brintos

brintos / llvm-project-archived public Read only

0
0
Text · 39.1 KiB · 574b365 Raw
603 lines · plain
1// RUN: mlir-opt --xegpu-wg-to-sg-distribute -split-input-file %s | FileCheck %s2 3gpu.module @test_distribution {4  // CHECK-LABEL: create_nd_tdesc_no_offset5  // CHECK-SAME: %[[ARG_0:.*]]: memref<256x128xf32>6  gpu.func @create_nd_tdesc_no_offset(%src: memref<256x128xf32>) {7    // CHECK: xegpu.create_nd_tdesc %[[ARG_0]] : memref<256x128xf32>8    // CHECK-SAME: -> !xegpu.tensor_desc<32x32xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>9    %tdesc = xegpu.create_nd_tdesc %src : memref<256x128xf32>10        -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>>11      gpu.return12  }13 14  // CHECK-LABEL: create_nd_tdesc_with_ptr15  // CHECK-SAME: %[[ARG_0:.*]]: ui6416  gpu.func @create_nd_tdesc_with_ptr(%src: ui64, %w : index, %h : index, %x : index, %y : index) {17    // CHECK: xegpu.create_nd_tdesc %[[ARG_0]], shape : [{{.*}}, {{.*}}], strides : [{{.*}}, {{.*}}] : ui6418    // CHECK-SAME: -> !xegpu.tensor_desc<32x32xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>19    %c1 = arith.constant 1 : index20    %tdesc = xegpu.create_nd_tdesc %src, shape:[%h, %w], strides: [%w, %c1] : ui6421        -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>>22      gpu.return23  }24 25  // CHECK-LABEL: load_nd_tdesc_with_offset26  gpu.func @load_nd_tdesc_with_offset(%src: memref<256x128xf32>) {27    //CHECK: %[[TDESC:.*]] = xegpu.create_nd_tdesc %{{.*}} : memref<256x128xf32> -> !xegpu.tensor_desc<32x32xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>28    //CHECK-DAG: %[[SGID:.*]] = gpu.subgroup_id : index29    //CHECK-DAG: %[[C4:.*]] = arith.constant 4 : index30    //CHECK-DAG: %[[SGIDX:.*]] = index.remu %[[SGID]], %[[C4]]31    //CHECK-DAG: %[[SGIDY_TMP:.*]] = index.divu %[[SGID]], %[[C4]]32    //CHECK-DAG: %[[C8:.*]] = arith.constant 8 : index33    //CHECK-DAG: %[[SGIDY:.*]] = index.remu %[[SGIDY_TMP]], %[[C8]]34    //CHECK-DAG: %[[C32:.*]] = arith.constant 32 : index35    //CHECK-DAG: %[[L_OFF_Y:.*]] = index.mul %[[SGIDY]], %[[C32]]36    //CHECK-DAG: %[[L_OFF_X:.*]] = index.mul %[[SGIDX]], %[[C32]]37    //CHECK-DAG: %[[C256:.*]] = arith.constant 256 : index38    //CHECK-DAG: %[[OFF_Y:.*]] = index.remu %[[L_OFF_Y]], %[[C256]]39    //CHECK-DAG: %[[C128:.*]] = arith.constant 128 : index40    //CHECK-DAG: %[[OFF_X:.*]] = index.remu %[[L_OFF_X]], %[[C128]]41    //CHECK-DAG: %[[LOAD:.*]] = xegpu.load_nd %[[TDESC]][{{%.*}}, {{%.*}}] : !xegpu.tensor_desc<32x32xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>> -> vector<32x32xf32>42    %tdesc = xegpu.create_nd_tdesc %src : memref<256x128xf32>43      -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>>44    %load =  xegpu.load_nd %tdesc[0, 0]45      : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>>46      -> vector<256x128xf32>47    gpu.return48  }49 50  // CHECK-LABEL: store_nd_with_offsets51  // CHECK-SAME: %[[ARG_0:.*]]: memref<256x128xf32>52  gpu.func @store_nd_with_offsets(%src: memref<256x128xf32>) {53    //CHECK: xegpu.store_nd %{{.*}}, {{%.*}}[{{%.*}}, {{%.*}}]  : vector<32x32xf32>, !xegpu.tensor_desc<32x32xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>54    %tdesc = xegpu.create_nd_tdesc %src: memref<256x128xf32>55      -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>>56    %load =  xegpu.load_nd %tdesc[0, 0]57      : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>>58      -> vector<256x128xf32>59    xegpu.store_nd %load, %tdesc[0, 0]60      : vector<256x128xf32>, !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>>61    gpu.return62}63 64  // CHECK-LABEL: prefetch_nd_tdesc_with_offset65  // CHECK-SAME: %[[ARG_0:.*]]: memref<256x128xf32>66  gpu.func @prefetch_nd_tdesc_with_offset(%src: memref<256x128xf32>) {67    //CHECK: xegpu.prefetch_nd %{{.*}}[{{%.*}}, {{%.*}}] : !xegpu.tensor_desc<32x32xf32, #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>>68    %cst0 = arith.constant 0 : index69    %tdesc = xegpu.create_nd_tdesc %src : memref<256x128xf32>70      -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>>71    xegpu.prefetch_nd %tdesc[%cst0, %cst0]72      : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>>73    gpu.return74  }75 76  // CHECK-LABEL: dpas77  gpu.func @dpas(%a: memref<128x128xf16>, %b: memref<128x128xf16>) {78    // CHECK: %[[DPAS:.*]] = xegpu.dpas %{{.*}}, %{{.*}} {layout_result_0 =  #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>} : vector<16x128xf16>, vector<128x16xf16> -> vector<16x16xf32>79    %tdesc_a = xegpu.create_nd_tdesc %a : memref<128x128xf16>80      -> !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 128], lane_layout = [1, 16], lane_data = [1, 1]>>81    %load_a =  xegpu.load_nd %tdesc_a[0, 0]82      : !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 128], lane_layout = [1, 16], lane_data = [1, 1]>>83      -> vector<128x128xf16>84    %tdesc_b = xegpu.create_nd_tdesc %b : memref<128x128xf16>85      -> !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [128, 16], lane_layout = [1, 16], lane_data = [2, 1]>>86    %load_b =  xegpu.load_nd %tdesc_b[0, 0]87      : !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [128, 16], lane_layout = [1, 16], lane_data = [2, 1]>>88      -> vector<128x128xf16>89    %dpas = xegpu.dpas %load_a, %load_b90      {layout_result_0 =  #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>}91      : vector<128x128xf16>, vector<128x128xf16> -> vector<128x128xf32>92    gpu.return93  }94 95  // CHECK-LABEL: dpas_no_sg_data96  gpu.func @dpas_no_sg_data(%a: memref<128x128xf16>, %b: memref<128x128xf16>) {97    // CHECK: %[[DPAS:.*]] = xegpu.dpas %{{.*}}, %{{.*}} {layout_result_0 =  #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1], order = [1, 0]>} : vector<16x16xf16>, vector<16x16xf16> -> vector<16x16xf32>98    %tdesc_a = xegpu.create_nd_tdesc %a : memref<128x128xf16>99      -> !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], lane_layout = [1, 16], lane_data = [1, 1],100      order = [1, 0]>>101    %load_a =  xegpu.load_nd %tdesc_a[0, 0]102      : !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], lane_layout = [1, 16], lane_data = [1, 1],103      order = [1, 0]>>104      -> vector<128x128xf16>105    %tdesc_b = xegpu.create_nd_tdesc %b : memref<128x128xf16>106      -> !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], lane_layout = [1, 16], lane_data = [2, 1],107      order = [1, 0]>>108    %load_b =  xegpu.load_nd %tdesc_b[0, 0]109      : !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], lane_layout = [1, 16], lane_data = [2, 1],110      order = [1, 0]>>111      -> vector<128x128xf16>112    %dpas = xegpu.dpas %load_a, %load_b113      {layout_result_0 =  #xegpu.layout<sg_layout = [8, 8], lane_layout = [1, 16], lane_data = [1, 1], order = [1, 0]>}114      : vector<128x128xf16>, vector<128x128xf16> -> vector<128x128xf32>115    gpu.return116  }117 118  // CHECK-LABEL: dpas_with_no_create_nd_desc119  gpu.func @dpas_with_no_create_nd_desc(%a: vector<256x128xf32>, %b: vector<128x256xf32>) {120    // CHECK-NOT: vector<32x32xf32>121    %dpas = xegpu.dpas %a, %b122      {layout =  #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 16], lane_layout = [1, 16], lane_data = [1, 1]>}123      : vector<256x128xf32>, vector<128x256xf32> -> vector<256x256xf32>124    gpu.return125  }126 127  // CHECK-LABEL: broadcast_dim1128  // CHECK-SAME: %[[ARG_0:.*]]: memref<256x1xf32>129  gpu.func @broadcast_dim1(%src: memref<256x1xf32>) {130    %tdesc = xegpu.create_nd_tdesc %src : memref<256x1xf32>131      -> !xegpu.tensor_desc<256x1xf32, #xegpu.layout<sg_layout = [8, 1], sg_data = [32, 1], lane_layout = [8, 1], lane_data = [1, 1]>>132    %load =  xegpu.load_nd %tdesc[0, 0]133      : !xegpu.tensor_desc<256x1xf32, #xegpu.layout<sg_layout = [8, 1], sg_data = [32, 1], lane_layout = [8, 1], lane_data = [1, 1]>>134      -> vector<256x1xf32>135    // CHECK: vector.broadcast {{.*}} {layout_result_0 = #xegpu.layout<lane_layout = [8, 1], lane_data = [1, 1]>}136    // CHECK-SAME: : vector<32x1xf32> to vector<32x32xf32>137    %broadcast = vector.broadcast %load138      {layout_result_0 = #xegpu.layout<sg_layout = [8, 1], sg_data = [32, 32], lane_layout = [8, 1], lane_data = [1, 1]>}139      : vector<256x1xf32> to vector<256x32xf32>140    gpu.return141  }142 143  // CHECK-LABEL: broadcast_dim0144  // CHECK-SAME: %[[ARG_0:.*]]: memref<1x128xf32>145  gpu.func @broadcast_dim0(%src: memref<1x128xf32>) {146    %tdesc = xegpu.create_nd_tdesc %src : memref<1x128xf32>147      -> !xegpu.tensor_desc<1x128xf32, #xegpu.layout<sg_layout = [1, 4], sg_data = [1, 32], lane_layout = [1, 16], lane_data = [1, 1]>>148    %load =  xegpu.load_nd %tdesc[0, 0]149      : !xegpu.tensor_desc<1x128xf32, #xegpu.layout<sg_layout = [1, 4], sg_data = [1, 32], lane_layout = [1, 16], lane_data = [1, 1]>>150      -> vector<1x128xf32>151    // CHECK: vector.broadcast {{.*}} {layout_result_0 = #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1]>}152    // CHECK-SAME: : vector<1x32xf32> to vector<32x32xf32>153    %broadcast = vector.broadcast %load154      {layout_result_0 = #xegpu.layout<sg_layout = [1, 4], sg_data = [32, 32], lane_layout = [1, 16], lane_data = [1, 1]>}155      : vector<1x128xf32> to vector<32x128xf32>156    gpu.return157  }158 159  // CHECK-LABEL: gemm_with_load_store_offset160  // CHECK-SAME: %[[ARG_0:.*]]: memref<1024x1024xf16>, %[[ARG_1:.*]]: memref<1024x1024xf16>, %[[ARG_2:.*]]: memref<1024x1024xf32>161  gpu.func @gemm_with_load_store_offset(%arg0: memref<1024x1024xf16>, %arg1: memref<1024x1024xf16>, %arg2: memref<1024x1024xf32>) {162    //CHECK: [[c0:%.+]] = arith.constant 0 : index163    //CHECK: [[c128:%.+]] = arith.constant 128 : index164    //CHECK: [[c1024:%.+]] = arith.constant 1024 : index165    %c0 = arith.constant 0 : index166    %c128 = arith.constant 128 : index167    %c1024 = arith.constant 1024 : index168    %block_id_x = gpu.block_id  x169    %block_id_y = gpu.block_id  y170    %0 = arith.muli %block_id_x, %c128 : index171    %1 = arith.muli %block_id_y, %c128 : index172    %2 = xegpu.create_nd_tdesc %arg2 : memref<1024x1024xf32> -> !xegpu.tensor_desc<128x128xf32, #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 16]>>173    // CHECK: [[DESC_A:%.+]] = xegpu.create_nd_tdesc %[[ARG_0]] : memref<1024x1024xf16> -> !xegpu.tensor_desc<16x128xf16>174    // CHECK: [[DESC_B:%.+]] = xegpu.create_nd_tdesc %[[ARG_1]] : memref<1024x1024xf16> -> !xegpu.tensor_desc<128x16xf16>175    %3 = xegpu.create_nd_tdesc %arg0 : memref<1024x1024xf16> -> !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 128]>>176    %4 = xegpu.create_nd_tdesc %arg1 : memref<1024x1024xf16> -> !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [128, 16]>>177    // load_nd with offset178    %5 = xegpu.load_nd %2[%0, %1] : !xegpu.tensor_desc<128x128xf32, #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 16]>> -> vector<128x128xf32>179    %6 = xegpu.load_nd %3[%0, %c0] : !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 128]>> -> vector<128x128xf16>180    %7 = xegpu.load_nd %4[%c0, %1] : !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [128, 16]>> -> vector<128x128xf16>181    // scf.for loop182    //      CHECK: [[scf:%.+]]:3 = scf.for [[arg3:%.+]] = [[c0]] to [[c1024]] step [[c128]]183    // CHECK-SAME: iter_args([[arg4:%.+]] = {{.*}}, [[arg5:%.+]] = {{.*}}, [[arg6:%.+]] = {{.*}}) ->184    // CHECK-SAME: (vector<16x128xf16>, vector<128x16xf16>, vector<16x16xf32>)185    //      CHECK: [[c:%.+]] = xegpu.dpas [[arg4]], [[arg5]], [[arg6]] : vector<16x128xf16>, vector<128x16xf16>, vector<16x16xf32> -> vector<16x16xf32>186    //      CHECK: [[a:%.+]] = xegpu.load_nd [[DESC_A]][{{%.*}}, {{%.*}}]  : !xegpu.tensor_desc<16x128xf16> -> vector<16x128xf16>187    //      CHECK: [[b:%.+]] = xegpu.load_nd [[DESC_B]][{{%.*}}, {{%.*}}]  : !xegpu.tensor_desc<128x16xf16> -> vector<128x16xf16>188    //      CHECK: scf.yield [[a]], [[b]], [[c]] : vector<16x128xf16>, vector<128x16xf16>, vector<16x16xf32>189    %8:3 = scf.for %arg3 = %c0 to %c1024 step %c128 iter_args(%arg4 = %6, %arg5 = %7, %arg6 = %5)190        -> (vector<128x128xf16>, vector<128x128xf16>, vector<128x128xf32>) {191      // load_nd with offset inside loop192      %9 = xegpu.dpas %arg4, %arg5, %arg6 {layout_result_0 = #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 16]>}193                          : vector<128x128xf16>, vector<128x128xf16>, vector<128x128xf32> -> vector<128x128xf32>194      %10 = xegpu.load_nd %3[%arg3, %c0] : !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 128]>> -> vector<128x128xf16>195      %11 = xegpu.load_nd %4[%c0, %arg3] : !xegpu.tensor_desc<128x128xf16, #xegpu.layout<sg_layout = [8, 8], sg_data = [128, 16]>> -> vector<128x128xf16>196      scf.yield %10, %11, %9 : vector<128x128xf16>, vector<128x128xf16>, vector<128x128xf32>197    }198    // store_nd with offset199    xegpu.store_nd %8#2, %2[%0, %1] : vector<128x128xf32>, !xegpu.tensor_desc<128x128xf32, #xegpu.layout<sg_layout = [8, 8], sg_data = [16, 16]>>200    gpu.return201  }202 203  // CHECK-LABEL: @subgroup_id_range204  gpu.func @subgroup_id_range(%src: memref<256x128xf32>, %src1: memref<128x256xf32>, %src2: memref<128x64xf32>) {205    %sg_id = gpu.subgroup_id : index206    %c0 = arith.constant 0 : index207    %c1 = arith.constant 1 : index208    %c2 = arith.constant 2 : index209    %c31 = arith.constant 31 : index210    %c3 = arith.constant 3 : index211    %cond1 = arith.cmpi sge, %sg_id, %c0 : index212    %cond2 = arith.cmpi slt, %sg_id, %c1 : index213    %cond = arith.andi %cond1, %cond2 : i1214    scf.if %cond {215        // CHECK-NOT: index.sub216        %tdesc = xegpu.create_nd_tdesc %src : memref<256x128xf32>217          -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [8, 4], lane_data = [1, 1]>>218        %load =  xegpu.load_nd %tdesc[0, 0]219          : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [8, 4], lane_data = [1, 1]>>220          -> vector<256x128xf32>221    } {sg_id_range = #xegpu.range<[0, 32]>}222    %cond3 = arith.cmpi sge, %sg_id, %c2 : index223    %cond4 = arith.cmpi slt, %sg_id, %c31 : index224    %cond5 = arith.andi %cond3, %cond4 : i1225    scf.if %cond5 {226      // CHECK: %[[SGID:.*]] = gpu.subgroup_id : index227      // CHECK: %[[C2:.*]] = arith.constant 2 : index228      // CHECK: %[[SUB:.*]] = index.sub %{{.*}}, %[[C2]]229      %tdesc = xegpu.create_nd_tdesc %src2 : memref<128x64xf32>230        -> !xegpu.tensor_desc<128x64xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [32, 16], lane_layout = [8, 4], lane_data = [1, 1]>>231      %load =  xegpu.load_nd %tdesc[0, 0]232        : !xegpu.tensor_desc<128x64xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [32, 16], lane_layout = [8, 4], lane_data = [1, 1]>>233        -> vector<128x64xf32>234      %exp = math.exp %load {layout_result_0 = #xegpu.layout<sg_layout = [4, 4], sg_data = [32, 16], lane_layout = [8, 4], lane_data = [1, 1]>} : vector<128x64xf32>235    }{sg_id_range = #xegpu.range<[2, 18]>}236    gpu.return237  }238 239  // CHECK-LABEL: @subgroup_id_range_nested_if240  gpu.func @subgroup_id_range_nested_if(%src: memref<256x128xf32>, %src1: memref<128x64xf32>) {241    %sg_id = gpu.subgroup_id : index242    %c1 = arith.constant 1 : i1243    %c3 = arith.constant 3 : index244    %c32 = arith.constant 32 : index245    %tdesc = xegpu.create_nd_tdesc %src : memref<256x128xf32>246      -> !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [8, 4], lane_data = [1, 1]>>247    %load =  xegpu.load_nd %tdesc[0, 0]248      : !xegpu.tensor_desc<256x128xf32, #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32], lane_layout = [8, 4], lane_data = [1, 1]>>249      -> vector<256x128xf32>250    %cond1 = arith.cmpi sge, %sg_id, %c3 : index251    %cond2 = arith.cmpi slt, %sg_id, %c32 : index252    %cond = arith.andi %cond1, %cond2 : i1253    scf.if %c1 {254      scf.if %cond {255        // CHECK: %[[SGID:.*]] = gpu.subgroup_id : index256        // CHECK: %[[C3:.*]] = arith.constant 3 : index257        // CHECK: %[[SUB:.*]] = index.sub %{{.*}}, %[[C3]]258        %td = xegpu.create_nd_tdesc %src1 : memref<128x64xf32>259          -> !xegpu.tensor_desc<128x64xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [32, 16], lane_layout = [8, 4], lane_data = [1, 1]>>260        %ld =  xegpu.load_nd %td[0, 0]261          : !xegpu.tensor_desc<128x64xf32, #xegpu.layout<sg_layout = [4, 4], sg_data = [32, 16], lane_layout = [8, 4], lane_data = [1, 1]>>262          -> vector<128x64xf32>263        %exp = math.exp %ld {layout_result_0 = #xegpu.layout<sg_layout = [4, 4], sg_data = [32, 16], lane_layout = [8, 4], lane_data = [1, 1]>} : vector<128x64xf32>264    }265  } {sg_id_range = #xegpu.range<[3, 19]>}266  gpu.return267  }268 269  // CHECK-LABEL: @load_gather270  // CHECK-SAME: %[[ARG0:.*]]: memref<?xf16>271  gpu.func @load_gather(%src : memref<?xf16>) {272    // CHECK: %[[CST:.*]] = arith.constant dense<0> : vector<32x4xindex>273    // CHECK: %[[MASK:.*]] = arith.constant dense<true> : vector<32x4xi1>274    // CHECK: %[[LOAD:.*]] = xegpu.load %[[ARG0]][%[[CST]]], %[[MASK]] <{chunk_size = 1 : i64, l1_hint = #xegpu.cache_hint<cached>}>275    // CHECK-SAME: : memref<?xf16>, vector<32x4xindex>, vector<32x4xi1> -> vector<32x4xf16>276    %offset =  arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 4]>} dense<0> : vector<256x16xindex>277    %mask = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 4]>} dense<1> : vector<256x16xi1>278    %load = xegpu.load %src[%offset], %mask {chunk_size = 1, layout_result_0 = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 4]>, l1_hint = #xegpu.cache_hint<cached>}279      : memref<?xf16>, vector<256x16xindex>, vector<256x16xi1> -> vector<256x16xf16>280    gpu.return281  }282 283  // CHECK-LABEL: @store_scatter284  // CHECK-SAME: %[[ARG0:.*]]: memref<256xf16>285  gpu.func @store_scatter(%dest : memref<256xf16>) {286    // CHECK: %[[VAL:.*]] = arith.constant {layout_result_0 = #xegpu.layout<inst_data = [8]>} dense<2.550000e+01> : vector<8xf16>287    // CHECK: %[[CST:.*]] = arith.constant {layout_result_0 = #xegpu.layout<inst_data = [8]>} dense<0> : vector<8xindex>288    // CHECK: %[[MASK:.*]] = arith.constant {layout_result_0 = #xegpu.layout<inst_data = [8]>} dense<true> : vector<8xi1>289    // CHECK: xegpu.store %[[VAL]], %[[ARG0]][%[[CST]]], %[[MASK]] <{chunk_size = 1 : i64, l1_hint = #xegpu.cache_hint<cached>, layout = #xegpu.layout<inst_data = [8]>}>290    // CHECK-SAME: {layout_operand_0 = #xegpu.layout<inst_data = [8]>, layout_operand_2 = #xegpu.layout<inst_data = [8]>,291    // CHECK-SAME: layout_operand_3 = #xegpu.layout<inst_data = [8]>}292    // CHECK-SAME: : vector<8xf16>, memref<256xf16>, vector<8xindex>, vector<8xi1>293    %val = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>} dense<25.5> : vector<256xf16>294    %offset = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>} dense<0> : vector<256xindex>295    %mask = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>} dense<1> : vector<256xi1>296    xegpu.store %val, %dest[%offset], %mask {chunk_size = 1, layout_operand_0 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>, 297                                             layout_operand_2 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>,298                                             layout_operand_3 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>,299                                             l1_hint = #xegpu.cache_hint<cached>}300      : vector<256xf16>, memref<256xf16>, vector<256xindex>, vector<256xi1>301    gpu.return302  }303 304  // CHECK-LABEL: @load_with_non_unit_chunk_size305  // CHECK-SAME: %[[ARG0:.*]]: memref<?xf16>306  gpu.func @load_with_non_unit_chunk_size(%src : memref<?xf16>) {307    // CHECK: %[[CST:.*]] = arith.constant dense<0> : vector<8xindex>308    // CHECK: %[[MASK:.*]] = arith.constant dense<true> : vector<8xi1>309    // CHECK: %[[LOAD:.*]] = xegpu.load %[[ARG0]][%[[CST]]], %[[MASK]] <{chunk_size = 4 : i64, l1_hint = #xegpu.cache_hint<cached>}>310    // CHECK-SAME: : memref<?xf16>, vector<8xindex>, vector<8xi1> -> vector<8x4xf16>311    %offset =  arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8]>} dense<0> : vector<256xindex>312    %mask = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8]>} dense<1> : vector<256xi1>313    %load = xegpu.load %src[%offset], %mask {chunk_size = 4, layout_result_0 = #xegpu.layout<sg_layout = [32, 1], sg_data = [8, 4]>, l1_hint = #xegpu.cache_hint<cached>}314      : memref<?xf16>, vector<256xindex>, vector<256xi1> -> vector<256x4xf16>315    gpu.return316  }317 318  // CHECK-LABEL: distribute_load_matrix319  // CHECK-SAME: [[arg0:%.+]]: memref<32768xi8, 3>320  gpu.func @distribute_load_matrix(%arg0: memref<32768xi8, 3>) {321    //CHECK: [[mdesc:%.+]] = xegpu.create_mem_desc [[arg0]] : memref<32768xi8, 3> -> !xegpu.mem_desc<64x128xf32>322    //CHECK: [[sgid:%.+]] = gpu.subgroup_id : index323    //CHECK: [[c4:%.+]] = arith.constant 4 : index324    //CHECK: [[sgidx:%.+]] = index.remu [[sgid]], [[c4]]325    //CHECK: [[sgidy_tmp:%.+]] = index.divu [[sgid]], [[c4]]326    //CHECK: [[c2:%.+]] = arith.constant 2 : index327    //CHECK: [[sgidy:%.+]] = index.remu [[sgidy_tmp]], [[c2]]328    //CHECK: [[c32:%.+]] = arith.constant 32 : index329    //CHECK: [[l_off_y:%.+]] = index.mul [[sgidy]], [[c32]]330    //CHECK: [[c32_0:%.+]] = arith.constant 32 : index331    //CHECK: [[l_off_x:%.+]] = index.mul [[sgidx]], [[c32_0]]332    //CHECK: [[c64:%.+]] = arith.constant 64 : index333    //CHECK: [[off_y:%.+]] = index.remu [[l_off_y]], [[c64]]334    //CHECK: [[c128:%.+]] = arith.constant 128 : index335    //CHECK: [[off_x:%.+]] = index.remu [[l_off_x]], [[c128]]336    //CHECK: xegpu.load_matrix [[mdesc]][[[off_y]], [[off_x]]] <{layout = #xegpu.layout<lane_layout = [2, 8], lane_data = [1, 1]>}>: !xegpu.mem_desc<64x128xf32>, index, index -> vector<32x32xf32>337    %0 = xegpu.create_mem_desc %arg0 : memref<32768xi8, 3> -> !xegpu.mem_desc<64x128xf32>338    %1 = xegpu.load_matrix %0[0, 0] <{layout = #xegpu.layout<sg_layout = [2, 4], sg_data = [32, 32], lane_layout = [2, 8], lane_data = [1, 1]>}>: !xegpu.mem_desc<64x128xf32> -> vector<64x128xf32>339    gpu.return340  }341 342  //CHECK-LABEL: distribute_store_matrix343  //CHECK-SAME: [[arg0:%.+]]: memref<32768xi8, 3>344  gpu.func @distribute_store_matrix(%arg0 : memref<32768xi8, 3>) {345    //CHECK: [[cst:%.+]] = arith.constant dense<1.000000e+00> : vector<32x32xf32>346    //CHECK: [[mdesc:%.+]] = xegpu.create_mem_desc [[arg0]] : memref<32768xi8, 3> -> !xegpu.mem_desc<64x128xf32>347    //CHECK: [[sgid:%.+]] = gpu.subgroup_id : index348    //CHECK: [[c4:%.+]] = arith.constant 4 : index349    //CHECK: [[sgidx:%.+]] = index.remu [[sgid]], [[c4]]350    //CHECK: [[sgidy_tmp:%.+]] = index.divu [[sgid]], [[c4]]351    //CHECK: [[c2:%.+]] = arith.constant 2 : index352    //CHECK: [[sgidy:%.+]] = index.remu [[sgidy_tmp]], [[c2]]353    //CHECK: [[c32:%.+]] = arith.constant 32 : index354    //CHECK: [[l_off_y:%.+]] = index.mul [[sgidy]], [[c32]]355    //CHECK: [[c32_0:%.+]] = arith.constant 32 : index356    //CHECK: [[l_off_x:%.+]] = index.mul [[sgidx]], [[c32_0]]357    //CHECK: [[c64:%.+]] = arith.constant 64 : index358    //CHECK: [[off_y:%.+]] = index.remu [[l_off_y]], [[c64]]359    //CHECK: [[c128:%.+]] = arith.constant 128 : index360    //CHECK: [[off_x:%.+]] = index.remu [[l_off_x]], [[c128]]361    //CHECK: xegpu.store_matrix [[cst]], [[mdesc]][[[off_y]], [[off_x]]] : vector<32x32xf32>, !xegpu.mem_desc<64x128xf32>, index, index362    %cst = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [2, 4], sg_data = [32, 32]>} dense<1.0> : vector<64x128xf32>363    %mdesc = xegpu.create_mem_desc %arg0 : memref<32768xi8, 3> -> !xegpu.mem_desc<64x128xf32>364    xegpu.store_matrix %cst, %mdesc[0, 0] {layout = #xegpu.layout<sg_layout = [2, 4], sg_data = [32, 32]>} : vector<64x128xf32>, !xegpu.mem_desc<64x128xf32>365    gpu.return366  }367 368  // CHECK-LABEL: @vector_reduce_dim_0369  gpu.func @vector_reduce_dim_0(%src: memref<4x128xf32>) {370    %cst = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [1, 32], sg_data = [4, 4]>, dims = [0]>} dense<1.0> : vector<128xf32>371    %tdesc = xegpu.create_nd_tdesc %src : memref<4x128xf32>372      -> !xegpu.tensor_desc<4x128xf32, #xegpu.layout<sg_layout = [1, 32], sg_data = [4, 4]>>373    %load =  xegpu.load_nd %tdesc[0, 0]374      : !xegpu.tensor_desc<4x128xf32, #xegpu.layout<sg_layout = [1, 32], sg_data = [4, 4]>>375      -> vector<4x128xf32>376    // CHECK: vector.multi_reduction <add>, {{.*}}, {{.*}} [0] : vector<4x4xf32> to vector<4xf32>377    %reduce = vector.multi_reduction <add>, %load, %cst {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [1, 32], sg_data = [4, 4]>, dims = [0]>} [0]378      : vector<4x128xf32> to vector<128xf32>379    gpu.return380  }381 382  // CHECK-LABEL: @vector_reduce_dim_1383  gpu.func @vector_reduce_dim_1(%src: memref<256x64xf32>) {384    %cst = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [16, 1], sg_data = [16, 64]>, dims = [1]>} dense<1.0> : vector<256xf32>385    %tdesc = xegpu.create_nd_tdesc %src : memref<256x64xf32>386      -> !xegpu.tensor_desc<256x64xf32, #xegpu.layout<sg_layout = [16, 1], sg_data = [16, 64]>>387    %load =  xegpu.load_nd %tdesc[0, 0]388      : !xegpu.tensor_desc<256x64xf32, #xegpu.layout<sg_layout = [16, 1], sg_data = [16, 64]>>389      -> vector<256x64xf32>390    // CHECK: vector.multi_reduction <add>, {{.*}}, {{.*}} [1] : vector<16x64xf32> to vector<16xf32>391    %reduce = vector.multi_reduction <add>, %load, %cst {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [16, 1], sg_data = [16, 64]>, dims = [1]>} [1]392      : vector<256x64xf32> to vector<256xf32>393    gpu.return394  }395 396  // CHECK-LABEL: @vector_reduce_4D397   gpu.func @vector_reduce_4D(%src: ui64) {398      %cst_acc = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [4, 2, 6, 1], sg_data = [1, 1, 1, 32]>, dims = [3]>} dense<0.0> : vector<4x2x6xf16>399      %offset = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [4, 2, 6, 1], sg_data = [1, 1, 1, 32]>} dense<0>  : vector<4x2x6x32xindex>400      %mask = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [4, 2, 6, 1], sg_data = [1, 1, 1, 32]>} dense<true> : vector<4x2x6x32xi1>401      %load = xegpu.load %src[%offset], %mask  {layout_result_0 = #xegpu.layout<sg_layout = [4, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : ui64, vector<4x2x6x32xindex>, vector<4x2x6x32xi1> -> vector<4x2x6x32xf16>402      // CHECK: vector.multi_reduction <add>, {{.*}}, {{.*}} [3] : vector<1x1x1x32xf16> to vector<1x1x1xf16>403      %reduce = vector.multi_reduction <add>, %load, %cst_acc {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [4, 2, 6, 1], sg_data = [1, 1, 1, 32]>, dims = [3]>} [3]404      : vector<4x2x6x32xf16> to vector<4x2x6xf16>405      gpu.return406    }407 408  // CHECK-LABEL: vector_step_op409  gpu.func @vector_step_op_slice_attr() {410    //CHECK: [[sgId:%.+]] = gpu.subgroup_id : index411    //CHECK: [[c8:%.+]] = arith.constant 8 : index412    //CHECK: [[sgidx:%.+]] = index.remu [[sgId]], [[c8]]413    //CHECK: [[sgidy_tmp:%.+]] = index.divu [[sgId]], [[c8]]414    //CHECK: [[c4:%.+]] = arith.constant 4 : index415    //CHECK: [[sgidy:%.+]] = index.remu [[sgidy_tmp]], [[c4]]416    //CHECK: [[c32:%.+]] = arith.constant 32 : index417    //CHECK: [[LY:%.+]] = index.mul [[sgidy]], [[c32]]418    //CHECK: [[c128:%.+]] = arith.constant 128 : index419    //CHECK: [[MODY:%.+]] = index.remu [[LY]], [[c128]]420    //CHECK: [[BASE:%.+]] = vector.step : vector<32xindex>421    //CHECK: [[CAST:%.+]] = vector.broadcast [[MODY]] : index to vector<32xindex>422    //CHECK: [[ADD:%.+]] = arith.addi [[BASE]], [[CAST]] : vector<32xindex>423    %step = vector.step {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [4, 8], sg_data = [32, 32]>, dims = [1]>}: vector<128xindex>424    gpu.return425  }426 427  gpu.func @vector_step_op_layout_attr() {428    //CHECK: [[sgId:%.+]] = gpu.subgroup_id : index429    //CHECK: [[c16:%.+]] = arith.constant 16 : index430    //CHECK: [[sgidx:%.+]] = index.remu [[sgId]], [[c16]]431    //CHECK: [[c8:%.+]] = arith.constant 8 : index432    //CHECK: [[LOCALY:%.+]] = index.mul [[sgidx]], [[c8]]433    //CHECK: [[c128:%.+]] = arith.constant 128 : index434    //CHECK: [[MODY:%.+]] = index.remu [[LOCALY]], [[c128]]435    //CHECK: [[BASE:%.+]] = vector.step : vector<8xindex>436    //CHECK: [[CAST:%.+]] = vector.broadcast [[MODY]] : index to vector<8xindex>437    //CHECK: [[ADD:%.+]] = arith.addi [[BASE]], [[CAST]] : vector<8xindex>438    %step = vector.step {layout_result_0 = #xegpu.layout<sg_layout = [16], sg_data = [8]>}: vector<128xindex>439    gpu.return440  }441 442  // CHECK-LABEL: constant_with_slice_attr443  gpu.func @constant_with_slice_attr() {444    //CHECK: [[cst:%.+]] = arith.constant dense<10> : vector<1xindex>445    %cst = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [4, 2, 6, 1], sg_data = [1, 1, 1, 1]>, dims = [1, 2, 3]>} dense<10> : vector<4xindex>446    gpu.return447  }448 449  // CHECK-LABEL: vector_shape_cast450  gpu.func @vector_shape_cast() {451    %cst = arith.constant {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [8, 1, 1, 4], sg_data = [1, 1, 1, 32]>, dims = [0, 1, 2]>} dense<10> : vector<128xindex>452    %step = vector.step {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [8, 1, 1, 4], sg_data = [1, 1, 1, 32]>, dims = [0, 1, 2]>} : vector<128xindex>453    %muli = arith.muli %cst, %step {layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [8, 1, 1, 4], sg_data = [1, 1, 1, 32]>, dims = [0, 1, 2]>} : vector<128xindex>454    //CHECK: vector.shape_cast {{.*}} : vector<32xindex> to vector<1x1x1x32xindex>455    %shape_cast = vector.shape_cast %muli {layout_result_0 = #xegpu.layout<sg_layout = [8, 1, 1, 4], sg_data = [1, 1, 1, 32]>} : vector<128xindex> to vector<1x1x1x128xindex>456    gpu.return457  }458 459  // CHECK-LABEL: vector_broadcast460  gpu.func @vector_broadcast(%arg0: index, %arg1: index) {461    %muli = arith.muli %arg0, %arg1 : index462    // CHECK: vector.broadcast {{.*}} : index to vector<1x1x1x32xindex>463    %broadcast = vector.broadcast %muli {layout_result_0 = #xegpu.layout<sg_layout = [4, 2, 6, 1], sg_data = [1, 1, 1, 32]>} : index to vector<4x2x6x32xindex>464    gpu.return465  }466 467  // CHECK-LABEL: vector_transpose468  gpu.func @vector_transpose(%src: memref<256x32xf32>) {469    %tdesc = xegpu.create_nd_tdesc %src : memref<256x32xf32>470        -> !xegpu.tensor_desc<256x32xf32, #xegpu.layout<sg_layout = [4, 8], sg_data = [64, 32], lane_layout = [16, 1], lane_data = [1, 1], order =[0, 1]>>471    %load = xegpu.load_nd %tdesc[0, 0]472        : !xegpu.tensor_desc<256x32xf32, #xegpu.layout<sg_layout = [4, 8], sg_data = [64, 32], lane_layout = [16, 1], lane_data = [1, 1], order =[0, 1]>>473        -> vector<256x32xf32>474    //CHECK: vector.transpose {{.*}}, [1, 0] {layout_result_0 = #xegpu.layout<lane_layout = [1, 16], lane_data = [1, 1], order = [1, 0]>} : vector<64x32xf32> to vector<32x64xf32>475    %trans = vector.transpose %load, [1, 0] {layout_result_0 = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 64], lane_layout = [1, 16], lane_data = [1, 1], order =[1, 0]>} : vector<256x32xf32> to vector<32x256xf32>476      gpu.return477  }478 479  // CHECK-LABEL: non_splat_constant_2D480  gpu.func @non_splat_constant_2D() {481    // CHECK-DAG: %[[CST:.*]] = arith.constant dense<0> : vector<1x1xindex>482    // CHECK-DAG: %[[SGID:.*]] = gpu.subgroup_id : index483    // CHECK-DAG: %[[SGIDX:.*]] = index.remu %[[SGID]], %{{.*}}484    // CHECK-DAG: %[[SGIDY_TMP:.*]] = index.divu %[[SGID]], %{{.*}}485    // CHECK-DAG: %[[SGIDY:.*]] = index.remu %[[SGIDY_TMP]], %{{.*}}486    // CHECK-DAG: %[[IDY:.*]] = index.remu %[[SGIDY]], %{{.*}}487    // CHECK-DAG: %[[IDX:.*]] = index.remu %[[SGIDX]], %{{.*}}488    // CHECK-DAG: %[[STRIDECOL:.*]] = arith.muli %[[IDY]], %[[C16:.*]] : index489    // CHECK-DAG: %[[ADD:.*]] = arith.addi %[[C0:.*]], %[[STRIDECOL]] : index490    // CHECK-DAG: %[[STRIDEROW:.*]] = arith.muli %[[IDX]], %[[C0:.*]] : index491    // CHECK-DAG: %[[ADDSTRIDES:.*]] = arith.addi %[[ADD]], %[[STRIDEROW]] : index492    // CHECK-DAG: %[[BCAST:.*]] = vector.broadcast %[[ADDSTRIDES]] : index to vector<1x1xindex>493    // CHECK-DAG: arith.addi %[[CST]], %[[BCAST]] : vector<1x1xindex>494    %cst = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32, 1], sg_data = [1, 1]>} dense<[[0], [16], [32], [48], [64], [80], [96], [112], [128], [144], [160], [176], [192], [208], [224], [240], [256], [272], [288], [304], [320], [336], [352], [368], [384], [400], [416], [432], [448], [464], [480], [496]]> : vector<32x1xindex>495    gpu.return496  }497 498  // CHECK-LABEL: non_splat_constant_2D_non_unit_dim499  gpu.func @non_splat_constant_2D_non_unit_dim() {500    // CHECK-DAG: %[[BASECST:.*]] = arith.constant dense<{{\[}}{{\[}}0, 16{{\]}}, {{\[}}8, 24{{\]}}{{\]}}> : vector<2x2xindex>501    // CHECK-DAG: %[[SGID:.*]] = gpu.subgroup_id : index502    // CHECK-DAG: %[[SGIDX:.*]] = index.remu %[[SGID]], %{{.*}}503    // CHECK-DAG: %[[SGIDY_TMP:.*]] = index.divu %[[SGID]], %{{.*}}504    // CHECK-DAG: %[[SGIDY:.*]] = index.remu %[[SGIDY_TMP]], %{{.*}}505    // CHECK-DAG: %[[MULY:.*]] = index.mul %[[SGIDY]], %[[C2:.*]]506    // CHECK-DAG: %[[MULX:.*]] = index.mul %[[SGIDX]], %{{.*}}507    // CHECK-DAG: %[[REMU_Y:.*]] = index.remu %[[MULY]], %[[C8:.*]]508    // CHECK-DAG: %[[REMU_X:.*]] = index.remu %[[MULX]], %{{.*}}509    // CHECK-DAG: %[[MUL5:.*]] = arith.muli %[[REMU_Y]], %{{.*}} : index510    // CHECK-DAG: %[[ADD:.*]] = arith.addi %[[C0:.*]], %[[MUL5]] : index511    // CHECK-DAG: %[[MUL6:.*]] = arith.muli %[[REMU_X]], %[[C16:.*]] : index512    // CHECK-DAG: %[[ADDSTRIDES:.*]] = arith.addi %[[ADD]], %[[MUL6]] : index513    // CHECK-DAG: %[[BCAST:.*]] = vector.broadcast %[[ADDSTRIDES]] : index to vector<2x2xindex>514    // CHECK-DAG: %[[ADDCST:.*]] = arith.addi %[[BASECST]], %[[BCAST]] : vector<2x2xindex>515    %cst_8x8 = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [4, 4], sg_data = [2, 2]>} dense<[516         [0, 16, 32, 48, 64, 80, 96, 112],517         [8, 24, 40, 56, 72, 88, 104, 120],518         [16, 32, 48, 64, 80, 96, 112, 128],519         [24, 40, 56, 72, 88, 104, 120, 136],520         [32, 48, 64, 80, 96, 112, 128, 144],521         [40, 56, 72, 88, 104, 120, 136, 152],522         [48, 64, 80, 96, 112, 128, 144, 160],523         [56, 72, 88, 104, 120, 136, 152, 168]524      ]> : vector<8x8xindex>525      gpu.return526  }527 528  // CHECK-LABEL: non_splat_constant529  gpu.func @non_splat_constant() {530    // CHECK-DAG: %[[CST:.*]] = arith.constant dense<0> : vector<1xindex>531    // CHECK-DAG: %[[SGID:.*]] = gpu.subgroup_id : index532    // CHECK-DAG: %[[REMU:.*]] = index.remu %[[SGID]], %{{.*}}533    // CHECK-DAG: %[[REMU2:.*]] = index.remu %[[REMU]], %{{.*}}534    // CHECK-DAG: %[[MUL:.*]] = arith.muli %[[REMU2]], %[[C16:.*]] : index535    // CHECK-DAG: %[[ADDSTRIDES:.*]] = arith.addi %[[C0:.*]], %[[MUL]] : index536    // CHECK-DAG: %[[BCAST:.*]] = vector.broadcast %[[ADDSTRIDES]] : index to vector<1xindex>537    // CHECK-DAG: %[[ADD:.*]] = arith.addi %[[CST]], %[[BCAST]] : vector<1xindex>538    %cst = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [1]>} dense<[0, 16, 32, 48, 64, 80, 96, 112, 128, 144, 160, 176, 192, 208, 224, 240, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464, 480, 496]> : vector<32xindex>539    // CHECK: arith.constant dense<{{\[}}{{\[}}0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15{{\]}}{{\]}}> : vector<1x16xindex>540    %cst_1 = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32, 1], sg_data = [1, 16]>} dense<[[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]]> : vector<1x16xindex>541    gpu.return542  }543 544  // CHECK-LABEL: scalar_broadcast545  gpu.func @scalar_broadcast(%arg0: index) {546    // CHECK: vector.broadcast {{.*}} : index to vector<1x1x1xindex>547    %broadcast = vector.broadcast %arg0 {layout_result_0 = #xegpu.layout<sg_layout = [4, 8, 1], sg_data = [1, 1, 1]>} : index to vector<4x1x1xindex>548    gpu.return549  }550 551  // CHECK-LABEL: vector_mask_1D552  gpu.func @vector_mask_1D() {553    // CHECK-DAG: %[[SGID:.*]] = gpu.subgroup_id : index554    // CHECK-DAG: %[[REMU:.*]] = index.remu %[[SGID]], %[[C2:.*]]555    // CHECK-DAG: %[[MUL:.*]] = index.mul %[[REMU]], %[[C16:.*]]556    // CHECK-DAG: %[[REMU2:.*]] = index.remu %[[MUL]], %[[C32:.*]]557    // CHECK-DAG: %[[SUB:.*]] = arith.subi %[[C8:.*]], %[[REMU2]] : index558    // CHECK-DAG: %[[MAX:.*]] = arith.maxsi %[[SUB]], %[[C0:.*]] : index559    // CHECK-DAG: %[[MIN:.*]] = arith.minsi %[[MAX]], %[[C16:.*]] : index560    // CHECK-DAG: %[[MASK:.*]] = vector.create_mask %[[MIN]] : vector<16xi1>561    %constant_mask = vector.constant_mask [8] {layout_result_0 = #xegpu.layout<sg_layout = [2], sg_data = [16]>} : vector<32xi1>562    gpu.return563  }564 565  // CHECK-LABEL: vector_mask_2D566  gpu.func @vector_mask_2D() {567    // CHECK-DAG: %[[SGID:.*]] = gpu.subgroup_id : index568    // CHECK-DAG: %[[SGIDX:.*]] = index.remu %[[SGID]], %[[C4:.*]]569    // CHECK-DAG: %[[SGIDY_TMP:.*]] = index.divu %[[SGID]], %[[C4:.*]]570    // CHECK-DAG: %[[SGIDY:.*]] = index.remu %[[SGIDY_TMP]], %[[C8:.*]]571    // CHECK-DAG: %[[ROW:.*]] = index.mul %[[SGIDY]], %[[C32:.*]]572    // CHECK-DAG: %[[COL:.*]] = index.mul %[[SGIDX]], %[[C32:.*]]573    // CHECK-DAG: %[[MODROW:.*]] = index.remu %[[ROW]], %[[C256:.*]]574    // CHECK-DAG: %[[MODCOL:.*]] = index.remu %[[COL]], %[[C128:.*]]575    // CHECK-DAG: %[[SUBROW:.*]] = arith.subi %[[C16:.*]], %[[MODROW]] : index576    // CHECK-DAG: %[[MAXROW:.*]] = arith.maxsi %[[SUBROW]], %[[C4:.*]] : index577    // CHECK-DAG: %[[MINROW:.*]] = arith.minsi %[[MAXROW]], %[[C32:.*]] : index578    // CHECK-DAG: %[[SUBCOL:.*]] = arith.subi %[[C16:.*]], %[[MODCOL]] : index579    // CHECK-DAG: %[[MAXCOL:.*]] = arith.maxsi %[[SUBCOL]], %[[C7:.*]] : index580    // CHECK-DAG: %[[MINCOL:.*]] = arith.minsi %[[MAXCOL]], %[[C32:.*]] : index581    // CHECK-DAG: %[[MASK:.*]] = vector.create_mask %[[MINROW]], %[[MINCOL]] : vector<32x32xi1>582    %constant_mask = vector.constant_mask [16, 16] {layout_result_0 = #xegpu.layout<sg_layout = [8, 4], sg_data = [32, 32]>} : vector<256x128xi1>583    gpu.return584  }585 586  // CHECK-LABEL: distribute_load_slice_attr587  gpu.func @distribute_load_slice_attr() {588    %2 = memref.alloca() {alignment = 1024} : memref<4096xf32>589    %offset =  arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [8], sg_data = [32], inst_data = [16]> } dense<0> : vector<256xindex>590    %mask = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [8], sg_data = [32], inst_data = [16]> } dense<1> : vector<256xi1>591 592    // CHECK: %[[LOAD:.*]] = xegpu.load {{.*}} <{chunk_size = 1 : i64, layout = #xegpu.slice<#xegpu.layout<inst_data = [8, 16]>, dims = [0]>}>593    // CHECK-SAME: {layout_result_0 = #xegpu.slice<#xegpu.layout<inst_data = [8, 16]>, dims = [0]>} :594    // CHECK-SAME: memref<4096xf32>, vector<32xindex>, vector<32xi1> -> vector<32xf32>595    %3 = xegpu.load %2[%offset], %mask {chunk_size = 1, layout_result_0 = #xegpu.slice<#xegpu.layout<sg_layout = [8, 8], sg_data = [32, 32], inst_data = [8, 16]>, dims = [0]> } : memref<4096xf32>, vector<256xindex>, vector<256xi1> -> vector<256xf32>596 597    // CHECK: %[[BROADCAST:.*]] = vector.broadcast %[[LOAD]] {layout_result_0 = #xegpu.layout<inst_data = [8, 16]>} : vector<32xf32> to vector<32x32xf32>598    %4 = vector.broadcast %3 {layout_result_0 =599        #xegpu.layout<sg_layout = [8, 8], sg_data = [32, 32], inst_data = [8, 16]>} : vector<256xf32> to vector<256x256xf32>600    gpu.return601  }602}603