1// RUN: mlir-opt %s -convert-linalg-to-loops | FileCheck %s 2// RUN: mlir-opt %s -convert-linalg-to-parallel-loops | FileCheck --check-prefix=CHECKPARALLEL %s 3 4// Test that we can lower all the way to LLVM without crashing, don't check results here. 5// RUN: mlir-opt %s -convert-linalg-to-loops -convert-linalg-to-llvm -o=/dev/null 2>&1 6 7// CHECK-DAG: #[[$strided1D:.*]] = affine_map<(d0)[s0] -> (d0 + s0)> 8// CHECK-DAG: #[[$strided2D:.*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)> 9// CHECK-DAG: #[[$strided3D:.*]] = affine_map<(d0, d1, d2)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2 + d2)> 10// CHECK-DAG: #[[$stride1Dilation1:.*]] = affine_map<(d0, d1) -> (d0 + d1)> 11 12// CHECKPARALLEL-DAG: #[[$strided1D:.*]] = affine_map<(d0)[s0] -> (d0 + s0)> 13// CHECKPARALLEL-DAG: #[[$strided2D:.*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)> 14// CHECKPARALLEL-DAG: #[[$strided3D:.*]] = affine_map<(d0, d1, d2)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2 + d2)> 15// CHECKPARALLEL-DAG: #[[$stride1Dilation1:.*]] = affine_map<(d0, d1) -> (d0 + d1)> 16 17func.func @matmul(%arg0: memref<?xi8>, %M: index, %N: index, %K: index) { 18 %c0 = arith.constant 0 : index 19 %c1 = arith.constant 1 : index 20 %A = memref.view %arg0[%c0][%M, %K] : memref<?xi8> to memref<?x?xf32> 21 %B = memref.view %arg0[%c0][%K, %N] : memref<?xi8> to memref<?x?xf32> 22 %C = memref.view %arg0[%c0][%M, %N] : memref<?xi8> to memref<?x?xf32> 23 linalg.matmul ins(%A, %B: memref<?x?xf32>, memref<?x?xf32>) 24 outs(%C: memref<?x?xf32>) 25 return 26} 27// CHECK-LABEL: func @matmul(%{{.*}}: memref<?xi8>, 28// CHECK-SAME: [[M:arg[0-9]+]]: index 29// CHECK-SAME: [[N:arg[0-9]+]]: index 30// CHECK-SAME: [[K:arg[0-9]+]]: index 31// CHECK: %[[A:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 32// CHECK: %[[B:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 33// CHECK: %[[C:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 34// CHECK: scf.for {{.*}} to %[[M]] 35// CHECK: scf.for {{.*}} to %[[N]] 36// CHECK: scf.for {{.*}} to %[[K]] 37// CHECK-DAG: %[[a:.*]] = memref.load %[[A]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 38// CHECK-DAG: %[[b:.*]] = memref.load %[[B]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 39// CHECK-DAG: %[[inc:.*]] = arith.mulf %[[a]], %[[b]] : f32 40// CHECK-DAG: %[[c:.*]] = memref.load %[[C]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 41// CHECK-DAG: %[[res:.*]] = arith.addf %[[c]], %[[inc]] : f32 42// CHECK: store %[[res]], %[[C]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 43 44// CHECKPARALLEL-LABEL: func @matmul(%{{.*}}: memref<?xi8>, 45// CHECKPARALLEL-SAME: [[M:arg[0-9]+]]: index 46// CHECKPARALLEL-SAME: [[N:arg[0-9]+]]: index 47// CHECKPARALLEL-SAME: [[K:arg[0-9]+]]: index 48// CHECKPARALLEL: %[[A:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 49// CHECKPARALLEL: %[[B:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 50// CHECKPARALLEL: %[[C:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 51// CHECKPARALLEL: scf.parallel {{.*}} to (%[[M]], %[[N]]) step (%{{.*}}, %{{.*}} { 52// CHECKPARALLEL: scf.for {{.*}} to %[[K]] 53// CHECKPARALLEL-DAG: %[[a:.*]] = memref.load %[[A]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 54// CHECKPARALLEL-DAG: %[[b:.*]] = memref.load %[[B]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 55// CHECKPARALLEL-DAG: %[[inc:.*]] = arith.mulf %[[a]], %[[b]] : f32 56// CHECKPARALLEL-DAG: %[[c:.*]] = memref.load %[[C]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 57// CHECKPARALLEL-DAG: %[[res:.*]] = arith.addf %[[c]], %[[inc]] : f32 58// CHECKPARALLEL: store %[[res]], %[[C]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 59 60 61 62func.func @matvec(%arg0: memref<?xi8>, %M: index, %N: index) { 63 %c0 = arith.constant 0 : index 64 %c1 = arith.constant 1 : index 65 %2 = memref.view %arg0[%c0][%M, %N] : memref<?xi8> to memref<?x?xf32> 66 %3 = memref.view %arg0[%c0][%M] : memref<?xi8> to memref<?xf32> 67 %4 = memref.view %arg0[%c0][%N] : memref<?xi8> to memref<?xf32> 68 linalg.matvec ins(%2, %3: memref<?x?xf32>, memref<?xf32>) 69 outs(%4 : memref<?xf32>) 70 return 71} 72// CHECK-LABEL: func @matvec(%{{.*}}: memref<?xi8>, 73// CHECK-SAME: [[M:arg[0-9]+]]: index 74// CHECK-SAME: [[K:arg[0-9]+]]: index 75// CHECK: %[[A:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 76// CHECK: %[[B:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?xf32> 77// CHECK: %[[C:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?xf32> 78// CHECK: scf.for {{.*}} to %[[M]] 79// CHECK: scf.for {{.*}} to %[[K]] 80// CHECK-DAG: %[[a:.*]] = memref.load %[[A]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 81// CHECK-DAG: %[[b:.*]] = memref.load %[[B]][%{{.*}}] : memref<?xf32> 82// CHECK-DAG: %[[inc:.*]] = arith.mulf %[[a]], %[[b]] : f32 83// CHECK-DAG: %[[c:.*]] = memref.load %[[C]][%{{.*}}] : memref<?xf32> 84// CHECK-DAG: %[[res:.*]] = arith.addf %[[c]], %[[inc]] : f32 85// CHECK: store %[[res]], %[[C]][%{{.*}}] : memref<?xf32> 86 87// CHECKPARALLEL-LABEL: func @matvec(%{{.*}}: memref<?xi8>, 88// CHECKPARALLEL-SAME: [[M:arg[0-9]+]]: index 89// CHECKPARALLEL-SAME: [[K:arg[0-9]+]]: index 90// CHECKPARALLEL: %[[A:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?x?xf32> 91// CHECKPARALLEL: %[[B:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?xf32> 92// CHECKPARALLEL: %[[C:.*]] = memref.view %{{.*}}[{{.*}}] : memref<?xi8> to memref<?xf32> 93// CHECKPARALLEL: scf.parallel (%{{.*}}) = (%{{.*}}) to (%[[M]]) step (%{{.*}}) { 94// CHECKPARALLEL: scf.for {{.*}} to %[[K]] 95// CHECKPARALLEL-DAG: %[[a:.*]] = memref.load %[[A]][%{{.*}}, %{{.*}}] : memref<?x?xf32> 96// CHECKPARALLEL-DAG: %[[b:.*]] = memref.load %[[B]][%{{.*}}] : memref<?xf32> 97// CHECKPARALLEL-DAG: %[[inc:.*]] = arith.mulf %[[a]], %[[b]] : f32 98// CHECKPARALLEL-DAG: %[[c:.*]] = memref.load %[[C]][%{{.*}}] : memref<?xf32> 99// CHECKPARALLEL-DAG: %[[res:.*]] = arith.addf %[[c]], %[[inc]] : f32 100// CHECKPARALLEL: store %[[res]], %[[C]][%{{.*}}] : memref<?xf32> 101 102 103func.func @dot(%arg0: memref<?xi8>, %M: index) { 104 %c0 = arith.constant 0 : index 105 %c1 = arith.constant 1 : index 106 %1 = memref.view %arg0[%c0][%M] : memref<?xi8> to memref<?xf32> 107 %2 = memref.view %arg0[%c0][%M] : memref<?xi8> to memref<?xf32> 108 %3 = memref.view %arg0[%c0][] : memref<?xi8> to memref<f32> 109 linalg.dot ins(%1, %2 : memref<?xf32>, memref<?xf32>) 110 outs(%3 : memref<f32>) 111 return 112} 113// CHECK-LABEL: func @dot(%{{.*}}: memref<?xi8>, 114// CHECK-SAME: [[K:arg[0-9]+]]: index 115// CHECK: %[[A:.*]] = memref.view %{{.*}}[{{.*}}][{{.*}}] : memref<?xi8> to memref<?xf32> 116// CHECK: %[[B:.*]] = memref.view %{{.*}}[{{.*}}][{{.*}}] : memref<?xi8> to memref<?xf32> 117// CHECK: %[[C:.*]] = memref.view %{{.*}}[{{.*}}][] : memref<?xi8> to memref<f32> 118// CHECK: scf.for {{.*}} to %[[K]] 119// CHECK-DAG: %[[a:.*]] = memref.load %[[A]][%{{.*}}] : memref<?xf32> 120// CHECK-DAG: %[[b:.*]] = memref.load %[[B]][%{{.*}}] : memref<?xf32> 121// CHECK-DAG: %[[inc:.*]] = arith.mulf %[[a]], %[[b]] : f32 122// CHECK-DAG: %[[c:.*]] = memref.load %[[C]][] : memref<f32> 123// CHECK-DAG: %[[res:.*]] = arith.addf %[[c]], %[[inc]] : f32 124// CHECK: store %[[res]], %[[C]][] : memref<f32> 125 126// CHECKPARALLEL-LABEL: func @dot(%{{.*}}: memref<?xi8>, 127// CHECKPARALLEL-SAME: [[K:arg[0-9]+]]: index 128// CHECKPARALLEL: %[[A:.*]] = memref.view %{{.*}}[{{.*}}][{{.*}}] : memref<?xi8> to memref<?xf32> 129// CHECKPARALLEL: %[[B:.*]] = memref.view %{{.*}}[{{.*}}][{{.*}}] : memref<?xi8> to memref<?xf32> 130// CHECKPARALLEL: %[[C:.*]] = memref.view %{{.*}}[{{.*}}][] : memref<?xi8> to memref<f32> 131// CHECKPARALLEL: scf.for {{.*}} to %[[K]] 132// CHECKPARALLEL-DAG: %[[a:.*]] = memref.load %[[A]][%{{.*}}] : memref<?xf32> 133// CHECKPARALLEL-DAG: %[[b:.*]] = memref.load %[[B]][%{{.*}}] : memref<?xf32> 134// CHECKPARALLEL-DAG: %[[inc:.*]] = arith.mulf %[[a]], %[[b]] : f32 135// CHECKPARALLEL-DAG: %[[c:.*]] = memref.load %[[C]][] : memref<f32> 136// CHECKPARALLEL-DAG: %[[res:.*]] = arith.addf %[[c]], %[[inc]] : f32 137// CHECKPARALLEL: store %[[res]], %[[C]][] : memref<f32> 138 139 140func.func @dot_int(%arg0: memref<?xi32>, %arg1: memref<?xi32>, 141 %arg3: memref<i32>) { 142 // Verifies that we use the correct arith operations for integers. 143 linalg.dot ins(%arg0, %arg1 : memref<?xi32>, memref<?xi32>) 144 outs(%arg3 : memref<i32>) 145 return 146} 147// CHECK-LABEL: func @dot_int( 148// CHECK: %[[inc:.*]] = arith.muli {{.*}} : i32 149// CHECK-NEXT: %[[res:.*]] = arith.addi {{.*}}, %[[inc]] : i32 150// CHECK-NEXT: store %[[res]], {{.*}} : memref<i32> 151 152 153func.func @dot_bool(%arg0: memref<?xi1>, %arg1: memref<?xi1>, 154 %arg3: memref<i1>) { 155 // Verifies that we use the correct (saturating) arith operations for booleans. 156 linalg.dot ins(%arg0, %arg1 : memref<?xi1>, memref<?xi1>) 157 outs(%arg3 : memref<i1>) 158 return 159} 160// CHECK-LABEL: func @dot_bool( 161// CHECK: %[[inc:.*]] = arith.andi {{.*}} : i1 162// CHECK-NEXT: %[[res:.*]] = arith.ori {{.*}}, %[[inc]] : i1 163// CHECK-NEXT: store %[[res]], {{.*}} : memref<i1> 164 165 166func.func @dot_view(%arg0: memref<?xf32, offset: ?, strides: [1]>, %arg1: memref<?xf32, offset: ?, strides: [1]>, %arg2: memref<f32>) { 167 linalg.dot ins(%arg0, %arg1 : memref<?xf32, offset: ?, strides: [1]>, 168 memref<?xf32, offset: ?, strides: [1]>) 169 outs(%arg2: memref<f32>) 170 return 171} 172// CHECK-LABEL: func @dot_view( 173// CHECK: %{{.*}}: memref<?xf32, #[[$strided1D]]>, %{{.*}}: memref<?xf32, #[[$strided1D]]>, %{{.*}}: memref<f32>) { 174// CHECK: %[[K:.*]] = memref.dim %arg0, %c0 : memref<?xf32, #[[$strided1D]]> 175// CHECK: scf.for {{.*}} to %[[K]] 176// CHECK-DAG: %[[a:.*]] = memref.load %arg0[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 177// CHECK-DAG: %[[b:.*]] = memref.load %{{.*}}[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 178// CHECK-DAG: %[[inc:.*]] = arith.mulf %[[a]], %[[b]] : f32 179// CHECK-DAG: %[[c:.*]] = memref.load %{{.*}}[] : memref<f32> 180// CHECK-DAG: %[[res:.*]] = arith.addf %[[c]], %[[inc]] : f32 181// CHECK: store %[[res]], %{{.*}}[] : memref<f32> 182 183// CHECKPARALLEL-LABEL: func @dot_view( 184// CHECKPARALLEL: %{{.*}}: memref<?xf32, #[[$strided1D]]>, %{{.*}}: memref<?xf32, #[[$strided1D]]>, %{{.*}}: memref<f32>) { 185// CHECKPARALLEL: %[[K:.*]] = memref.dim %arg0, %c0 : memref<?xf32, #[[$strided1D]]> 186// CHECKPARALLEL: scf.for {{.*}} to %[[K]] 187// CHECKPARALLEL-DAG: %[[a:.*]] = memref.load %arg0[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 188// CHECKPARALLEL-DAG: %[[b:.*]] = memref.load %{{.*}}[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 189// CHECKPARALLEL-DAG: %[[inc:.*]] = arith.mulf %[[a]], %[[b]] : f32 190// CHECKPARALLEL-DAG: %[[c:.*]] = memref.load %{{.*}}[] : memref<f32> 191// CHECKPARALLEL-DAG: %[[res:.*]] = arith.addf %[[c]], %[[inc]] : f32 192// CHECKPARALLEL: store %[[res]], %{{.*}}[] : memref<f32> 193 194func.func @fill_view(%arg0: memref<?xf32, offset: ?, strides: [1]>, %arg1: f32) { 195 linalg.fill ins(%arg1 : f32) outs(%arg0 : memref<?xf32, offset: ?, strides: [1]>) 196 return 197} 198// CHECK-LABEL: func @fill_view( 199// CHECK: %{{.*}}: memref<?xf32, #[[$strided1D]]>, %{{.*}}: f32) { 200// CHECK: scf.for {{.*}} to %{{.*}} 201// CHECK: store %{{.*}}, %{{.*}}[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 202 203// CHECKPARALLEL-LABEL: func @fill_view( 204// CHECKPARALLEL: %{{.*}}: memref<?xf32, #[[$strided1D]]>, %{{.*}}: f32) { 205// CHECKPARALLEL: scf.parallel (%{{.*}}) = (%{{.*}}) to (%{{.*}}) step (%{{.*}}) { 206// CHECKPARALLEL: store %{{.*}}, %{{.*}}[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 207 208func.func @fill_view0(%arg0: memref<f32>, %arg1: f32) { 209 linalg.fill ins(%arg1 : f32) outs(%arg0 : memref<f32>) 210 return 211} 212// CHECK-LABEL: func @fill_view0(%{{.*}}: memref<f32>, %{{.*}}: f32) { 213// CHECK: store %{{.*}}, %{{.*}}[] : memref<f32> 214 215// CHECKPARALLEL-LABEL: func @fill_view0(%{{.*}}: memref<f32>, %{{.*}}: f32) { 216// CHECKPARALLEL: store %{{.*}}, %{{.*}}[] : memref<f32> 217 218func.func @fill_view3(%arg0: memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>, %arg1: f32) { 219 linalg.fill ins(%arg1 : f32) outs(%arg0 : memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>) 220 return 221} 222// CHECK-LABEL: func @fill_view3( 223// CHECK: %{{.*}}: memref<?x?x?xf32, #[[$strided3D]]>, %{{.*}}: f32) { 224// CHECK: scf.for {{.*}} to %{{.*}} 225// CHECK: scf.for {{.*}} to %{{.*}} 226// CHECK: scf.for {{.*}} to %{{.*}} 227// CHECK: store %{{.*}}, {{.*}} : memref<?x?x?xf32, #[[$strided3D]]> 228 229// CHECKPARALLEL-LABEL: func @fill_view3( 230// CHECKPARALLEL: %{{.*}}: memref<?x?x?xf32, #[[$strided3D]]>, %{{.*}}: f32) { 231// CHECKPARALLEL: scf.parallel (%{{.*}}, %{{.*}}, %{{.*}}) = (%{{.*}}, %{{.*}}, %{{.*}}) to (%{{.*}}, %{{.*}}, %{{.*}}) step (%{{.*}}, %{{.*}}, %{{.*}}) { 232// CHECKPARALLEL: store %{{.*}}, {{.*}} : memref<?x?x?xf32, #[[$strided3D]]> 233 234func.func @copy_view(%arg0: memref<?xf32, offset: ?, strides: [1]>, %arg1: memref<?xf32, offset: ?, strides: [1]>) { 235 linalg.generic { 236 iterator_types = ["parallel"], 237 indexing_maps = [ affine_map<(i) -> (i)>, affine_map<(i) -> (i)>] } 238 ins(%arg0: memref<?xf32, offset: ?, strides: [1]>) 239 outs(%arg1: memref<?xf32, offset: ?, strides : [1]>) { 240 ^bb0(%a: f32, %b: f32): 241 linalg.yield %a : f32 242 } 243 return 244} 245// CHECK-LABEL: func @copy_view( 246// CHECK: %{{.*}}: memref<?xf32, #[[$strided1D]]>, %{{.*}}: memref<?xf32, #[[$strided1D]]>) { 247// CHECK: scf.for {{.*}} to %{{.*}} 248// CHECK: %[[L:.*]] = memref.load %{{.*}}[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 249// CHECK: store %[[L]], %{{.*}}[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 250 251// CHECKPARALLEL-LABEL: func @copy_view( 252// CHECKPARALLEL: %{{.*}}: memref<?xf32, #[[$strided1D]]>, %{{.*}}: memref<?xf32, #[[$strided1D]]>) { 253// CHECKPARALLEL: scf.parallel (%{{.*}}) = (%{{.*}}) to (%{{.*}}) step (%{{.*}}) { 254// CHECKPARALLEL: %[[L:.*]] = memref.load %{{.*}}[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 255// CHECKPARALLEL: store %[[L]], %{{.*}}[%{{.*}}] : memref<?xf32, #[[$strided1D]]> 256 257#accesses = [ 258 affine_map<(i, j, k) -> (i, j)>, 259 affine_map<(i, j, k) -> (i, j, k)>, 260 affine_map<(i, j, k) -> (i, k, j)> 261] 262#trait2 = { 263 args_in = 1, 264 args_out = 2, 265 iterator_types = ["parallel", "parallel", "parallel"], 266 indexing_maps = #accesses, 267 library_call = "some_external_function_name_2", 268 doc = "B(i,j,k), C(i,k,j) = foo(A(i, j), B(i,j,k), C(i,k,j))" 269} 270func.func @generic_region(%arg0: memref<?x?xf32, offset: ?, strides: [?, 1]>, %arg1: memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>, %arg2: memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>) { 271 linalg.generic #trait2 272 ins(%arg0: memref<?x?xf32, offset: ?, strides: [?, 1]>) 273 outs(%arg1, %arg2 : memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>, 274 memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>) { 275 ^bb0(%a: f32, %b: f32, %c: f32): 276 %d = arith.mulf %a, %b : f32 277 %e = arith.addf %c, %d : f32 278 linalg.yield %d, %e : f32, f32 279 } 280 return 281} 282// CHECK-LABEL: @generic_region 283// CHECK: scf.for %[[i:.*]] = {{.*}} 284// CHECK: scf.for %[[j:.*]] = {{.*}} 285// CHECK: scf.for %[[k:.*]] = {{.*}} 286// CHECK: %[[a:.*]] = memref.load %{{.*}}[%[[i]], %[[j]]] : memref<?x?xf32, #[[$strided2D]]> 287// CHECK: %[[b:.*]] = memref.load %{{.*}}[%[[i]], %[[j]], %[[k]]] : memref<?x?x?xf32, #[[$strided3D]]> 288// CHECK: %[[c:.*]] = memref.load %{{.*}}[%[[i]], %[[k]], %[[j]]] : memref<?x?x?xf32, #[[$strided3D]]> 289// CHECK: %[[d:.*]] = arith.mulf %[[a]], %[[b]] : f32 290// CHECK: %[[e:.*]] = arith.addf %[[c]], %[[d]] : f32 291// CHECK: store %[[d]], %{{.*}}[%[[i]], %[[j]], %[[k]]] : memref<?x?x?xf32, #[[$strided3D]]> 292// CHECK: store %[[e]], %{{.*}}[%[[i]], %[[k]], %[[j]]] : memref<?x?x?xf32, #[[$strided3D]]> 293 294// CHECKPARALLEL-LABEL: @generic_region 295// CHECKPARALLEL: scf.parallel (%[[i:[a-zA-Z0-9_]*]], %[[j:[a-zA-Z0-9_]*]], %[[k:[a-zA-Z0-9_]*]]) 296// CHECKPARALLEL: %[[a:.*]] = memref.load %{{.*}}[%[[i]], %[[j]]] : memref<?x?xf32, #[[$strided2D]]> 297// CHECKPARALLEL: %[[b:.*]] = memref.load %{{.*}}[%[[i]], %[[j]], %[[k]]] : memref<?x?x?xf32, #[[$strided3D]]> 298// CHECKPARALLEL: %[[c:.*]] = memref.load %{{.*}}[%[[i]], %[[k]], %[[j]]] : memref<?x?x?xf32, #[[$strided3D]]> 299// CHECKPARALLEL: %[[d:.*]] = arith.mulf %[[a]], %[[b]] : f32 300// CHECKPARALLEL: %[[e:.*]] = arith.addf %[[c]], %[[d]] : f32 301// CHECKPARALLEL: store %[[d]], %{{.*}}[%[[i]], %[[j]], %[[k]]] : memref<?x?x?xf32, #[[$strided3D]]> 302// CHECKPARALLEL: store %[[e]], %{{.*}}[%[[i]], %[[k]], %[[j]]] : memref<?x?x?xf32, #[[$strided3D]]> 303 304#trait4 = { 305 args_in = 1, 306 args_out = 2, 307 iterator_types = ["parallel", "parallel", "parallel"], 308 indexing_maps = #accesses, 309 library_call = "some_external_function_name_2", 310 doc = "B(i,j,k), C(i,k,j) = foo(A(i, j) * B(i,j,k), i * j * k + C(i,k,j))" 311} 312func.func @generic_index_region( 313 %arg0: memref<?x?xf32, offset: ?, strides: [?, 1]>, 314 %arg1: memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>, 315 %arg2: memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>) { 316 linalg.generic #trait4 317 ins(%arg0 : memref<?x?xf32, offset: ?, strides: [?, 1]>) 318 outs(%arg1, %arg2 : memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>, 319 memref<?x?x?xf32, offset: ?, strides: [?, ?, 1]>) { 320 ^bb0(%a: f32, %b: f32, %c: f32): 321 %i = linalg.index 0 : index 322 %j = linalg.index 1 : index 323 %k = linalg.index 2 : index 324 %result_1 = arith.mulf %a, %b : f32 325 326 %ij = arith.addi %i, %j : index 327 %ijk = arith.addi %ij, %k : index 328 %ijk_int = arith.index_cast %ijk : index to i32 329 %ijk_float = arith.sitofp %ijk_int : i32 to f32 330 331 %result_2 = arith.addf %c, %ijk_float : f32 332 linalg.yield %result_1, %result_2 : f32, f32 333 } 334 return 335} 336 337// CHECK-LABEL: @generic_index_region 338// CHECK: scf.for %[[i:.*]] = {{.*}} 339// CHECK: scf.for %[[j:.*]] = {{.*}} 340// CHECK: scf.for %[[k:.*]] = {{.*}} 341// CHECK: %[[a:.*]] = memref.load %{{.*}}[%[[i]], %[[j]]] 342// CHECK: %[[b:.*]] = memref.load %{{.*}}[%[[i]], %[[j]], %[[k]]] 343// CHECK: %[[c:.*]] = memref.load %{{.*}}[%[[i]], %[[k]], %[[j]]] 344// CHECK: %[[result_1:.*]] = arith.mulf %[[a]], %[[b]] : f32 345// CHECK: %[[ij:.*]] = arith.addi %[[i]], %[[j]] : index 346// CHECK: %[[ijk:.*]] = arith.addi %[[ij]], %[[k]] : index 347// CHECK: %[[ijk_int:.*]] = arith.index_cast %[[ijk]] : index to i32 348// CHECK: %[[ijk_float:.*]] = arith.sitofp %[[ijk_int]] : i32 to f32 349// CHECK: %[[result_2:.*]] = arith.addf %[[c]], %[[ijk_float]] : f32 350// CHECK: store %[[result_1]], %{{.*}}[%[[i]], %[[j]], %[[k]]] 351// CHECK: store %[[result_2]], %{{.*}}[%[[i]], %[[k]], %[[j]]] 352 353// CHECKPARALLEL-LABEL: @generic_index_region 354// CHECKPARALLEL: scf.parallel (%[[i:[a-zA-Z0-9_]*]], %[[j:[a-zA-Z0-9_]*]], %[[k:[a-zA-Z0-9_]*]]) 355// CHECKPARALLEL: %[[a:.*]] = memref.load %{{.*}}[%[[i]], %[[j]]] 356// CHECKPARALLEL: %[[b:.*]] = memref.load %{{.*}}[%[[i]], %[[j]], %[[k]]] 357// CHECKPARALLEL: %[[c:.*]] = memref.load %{{.*}}[%[[i]], %[[k]], %[[j]]] 358// CHECKPARALLEL: %[[result_1:.*]] = arith.mulf %[[a]], %[[b]] : f32 359// CHECKPARALLEL: %[[ij:.*]] = arith.addi %[[i]], %[[j]] : index 360// CHECKPARALLEL: %[[ijk:.*]] = arith.addi %[[ij]], %[[k]] : index 361// CHECKPARALLEL: %[[ijk_int:.*]] = arith.index_cast %[[ijk]] : index to i32 362// CHECKPARALLEL: %[[ijk_float:.*]] = arith.sitofp %[[ijk_int]] : i32 to f32 363// CHECKPARALLEL: %[[result_2:.*]] = arith.addf %[[c]], %[[ijk_float]] : f32 364// CHECKPARALLEL: store %[[result_1]], %{{.*}}[%[[i]], %[[j]], %[[k]]] 365// CHECKPARALLEL: store %[[result_2]], %{{.*}}[%[[i]], %[[k]], %[[j]]] 366 367// ----- 368 369#broadcast_access = [ 370 affine_map<(i, j) -> ()>, 371 affine_map<(i, j) -> (i, j)> 372] 373 374#trait_broadcast = { 375 args_in = 1, 376 args_out = 1, 377 indexing_maps = #broadcast_access, 378 iterator_types = ["parallel", "parallel"], 379 library_call = "some_broadcast_external_fn" 380} 381 382func.func @generic_op_zero_rank(%arg0: memref<f32>, %arg1: memref<3x4xf32>) 383{ 384 linalg.generic #trait_broadcast 385 ins(%arg0 : memref<f32>) 386 outs(%arg1 : memref<3x4xf32>) { 387 ^bb(%a: f32, %b: f32) : 388 linalg.yield %a : f32 389 } 390 return 391} 392 393// CHECK-LABEL: @generic_op_zero_rank 394// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<f32> 395// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<3x4xf32> 396// CHECK: scf.for %[[i:.*]] = {{.*}} 397// CHECK: scf.for %[[j:.*]] = {{.*}} 398// CHECK: %[[a:.*]] = memref.load %[[ARG0]][] 399// CHECK: store %[[a]], %[[ARG1]][%[[i]], %[[j]]] 400 401// CHECKPARALLEL-LABEL: @generic_op_zero_rank 402// CHECKPARALLEL-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<f32> 403// CHECKPARALLEL-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<3x4xf32> 404// CHECKPARALLEL: scf.parallel (%[[i:[a-zA-Z0-9_]*]], %[[j:[a-zA-Z0-9_]*]]) 405// CHECKPARALLEL: %[[a:.*]] = memref.load %[[ARG0]][] 406// CHECKPARALLEL: store %[[a]], %[[ARG1]][%[[i]], %[[j]]] 407 408func.func @generic_op_scalar(%arg0: f32, %arg1: memref<3x4xf32>) 409{ 410 linalg.generic #trait_broadcast 411 ins(%arg0 : f32) 412 outs(%arg1 : memref<3x4xf32>) { 413 ^bb(%a: f32, %b: f32) : 414 linalg.yield %a : f32 415 } 416 return 417} 418 419// CHECK-LABEL: @generic_op_scalar 420// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: f32 421// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<3x4xf32> 422// CHECK: scf.for %[[i:.*]] = {{.*}} 423// CHECK: scf.for %[[j:.*]] = {{.*}} 424// CHECK: store %[[ARG0]], %[[ARG1]][%[[i]], %[[j]]] 425 426// CHECKPARALLEL-LABEL: @generic_op_scalar 427// CHECKPARALLEL-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: f32 428// CHECKPARALLEL-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<3x4xf32> 429// CHECKPARALLEL: scf.parallel (%[[i:[a-zA-Z0-9_]*]], %[[j:[a-zA-Z0-9_]*]]) 430// CHECKPARALLEL: store %[[ARG0]], %[[ARG1]][%[[i]], %[[j]]] 431 432func.func @generic_index_op_zero_rank(%arg0: memref<i32>, %arg1: memref<3x4xi32>) 433{ 434 linalg.generic #trait_broadcast 435 ins(%arg0 : memref<i32>) 436 outs(%arg1 : memref<3x4xi32>) { 437 ^bb(%a: i32, %b: i32) : 438 %i = linalg.index 0 : index 439 %j = linalg.index 1 : index 440 %ij = arith.addi %i, %j : index 441 %ij_int = arith.index_cast %ij : index to i32 442 %result = arith.addi %a, %ij_int : i32 443 linalg.yield %result : i32 444 } 445 return 446} 447 448// CHECK-LABEL: @generic_index_op_zero_rank 449// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<i32> 450// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<3x4xi32> 451// CHECK: scf.for %[[i:.*]] = {{.*}} 452// CHECK: scf.for %[[j:.*]] = {{.*}} 453// CHECK: %[[a:.*]] = memref.load %[[ARG0]][ 454// CHECK: %[[ij:.*]] = arith.addi %[[i]], %[[j]] : index 455// CHECK: %[[ij_int:.*]] = arith.index_cast %[[ij]] : index to i32 456// CHECK: %[[result:.*]] = arith.addi %[[a]], %[[ij_int]] : i32 457// CHECK: store %[[result]], %[[ARG1]][%[[i]], %[[j]]] 458 459// CHECKPARALLEL-LABEL: @generic_index_op_zero_rank 460// CHECKPARALLEL-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<i32> 461// CHECKPARALLEL-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<3x4xi32> 462// CHECKPARALLEL: scf.parallel (%[[i:[a-zA-Z0-9_]*]], %[[j:[a-zA-Z0-9_]*]]) 463// CHECKPARALLEL: %[[a:.*]] = memref.load %[[ARG0]][ 464// CHECKPARALLEL: %[[ij:.*]] = arith.addi %[[i]], %[[j]] : index 465// CHECKPARALLEL: %[[ij_int:.*]] = arith.index_cast %[[ij]] : index to i32 466// CHECKPARALLEL: %[[result:.*]] = arith.addi %[[a]], %[[ij_int]] : i32 467// CHECKPARALLEL: store %[[result]], %[[ARG1]][%[[i]], %[[j]]] 468 469#reduce_1D_access = [ 470 affine_map<(i) -> (i)>, 471 affine_map<(i) -> ()> 472] 473 474#trait_reduce_1D = { 475 args_in = 1, 476 args_out = 1, 477 indexing_maps = #reduce_1D_access, 478 iterator_types = ["reduction"], 479 library_call = "some_reduce_external_fn" 480} 481 482func.func @generic_op_1D_reduce(%arg0: memref<?xf32>, %arg1: memref<f32>) 483{ 484 linalg.generic #trait_reduce_1D 485 ins(%arg0 : memref<?xf32>) 486 outs(%arg1 : memref<f32>) { 487 ^bb(%a: f32, %b: f32) : 488 %0 = arith.addf %a, %b : f32 489 linalg.yield %0 : f32 490 } 491 return 492} 493// CHECK-LABEL: @generic_op_1D_reduce 494// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<?xf32> 495// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<f32> 496// CHECK: scf.for %[[i:.*]] = {{.*}} 497// CHECK: %[[a:.*]] = memref.load %[[ARG0]][%[[i]]] 498// CHECK: %[[b:.*]] = memref.load %[[ARG1]][] 499// CHECK: %[[c:.*]] = arith.addf %[[a]], %[[b]] : f32 500// CHECK: store %[[c]], %[[ARG1]][] 501 502// CHECKPARALLEL-LABEL: @generic_op_1D_reduce 503// CHECKPARALLEL-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<?xf32> 504// CHECKPARALLEL-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<f32> 505// CHECKPARALLEL: scf.for %[[i:.*]] = {{.*}} 506// CHECKPARALLEL: %[[a:.*]] = memref.load %[[ARG0]][%[[i]]] 507// CHECKPARALLEL: %[[b:.*]] = memref.load %[[ARG1]][] 508// CHECKPARALLEL: %[[c:.*]] = arith.addf %[[a]], %[[b]] : f32 509// CHECKPARALLEL: store %[[c]], %[[ARG1]][] 510 511 512#reduce_init_1D_access = [ 513 affine_map<(i) -> (i)>, 514 affine_map<(i) -> ()>, 515 affine_map<(i) -> ()> 516] 517 518#trait_reduce_init_1D = { 519 args_in = 2, 520 args_out = 1, 521 indexing_maps = #reduce_init_1D_access, 522 iterator_types = ["reduction"], 523 library_call = "some_reduce_external_fn" 524} 525 526func.func @generic_index_op_1D_reduce(%arg0: memref<?xf32>, 527 %arg1: memref<f32>, 528 %arg2: memref<f32>) 529{ 530 linalg.generic #trait_reduce_init_1D 531 ins(%arg0, %arg1 : memref<?xf32>, memref<f32>) 532 outs(%arg2 : memref<f32>) { 533 ^bb(%a: f32, %b: f32, %c: f32) : 534 %i = linalg.index 0 : index 535 %0 = arith.constant 0 : index 536 %1 = arith.cmpi eq, %0, %i : index 537 %2 = arith.select %1, %b, %c : f32 538 %3 = arith.addf %a, %2 : f32 539 linalg.yield %3 : f32 540 } 541 return 542} 543// CHECK-LABEL: @generic_index_op_1D_reduce 544// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<?xf32> 545// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<f32> 546// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]*]]: memref<f32> 547// CHECK: scf.for %[[i:.*]] = {{.*}} 548// CHECK: %[[a:.*]] = memref.load %[[ARG0]][%[[i]]] 549// CHECK: %[[b:.*]] = memref.load %[[ARG1]][] 550// CHECK: %[[c:.*]] = memref.load %[[ARG2]][] 551// CHECK: %[[d:.*]] = arith.select %{{.*}}, %[[b]], %[[c]] 552// CHECK: %[[e:.*]] = arith.addf %[[a]], %[[d]] 553// CHECK: store %[[e]], %[[ARG2]][] 554 555// CHECKPARALLEL-LABEL: @generic_index_op_1D_reduce 556// CHECKPARALLEL-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<?xf32> 557// CHECKPARALLEL-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<f32> 558// CHECKPARALLEL-SAME: %[[ARG2:[a-zA-Z0-9_]*]]: memref<f32> 559// CHECKPARALLEL: scf.for %[[i:.*]] = {{.*}} 560// CHECKPARALLEL: %[[a:.*]] = memref.load %[[ARG0]][%[[i]]] 561// CHECKPARALLEL: %[[b:.*]] = memref.load %[[ARG1]][] 562// CHECKPARALLEL: %[[c:.*]] = memref.load %[[ARG2]][] 563// CHECKPARALLEL: %[[d:.*]] = arith.select %{{.*}}, %[[b]], %[[c]] 564// CHECKPARALLEL: %[[e:.*]] = arith.addf %[[a]], %[[d]] 565// CHECKPARALLEL: store %[[e]], %[[ARG2]][] 566 567#trait_const_fill = { 568 args_in = 0, 569 args_out = 1, 570 indexing_maps = [affine_map<(i) -> (i)>], 571 iterator_types = ["parallel"], 572 library_call = "some_external_fn" 573} 574func.func @generic_const_init(%arg0: memref<?xf32>) { 575 %cst = arith.constant 1.0 : f32 576 linalg.generic #trait_const_fill outs(%arg0 : memref<?xf32>) { 577 ^bb0(%arg1: f32): 578 linalg.yield %cst : f32 579 } 580 return 581} 582// CHECK-LABEL: @generic_const_init 583// CHECK-SAME: %[[ARG0:.*]]: memref<?xf32> 584// CHECK: %[[CONST:.*]] = arith.constant 1.000000e+00 : f32 585// CHECK: scf.for %[[i:.*]] = {{.*}} 586// CHECK: store %[[CONST]], %[[ARG0]] 587 588// CHECKPARALLEL-LABEL: @generic_const_init 589// CHECKPARALLEL-SAME: %[[ARG0:.*]]: memref<?xf32> 590// CHECKPARALLEL: %[[CONST:.*]] = arith.constant 1.000000e+00 : f32 591// CHECKPARALLEL: scf.parallel (%[[i:.*]]) 592// CHECKPARALLEL: store %[[CONST]], %[[ARG0]] 593 594#scalar_access = [ 595 affine_map<() -> ()>, 596 affine_map<() -> ()>, 597 affine_map<() -> ()> 598] 599#scalar_trait = { 600 args_in = 2, 601 args_out = 1, 602 iterator_types = [], 603 indexing_maps = #scalar_access, 604 library_call = "some_external_fn" 605} 606func.func @scalar_code(%arg0: memref<f32>, %arg1 : memref<f32>, %arg2 : memref<f32>, %arg3 : i1) 607{ 608 linalg.generic #scalar_trait 609 ins(%arg0, %arg1 : memref<f32>, memref<f32>) 610 outs(%arg2 : memref<f32>) { 611 ^bb(%a : f32, %b : f32, %c : f32) : 612 %result = scf.if %arg3 -> (f32) { 613 scf.yield %a : f32 614 } else { 615 scf.yield %b : f32 616 } 617 linalg.yield %result : f32 618 } 619 return 620} 621// CHECK-LABEL: @scalar_code 622// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<f32> 623// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<f32> 624// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]*]]: memref<f32> 625// CHECK-NOT: scf.for 626// CHECK: memref.load %[[ARG0]][] 627// CHECK: memref.load %[[ARG1]][] 628// CHECK: scf.if 629// CHECK: scf.yield 630// CHECK: else 631// CHECK: scf.yield 632// CHECK: store %{{.*}}, %[[ARG2]][] 633 634// CHECKPARALLEL-LABEL: @scalar_code 635// CHECKPARALLEL-SAME: %[[ARG0:[a-zA-Z0-9_]*]]: memref<f32> 636// CHECKPARALLEL-SAME: %[[ARG1:[a-zA-Z0-9_]*]]: memref<f32> 637// CHECKPARALLEL-SAME: %[[ARG2:[a-zA-Z0-9_]*]]: memref<f32> 638// CHECKPARALLEL-NOT: scf.for 639// CHECKPARALLEL: memref.load %[[ARG0]][] 640// CHECKPARALLEL: memref.load %[[ARG1]][] 641// CHECKPARALLEL: scf.if 642// CHECKPARALLEL: scf.yield 643// CHECKPARALLEL: else 644// CHECKPARALLEL: scf.yield 645// CHECKPARALLEL: store %{{.*}}, %[[ARG2]][] 646 647//----------------------------------------------------------------------------// 648// Named ops to loops. 649//----------------------------------------------------------------------------// 650func.func @named_batch_matmul(%A: memref<?x?x?xf32>, %B: memref<?x?x?xf32>, %C: memref<?x?x?xf32>) { 651 linalg.batch_matmul ins(%A, %B : memref<?x?x?xf32>, memref<?x?x?xf32>) 652 outs(%C : memref<?x?x?xf32>) 653 return 654} 655// CHECK-LABEL: @named_batch_matmul 656// CHECK-SAME: %[[mA:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 657// CHECK-SAME: %[[mB:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 658// CHECK-SAME: %[[mC:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 659// CHECK: %[[B:.*]] = memref.dim %[[mA]], %c0 : memref<?x?x?xf32> 660// CHECK: %[[M:.*]] = memref.dim %[[mA]], %c1 : memref<?x?x?xf32> 661// CHECK: %[[K:.*]] = memref.dim %[[mA]], %c2 : memref<?x?x?xf32> 662// CHECK: %[[N:.*]] = memref.dim %[[mB]], %c2 : memref<?x?x?xf32> 663// CHECK: scf.for %[[b:.*]] = %{{.*}} to %[[B]] 664// CHECK: scf.for %[[m:.*]] = %{{.*}} to %[[M]] 665// CHECK: scf.for %[[n:.*]] = %{{.*}} to %[[N]] 666// CHECK: scf.for %[[k:.*]] = %{{.*}} to %[[K]] 667// CHECK: %[[va:.*]] = memref.load %[[mA]][%[[b]], %[[m]], %[[k]]] : memref<?x?x?xf32> 668// CHECK: %[[vb:.*]] = memref.load %[[mB]][%[[b]], %[[k]], %[[n]]] : memref<?x?x?xf32> 669// CHECK: %[[vc:.*]] = memref.load %[[mC]][%[[b]], %[[m]], %[[n]]] : memref<?x?x?xf32> 670// CHECK: %[[inc:.*]] = arith.mulf %[[va]], %[[vb]] : f32 671// CHECK: %[[res:.*]] = arith.addf %[[vc]], %[[inc]] : f32 672// CHECK: store %[[res]], %[[mC]][%[[b]], %[[m]], %[[n]]] : memref<?x?x?xf32> 673 674// CHECKPARALLEL-LABEL: @named_batch_matmul 675// CHECKPARALLEL-SAME: %[[mA:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 676// CHECKPARALLEL-SAME: %[[mB:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 677// CHECKPARALLEL-SAME: %[[mC:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 678// CHECKPARALLEL: %[[B:.*]] = memref.dim %[[mA]], %c0 : memref<?x?x?xf32> 679// CHECKPARALLEL: %[[M:.*]] = memref.dim %[[mA]], %c1 : memref<?x?x?xf32> 680// CHECKPARALLEL: %[[K:.*]] = memref.dim %[[mA]], %c2 : memref<?x?x?xf32> 681// CHECKPARALLEL: %[[N:.*]] = memref.dim %[[mB]], %c2 : memref<?x?x?xf32> 682// CHECKPARALLEL: scf.parallel (%[[b:.*]], %[[m:.*]], %[[n:.*]]) = ({{.*}}) to (%[[B]], %[[M]], %[[N]]) step ({{.*}}) { 683// CHECKPARALLEL: scf.for %[[k:.*]] = %{{.*}} to %[[K]] 684// CHECKPARALLEL: %[[va:.*]] = memref.load %[[mA]][%[[b]], %[[m]], %[[k]]] : memref<?x?x?xf32> 685// CHECKPARALLEL: %[[vb:.*]] = memref.load %[[mB]][%[[b]], %[[k]], %[[n]]] : memref<?x?x?xf32> 686// CHECKPARALLEL: %[[vc:.*]] = memref.load %[[mC]][%[[b]], %[[m]], %[[n]]] : memref<?x?x?xf32> 687// CHECKPARALLEL: %[[inc:.*]] = arith.mulf %[[va]], %[[vb]] : f32 688// CHECKPARALLEL: %[[res:.*]] = arith.addf %[[vc]], %[[inc]] : f32 689// CHECKPARALLEL: store %[[res]], %[[mC]][%[[b]], %[[m]], %[[n]]] : memref<?x?x?xf32> 690 691 692func.func @conv1d_no_symbols(%in : memref<?xf32>, %filter : memref<?xf32>, %out : memref<?xf32>) -> () { 693 linalg.conv_1d ins(%in, %filter : memref<?xf32>, memref<?xf32>) 694 outs(%out : memref<?xf32>) 695 return 696} 697 698// CHECK-LABEL: @conv1d_no_symbols 699// CHECK-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?xf32> 700// CHECK-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?xf32> 701// CHECK-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?xf32> 702// CHECK: %[[c0:.*]] = arith.constant 0 : index 703// CHECK: %[[c1:.*]] = arith.constant 1 : index 704// CHECK: %[[dim0:.*]] = memref.dim %[[arg1]], %[[c0]] : memref<?xf32> 705// CHECK: %[[dim1:.*]] = memref.dim %[[arg2]], %[[c0]] : memref<?xf32> 706// CHECK: scf.for %[[b:.*]] = %[[c0]] to %[[dim1]] step %[[c1]] { 707// CHECK: scf.for %[[m:.*]] = %[[c0]] to %[[dim0]] step %[[c1]] { 708// CHECK: %[[aff:.*]] = affine.apply #[[$stride1Dilation1]](%[[b]], %[[m]]) 709// CHECK: %[[vb:.*]] = memref.load %[[arg0]][%[[aff]]] : memref<?xf32> 710// CHECK: %[[va:.*]] = memref.load %[[arg1]][%[[m]]] : memref<?xf32> 711// CHECK: %[[vc:.*]] = memref.load %[[arg2]][%[[b]]] : memref<?xf32> 712// CHECK: %[[inc:.*]] = arith.mulf %[[vb]], %[[va]] : f32 713// CHECK: %[[res:.*]] = arith.addf %[[vc]], %[[inc]] : f32 714// CHECK: store %[[res]], %[[arg2]][%[[b]]] : memref<?xf32> 715 716// CHECKPARALLEL-LABEL: @conv1d_no_symbols 717// CHECKPARALLEL-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?xf32> 718// CHECKPARALLEL-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?xf32> 719// CHECKPARALLEL-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?xf32> 720// CHECKPARALLEL: %[[c0:.*]] = arith.constant 0 : index 721// CHECKPARALLEL: %[[c1:.*]] = arith.constant 1 : index 722// CHECKPARALLEL: %[[dim0:.*]] = memref.dim %[[arg1]], %[[c0]] : memref<?xf32> 723// CHECKPARALLEL: %[[dim1:.*]] = memref.dim %[[arg2]], %[[c0]] : memref<?xf32> 724// CHECKPARALLEL: scf.parallel (%[[b:.*]]) = (%[[c0]]) to (%[[dim1]]) step (%[[c1]]) { 725// CHECKPARALLEL: scf.for %[[m:.*]] = %[[c0]] to %[[dim0]] step %[[c1]] { 726// CHECKPARALLEL: %[[aff:.*]] = affine.apply #[[$stride1Dilation1]](%[[b]], %[[m]]) 727// CHECKPARALLEL: %[[vb:.*]] = memref.load %[[arg0]][%[[aff]]] : memref<?xf32> 728// CHECKPARALLEL: %[[va:.*]] = memref.load %[[arg1]][%[[m]]] : memref<?xf32> 729// CHECKPARALLEL: %[[vc:.*]] = memref.load %[[arg2]][%[[b]]] : memref<?xf32> 730// CHECKPARALLEL: %[[inc:.*]] = arith.mulf %[[vb]], %[[va]] : f32 731// CHECKPARALLEL: %[[res:.*]] = arith.addf %[[vc]], %[[inc]] : f32 732// CHECKPARALLEL: store %[[res]], %[[arg2]][%[[b]]] : memref<?xf32> 733 734 735func.func @conv2d_no_symbols(%in : memref<?x?xf32>, %filter : memref<?x?xf32>, %out : memref<?x?xf32>) -> () { 736 linalg.conv_2d ins(%in, %filter : memref<?x?xf32>, memref<?x?xf32>) 737 outs(%out: memref<?x?xf32>) 738 return 739} 740// CHECK-LABEL: @conv2d_no_symbols 741// CHECK-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?xf32> 742// CHECK-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?xf32> 743// CHECK-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?xf32> 744// CHECK: %[[c0:.*]] = arith.constant 0 : index 745// CHECK: %[[c1:.*]] = arith.constant 1 : index 746// CHECK: %[[dim0:.*]] = memref.dim %[[arg1]], %[[c0]] : memref<?x?xf32> 747// CHECK: %[[dim1:.*]] = memref.dim %[[arg1]], %[[c1]] : memref<?x?xf32> 748// CHECK: %[[dim2:.*]] = memref.dim %[[arg2]], %[[c0]] : memref<?x?xf32> 749// CHECK: %[[dim3:.*]] = memref.dim %[[arg2]], %[[c1]] : memref<?x?xf32> 750// CHECK: scf.for %[[arg3:.*]] = %[[c0]] to %[[dim2]] step %[[c1]] { 751// CHECK: scf.for %[[arg4:.*]] = %[[c0]] to %[[dim3]] step %[[c1]] { 752// CHECK: scf.for %[[arg5:.*]] = %[[c0]] to %[[dim0]] step %[[c1]] { 753// CHECK: scf.for %[[arg6:.*]] = %[[c0]] to %[[dim1]] step %[[c1]] { 754// CHECK: %[[aff:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg3]], %[[arg5]]) 755// CHECK: %[[aff2:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg4]], %[[arg6]]) 756// CHECK: %[[vb:.*]] = memref.load %[[arg0]][%[[aff]], %[[aff2]]] : memref<?x?xf32> 757 758// CHECK: %[[va:.*]] = memref.load %[[arg1]][%[[arg5]], %[[arg6]]] : memref<?x?xf32> 759// CHECK: %[[vc:.*]] = memref.load %[[arg2]][%[[arg3]], %[[arg4]]] : memref<?x?xf32> 760 761// CHECK: %[[inc:.*]] = arith.mulf %[[vb]], %[[va]] : f32 762// CHECK: %[[res:.*]] = arith.addf %[[vc]], %[[inc]] : f32 763// CHECK: store %[[res]], %[[arg2]][%[[arg3]], %[[arg4]]] : memref<?x?xf32> 764 765// CHECKPARALLEL-LABEL: @conv2d_no_symbols 766// CHECKPARALLEL-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?xf32> 767// CHECKPARALLEL-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?xf32> 768// CHECKPARALLEL-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?xf32> 769// CHECKPARALLEL: %[[c0:.*]] = arith.constant 0 : index 770// CHECKPARALLEL: %[[c1:.*]] = arith.constant 1 : index 771// CHECKPARALLEL: %[[dim0:.*]] = memref.dim %[[arg1]], %[[c0]] : memref<?x?xf32> 772// CHECKPARALLEL: %[[dim1:.*]] = memref.dim %[[arg1]], %[[c1]] : memref<?x?xf32> 773// CHECKPARALLEL: %[[dim2:.*]] = memref.dim %[[arg2]], %[[c0]] : memref<?x?xf32> 774// CHECKPARALLEL: %[[dim3:.*]] = memref.dim %[[arg2]], %[[c1]] : memref<?x?xf32> 775// CHECKPARALLEL: scf.parallel (%[[arg3:.*]], %[[arg4:.*]]) = (%[[c0]], %[[c0]]) to (%[[dim2]], %[[dim3]]) step (%[[c1]], %[[c1]]) { 776// CHECKPARALLEL: scf.for %[[arg5:.*]] = %[[c0]] to %[[dim0]] step %[[c1]] { 777// CHECKPARALLEL: scf.for %[[arg6:.*]] = %[[c0]] to %[[dim1]] step %[[c1]] { 778// CHECKPARALLEL: %[[aff:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg3]], %[[arg5]]) 779// CHECKPARALLEL: %[[aff2:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg4]], %[[arg6]]) 780// CHECKPARALLEL: %[[vb:.*]] = memref.load %[[arg0]][%[[aff]], %[[aff2]]] : memref<?x?xf32> 781// CHECKPARALLEL: %[[va:.*]] = memref.load %[[arg1]][%[[arg5]], %[[arg6]]] : memref<?x?xf32> 782// CHECKPARALLEL: %[[vc:.*]] = memref.load %[[arg2]][%[[arg3]], %[[arg4]]] : memref<?x?xf32> 783// CHECKPARALLEL: %[[inc:.*]] = arith.mulf %[[vb]], %[[va]] : f32 784// CHECKPARALLEL: %[[res:.*]] = arith.addf %[[vc]], %[[inc]] : f32 785// CHECKPARALLEL: store %[[res]], %[[arg2]][%[[arg3]], %[[arg4]]] : memref<?x?xf32> 786 787 788func.func @conv3d_no_symbols(%in : memref<?x?x?xf32>, %filter : memref<?x?x?xf32>, %out : memref<?x?x?xf32>) -> () { 789 linalg.conv_3d ins(%in, %filter : memref<?x?x?xf32>, memref<?x?x?xf32>) 790 outs(%out : memref<?x?x?xf32>) 791 return 792} 793 794// CHECK-LABEL: @conv3d_no_symbols 795// CHECK-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 796// CHECK-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 797// CHECK-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 798// CHECK-DAG: %[[c2:.*]] = arith.constant 2 : index 799// CHECK-DAG: %[[c0:.*]] = arith.constant 0 : index 800// CHECK-DAG: %[[c1:.*]] = arith.constant 1 : index 801// CHECK: %[[dim0:.*]] = memref.dim %[[arg1]], %[[c0]] : memref<?x?x?xf32> 802// CHECK: %[[dim1:.*]] = memref.dim %[[arg1]], %[[c1]] : memref<?x?x?xf32> 803// CHECK: %[[dim2:.*]] = memref.dim %[[arg1]], %[[c2]] : memref<?x?x?xf32> 804// CHECK: %[[dim3:.*]] = memref.dim %[[arg2]], %[[c0]] : memref<?x?x?xf32> 805// CHECK: %[[dim4:.*]] = memref.dim %[[arg2]], %[[c1]] : memref<?x?x?xf32> 806// CHECK: %[[dim5:.*]] = memref.dim %[[arg2]], %[[c2]] : memref<?x?x?xf32> 807// CHECK: scf.for %[[arg3:.*]] = %[[c0]] to %[[dim3]] step %[[c1]] { 808// CHECK: scf.for %[[arg4:.*]] = %[[c0]] to %[[dim4]] step %[[c1]] { 809// CHECK: scf.for %[[arg5:.*]] = %[[c0]] to %[[dim5]] step %[[c1]] { 810// CHECK: scf.for %[[arg6:.*]] = %[[c0]] to %[[dim0]] step %[[c1]] { 811// CHECK: scf.for %[[arg7:.*]] = %[[c0]] to %[[dim1]] step %[[c1]] { 812// CHECK: scf.for %[[arg8:.*]] = %[[c0]] to %[[dim2]] step %[[c1]] { 813// CHECK: %[[aff:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg3]], %[[arg6]]) 814// CHECK: %[[aff2:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg4]], %[[arg7]]) 815// CHECK: %[[aff3:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg5]], %[[arg8]]) 816// CHECK: %[[vb:.*]] = memref.load %[[arg0]][%[[aff]], %[[aff2]], %[[aff3]]] : memref<?x?x?xf32> 817 818// CHECK: %[[va:.*]] = memref.load %[[arg1]][%[[arg6]], %[[arg7]], %[[arg8]]] : memref<?x?x?xf32> 819// CHECK: %[[vc:.*]] = memref.load %[[arg2]][%[[arg3]], %[[arg4]], %[[arg5]]] : memref<?x?x?xf32> 820 821// CHECK: %[[inc:.*]] = arith.mulf %[[vb]], %[[va]] : f32 822// CHECK: %[[res:.*]] = arith.addf %[[vc]], %[[inc]] : f32 823// CHECK: store %[[res]], %[[arg2]][%[[arg3]], %[[arg4]], %[[arg5]]] : memref<?x?x?xf32> 824 825// CHECKPARALLEL-LABEL: @conv3d_no_symbols 826// CHECKPARALLEL-SAME: %[[arg0:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 827// CHECKPARALLEL-SAME: %[[arg1:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 828// CHECKPARALLEL-SAME: %[[arg2:[a-zA-Z0-9]+]]: memref<?x?x?xf32> 829// CHECKPARALLEL-DAG: %[[c2:.*]] = arith.constant 2 : index 830// CHECKPARALLEL-DAG: %[[c0:.*]] = arith.constant 0 : index 831// CHECKPARALLEL-DAG: %[[c1:.*]] = arith.constant 1 : index 832// CHECKPARALLEL: %[[dim0:.*]] = memref.dim %[[arg1]], %[[c0]] : memref<?x?x?xf32> 833// CHECKPARALLEL: %[[dim1:.*]] = memref.dim %[[arg1]], %[[c1]] : memref<?x?x?xf32> 834// CHECKPARALLEL: %[[dim2:.*]] = memref.dim %[[arg1]], %[[c2]] : memref<?x?x?xf32> 835// CHECKPARALLEL: %[[dim3:.*]] = memref.dim %[[arg2]], %[[c0]] : memref<?x?x?xf32> 836// CHECKPARALLEL: %[[dim4:.*]] = memref.dim %[[arg2]], %[[c1]] : memref<?x?x?xf32> 837// CHECKPARALLEL: %[[dim5:.*]] = memref.dim %[[arg2]], %[[c2]] : memref<?x?x?xf32> 838// CHECKPARALLEL: scf.parallel (%[[arg3:.*]], %[[arg4:.*]], %[[arg5:.*]]) = (%[[c0]], %[[c0]], %[[c0]]) to (%[[dim3]], %[[dim4]], %[[dim5]]) step (%[[c1]], %[[c1]], %[[c1]]) { 839// CHECKPARALLEL: scf.for %[[arg6:.*]] = %[[c0]] to %[[dim0]] step %[[c1]] { 840// CHECKPARALLEL: scf.for %[[arg7:.*]] = %[[c0]] to %[[dim1]] step %[[c1]] { 841// CHECKPARALLEL: scf.for %[[arg8:.*]] = %[[c0]] to %[[dim2]] step %[[c1]] { 842// CHECKPARALLEL: %[[aff:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg3]], %[[arg6]]) 843// CHECKPARALLEL: %[[aff2:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg4]], %[[arg7]]) 844// CHECKPARALLEL: %[[aff3:.*]] = affine.apply #[[$stride1Dilation1]](%[[arg5]], %[[arg8]]) 845// CHECKPARALLEL: %[[vb:.*]] = memref.load %[[arg0]][%[[aff]], %[[aff2]], %[[aff3]]] : memref<?x?x?xf32> 846// CHECKPARALLEL: %[[va:.*]] = memref.load %[[arg1]][%[[arg6]], %[[arg7]], %[[arg8]]] : memref<?x?x?xf32> 847// CHECKPARALLEL: %[[vc:.*]] = memref.load %[[arg2]][%[[arg3]], %[[arg4]], %[[arg5]]] : memref<?x?x?xf32> 848// CHECKPARALLEL: %[[inc:.*]] = arith.mulf %[[vb]], %[[va]] : f32 849// CHECKPARALLEL: %[[res:.*]] = arith.addf %[[vc]], %[[inc]] : f32 850// CHECKPARALLEL: store %[[res]], %[[arg2]][%[[arg3]], %[[arg4]], %[[arg5]]] : memref<?x?x?xf32> 851 852// ----- 853 854func.func @lower_to_loops_with_rank_reducing_subviews( 855 %arg0 : memref<?xi32>, %arg1 : memref<?x?xi32>, %arg2 : index, 856 %arg3 : index, %arg4 : index) { 857 %0 = memref.subview %arg0[%arg2] [%arg3] [1] 858 : memref<?xi32> to memref<?xi32, offset: ?, strides: [1]> 859 %1 = memref.subview %arg1[0, %arg4] [1, %arg3] [1, 1] 860 : memref<?x?xi32> to memref<?xi32, offset: ?, strides : [1]> 861 linalg.generic { 862 iterator_types = ["parallel"], 863 indexing_maps = [affine_map<(i) -> (i)>, affine_map<(i) -> (i)>]} 864 ins(%0: memref<?xi32, offset: ?, strides: [1]>) 865 outs(%1: memref<?xi32, offset: ?, strides : [1]>) { 866 ^bb0(%a: i32, %b: i32): 867 linalg.yield %a : i32 868 } 869 return 870} 871// CHECK-LABEL: func @lower_to_loops_with_rank_reducing_subviews 872// CHECK: scf.for %[[IV:.+]] = %{{.+}} to %{{.+}} step %{{.+}} { 873// CHECK: %[[VAL:.+]] = memref.load %{{.+}}[%[[IV]]] 874// CHECK: memref.store %[[VAL]], %{{.+}}[%[[IV]]] 875// CHECK: } 876 877// CHECKPARALLEL-LABEL: func @lower_to_loops_with_rank_reducing_subviews 878// CHECKPARALLEL: scf.parallel (%[[IV:.+]]) = (%{{.+}}) to (%{{.+}}) step (%{{.+}}) { 879// CHECKPARALLEL: %[[VAL:.+]] = memref.load %{{.+}}[%[[IV]]] 880// CHECKPARALLEL: memref.store %[[VAL]], %{{.+}}[%[[IV]]] 881// CHECKPARALLEL: } 882