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 enable-buffer-initialization=true 25// RUN: %{compile} | %{run} | FileCheck %s 26// 27// Do the same run, but now with vectorization. 28// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false enable-buffer-initialization=true 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 VLA vectorization. 32// RUN: %if mlir_arm_sve_tests %{ %{compile_sve} | %{run_sve} | FileCheck %s %} 33 34#SparseVector = #sparse_tensor.encoding<{map = (d0) -> (d0 : compressed)}> 35#DenseVector = #sparse_tensor.encoding<{map = (d0) -> (d0 : dense)}> 36 37// 38// Traits for 1-d tensor (aka vector) operations. 39// 40#trait_scale = { 41 indexing_maps = [ 42 affine_map<(i) -> (i)>, // a (in) 43 affine_map<(i) -> (i)> // x (out) 44 ], 45 iterator_types = ["parallel"], 46 doc = "x(i) = a(i) * 2.0" 47} 48#trait_scale_inpl = { 49 indexing_maps = [ 50 affine_map<(i) -> (i)> // x (out) 51 ], 52 iterator_types = ["parallel"], 53 doc = "x(i) *= 2.0" 54} 55#trait_op = { 56 indexing_maps = [ 57 affine_map<(i) -> (i)>, // a (in) 58 affine_map<(i) -> (i)>, // b (in) 59 affine_map<(i) -> (i)> // x (out) 60 ], 61 iterator_types = ["parallel"], 62 doc = "x(i) = a(i) OP b(i)" 63} 64#trait_dot = { 65 indexing_maps = [ 66 affine_map<(i) -> (i)>, // a (in) 67 affine_map<(i) -> (i)>, // b (in) 68 affine_map<(i) -> ()> // x (out) 69 ], 70 iterator_types = ["parallel"], 71 doc = "x(i) += a(i) * b(i)" 72} 73 74module { 75 // Scales a sparse vector into a new sparse vector. 76 func.func @vector_scale(%arga: tensor<?xf64, #SparseVector>) -> tensor<?xf64, #SparseVector> { 77 %s = arith.constant 2.0 : f64 78 %c = arith.constant 0 : index 79 %d = tensor.dim %arga, %c : tensor<?xf64, #SparseVector> 80 %xv = tensor.empty(%d) : tensor<?xf64, #SparseVector> 81 %0 = linalg.generic #trait_scale 82 ins(%arga: tensor<?xf64, #SparseVector>) 83 outs(%xv: tensor<?xf64, #SparseVector>) { 84 ^bb(%a: f64, %x: f64): 85 %1 = arith.mulf %a, %s : f64 86 linalg.yield %1 : f64 87 } -> tensor<?xf64, #SparseVector> 88 return %0 : tensor<?xf64, #SparseVector> 89 } 90 91 // Scales a sparse vector in place. 92 func.func @vector_scale_inplace(%argx: tensor<?xf64, #SparseVector>) -> tensor<?xf64, #SparseVector> { 93 %s = arith.constant 2.0 : f64 94 %0 = linalg.generic #trait_scale_inpl 95 outs(%argx: tensor<?xf64, #SparseVector>) { 96 ^bb(%x: f64): 97 %1 = arith.mulf %x, %s : f64 98 linalg.yield %1 : f64 99 } -> tensor<?xf64, #SparseVector> 100 return %0 : tensor<?xf64, #SparseVector> 101 } 102 103 // Adds two sparse vectors into a new sparse vector. 104 func.func @vector_add(%arga: tensor<?xf64, #SparseVector>, 105 %argb: tensor<?xf64, #SparseVector>) -> tensor<?xf64, #SparseVector> { 106 %c = arith.constant 0 : index 107 %d = tensor.dim %arga, %c : tensor<?xf64, #SparseVector> 108 %xv = tensor.empty(%d) : tensor<?xf64, #SparseVector> 109 %0 = linalg.generic #trait_op 110 ins(%arga, %argb: tensor<?xf64, #SparseVector>, tensor<?xf64, #SparseVector>) 111 outs(%xv: tensor<?xf64, #SparseVector>) { 112 ^bb(%a: f64, %b: f64, %x: f64): 113 %1 = arith.addf %a, %b : f64 114 linalg.yield %1 : f64 115 } -> tensor<?xf64, #SparseVector> 116 return %0 : tensor<?xf64, #SparseVector> 117 } 118 119 // Multiplies two sparse vectors into a new sparse vector. 120 func.func @vector_mul(%arga: tensor<?xf64, #SparseVector>, 121 %argb: tensor<?xf64, #SparseVector>) -> tensor<?xf64, #SparseVector> { 122 %c = arith.constant 0 : index 123 %d = tensor.dim %arga, %c : tensor<?xf64, #SparseVector> 124 %xv = tensor.empty(%d) : tensor<?xf64, #SparseVector> 125 %0 = linalg.generic #trait_op 126 ins(%arga, %argb: tensor<?xf64, #SparseVector>, tensor<?xf64, #SparseVector>) 127 outs(%xv: tensor<?xf64, #SparseVector>) { 128 ^bb(%a: f64, %b: f64, %x: f64): 129 %1 = arith.mulf %a, %b : f64 130 linalg.yield %1 : f64 131 } -> tensor<?xf64, #SparseVector> 132 return %0 : tensor<?xf64, #SparseVector> 133 } 134 135 // Multiplies two sparse vectors into a new "annotated" dense vector. 136 func.func @vector_mul_d(%arga: tensor<?xf64, #SparseVector>, 137 %argb: tensor<?xf64, #SparseVector>) -> tensor<?xf64, #DenseVector> { 138 %c = arith.constant 0 : index 139 %d = tensor.dim %arga, %c : tensor<?xf64, #SparseVector> 140 %xv = tensor.empty(%d) : tensor<?xf64, #DenseVector> 141 %0 = linalg.generic #trait_op 142 ins(%arga, %argb: tensor<?xf64, #SparseVector>, tensor<?xf64, #SparseVector>) 143 outs(%xv: tensor<?xf64, #DenseVector>) { 144 ^bb(%a: f64, %b: f64, %x: f64): 145 %1 = arith.mulf %a, %b : f64 146 linalg.yield %1 : f64 147 } -> tensor<?xf64, #DenseVector> 148 return %0 : tensor<?xf64, #DenseVector> 149 } 150 151 // Sum reduces dot product of two sparse vectors. 152 func.func @vector_dotprod(%arga: tensor<?xf64, #SparseVector>, 153 %argb: tensor<?xf64, #SparseVector>, 154 %argx: tensor<f64>) -> tensor<f64> { 155 %0 = linalg.generic #trait_dot 156 ins(%arga, %argb: tensor<?xf64, #SparseVector>, tensor<?xf64, #SparseVector>) 157 outs(%argx: tensor<f64>) { 158 ^bb(%a: f64, %b: f64, %x: f64): 159 %1 = arith.mulf %a, %b : f64 160 %2 = arith.addf %x, %1 : f64 161 linalg.yield %2 : f64 162 } -> tensor<f64> 163 return %0 : tensor<f64> 164 } 165 166 // Driver method to call and verify vector kernels. 167 func.func @main() { 168 %c0 = arith.constant 0 : index 169 %d1 = arith.constant 1.1 : f64 170 171 // Setup sparse vectors. 172 %v1 = arith.constant sparse< 173 [ [0], [3], [11], [17], [20], [21], [28], [29], [31] ], 174 [ 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0 ] 175 > : tensor<32xf64> 176 %v2 = arith.constant sparse< 177 [ [1], [3], [4], [10], [16], [18], [21], [28], [29], [31] ], 178 [11.0, 12.0, 13.0, 14.0, 15.0, 16.0, 17.0, 18.0, 19.0, 20.0 ] 179 > : tensor<32xf64> 180 %sv1 = sparse_tensor.convert %v1 : tensor<32xf64> to tensor<?xf64, #SparseVector> 181 // TODO: Use %sv1 when copying sparse tensors is supported. 182 %sv1_dup = sparse_tensor.convert %v1 : tensor<32xf64> to tensor<?xf64, #SparseVector> 183 %sv2 = sparse_tensor.convert %v2 : tensor<32xf64> to tensor<?xf64, #SparseVector> 184 185 // Setup memory for a single reduction scalar. 186 %x = tensor.from_elements %d1 : tensor<f64> 187 188 // Call sparse vector kernels. 189 %0 = call @vector_scale(%sv1) 190 : (tensor<?xf64, #SparseVector>) -> tensor<?xf64, #SparseVector> 191 %1 = call @vector_scale_inplace(%sv1_dup) 192 : (tensor<?xf64, #SparseVector>) -> tensor<?xf64, #SparseVector> 193 %2 = call @vector_add(%1, %sv2) 194 : (tensor<?xf64, #SparseVector>, 195 tensor<?xf64, #SparseVector>) -> tensor<?xf64, #SparseVector> 196 %3 = call @vector_mul(%1, %sv2) 197 : (tensor<?xf64, #SparseVector>, 198 tensor<?xf64, #SparseVector>) -> tensor<?xf64, #SparseVector> 199 %4 = call @vector_mul_d(%1, %sv2) 200 : (tensor<?xf64, #SparseVector>, 201 tensor<?xf64, #SparseVector>) -> tensor<?xf64, #DenseVector> 202 %5 = call @vector_dotprod(%1, %sv2, %x) 203 : (tensor<?xf64, #SparseVector>, 204 tensor<?xf64, #SparseVector>, tensor<f64>) -> tensor<f64> 205 206 // 207 // Verify the results. 208 // 209 // CHECK: ---- Sparse Tensor ---- 210 // CHECK-NEXT: nse = 9 211 // CHECK-NEXT: dim = ( 32 ) 212 // CHECK-NEXT: lvl = ( 32 ) 213 // CHECK-NEXT: pos[0] : ( 0, 9 ) 214 // CHECK-NEXT: crd[0] : ( 0, 3, 11, 17, 20, 21, 28, 29, 31 ) 215 // CHECK-NEXT: values : ( 1, 2, 3, 4, 5, 6, 7, 8, 9 ) 216 // CHECK-NEXT: ---- 217 // CHECK: ---- Sparse Tensor ---- 218 // CHECK-NEXT: nse = 10 219 // CHECK-NEXT: dim = ( 32 ) 220 // CHECK-NEXT: lvl = ( 32 ) 221 // CHECK-NEXT: pos[0] : ( 0, 10 ) 222 // CHECK-NEXT: crd[0] : ( 1, 3, 4, 10, 16, 18, 21, 28, 29, 31 ) 223 // CHECK-NEXT: values : ( 11, 12, 13, 14, 15, 16, 17, 18, 19, 20 ) 224 // CHECK-NEXT: ---- 225 // CHECK: ---- Sparse Tensor ---- 226 // CHECK-NEXT: nse = 9 227 // CHECK-NEXT: dim = ( 32 ) 228 // CHECK-NEXT: lvl = ( 32 ) 229 // CHECK-NEXT: pos[0] : ( 0, 9 ) 230 // CHECK-NEXT: crd[0] : ( 0, 3, 11, 17, 20, 21, 28, 29, 31 ) 231 // CHECK-NEXT: values : ( 2, 4, 6, 8, 10, 12, 14, 16, 18 ) 232 // CHECK-NEXT: ---- 233 // CHECK: ---- Sparse Tensor ---- 234 // CHECK-NEXT: nse = 9 235 // CHECK-NEXT: dim = ( 32 ) 236 // CHECK-NEXT: lvl = ( 32 ) 237 // CHECK-NEXT: pos[0] : ( 0, 9 ) 238 // CHECK-NEXT: crd[0] : ( 0, 3, 11, 17, 20, 21, 28, 29, 31 ) 239 // CHECK-NEXT: values : ( 2, 4, 6, 8, 10, 12, 14, 16, 18 ) 240 // CHECK-NEXT: ---- 241 // CHECK: ---- Sparse Tensor ---- 242 // CHECK-NEXT: nse = 14 243 // CHECK-NEXT: dim = ( 32 ) 244 // CHECK-NEXT: lvl = ( 32 ) 245 // CHECK-NEXT: pos[0] : ( 0, 14 ) 246 // CHECK-NEXT: crd[0] : ( 0, 1, 3, 4, 10, 11, 16, 17, 18, 20, 21, 28, 29, 31 ) 247 // CHECK-NEXT: values : ( 2, 11, 16, 13, 14, 6, 15, 8, 16, 10, 29, 32, 35, 38 ) 248 // CHECK-NEXT: ---- 249 // CHECK: ---- Sparse Tensor ---- 250 // CHECK-NEXT: nse = 5 251 // CHECK-NEXT: dim = ( 32 ) 252 // CHECK-NEXT: lvl = ( 32 ) 253 // CHECK-NEXT: pos[0] : ( 0, 5 ) 254 // CHECK-NEXT: crd[0] : ( 3, 21, 28, 29, 31 ) 255 // CHECK-NEXT: values : ( 48, 204, 252, 304, 360 ) 256 // CHECK-NEXT: ---- 257 // CHECK: ---- Sparse Tensor ---- 258 // CHECK-NEXT: nse = 32 259 // CHECK-NEXT: dim = ( 32 ) 260 // CHECK-NEXT: lvl = ( 32 ) 261 // CHECK-NEXT: values : ( 0, 0, 0, 48, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 204, 0, 0, 0, 0, 0, 0, 252, 304, 0, 360 ) 262 // CHECK-NEXT: ---- 263 // CHECK-NEXT: 1169.1 264 // 265 sparse_tensor.print %sv1 : tensor<?xf64, #SparseVector> 266 sparse_tensor.print %sv2 : tensor<?xf64, #SparseVector> 267 sparse_tensor.print %0 : tensor<?xf64, #SparseVector> 268 sparse_tensor.print %1 : tensor<?xf64, #SparseVector> 269 sparse_tensor.print %2 : tensor<?xf64, #SparseVector> 270 sparse_tensor.print %3 : tensor<?xf64, #SparseVector> 271 sparse_tensor.print %4 : tensor<?xf64, #DenseVector> 272 %v5 = tensor.extract %5[] : tensor<f64> 273 vector.print %v5 : f64 274 275 // Release the resources. 276 bufferization.dealloc_tensor %sv1 : tensor<?xf64, #SparseVector> 277 bufferization.dealloc_tensor %sv1_dup : tensor<?xf64, #SparseVector> 278 bufferization.dealloc_tensor %sv2 : tensor<?xf64, #SparseVector> 279 bufferization.dealloc_tensor %0 : tensor<?xf64, #SparseVector> 280 // Note: No dealloc for %1 because it was inplace! 281 bufferization.dealloc_tensor %2 : tensor<?xf64, #SparseVector> 282 bufferization.dealloc_tensor %3 : tensor<?xf64, #SparseVector> 283 bufferization.dealloc_tensor %4 : tensor<?xf64, #DenseVector> 284 bufferization.dealloc_tensor %5 : tensor<f64> 285 return 286 } 287} 288