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#DCSR = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : compressed, d1 : compressed) }> 35#CSR = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : dense, d1 : compressed) }> 36#CDR = #sparse_tensor.encoding<{map = (d0, d1) -> (d0 : compressed, d1 : dense)}> 37#CSC = #sparse_tensor.encoding<{ 38 map = (d0, d1) -> (d1 : dense, d0 : compressed) 39}> 40 41#map = affine_map<(d0, d1, d2, d3) -> (d0 + d1, d3 + d2)> 42#map1 = affine_map<(d0, d1, d2, d3) -> (d1, d2)> 43#map2 = affine_map<(d0, d1, d2, d3) -> (d0, d3)> 44 45// An example of a 2D convolution with a sparse filter. 46module { 47 48 func.func @conv2d(%input: tensor<8x8xi32>, 49 %filter: tensor<3x3xi32>, 50 %output: tensor<6x6xi32>) -> tensor<6x6xi32> { 51 %0 = linalg.conv_2d 52 ins (%input, %filter: tensor<8x8xi32>, tensor<3x3xi32>) 53 outs (%output: tensor<6x6xi32>) -> tensor<6x6xi32> 54 return %0 : tensor<6x6xi32> 55 } 56 57 func.func @conv2d_CSR_dense_rotated(%arg0: tensor<8x8xi32, #CSR>, 58 %arg1: tensor<3x3xi32>) -> tensor<6x6xi32> { 59 %s = arith.constant dense<0> : tensor<6x6xi32> 60 %0 = linalg.generic {indexing_maps = [#map, #map1, #map2], 61 iterator_types = ["parallel", "reduction", "reduction", "parallel"]} 62 ins(%arg0, %arg1 : tensor<8x8xi32, #CSR>, tensor<3x3xi32>) 63 outs(%s : tensor<6x6xi32>) attrs = {sorted = true} { 64 ^bb0(%in: i32, %in_0: i32, %out: i32): 65 %1 = arith.muli %in, %in_0 : i32 66 %2 = arith.addi %out, %1 : i32 67 linalg.yield %2 : i32 68 } -> tensor<6x6xi32> 69 return %0 : tensor<6x6xi32> 70 } 71 72 func.func @conv2d_sparse_out(%input: tensor<8x8xi32>, 73 %filter: tensor<3x3xi32>) -> tensor<6x6xi32, #DCSR> { 74 %s = tensor.empty() : tensor<6x6xi32, #DCSR> 75 %0 = linalg.conv_2d 76 ins (%input, %filter: tensor<8x8xi32>, tensor<3x3xi32>) 77 outs (%s: tensor<6x6xi32, #DCSR>) -> tensor<6x6xi32, #DCSR> 78 return %0 : tensor<6x6xi32, #DCSR> 79 } 80 81 func.func @conv2d_all_sparse_DCSR(%input: tensor<8x8xi32, #DCSR>, 82 %filter: tensor<3x3xi32>) -> tensor<6x6xi32, #DCSR> { 83 %s = tensor.empty() : tensor<6x6xi32, #DCSR> 84 %0 = linalg.conv_2d 85 ins (%input, %filter: tensor<8x8xi32, #DCSR>, tensor<3x3xi32>) 86 outs (%s: tensor<6x6xi32, #DCSR>) -> tensor<6x6xi32, #DCSR> 87 return %0 : tensor<6x6xi32, #DCSR> 88 } 89 90 func.func @conv2d_all_sparse_CSR(%input: tensor<8x8xi32, #CSR>, 91 %filter: tensor<3x3xi32>) -> tensor<6x6xi32, #CSR> { 92 %s = tensor.empty() : tensor<6x6xi32, #CSR> 93 %0 = linalg.conv_2d 94 ins (%input, %filter: tensor<8x8xi32, #CSR>, tensor<3x3xi32>) 95 outs (%s: tensor<6x6xi32, #CSR>) -> tensor<6x6xi32, #CSR> 96 return %0 : tensor<6x6xi32, #CSR> 97 } 98 99 func.func @conv2d_all_sparse_CD(%input: tensor<8x8xi32, #CDR>, 100 %filter: tensor<3x3xi32>) -> tensor<6x6xi32, #CDR> { 101 %s = tensor.empty() : tensor<6x6xi32, #CDR> 102 %0 = linalg.conv_2d 103 ins (%input, %filter: tensor<8x8xi32, #CDR>, tensor<3x3xi32>) 104 outs (%s: tensor<6x6xi32, #CDR>) -> tensor<6x6xi32, #CDR> 105 return %0 : tensor<6x6xi32, #CDR> 106 } 107 108 func.func @conv2d_all_sparse_CSC(%input: tensor<8x8xi32, #CSC>, 109 %filter: tensor<3x3xi32>) -> tensor<6x6xi32, #CSC> { 110 %s = tensor.empty() : tensor<6x6xi32, #CSC> 111 %0 = linalg.conv_2d 112 ins (%input, %filter: tensor<8x8xi32, #CSC>, tensor<3x3xi32>) 113 outs (%s: tensor<6x6xi32, #CSC>) -> tensor<6x6xi32, #CSC> 114 return %0 : tensor<6x6xi32, #CSC> 115 } 116 117 func.func @main() { 118 %c0 = arith.constant 0 : index 119 %i0 = arith.constant 0 : i32 120 121 // A typical edge detection filter. 122 %filter = arith.constant dense<[ 123 [ 1, 0, -1 ], 124 [ 0, 0, 0 ], 125 [ -1, 0, 1 ] 126 ]> : tensor<3x3xi32> 127 128 %input = arith.constant dense<[ 129 [ 1, 2, 3, 4, 0, 6, 7, 8 ], 130 [ 2, 2, 4, 4, 0, 0, 6, 8 ], 131 [ 2, 2, 4, 4, 0, 0, 6, 8 ], 132 [ 2, 2, 3, 4, 0, 0, 7, 8 ], 133 [ 1, 3, 3, 4, 0, 0, 6, 8 ], 134 [ 3, 2, 3, 4, 0, 0, 7, 8 ], 135 [ 1, 3, 3, 4, 3, 6, 6, 8 ], 136 [ 1, 3, 3, 4, 3, 0, 7, 8 ] 137 ]> : tensor<8x8xi32> 138 %sparse_input_DCSR = sparse_tensor.convert %input 139 : tensor<8x8xi32> to tensor<8x8xi32, #DCSR> 140 %sparse_input_CSR = sparse_tensor.convert %input 141 : tensor<8x8xi32> to tensor<8x8xi32, #CSR> 142 %sparse_input_CD = sparse_tensor.convert %input 143 : tensor<8x8xi32> to tensor<8x8xi32, #CDR> 144 %sparse_input_CSC = sparse_tensor.convert %input 145 : tensor<8x8xi32> to tensor<8x8xi32, #CSC> 146 147 // Call the kernel. 148 %output = arith.constant dense<0> : tensor<6x6xi32> 149 %0 = call @conv2d(%input, %filter, %output) 150 : (tensor<8x8xi32>, 151 tensor<3x3xi32>, tensor<6x6xi32>) -> tensor<6x6xi32> 152 %1 = call @conv2d_sparse_out(%input, %filter) 153 : (tensor<8x8xi32>, 154 tensor<3x3xi32>) -> tensor<6x6xi32, #DCSR> 155 %2 = call @conv2d_all_sparse_DCSR(%sparse_input_DCSR, %filter) 156 : (tensor<8x8xi32, #DCSR>, 157 tensor<3x3xi32>) -> tensor<6x6xi32, #DCSR> 158 %3 = call @conv2d_all_sparse_CSR(%sparse_input_CSR, %filter) 159 : (tensor<8x8xi32, #CSR>, 160 tensor<3x3xi32>) -> tensor<6x6xi32, #CSR> 161 %4 = call @conv2d_all_sparse_CD(%sparse_input_CD, %filter) 162 : (tensor<8x8xi32, #CDR>, 163 tensor<3x3xi32>) -> tensor<6x6xi32, #CDR> 164 %5 = call @conv2d_all_sparse_CSC(%sparse_input_CSC, %filter) 165 : (tensor<8x8xi32, #CSC>, 166 tensor<3x3xi32>) -> tensor<6x6xi32, #CSC> 167 %6 = call @conv2d_CSR_dense_rotated(%sparse_input_CSR, %filter) 168 : (tensor<8x8xi32, #CSR>, 169 tensor<3x3xi32>) -> tensor<6x6xi32> 170 171 // Verify the output. 172 // 173 // CHECK: ( ( 0, 0, -1, -6, -1, 6 ), 174 // CHECK-SAME: ( -1, 0, 1, 0, 1, 0 ), 175 // CHECK-SAME: ( 0, -1, 1, 0, 0, 0 ), 176 // CHECK-SAME: ( -1, 0, 0, 0, 0, 0 ), 177 // CHECK-SAME: ( 0, 0, 3, 6, -3, -6 ), 178 // CHECK-SAME: ( 2, -1, 3, 0, -3, 0 ) ) 179 // 180 %v = vector.transfer_read %0[%c0, %c0], %i0 181 : tensor<6x6xi32>, vector<6x6xi32> 182 vector.print %v : vector<6x6xi32> 183 184 // 185 // Should be the same as dense output. 186 // 187 // CHECK: ---- Sparse Tensor ---- 188 // CHECK-NEXT: nse = 36 189 // CHECK-NEXT: dim = ( 6, 6 ) 190 // CHECK-NEXT: lvl = ( 6, 6 ) 191 // CHECK-NEXT: pos[0] : ( 0, 6 ) 192 // CHECK-NEXT: crd[0] : ( 0, 1, 2, 3, 4, 5 ) 193 // CHECK-NEXT: pos[1] : ( 0, 6, 12, 18, 24, 30, 36 ) 194 // CHECK-NEXT: crd[1] : ( 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5 ) 195 // CHECK-NEXT: values : ( 0, 0, -1, -6, -1, 6, -1, 0, 1, 0, 1, 0, 0, -1, 1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 3, 6, -3, -6, 2, -1, 3, 0, -3, 0 ) 196 // CHECK-NEXT: ---- 197 // 198 sparse_tensor.print %1 : tensor<6x6xi32, #DCSR> 199 200 // 201 // Should be the same as dense output. 202 // 203 // CHECK: ---- Sparse Tensor ---- 204 // CHECK-NEXT: nse = 36 205 // CHECK-NEXT: dim = ( 6, 6 ) 206 // CHECK-NEXT: lvl = ( 6, 6 ) 207 // CHECK-NEXT: pos[0] : ( 0, 6 ) 208 // CHECK-NEXT: crd[0] : ( 0, 1, 2, 3, 4, 5 ) 209 // CHECK-NEXT: pos[1] : ( 0, 6, 12, 18, 24, 30, 36 ) 210 // CHECK-NEXT: crd[1] : ( 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5 ) 211 // CHECK-NEXT: values : ( 0, 0, -1, -6, -1, 6, -1, 0, 1, 0, 1, 0, 0, -1, 1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 3, 6, -3, -6, 2, -1, 3, 0, -3, 0 ) 212 // CHECK-NEXT: ---- 213 // 214 sparse_tensor.print %2 : tensor<6x6xi32, #DCSR> 215 216 // 217 // Should be the same as dense output. 218 // 219 // CHECK: ---- Sparse Tensor ---- 220 // CHECK-NEXT: nse = 36 221 // CHECK-NEXT: dim = ( 6, 6 ) 222 // CHECK-NEXT: lvl = ( 6, 6 ) 223 // CHECK-NEXT: pos[1] : ( 0, 6, 12, 18, 24, 30, 36 ) 224 // CHECK-NEXT: crd[1] : ( 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5 ) 225 // CHECK-NEXT: values : ( 0, 0, -1, -6, -1, 6, -1, 0, 1, 0, 1, 0, 0, -1, 1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 3, 6, -3, -6, 2, -1, 3, 0, -3, 0 ) 226 // CHECK-NEXT: ---- 227 // 228 sparse_tensor.print %3 : tensor<6x6xi32, #CSR> 229 230 // 231 // Should be the same as dense output. 232 // 233 // CHECK: ---- Sparse Tensor ---- 234 // CHECK-NEXT: nse = 36 235 // CHECK-NEXT: dim = ( 6, 6 ) 236 // CHECK-NEXT: lvl = ( 6, 6 ) 237 // CHECK-NEXT: pos[0] : ( 0, 6 ) 238 // CHECK-NEXT: crd[0] : ( 0, 1, 2, 3, 4, 5 ) 239 // CHECK-NEXT: values : ( 0, 0, -1, -6, -1, 6, -1, 0, 1, 0, 1, 0, 0, -1, 1, 0, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 3, 6, -3, -6, 2, -1, 3, 0, -3, 0 ) 240 // CHECK-NEXT: ---- 241 // 242 sparse_tensor.print %4 : tensor<6x6xi32, #CDR> 243 244 // 245 // Should be the same as dense output. 246 // 247 // CHECK: ---- Sparse Tensor ---- 248 // CHECK-NEXT: nse = 36 249 // CHECK-NEXT: dim = ( 6, 6 ) 250 // CHECK-NEXT: lvl = ( 6, 6 ) 251 // CHECK-NEXT: pos[1] : ( 0, 6, 12, 18, 24, 30, 36 ) 252 // CHECK-NEXT: crd[1] : ( 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5 ) 253 // CHECK-NEXT: values : ( 0, -1, 0, -1, 0, 2, 0, 0, -1, 0, 0, -1, -1, 1, 1, 0, 3, 3, -6, 0, 0, 0, 6, 0, -1, 1, 0, 0, -3, -3, 6, 0, 0, 0, -6, 0 ) 254 // CHECK-NEXT: ---- 255 // 256 sparse_tensor.print %5 : tensor<6x6xi32, #CSC> 257 258 // 259 // Should be the same as dense output. 260 // CHECK: ( ( 0, 0, -1, -6, -1, 6 ), 261 // CHECK-SAME: ( -1, 0, 1, 0, 1, 0 ), 262 // CHECK-SAME: ( 0, -1, 1, 0, 0, 0 ), 263 // CHECK-SAME: ( -1, 0, 0, 0, 0, 0 ), 264 // CHECK-SAME: ( 0, 0, 3, 6, -3, -6 ), 265 // CHECK-SAME: ( 2, -1, 3, 0, -3, 0 ) ) 266 // 267 %v6 = vector.transfer_read %6[%c0, %c0], %i0 268 : tensor<6x6xi32>, vector<6x6xi32> 269 vector.print %v : vector<6x6xi32> 270 271 // Release the resources. 272 bufferization.dealloc_tensor %sparse_input_DCSR : tensor<8x8xi32, #DCSR> 273 bufferization.dealloc_tensor %sparse_input_CSR : tensor<8x8xi32, #CSR> 274 bufferization.dealloc_tensor %sparse_input_CSC : tensor<8x8xi32, #CSC> 275 bufferization.dealloc_tensor %sparse_input_CD : tensor<8x8xi32, #CDR> 276 277 bufferization.dealloc_tensor %0 : tensor<6x6xi32> 278 bufferization.dealloc_tensor %1 : tensor<6x6xi32, #DCSR> 279 bufferization.dealloc_tensor %2 : tensor<6x6xi32, #DCSR> 280 bufferization.dealloc_tensor %3 : tensor<6x6xi32, #CSR> 281 bufferization.dealloc_tensor %4 : tensor<6x6xi32, #CDR> 282 bufferization.dealloc_tensor %5 : tensor<6x6xi32, #CSC> 283 bufferization.dealloc_tensor %6 : tensor<6x6xi32> 284 285 return 286 } 287} 288