1// RUN: mlir-opt -linalg-bufferize -canonicalize -cse -split-input-file %s | FileCheck %s 2 3#map0 = affine_map<(d0) -> (d0)> 4 5// In-depth checking of a basic case, this is testing 6// - memref.buffer_cast / memref.tensor_load materializations are properly inserted 7// - payload is correctly carried over 8// - affine maps are correctly carried over 9// Later tests will not check all these details. 10 11// CHECK: #map = affine_map<(d0) -> (d0)> 12// CHECK-LABEL: func @basic( 13// CHECK-SAME: %[[TENSOR:.*]]: tensor<4xf32>) -> tensor<4xf32> { 14// CHECK: %[[MEMREF:.*]] = memref.buffer_cast %[[TENSOR]] : memref<4xf32> 15// CHECK: %[[RESULT_MEMREF:.*]] = memref.alloc() : memref<4xf32> 16// CHECK: linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} 17// CHECK-SAME: ins(%[[MEMREF]] : memref<4xf32>) 18// CHECK-SAME: outs(%[[RESULT_MEMREF]] : memref<4xf32>) { 19// CHECK: ^bb0(%[[RESULT1:.*]]: f32, %[[UNUSED:.*]]: f32): 20// CHECK: %[[DIM1:.*]] = math.exp %[[RESULT1]] : f32 21// CHECK: linalg.yield %[[DIM1]] : f32 22// CHECK: } 23// CHECK: %[[RESULT:.*]] = memref.tensor_load %[[RESULT_MEMREF]] : memref<4xf32> 24// CHECK: return %[[RESULT]] : tensor<4xf32> 25func @basic(%arg0: tensor<4xf32>) -> tensor<4xf32> { 26 %0 = linalg.generic { 27 indexing_maps = [#map0, #map0], 28 iterator_types = ["parallel"] 29 } ins(%arg0 : tensor<4xf32>) 30 outs(%arg0 : tensor<4xf32>) { 31 ^bb0(%gen_arg1: f32, %out: f32): 32 %tmp1 = math.exp %gen_arg1 : f32 33 linalg.yield %tmp1 : f32 34 } -> tensor<4xf32> 35 return %0 : tensor<4xf32> 36} 37 38 39// ----- 40 41#map0 = affine_map<(d0) -> (d0)> 42 43// Same as above but with linalg.init_tensor op. 44 45// CHECK: #map = affine_map<(d0) -> (d0)> 46// CHECK-LABEL: func @init_tensor( 47// CHECK-SAME: %[[IN:.*]]: tensor<?xf32>, %[[SIZE:.*]]: index) 48// CHECK: %[[MEMREF:.*]] = memref.buffer_cast %[[IN]] : memref<?xf32> 49// CHECK: %[[OUT_BUF:.*]] = memref.alloc(%[[SIZE]]) : memref<?xf32> 50// CHECK: linalg.generic 51// CHECK-SAME: ins(%[[MEMREF]] : memref<?xf32>) 52// CHECK-SAME: outs(%[[OUT_BUF]] : memref<?xf32>) { 53func @init_tensor(%in : tensor<?xf32>, %size: index) -> tensor<?xf32> { 54 %init = linalg.init_tensor [%size] : tensor<?xf32> 55 %0 = linalg.generic { 56 indexing_maps = [#map0, #map0], 57 iterator_types = ["parallel"] 58 } ins(%in : tensor<?xf32>) 59 outs(%init : tensor<?xf32>) { 60 ^bb0(%gen_arg1: f32, %out: f32): 61 %tmp1 = math.exp %gen_arg1 : f32 62 linalg.yield %tmp1 : f32 63 } -> tensor<?xf32> 64 return %0 : tensor<?xf32> 65} 66 67 68// ----- 69 70#map0 = affine_map<(d0) -> (d0)> 71 72// CHECK-LABEL: func @multiple_results 73// CHECK: %[[RESULT0:.*]] = memref.alloc() : memref<4xf32> 74// CHECK: %[[RESULT1:.*]] = memref.alloc() : memref<4xf32> 75// CHECK: linalg.generic 76// CHECK-SAME: ins(%{{.*}} : memref<4xf32>) 77// CHECK-SAME: outs(%[[RESULT0]], %[[RESULT1]] : memref<4xf32>, memref<4xf32>) 78// CHECK-NEXT: ^bb0(%{{.*}}: f32, %{{.*}}: f32, %{{.*}}: f32): 79func @multiple_results(%arg0: tensor<4xf32>) -> (tensor<4xf32>, tensor<4xf32>) { 80 %0, %1 = linalg.generic { 81 indexing_maps = [#map0, #map0, #map0], 82 iterator_types = ["parallel"] 83 } ins(%arg0 : tensor<4xf32>) 84 outs (%arg0, %arg0 : tensor<4xf32>, tensor<4xf32>) { 85 ^bb0(%gen_arg1: f32, %out1: f32, %out2: f32): 86 %tmp1 = math.exp %gen_arg1 : f32 87 linalg.yield %tmp1, %tmp1 : f32, f32 88 } -> (tensor<4xf32>, tensor<4xf32>) 89 return %0, %1 : tensor<4xf32>, tensor<4xf32> 90} 91 92// ----- 93 94#map_2d = affine_map<(d0, d1) -> (d0, d1)> 95 96// Check that the allocs properly consider the different shapes of the output 97// operands. The permuted indexing maps translate to different output shapes. 98 99// CHECK-LABEL: func @dynamic_results( 100// CHECK-SAME: %[[ARG:.*]]: tensor<?x?xf32> 101// CHECK-DAG: %[[C0:.*]] = constant 0 : index 102// CHECK-DAG: %[[C1:.*]] = constant 1 : index 103// CHECK: %[[MEMREF_ARG:.*]] = memref.buffer_cast %[[ARG]] : memref<?x?xf32> 104// CHECK: %[[DIM0:.*]] = tensor.dim %[[ARG]], %[[C0]] : tensor<?x?xf32> 105// CHECK: %[[DIM1:.*]] = tensor.dim %[[ARG]], %[[C1]] : tensor<?x?xf32> 106// CHECK: %[[RESULT0:.*]] = memref.alloc(%[[DIM0]], %[[DIM1]]) : memref<?x?xf32> 107// CHECK: %[[RESULT1:.*]] = memref.alloc(%[[DIM0]], %[[DIM1]]) : memref<?x?xf32> 108// CHECK: linalg.generic 109// CHECK-SAME: ins(%[[MEMREF_ARG]] : memref<?x?xf32>) 110// CHECK-SAME: outs(%[[RESULT0]], %[[RESULT1]] : memref<?x?xf32>, memref<?x?xf32>) 111func @dynamic_results(%arg0: tensor<?x?xf32>) 112 -> (tensor<?x?xf32>, tensor<?x?xf32>) { 113 %0, %1 = linalg.generic { 114 indexing_maps = [#map_2d, #map_2d, #map_2d], 115 iterator_types = ["parallel", "parallel"] 116 } ins(%arg0 : tensor<?x?xf32>) 117 outs (%arg0, %arg0 : tensor<?x?xf32>, tensor<?x?xf32>) { 118 ^bb0(%gen_arg1: f32, %out1: f32, %out2: f32): 119 %tmp1 = math.exp %gen_arg1 : f32 120 linalg.yield %tmp1, %tmp1 : f32, f32 121 } -> (tensor<?x?xf32>, tensor<?x?xf32>) 122 return %0, %1 : tensor<?x?xf32>, tensor<?x?xf32> 123} 124 125// ----- 126 127#accesses = [ 128 affine_map<(i, j, k) -> (j, i, k)>, 129 affine_map<(i, j, k) -> (i, j)> 130] 131 132#trait = { 133 indexing_maps = #accesses, 134 iterator_types = ["parallel", "parallel", "reduction"] 135} 136 137// Check the bufferization of init tensors. 138 139// CHECK-LABEL: func @generic_with_init_tensor( 140// CHECK-SAME: %[[ARG0_TENSOR:.*]]: tensor<2x3x4xvector<3x4xi4>>, 141// CHECK-SAME: %[[ARG1_TENSOR:.*]]: tensor<3x2xf32>) -> tensor<3x2xf32> { 142// CHECK: %[[ARG0_MEMREF:.*]] = memref.buffer_cast %[[ARG0_TENSOR]] : memref<2x3x4xvector<3x4xi4>> 143// CHECK: %[[ARG1_MEMREF:.*]] = memref.buffer_cast %[[ARG1_TENSOR]] : memref<3x2xf32> 144// CHECK: %[[INIT_BUFFER:.*]] = memref.alloc() : memref<3x2xf32> 145// CHECK: linalg.copy(%[[ARG1_MEMREF]], %[[INIT_BUFFER]]) : memref<3x2xf32>, memref<3x2xf32> 146// CHECK: linalg.generic 147// CHECK-SAME: ins(%[[ARG0_MEMREF]] : memref<2x3x4xvector<3x4xi4>>) 148// CHECK-SAME: outs(%[[INIT_BUFFER]] : memref<3x2xf32>) { 149func @generic_with_init_tensor(%arg0: tensor<2x3x4xvector<3x4xi4>>, 150 %arg1: tensor<3x2xf32>) -> (tensor<3x2xf32>) { 151 152 %0 = linalg.generic #trait 153 ins(%arg0 : tensor<2x3x4xvector<3x4xi4>>) 154 outs(%arg1 : tensor<3x2xf32>) { 155 ^bb(%v0: vector<3x4xi4>, %v1: f32) : 156 linalg.yield %v1 : f32 157 } -> tensor<3x2xf32> 158 159 return %0 : tensor<3x2xf32> 160} 161 162// ----- 163 164// CHECK-DAG: #[[$MAP0:[0-9a-z]*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)> 165// CHECK-DAG: #[[$MAP1:[0-9a-z]*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1 * 2)> 166 167func private @make_index() -> index 168 169// CHECK-LABEL: func @bufferize_slice( 170// CHECK-SAME: %[[T:[0-9a-z]*]]: tensor<?x?xf32> 171func @bufferize_slice(%t : tensor<?x?xf32>) -> (tensor<2x3xf32>, tensor<2x?xf32>) { 172 // CHECK: %[[IDX:.*]] = call @make_index() : () -> index 173 %i0 = call @make_index() : () -> index 174 175 // CHECK: %[[M:.*]] = memref.buffer_cast %[[T]] : memref<?x?xf32> 176 // CHECK-NEXT: %[[A0:.*]] = memref.alloc() : memref<2x3xf32> 177 // CHECK-NEXT: %[[SM0:.*]] = memref.subview %[[M]][0, 0] [2, 3] [1, 1] 178 // CHECK-SAME: memref<?x?xf32> to memref<2x3xf32, #[[$MAP0]]> 179 // CHECK-NEXT: linalg.copy(%[[SM0]], %[[A0]]) : memref<2x3xf32, #[[$MAP0]]>, memref<2x3xf32> 180 // CHECK-NEXT: %[[RT0:.*]] = memref.tensor_load %[[A0]] : memref<2x3xf32> 181 %st0 = tensor.extract_slice %t[0, 0][2, 3][1, 1] : tensor<?x?xf32> to tensor<2x3xf32> 182 183 // CHECK-NEXT: %[[A1:.*]] = memref.alloc(%[[IDX]]) : memref<2x?xf32> 184 // CHECK-NEXT: %[[SM1:.*]] = memref.subview %[[M]][0, %[[IDX]]] [2, %[[IDX]]] [1, 2] 185 // CHECK-SAME: memref<?x?xf32> to memref<2x?xf32, #[[$MAP1]]> 186 // CHECK-NEXT: linalg.copy(%[[SM1]], %[[A1]]) : memref<2x?xf32, #[[$MAP1]]>, memref<2x?xf32> 187 // CHECK-NEXT: %[[RT1:.*]] = memref.tensor_load %[[A1]] : memref<2x?xf32> 188 %st1 = tensor.extract_slice %t[0, %i0][2, %i0][1, 2] : tensor<?x?xf32> to tensor<2x?xf32> 189 190 // CHECK-NEXT: return %[[RT0]], %[[RT1]] 191 return %st0, %st1 : tensor<2x3xf32>, tensor<2x?xf32> 192} 193 194// ----- 195 196// CHECK-DAG: #[[$MAP0:[0-9a-z]*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)> 197// CHECK-DAG: #[[$MAP1:[0-9a-z]*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1 * 2)> 198 199func private @make_index() -> index 200 201// CHECK-LABEL: func @bufferize_insert_slice( 202// CHECK-SAME: %[[T:[0-9a-z]*]]: tensor<?x?xf32> 203// CHECK-SAME: %[[ST0:[0-9a-z]*]]: tensor<2x3xf32> 204// CHECK-SAME: %[[ST1:[0-9a-z]*]]: tensor<2x?xf32> 205func @bufferize_insert_slice(%t : tensor<?x?xf32>, %st0 : tensor<2x3xf32>, %st1 : tensor<2x?xf32>) -> 206 (tensor<?x?xf32>, tensor<?x?xf32>) { 207 %c0 = constant 0 : index 208 %c1 = constant 1 : index 209 // CHECK-DAG: %[[C0:.*]] = constant 0 : index 210 // CHECK-DAG: %[[C1:.*]] = constant 1 : index 211 %i0 = call @make_index() : () -> index 212 // CHECK: %[[IDX:.*]] = call @make_index() : () -> index 213 214 215 // CHECK-DAG: %[[M:.*]] = memref.buffer_cast %[[T]] : memref<?x?xf32> 216 // CHECK-DAG: %[[SM0:.*]] = memref.buffer_cast %[[ST0]] : memref<2x3xf32> 217 // CHECK-NEXT: %[[DIM0:.*]] = tensor.dim %[[T]], %[[C0]] : tensor<?x?xf32> 218 // CHECK-NEXT: %[[DIM1:.*]] = tensor.dim %[[T]], %[[C1]] : tensor<?x?xf32> 219 // CHECK-NEXT: %[[M_COPY0:.*]] = memref.alloc(%[[DIM0]], %[[DIM1]]) : memref<?x?xf32> 220 // CHECK-NEXT: linalg.copy(%[[M]], %[[M_COPY0]]) : memref<?x?xf32>, memref<?x?xf32> 221 // CHECK-NEXT: %[[SUBVIEW0:.*]] = memref.subview %[[M_COPY0]][0, 0] [2, 3] [1, 1] 222 // CHECK-SAME: memref<?x?xf32> to memref<2x3xf32, #[[$MAP0]]> 223 // CHECK-NEXT: linalg.copy(%[[SM0]], %[[SUBVIEW0]]) : memref<2x3xf32>, memref<2x3xf32, #[[$MAP0]]> 224 // CHECK-NEXT: %[[RT0:.*]] = memref.tensor_load %[[M_COPY0]] : memref<?x?xf32> 225 %t0 = tensor.insert_slice %st0 into %t[0, 0][2, 3][1, 1] : tensor<2x3xf32> into tensor<?x?xf32> 226 227 // CHECK-DAG: %[[SM1:.*]] = memref.buffer_cast %[[ST1]] : memref<2x?xf32> 228 // CHECK-NEXT: %[[M_COPY1:.*]] = memref.alloc(%[[DIM0]], %[[DIM1]]) : memref<?x?xf32> 229 // CHECK-NEXT: linalg.copy(%[[M]], %[[M_COPY1]]) : memref<?x?xf32>, memref<?x?xf32> 230 // CHECK-NEXT: %[[SUBVIEW1:.*]] = memref.subview %[[M_COPY1]][0, %[[IDX]]] [2, %[[IDX]]] [1, 2] 231 // CHECK-SAME: memref<?x?xf32> to memref<2x?xf32, #[[$MAP1]]> 232 // CHECK-NEXT: linalg.copy(%[[SM1]], %[[SUBVIEW1]]) : memref<2x?xf32>, memref<2x?xf32, #[[$MAP1]]> 233 // CHECK-NEXT: %[[RT1:.*]] = memref.tensor_load %[[M_COPY1]] : memref<?x?xf32> 234 %t1 = tensor.insert_slice %st1 into %t[0, %i0][2, %i0][1, 2] : tensor<2x?xf32> into tensor<?x?xf32> 235 236 // CHECK: return %[[RT0]], %[[RT1]] 237 return %t0, %t1: tensor<?x?xf32>, tensor<?x?xf32> 238} 239 240// ----- 241 242// CHECK-LABEL: func @bufferize_fill( 243// CHECK-SAME: %[[IN:.*]]: tensor<?xf32> 244func @bufferize_fill(%arg0: tensor<?xf32>) -> tensor<?xf32> { 245 %c0 = constant 0.0 : f32 246 // CHECK: %[[MEMREF:.*]] = memref.buffer_cast %[[IN]] : memref<?xf32> 247 // CHECK: linalg.fill(%cst, %[[MEMREF]]) : f32, memref<?xf32> 248 // CHECK: %[[TENSOR:.*]] = memref.tensor_load %[[MEMREF]] : memref<?xf32> 249 // CHECK: return %[[TENSOR]] 250 %0 = linalg.fill(%c0, %arg0) : f32, tensor<?xf32> -> tensor<?xf32> 251 return %0 : tensor<?xf32> 252} 253 254// ----- 255 256// CHECK-LABEL: func @bufferize_tensor_collapse_shape( 257// CHECK-SAME: %[[IN:.*]]: tensor<4x5xf32> 258func @bufferize_tensor_collapse_shape(%arg0: tensor<4x5xf32>) -> tensor<20xf32> { 259 %out = linalg.tensor_collapse_shape %arg0 [[0, 1]] : 260 tensor<4x5xf32> into tensor<20xf32> 261 return %out : tensor<20xf32> 262} 263// CHECK: %[[MEMREF:.*]] = memref.buffer_cast %[[IN]] : memref<4x5xf32> 264// CHECK: %[[RESHAPE:.*]] = memref.collapse_shape %[[MEMREF]] {{\[}}[0, 1]] 265// CHECK-SAME: : memref<4x5xf32> into memref<20xf32> 266// CHECK: %[[TENSOR:.*]] = memref.tensor_load %[[RESHAPE]] : memref<20xf32> 267// CHECK: return %[[TENSOR]] 268 269// ----- 270 271// CHECK-LABEL: func @pad_tensor_dynamic_shape( 272// CHECK-SAME: %[[IN:.*]]: tensor<4x?x2x?xf32>, 273// CHECK-SAME: %[[OFFSET:.*]]: index) -> tensor<4x?x?x?xf32> { 274func @pad_tensor_dynamic_shape(%arg0: tensor<4x?x2x?xf32>, %arg1: index) -> tensor<4x?x?x?xf32> { 275 %c0 = constant 0 : index 276 %cst = constant 0.0 : f32 277 %out = linalg.pad_tensor %arg0 low[%c0, %c0, %arg1, %c0] high[%c0, %c0, %c0, %arg1] { 278 ^bb0(%gen_arg1: index, %gen_arg2: index, %gen_arg3: index, %gen_arg4: index): // no predecessors 279 linalg.yield %cst : f32 280 } : tensor<4x?x2x?xf32> to tensor<4x?x?x?xf32> 281 return %out : tensor<4x?x?x?xf32> 282} 283 284// CHECK-DAG: %[[C3:.*]] = constant 3 : index 285// CHECK-DAG: %[[C2:.*]] = constant 2 : index 286// CHECK-DAG: %[[C1:.*]] = constant 1 : index 287// CHECK-DAG: %[[CST:.*]] = constant 0.000000e+00 : f32 288// CHECK: %[[DIM1:.*]] = tensor.dim %[[IN]], %[[C1]] : tensor<4x?x2x?xf32> 289// CHECK: %[[OUT_DIM2:.*]] = addi %[[OFFSET]], %[[C2]] : index 290// CHECK: %[[DIM3:.*]] = tensor.dim %[[IN]], %[[C3]] : tensor<4x?x2x?xf32> 291// CHECK: %[[OUT_DIM3:.*]] = addi %[[DIM3]], %[[OFFSET]] : index 292// CHECK: %[[FILLED:.*]] = memref.alloc(%[[DIM1]], %[[OUT_DIM2]], %[[OUT_DIM3]]) : memref<4x?x?x?xf32> 293// CHECK: linalg.fill(%[[CST]], %[[FILLED]]) : f32, memref<4x?x?x?xf32> 294// CHECK: %[[IN_MEMREF:.*]] = memref.buffer_cast %[[IN]] : memref<4x?x2x?xf32> 295// CHECK: %[[OUT:.*]] = memref.alloc(%[[DIM1]], %[[OUT_DIM2]], %[[OUT_DIM3]]) : memref<4x?x?x?xf32> 296// CHECK: linalg.copy(%[[FILLED]], %[[OUT]]) : memref<4x?x?x?xf32>, memref<4x?x?x?xf32> 297// CHECK: %[[INTERIOR:.*]] = memref.subview %[[OUT]][0, 0, %[[OFFSET]], 0] [4, %[[DIM1]], 2, %[[DIM3]]] [1, 1, 1, 1] : memref<4x?x?x?xf32> to memref<4x?x2x?xf32, #map> 298// CHECK: linalg.copy(%[[IN_MEMREF]], %[[INTERIOR]]) : memref<4x?x2x?xf32>, memref<4x?x2x?xf32, #map> 299// CHECK: %[[OUT_TENSOR:.*]] = memref.tensor_load %[[OUT]] : memref<4x?x?x?xf32> 300// CHECK: return %[[OUT_TENSOR]] : tensor<4x?x?x?xf32> 301// CHECK: } 302 303 304// ----- 305 306// CHECK-LABEL: func @vector_transfer 307func @vector_transfer(%in: tensor<4xf32>, %out: tensor<4xf32>) { 308 %c0 = constant 0 : index 309 %cst = constant 0.000000e+00 : f32 310 %read = vector.transfer_read %in[%c0], %cst {in_bounds = [true]} 311 : tensor<4xf32>, vector<4xf32> 312 %tanh = math.tanh %read : vector<4xf32> 313 %write = vector.transfer_write %tanh, %out[%c0] {in_bounds = [true]} 314 : vector<4xf32>, tensor<4xf32> 315 return 316 // CHECK: vector.transfer_read {{.*}} : memref<4xf32>, vector<4xf32> 317 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, memref<4xf32> 318} 319