brintos

brintos / llvm-project-archived public Read only

0
0
Text · 20.0 KiB · 45fa35d Raw
460 lines · plain
1// REQUIRES: arm-emulator2 3// DEFINE: %{compile} = mlir-opt %s \4// DEFINE:   --convert-vector-to-scf --convert-scf-to-cf  --convert-vector-to-llvm='enable-arm-sve enable-arm-i8mm' \5// DEFINE:   --expand-strided-metadata --convert-to-llvm --finalize-memref-to-llvm  \6// DEFINE:   --lower-affine --convert-arith-to-llvm --reconcile-unrealized-casts \7// DEFINE: -o %t8 9// DEFINE: %{entry_point} = main10 11// DEFINE: %{run} = %mcr_aarch64_cmd %t -e %{entry_point} -entry-point-result=void  --march=aarch64 --mattr="+sve,+i8mm" \12// DEFINE:    -shared-libs=%mlir_runner_utils,%mlir_c_runner_utils,%native_mlir_arm_runner_utils13 14// RUN: rm -f %t && %{compile} && FileCheck %s --input-file=%t -check-prefix CHECK-IR && %{run} | FileCheck %s15 16#packed_maps = [17  affine_map<(m, n, k) -> (m, k)>,18  affine_map<(m, n, k) -> (n, k)>,19  affine_map<(m, n, k) -> (m, n)>20]21 22//23// Test the lowering of `vector.contract` using the `LowerContractionToSVEBFMMLAPattern`24//25// The operation that the `vector.contract` in this test performs is matrix26// multiplication with accumulate27//     OUT = ACC + LHS * RHS28// of two 8-bit integer matrices LHS and RHS, and a 32-bit integer matrix ACC29// into a 32-bit integer matrix OUT. The LHS and RHS can be sign- or zero- extended,30// this test covers all the possible variants.31//32// Tested are calculations as well as that the relevant `ArmSVE` dialect33// operations ('arm_sve.smmla`, arm_sve.ummla`, etc) are emitted.34//35// That pattern above handles (therefore this test prepares) input/output vectors with36// specific shapes:37//   * LHS:      vector<Mx8xi8>38//   * RHS:      vector<[N]x8xi8>39//   * ACC, OUT: vector<Mx[N]xi32>40// Note that the RHS is transposed.41// This data layout makes it efficient to load data into SVE42// registers in the layout expected by FEAT_I8MM instructions.43// Such a `vector.contract` is representative of the code we aim to generate44// by scalable vectorisation of `linalg.mmt4d`.45// See mlir/lib/Dialect/ArmSVE/Transforms/LowerContractToSVEPatterns.cpp46// for more information and rationale about these shapes.47//48// In this specific test we use M == 4 and N == 449//50 51// Allocate and initialise a memref containing test data for use as the ACC52// operand. The memref has one dynamic dimension whose extent depends on the53// runtime value of VSCALE.54//55// The input parameter `%in` is a vector that is replicated VSCALE times56// across the columns of the memref.57func.func private @prepareAccTestData(%in: vector<4x4xi32>) -> memref<4x?xi32> {58  %c0 = arith.constant 0 : index59  %c1 = arith.constant 1 : index60  %c4 = arith.constant 4 : index61  %c0_i32 = arith.constant 0 : i3262 63  %vs = vector.vscale64  %d = arith.muli %c4, %vs : index65  %mem = memref.alloc(%d) : memref<4x?xi32>66 67  scf.for %j = %c0 to %d step %c4 {68    vector.transfer_write %in, %mem[%c0, %j] {in_bounds = [true, true]} : vector<4x4xi32>, memref<4x?xi32>69  }70 71  return %mem : memref<4x?xi32>72}73 74// Allocate and initialise a memref containing test data for use as the LHS75// operand. This function just writes the parameter `%in` into the memref.76// The size of the LHS does not depends on VSCALE.77func.func private @prepareLHSTestData(%in: vector<4x8xi8>) -> memref<4x8xi8> {78  %c0 = arith.constant 0 : index79  %c0_i8 = arith.constant 0 : i880 81  %mem = memref.alloc() : memref<4x8xi8>82  vector.transfer_write %in, %mem[%c0, %c0] {in_bounds = [true, true]} : vector<4x8xi8>, memref<4x8xi8>83 84  return %mem : memref<4x8xi8>85}86 87// Allocate and initialise a memref containing test data for use as the RHS88// operand. The memref has one dynamic dimension whose extent depends on the89// runtime value of VSCALE.90//91// The input parameter `%in` is a vector that is replicated VSCALE times92// across the rows of the memref.93//94// For convenience, flatten the memref, since the RHS vector is read first as a95// single-dimensional scalable vector and then cast into [N]x8 shape.96func.func private @prepareRHSTestData(%in: vector<4x8xi8>) -> memref<?xi8> {97  %c0 = arith.constant 0 : index98  %c1 = arith.constant 1 : index99  %c4 = arith.constant 4 : index100  %c0_i8 = arith.constant 0 : i8101 102  %vs = vector.vscale103  %d = arith.muli %c4, %vs : index104  %mem = memref.alloc(%d) : memref<?x8xi8>105 106  scf.for %i = %c0 to %d step %c4 {107    vector.transfer_write %in, %mem[%i, %c0] {in_bounds = [true, true]} : vector<4x8xi8>, memref<?x8xi8>108  }109 110  %mem_out = memref.collapse_shape %mem [[0, 1]] : memref<?x8xi8> into memref<?xi8>111  return %mem_out : memref<?xi8>112}113 114// Test the operation where both LHS and RHS are interpreted as signed, hence115// we ultimately emit and execute the `smmla` instruction.116 117// CHECK-IR-LABEL: llvm.func @test_smmla118// CHECK-IR-COUNT-4: arm_sve.intr.smmla119func.func @test_smmla() attributes {no_inline} {120 121  %c0 = arith.constant 0 : index122  %c0_i32 = arith.constant 0 : i32123  %c0_i8 = arith.constant 0 : i8124 125  // Accumulator test data126  %acc_cst = arith.constant dense<[[-44,  20,  44, -46],127                                   [ -8,  25, -34,  26],128                                   [-20, -36,  -3,  39],129                                   [-48, -31, -25, -21]]> : vector<4x4xi32>130 131  %acc_mem = func.call @prepareAccTestData(%acc_cst) : (vector<4x4xi32>) -> memref<4x?xi32>132  %acc = vector.transfer_read %acc_mem[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : memref<4x?xi32>, vector<4x[4]xi32>133 134  // LHS test data135  %lhs_cst = arith.constant dense<[[-35, -27, -36, -31,  23, -34,  -8, -33],136                                   [-20,  17, -32, -47,  37,  22,  -7, -21],137                                   [ -7, -35,  20,  -4,  39,  46, -23,  40],138                                   [ 40,  27,  37,  43,  38,  -6,  37,  49]]> : vector<4x8xi8>139 140  %lhs_mem = func.call @prepareLHSTestData(%lhs_cst) : (vector<4x8xi8>) -> memref<4x8xi8>141  %lhs = vector.transfer_read %lhs_mem[%c0, %c0], %c0_i8 {in_bounds = [true, true]} : memref<4x8xi8>, vector<4x8xi8>142 143  // RHS test data144  %rhs_cst = arith.constant dense<[[-17, -50,  -1,  48, -13,  22,  39,  33],145                                   [-35, -24,  37, -32,  33,  30, -11, -17],146                                   [-28,  31,   3, -44, -15, -27,  22,  35],147                                   [-23,  39,  48,  26, -23,  32, -39, -38]]> : vector<4x8xi8>148 149  %rhs_mem = func.call @prepareRHSTestData(%rhs_cst) : (vector<4x8xi8>) -> memref<?xi8>150  %rhs_flat = vector.transfer_read %rhs_mem[%c0], %c0_i8 {in_bounds = [true]} :  memref<?xi8>, vector<[32]xi8>151  %rhs = vector.shape_cast %rhs_flat : vector<[32]xi8> to vector<[4]x8xi8>152 153  // Matrix multiplication and accumulate with transposed RHS.154  %0 = arith.extsi %lhs : vector<4x8xi8> to vector<4x8xi32>155  %1 = arith.extsi %rhs : vector<[4]x8xi8> to vector<[4]x8xi32>156  %2 = vector.contract {indexing_maps = #packed_maps,157                        iterator_types = ["parallel", "parallel", "reduction"],158                        kind = #vector.kind<add>} %0, %1, %acc159    : vector<4x8xi32>, vector<[4]x8xi32> into vector<4x[4]xi32>160 161  // Display the result of the multiplication162  vector.print str "Result(SMMLA):\n"163  %u0 = vector.extract %2[0] : vector<[4]xi32> from vector<4x[4]xi32>164  %u1 = vector.extract %2[1] : vector<[4]xi32> from vector<4x[4]xi32>165  %u2 = vector.extract %2[2] : vector<[4]xi32> from vector<4x[4]xi32>166  %u3 = vector.extract %2[3] : vector<[4]xi32> from vector<4x[4]xi32>167  vector.print %u0 : vector<[4]xi32>168  vector.print %u1 : vector<[4]xi32>169  vector.print %u2 : vector<[4]xi32>170  vector.print %u3 : vector<[4]xi32>171 172  // Deallocate the buffers.173  memref.dealloc %acc_mem : memref<4x?xi32>174  memref.dealloc %lhs_mem : memref<4x8xi8>175  memref.dealloc %rhs_mem : memref<?xi8>176 177  return178}179 180// Test the operation where both LHS and RHS are interpreted as unsigned, hence181// we ultimately emit and execute the `ummla` instruction.182 183// CHECK-IR-LABEL: llvm.func @test_ummla184// CHECK-IR-COUNT-4: arm_sve.intr.ummla185func.func @test_ummla() attributes {no_inline} {186 187  %c0 = arith.constant 0 : index188  %c0_i32 = arith.constant 0 : i32189  %c0_i8 = arith.constant 0 : i8190 191  // Accumulator test data192  %acc_cst = arith.constant dense<[[16, 16, 48, 40],193                                   [40, 24, 35, 12],194                                   [33, 24, 29, 19],195                                   [28, 13, 33, 18]]> : vector<4x4xi32>196 197  %acc_mem = func.call @prepareAccTestData(%acc_cst) : (vector<4x4xi32>) -> memref<4x?xi32>198  %acc = vector.transfer_read %acc_mem[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : memref<4x?xi32>, vector<4x[4]xi32>199 200  // LHS test data201  %lhs_cst = arith.constant dense<[[35, 42, 37, 49, 36, 36, 23, 33],202                                   [39, 34, 33, 45, 43, 10, 44, 47],203                                   [18, 35, 29, 25, 36, 33, 28, 29],204                                   [26, 49, 43, 32, 27, 16, 45, 33]]> : vector<4x8xi8>205 206  %lhs_mem = func.call @prepareLHSTestData(%lhs_cst) : (vector<4x8xi8>) -> memref<4x8xi8>207  %lhs = vector.transfer_read %lhs_mem[%c0, %c0], %c0_i8 {in_bounds = [true, true]} : memref<4x8xi8>, vector<4x8xi8>208 209  // RHS test data210  %rhs_cst = arith.constant dense<[[18, 31, 37, 35, 44, 22, 37, 28],211                                   [21, 22, 49, 39, 30, 28, 35, 37],212                                   [21, 47, 39, 35, 23, 43, 24, 49],213                                   [49, 49, 40, 32, 37, 20, 47, 40]]> : vector<4x8xi8>214 215  %rhs_mem = func.call @prepareRHSTestData(%rhs_cst) : (vector<4x8xi8>) -> memref<?xi8>216  %rhs_flat = vector.transfer_read %rhs_mem[%c0], %c0_i8 {in_bounds = [true]} :  memref<?xi8>, vector<[32]xi8>217  %rhs = vector.shape_cast %rhs_flat : vector<[32]xi8> to vector<[4]x8xi8>218 219  // Matrix multiplication and accumulate with transposed RHS.220  %0 = arith.extui %lhs : vector<4x8xi8> to vector<4x8xi32>221  %1 = arith.extui %rhs : vector<[4]x8xi8> to vector<[4]x8xi32>222  %2 = vector.contract {indexing_maps = #packed_maps,223                        iterator_types = ["parallel", "parallel", "reduction"],224                        kind = #vector.kind<add>} %0, %1, %acc225    : vector<4x8xi32>, vector<[4]x8xi32> into vector<4x[4]xi32>226 227  // Display the result of the multiplication228  vector.print str "Result(UMMLA):\n"229  %u0 = vector.extract %2[0] : vector<[4]xi32> from vector<4x[4]xi32>230  %u1 = vector.extract %2[1] : vector<[4]xi32> from vector<4x[4]xi32>231  %u2 = vector.extract %2[2] : vector<[4]xi32> from vector<4x[4]xi32>232  %u3 = vector.extract %2[3] : vector<[4]xi32> from vector<4x[4]xi32>233  vector.print %u0 : vector<[4]xi32>234  vector.print %u1 : vector<[4]xi32>235  vector.print %u2 : vector<[4]xi32>236  vector.print %u3 : vector<[4]xi32>237 238  // Deallocate the buffers.239  memref.dealloc %acc_mem : memref<4x?xi32>240  memref.dealloc %lhs_mem : memref<4x8xi8>241  memref.dealloc %rhs_mem : memref<?xi8>242 243  return244}245 246// Test the operation where LHS is interpreted as unsigned and RHS is247// interpreted as signed, hence we ultimately emit and execute the `usmmla`248// instruction.249 250// CHECK-IR-LABEL: llvm.func @test_usmmla251// CHECK-IR-COUNT-4: arm_sve.intr.usmmla252func.func @test_usmmla() attributes {no_inline} {253 254  %c0 = arith.constant 0 : index255  %c0_i32 = arith.constant 0 : i32256  %c0_i8 = arith.constant 0 : i8257 258  // Accumulator test data259  %acc_cst = arith.constant dense<[[-44,  20,  44, -46],260                                   [ -8,  25, -34,  26],261                                   [-20, -36,  -3,  39],262                                   [-48, -31, -25, -21]]> : vector<4x4xi32>263 264  %acc_mem = func.call @prepareAccTestData(%acc_cst) : (vector<4x4xi32>) -> memref<4x?xi32>265  %acc = vector.transfer_read %acc_mem[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : memref<4x?xi32>, vector<4x[4]xi32>266 267  // LHS test data268  %lhs_cst = arith.constant dense<[[153, 161,  24, 157, 211, 154,  52,  27],269                                   [168,  77, 136, 124, 249,  28,  13, 122],270                                   [ 97,  82, 181,  39,  53,  25,  80, 240],271                                   [184, 227, 106, 165, 126, 113, 121, 228]]> : vector<4x8xi8>272 273  %lhs_mem = func.call @prepareLHSTestData(%lhs_cst) : (vector<4x8xi8>) -> memref<4x8xi8>274  %lhs = vector.transfer_read %lhs_mem[%c0, %c0], %c0_i8 {in_bounds = [true, true]} : memref<4x8xi8>, vector<4x8xi8>275 276  // RHS test data277  %rhs_cst = arith.constant dense<[[ 40,  27,  37,  43,  38,  -6,  37,  49],278                                   [-17, -50,  -1,  48, -13,  22,  39,  33],279                                   [-35, -24,  37, -32,  33,  30, -11, -17],280                                   [-28,  31,   3, -44, -15, -27,  22,  35]]> : vector<4x8xi8>281 282  %rhs_mem = func.call @prepareRHSTestData(%rhs_cst) : (vector<4x8xi8>) -> memref<?xi8>283  %rhs_flat = vector.transfer_read %rhs_mem[%c0], %c0_i8 {in_bounds = [true]} :  memref<?xi8>, vector<[32]xi8>284  %rhs = vector.shape_cast %rhs_flat : vector<[32]xi8> to vector<[4]x8xi8>285 286  // Matrix multiplication and accumulate with transposed RHS.287  %0 = arith.extui %lhs : vector<4x8xi8> to vector<4x8xi32>288  %1 = arith.extsi %rhs : vector<[4]x8xi8> to vector<[4]x8xi32>289  %2 = vector.contract {indexing_maps = #packed_maps,290                        iterator_types = ["parallel", "parallel", "reduction"],291                        kind = #vector.kind<add>} %0, %1, %acc292    : vector<4x8xi32>, vector<[4]x8xi32> into vector<4x[4]xi32>293 294  // Display the result of the multiplication295  vector.print str "Result(USMMLA):\n"296  %u0 = vector.extract %2[0] : vector<[4]xi32> from vector<4x[4]xi32>297  %u1 = vector.extract %2[1] : vector<[4]xi32> from vector<4x[4]xi32>298  %u2 = vector.extract %2[2] : vector<[4]xi32> from vector<4x[4]xi32>299  %u3 = vector.extract %2[3] : vector<[4]xi32> from vector<4x[4]xi32>300  vector.print %u0 : vector<[4]xi32>301  vector.print %u1 : vector<[4]xi32>302  vector.print %u2 : vector<[4]xi32>303  vector.print %u3 : vector<[4]xi32>304 305  // Deallocate the buffers.306  memref.dealloc %acc_mem : memref<4x?xi32>307  memref.dealloc %lhs_mem : memref<4x8xi8>308  memref.dealloc %rhs_mem : memref<?xi8>309 310  return311}312 313// Test the operation where LHS is interpreted as signed and RHS is interpreted314// as unsigned. In this test we ultimately emit end execute the `usmmla`315// instruction with reversed operands, see `LowerContractToSVEPatterns.cpp`316// for more details.317 318// CHECK-IR-LABEL: llvm.func @test_summla319// CHECK-IR-COUNT-4: arm_sve.intr.usmmla320func.func @test_summla() attributes {no_inline} {321 322  %c0 = arith.constant 0 : index323  %c0_i32 = arith.constant 0 : i32324  %c0_i8 = arith.constant 0 : i8325 326  // Accumulator test data327  %acc_cst = arith.constant dense<[[-44,  20,  44, -46],328                                   [ -8,  25, -34,  26],329                                   [-20, -36,  -3,  39],330                                   [-48, -31, -25, -21]]> : vector<4x4xi32>331 332  %acc_mem = func.call @prepareAccTestData(%acc_cst) : (vector<4x4xi32>) -> memref<4x?xi32>333  %acc = vector.transfer_read %acc_mem[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : memref<4x?xi32>, vector<4x[4]xi32>334 335  // LHS test data336  %lhs_cst = arith.constant dense<[[-35, -27, -36, -31,  23, -34,  -8, -33],337                                   [-20,  17, -32, -47,  37,  22,  -7, -21],338                                   [ -7, -35,  20,  -4,  39,  46, -23,  40],339                                   [ 40,  27,  37,  43,  38,  -6,  37,  49]]> : vector<4x8xi8>340 341  %lhs_mem = func.call @prepareLHSTestData(%lhs_cst) : (vector<4x8xi8>) -> memref<4x8xi8>342  %lhs = vector.transfer_read %lhs_mem[%c0, %c0], %c0_i8 {in_bounds = [true, true]} : memref<4x8xi8>, vector<4x8xi8>343 344  // RHS test data345  %rhs_cst = arith.constant dense<[[125, 171, 138, 187, 108, 175,  82,  99],346                                   [221,  25, 164,  97, 156, 221, 218, 177],347                                   [171, 160, 219, 191, 144,  45, 161, 210],348                                   [223, 165, 123,  99, 108,  86,  37,  92]]> : vector<4x8xi8>349 350  %rhs_mem = func.call @prepareRHSTestData(%rhs_cst) : (vector<4x8xi8>) -> memref<?xi8>351  %rhs_flat = vector.transfer_read %rhs_mem[%c0], %c0_i8 {in_bounds = [true]} :  memref<?xi8>, vector<[32]xi8>352  %rhs = vector.shape_cast %rhs_flat : vector<[32]xi8> to vector<[4]x8xi8>353 354  // Matrix multiplication and accumulate with transposed RHS.355  %0 = arith.extsi %lhs : vector<4x8xi8> to vector<4x8xi32>356  %1 = arith.extui %rhs : vector<[4]x8xi8> to vector<[4]x8xi32>357  %2 = vector.contract {indexing_maps = #packed_maps,358                        iterator_types = ["parallel", "parallel", "reduction"],359                        kind = #vector.kind<add>} %0, %1, %acc360    : vector<4x8xi32>, vector<[4]x8xi32> into vector<4x[4]xi32>361 362  // Display the result of the multiplication363  vector.print str "Result(SUMMLA (i.e. USMMLA transposed)):\n"364  %u0 = vector.extract %2[0] : vector<[4]xi32> from vector<4x[4]xi32>365  %u1 = vector.extract %2[1] : vector<[4]xi32> from vector<4x[4]xi32>366  %u2 = vector.extract %2[2] : vector<[4]xi32> from vector<4x[4]xi32>367  %u3 = vector.extract %2[3] : vector<[4]xi32> from vector<4x[4]xi32>368  vector.print %u0 : vector<[4]xi32>369  vector.print %u1 : vector<[4]xi32>370  vector.print %u2 : vector<[4]xi32>371  vector.print %u3 : vector<[4]xi32>372 373  // Deallocate the buffers.374  memref.dealloc %acc_mem : memref<4x?xi32>375  memref.dealloc %lhs_mem : memref<4x8xi8>376  memref.dealloc %rhs_mem : memref<?xi8>377 378  return379}380 381// Perform each test with SVE vector lengths 128 bits and 256 bits (i.e. VSCALEs382// 1 and 2, respectively). The vector length is set via the `setArmVLBits`383// function. The effect of setting a different vector length is that the tests384// allocate and operate on different sized buffers (see `prepare<X>TestData`385// functions).386 387func.func @main() {388  %c128 = arith.constant 128 : i32389  %c256 = arith.constant 256 : i32390 391// CHECK-LABEL: Result(SMMLA):392// CHECK: ( -1999,  1941,   685, -2879 )393// CHECK: ( -3705,  2952,   987,  -685 )394// CHECK: (  2565,  4157, -1589,  -357 )395// CHECK: (  2383, -2252,    32, -1365 )396  func.call @setArmVLBits(%c128) : (i32) -> ()397  func.call @test_smmla() : () -> ()398 399// CHECK: Result(SMMLA):400// CHECK: ( -1999,  1941,   685, -2879, -1999,  1941,   685, -2879 )401// CHECK: ( -3705,  2952,   987,  -685, -3705,  2952,   987,  -685 )402// CHECK: (  2565,  4157, -1589,  -357,  2565,  4157, -1589,  -357 )403// CHECK: (  2383, -2252,    32, -1365,  2383, -2252,    32, -1365 )404  func.call @setArmVLBits(%c256) : (i32) -> ()405  func.call @test_smmla() : () -> ()406 407// CHECK-LABEL: Result(UMMLA):408// CHECK: ( 9183, 9513, 10460, 11314 )409// CHECK: ( 9648, 9812, 10092, 12088 )410// CHECK: ( 7548, 7625,  8398,  9044 )411// CHECK: ( 8855, 9046,  9685, 11191 )412  func.call @setArmVLBits(%c128) : (i32) -> ()413  func.call @test_ummla() : () -> ()414 415// CHECK: Result(UMMLA):416// CHECK: ( 9183, 9513, 10460, 11314, 9183, 9513, 10460, 11314 )417// CHECK: ( 9648, 9812, 10092, 12088, 9648, 9812, 10092, 12088 )418// CHECK: ( 7548, 7625,  8398,  9044, 7548, 7625,  8398,  9044 )419// CHECK: ( 8855, 9046,  9685, 11191, 8855, 9046,  9685, 11191 )420  func.call @setArmVLBits(%c256) : (i32) -> ()421  func.call @test_ummla() : () -> ()422 423// CHECK-LABEL: Result(USMMLA):424// CHECK: ( 28403,  445,  -2759, -11409 )425// CHECK: ( 34908, 1047,    142,  -7274 )426// CHECK: ( 31032, 6807,  -2378,   7382 )427// CHECK: ( 44217, 6396, -10930,    623 )428  func.call @setArmVLBits(%c128) : (i32) -> ()429  func.call @test_usmmla() : () -> ()430 431// CHECK: Result(USMMLA):432// CHECK: ( 28403,  445,  -2759, -11409, 28403,  445,  -2759, -11409 )433// CHECK: ( 34908, 1047,    142,  -7274, 34908, 1047,    142,  -7274 )434// CHECK: ( 31032, 6807,  -2378,   7382, 31032, 6807,  -2378,   7382 )435// CHECK: ( 44217, 6396, -10930,    623, 44217, 6396, -10930,    623 )436  func.call @setArmVLBits(%c256) : (i32) -> ()437  func.call @test_usmmla() : () -> ()438 439// CHECK-LABEL: Result(SUMMLA (i.e. USMMLA transposed)):440// CHECK: ( -27190, -28812, -30502, -23575 )441// CHECK: (  -7613,  -8386, -15938,  -6521 )442// CHECK: (   9468,  18750,   9199,   5764 )443// CHECK: (  33655,  41064,  48900,  31627 )444  func.call @setArmVLBits(%c128) : (i32) -> ()445  func.call @test_summla() : () -> ()446 447// CHECK: Result(SUMMLA (i.e. USMMLA transposed)):448// CHECK: ( -27190, -28812, -30502, -23575, -27190, -28812, -30502, -23575 )449// CHECK: (  -7613,  -8386, -15938,  -6521,  -7613,  -8386, -15938,  -6521 )450// CHECK: (   9468,  18750,   9199,   5764,   9468,  18750,   9199,   5764 )451// CHECK: (  33655,  41064,  48900,  31627,  33655,  41064,  48900,  31627 )452  func.call @setArmVLBits(%c256) : (i32) -> ()453  func.call @test_summla() : () -> ()454 455  return456}457 458func.func private @setArmVLBits(%bits : i32)459func.func private @printMemrefI32(%ptr : memref<*xi32>)460