1 // Copyright 2020 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <algorithm>
7 #include <cmath>
8 #include <functional>
9 #include <limits>
10 #include <random>
11 #include <vector>
12
13 #include <xnnpack.h>
14
15 #include <benchmark/benchmark.h>
16 #include "bench/utils.h"
17
18
xnnpack_truncation_f32(benchmark::State & state)19 static void xnnpack_truncation_f32(benchmark::State& state) {
20 const size_t batch_size = state.range(0);
21
22 std::random_device random_device;
23 auto rng = std::mt19937(random_device());
24 auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng));
25
26 std::vector<float> input(batch_size + XNN_EXTRA_BYTES / sizeof(float));
27 std::vector<float> output(batch_size);
28 std::generate(input.begin(), input.end(), std::ref(f32rng));
29 std::fill(output.begin(), output.end(), std::nanf(""));
30
31 xnn_status status = xnn_initialize(nullptr /* allocator */);
32 if (status != xnn_status_success) {
33 state.SkipWithError("failed to initialize XNNPACK");
34 return;
35 }
36
37 xnn_operator_t truncation_op = nullptr;
38 status = xnn_create_truncation_nc_f32(
39 1 /* channels */, 1 /* input stride */, 1 /* output stride */,
40 0 /* flags */, &truncation_op);
41 if (status != xnn_status_success || truncation_op == nullptr) {
42 state.SkipWithError("failed to create Truncation operator");
43 return;
44 }
45
46 status = xnn_setup_truncation_nc_f32(
47 truncation_op, batch_size,
48 input.data(), output.data(),
49 nullptr /* thread pool */);
50 if (status != xnn_status_success) {
51 state.SkipWithError("failed to setup Truncation operator");
52 return;
53 }
54
55 for (auto _ : state) {
56 status = xnn_run_operator(truncation_op, nullptr /* thread pool */);
57 if (status != xnn_status_success) {
58 state.SkipWithError("failed to run Truncation operator");
59 return;
60 }
61 }
62
63 status = xnn_delete_operator(truncation_op);
64 if (status != xnn_status_success) {
65 state.SkipWithError("failed to delete Truncation operator");
66 return;
67 }
68
69 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
70 if (cpu_frequency != 0) {
71 state.counters["cpufreq"] = cpu_frequency;
72 }
73
74 state.counters["elements"] =
75 benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate);
76
77 const size_t bytes_per_iteration = 2 * batch_size * sizeof(float);
78 state.counters["bytes"] =
79 benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
80 }
81
82 BENCHMARK(xnnpack_truncation_f32)
83 ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
84 ->UseRealTime();
85
86 #ifndef XNNPACK_BENCHMARK_NO_MAIN
87 BENCHMARK_MAIN();
88 #endif
89