brintos

brintos / llvm-project-archived public Read only

0
0
Text · 12.1 KiB · 42ed78d Raw
288 lines · plain
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, this5// config could be moved to lit.local.cfg. However, there are downstream users that6//  do not use these LIT config files. Hence why this is kept inline.7//8// DEFINE: %{sparsifier_opts} = enable-runtime-library=true9// 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_utils13// DEFINE: %{run_libs_sve} = -shared-libs=%native_mlir_runner_utils,%native_mlir_c_runner_utils14// DEFINE: %{run_opts} = -e main -entry-point-result=void15// 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 %s22//23// Do the same run, but now with direct IR generation.24// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false25// RUN: %{compile} | %{run} | FileCheck %s26//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=true29// RUN: %{compile} | %{run} | FileCheck %s30//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_2d52      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 : i3266        %2 = arith.addi %out, %1 : i3267        linalg.yield %2 : i3268    } -> 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_2d76      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_2d85      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_2d94      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_2d103      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_2d112      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 : index119    %i0 = arith.constant 0 : i32120 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 %input139      : tensor<8x8xi32> to tensor<8x8xi32, #DCSR>140    %sparse_input_CSR = sparse_tensor.convert %input141      : tensor<8x8xi32> to tensor<8x8xi32, #CSR>142    %sparse_input_CD = sparse_tensor.convert %input143      : tensor<8x8xi32> to tensor<8x8xi32, #CDR>144    %sparse_input_CSC = sparse_tensor.convert %input145      : 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], %i0181      : 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 = 36189    // 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 = 36205    // 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 = 36221    // 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 = 36235    // 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 = 36249    // 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], %i0268      : 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    return286  }287}288