142 lines · plain
1// RUN: mlir-opt %s -async-parallel-for \2// RUN: -async-to-async-runtime \3// RUN: -async-runtime-ref-counting \4// RUN: -async-runtime-ref-counting-opt \5// RUN: -convert-async-to-llvm \6// RUN: -convert-scf-to-cf \7// RUN: -finalize-memref-to-llvm \8// RUN: -arith-expand \9// RUN: -memref-expand \10// RUN: -convert-func-to-llvm \11// RUN: -convert-arith-to-llvm \12// RUN: -convert-cf-to-llvm \13// RUN: -reconcile-unrealized-casts \14// RUN: | mlir-runner \15// RUN: -e entry -entry-point-result=void -O0 \16// RUN: -shared-libs=%mlir_runner_utils \17// RUN: -shared-libs=%mlir_async_runtime\18// RUN: | FileCheck %s --dump-input=always19 20// RUN: mlir-opt %s -async-parallel-for \21// RUN: -async-to-async-runtime \22// RUN: -async-runtime-policy-based-ref-counting \23// RUN: -convert-async-to-llvm \24// RUN: -convert-scf-to-cf \25// RUN: -finalize-memref-to-llvm \26// RUN: -arith-expand \27// RUN: -memref-expand \28// RUN: -convert-func-to-llvm \29// RUN: -convert-arith-to-llvm \30// RUN: -convert-cf-to-llvm \31// RUN: -reconcile-unrealized-casts \32// RUN: | mlir-runner \33// RUN: -e entry -entry-point-result=void -O0 \34// RUN: -shared-libs=%mlir_runner_utils \35// RUN: -shared-libs=%mlir_async_runtime\36// RUN: | FileCheck %s --dump-input=always37 38// RUN: mlir-opt %s -async-parallel-for="async-dispatch=false \39// RUN: num-workers=20 \40// RUN: min-task-size=1" \41// RUN: -async-to-async-runtime \42// RUN: -async-runtime-ref-counting \43// RUN: -async-runtime-ref-counting-opt \44// RUN: -convert-async-to-llvm \45// RUN: -convert-scf-to-cf \46// RUN: -finalize-memref-to-llvm \47// RUN: -arith-expand \48// RUN: -memref-expand \49// RUN: -convert-func-to-llvm \50// RUN: -convert-arith-to-llvm \51// RUN: -convert-cf-to-llvm \52// RUN: -reconcile-unrealized-casts \53// RUN: | mlir-runner \54// RUN: -e entry -entry-point-result=void -O0 \55// RUN: -shared-libs=%mlir_runner_utils \56// RUN: -shared-libs=%mlir_async_runtime\57// RUN: | FileCheck %s --dump-input=always58 59// Suppress constant folding by introducing "dynamic" zero value at runtime.60func.func private @zero() -> index {61 %0 = arith.constant 0 : index62 return %0 : index63}64 65func.func @entry() {66 %c0 = arith.constant 0.0 : f3267 %c1 = arith.constant 1 : index68 %c2 = arith.constant 2 : index69 %c3 = arith.constant 3 : index70 71 %lb = arith.constant 0 : index72 %ub = arith.constant 9 : index73 74 %A = memref.alloc() : memref<9xf32>75 %U = memref.cast %A : memref<9xf32> to memref<*xf32>76 77 // Initialize memref with zeros because we do load and store to in every test78 // to verify that we process each element of the iteration space once.79 scf.parallel (%i) = (%lb) to (%ub) step (%c1) {80 memref.store %c0, %A[%i] : memref<9xf32>81 }82 83 // 1. %i = (0) to (9) step (1)84 scf.parallel (%i) = (%lb) to (%ub) step (%c1) {85 %0 = arith.index_cast %i : index to i3286 %1 = arith.sitofp %0 : i32 to f3287 %2 = memref.load %A[%i] : memref<9xf32>88 %3 = arith.addf %1, %2 : f3289 memref.store %3, %A[%i] : memref<9xf32>90 }91 // CHECK: [0, 1, 2, 3, 4, 5, 6, 7, 8]92 call @printMemrefF32(%U): (memref<*xf32>) -> ()93 94 scf.parallel (%i) = (%lb) to (%ub) step (%c1) {95 memref.store %c0, %A[%i] : memref<9xf32>96 }97 98 // 2. %i = (0) to (9) step (2)99 scf.parallel (%i) = (%lb) to (%ub) step (%c2) {100 %0 = arith.index_cast %i : index to i32101 %1 = arith.sitofp %0 : i32 to f32102 %2 = memref.load %A[%i] : memref<9xf32>103 %3 = arith.addf %1, %2 : f32104 memref.store %3, %A[%i] : memref<9xf32>105 }106 // CHECK: [0, 0, 2, 0, 4, 0, 6, 0, 8]107 call @printMemrefF32(%U): (memref<*xf32>) -> ()108 109 scf.parallel (%i) = (%lb) to (%ub) step (%c1) {110 memref.store %c0, %A[%i] : memref<9xf32>111 }112 113 // 3. %i = (-20) to (-11) step (3)114 %lb0 = arith.constant -20 : index115 %ub0 = arith.constant -11 : index116 scf.parallel (%i) = (%lb0) to (%ub0) step (%c3) {117 %0 = arith.index_cast %i : index to i32118 %1 = arith.sitofp %0 : i32 to f32119 %2 = arith.constant 20 : index120 %3 = arith.addi %i, %2 : index121 %4 = memref.load %A[%3] : memref<9xf32>122 %5 = arith.addf %1, %4 : f32123 memref.store %5, %A[%3] : memref<9xf32>124 }125 // CHECK: [-20, 0, 0, -17, 0, 0, -14, 0, 0]126 call @printMemrefF32(%U): (memref<*xf32>) -> ()127 128 // 4. Check that loop with zero iterations doesn't crash at runtime.129 %lb1 = call @zero(): () -> (index)130 %ub1 = call @zero(): () -> (index)131 132 scf.parallel (%i) = (%lb1) to (%ub1) step (%c1) {133 %false = arith.constant 0 : i1134 cf.assert %false, "should never be executed"135 }136 137 memref.dealloc %A : memref<9xf32>138 return139}140 141func.func private @printMemrefF32(memref<*xf32>) attributes { llvm.emit_c_interface }142