165 lines · plain
1// RUN: mlir-opt %s -acc-legalize-serial | FileCheck %s2 3acc.private.recipe @privatization_memref_10_f32 : memref<10xf32> init {4^bb0(%arg0: memref<10xf32>):5 %0 = memref.alloc() : memref<10xf32>6 acc.yield %0 : memref<10xf32>7} destroy {8^bb0(%arg0: memref<10xf32>):9 memref.dealloc %arg0 : memref<10xf32> 10 acc.terminator11}12 13acc.private.recipe @privatization_memref_10_10_f32 : memref<10x10xf32> init {14^bb0(%arg0: memref<10x10xf32>):15 %0 = memref.alloc() : memref<10x10xf32>16 acc.yield %0 : memref<10x10xf32>17} destroy {18^bb0(%arg0: memref<10x10xf32>):19 memref.dealloc %arg0 : memref<10x10xf32> 20 acc.terminator21}22 23acc.firstprivate.recipe @firstprivatization_memref_10xf32 : memref<10xf32> init {24^bb0(%arg0: memref<10xf32>):25 %0 = memref.alloc() : memref<10xf32>26 acc.yield %0 : memref<10xf32>27} copy {28^bb0(%arg0: memref<10xf32>, %arg1: memref<10xf32>):29 acc.terminator30} destroy {31^bb0(%arg0: memref<10xf32>):32 memref.dealloc %arg0 : memref<10xf32> 33 acc.terminator34}35 36acc.reduction.recipe @reduction_add_i64 : i64 reduction_operator<add> init {37^bb0(%0: i64):38 %1 = arith.constant 0 : i6439 acc.yield %1 : i6440} combiner {41^bb0(%0: i64, %1: i64):42 %2 = arith.addi %0, %1 : i6443 acc.yield %2 : i6444}45 46acc.reduction.recipe @reduction_add_memref_i64 : memref<i64> reduction_operator<add> init {47^bb0(%arg0: memref<i64>):48 %0 = memref.alloca() : memref<i64>49 %c0 = arith.constant 0 : i6450 memref.store %c0, %0[] : memref<i64>51 acc.yield %0 : memref<i64>52} combiner {53^bb0(%arg0: memref<i64>, %arg1: memref<i64>):54 %0 = memref.load %arg0[] : memref<i64>55 %1 = memref.load %arg1[] : memref<i64>56 %2 = arith.addi %0, %1 : i6457 memref.store %2, %arg0[] : memref<i64>58 acc.terminator59}60 61// CHECK: func.func @testserialop(%[[VAL_0:.*]]: memref<10xf32>, %[[VAL_1:.*]]: memref<10xf32>, %[[VAL_2:.*]]: memref<10x10xf32>) {62// CHECK: %[[VAL_3:.*]] = arith.constant 1 : i6463// CHECK: %[[VAL_4:.*]] = arith.constant 1 : i3264// CHECK: %[[VAL_5:.*]] = arith.constant 1 : index65// CHECK: acc.parallel async(%[[VAL_3]] : i64) num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {66// CHECK: }67// CHECK: acc.parallel async(%[[VAL_4]] : i32) num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {68// CHECK: }69// CHECK: acc.parallel async(%[[VAL_5]] : index) num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {70// CHECK: }71// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) wait({%[[VAL_3]] : i64}) {72// CHECK: }73// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) wait({%[[VAL_4]] : i32}) {74// CHECK: }75// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) wait({%[[VAL_5]] : index}) {76// CHECK: }77// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) wait({%[[VAL_3]] : i64, %[[VAL_4]] : i32, %[[VAL_5]] : index}) {78// CHECK: }79// CHECK: %[[VAL_6:.*]] = acc.firstprivate varPtr(%[[VAL_1]] : memref<10xf32>) recipe(@firstprivatization_memref_10xf32) -> memref<10xf32>80// CHECK: %[[VAL_9:.*]] = acc.private varPtr(%[[VAL_2]] : memref<10x10xf32>) recipe(@privatization_memref_10_10_f32) -> memref<10x10xf32>81// CHECK: acc.parallel firstprivate(%[[VAL_6]] : memref<10xf32>) num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) private(%[[VAL_9]] : memref<10x10xf32>) vector_length(%[[VAL_4]] : i32) {82// CHECK: }83// CHECK: %[[VAL_7:.*]] = acc.copyin varPtr(%[[VAL_0]] : memref<10xf32>) -> memref<10xf32> {dataClause = #acc<data_clause acc_copy>}84// CHECK: acc.parallel dataOperands(%[[VAL_7]] : memref<10xf32>) num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {85// CHECK: }86// CHECK: %[[I64MEM:.*]] = memref.alloca() : memref<i64>87// CHECK: memref.store %[[VAL_3]], %[[I64MEM]][] : memref<i64>88// CHECK: %[[VAL_10:.*]] = acc.reduction varPtr(%[[I64MEM]] : memref<i64>) recipe(@reduction_add_memref_i64) -> memref<i64>89// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) reduction(%[[VAL_10]] : memref<i64>) {90// CHECK: }91// CHECK: acc.parallel combined(loop) num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {92// CHECK: acc.loop combined(serial) control(%{{.*}} : index) = (%[[VAL_5]] : index) to (%[[VAL_5]] : index) step (%[[VAL_5]] : index) {93// CHECK: acc.yield94// CHECK: } attributes {seq = [#acc.device_type<none>]}95// CHECK: acc.terminator96// CHECK: }97// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {98// CHECK: } attributes {defaultAttr = #acc<defaultvalue none>}99// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {100// CHECK: } attributes {defaultAttr = #acc<defaultvalue present>}101// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {102// CHECK: }103// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {104// CHECK: }105// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {106// CHECK: } attributes {selfAttr}107// CHECK: acc.parallel num_gangs({%[[VAL_4]] : i32}) num_workers(%[[VAL_4]] : i32) vector_length(%[[VAL_4]] : i32) {108// CHECK: acc.yield109// CHECK: } attributes {selfAttr}110// CHECK: return111// CHECK: }112 113func.func @testserialop(%a: memref<10xf32>, %b: memref<10xf32>, %c: memref<10x10xf32>) -> () {114 %i64value = arith.constant 1 : i64115 %i32value = arith.constant 1 : i32116 %idxValue = arith.constant 1 : index117 acc.serial async(%i64value: i64) {118 }119 acc.serial async(%i32value: i32) {120 }121 acc.serial async(%idxValue: index) {122 }123 acc.serial wait({%i64value: i64}) {124 }125 acc.serial wait({%i32value: i32}) {126 }127 acc.serial wait({%idxValue: index}) {128 }129 acc.serial wait({%i64value : i64, %i32value : i32, %idxValue : index}) {130 }131 %firstprivate = acc.firstprivate varPtr(%b : memref<10xf32>) recipe(@firstprivatization_memref_10xf32) -> memref<10xf32>132 %c_private = acc.private varPtr(%c : memref<10x10xf32>) recipe(@privatization_memref_10_10_f32) -> memref<10x10xf32>133 acc.serial private(%c_private : memref<10x10xf32>) firstprivate(%firstprivate : memref<10xf32>) {134 }135 %copyinfromcopy = acc.copyin varPtr(%a : memref<10xf32>) -> memref<10xf32> {dataClause = #acc<data_clause acc_copy>}136 acc.serial dataOperands(%copyinfromcopy : memref<10xf32>) {137 }138 %i64mem = memref.alloca() : memref<i64>139 memref.store %i64value, %i64mem[] : memref<i64>140 %i64reduction = acc.reduction varPtr(%i64mem : memref<i64>) recipe(@reduction_add_memref_i64) -> memref<i64>141 acc.serial reduction(%i64reduction : memref<i64>) {142 }143 acc.serial combined(loop) {144 acc.loop combined(serial) control(%arg3 : index) = (%idxValue : index) to (%idxValue : index) step (%idxValue : index) {145 acc.yield146 } attributes {seq = [#acc.device_type<none>]}147 acc.terminator148 }149 acc.serial {150 } attributes {defaultAttr = #acc<defaultvalue none>}151 acc.serial {152 } attributes {defaultAttr = #acc<defaultvalue present>}153 acc.serial {154 } attributes {asyncAttr}155 acc.serial {156 } attributes {waitAttr}157 acc.serial {158 } attributes {selfAttr}159 acc.serial {160 acc.yield161 } attributes {selfAttr}162 return163}164 165