Hans Boehm | 3f55787 | 2017-01-23 17:30:44 -0800 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (C) 2017 The Android Open Source Project |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | // Our goal is to measure the cost of various C++ atomic operations. |
| 18 | // Android doesn't really control those. But since some of these operations can be quite |
| 19 | // expensive, this may be useful input for development of higher level code. |
| 20 | // Expected mappings from C++ atomics to hardware primitives can be found at |
| 21 | // http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html . |
| 22 | |
| 23 | #include <benchmark/benchmark.h> |
| 24 | #include <atomic> |
| 25 | #include <mutex> |
| 26 | |
| 27 | // We time atomic operations separated by a volatile (not atomic!) increment. This ensures |
| 28 | // that the compiler emits memory instructions (e.g. load or store) prior to any fence or the |
| 29 | // like. That in turn ensures that the CPU has outstanding memory operations when the fence |
| 30 | // is executed. |
| 31 | |
| 32 | // In most respects, we compute best case values. Since there is only one thread, there are no |
| 33 | // coherence misses. |
| 34 | |
| 35 | // We assume that the compiler is not smart enough to optimize away fences in a single-threaded |
| 36 | // program. If that changes, we'll need to add a second thread. |
| 37 | |
| 38 | volatile unsigned counter; |
| 39 | |
| 40 | std::atomic<int> test_loc(0); |
| 41 | |
| 42 | volatile unsigned sink; |
| 43 | |
| 44 | std::mutex mtx; |
| 45 | |
| 46 | void BM_empty(benchmark::State& state) { |
| 47 | while (state.KeepRunning()) { |
| 48 | ++counter; |
| 49 | } |
| 50 | } |
| 51 | BENCHMARK(BM_empty); |
| 52 | |
| 53 | static void BM_load_relaxed(benchmark::State& state) { |
| 54 | unsigned result = 0; |
| 55 | while (state.KeepRunning()) { |
| 56 | result += test_loc.load(std::memory_order_relaxed); |
| 57 | ++counter; |
| 58 | } |
| 59 | sink = result; |
| 60 | } |
| 61 | BENCHMARK(BM_load_relaxed); |
| 62 | |
| 63 | static void BM_load_acquire(benchmark::State& state) { |
| 64 | unsigned result = 0; |
| 65 | while (state.KeepRunning()) { |
| 66 | result += test_loc.load(std::memory_order_acquire); |
| 67 | ++counter; |
| 68 | } |
| 69 | sink = result; |
| 70 | } |
| 71 | BENCHMARK(BM_load_acquire); |
| 72 | |
| 73 | static void BM_store_release(benchmark::State& state) { |
| 74 | int i = counter; |
| 75 | while (state.KeepRunning()) { |
| 76 | test_loc.store(++i, std::memory_order_release); |
| 77 | ++counter; |
| 78 | } |
| 79 | } |
| 80 | BENCHMARK(BM_store_release); |
| 81 | |
| 82 | static void BM_store_seq_cst(benchmark::State& state) { |
| 83 | int i = counter; |
| 84 | while (state.KeepRunning()) { |
| 85 | test_loc.store(++i, std::memory_order_seq_cst); |
| 86 | ++counter; |
| 87 | } |
| 88 | } |
| 89 | BENCHMARK(BM_store_seq_cst); |
| 90 | |
| 91 | static void BM_fetch_add_relaxed(benchmark::State& state) { |
| 92 | unsigned result = 0; |
| 93 | while (state.KeepRunning()) { |
| 94 | result += test_loc.fetch_add(1, std::memory_order_relaxed); |
| 95 | ++counter; |
| 96 | } |
| 97 | sink = result; |
| 98 | } |
| 99 | BENCHMARK(BM_fetch_add_relaxed); |
| 100 | |
| 101 | static void BM_fetch_add_seq_cst(benchmark::State& state) { |
| 102 | unsigned result = 0; |
| 103 | while (state.KeepRunning()) { |
| 104 | result += test_loc.fetch_add(1, std::memory_order_seq_cst); |
| 105 | ++counter; |
| 106 | } |
| 107 | sink = result; |
| 108 | } |
| 109 | BENCHMARK(BM_fetch_add_seq_cst); |
| 110 | |
| 111 | // The fence benchmarks include a relaxed load to make it much harder to optimize away |
| 112 | // the fence. |
| 113 | |
| 114 | static void BM_acquire_fence(benchmark::State& state) { |
| 115 | unsigned result = 0; |
| 116 | while (state.KeepRunning()) { |
| 117 | result += test_loc.load(std::memory_order_relaxed); |
| 118 | std::atomic_thread_fence(std::memory_order_acquire); |
| 119 | ++counter; |
| 120 | } |
| 121 | sink = result; |
| 122 | } |
| 123 | BENCHMARK(BM_acquire_fence); |
| 124 | |
| 125 | static void BM_seq_cst_fence(benchmark::State& state) { |
| 126 | unsigned result = 0; |
| 127 | while (state.KeepRunning()) { |
| 128 | result += test_loc.load(std::memory_order_relaxed); |
| 129 | std::atomic_thread_fence(std::memory_order_seq_cst); |
| 130 | ++counter; |
| 131 | } |
| 132 | sink = result; |
| 133 | } |
| 134 | BENCHMARK(BM_seq_cst_fence); |
| 135 | |
| 136 | // For comparison, also throw in a critical section version: |
| 137 | |
| 138 | static void BM_fetch_add_cs(benchmark::State& state) { |
| 139 | unsigned result = 0; |
| 140 | while (state.KeepRunning()) { |
| 141 | { |
| 142 | std::lock_guard<std::mutex> _(mtx); |
| 143 | result += ++counter; |
| 144 | } |
| 145 | } |
| 146 | sink = result; |
| 147 | } |
| 148 | BENCHMARK(BM_fetch_add_cs); |