1// RUN: mlir-opt -test-gpu-rewrite %s | FileCheck %s 2 3// NOTE: Assertions have been autogenerated by utils/generate-test-checks.py 4// CHECK: gpu.module @kernels { 5gpu.module @kernels { 6 7 // CHECK-LABEL: gpu.func @kernel( 8 // CHECK-SAME: [[VAL_0:%.*]]: f32) workgroup([[VAL_1:%.*]] : memref<32xf32, 3>) kernel { 9 gpu.func @kernel(%arg0 : f32) kernel { 10 // CHECK: [[VAL_2:%.*]] = constant 31 : i32 11 // CHECK: [[VAL_3:%.*]] = constant 0 : i32 12 // CHECK: [[VAL_4:%.*]] = constant 0 : index 13 // CHECK: [[VAL_5:%.*]] = constant 32 : i32 14 // CHECK: [[VAL_6:%.*]] = constant 1 : i32 15 // CHECK: [[VAL_7:%.*]] = constant 2 : i32 16 // CHECK: [[VAL_8:%.*]] = constant 4 : i32 17 // CHECK: [[VAL_9:%.*]] = constant 8 : i32 18 // CHECK: [[VAL_10:%.*]] = constant 16 : i32 19 // CHECK: [[VAL_11:%.*]] = "gpu.block_dim"() {dimension = "x"} : () -> index 20 // CHECK: [[VAL_12:%.*]] = index_cast [[VAL_11]] : index to i32 21 // CHECK: [[VAL_13:%.*]] = "gpu.block_dim"() {dimension = "y"} : () -> index 22 // CHECK: [[VAL_14:%.*]] = index_cast [[VAL_13]] : index to i32 23 // CHECK: [[VAL_15:%.*]] = "gpu.block_dim"() {dimension = "z"} : () -> index 24 // CHECK: [[VAL_16:%.*]] = index_cast [[VAL_15]] : index to i32 25 // CHECK: [[VAL_17:%.*]] = "gpu.thread_id"() {dimension = "x"} : () -> index 26 // CHECK: [[VAL_18:%.*]] = index_cast [[VAL_17]] : index to i32 27 // CHECK: [[VAL_19:%.*]] = "gpu.thread_id"() {dimension = "y"} : () -> index 28 // CHECK: [[VAL_20:%.*]] = index_cast [[VAL_19]] : index to i32 29 // CHECK: [[VAL_21:%.*]] = "gpu.thread_id"() {dimension = "z"} : () -> index 30 // CHECK: [[VAL_22:%.*]] = index_cast [[VAL_21]] : index to i32 31 // CHECK: [[VAL_23:%.*]] = muli [[VAL_22]], [[VAL_14]] : i32 32 // CHECK: [[VAL_24:%.*]] = addi [[VAL_23]], [[VAL_20]] : i32 33 // CHECK: [[VAL_25:%.*]] = muli [[VAL_24]], [[VAL_12]] : i32 34 // CHECK: [[VAL_26:%.*]] = muli [[VAL_12]], [[VAL_14]] : i32 35 // CHECK: [[VAL_27:%.*]] = addi [[VAL_25]], [[VAL_18]] : i32 36 // CHECK: [[VAL_28:%.*]] = muli [[VAL_26]], [[VAL_16]] : i32 37 // CHECK: [[VAL_29:%.*]] = and [[VAL_27]], [[VAL_2]] : i32 38 // CHECK: [[VAL_30:%.*]] = cmpi "eq", [[VAL_29]], [[VAL_3]] : i32 39 // CHECK: [[VAL_31:%.*]] = subi [[VAL_27]], [[VAL_29]] : i32 40 // CHECK: [[VAL_32:%.*]] = subi [[VAL_28]], [[VAL_31]] : i32 41 // CHECK: [[VAL_33:%.*]] = cmpi "slt", [[VAL_32]], [[VAL_5]] : i32 42 // CHECK: cond_br [[VAL_33]], ^bb1, ^bb17 43 // CHECK: ^bb1: 44 // CHECK: [[VAL_34:%.*]], [[VAL_35:%.*]] = gpu.shuffle [[VAL_0]], [[VAL_6]], [[VAL_32]] xor : f32 45 // CHECK: cond_br [[VAL_35]], ^bb2, ^bb3 46 // CHECK: ^bb2: 47 // CHECK: [[VAL_36:%.*]] = cmpf "ugt", [[VAL_0]], [[VAL_34]] : f32 48 // CHECK: [[VAL_37:%.*]] = select [[VAL_36]], [[VAL_0]], [[VAL_34]] : f32 49 // CHECK: br ^bb4([[VAL_37]] : f32) 50 // CHECK: ^bb3: 51 // CHECK: br ^bb4([[VAL_0]] : f32) 52 // CHECK: ^bb4([[VAL_38:%.*]]: f32): 53 // CHECK: [[VAL_39:%.*]], [[VAL_40:%.*]] = gpu.shuffle [[VAL_38]], [[VAL_7]], [[VAL_32]] xor : f32 54 // CHECK: cond_br [[VAL_40]], ^bb5, ^bb6 55 // CHECK: ^bb5: 56 // CHECK: [[VAL_41:%.*]] = cmpf "ugt", [[VAL_38]], [[VAL_39]] : f32 57 // CHECK: [[VAL_42:%.*]] = select [[VAL_41]], [[VAL_38]], [[VAL_39]] : f32 58 // CHECK: br ^bb7([[VAL_42]] : f32) 59 // CHECK: ^bb6: 60 // CHECK: br ^bb7([[VAL_38]] : f32) 61 // CHECK: ^bb7([[VAL_43:%.*]]: f32): 62 // CHECK: [[VAL_44:%.*]], [[VAL_45:%.*]] = gpu.shuffle [[VAL_43]], [[VAL_8]], [[VAL_32]] xor : f32 63 // CHECK: cond_br [[VAL_45]], ^bb8, ^bb9 64 // CHECK: ^bb8: 65 // CHECK: [[VAL_46:%.*]] = cmpf "ugt", [[VAL_43]], [[VAL_44]] : f32 66 // CHECK: [[VAL_47:%.*]] = select [[VAL_46]], [[VAL_43]], [[VAL_44]] : f32 67 // CHECK: br ^bb10([[VAL_47]] : f32) 68 // CHECK: ^bb9: 69 // CHECK: br ^bb10([[VAL_43]] : f32) 70 // CHECK: ^bb10([[VAL_48:%.*]]: f32): 71 // CHECK: [[VAL_49:%.*]], [[VAL_50:%.*]] = gpu.shuffle [[VAL_48]], [[VAL_9]], [[VAL_32]] xor : f32 72 // CHECK: cond_br [[VAL_50]], ^bb11, ^bb12 73 // CHECK: ^bb11: 74 // CHECK: [[VAL_51:%.*]] = cmpf "ugt", [[VAL_48]], [[VAL_49]] : f32 75 // CHECK: [[VAL_52:%.*]] = select [[VAL_51]], [[VAL_48]], [[VAL_49]] : f32 76 // CHECK: br ^bb13([[VAL_52]] : f32) 77 // CHECK: ^bb12: 78 // CHECK: br ^bb13([[VAL_48]] : f32) 79 // CHECK: ^bb13([[VAL_53:%.*]]: f32): 80 // CHECK: [[VAL_54:%.*]], [[VAL_55:%.*]] = gpu.shuffle [[VAL_53]], [[VAL_10]], [[VAL_32]] xor : f32 81 // CHECK: cond_br [[VAL_55]], ^bb14, ^bb15 82 // CHECK: ^bb14: 83 // CHECK: [[VAL_56:%.*]] = cmpf "ugt", [[VAL_53]], [[VAL_54]] : f32 84 // CHECK: [[VAL_57:%.*]] = select [[VAL_56]], [[VAL_53]], [[VAL_54]] : f32 85 // CHECK: br ^bb16([[VAL_57]] : f32) 86 // CHECK: ^bb15: 87 // CHECK: br ^bb16([[VAL_53]] : f32) 88 // CHECK: ^bb16([[VAL_58:%.*]]: f32): 89 // CHECK: br ^bb18([[VAL_58]] : f32) 90 // CHECK: ^bb17: 91 // CHECK: [[VAL_59:%.*]], [[VAL_60:%.*]] = gpu.shuffle [[VAL_0]], [[VAL_6]], [[VAL_5]] xor : f32 92 // CHECK: [[VAL_61:%.*]] = cmpf "ugt", [[VAL_0]], [[VAL_59]] : f32 93 // CHECK: [[VAL_62:%.*]] = select [[VAL_61]], [[VAL_0]], [[VAL_59]] : f32 94 // CHECK: [[VAL_63:%.*]], [[VAL_64:%.*]] = gpu.shuffle [[VAL_62]], [[VAL_7]], [[VAL_5]] xor : f32 95 // CHECK: [[VAL_65:%.*]] = cmpf "ugt", [[VAL_62]], [[VAL_63]] : f32 96 // CHECK: [[VAL_66:%.*]] = select [[VAL_65]], [[VAL_62]], [[VAL_63]] : f32 97 // CHECK: [[VAL_67:%.*]], [[VAL_68:%.*]] = gpu.shuffle [[VAL_66]], [[VAL_8]], [[VAL_5]] xor : f32 98 // CHECK: [[VAL_69:%.*]] = cmpf "ugt", [[VAL_66]], [[VAL_67]] : f32 99 // CHECK: [[VAL_70:%.*]] = select [[VAL_69]], [[VAL_66]], [[VAL_67]] : f32 100 // CHECK: [[VAL_71:%.*]], [[VAL_72:%.*]] = gpu.shuffle [[VAL_70]], [[VAL_9]], [[VAL_5]] xor : f32 101 // CHECK: [[VAL_73:%.*]] = cmpf "ugt", [[VAL_70]], [[VAL_71]] : f32 102 // CHECK: [[VAL_74:%.*]] = select [[VAL_73]], [[VAL_70]], [[VAL_71]] : f32 103 // CHECK: [[VAL_75:%.*]], [[VAL_76:%.*]] = gpu.shuffle [[VAL_74]], [[VAL_10]], [[VAL_5]] xor : f32 104 // CHECK: [[VAL_77:%.*]] = cmpf "ugt", [[VAL_74]], [[VAL_75]] : f32 105 // CHECK: [[VAL_78:%.*]] = select [[VAL_77]], [[VAL_74]], [[VAL_75]] : f32 106 // CHECK: br ^bb18([[VAL_78]] : f32) 107 // CHECK: ^bb18([[VAL_79:%.*]]: f32): 108 // CHECK: cond_br [[VAL_30]], ^bb19, ^bb20 109 // CHECK: ^bb19: 110 // CHECK: [[VAL_80:%.*]] = divi_signed [[VAL_27]], [[VAL_5]] : i32 111 // CHECK: [[VAL_81:%.*]] = index_cast [[VAL_80]] : i32 to index 112 // CHECK: store [[VAL_79]], [[VAL_1]]{{\[}}[[VAL_81]]] : memref<32xf32, 3> 113 // CHECK: br ^bb21 114 // CHECK: ^bb20: 115 // CHECK: br ^bb21 116 // CHECK: ^bb21: 117 // CHECK: gpu.barrier 118 // CHECK: [[VAL_82:%.*]] = addi [[VAL_28]], [[VAL_2]] : i32 119 // CHECK: [[VAL_83:%.*]] = divi_signed [[VAL_82]], [[VAL_5]] : i32 120 // CHECK: [[VAL_84:%.*]] = cmpi "slt", [[VAL_27]], [[VAL_83]] : i32 121 // CHECK: cond_br [[VAL_84]], ^bb22, ^bb41 122 // CHECK: ^bb22: 123 // CHECK: [[VAL_85:%.*]] = index_cast [[VAL_27]] : i32 to index 124 // CHECK: [[VAL_86:%.*]] = load [[VAL_1]]{{\[}}[[VAL_85]]] : memref<32xf32, 3> 125 // CHECK: [[VAL_87:%.*]] = cmpi "slt", [[VAL_83]], [[VAL_5]] : i32 126 // CHECK: cond_br [[VAL_87]], ^bb23, ^bb39 127 // CHECK: ^bb23: 128 // CHECK: [[VAL_88:%.*]], [[VAL_89:%.*]] = gpu.shuffle [[VAL_86]], [[VAL_6]], [[VAL_83]] xor : f32 129 // CHECK: cond_br [[VAL_89]], ^bb24, ^bb25 130 // CHECK: ^bb24: 131 // CHECK: [[VAL_90:%.*]] = cmpf "ugt", [[VAL_86]], [[VAL_88]] : f32 132 // CHECK: [[VAL_91:%.*]] = select [[VAL_90]], [[VAL_86]], [[VAL_88]] : f32 133 // CHECK: br ^bb26([[VAL_91]] : f32) 134 // CHECK: ^bb25: 135 // CHECK: br ^bb26([[VAL_86]] : f32) 136 // CHECK: ^bb26([[VAL_92:%.*]]: f32): 137 // CHECK: [[VAL_93:%.*]], [[VAL_94:%.*]] = gpu.shuffle [[VAL_92]], [[VAL_7]], [[VAL_83]] xor : f32 138 // CHECK: cond_br [[VAL_94]], ^bb27, ^bb28 139 // CHECK: ^bb27: 140 // CHECK: [[VAL_95:%.*]] = cmpf "ugt", [[VAL_92]], [[VAL_93]] : f32 141 // CHECK: [[VAL_96:%.*]] = select [[VAL_95]], [[VAL_92]], [[VAL_93]] : f32 142 // CHECK: br ^bb29([[VAL_96]] : f32) 143 // CHECK: ^bb28: 144 // CHECK: br ^bb29([[VAL_92]] : f32) 145 // CHECK: ^bb29([[VAL_97:%.*]]: f32): 146 // CHECK: [[VAL_98:%.*]], [[VAL_99:%.*]] = gpu.shuffle [[VAL_97]], [[VAL_8]], [[VAL_83]] xor : f32 147 // CHECK: cond_br [[VAL_99]], ^bb30, ^bb31 148 // CHECK: ^bb30: 149 // CHECK: [[VAL_100:%.*]] = cmpf "ugt", [[VAL_97]], [[VAL_98]] : f32 150 // CHECK: [[VAL_101:%.*]] = select [[VAL_100]], [[VAL_97]], [[VAL_98]] : f32 151 // CHECK: br ^bb32([[VAL_101]] : f32) 152 // CHECK: ^bb31: 153 // CHECK: br ^bb32([[VAL_97]] : f32) 154 // CHECK: ^bb32([[VAL_102:%.*]]: f32): 155 // CHECK: [[VAL_103:%.*]], [[VAL_104:%.*]] = gpu.shuffle [[VAL_102]], [[VAL_9]], [[VAL_83]] xor : f32 156 // CHECK: cond_br [[VAL_104]], ^bb33, ^bb34 157 // CHECK: ^bb33: 158 // CHECK: [[VAL_105:%.*]] = cmpf "ugt", [[VAL_102]], [[VAL_103]] : f32 159 // CHECK: [[VAL_106:%.*]] = select [[VAL_105]], [[VAL_102]], [[VAL_103]] : f32 160 // CHECK: br ^bb35([[VAL_106]] : f32) 161 // CHECK: ^bb34: 162 // CHECK: br ^bb35([[VAL_102]] : f32) 163 // CHECK: ^bb35([[VAL_107:%.*]]: f32): 164 // CHECK: [[VAL_108:%.*]], [[VAL_109:%.*]] = gpu.shuffle [[VAL_107]], [[VAL_10]], [[VAL_83]] xor : f32 165 // CHECK: cond_br [[VAL_109]], ^bb36, ^bb37 166 // CHECK: ^bb36: 167 // CHECK: [[VAL_110:%.*]] = cmpf "ugt", [[VAL_107]], [[VAL_108]] : f32 168 // CHECK: [[VAL_111:%.*]] = select [[VAL_110]], [[VAL_107]], [[VAL_108]] : f32 169 // CHECK: br ^bb38([[VAL_111]] : f32) 170 // CHECK: ^bb37: 171 // CHECK: br ^bb38([[VAL_107]] : f32) 172 // CHECK: ^bb38([[VAL_112:%.*]]: f32): 173 // CHECK: br ^bb40([[VAL_112]] : f32) 174 // CHECK: ^bb39: 175 // CHECK: [[VAL_113:%.*]], [[VAL_114:%.*]] = gpu.shuffle [[VAL_86]], [[VAL_6]], [[VAL_5]] xor : f32 176 // CHECK: [[VAL_115:%.*]] = cmpf "ugt", [[VAL_86]], [[VAL_113]] : f32 177 // CHECK: [[VAL_116:%.*]] = select [[VAL_115]], [[VAL_86]], [[VAL_113]] : f32 178 // CHECK: [[VAL_117:%.*]], [[VAL_118:%.*]] = gpu.shuffle [[VAL_116]], [[VAL_7]], [[VAL_5]] xor : f32 179 // CHECK: [[VAL_119:%.*]] = cmpf "ugt", [[VAL_116]], [[VAL_117]] : f32 180 // CHECK: [[VAL_120:%.*]] = select [[VAL_119]], [[VAL_116]], [[VAL_117]] : f32 181 // CHECK: [[VAL_121:%.*]], [[VAL_122:%.*]] = gpu.shuffle [[VAL_120]], [[VAL_8]], [[VAL_5]] xor : f32 182 // CHECK: [[VAL_123:%.*]] = cmpf "ugt", [[VAL_120]], [[VAL_121]] : f32 183 // CHECK: [[VAL_124:%.*]] = select [[VAL_123]], [[VAL_120]], [[VAL_121]] : f32 184 // CHECK: [[VAL_125:%.*]], [[VAL_126:%.*]] = gpu.shuffle [[VAL_124]], [[VAL_9]], [[VAL_5]] xor : f32 185 // CHECK: [[VAL_127:%.*]] = cmpf "ugt", [[VAL_124]], [[VAL_125]] : f32 186 // CHECK: [[VAL_128:%.*]] = select [[VAL_127]], [[VAL_124]], [[VAL_125]] : f32 187 // CHECK: [[VAL_129:%.*]], [[VAL_130:%.*]] = gpu.shuffle [[VAL_128]], [[VAL_10]], [[VAL_5]] xor : f32 188 // CHECK: [[VAL_131:%.*]] = cmpf "ugt", [[VAL_128]], [[VAL_129]] : f32 189 // CHECK: [[VAL_132:%.*]] = select [[VAL_131]], [[VAL_128]], [[VAL_129]] : f32 190 // CHECK: br ^bb40([[VAL_132]] : f32) 191 // CHECK: ^bb40([[VAL_133:%.*]]: f32): 192 // CHECK: store [[VAL_133]], [[VAL_1]]{{\[}}[[VAL_4]]] : memref<32xf32, 3> 193 // CHECK: br ^bb42 194 // CHECK: ^bb41: 195 // CHECK: br ^bb42 196 // CHECK: ^bb42: 197 // CHECK: gpu.barrier 198 %sum = "gpu.all_reduce"(%arg0) ({}) {op = "max"} : (f32) -> (f32) 199 gpu.return 200 } 201 202} 203