1// RUN: mlir-opt %s -test-tensor-copy-insertion=allow-return-allocs-from-loops -allow-unregistered-dialect -split-input-file | FileCheck %s 2// RUN: mlir-opt %s -test-tensor-copy-insertion="allow-return-allocs-from-loops bufferize-function-boundaries" -split-input-file | FileCheck %s --check-prefix=CHECK-FUNC 3 4// CHECK-LABEL: func @scf_for( 5// CHECK-SAME: %[[A:.*]]: tensor<?xf32>, %[[B:.*]]: tensor<?xf32> 6func.func @scf_for(%A : tensor<?xf32>, %B : tensor<?xf32>, 7 %lb : index, %ub : index, %step : index) 8 -> (tensor<?xf32>, tensor<?xf32>) 9{ 10 // CHECK: %[[A_copy:.*]] = bufferization.alloc_tensor() copy(%[[A]]) : tensor<?xf32> 11 // CHECK: %[[B_copy:.*]] = bufferization.alloc_tensor() copy(%[[B]]) : tensor<?xf32> 12 // CHECK: %[[for:.*]]:2 = scf.for {{.*}} iter_args(%[[iter1:.*]] = %[[A_copy]], %[[iter2:.*]] = %[[B_copy]]) 13 %r0:2 = scf.for %i = %lb to %ub step %step iter_args(%tA = %A, %tB = %B) 14 -> (tensor<?xf32>, tensor<?xf32>) 15 { 16 // CHECK: scf.yield %[[iter1]], %[[iter2]] 17 scf.yield %tA, %tB : tensor<?xf32>, tensor<?xf32> 18 } 19 20 return %r0#0, %r0#1 : tensor<?xf32>, tensor<?xf32> 21} 22 23// ----- 24 25// CHECK-LABEL: func @scf_for_swapping_yields( 26// CHECK-SAME: %[[A:.*]]: tensor<?xf32>, %[[B:.*]]: tensor<?xf32> 27func.func @scf_for_swapping_yields(%A : tensor<?xf32>, %B : tensor<?xf32>, 28 %lb : index, %ub : index, %step : index) 29 -> (tensor<?xf32>, tensor<?xf32>) 30{ 31 // CHECK: %[[A_copy:.*]] = bufferization.alloc_tensor() copy(%[[A]]) : tensor<?xf32> 32 // CHECK: %[[B_copy:.*]] = bufferization.alloc_tensor() copy(%[[B]]) : tensor<?xf32> 33 // CHECK: %[[for:.*]]:2 = scf.for {{.*}} iter_args(%[[iter1:.*]] = %[[A_copy]], %[[iter2:.*]] = %[[B_copy]]) 34 %r0:2 = scf.for %i = %lb to %ub step %step iter_args(%tA = %A, %tB = %B) 35 -> (tensor<?xf32>, tensor<?xf32>) 36 { 37 // Yield tensors in different order. 38 // CHECK-DAG: %[[yield1:.*]] = bufferization.alloc_tensor() copy(%[[iter2]]) : tensor<?xf32> 39 // CHECK-DAG: %[[yield2:.*]] = bufferization.alloc_tensor() copy(%[[iter1]]) : tensor<?xf32> 40 // CHECK: scf.yield %[[yield1]], %[[yield2]] 41 scf.yield %tB, %tA : tensor<?xf32>, tensor<?xf32> 42 } 43 44 return %r0#0, %r0#1 : tensor<?xf32>, tensor<?xf32> 45} 46 47// ----- 48 49// CHECK-LABEL: func @scf_while( 50// CHECK-SAME: %[[A:.*]]: tensor<5xi1>, %[[B:.*]]: tensor<5xi1> 51func.func @scf_while(%A: tensor<5xi1>, %B: tensor<5xi1>, %idx: index) 52 -> (tensor<5xi1>, tensor<5xi1>) 53{ 54 // CHECK: %[[A_copy:.*]] = bufferization.alloc_tensor() copy(%[[A]]) : tensor<5xi1> 55 // CHECK: %[[B_copy:.*]] = bufferization.alloc_tensor() copy(%[[B]]) : tensor<5xi1> 56 // CHECK: %[[loop:.*]]:2 = scf.while (%[[w0:.*]] = %[[A_copy]], %[[w1:.*]] = %[[B_copy]]) {{.*}} { 57 %r0, %r1 = scf.while (%w0 = %A, %w1 = %B) 58 : (tensor<5xi1>, tensor<5xi1>) -> (tensor<5xi1>, tensor<5xi1>) { 59 // CHECK: %[[condition:.*]] = tensor.extract %[[w0]] 60 %condition = tensor.extract %w0[%idx] : tensor<5xi1> 61 // Yield tensors in different order. 62 // CHECK: scf.condition(%[[condition]]) %[[w0]], %[[w1]] 63 scf.condition(%condition) %w0, %w1 : tensor<5xi1>, tensor<5xi1> 64 } do { 65 ^bb0(%b0: tensor<5xi1>, %b1: tensor<5xi1>): 66 // CHECK: } do { 67 // CHECK: ^bb0(%[[b0:.*]]: tensor<5xi1>, %[[b1:.*]]: tensor<5xi1>): 68 // CHECK: scf.yield %[[b0]], %[[b1]] 69 // CHECK: } 70 scf.yield %b0, %b1 : tensor<5xi1>, tensor<5xi1> 71 } 72 73 return %r0, %r1 : tensor<5xi1>, tensor<5xi1> 74} 75 76// ----- 77 78// CHECK-LABEL: func @scf_while_non_equiv_condition_and_body( 79// CHECK-SAME: %[[A:.*]]: tensor<5xi1>, %[[B:.*]]: tensor<5xi1> 80func.func @scf_while_non_equiv_condition_and_body(%A: tensor<5xi1>, 81 %B: tensor<5xi1>, 82 %idx: index) 83 -> (tensor<5xi1>, tensor<5xi1>) 84{ 85 // CHECK: %[[A_copy:.*]] = bufferization.alloc_tensor() copy(%[[A]]) : tensor<5xi1> 86 // CHECK: %[[B_copy:.*]] = bufferization.alloc_tensor() copy(%[[B]]) : tensor<5xi1> 87 // CHECK: %[[loop:.*]]:2 = scf.while (%[[w0:.*]] = %[[A_copy]], %[[w1:.*]] = %[[B_copy]]) {{.*}} { 88 %r0, %r1 = scf.while (%w0 = %A, %w1 = %B) 89 : (tensor<5xi1>, tensor<5xi1>) -> (tensor<5xi1>, tensor<5xi1>) { 90 // CHECK: %[[condition:.*]] = tensor.extract %[[w0]] 91 %condition = tensor.extract %w0[%idx] : tensor<5xi1> 92 // Yield tensors in different order. 93 // CHECK-DAG: %[[yield0:.*]] = bufferization.alloc_tensor() copy(%[[w1]]) : tensor<5xi1> 94 // CHECK-DAG: %[[yield1:.*]] = bufferization.alloc_tensor() copy(%[[w0]]) : tensor<5xi1> 95 // CHECK: scf.condition(%[[condition]]) %[[yield0]], %[[yield1]] 96 scf.condition(%condition) %w1, %w0 : tensor<5xi1>, tensor<5xi1> 97 } do { 98 ^bb0(%b0: tensor<5xi1>, %b1: tensor<5xi1>): 99 // CHECK: } do { 100 // CHECK: ^bb0(%[[b0:.*]]: tensor<5xi1>, %[[b1:.*]]: tensor<5xi1>): 101 // CHECK: scf.yield %[[b1]], %[[b0]] 102 // CHECK: } 103 scf.yield %b1, %b0 : tensor<5xi1>, tensor<5xi1> 104 } 105 106 return %r0, %r1 : tensor<5xi1>, tensor<5xi1> 107} 108 109// ----- 110 111// CHECK-LABEL: func @scf_forall_out_of_place( 112// CHECK-SAME: %[[arg0:.*]]: tensor<100xf32>, %[[arg1:.*]]: tensor<100xf32> 113// CHECK-FUNC-LABEL: func @scf_forall_out_of_place( 114func.func @scf_forall_out_of_place(%in: tensor<100xf32>, 115 %out: tensor<100xf32>) { 116 %c1 = arith.constant 1 : index 117 %num_threads = arith.constant 100 : index 118 119 // CHECK-FUNC-NOT: alloc_tensor 120 // CHECK: %[[alloc:.*]] = bufferization.alloc_tensor() copy(%[[arg1]]) : tensor<100xf32> 121 // CHECK: scf.forall {{.*}} shared_outs(%[[o:.*]] = %[[alloc]]) 122 %result = scf.forall (%thread_idx) in (%num_threads) shared_outs(%o = %out) -> tensor<100xf32> { 123 // CHECK: tensor.extract_slice 124 // CHECK: scf.forall.in_parallel 125 // CHECK: tensor.parallel_insert_slice %{{.*}} into %[[o]] 126 %1 = tensor.extract_slice %in[%thread_idx][1][1] : tensor<100xf32> to tensor<1xf32> 127 scf.forall.in_parallel { 128 tensor.parallel_insert_slice %1 into %o[%thread_idx][1][1] : 129 tensor<1xf32> into tensor<100xf32> 130 } 131 // CHECK: } {mapping = [#gpu.thread<x>]} 132 } {mapping = [#gpu.thread<x>]} 133 return 134} 135