1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2022 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker //
3*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker
6*4bdc9457SAndroid Build Coastguard Worker #include <algorithm>
7*4bdc9457SAndroid Build Coastguard Worker #include <cmath>
8*4bdc9457SAndroid Build Coastguard Worker #include <functional>
9*4bdc9457SAndroid Build Coastguard Worker #include <numeric>
10*4bdc9457SAndroid Build Coastguard Worker #include <vector>
11*4bdc9457SAndroid Build Coastguard Worker
12*4bdc9457SAndroid Build Coastguard Worker #include <benchmark/benchmark.h>
13*4bdc9457SAndroid Build Coastguard Worker #include "bench/utils.h"
14*4bdc9457SAndroid Build Coastguard Worker
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack.h>
16*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/aligned-allocator.h>
17*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
18*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/microfnptr.h>
19*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/transpose.h>
20*4bdc9457SAndroid Build Coastguard Worker
21*4bdc9457SAndroid Build Coastguard Worker
transpose(benchmark::State & state,xnn_transposev_ukernel_function transpose,benchmark::utils::IsaCheckFunction isa_check=nullptr)22*4bdc9457SAndroid Build Coastguard Worker void transpose(
23*4bdc9457SAndroid Build Coastguard Worker benchmark::State& state,
24*4bdc9457SAndroid Build Coastguard Worker xnn_transposev_ukernel_function transpose,
25*4bdc9457SAndroid Build Coastguard Worker benchmark::utils::IsaCheckFunction isa_check = nullptr)
26*4bdc9457SAndroid Build Coastguard Worker {
27*4bdc9457SAndroid Build Coastguard Worker if (isa_check && !isa_check(state)) {
28*4bdc9457SAndroid Build Coastguard Worker return;
29*4bdc9457SAndroid Build Coastguard Worker }
30*4bdc9457SAndroid Build Coastguard Worker const size_t height = state.range(0);
31*4bdc9457SAndroid Build Coastguard Worker const size_t width = state.range(1);
32*4bdc9457SAndroid Build Coastguard Worker const size_t element_size = state.range(2);
33*4bdc9457SAndroid Build Coastguard Worker const size_t tile_hbytes = height * 3;
34*4bdc9457SAndroid Build Coastguard Worker const size_t tile_wbytes = width * 3;
35*4bdc9457SAndroid Build Coastguard Worker
36*4bdc9457SAndroid Build Coastguard Worker std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> x(
37*4bdc9457SAndroid Build Coastguard Worker height * width * element_size + XNN_EXTRA_BYTES / sizeof(uint8_t));
38*4bdc9457SAndroid Build Coastguard Worker std::vector<uint8_t, AlignedAllocator<uint8_t, 64>> y(
39*4bdc9457SAndroid Build Coastguard Worker height * width * element_size + XNN_EXTRA_BYTES / sizeof(uint8_t));
40*4bdc9457SAndroid Build Coastguard Worker std::iota(x.begin(), x.end(), 0);
41*4bdc9457SAndroid Build Coastguard Worker std::fill(y.begin(), y.end(), 0);
42*4bdc9457SAndroid Build Coastguard Worker
43*4bdc9457SAndroid Build Coastguard Worker for (auto _ : state) {
44*4bdc9457SAndroid Build Coastguard Worker transpose(x.data(), y.data(), tile_wbytes, tile_hbytes, element_size,
45*4bdc9457SAndroid Build Coastguard Worker element_size, element_size, width, height);
46*4bdc9457SAndroid Build Coastguard Worker }
47*4bdc9457SAndroid Build Coastguard Worker
48*4bdc9457SAndroid Build Coastguard Worker const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
49*4bdc9457SAndroid Build Coastguard Worker if (cpu_frequency != 0) {
50*4bdc9457SAndroid Build Coastguard Worker state.counters["cpufreq"] = cpu_frequency;
51*4bdc9457SAndroid Build Coastguard Worker }
52*4bdc9457SAndroid Build Coastguard Worker }
53*4bdc9457SAndroid Build Coastguard Worker
BenchmarkKernelSize(benchmark::internal::Benchmark * b)54*4bdc9457SAndroid Build Coastguard Worker static void BenchmarkKernelSize(benchmark::internal::Benchmark* b)
55*4bdc9457SAndroid Build Coastguard Worker {
56*4bdc9457SAndroid Build Coastguard Worker b->ArgNames({"height", "width", "element_size"});
57*4bdc9457SAndroid Build Coastguard Worker //b->Args({32, 32, 3});
58*4bdc9457SAndroid Build Coastguard Worker //b->Args({64, 64, 3});
59*4bdc9457SAndroid Build Coastguard Worker //b->Args({117, 117, 3});
60*4bdc9457SAndroid Build Coastguard Worker //b->Args({1024, 1024, 3});
61*4bdc9457SAndroid Build Coastguard Worker b->Args({49153, 8, 128});
62*4bdc9457SAndroid Build Coastguard Worker }
63*4bdc9457SAndroid Build Coastguard Worker
64*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_CAPTURE(transpose, 1x1_mmemcpy, xnn_xx_transposev_ukernel__1x1_memcpy)
65*4bdc9457SAndroid Build Coastguard Worker ->Apply(BenchmarkKernelSize)->UseRealTime();
66*4bdc9457SAndroid Build Coastguard Worker
67*4bdc9457SAndroid Build Coastguard Worker #ifndef XNNPACK_BENCHMARK_NO_MAIN
68*4bdc9457SAndroid Build Coastguard Worker BENCHMARK_MAIN();
69*4bdc9457SAndroid Build Coastguard Worker #endif
70