1//-------------------------------------------------------------------------------------------------- 2// WHEN CREATING A NEW TEST, PLEASE JUST COPY & PASTE WITHOUT EDITS. 3// 4// Set-up that's shared across all tests in this directory. In principle, this 5// config could be moved to lit.local.cfg. However, there are downstream users that 6// do not use these LIT config files. Hence why this is kept inline. 7// 8// DEFINE: %{sparsifier_opts} = enable-runtime-library=true 9// DEFINE: %{sparsifier_opts_sve} = enable-arm-sve=true %{sparsifier_opts} 10// DEFINE: %{compile} = mlir-opt %s --sparsifier="%{sparsifier_opts}" 11// DEFINE: %{compile_sve} = mlir-opt %s --sparsifier="%{sparsifier_opts_sve}" 12// DEFINE: %{run_libs} = -shared-libs=%mlir_c_runner_utils,%mlir_runner_utils 13// DEFINE: %{run_libs_sve} = -shared-libs=%native_mlir_runner_utils,%native_mlir_c_runner_utils 14// DEFINE: %{run_opts} = -e main -entry-point-result=void 15// DEFINE: %{run} = mlir-runner %{run_opts} %{run_libs} 16// DEFINE: %{run_sve} = %mcr_aarch64_cmd --march=aarch64 --mattr="+sve" %{run_opts} %{run_libs_sve} 17// 18// DEFINE: %{env} = 19//-------------------------------------------------------------------------------------------------- 20 21// RUN: %{compile} | %{run} | FileCheck %s 22// 23// Do the same run, but now with direct IR generation. 24// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false 25// RUN: %{compile} | %{run} | FileCheck %s 26// 27// Do the same run, but now with direct IR generation and vectorization. 28// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false vl=2 reassociate-fp-reductions=true enable-index-optimizations=true 29// RUN: %{compile} | %{run} | FileCheck %s 30// 31// Do the same run, but now with direct IR generation and VLA vectorization. 32// RUN: %if mlir_arm_sve_tests %{ %{compile_sve} | %{run_sve} | FileCheck %s %} 33 34 35#SparseVector = #sparse_tensor.encoding<{ 36 map = (d0) -> (d0 : compressed) 37}> 38 39#SparseMatrix = #sparse_tensor.encoding<{ 40 map = (d0, d1) -> (d0 : compressed, d1 : compressed) 41}> 42 43#trait_1d = { 44 indexing_maps = [ 45 affine_map<(i) -> (i)>, // a 46 affine_map<(i) -> (i)> // x (out) 47 ], 48 iterator_types = ["parallel"], 49 doc = "X(i) = a(i) op i" 50} 51 52#trait_2d = { 53 indexing_maps = [ 54 affine_map<(i,j) -> (i,j)>, // A 55 affine_map<(i,j) -> (i,j)> // X (out) 56 ], 57 iterator_types = ["parallel", "parallel"], 58 doc = "X(i,j) = A(i,j) op i op j" 59} 60 61// 62// Test with indices. Note that a lot of results are actually 63// dense, but this is done to stress test all the operations. 64// 65module { 66 67 // 68 // Kernel that uses index in the index notation (conjunction). 69 // 70 func.func @sparse_index_1d_conj(%arga: tensor<8xi64, #SparseVector>) 71 -> tensor<8xi64, #SparseVector> { 72 %init = tensor.empty() : tensor<8xi64, #SparseVector> 73 %r = linalg.generic #trait_1d 74 ins(%arga: tensor<8xi64, #SparseVector>) 75 outs(%init: tensor<8xi64, #SparseVector>) { 76 ^bb(%a: i64, %x: i64): 77 %i = linalg.index 0 : index 78 %ii = arith.index_cast %i : index to i64 79 %m1 = arith.muli %a, %ii : i64 80 linalg.yield %m1 : i64 81 } -> tensor<8xi64, #SparseVector> 82 return %r : tensor<8xi64, #SparseVector> 83 } 84 85 // 86 // Kernel that uses index in the index notation (disjunction). 87 // 88 func.func @sparse_index_1d_disj(%arga: tensor<8xi64, #SparseVector>) 89 -> tensor<8xi64, #SparseVector> { 90 %init = tensor.empty() : tensor<8xi64, #SparseVector> 91 %r = linalg.generic #trait_1d 92 ins(%arga: tensor<8xi64, #SparseVector>) 93 outs(%init: tensor<8xi64, #SparseVector>) { 94 ^bb(%a: i64, %x: i64): 95 %i = linalg.index 0 : index 96 %ii = arith.index_cast %i : index to i64 97 %m1 = arith.addi %a, %ii : i64 98 linalg.yield %m1 : i64 99 } -> tensor<8xi64, #SparseVector> 100 return %r : tensor<8xi64, #SparseVector> 101 } 102 103 // 104 // Kernel that uses indices in the index notation (conjunction). 105 // 106 func.func @sparse_index_2d_conj(%arga: tensor<3x4xi64, #SparseMatrix>) 107 -> tensor<3x4xi64, #SparseMatrix> { 108 %init = tensor.empty() : tensor<3x4xi64, #SparseMatrix> 109 %r = linalg.generic #trait_2d 110 ins(%arga: tensor<3x4xi64, #SparseMatrix>) 111 outs(%init: tensor<3x4xi64, #SparseMatrix>) { 112 ^bb(%a: i64, %x: i64): 113 %i = linalg.index 0 : index 114 %j = linalg.index 1 : index 115 %ii = arith.index_cast %i : index to i64 116 %jj = arith.index_cast %j : index to i64 117 %m1 = arith.muli %ii, %a : i64 118 %m2 = arith.muli %jj, %m1 : i64 119 linalg.yield %m2 : i64 120 } -> tensor<3x4xi64, #SparseMatrix> 121 return %r : tensor<3x4xi64, #SparseMatrix> 122 } 123 124 // 125 // Kernel that uses indices in the index notation (disjunction). 126 // 127 func.func @sparse_index_2d_disj(%arga: tensor<3x4xi64, #SparseMatrix>) 128 -> tensor<3x4xi64, #SparseMatrix> { 129 %init = tensor.empty() : tensor<3x4xi64, #SparseMatrix> 130 %r = linalg.generic #trait_2d 131 ins(%arga: tensor<3x4xi64, #SparseMatrix>) 132 outs(%init: tensor<3x4xi64, #SparseMatrix>) { 133 ^bb(%a: i64, %x: i64): 134 %i = linalg.index 0 : index 135 %j = linalg.index 1 : index 136 %ii = arith.index_cast %i : index to i64 137 %jj = arith.index_cast %j : index to i64 138 %m1 = arith.addi %ii, %a : i64 139 %m2 = arith.addi %jj, %m1 : i64 140 linalg.yield %m2 : i64 141 } -> tensor<3x4xi64, #SparseMatrix> 142 return %r : tensor<3x4xi64, #SparseMatrix> 143 } 144 145 func.func @add_outer_2d(%arg0: tensor<2x3xf32, #SparseMatrix>) 146 -> tensor<2x3xf32, #SparseMatrix> { 147 %0 = tensor.empty() : tensor<2x3xf32, #SparseMatrix> 148 %1 = linalg.generic #trait_2d 149 ins(%arg0 : tensor<2x3xf32, #SparseMatrix>) 150 outs(%0 : tensor<2x3xf32, #SparseMatrix>) { 151 ^bb0(%arg1: f32, %arg2: f32): 152 %2 = linalg.index 0 : index 153 %3 = arith.index_cast %2 : index to i64 154 %4 = arith.uitofp %3 : i64 to f32 155 %5 = arith.addf %arg1, %4 : f32 156 linalg.yield %5 : f32 157 } -> tensor<2x3xf32, #SparseMatrix> 158 return %1 : tensor<2x3xf32, #SparseMatrix> 159 } 160 161 // 162 // Main driver. 163 // 164 func.func @main() { 165 %c0 = arith.constant 0 : index 166 %du = arith.constant -1 : i64 167 %df = arith.constant -1.0 : f32 168 169 // Setup input sparse vector. 170 %v1 = arith.constant sparse<[[2], [4]], [ 10, 20]> : tensor<8xi64> 171 %sv = sparse_tensor.convert %v1 : tensor<8xi64> to tensor<8xi64, #SparseVector> 172 173 // Setup input "sparse" vector. 174 %v2 = arith.constant dense<[ 1, 2, 4, 8, 16, 32, 64, 128 ]> : tensor<8xi64> 175 %dv = sparse_tensor.convert %v2 : tensor<8xi64> to tensor<8xi64, #SparseVector> 176 177 // Setup input sparse matrix. 178 %m1 = arith.constant sparse<[[1,1], [2,3]], [10, 20]> : tensor<3x4xi64> 179 %sm = sparse_tensor.convert %m1 : tensor<3x4xi64> to tensor<3x4xi64, #SparseMatrix> 180 181 // Setup input "sparse" matrix. 182 %m2 = arith.constant dense <[ [ 1, 1, 1, 1 ], 183 [ 1, 2, 1, 1 ], 184 [ 1, 1, 3, 4 ] ]> : tensor<3x4xi64> 185 %dm = sparse_tensor.convert %m2 : tensor<3x4xi64> to tensor<3x4xi64, #SparseMatrix> 186 187 // Setup input sparse f32 matrix. 188 %mf32 = arith.constant sparse<[[0,1], [1,2]], [10.0, 41.0]> : tensor<2x3xf32> 189 %sf32 = sparse_tensor.convert %mf32 : tensor<2x3xf32> to tensor<2x3xf32, #SparseMatrix> 190 191 // Call the kernels. 192 %0 = call @sparse_index_1d_conj(%sv) : (tensor<8xi64, #SparseVector>) 193 -> tensor<8xi64, #SparseVector> 194 %1 = call @sparse_index_1d_disj(%sv) : (tensor<8xi64, #SparseVector>) 195 -> tensor<8xi64, #SparseVector> 196 %2 = call @sparse_index_1d_conj(%dv) : (tensor<8xi64, #SparseVector>) 197 -> tensor<8xi64, #SparseVector> 198 %3 = call @sparse_index_1d_disj(%dv) : (tensor<8xi64, #SparseVector>) 199 -> tensor<8xi64, #SparseVector> 200 %4 = call @sparse_index_2d_conj(%sm) : (tensor<3x4xi64, #SparseMatrix>) 201 -> tensor<3x4xi64, #SparseMatrix> 202 %5 = call @sparse_index_2d_disj(%sm) : (tensor<3x4xi64, #SparseMatrix>) 203 -> tensor<3x4xi64, #SparseMatrix> 204 %6 = call @sparse_index_2d_conj(%dm) : (tensor<3x4xi64, #SparseMatrix>) 205 -> tensor<3x4xi64, #SparseMatrix> 206 %7 = call @sparse_index_2d_disj(%dm) : (tensor<3x4xi64, #SparseMatrix>) 207 -> tensor<3x4xi64, #SparseMatrix> 208 209 // 210 // Verify result. 211 // 212 // CHECK: ---- Sparse Tensor ---- 213 // CHECK-NEXT: nse = 2 214 // CHECK-NEXT: dim = ( 8 ) 215 // CHECK-NEXT: lvl = ( 8 ) 216 // CHECK-NEXT: pos[0] : ( 0, 2 ) 217 // CHECK-NEXT: crd[0] : ( 2, 4 ) 218 // CHECK-NEXT: values : ( 20, 80 ) 219 // CHECK-NEXT: ---- 220 // 221 // CHECK: ---- Sparse Tensor ---- 222 // CHECK-NEXT: nse = 8 223 // CHECK-NEXT: dim = ( 8 ) 224 // CHECK-NEXT: lvl = ( 8 ) 225 // CHECK-NEXT: pos[0] : ( 0, 8 ) 226 // CHECK-NEXT: crd[0] : ( 0, 1, 2, 3, 4, 5, 6, 7 ) 227 // CHECK-NEXT: values : ( 0, 1, 12, 3, 24, 5, 6, 7 ) 228 // CHECK-NEXT: ---- 229 // 230 // CHECK: ---- Sparse Tensor ---- 231 // CHECK-NEXT: nse = 8 232 // CHECK-NEXT: dim = ( 8 ) 233 // CHECK-NEXT: lvl = ( 8 ) 234 // CHECK-NEXT: pos[0] : ( 0, 8 ) 235 // CHECK-NEXT: crd[0] : ( 0, 1, 2, 3, 4, 5, 6, 7 ) 236 // CHECK-NEXT: values : ( 0, 2, 8, 24, 64, 160, 384, 896 ) 237 // CHECK-NEXT: ---- 238 // 239 // CHECK: ---- Sparse Tensor ---- 240 // CHECK-NEXT: nse = 8 241 // CHECK-NEXT: dim = ( 8 ) 242 // CHECK-NEXT: lvl = ( 8 ) 243 // CHECK-NEXT: pos[0] : ( 0, 8 ) 244 // CHECK-NEXT: crd[0] : ( 0, 1, 2, 3, 4, 5, 6, 7 ) 245 // CHECK-NEXT: values : ( 1, 3, 6, 11, 20, 37, 70, 135 ) 246 // CHECK-NEXT: ---- 247 // 248 // CHECK: ---- Sparse Tensor ---- 249 // CHECK-NEXT: nse = 2 250 // CHECK-NEXT: dim = ( 3, 4 ) 251 // CHECK-NEXT: lvl = ( 3, 4 ) 252 // CHECK-NEXT: pos[0] : ( 0, 2 ) 253 // CHECK-NEXT: crd[0] : ( 1, 2 ) 254 // CHECK-NEXT: pos[1] : ( 0, 1, 2 ) 255 // CHECK-NEXT: crd[1] : ( 1, 3 ) 256 // CHECK-NEXT: values : ( 10, 120 ) 257 // CHECK-NEXT: ---- 258 // 259 // CHECK: ---- Sparse Tensor ---- 260 // CHECK-NEXT: nse = 12 261 // CHECK-NEXT: dim = ( 3, 4 ) 262 // CHECK-NEXT: lvl = ( 3, 4 ) 263 // CHECK-NEXT: pos[0] : ( 0, 3 ) 264 // CHECK-NEXT: crd[0] : ( 0, 1, 2 ) 265 // CHECK-NEXT: pos[1] : ( 0, 4, 8, 12 ) 266 // CHECK-NEXT: crd[1] : ( 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3 ) 267 // CHECK-NEXT: values : ( 0, 1, 2, 3, 1, 12, 3, 4, 2, 3, 4, 25 ) 268 // CHECK-NEXT: ---- 269 // 270 // CHECK: ---- Sparse Tensor ---- 271 // CHECK-NEXT: nse = 12 272 // CHECK-NEXT: dim = ( 3, 4 ) 273 // CHECK-NEXT: lvl = ( 3, 4 ) 274 // CHECK-NEXT: pos[0] : ( 0, 3 ) 275 // CHECK-NEXT: crd[0] : ( 0, 1, 2 ) 276 // CHECK-NEXT: pos[1] : ( 0, 4, 8, 12 ) 277 // CHECK-NEXT: crd[1] : ( 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3 ) 278 // CHECK-NEXT: values : ( 0, 0, 0, 0, 0, 2, 2, 3, 0, 2, 12, 24 ) 279 // CHECK-NEXT: ---- 280 // 281 // CHECK: ---- Sparse Tensor ---- 282 // CHECK-NEXT: nse = 12 283 // CHECK-NEXT: dim = ( 3, 4 ) 284 // CHECK-NEXT: lvl = ( 3, 4 ) 285 // CHECK-NEXT: pos[0] : ( 0, 3 ) 286 // CHECK-NEXT: crd[0] : ( 0, 1, 2 ) 287 // CHECK-NEXT: pos[1] : ( 0, 4, 8, 12 ) 288 // CHECK-NEXT: crd[1] : ( 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3 ) 289 // CHECK-NEXT: values : ( 1, 2, 3, 4, 2, 4, 4, 5, 3, 4, 7, 9 ) 290 // CHECK-NEXT: ---- 291 // 292 sparse_tensor.print %0 : tensor<8xi64, #SparseVector> 293 sparse_tensor.print %1 : tensor<8xi64, #SparseVector> 294 sparse_tensor.print %2 : tensor<8xi64, #SparseVector> 295 sparse_tensor.print %3 : tensor<8xi64, #SparseVector> 296 sparse_tensor.print %4 : tensor<3x4xi64, #SparseMatrix> 297 sparse_tensor.print %5 : tensor<3x4xi64, #SparseMatrix> 298 sparse_tensor.print %6 : tensor<3x4xi64, #SparseMatrix> 299 sparse_tensor.print %7 : tensor<3x4xi64, #SparseMatrix> 300 301 // 302 // Call the f32 kernel, verify the result. 303 // 304 // CHECK: ---- Sparse Tensor ---- 305 // CHECK-NEXT: nse = 6 306 // CHECK-NEXT: dim = ( 2, 3 ) 307 // CHECK-NEXT: lvl = ( 2, 3 ) 308 // CHECK-NEXT: pos[0] : ( 0, 2 ) 309 // CHECK-NEXT: crd[0] : ( 0, 1 ) 310 // CHECK-NEXT: pos[1] : ( 0, 3, 6 ) 311 // CHECK-NEXT: crd[1] : ( 0, 1, 2, 0, 1, 2 ) 312 // CHECK-NEXT: values : ( 0, 10, 0, 1, 1, 42 ) 313 // CHECK-NEXT: ---- 314 // 315 %100 = call @add_outer_2d(%sf32) : (tensor<2x3xf32, #SparseMatrix>) 316 -> tensor<2x3xf32, #SparseMatrix> 317 sparse_tensor.print %100 : tensor<2x3xf32, #SparseMatrix> 318 319 // Release resources. 320 bufferization.dealloc_tensor %sv : tensor<8xi64, #SparseVector> 321 bufferization.dealloc_tensor %dv : tensor<8xi64, #SparseVector> 322 bufferization.dealloc_tensor %0 : tensor<8xi64, #SparseVector> 323 bufferization.dealloc_tensor %1 : tensor<8xi64, #SparseVector> 324 bufferization.dealloc_tensor %2 : tensor<8xi64, #SparseVector> 325 bufferization.dealloc_tensor %3 : tensor<8xi64, #SparseVector> 326 bufferization.dealloc_tensor %sm : tensor<3x4xi64, #SparseMatrix> 327 bufferization.dealloc_tensor %dm : tensor<3x4xi64, #SparseMatrix> 328 bufferization.dealloc_tensor %4 : tensor<3x4xi64, #SparseMatrix> 329 bufferization.dealloc_tensor %5 : tensor<3x4xi64, #SparseMatrix> 330 bufferization.dealloc_tensor %6 : tensor<3x4xi64, #SparseMatrix> 331 bufferization.dealloc_tensor %7 : tensor<3x4xi64, #SparseMatrix> 332 bufferization.dealloc_tensor %sf32 : tensor<2x3xf32, #SparseMatrix> 333 bufferization.dealloc_tensor %100 : tensor<2x3xf32, #SparseMatrix> 334 335 return 336 } 337} 338