brintos

brintos / llvm-project-archived public Read only

0
0
Text · 21.8 KiB · 3ae1883 Raw
494 lines · plain
1// RUN: mlir-opt %s --transform-interpreter --split-input-file | FileCheck %s2 3// TODO: Review the usage of `in_bounds` and remove where not affecting the4// generated output.5 6///  CHECK: #[[$MAP:.*]] = affine_map<(d0, d1, d2, d3) -> (d1, 0, d3)>7 8///----------------------------------------------------------------------------------------9/// [Pattern: TransferWritePermutationLowering]10///11/// IN: vector.transfer_write (_transposed_ minor identity permutation map)12/// OUT: vector.transpose + vector.transfer_write (minor identity permutation map)13///----------------------------------------------------------------------------------------14 15// CHECK-LABEL:   func.func @xfer_write_minor_identity_transposed16// CHECK-SAME:      %[[VEC:.*]]: vector<4x8xi16>,17// CHECK-SAME:      %[[MEM:.*]]: memref<2x2x8x4xi16>18// CHECK:           %[[TR:.*]] = vector.transpose %[[VEC]], [1, 0] : vector<4x8xi16> to vector<8x4xi16>19// CHECK:           vector.transfer_write20// CHECK-NOT:       permutation_map21// CHECK-SAME:      %[[TR]], %[[MEM]]{{.*}} {in_bounds = [true, true]} : vector<8x4xi16>, memref<2x2x8x4xi16>22func.func @xfer_write_minor_identity_transposed(23    %vec: vector<4x8xi16>,24    %mem: memref<2x2x8x4xi16>,25    %idx: index) {26 27  vector.transfer_write %vec, %mem[%idx, %idx, %idx, %idx] {28    in_bounds = [true, true],29    permutation_map = affine_map<(d0, d1, d2, d3) -> (d3, d2)>30  } : vector<4x8xi16>, memref<2x2x8x4xi16>31 32  return33}34 35// Even with out-of-bounds accesses, it is safe to apply this pattern36 37// CHECK-LABEL:   func.func @xfer_write_minor_identity_transposed_out_of_bounds38// CHECK-SAME:      %[[VEC:.*]]: vector<4x8xi16>,39// CHECK-SAME:      %[[MEM:.*]]: memref<2x2x?x?xi16>,40// CHECK-SAME:      %[[IDX:.*]]: index) {41// CHECK:           %[[TR:.*]] = vector.transpose %[[VEC]], [1, 0] : vector<4x8xi16> to vector<8x4xi16>42// Expect the in_bounds attribute to be preserved. Since we don't print it when43// all flags are "false", it should not appear in the output.44// CHECK-NOT:       in_bounds45// CHECK:           vector.transfer_write46// CHECK-NOT:       permutation_map47// CHECK-SAME:      %[[TR]], %[[MEM]][%[[IDX]], %[[IDX]], %[[IDX]], %[[IDX]]] : vector<8x4xi16>, memref<2x2x?x?xi16>48func.func @xfer_write_minor_identity_transposed_out_of_bounds(49    %vec: vector<4x8xi16>,50    %mem: memref<2x2x?x?xi16>,51    %idx: index) {52 53  vector.transfer_write %vec, %mem[%idx, %idx, %idx, %idx] {54    in_bounds = [false, false],55    permutation_map = affine_map<(d0, d1, d2, d3) -> (d3, d2)>56  } : vector<4x8xi16>, memref<2x2x?x?xi16>57 58  return59}60 61// CHECK-LABEL:   func.func @xfer_write_minor_identity_transposed_with_mask_scalable62// CHECK-SAME:      %[[VEC:.*]]: vector<4x[8]xi16>,63// CHECK-SAME:      %[[MEM:.*]]: memref<2x2x?x4xi16>,64// CHECK-SAME:      %[[MASK:.*]]: vector<[8]x4xi1>65// CHECK:           %[[TR:.*]] = vector.transpose %[[VEC]], [1, 0] : vector<4x[8]xi16> to vector<[8]x4xi16>66// CHECK:           vector.transfer_write67// CHECK-NOT:       permutation_map68// CHECK-SAME:      %[[TR]], %[[MEM]]{{.*}}, %[[MASK]] {in_bounds = [true, true]} : vector<[8]x4xi16>, memref<2x2x?x4xi16>69func.func @xfer_write_minor_identity_transposed_with_mask_scalable(70    %vec: vector<4x[8]xi16>,71    %mem: memref<2x2x?x4xi16>,72    %mask: vector<[8]x4xi1>,73    %idx: index) {74 75  %c0 = arith.constant 0 : index76  vector.transfer_write %vec, %mem[%idx, %idx, %idx, %idx], %mask {77    in_bounds = [true, true],78    permutation_map = affine_map<(d0, d1, d2, d3) -> (d3, d2)>79  } : vector<4x[8]xi16>, memref<2x2x?x4xi16>80 81  return82}83 84// Masked version is not supported85 86// CHECK-LABEL:   func.func @xfer_write_minor_identity_transposed_map_masked87// CHECK-NOT: vector.transpose88func.func @xfer_write_minor_identity_transposed_map_masked(89    %vec: vector<4x8xi16>,90    %mem: memref<2x2x8x4xi16>,91    %mask: vector<8x4xi1>,92    %idx: index) {93 94  %c0 = arith.constant 0 : index95  vector.mask %mask {96    vector.transfer_write %vec, %mem[%idx, %idx, %idx, %idx] {97      in_bounds = [true, true],98      permutation_map = affine_map<(d0, d1, d2, d3) -> (d3, d2)>99    } : vector<4x8xi16>, memref<2x2x8x4xi16>100  } : vector<8x4xi1>101 102  return103}104 105///----------------------------------------------------------------------------------------106/// [Patterns: TransferWriteNonPermutationLowering + TransferWritePermutationLowering]107///108/// IN: vector.transfer_write109///         (neither a minor identity nor transposed minor identity map)110/// OUT 1: vector.broadcast + vector.transfer_write111///         (transposed minor identity)112/// OUT 2: vector.transfer_write -> vector.broadcast + vector.transpose113///        + vector.transfer_write (minor identity)114///----------------------------------------------------------------------------------------115 116// CHECK-LABEL:   func.func @xfer_write_non_minor_identity(117// CHECK-SAME:      %[[MEM:.*]]: memref<?x?xf32>,118// CHECK-SAME:      %[[VEC:.*]]: vector<7xf32>,119// CHECK-SAME:      %[[IDX_1:.*]]: index, %[[IDX_2:.*]]: index) {120// CHECK:           %[[BC:.*]] = vector.broadcast %[[VEC]] : vector<7xf32> to vector<1x7xf32>121// CHECK:           %[[TR:.*]] = vector.transpose %[[BC]], [1, 0] : vector<1x7xf32> to vector<7x1xf32>122// CHECK:           vector.transfer_write %[[TR]], %[[MEM]]{{\[}}%[[IDX_1]], %[[IDX_2]]] {in_bounds = [false, true]} : vector<7x1xf32>, memref<?x?xf32>123func.func @xfer_write_non_minor_identity(124    %mem : memref<?x?xf32>,125    %vec : vector<7xf32>,126    %idx_1 : index,127    %idx_2 : index) {128 129  vector.transfer_write %vec, %mem[%idx_1, %idx_2] {130    permutation_map = affine_map<(d0, d1) -> (d0)>131  } : vector<7xf32>, memref<?x?xf32>132 133  return134}135 136// Even with out-of-bounds accesses, it is safe to apply this pattern137 138// CHECK-LABEL:   func.func @xfer_write_non_minor_identity_with_mask_out_of_bounds(139// CHECK-SAME:      %[[MEM:.*]]: memref<?x?xf32>,140// CHECK-SAME:      %[[VEC:.*]]: vector<7xf32>,141// CHECK-SAME:      %[[IDX_1:.*]]: index, %[[IDX_2:.*]]: index,142// CHECK-SAME:      %[[MASK:.*]]: vector<7xi1>) {143// CHECK:           %[[BC_VEC:.*]] = vector.broadcast %[[VEC]] : vector<7xf32> to vector<1x7xf32>144// CHECK:           %[[BC_MASK:.*]] = vector.broadcast %[[MASK]] : vector<7xi1> to vector<1x7xi1>145// CHECK:           %[[TR_MASK:.*]] = vector.transpose %[[BC_MASK]], [1, 0] : vector<1x7xi1> to vector<7x1xi1>146// CHECK:           %[[TR_VEC:.*]] = vector.transpose %[[BC_VEC]], [1, 0] : vector<1x7xf32> to vector<7x1xf32>147// CHECK:           vector.transfer_write %[[TR_VEC]], %[[MEM]]{{\[}}%[[IDX_1]], %[[IDX_2]]], %[[TR_MASK]] {in_bounds = [false, true]} : vector<7x1xf32>, memref<?x?xf32>148func.func @xfer_write_non_minor_identity_with_mask_out_of_bounds(149    %mem : memref<?x?xf32>,150    %vec : vector<7xf32>,151    %idx_1 : index,152    %idx_2 : index,153    %mask : vector<7xi1>) {154 155  vector.transfer_write %vec, %mem[%idx_1, %idx_2], %mask {156    permutation_map = affine_map<(d0, d1) -> (d0)>,157    in_bounds = [false]158  } : vector<7xf32>, memref<?x?xf32>159 160  return161}162 163// CHECK-LABEL:     func.func @xfer_write_non_minor_identity_with_mask_scalable(164// CHECK-SAME:        %[[VEC:.*]]: vector<4x[8]xi16>,165// CHECK-SAME:        %[[MEM:.*]]: memref<1x4x?x1xi16>,166// CHECK-SAME:        %[[MASK:.*]]: vector<4x[8]xi1>167// CHECK:             %[[BC_1:.*]] = vector.broadcast %[[VEC]] : vector<4x[8]xi16> to vector<1x4x[8]xi16>168// CHECK:             %[[BC_2:.*]] = vector.broadcast %[[MASK]] : vector<4x[8]xi1> to vector<1x4x[8]xi1>169// CHECK:             %[[TRANSPOSE_1:.*]] =  vector.transpose %[[BC_2]], [1, 2, 0] : vector<1x4x[8]xi1> to vector<4x[8]x1xi1>170// CHECK:             %[[TRANSPOSE_2:.*]] =  vector.transpose %[[BC_1]], [1, 2, 0] : vector<1x4x[8]xi16> to vector<4x[8]x1xi16>171// CHECK:             vector.transfer_write %[[TRANSPOSE_2]], %[[MEM]]{{.*}}, %[[TRANSPOSE_1]] {in_bounds = [true, true, true]} : vector<4x[8]x1xi16>, memref<1x4x?x1xi16>172func.func @xfer_write_non_minor_identity_with_mask_scalable(173    %vec: vector<4x[8]xi16>,174    %mem: memref<1x4x?x1xi16>,175    %mask: vector<4x[8]xi1>,176    %idx: index){177 178  vector.transfer_write %vec, %mem[%idx, %idx, %idx, %idx], %mask {179    in_bounds = [true, true],180    permutation_map = affine_map<(d0, d1, d2, d3) -> (d1, d2)>181  } : vector<4x[8]xi16>, memref<1x4x?x1xi16>182 183  return184}185 186// Masked version is not supported187 188// CHECK-LABEL: func @xfer_write_non_minor_identity_masked189//  CHECK-SAME:   %[[DEST:.*]]: tensor<?x?xf32>,190//  CHECK-SAME:   %[[VEC:.*]]: vector<16xf32>,191//  CHECK-SAME:   %[[IDX:.*]]: index,192//  CHECK-SAME:   %[[MASK:.*]]: vector<16xi1>193//   CHECK-NOT:   vector.transpose194//       CHECK:   vector.mask %[[MASK]] { vector.transfer_write %[[VEC]], %[[DEST]]{{.*}} vector<16xf32>, tensor<?x?xf32> } : vector<16xi1> -> tensor<?x?xf32>195func.func @xfer_write_non_minor_identity_masked(196    %dest: tensor<?x?xf32>,197    %vec: vector<16xf32>,198    %idx: index,199    %mask: vector<16xi1>) -> tensor<?x?xf32> {200 201  %res = vector.mask %mask {202    vector.transfer_write %vec, %dest[%idx, %idx] {203      permutation_map = affine_map<(d0, d1) -> (d0)>204    } : vector<16xf32>, tensor<?x?xf32>205  } : vector<16xi1> -> tensor<?x?xf32>206 207  return %res : tensor<?x?xf32>208}209 210// CHECK-LABEL: func.func @xfer_write_non_minor_identity_masked_scalable211//  CHECK-SAME:   %[[VEC:.*]]: vector<4x[8]xi16>,212//  CHECK-SAME:   %[[DEST:.*]]: tensor<?x?x?x?xf32>,213//  CHECK-SAME:   %[[MASK:.*]]: vector<4x[8]xi1>214//  CHECK-SAME:   -> tensor<?x?x?x?xf32> {215//   CHECK-NOT:   vector.transpose216//       CHECK:   vector.mask %[[MASK]] { vector.transfer_write %[[VEC]], %[[DEST]]{{.*}} : vector<4x[8]xi16>, tensor<?x?x?x?xf32> } : vector<4x[8]xi1> -> tensor<?x?x?x?xf32>217func.func @xfer_write_non_minor_identity_masked_scalable(218    %vec: vector<4x[8]xi16>,219    %dest: tensor<?x?x?x?xf32>,220    %mask:  vector<4x[8]xi1>,221    %idx: index) -> tensor<?x?x?x?xf32> {222 223  %c0 = arith.constant 0 : index224  %res = vector.mask %mask {225    vector.transfer_write %vec, %dest[%idx, %idx, %idx, %idx] {226      in_bounds = [true, true],227      permutation_map = affine_map<(d0, d1, d2, d3) -> (d1, d2)>228    } : vector<4x[8]xi16>, tensor<?x?x?x?xf32>229  } : vector<4x[8]xi1> -> tensor<?x?x?x?xf32>230 231  return %res : tensor<?x?x?x?xf32>232}233 234// Masked version is not supported235 236// CHECK-LABEL: func @xfer_write_non_minor_identity_masked_2237//  CHECK-SAME:   %[[DEST:.*]]: tensor<?x?x?x?xf32>238//  CHECK-SAME:   %[[VEC:.*]]: vector<14x8x16xf32>239//  CHECK-SAME:   %[[MASK:.*]]: vector<14x8x16xi1>240//  CHECK-SAME:   %[[DIM:.*]]: index241//   CHECK-NOT:   vector.broadcast242//       CHECK:   vector.mask %[[MASK]] { vector.transfer_write %[[VEC]], %[[DEST]]{{.*}} : vector<14x8x16xf32>, tensor<?x?x?x?xf32> } : vector<14x8x16xi1> -> tensor<?x?x?x?xf32>243func.func @xfer_write_non_minor_identity_masked_2(244    %dest : tensor<?x?x?x?xf32>,245    %vec : vector<14x8x16xf32>,246    %mask:  vector<14x8x16xi1>,247    %idx: index) -> tensor<?x?x?x?xf32> {248 249  %res = vector.mask %mask {250    vector.transfer_write %vec, %dest[%idx, %idx, %idx, %idx] {251      in_bounds = [false, false, true],252      permutation_map = affine_map<(d0, d1, d2, d3) -> (d0, d1, d3)>253    } : vector<14x8x16xf32>, tensor<?x?x?x?xf32>254  } : vector<14x8x16xi1> -> tensor<?x?x?x?xf32>255 256  return %res : tensor<?x?x?x?xf32>257}258 259///----------------------------------------------------------------------------------------260/// [Pattern: TransferOpReduceRank (for leading 0 dim) + 261///           TransferReadPermutationLowering (for transposed minor identity)]262///263/// IN: vector.transfer_read264///     (_transposed_ minor identity permutation map, with 0 or more broadcast dims)265/// OUT: vector.transfer_read + vector.broadcast + vector.transpose266///     (minor identity permutation map with 0 or more leading broadcast dims)267///----------------------------------------------------------------------------------------268/// TODO: Inner broadcast dim - see also the block at the bottom of this file269 270// CHECK-LABEL:   func.func @xfer_read_minor_identity_transposed_with_mask271// CHECK-SAME:      %[[MEM:.*]]: memref<?x?xf32>,272// CHECK-SAME:      %[[MASK:.*]]: vector<2x4xi1>273// CHECK-SAME:      %[[IDX:.*]]: index274// CHECK:           %[[PASS_THROUGH:.*]] = arith.constant 0.000000e+00 : f32275// CHECK:           %[[T_READ:.*]] = vector.transfer_read %[[MEM]]{{\[}}%[[IDX]], %[[IDX]]], %[[PASS_THROUGH]], %[[MASK]] {in_bounds = [true, true]} : memref<?x?xf32>, vector<2x4xf32>276// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[T_READ]] : vector<2x4xf32> to vector<8x2x4xf32>277// CHECK:           %[[TRANSPOSE:.*]] = vector.transpose %[[BCAST]], [0, 2, 1] : vector<8x2x4xf32> to vector<8x4x2xf32>278// CHECK:           return %[[TRANSPOSE]] : vector<8x4x2xf32>279func.func @xfer_read_minor_identity_transposed_with_mask(280    %mem: memref<?x?xf32>,281    %mask: vector<2x4xi1>,282    %idx: index) -> (vector<8x4x2xf32>) {283 284  %pad = arith.constant 0.000000e+00 : f32285 286  %res = vector.transfer_read %mem[%idx, %idx], %pad, %mask {287    in_bounds = [true, true, true],288    permutation_map = affine_map<(d0, d1) -> (0, d1, d0)>289  } : memref<?x?xf32>, vector<8x4x2xf32>290 291  return %res : vector<8x4x2xf32>292}293 294// CHECK-LABEL:   func.func @xfer_read_minor_identity_transposed_with_mask_scalable(295// CHECK-SAME:      %[[MEM:.*]]: memref<?x?xf32>,296// CHECK-SAME:      %[[MASK:.*]]: vector<2x[4]xi1>297// CHECK-SAME:      %[[IDX:.*]]: index298// CHECK:           %[[PAD:.*]] = arith.constant 0.000000e+00 : f32299// CHECK:           %[[T_READ:.*]] = vector.transfer_read %[[MEM]]{{\[}}%[[IDX]], %[[IDX]]], %[[PAD]], %[[MASK]] {in_bounds = [true, true]} : memref<?x?xf32>, vector<2x[4]xf32>300// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[T_READ]] : vector<2x[4]xf32> to vector<8x2x[4]xf32>301// CHECK:           %[[TRANSPOSE:.*]] = vector.transpose %[[BCAST]], [0, 2, 1] : vector<8x2x[4]xf32> to vector<8x[4]x2xf32>302// CHECK:           return %[[TRANSPOSE]] : vector<8x[4]x2xf32>303func.func @xfer_read_minor_identity_transposed_with_mask_scalable(304    %mem: memref<?x?xf32>,305    %mask: vector<2x[4]xi1>,306    %idx: index) -> (vector<8x[4]x2xf32>) {307 308  %pad = arith.constant 0.000000e+00 : f32309 310  %res = vector.transfer_read %mem[%idx, %idx], %pad, %mask {311    in_bounds = [true, true, true],312    permutation_map = affine_map<(d0, d1) -> (0, d1, d0)>313  } : memref<?x?xf32>, vector<8x[4]x2xf32>314 315  return %res : vector<8x[4]x2xf32>316}317 318// Masked version is not supported319 320// CHECK-LABEL: func @xfer_read_minor_identity_transposed_masked(321//  CHECK-SAME:   %[[DEST:.*]]: tensor<?x?xf32>,322//  CHECK-SAME:   %[[MASK:.*]]: vector<2x4xi1>323//  CHECK-SAME:   %[[IDX:.*]]: index324//   CHECK-NOT:   vector.transpose325//       CHECK:   vector.mask %[[MASK]] { vector.transfer_read %[[DEST]]{{.*}}: tensor<?x?xf32>, vector<8x4x2xf32> } : vector<2x4xi1> -> vector<8x4x2xf32>326func.func @xfer_read_minor_identity_transposed_masked(327    %dest: tensor<?x?xf32>,328    %mask: vector<2x4xi1>,329    %idx: index) -> (vector<8x4x2xf32>) {330 331  %pad = arith.constant 0.000000e+00 : f32332 333  %res = vector.mask %mask {334    vector.transfer_read %dest[%idx, %idx], %pad {335    in_bounds = [true, true, true],336    permutation_map = affine_map<(d0, d1) -> (0, d1, d0)>337    } : tensor<?x?xf32>, vector<8x4x2xf32>338  } : vector<2x4xi1> -> vector<8x4x2xf32>339 340  return %res : vector<8x4x2xf32>341}342 343// CHECK-LABEL:  func.func @xfer_read_minor_identity_transposed_masked_scalable(344//  CHECK-SAME:    %[[DEST:.*]]: tensor<?x?xf32>,345//  CHECK-SAME:    %[[MASK:.*]]: vector<2x[4]xi1>346//   CHECK-NOT:    vector.transpose347//       CHECK:    %[[T_READ:.*]] = vector.mask %[[MASK]] { vector.transfer_read %[[DEST]]{{.*}} : tensor<?x?xf32>, vector<8x[4]x2xf32> } : vector<2x[4]xi1> -> vector<8x[4]x2xf32>348func.func @xfer_read_minor_identity_transposed_masked_scalable(349  %dest: tensor<?x?xf32>,350  %mask: vector<2x[4]xi1>,351  %idx: index) -> vector<8x[4]x2xf32> {352 353  %pad = arith.constant 0.000000e+00 : f32354 355  %res = vector.mask %mask {356    vector.transfer_read %dest[%idx, %idx], %pad {357      in_bounds = [true, true, true],358      permutation_map = affine_map<(d0, d1) -> (0, d1, d0)>359    } : tensor<?x?xf32>, vector<8x[4]x2xf32>360  } :vector<2x[4]xi1> -> vector<8x[4]x2xf32>361 362  return %res : vector<8x[4]x2xf32>363}364 365///----------------------------------------------------------------------------------------366/// [Pattern: TransferOpReduceRank]367///368/// IN: vector.transfer_read (minor identity map + broadcast)369/// OUT: vector.transfer_read + vector.broadcast370///----------------------------------------------------------------------------------------371 372// CHECK-LABEL:   func.func @xfer_read_minor_identitiy_bcast_dims373//  CHECK-SAME:     %[[MEM:.*]]: memref<?x?x?x?xf32>, %[[IDX:.*]]: index) -> vector<8x4x2x3xf32> {374//       CHECK:     %[[T_READ:.*]] = vector.transfer_read %[[MEM]][%[[IDX]], %[[IDX]], %[[IDX]], %[[IDX]]]{{.*}} permutation_map = #[[$MAP]]} : memref<?x?x?x?xf32>, vector<4x2x3xf32>375//       CHECK:     %[[BC:.*]] = vector.broadcast %[[T_READ]] : vector<4x2x3xf32> to vector<8x4x2x3xf32>376//       CHECK:     return %[[BC]] : vector<8x4x2x3xf32>377func.func @xfer_read_minor_identitiy_bcast_dims(378    %mem: memref<?x?x?x?xf32>,379    %idx: index) -> vector<8x4x2x3xf32> {380 381  %pad = arith.constant 0.000000e+00 : f32382 383  %res = vector.transfer_read %mem[%idx, %idx, %idx, %idx], %pad {384    in_bounds = [true, true, true, true],385    permutation_map = affine_map<(d0, d1, d2, d3) -> (0, d1, 0, d3)>386  } : memref<?x?x?x?xf32>, vector<8x4x2x3xf32>387 388  return %res : vector<8x4x2x3xf32>389}390 391// CHECK-LABEL:   func.func @xfer_read_minor_identitiy_bcast_dims_scalable392//  CHECK-SAME:     %[[MEM:.*]]: memref<?x?x?x?xf32>, %[[IDX:.*]]: index) -> vector<8x[4]x2x3xf32> {393//       CHECK:     %[[T_READ:.*]] = vector.transfer_read %[[MEM]][%[[IDX]], %[[IDX]], %[[IDX]], %[[IDX]]]{{.*}} permutation_map = #[[$MAP]]} : memref<?x?x?x?xf32>, vector<[4]x2x3xf32>394//       CHECK:     %[[BC:.*]] = vector.broadcast %[[T_READ]] : vector<[4]x2x3xf32> to vector<8x[4]x2x3xf32>395//       CHECK:     return %[[BC]] : vector<8x[4]x2x3xf32>396func.func @xfer_read_minor_identitiy_bcast_dims_scalable(397    %mem: memref<?x?x?x?xf32>,398    %idx: index) -> vector<8x[4]x2x3xf32> {399 400  %pad = arith.constant 0.000000e+00 : f32401 402  %res = vector.transfer_read %mem[%idx, %idx, %idx, %idx], %pad {403    in_bounds = [true, true, true, true],404    permutation_map = affine_map<(d0, d1, d2, d3) -> (0, d1, 0, d3)>405  } : memref<?x?x?x?xf32>, vector<8x[4]x2x3xf32>406 407  return %res : vector<8x[4]x2x3xf32>408}409 410// CHECK-LABEL:   func.func @xfer_read_minor_identitiy_bcast_dims_with_mask411//  CHECK-SAME:     %[[MEM:.*]]: memref<?x?x?x?xf32>412//  CHECK-SAME:     %[[MASK:.*]]: vector<4x3xi1>413//  CHECK-SAME:     %[[IDX:.*]]: index) -> vector<8x4x2x3xf32>414//       CHECK:     %[[PASS_THROUGH:.*]] = arith.constant 0.000000e+00 : f32415//       CHECK:     %[[T_READ:.*]] = vector.transfer_read %[[MEM]][%[[IDX]], %[[IDX]], %[[IDX]], %[[IDX]]], %[[PASS_THROUGH]], %[[MASK]]{{.*}} permutation_map = #[[$MAP]]} : memref<?x?x?x?xf32>, vector<4x2x3xf32>416//       CHECK:     %[[BC:.*]] = vector.broadcast %[[T_READ]] : vector<4x2x3xf32> to vector<8x4x2x3xf32>417//       CHECK:     return %[[BC]] : vector<8x4x2x3xf32>418func.func @xfer_read_minor_identitiy_bcast_dims_with_mask(419    %mem: memref<?x?x?x?xf32>,420    %mask: vector<4x3xi1>,421    %idx: index) -> vector<8x4x2x3xf32> {422 423  %pad = arith.constant 0.000000e+00 : f32424 425  %res = vector.transfer_read %mem[%idx, %idx, %idx, %idx], %pad, %mask {426    in_bounds = [true, true, true, true],427    permutation_map = affine_map<(d0, d1, d2, d3) -> (0, d1, 0, d3)>428  } : memref<?x?x?x?xf32>, vector<8x4x2x3xf32>429 430  return %res : vector<8x4x2x3xf32>431}432 433// CHECK-LABEL:   func.func @xfer_read_minor_identitiy_bcast_dims_with_mask_scalable434//  CHECK-SAME:     %[[MEM:.*]]: memref<?x?x?x?xf32>435//  CHECK-SAME:     %[[MASK:.*]]: vector<[4]x3xi1>436//  CHECK-SAME:     %[[IDX:.*]]: index) -> vector<8x[4]x2x3xf32>437//       CHECK:     %[[PASS_THROUGH:.*]] = arith.constant 0.000000e+00 : f32438//       CHECK:     %[[T_READ:.*]] = vector.transfer_read %[[MEM]][%[[IDX]], %[[IDX]], %[[IDX]], %[[IDX]]], %[[PASS_THROUGH]], %[[MASK]]{{.*}} permutation_map = #[[$MAP]]} : memref<?x?x?x?xf32>, vector<[4]x2x3xf32>439//       CHECK:     %[[BC:.*]] = vector.broadcast %[[T_READ]] : vector<[4]x2x3xf32> to vector<8x[4]x2x3xf32>440//       CHECK:     return %[[BC]] : vector<8x[4]x2x3xf32>441func.func @xfer_read_minor_identitiy_bcast_dims_with_mask_scalable(442    %mem: memref<?x?x?x?xf32>,443    %mask: vector<[4]x3xi1>,444    %idx: index) -> vector<8x[4]x2x3xf32> {445 446  %pad = arith.constant 0.000000e+00 : f32447 448  %res = vector.transfer_read %mem[%idx, %idx, %idx, %idx], %pad, %mask {449    in_bounds = [true, true, true, true],450    permutation_map = affine_map<(d0, d1, d2, d3) -> (0, d1, 0, d3)>451  } : memref<?x?x?x?xf32>, vector<8x[4]x2x3xf32>452 453  return %res : vector<8x[4]x2x3xf32>454}455 456// Masked version is not supported457 458// CHECK-LABEL:   func.func @xfer_read_minor_identitiy_bcast_dims_masked459//  CHECK-SAME:     %[[MEM:.*]]: memref<?x?x?x?xf32>,460//  CHECK-SAME:     %[[MASK:.*]]: vector<4x3xi1>461//  CHECK-SAME:     %[[IDX:.*]]: index) -> vector<8x4x2x3xf32> {462//   CHECK-NOT:     vector.broadcast463//       CHECK:     vector.mask %[[MASK]] { vector.transfer_read %[[MEM]]{{.*}} : memref<?x?x?x?xf32>, vector<8x4x2x3xf32> } : vector<4x3xi1> -> vector<8x4x2x3xf32>464func.func @xfer_read_minor_identitiy_bcast_dims_masked(465    %mem: memref<?x?x?x?xf32>,466    %mask: vector<4x3xi1>,467    %idx: index) -> vector<8x4x2x3xf32> {468 469  %pad = arith.constant 0.000000e+00 : f32470 471  %res = vector.mask %mask {472    vector.transfer_read %mem[%idx, %idx, %idx, %idx], %pad {473      in_bounds = [true, true, true, true],474      permutation_map = affine_map<(d0, d1, d2, d3) -> (0, d1, 0, d3)>475    } : memref<?x?x?x?xf32>, vector<8x4x2x3xf32>476  } : vector<4x3xi1> -> vector<8x4x2x3xf32>477 478  return %res : vector<8x4x2x3xf32>479}480 481///----------------------------------------------------------------------------------------482//  TD sequence483///----------------------------------------------------------------------------------------484module attributes {transform.with_named_sequence} {485  transform.named_sequence @__transform_main(%module_op: !transform.any_op {transform.readonly}) {486    %f = transform.structured.match ops{["func.func"]} in %module_op487      : (!transform.any_op) -> !transform.any_op488    transform.apply_patterns to %f {489      transform.apply_patterns.vector.transfer_permutation_patterns490    } : !transform.any_op491    transform.yield492  }493}494