1 // Copyright 2022 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 <numeric>
10 #include <vector>
11
12 #include "bench/utils.h"
13 #include <benchmark/benchmark.h>
14
15 #include <xnnpack.h>
16 #include <xnnpack/aligned-allocator.h>
17 #include <xnnpack/common.h>
18 #include <xnnpack/fft.h>
19 #include <xnnpack/microfnptr.h>
20 #include <xnnpack/microparams-init.h>
21
22
cs16_fftr(benchmark::State & state,xnn_cs16_fftr_ukernel_function fftr,benchmark::utils::IsaCheckFunction isa_check=nullptr)23 void cs16_fftr(
24 benchmark::State& state,
25 xnn_cs16_fftr_ukernel_function fftr,
26 benchmark::utils::IsaCheckFunction isa_check = nullptr)
27 {
28 if (isa_check && !isa_check(state)) {
29 return;
30 }
31 const size_t samples = state.range(0);
32
33 assert(samples % 2 == 0);
34 const size_t sample_size = samples * 2 + 2;
35
36 std::vector<int16_t, AlignedAllocator<int16_t, 64>> data(sample_size + XNN_EXTRA_BYTES / sizeof(int16_t));
37 std::vector<int16_t, AlignedAllocator<int16_t, 64>> twiddle(samples);
38
39 std::iota(data.begin(), data.end(), 0);
40 std::iota(twiddle.begin(), twiddle.end(), 2);
41
42 for (auto _ : state) {
43 fftr(samples, data.data(), twiddle.data());
44 }
45
46 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
47 if (cpu_frequency != 0) {
48 state.counters["cpufreq"] = cpu_frequency;
49 }
50 }
51
BenchmarkKernelSize(benchmark::internal::Benchmark * b)52 static void BenchmarkKernelSize(benchmark::internal::Benchmark* b)
53 {
54 b->ArgNames({"samples"});
55 b->Args({256});
56 b->Args({1024});
57 }
58
59 BENCHMARK_CAPTURE(cs16_fftr, cs16_scalar_x1, xnn_cs16_fftr_ukernel__scalar_x1)->Apply(BenchmarkKernelSize)->UseRealTime();
60 BENCHMARK_CAPTURE(cs16_fftr, cs16_scalar_x2, xnn_cs16_fftr_ukernel__scalar_x2)->Apply(BenchmarkKernelSize)->UseRealTime();
61 BENCHMARK_CAPTURE(cs16_fftr, cs16_scalar_x3, xnn_cs16_fftr_ukernel__scalar_x3)->Apply(BenchmarkKernelSize)->UseRealTime();
62 BENCHMARK_CAPTURE(cs16_fftr, cs16_scalar_x4, xnn_cs16_fftr_ukernel__scalar_x4)->Apply(BenchmarkKernelSize)->UseRealTime();
63
64 #ifndef XNNPACK_BENCHMARK_NO_MAIN
65 BENCHMARK_MAIN();
66 #endif
67