1// RUN: mlir-opt %s \ 2// RUN: -async-parallel-for \ 3// RUN: -async-to-async-runtime \ 4// RUN: -async-runtime-ref-counting \ 5// RUN: -async-runtime-ref-counting-opt \ 6// RUN: -convert-async-to-llvm \ 7// RUN: -convert-linalg-to-loops \ 8// RUN: -convert-scf-to-cf \ 9// RUN: -arith-expand \ 10// RUN: -memref-expand \ 11// RUN: -convert-vector-to-llvm \ 12// RUN: -convert-memref-to-llvm \ 13// RUN: -convert-func-to-llvm \ 14// RUN: -reconcile-unrealized-casts \ 15// RUN: | mlir-cpu-runner \ 16// RUN: -e entry -entry-point-result=void -O3 \ 17// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_runner_utils%shlibext \ 18// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext\ 19// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_async_runtime%shlibext \ 20// RUN: | FileCheck %s --dump-input=always 21 22// RUN: mlir-opt %s \ 23// RUN: -async-parallel-for=async-dispatch=false \ 24// RUN: -async-to-async-runtime \ 25// RUN: -async-runtime-ref-counting \ 26// RUN: -async-runtime-ref-counting-opt \ 27// RUN: -convert-async-to-llvm \ 28// RUN: -convert-linalg-to-loops \ 29// RUN: -convert-scf-to-cf \ 30// RUN: -arith-expand \ 31// RUN: -memref-expand \ 32// RUN: -convert-vector-to-llvm \ 33// RUN: -convert-memref-to-llvm \ 34// RUN: -convert-func-to-llvm \ 35// RUN: -reconcile-unrealized-casts \ 36// RUN: | mlir-cpu-runner \ 37// RUN: -e entry -entry-point-result=void -O3 \ 38// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_runner_utils%shlibext \ 39// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext\ 40// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_async_runtime%shlibext \ 41// RUN: | FileCheck %s --dump-input=always 42 43// RUN: mlir-opt %s \ 44// RUN: -convert-linalg-to-loops \ 45// RUN: -convert-scf-to-cf \ 46// RUN: -convert-vector-to-llvm \ 47// RUN: -convert-memref-to-llvm \ 48// RUN: -convert-func-to-llvm \ 49// RUN: -reconcile-unrealized-casts \ 50// RUN: | mlir-cpu-runner \ 51// RUN: -e entry -entry-point-result=void -O3 \ 52// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_runner_utils%shlibext \ 53// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext\ 54// RUN: -shared-libs=%mlir_integration_test_dir/libmlir_async_runtime%shlibext \ 55// RUN: | FileCheck %s --dump-input=always 56 57#map0 = affine_map<(d0, d1) -> (d0, d1)> 58 59func @scf_parallel(%lhs: memref<?x?xf32>, 60 %rhs: memref<?x?xf32>, 61 %sum: memref<?x?xf32>) { 62 %c0 = arith.constant 0 : index 63 %c1 = arith.constant 1 : index 64 65 %d0 = memref.dim %lhs, %c0 : memref<?x?xf32> 66 %d1 = memref.dim %lhs, %c1 : memref<?x?xf32> 67 68 scf.parallel (%i, %j) = (%c0, %c0) to (%d0, %d1) step (%c1, %c1) { 69 %lv = memref.load %lhs[%i, %j] : memref<?x?xf32> 70 %rv = memref.load %lhs[%i, %j] : memref<?x?xf32> 71 %r = arith.addf %lv, %rv : f32 72 memref.store %r, %sum[%i, %j] : memref<?x?xf32> 73 } 74 75 return 76} 77 78func @entry() { 79 %f1 = arith.constant 1.0 : f32 80 %f4 = arith.constant 4.0 : f32 81 %c0 = arith.constant 0 : index 82 %c1 = arith.constant 1 : index 83 %cN = arith.constant 50 : index 84 85 // 86 // Sanity check for the function under test. 87 // 88 89 %LHS10 = memref.alloc() {alignment = 64} : memref<1x10xf32> 90 %RHS10 = memref.alloc() {alignment = 64} : memref<1x10xf32> 91 %DST10 = memref.alloc() {alignment = 64} : memref<1x10xf32> 92 93 linalg.fill ins(%f1 : f32) outs(%LHS10 : memref<1x10xf32>) 94 linalg.fill ins(%f1 : f32) outs(%RHS10 : memref<1x10xf32>) 95 96 %LHS = memref.cast %LHS10 : memref<1x10xf32> to memref<?x?xf32> 97 %RHS = memref.cast %RHS10 : memref<1x10xf32> to memref<?x?xf32> 98 %DST = memref.cast %DST10 : memref<1x10xf32> to memref<?x?xf32> 99 100 call @scf_parallel(%LHS, %RHS, %DST) 101 : (memref<?x?xf32>, memref<?x?xf32>, memref<?x?xf32>) -> () 102 103 // CHECK: [2, 2, 2, 2, 2, 2, 2, 2, 2, 2] 104 %U = memref.cast %DST10 : memref<1x10xf32> to memref<*xf32> 105 call @print_memref_f32(%U): (memref<*xf32>) -> () 106 107 memref.dealloc %LHS10: memref<1x10xf32> 108 memref.dealloc %RHS10: memref<1x10xf32> 109 memref.dealloc %DST10: memref<1x10xf32> 110 111 // 112 // Allocate data for microbenchmarks. 113 // 114 115 %LHS1024 = memref.alloc() {alignment = 64} : memref<1024x1024xf32> 116 %RHS1024 = memref.alloc() {alignment = 64} : memref<1024x1024xf32> 117 %DST1024 = memref.alloc() {alignment = 64} : memref<1024x1024xf32> 118 119 %LHS0 = memref.cast %LHS1024 : memref<1024x1024xf32> to memref<?x?xf32> 120 %RHS0 = memref.cast %RHS1024 : memref<1024x1024xf32> to memref<?x?xf32> 121 %DST0 = memref.cast %DST1024 : memref<1024x1024xf32> to memref<?x?xf32> 122 123 // 124 // Warm up. 125 // 126 127 call @scf_parallel(%LHS0, %RHS0, %DST0) 128 : (memref<?x?xf32>, memref<?x?xf32>, memref<?x?xf32>) -> () 129 130 // 131 // Measure execution time. 132 // 133 134 %t0 = call @rtclock() : () -> f64 135 scf.for %i = %c0 to %cN step %c1 { 136 call @scf_parallel(%LHS0, %RHS0, %DST0) 137 : (memref<?x?xf32>, memref<?x?xf32>, memref<?x?xf32>) -> () 138 } 139 %t1 = call @rtclock() : () -> f64 140 %t1024 = arith.subf %t1, %t0 : f64 141 142 // Print timings. 143 vector.print %t1024 : f64 144 145 // Free. 146 memref.dealloc %LHS1024: memref<1024x1024xf32> 147 memref.dealloc %RHS1024: memref<1024x1024xf32> 148 memref.dealloc %DST1024: memref<1024x1024xf32> 149 150 return 151} 152 153func private @rtclock() -> f64 154 155func private @print_memref_f32(memref<*xf32>) 156 attributes { llvm.emit_c_interface } 157