1// RUN: mlir-opt %s --sparse-compiler | \ 2// RUN: mlir-cpu-runner -e entry -entry-point-result=void \ 3// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext | \ 4// RUN: FileCheck %s 5 6#SparseVector = #sparse_tensor.encoding<{ 7 dimLevelType = ["compressed"] 8}> 9 10#SparseMatrix = #sparse_tensor.encoding<{ 11 dimLevelType = ["compressed", "compressed"] 12}> 13 14#trait_1d = { 15 indexing_maps = [ 16 affine_map<(i) -> (i)>, // a 17 affine_map<(i) -> (i)> // x (out) 18 ], 19 iterator_types = ["parallel"], 20 doc = "X(i) = a(i) op i" 21} 22 23#trait_2d = { 24 indexing_maps = [ 25 affine_map<(i,j) -> (i,j)>, // A 26 affine_map<(i,j) -> (i,j)> // X (out) 27 ], 28 iterator_types = ["parallel", "parallel"], 29 doc = "X(i,j) = A(i,j) op i op j" 30} 31 32// 33// Test with indices. Note that a lot of results are actually 34// dense, but this is done to stress test all the operations. 35// 36module { 37 38 // 39 // Kernel that uses index in the index notation (conjunction). 40 // 41 func.func @sparse_index_1d_conj(%arga: tensor<8xi64, #SparseVector>) 42 -> tensor<8xi64, #SparseVector> { 43 %d0 = arith.constant 8 : index 44 %init = sparse_tensor.init [%d0] : tensor<8xi64, #SparseVector> 45 %r = linalg.generic #trait_1d 46 ins(%arga: tensor<8xi64, #SparseVector>) 47 outs(%init: tensor<8xi64, #SparseVector>) { 48 ^bb(%a: i64, %x: i64): 49 %i = linalg.index 0 : index 50 %ii = arith.index_cast %i : index to i64 51 %m1 = arith.muli %a, %ii : i64 52 linalg.yield %m1 : i64 53 } -> tensor<8xi64, #SparseVector> 54 return %r : tensor<8xi64, #SparseVector> 55 } 56 57 // 58 // Kernel that uses index in the index notation (disjunction). 59 // 60 func.func @sparse_index_1d_disj(%arga: tensor<8xi64, #SparseVector>) 61 -> tensor<8xi64, #SparseVector> { 62 %d0 = arith.constant 8 : index 63 %init = sparse_tensor.init [%d0] : tensor<8xi64, #SparseVector> 64 %r = linalg.generic #trait_1d 65 ins(%arga: tensor<8xi64, #SparseVector>) 66 outs(%init: tensor<8xi64, #SparseVector>) { 67 ^bb(%a: i64, %x: i64): 68 %i = linalg.index 0 : index 69 %ii = arith.index_cast %i : index to i64 70 %m1 = arith.addi %a, %ii : i64 71 linalg.yield %m1 : i64 72 } -> tensor<8xi64, #SparseVector> 73 return %r : tensor<8xi64, #SparseVector> 74 } 75 76 // 77 // Kernel that uses indices in the index notation (conjunction). 78 // 79 func.func @sparse_index_2d_conj(%arga: tensor<3x4xi64, #SparseMatrix>) 80 -> tensor<3x4xi64, #SparseMatrix> { 81 %d0 = arith.constant 3 : index 82 %d1 = arith.constant 4 : index 83 %init = sparse_tensor.init [%d0, %d1] : tensor<3x4xi64, #SparseMatrix> 84 %r = linalg.generic #trait_2d 85 ins(%arga: tensor<3x4xi64, #SparseMatrix>) 86 outs(%init: tensor<3x4xi64, #SparseMatrix>) { 87 ^bb(%a: i64, %x: i64): 88 %i = linalg.index 0 : index 89 %j = linalg.index 1 : index 90 %ii = arith.index_cast %i : index to i64 91 %jj = arith.index_cast %j : index to i64 92 %m1 = arith.muli %ii, %a : i64 93 %m2 = arith.muli %jj, %m1 : i64 94 linalg.yield %m2 : i64 95 } -> tensor<3x4xi64, #SparseMatrix> 96 return %r : tensor<3x4xi64, #SparseMatrix> 97 } 98 99 // 100 // Kernel that uses indices in the index notation (disjunction). 101 // 102 func.func @sparse_index_2d_disj(%arga: tensor<3x4xi64, #SparseMatrix>) 103 -> tensor<3x4xi64, #SparseMatrix> { 104 %d0 = arith.constant 3 : index 105 %d1 = arith.constant 4 : index 106 %init = sparse_tensor.init [%d0, %d1] : tensor<3x4xi64, #SparseMatrix> 107 %r = linalg.generic #trait_2d 108 ins(%arga: tensor<3x4xi64, #SparseMatrix>) 109 outs(%init: tensor<3x4xi64, #SparseMatrix>) { 110 ^bb(%a: i64, %x: i64): 111 %i = linalg.index 0 : index 112 %j = linalg.index 1 : index 113 %ii = arith.index_cast %i : index to i64 114 %jj = arith.index_cast %j : index to i64 115 %m1 = arith.addi %ii, %a : i64 116 %m2 = arith.addi %jj, %m1 : i64 117 linalg.yield %m2 : i64 118 } -> tensor<3x4xi64, #SparseMatrix> 119 return %r : tensor<3x4xi64, #SparseMatrix> 120 } 121 122 func.func @add_outer_2d(%arg0: tensor<2x3xf32, #SparseMatrix>) 123 -> tensor<2x3xf32, #SparseMatrix> { 124 %c2 = arith.constant 2 : index 125 %c3 = arith.constant 3 : index 126 %0 = sparse_tensor.init[%c2, %c3] : tensor<2x3xf32, #SparseMatrix> 127 %1 = linalg.generic #trait_2d 128 ins(%arg0 : tensor<2x3xf32, #SparseMatrix>) 129 outs(%0 : tensor<2x3xf32, #SparseMatrix>) { 130 ^bb0(%arg1: f32, %arg2: f32): 131 %2 = linalg.index 0 : index 132 %3 = arith.index_cast %2 : index to i64 133 %4 = arith.uitofp %3 : i64 to f32 134 %5 = arith.addf %arg1, %4 : f32 135 linalg.yield %5 : f32 136 } -> tensor<2x3xf32, #SparseMatrix> 137 return %1 : tensor<2x3xf32, #SparseMatrix> 138 } 139 140 // 141 // Main driver. 142 // 143 func.func @entry() { 144 %c0 = arith.constant 0 : index 145 %du = arith.constant -1 : i64 146 %df = arith.constant -1.0 : f32 147 148 // Setup input sparse vector. 149 %v1 = arith.constant sparse<[[2], [4]], [ 10, 20]> : tensor<8xi64> 150 %sv = sparse_tensor.convert %v1 : tensor<8xi64> to tensor<8xi64, #SparseVector> 151 152 // Setup input "sparse" vector. 153 %v2 = arith.constant dense<[ 1, 2, 4, 8, 16, 32, 64, 128 ]> : tensor<8xi64> 154 %dv = sparse_tensor.convert %v2 : tensor<8xi64> to tensor<8xi64, #SparseVector> 155 156 // Setup input sparse matrix. 157 %m1 = arith.constant sparse<[[1,1], [2,3]], [10, 20]> : tensor<3x4xi64> 158 %sm = sparse_tensor.convert %m1 : tensor<3x4xi64> to tensor<3x4xi64, #SparseMatrix> 159 160 // Setup input "sparse" matrix. 161 %m2 = arith.constant dense <[ [ 1, 1, 1, 1 ], 162 [ 1, 2, 1, 1 ], 163 [ 1, 1, 3, 4 ] ]> : tensor<3x4xi64> 164 %dm = sparse_tensor.convert %m2 : tensor<3x4xi64> to tensor<3x4xi64, #SparseMatrix> 165 166 // Setup input sparse f32 matrix. 167 %mf32 = arith.constant sparse<[[0,1], [1,2]], [10.0, 41.0]> : tensor<2x3xf32> 168 %sf32 = sparse_tensor.convert %mf32 : tensor<2x3xf32> to tensor<2x3xf32, #SparseMatrix> 169 170 // Call the kernels. 171 %0 = call @sparse_index_1d_conj(%sv) : (tensor<8xi64, #SparseVector>) 172 -> tensor<8xi64, #SparseVector> 173 %1 = call @sparse_index_1d_disj(%sv) : (tensor<8xi64, #SparseVector>) 174 -> tensor<8xi64, #SparseVector> 175 %2 = call @sparse_index_1d_conj(%dv) : (tensor<8xi64, #SparseVector>) 176 -> tensor<8xi64, #SparseVector> 177 %3 = call @sparse_index_1d_disj(%dv) : (tensor<8xi64, #SparseVector>) 178 -> tensor<8xi64, #SparseVector> 179 %4 = call @sparse_index_2d_conj(%sm) : (tensor<3x4xi64, #SparseMatrix>) 180 -> tensor<3x4xi64, #SparseMatrix> 181 %5 = call @sparse_index_2d_disj(%sm) : (tensor<3x4xi64, #SparseMatrix>) 182 -> tensor<3x4xi64, #SparseMatrix> 183 %6 = call @sparse_index_2d_conj(%dm) : (tensor<3x4xi64, #SparseMatrix>) 184 -> tensor<3x4xi64, #SparseMatrix> 185 %7 = call @sparse_index_2d_disj(%dm) : (tensor<3x4xi64, #SparseMatrix>) 186 -> tensor<3x4xi64, #SparseMatrix> 187 188 // 189 // Verify result. 190 // 191 // CHECK: ( 20, 80, -1, -1, -1, -1, -1, -1 ) 192 // CHECK-NEXT: ( 0, 1, 12, 3, 24, 5, 6, 7 ) 193 // CHECK-NEXT: ( 0, 2, 8, 24, 64, 160, 384, 896 ) 194 // CHECK-NEXT: ( 1, 3, 6, 11, 20, 37, 70, 135 ) 195 // CHECK-NEXT: ( 10, 120, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1 ) 196 // CHECK-NEXT: ( 0, 1, 2, 3, 1, 12, 3, 4, 2, 3, 4, 25 ) 197 // CHECK-NEXT: ( 0, 0, 0, 0, 0, 2, 2, 3, 0, 2, 12, 24 ) 198 // CHECK-NEXT: ( 1, 2, 3, 4, 2, 4, 4, 5, 3, 4, 7, 9 ) 199 // 200 %8 = sparse_tensor.values %0 : tensor<8xi64, #SparseVector> to memref<?xi64> 201 %9 = sparse_tensor.values %1 : tensor<8xi64, #SparseVector> to memref<?xi64> 202 %10 = sparse_tensor.values %2 : tensor<8xi64, #SparseVector> to memref<?xi64> 203 %11 = sparse_tensor.values %3 : tensor<8xi64, #SparseVector> to memref<?xi64> 204 %12 = sparse_tensor.values %4 : tensor<3x4xi64, #SparseMatrix> to memref<?xi64> 205 %13 = sparse_tensor.values %5 : tensor<3x4xi64, #SparseMatrix> to memref<?xi64> 206 %14 = sparse_tensor.values %6 : tensor<3x4xi64, #SparseMatrix> to memref<?xi64> 207 %15 = sparse_tensor.values %7 : tensor<3x4xi64, #SparseMatrix> to memref<?xi64> 208 %16 = vector.transfer_read %8[%c0], %du: memref<?xi64>, vector<8xi64> 209 %17 = vector.transfer_read %9[%c0], %du: memref<?xi64>, vector<8xi64> 210 %18 = vector.transfer_read %10[%c0], %du: memref<?xi64>, vector<8xi64> 211 %19 = vector.transfer_read %11[%c0], %du: memref<?xi64>, vector<8xi64> 212 %20 = vector.transfer_read %12[%c0], %du: memref<?xi64>, vector<12xi64> 213 %21 = vector.transfer_read %13[%c0], %du: memref<?xi64>, vector<12xi64> 214 %22 = vector.transfer_read %14[%c0], %du: memref<?xi64>, vector<12xi64> 215 %23 = vector.transfer_read %15[%c0], %du: memref<?xi64>, vector<12xi64> 216 vector.print %16 : vector<8xi64> 217 vector.print %17 : vector<8xi64> 218 vector.print %18 : vector<8xi64> 219 vector.print %19 : vector<8xi64> 220 vector.print %20 : vector<12xi64> 221 vector.print %21 : vector<12xi64> 222 vector.print %22 : vector<12xi64> 223 vector.print %23 : vector<12xi64> 224 225 // Release resources. 226 sparse_tensor.release %sv : tensor<8xi64, #SparseVector> 227 sparse_tensor.release %dv : tensor<8xi64, #SparseVector> 228 sparse_tensor.release %0 : tensor<8xi64, #SparseVector> 229 sparse_tensor.release %1 : tensor<8xi64, #SparseVector> 230 sparse_tensor.release %2 : tensor<8xi64, #SparseVector> 231 sparse_tensor.release %3 : tensor<8xi64, #SparseVector> 232 sparse_tensor.release %sm : tensor<3x4xi64, #SparseMatrix> 233 sparse_tensor.release %dm : tensor<3x4xi64, #SparseMatrix> 234 sparse_tensor.release %4 : tensor<3x4xi64, #SparseMatrix> 235 sparse_tensor.release %5 : tensor<3x4xi64, #SparseMatrix> 236 sparse_tensor.release %6 : tensor<3x4xi64, #SparseMatrix> 237 sparse_tensor.release %7 : tensor<3x4xi64, #SparseMatrix> 238 239 // 240 // Call the f32 kernel, verify the result, release the resources. 241 // 242 // CHECK-NEXT: ( 0, 10, 0, 1, 1, 42 ) 243 // 244 %100 = call @add_outer_2d(%sf32) : (tensor<2x3xf32, #SparseMatrix>) 245 -> tensor<2x3xf32, #SparseMatrix> 246 %101 = sparse_tensor.values %100 : tensor<2x3xf32, #SparseMatrix> to memref<?xf32> 247 %102 = vector.transfer_read %101[%c0], %df: memref<?xf32>, vector<6xf32> 248 vector.print %102 : vector<6xf32> 249 sparse_tensor.release %sf32 : tensor<2x3xf32, #SparseMatrix> 250 sparse_tensor.release %100 : tensor<2x3xf32, #SparseMatrix> 251 252 return 253 } 254} 255