECK: arith.addf {{.*}} : vector<7x14x8x16xf32> 524// CHECK: vector.transfer_write {{.*}} : vector<7x14x8x16xf32>, memref<7x14x8x16xf32> 525func.func @vectorization_transpose(%A: memref<14x7xf32>, %B: memref<16x14xf32>, 526 %C: memref<16x14x7x8xf32>, %D: memref<7x14x8x16xf32>) { 527 linalg.generic #matmul_trait 528 ins(%A, %B, %C : memref<14x7xf32>, memref<16x14xf32>, memref<16x14x7x8xf32>) 529 outs(%D : memref<7x14x8x16xf32>) { 530 ^bb(%a: f32, %b: f32, %c: f32, %d: f32) : 531 %e = arith.addf %a, %b: f32 532 %f = arith.addf %e, %c: f32 533 linalg.yield %f : f32 534 } 535 return 536} 537 538// ----- 539 540// CHECK-LABEL: func @matmul_tensors 541// CHECK-SAME: (%[[ARG0:.*]]: tensor<8x4xf32>, %[[ARG1:.*]]: tensor<4x12xf32>, 542// CHECK-SAME: %[[ARG2:.*]]: tensor<8x12xf32>) -> tensor<8x12xf32> 543func.func @matmul_tensors( 544 %arg0: tensor<8x4xf32>, %arg1: tensor<4x12xf32>, %arg2: tensor<8x12xf32>) 545 -> tensor<8x12xf32> { 546 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 547 // CHECK-DAG: %[[V0:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<8x4xf32>, vector<8x12x4xf32> 548 // CHECK-DAG: %[[V1:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x12xf32>, vector<8x12x4xf32> 549 // CHECK-DAG: %[[V2:.*]] = vector.transfer_read %[[ARG2]][%[[C0]], %[[C0]]], {{.*}} : tensor<8x12xf32>, vector<8x12xf32> 550 // 551 // linalg matmul lowers gets expanded to a 3D reduction, canonicalization later 552 // convert it to a 2D contract. 553 // CHECK: %[[MUL:.*]] = arith.mulf %[[V0]], %[[V1]] : vector<8x12x4xf32> 554 // CHECK: %[[R:.*]] = vector.multi_reduction <add>, %[[MUL]], %[[V2]] [2] : vector<8x12x4xf32> to vector<8x12xf32> 555 // CHECK: %[[W:.*]] = vector.transfer_write %[[R]], %[[ARG2]][%[[C0]], %[[C0]]] {in_bounds = [true, true]} : vector<8x12xf32>, tensor<8x12xf32> 556 %0 = linalg.matmul ins(%arg0, %arg1: tensor<8x4xf32>, tensor<4x12xf32>) 557 outs(%arg2: tensor<8x12xf32>) 558 -> tensor<8x12xf32> 559 // CHECK: return %[[W]] : tensor<8x12xf32> 560 return %0 : tensor<8x12xf32> 561} 562 563// ----- 564 565// CHECK-LABEL: func @pad_static( 566// CHECK-SAME: %[[ARG0:.*]]: tensor<2x?x2xf32>, %[[PAD:.*]]: f32 567// CHECK-NOT: tensor.pad 568// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 569// CHECK-DAG: %[[C2:.*]] = arith.constant 2 : index 570// CHECK-DAG: %[[INIT:.*]] = linalg.init_tensor [2, 3, 4] : tensor<2x3x4xf32> 571// CHECK-DAG: %[[VEC:.*]] = vector.broadcast %[[PAD]] : f32 to vector<2x3x4xf32> 572// CHECK: %[[FILL:.*]] = vector.transfer_write %[[VEC]], %[[INIT]]{{.*}} : vector<2x3x4xf32>, tensor<2x3x4xf32> 573// CHECK: %[[READ:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]], %[[C0]]], %[[PAD]] {in_bounds = [true, false, true]} : tensor<2x?x2xf32>, vector<2x3x2xf32> 574// CHECK: %[[RESULT:.*]] = vector.transfer_write %[[READ]], %[[FILL]][%[[C0]], %[[C0]], %[[C2]]] {in_bounds = [true, true, true]} : vector<2x3x2xf32>, tensor<2x3x4xf32> 575// CHECK: return %[[RESULT]] 576func.func @pad_static(%arg0: tensor<2x?x2xf32>, %pad_value: f32) -> tensor<2x3x4xf32> { 577 %0 = tensor.pad %arg0 low[0, 0, 2] high[0, 1, 0] { 578 ^bb0(%arg1: index, %arg2: index, %arg3: index): 579 tensor.yield %pad_value : f32 580 } : tensor<2x?x2xf32> to tensor<2x3x4xf32> 581 return %0 : tensor<2x3x4xf32> 582} 583 584// ----- 585 586// CHECK-LABEL: func @pad_static_source( 587// CHECK-SAME: %[[ARG0:.*]]: tensor<2x5x2xf32>, %[[PAD:.*]]: f32 588// CHECK-NOT: tensor.pad 589// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 590// CHECK-DAG: %[[C2:.*]] = arith.constant 2 : index 591// CHECK: %[[INIT:.*]] = linalg.init_tensor [2, 6, 4] : tensor<2x6x4xf32> 592// CHECK: %[[VEC:.*]] = vector.broadcast %[[PAD]] : f32 to vector<2x6x4xf32> 593// CHECK: %[[FILL:.*]] = vector.transfer_write %[[VEC]], %[[INIT]][%[[C0]], %[[C0]], %[[C0]]] {in_bounds = [true, true, true]} : vector<2x6x4xf32>, tensor<2x6x4xf32> 594// CHECK: %[[READ:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]], %[[C0]]], %{{.*}} {in_bounds = [true, true, true]} : tensor<2x5x2xf32>, vector<2x5x2xf32> 595// CHECK: %[[WRITE:.*]] = vector.transfer_write %[[READ]], %[[FILL]][%[[C0]], %[[C0]], %[[C2]]] {in_bounds = [true, true, true]} : vector<2x5x2xf32>, tensor<2x6x4xf32> 596// CHECK: return %[[WRITE]] 597func.func @pad_static_source(%arg0: tensor<2x5x2xf32>, %pad_value: f32) -> tensor<2x6x4xf32> { 598 %0 = tensor.pad %arg0 low[0, 0, 2] high[0, 1, 0] { 599 ^bb0(%arg1: index, %arg2: index, %arg3: index): 600 tensor.yield %pad_value : f32 601 } : tensor<2x5x2xf32> to tensor<2x6x4xf32> 602 return %0 : tensor<2x6x4xf32> 603} 604 605// ----- 606 607// CHECK-LABEL: func @pad_static_dynamic( 608// CHECK-SAME: %[[SRC:.*]]: tensor<1x2x2x?xf32>, %[[LOW:.*]]: index, %[[HIGH:.*]]: index 609// CHECK-NOT: tensor.pad 610// CHECK-DAG: %[[C2:.*]] = arith.constant 2 : index 611// CHECK-DAG: %[[C3:.*]] = arith.constant 3 : index 612// CHECK-DAG: %[[C5:.*]] = arith.constant 5 : index 613// CHECK: %[[V0:.*]] = arith.addi %[[LOW]], %[[C2]] : index 614// CHECK: %[[V1:.*]] = arith.addi %[[V0]], %[[C3]] : index 615// CHECK: %[[V2:.*]] = arith.addi %[[HIGH]], %[[C5]] : index 616// CHECK: %[[DIM3:.*]] = tensor.dim %[[SRC]], %[[C3]] : tensor<1x2x2x?xf32> 617// CHECK: %[[V4:.*]] = arith.addi %[[DIM3]], %[[C3]] : index 618// CHECK: %[[V5:.*]] = arith.addi %[[V4]], %[[C2]] : index 619// CHECK: %[[INIT:.*]] = linalg.init_tensor [6, %[[V1]], %[[V2]], %[[V5]]] : tensor<6x?x?x?xf32> 620// CHECK: %[[FILL:.*]] = linalg.fill ins(%{{.*}} : f32) outs(%[[INIT]] : tensor<6x?x?x?xf32>) -> tensor<6x?x?x?xf32> 621// CHECK: %[[SRCDIM:.*]] = tensor.dim %[[SRC]], %[[C3]] : tensor<1x2x2x?xf32> 622// CHECK: %[[RESULT:.*]] = tensor.insert_slice %[[SRC]] into %[[FILL]][2, %[[LOW]], 3, 3] [1, 2, 2, %[[SRCDIM]]] [1, 1, 1, 1] : tensor<1x2x2x?xf32> into tensor<6x?x?x?xf32> 623// CHECK: return %[[RESULT]] 624func.func @pad_static_dynamic(%arg0: tensor<1x2x2x?xf32>, %low: index, %high: index, 625 %pad_value: f32) -> tensor<6x?x?x?xf32> { 626 %0 = tensor.pad %arg0 low[2, %low, 3, 3] high[3, 3, %high, 2] { 627 ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index): 628 tensor.yield %pad_value : f32 629 } : tensor<1x2x2x?xf32> to tensor<6x?x?x?xf32> 630 return %0 : tensor<6x?x?x?xf32> 631} 632 633// ----- 634 635// CHECK-LABEL: func @pad_and_transfer_read 636// CHECK-SAME: %[[ARG0:.*]]: tensor<5x6xf32> 637// CHECK-NOT: tensor.pad 638// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 639// CHECK-DAG: %[[C5:.*]] = arith.constant 5.0 640// CHECK: %[[RESULT:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], %[[C5]] : tensor<5x6xf32>, vector<7x9xf32> 641// CHECK: return %[[RESULT]] 642func.func @pad_and_transfer_read(%arg0: tensor<5x6xf32>) -> vector<7x9xf32> { 643 %c0 = arith.constant 0 : index 644 %c5 = arith.constant 5.0 : f32 645 %c6 = arith.constant 6.0 : f32 646 %0 = tensor.pad %arg0 low[0, 0] high[5, 7] { 647 ^bb0(%arg1: index, %arg2: index): 648 tensor.yield %c5 : f32 649 } : tensor<5x6xf32> to tensor<10x13xf32> 650 %1 = vector.transfer_read %0[%c0, %c0], %c6 651 : tensor<10x13xf32>, vector<7x9xf32> 652 return %1 : vector<7x9xf32> 653} 654 655// ----- 656 657func.func private @make_vector() -> vector<7x9xf32> 658 659// CHECK-LABEL: func @pad_and_transfer_write_static 660// CHECK-SAME: %[[ARG0:.*]]: tensor<5x6xf32> 661// CHECK-NOT: tensor.pad 662// CHECK: %[[C0:.*]] = arith.constant 0 : index 663// CHECK: %[[VEC0:.*]] = call @make_vector() : () -> vector<7x9xf32> 664// CHECK: %[[RESULT:.*]] = vector.transfer_write %[[VEC0]], %[[ARG0]][%[[C0]], %[[C0]]] : vector<7x9xf32>, tensor<5x6xf32> 665// CHECK: return %[[RESULT]] 666func.func @pad_and_transfer_write_static( 667 %arg0: tensor<5x6xf32>) -> tensor<5x6xf32> { 668 %c0 = arith.constant 0 : index 669 %c5 = arith.constant 5.0 : f32 670 %0 = tensor.pad %arg0 low[0, 0] high[5, 7] { 671 ^bb0(%arg2: index, %arg3: index): 672 tensor.yield %c5 : f32 673 } : tensor<5x6xf32> to tensor<10x13xf32> 674 %1 = call @make_vector() : () -> vector<7x9xf32> 675 %2 = vector.transfer_write %1, %0[%c0, %c0] 676 : vector<7x9xf32>, tensor<10x13xf32> 677 %3 = tensor.extract_slice %2[0, 0] [5, 6] [1, 1] : tensor<10x13xf32> to tensor<5x6xf32> 678 return %3 : tensor<5x6xf32> 679} 680 681// ----- 682 683func.func private @make_vector() -> vector<7x9xf32> 684 685// CHECK-LABEL: func @pad_and_transfer_write_dynamic_static 686// CHECK-SAME: %[[ARG0:.*]]: tensor<?x?xf32>, %[[SIZE:.*]]: index, %[[PADDING:.*]]: index 687// CHECK-NOT: tensor.pad 688// CHECK: %[[C0:.*]] = arith.constant 0 : index 689// CHECK: %[[SUB:.*]] = tensor.extract_slice %[[ARG0]][0, 0] [%[[SIZE]], 6] [1, 1] : tensor<?x?xf32> to tensor<?x6xf32> 690// CHECK: %[[VEC0:.*]] = call @make_vector() : () -> vector<7x9xf32> 691// CHECK: %[[RESULT:.*]] = vector.transfer_write %[[VEC0]], %[[SUB]][%[[C0]], %[[C0]]] : vector<7x9xf32>, tensor<?x6xf32> 692// CHECK: return %[[RESULT]] 693func.func @pad_and_transfer_write_dynamic_static( 694 %arg0: tensor<?x?xf32>, %size: index, %padding: index) -> tensor<?x6xf32> { 695 %c0 = arith.constant 0 : index 696 %c5 = arith.constant 5.0 : f32 697 %s = tensor.extract_slice %arg0[0, 0] [%size, 6] [1, 1] 698 : tensor<?x?xf32> to tensor<?x6xf32> 699 %0 = tensor.pad %s low[0, 0] high[%padding, 7] { 700 ^bb0(%arg2: index, %arg3: index): 701 tensor.yield %c5 : f32 702 } : tensor<?x6xf32> to tensor<?x13xf32> 703 %1 = call @make_vector() : () -> vector<7x9xf32> 704 %2 = vector.transfer_write %1, %0[%c0, %c0] 705 : vector<7x9xf32>, tensor<?x13xf32> 706 %3 = tensor.extract_slice %2[0, 0] [%size, 6] [1, 1] : tensor<?x13xf32> to tensor<?x6xf32> 707 return %3 : tensor<?x6xf32> 708} 709 710// ----- 711 712func.func private @make_vector() -> tensor<12x13xf32> 713 714// CHECK-LABEL: func @pad_and_insert_slice_source 715// CHECK-SAME: %[[ARG0:.*]]: tensor<5x6xf32> 716// CHECK-NOT: tensor.pad 717// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 718// CHECK-DAG: %[[C5:.*]] = arith.constant 5.0 719// CHECK: %[[VEC0:.*]] = call @make_vector() : () -> tensor<12x13xf32> 720// CHECK: %[[READ:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], %[[C5]] : tensor<5x6xf32>, vector<7x9xf32> 721// CHECK: %[[WRITE:.*]] = vector.transfer_write %[[READ]], %[[VEC0]][%[[C0]], %[[C0]]] {in_bounds = [true, true]} : vector<7x9xf32>, tensor<12x13xf32> 722// CHECK: return %[[WRITE]] 723func.func @pad_and_insert_slice_source( 724 %arg0: tensor<5x6xf32>) -> tensor<12x13xf32> { 725 %c0 = arith.constant 0 : index 726 %c5 = arith.constant 5.0 : f32 727 %0 = tensor.pad %arg0 low[0, 0] high[2, 3] { 728 ^bb0(%arg2: index, %arg3: index): 729 tensor.yield %c5 : f32 730 } : tensor<5x6xf32> to tensor<7x9xf32> 731 %1 = call @make_vector() : () -> tensor<12x13xf32> 732 %r = tensor.insert_slice %0 into %1[0, 0][7, 9][1, 1] : tensor<7x9xf32> into tensor<12x13xf32> 733 return %r : tensor<12x13xf32> 734} 735 736// ----- 737 738func.func private @make_vector() -> tensor<12x13xf32> 739 740// CHECK-LABEL: func @pad_and_insert_slice_dest 741// Check the insert slice is not rewritten if the padded result is used by the destination operand. 742// CHECK: %[[T1:.*]] = call @make_vector() : () -> tensor<12x13xf32> 743// CHECK: = tensor.insert_slice %[[T1]] into 744func.func @pad_and_insert_slice_dest( 745 %arg0: tensor<1x5x6xf32>) -> tensor<1x12x13xf32> { 746 %c5 = arith.constant 5.0 : f32 747 %0 = tensor.pad %arg0 low[0, 0, 0] high[0, 7, 7] { 748 ^bb0(%arg2: index, %arg3: index, %arg4: index): 749 tensor.yield %c5 : f32 750 } : tensor<1x5x6xf32> to tensor<1x12x13xf32> 751 %1 = call @make_vector() : () -> tensor<12x13xf32> 752 %r = tensor.insert_slice %1 into %0[0, 0, 0][1, 12, 13][1, 1, 1] : tensor<12x13xf32> into tensor<1x12x13xf32> 753 return %r : tensor<1x12x13xf32> 754} 755 756// ----- 757 758// CHECK-LABEL: func @pad_tensor_non_const_pad_value 759// CHECK-SAME: %[[ARG0:.*]]: tensor<5x6xf32> 760// CHECK-NOT: tensor.pad 761// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 762// CHECK-DAG: %[[C3:.*]] = arith.constant 3 : index 763// CHECK-DAG: %[[C4:.*]] = arith.constant 4 : index 764// CHECK: %[[FILL:.*]] = tensor.generate 765// CHECK: %[[RES:.*]] = arith.mulf 766// CHECK: tensor.yield %[[RES]] : f32 767// CHECK: %[[READ:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], %{{.*}} {in_bounds = [true, true]} : tensor<5x6xf32>, vector<5x6xf32> 768// CHECK: %[[WRITE:.*]] = vector.transfer_write %[[READ]], %[[FILL]][%[[C3]], %[[C4]]] {in_bounds = [true, true]} : vector<5x6xf32>, tensor<12x13xf32> 769// CHECK: return %[[WRITE]] 770func.func @pad_tensor_non_const_pad_value(%arg0: tensor<5x6xf32>) -> tensor<12x13xf32> { 771 %c0 = arith.constant 0 : index 772 %c5 = arith.constant 5.0 : f32 773 %0 = tensor.pad %arg0 low[3, 4] high[4, 3] { 774 ^bb0(%arg1: index, %arg2: index): 775 %i1 = arith.index_cast %arg1 : index to i32 776 %i2 = arith.index_cast %arg2 : index to i32 777 %f1 = arith.sitofp %i1 : i32 to f32 778 %f2 = arith.sitofp %i2 : i32 to f32 779 %m = arith.mulf %f1, %f2 : f32 780 tensor.yield %m : f32 781 } : tensor<5x6xf32> to tensor<12x13xf32> 782 return %0 : tensor<12x13xf32> 783} 784 785// ----- 786 787// CHECK-LABEL: func @sum_exp 788func.func @sum_exp(%input: tensor<4x16x8xf32>, %output: tensor<4x16xf32>) 789 -> tensor<4x16xf32> 790{ 791 // CHECK: vector.transfer_read {{.*}} : tensor<4x16x8xf32>, vector<4x16x8xf32> 792 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<4x16xf32>, vector<4x16xf32> 793 // CHECK: math.exp {{.*}} : vector<4x16x8xf32> 794 // CHECK: vector.multi_reduction <add>, %{{.*}}, %{{.*}} [2] : vector<4x16x8xf32> to vector<4x16xf32> 795 // CHECK: vector.transfer_write {{.*}} : vector<4x16xf32>, tensor<4x16xf32> 796 // CHECK: return {{.*}} : tensor<4x16xf32> 797 %0 = linalg.generic { 798 indexing_maps = [ 799 affine_map<(d0, d1, d2) -> (d0, d1, d2)>, 800 affine_map<(d0, d1, d2) -> (d0, d1)> 801 ], 802 iterator_types = ["parallel", "parallel", "reduction"] 803 } ins(%input : tensor<4x16x8xf32>) outs(%output : tensor<4x16xf32>) { 804 ^bb0(%arg0: f32, %arg1: f32): 805 %1 = math.exp %arg0 : f32 806 %2 = arith.addf %1, %arg1 : f32 807 linalg.yield %2 : f32 808 } -> tensor<4x16xf32> 809 return %0 : tensor<4x16xf32> 810} 811 812// ----- 813 814// CHECK-DAG: #[[$M1:.*]] = affine_map<(d0, d1) -> (d1, d0, 0, 0)> 815// CHECK-DAG: #[[$M2:.*]] = affine_map<(d0, d1) -> (0, 0, d1, d0)> 816// CHECK-DAG: #[[$M3:.*]] = affine_map<(d0, d1) -> (d1, d0)> 817 818// CHECK-LABEL: func @sum_exp_2 819func.func @sum_exp_2(%input: tensor<3x2xf32>, %input_2: tensor<5x4xf32>, %output: tensor<5x2xf32>) 820 -> tensor<5x2xf32> 821{ 822 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true, true, true], permutation_map = #[[$M1]]} : tensor<3x2xf32>, vector<2x3x4x5xf32> 823 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true, true, true], permutation_map = #[[$M2]]} : tensor<5x4xf32>, vector<2x3x4x5xf32> 824 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true], permutation_map = #[[$M3]]} : tensor<5x2xf32>, vector<2x5xf32> 825 // CHECK: math.exp {{.*}} : vector<2x3x4x5xf32> 826 // CHECK: math.exp {{.*}} : vector<2x3x4x5xf32> 827 // CHECK: addf {{.*}} : vector<2x3x4x5xf32> 828 // CHECK: vector.multi_reduction <add>, {{.*}}, %{{.*}} [1, 2] : vector<2x3x4x5xf32> to vector<2x5xf32> 829 // CHECK: vector.transfer_write {{.*}} {in_bounds = [true, true], permutation_map = #[[$M3]]} : vector<2x5xf32>, tensor<5x2xf32> 830 // CHECK: return {{.*}} : tensor<5x2xf32> 831 %0 = linalg.generic { 832 indexing_maps = [ 833 affine_map<(d0, d1, d2, d3) -> (d1, d0)>, 834 affine_map<(d0, d1, d2, d3) -> (d3, d2)>, 835 affine_map<(d0, d1, d2, d3) -> (d3, d0)> 836 ], 837 iterator_types = ["parallel", "reduction", "reduction", "parallel"] 838 } ins(%input, %input_2 : tensor<3x2xf32>, tensor<5x4xf32>) outs(%output : tensor<5x2xf32>) { 839 ^bb0(%arg0: f32, %arg1: f32, %arg2: f32): 840 %1 = math.exp %arg0 : f32 841 %2 = math.exp %arg1 : f32 842 %3 = arith.addf %1, %2 : f32 843 %4 = arith.addf %3, %arg2 : f32 844 linalg.yield %4 : f32 845 } -> tensor<5x2xf32> 846 return %0 : tensor<5x2xf32> 847} 848 849// ----- 850 851// CHECK-LABEL: func @red_max_2d( 852func.func @red_max_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> { 853 // CHECK: %[[CMINF:.+]] = arith.constant dense<-3.402820e+38> : vector<4xf32> 854 // CHECK: linalg.init_tensor [4] : tensor<4xf32> 855 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32> 856 // CHECK: vector.multi_reduction <maxf>, {{.*}}, %[[CMINF]] [1] : vector<4x4xf32> to vector<4xf32> 857 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32> 858 %ident = arith.constant -3.40282e+38 : f32 859 %init = linalg.init_tensor [4] : tensor<4xf32> 860 %fill = linalg.fill ins(%ident : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32> 861 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, 862 affine_map<(d0, d1) -> (d0)>], 863 iterator_types = ["parallel", "reduction"]} 864 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) { 865 ^bb0(%in0: f32, %out0: f32): 866 %max = arith.maxf %in0, %out0 : f32 867 linalg.yield %max : f32 868 } -> tensor<4xf32> 869 return %red : tensor<4xf32> 870} 871 872// ----- 873 874// CHECK-LABEL: func @red_min_2d( 875func.func @red_min_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> { 876 // CHECK: %[[CMAXF:.+]] = arith.constant dense<3.402820e+38> : vector<4xf32> 877 // CHECK: linalg.init_tensor [4] : tensor<4xf32> 878 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32> 879 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xf32>, vector<4x4xf32> 880 // CHECK: vector.multi_reduction <minf>, {{.*}}, %[[CMAXF]] [1] : vector<4x4xf32> to vector<4xf32> 881 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32> 882 %maxf32 = arith.constant 3.40282e+38 : f32 883 %init = linalg.init_tensor [4] : tensor<4xf32> 884 %fill = linalg.fill ins(%maxf32 : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32> 885 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, 886 affine_map<(d0, d1) -> (d0)>], 887 iterator_types = ["parallel", "reduction"]} 888 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) { 889 ^bb0(%in0: f32, %out0: f32): 890 %min = arith.minf %out0, %in0 : f32 891 linalg.yield %min : f32 892 } -> tensor<4xf32> 893 return %red : tensor<4xf32> 894} 895 896// ----- 897 898// CHECK-LABEL: func @red_mul_2d( 899func.func @red_mul_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> { 900 // CHECK: linalg.init_tensor [4] : tensor<4xf32> 901 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32> 902 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xf32>, vector<4x4xf32> 903 // CHECK: vector.multi_reduction <mul>, {{.*}}, {{.*}} [1] : vector<4x4xf32> to vector<4xf32> 904 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32> 905 %ident = arith.constant 1.0 : f32 906 %init = linalg.init_tensor [4] : tensor<4xf32> 907 %fill = linalg.fill ins(%ident : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32> 908 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, 909 affine_map<(d0, d1) -> (d0)>], 910 iterator_types = ["parallel", "reduction"]} 911 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) { 912 ^bb0(%in0: f32, %out0: f32): 913 %mul = arith.mulf %in0, %out0 : f32 914 linalg.yield %mul : f32 915 } -> tensor<4xf32> 916 return %red : tensor<4xf32> 917} 918 919// ----- 920 921// CHECK-LABEL: func @red_or_2d( 922func.func @red_or_2d(%arg0: tensor<4x4xi1>) -> tensor<4xi1> { 923 // CHECK: linalg.init_tensor [4] : tensor<4xi1> 924 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1> 925 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xi1>, vector<4x4xi1> 926 // CHECK: vector.multi_reduction <or>, {{.*}}, {{.*}} [1] : vector<4x4xi1> to vector<4xi1> 927 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1> 928 %ident = arith.constant false 929 %init = linalg.init_tensor [4] : tensor<4xi1> 930 %fill = linalg.fill ins(%ident : i1) outs(%init : tensor<4xi1>) -> tensor<4xi1> 931 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, 932 affine_map<(d0, d1) -> (d0)>], 933 iterator_types = ["parallel", "reduction"]} 934 ins(%arg0 : tensor<4x4xi1>) outs(%fill : tensor<4xi1>) { 935 ^bb0(%in0: i1, %out0: i1): 936 %or = arith.ori %in0, %out0 : i1 937 linalg.yield %or : i1 938 } -> tensor<4xi1> 939 return %red : tensor<4xi1> 940} 941 942// ----- 943 944// CHECK-LABEL: func @red_and_2d( 945func.func @red_and_2d(%arg0: tensor<4x4xi1>) -> tensor<4xi1> { 946 // CHECK: linalg.init_tensor [4] : tensor<4xi1> 947 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1> 948 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xi1>, vector<4x4xi1> 949 // CHECK: vector.multi_reduction <and>, {{.*}}, {{.*}} [1] : vector<4x4xi1> to vector<4xi1> 950 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1> 951 %ident = arith.constant true 952 %init = linalg.init_tensor [4] : tensor<4xi1> 953 %fill = linalg.fill ins(%ident : i1) outs(%init : tensor<4xi1>) -> tensor<4xi1> 954 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, 955 affine_map<(d0, d1) -> (d0)>], 956 iterator_types = ["parallel", "reduction"]} 957 ins(%arg0 : tensor<4x4xi1>) outs(%fill : tensor<4xi1>) { 958 ^bb0(%in0: i1, %out0: i1): 959 %and = arith.andi %in0, %out0 : i1 960 linalg.yield %and : i1 961 } -> tensor<4xi1> 962 return %red : tensor<4xi1> 963} 964 965// ----- 966 967// CHECK-LABEL: func @red_xor_2d( 968func.func @red_xor_2d(%arg0: tensor<4x4xi1>) -> tensor<4xi1> { 969 // CHECK: linalg.init_tensor [4] : tensor<4xi1> 970 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1> 971 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xi1>, vector<4x4xi1> 972 // CHECK: vector.multi_reduction <xor>, {{.*}}, {{.*}} [1] : vector<4x4xi1> to vector<4xi1> 973 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1> 974 %ident = arith.constant false 975 %init = linalg.init_tensor [4] : tensor<4xi1> 976 %fill = linalg.fill ins(%ident : i1) outs(%init : tensor<4xi1>) -> tensor<4xi1> 977 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, 978 affine_map<(d0, d1) -> (d0)>], 979 iterator_types = ["parallel", "reduction"]} 980 ins(%arg0 : tensor<4x4xi1>) outs(%fill : tensor<4xi1>) { 981 ^bb0(%in0: i1, %out0: i1): 982 %xor = arith.xori %in0, %out0 : i1 983 linalg.yield %xor : i1 984 } -> tensor<4xi1> 985 return %red : tensor<4xi1> 986} 987 988// ----- 989 990// CHECK-DAG: #[[$M5:.*]] = affine_map<(d0, d1) -> (d0, 0)> 991 992// CHECK-LABEL: func @explicit_broadcast( 993func.func @explicit_broadcast(%arg0: tensor<4x4xf32>, %arg1: tensor<4x1xf32>) -> tensor<4x4xf32> { 994 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<4x4xf32>, vector<4x4xf32> 995 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true], permutation_map = #[[$M5]]} : tensor<4x1xf32>, vector<4x4xf32> 996 // CHECK: subf {{.*}} : vector<4x4xf32> 997 // CHECK: vector.transfer_write {{.*}} {in_bounds = [true, true]} : vector<4x4xf32>, tensor<4x4xf32> 998 %c0 = arith.constant 0.0 : f32 999 %init = linalg.init_tensor [4, 4] : tensor<4x4xf32> 1000 %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<4x4xf32>) -> tensor<4x4xf32> 1001 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, 1002 affine_map<(d0, d1) -> (d0, 0)>, 1003 affine_map<(d0, d1) -> (d0, d1)>], 1004 iterator_types = ["parallel", "parallel"]} 1005 ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x1xf32>) 1006 outs(%fill : tensor<4x4xf32>) { 1007 ^bb0(%arg7: f32, %arg8: f32, %arg9: f32): 1008 %40 = arith.subf %arg7, %arg8 : f32 1009 linalg.yield %40 : f32 1010 } -> tensor<4x4xf32> 1011 return %red : tensor<4x4xf32> 1012} 1013 1014// ----- 1015 1016// CHECK-DAG: #[[$M6:.*]] = affine_map<(d0, d1) -> (d0, 0)> 1017 1018// CHECK-LABEL: func @fused_broadcast_red_2d 1019func.func @fused_broadcast_red_2d(%arg0: tensor<4x4xf32>, %arg1: tensor<4x1xf32>) -> tensor<4xf32> { 1020 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<4x4xf32>, vector<4x4xf32> 1021 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true], permutation_map = #[[$M6]]} : tensor<4x1xf32>, vector<4x4xf32> 1022 // CHECK: subf {{.*}} : vector<4x4xf32> 1023 // CHECK: math.exp {{.*}} : vector<4x4xf32> 1024 // CHECK: vector.multi_reduction <add>, {{.*}}, {{.*}} : vector<4x4xf32> to vector<4xf32> 1025 // CHECK: vector.transfer_write {{.*}} {in_bounds = [true]} : vector<4xf32>, tensor<4xf32> 1026 %c0 = arith.constant 0.0 : f32 1027 %init = linalg.init_tensor [4] : tensor<4xf32> 1028 %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32> 1029 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, 1030 affine_map<(d0, d1) -> (d0, 0)>, 1031 affine_map<(d0, d1) -> (d0)>], 1032 iterator_types = ["parallel", "reduction"]} 1033 ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x1xf32>) 1034 outs(%fill : tensor<4xf32>) { 1035 ^bb0(%arg7: f32, %arg8: f32, %arg9: f32): 1036 %40 = arith.subf %arg7, %arg8 : f32 1037 %41 = math.exp %40 : f32 1038 %42 = arith.addf %41, %arg9 : f32 1039 linalg.yield %42 : f32 1040 } -> tensor<4xf32> 1041 return %red : tensor<4xf32> 1042} 1043 1044// ----- 1045 1046// CHECK-LABEL: func @reduce_1d( 1047// CHECK-SAME: %[[A:.*]]: tensor<32xf32> 1048func.func @reduce_1d(%arg0: tensor<32xf32>) -> tensor<f32> { 1049 // CHECK-DAG: %[[vF0:.*]] = arith.constant dense<0.000000e+00> : vector<f32> 1050 // CHECK-DAG: %[[F0:.*]] = arith.constant 0.000000e+00 : f32 1051 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index 1052 %f0 = arith.constant 0.000000e+00 : f32 1053 1054 // CHECK: %[[init:.*]] = linalg.init_tensor [] : tensor<f32> 1055 %0 = linalg.init_tensor [] : tensor<f32> 1056 1057 // CHECK: %[[f:.*]] = vector.transfer_write %[[vF0]], %[[init]][] 1058 // CHECK-SAME: : vector<f32>, tensor<f32> 1059 %1 = linalg.fill ins(%f0 : f32) outs(%0 : tensor<f32>) -> tensor<f32> 1060 // CHECK: %[[r:.*]] = vector.transfer_read %[[A]][%[[C0]]] 1061 // CHECK-SAME: : tensor<32xf32>, vector<32xf32> 1062 // CHECK: %[[f0:.*]] = vector.extractelement %[[vF0]][] : vector<f32> 1063 // CHECK: %[[red:.*]] = vector.multi_reduction <add>, %[[r]], %[[f0]] [0] 1064 // CHECK-SAME: : vector<32xf32> to f32 1065 // CHECK: %[[red_v1:.*]] = vector.broadcast %[[red]] : f32 to vector<f32> 1066 // CHECK: %[[res:.*]] = vector.transfer_write %[[red_v1]], %[[f]][] 1067 // CHECK-SAME: : vector<f32>, tensor<f32> 1068 %2 = linalg.generic { 1069 indexing_maps = [affine_map<(d0) -> (d0)>, 1070 affine_map<(d0) -> ()>], 1071 iterator_types = ["reduction"]} 1072 ins(%arg0 : tensor<32xf32>) 1073 outs(%1 : tensor<f32>) { 1074 ^bb0(%a: f32, %b: f32): 1075 %3 = arith.addf %a, %b : f32 1076 linalg.yield %3 : f32 1077 } -> tensor<f32> 1078 1079 return %2 : tensor<f32> 1080} 1081 1082 1083// ----- 1084 1085// This test checks that vectorization does not occur when an input indexing map 1086// is not a projected permutation. In the future, this can be converted to a 1087// positive test when support is added. 1088 1089// CHECK-LABEL: func @not_projected_permutation 1090func.func @not_projected_permutation(%arg0: tensor<8x8xf32>) -> tensor<6x6x3x3xf32> { 1091 %c0 = arith.constant 0.0 : f32 1092 %init = linalg.init_tensor [6, 6, 3, 3] : tensor<6x6x3x3xf32> 1093 %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<6x6x3x3xf32>) -> tensor<6x6x3x3xf32> 1094 // CHECK: linalg.generic 1095 %result = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0 + d2, d1 + d3)>, 1096 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], 1097 iterator_types = ["parallel", "parallel", "parallel", "parallel"]} 1098 ins(%arg0 : tensor<8x8xf32>) 1099 outs(%fill : tensor<6x6x3x3xf32>) { 1100 ^bb0(%arg7: f32, %arg9: f32): 1101 linalg.yield %arg7 : f32 1102 } -> tensor<6x6x3x3xf32> 1103 return %result : tensor<6x6x3x3xf32> 1104} 1105