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/microfnptr.h>
19 #include <xnnpack/microparams-init.h>
20 #include <xnnpack/vsquareabs.h>
21
22
cs16_vsquareabs(benchmark::State & state,xnn_cs16_vsquareabs_ukernel_function vsquareabs,benchmark::utils::IsaCheckFunction isa_check=nullptr)23 void cs16_vsquareabs(
24 benchmark::State& state,
25 xnn_cs16_vsquareabs_ukernel_function vsquareabs,
26 benchmark::utils::IsaCheckFunction isa_check = nullptr)
27 {
28 if (isa_check && !isa_check(state)) {
29 return;
30 }
31 const size_t batch_size = state.range(0);
32
33 std::vector<int16_t, AlignedAllocator<int16_t, 64>> input(
34 batch_size * 2 + XNN_EXTRA_BYTES / sizeof(int16_t));
35 std::vector<uint32_t, AlignedAllocator<uint32_t, 64>> output(batch_size);
36 std::iota(input.begin(), input.end(), 0);
37 std::iota(output.begin(), output.end(), 0);
38
39 for (auto _ : state) {
40 vsquareabs(batch_size, input.data(), output.data());
41 }
42
43 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
44 if (cpu_frequency != 0) {
45 state.counters["cpufreq"] = cpu_frequency;
46 }
47 }
48
BenchmarkKernelSize(benchmark::internal::Benchmark * b)49 static void BenchmarkKernelSize(benchmark::internal::Benchmark* b)
50 {
51 b->ArgNames({"batch_size"});
52 b->Args({32});
53 b->Args({64});
54 b->Args({117});
55 b->Args({400});
56 b->Args({1000});
57 b->Args({10000});
58 }
59
60 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
61 BENCHMARK_CAPTURE(cs16_vsquareabs, cs16_neon_x4, xnn_cs16_vsquareabs_ukernel__neon_mlal_ld128_x4, benchmark::utils::CheckNEON)
62 ->Apply(BenchmarkKernelSize)->UseRealTime();
63 BENCHMARK_CAPTURE(cs16_vsquareabs, cs16_neon_x8, xnn_cs16_vsquareabs_ukernel__neon_mlal_ld128_x8, benchmark::utils::CheckNEON)
64 ->Apply(BenchmarkKernelSize)->UseRealTime();
65 BENCHMARK_CAPTURE(cs16_vsquareabs, cs16_neon_x12, xnn_cs16_vsquareabs_ukernel__neon_mlal_ld128_x12, benchmark::utils::CheckNEON)
66 ->Apply(BenchmarkKernelSize)->UseRealTime();
67 BENCHMARK_CAPTURE(cs16_vsquareabs, cs16_neon_x16, xnn_cs16_vsquareabs_ukernel__neon_mlal_ld128_x16, benchmark::utils::CheckNEON)
68 ->Apply(BenchmarkKernelSize)->UseRealTime();
69 #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
70
71 BENCHMARK_CAPTURE(cs16_vsquareabs, cs16_scalar_x1, xnn_cs16_vsquareabs_ukernel__scalar_x1)
72 ->Apply(BenchmarkKernelSize)->UseRealTime();
73 BENCHMARK_CAPTURE(cs16_vsquareabs, cs16_scalar_x2, xnn_cs16_vsquareabs_ukernel__scalar_x2)
74 ->Apply(BenchmarkKernelSize)->UseRealTime();
75 BENCHMARK_CAPTURE(cs16_vsquareabs, cs16_scalar_x3, xnn_cs16_vsquareabs_ukernel__scalar_x3)
76 ->Apply(BenchmarkKernelSize)->UseRealTime();
77 BENCHMARK_CAPTURE(cs16_vsquareabs, cs16_scalar_x4, xnn_cs16_vsquareabs_ukernel__scalar_x4)
78 ->Apply(BenchmarkKernelSize)->UseRealTime();
79
80 #ifndef XNNPACK_BENCHMARK_NO_MAIN
81 BENCHMARK_MAIN();
82 #endif
83