1 // Copyright 2019 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 <cfloat>
8 #include <cmath>
9 #include <functional>
10 #include <random>
11 #include <vector>
12
13 #include <cpuinfo.h>
14
15 #include <benchmark/benchmark.h>
16 #include <fp16/fp16.h>
17 #include "bench/dconv.h"
18 #include "bench/utils.h"
19
20 #include <xnnpack.h>
21 #include <xnnpack/aligned-allocator.h>
22 #include <xnnpack/common.h>
23 #include <xnnpack/conv.h>
24 #include <xnnpack/microfnptr.h>
25 #include <xnnpack/microparams-init.h>
26 #include <xnnpack/pack.h>
27
28
f16_conv_hwc2chw(benchmark::State & state,xnn_f16_conv_hwc2chw_ukernel_function conv,uint32_t output_channels_tile,xnn_init_f16_minmax_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)29 static void f16_conv_hwc2chw(benchmark::State& state,
30 xnn_f16_conv_hwc2chw_ukernel_function conv,
31 uint32_t output_channels_tile,
32 xnn_init_f16_minmax_params_fn init_params,
33 benchmark::utils::IsaCheckFunction isa_check = nullptr)
34 {
35 if (isa_check && !isa_check(state)) {
36 return;
37 }
38 const size_t input_height = state.range(0);
39 const size_t input_width = state.range(1);
40 const size_t output_channels = state.range(2);
41
42 std::random_device random_device;
43 auto rng = std::mt19937(random_device());
44 auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), std::ref(rng));
45 auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
46
47 const size_t input_channels = 3;
48 const size_t kernel_size = 3;
49 const size_t padding = 1;
50 const size_t subsampling = 2;
51
52 const size_t output_height = (input_height + 2 * padding - kernel_size) / subsampling + 1;
53 const size_t output_width = (input_width + 2 * padding - kernel_size) / subsampling + 1;
54
55 std::vector<uint16_t> input(input_height * input_width * input_channels + XNN_EXTRA_BYTES / sizeof(uint16_t));
56 std::generate(input.begin(), input.end(), std::ref(f16rng));
57 std::vector<uint16_t> kernel(output_channels * kernel_size * kernel_size * input_channels);
58 std::generate(kernel.begin(), kernel.end(), std::ref(f16rng));
59 std::vector<uint16_t> bias(output_channels);
60 std::generate(bias.begin(), bias.end(), std::ref(f16rng));
61
62 std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> zero(input_channels * input_width + XNN_EXTRA_BYTES / sizeof(uint16_t));
63
64 const size_t weights_elements = (kernel_size * kernel_size * input_channels + 1) *
65 benchmark::utils::RoundUp<size_t>(output_channels, output_channels_tile);
66 const size_t output_elements = output_height * output_width * output_channels;
67 const size_t num_buffers = 1 +
68 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
69 sizeof(uint16_t) * (weights_elements + output_elements));
70
71 std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> packed_weights(weights_elements * num_buffers);
72 std::fill(packed_weights.begin(), packed_weights.end(), 0.0f);
73 xnn_pack_f16_dconv_oki_w(
74 output_channels, input_channels, output_channels_tile,
75 kernel_size /* kernel height */, kernel_size /* kernel width */,
76 kernel.data(), bias.data(), packed_weights.data(), NULL);
77 for (size_t n = 1; n < num_buffers; n++) {
78 std::copy(packed_weights.cbegin(),
79 packed_weights.cbegin() + weights_elements,
80 packed_weights.begin() + n * weights_elements);
81 }
82
83 std::vector<uint16_t> output(output_elements * num_buffers);
84 std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */);
85
86 xnn_f16_minmax_params params;
87 init_params(¶ms, 0x7C00 /* inf */, 0xFC00 /* -inf */);
88
89 size_t buffer_index = 0;
90 for (auto _ : state) {
91 state.PauseTiming();
92 benchmark::utils::PrefetchToL1(input.data(), input.size() * sizeof(uint16_t));
93 buffer_index = (buffer_index + 1) % num_buffers;
94 state.ResumeTiming();
95
96 conv(
97 input_height, input_width,
98 0 /* output_y_start */, output_height /* output_y_end */,
99 input.data(), zero.data(),
100 packed_weights.data() + buffer_index * weights_elements,
101 output.data() + buffer_index * output_elements,
102 padding, output_channels,
103 output_channels * output_width * sizeof(uint16_t),
104 output_channels * sizeof(uint16_t),
105 ¶ms);
106 }
107
108 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
109 if (cpu_frequency != 0) {
110 state.counters["cpufreq"] = cpu_frequency;
111 }
112
113 state.counters["FLOPS"] = benchmark::Counter(
114 uint64_t(state.iterations()) * 2 *
115 output_height * output_width *
116 input_channels * output_channels *
117 kernel_size * kernel_size,
118 benchmark::Counter::kIsRate);
119 }
120
121 #if XNN_ENABLE_ARM_FP16 && XNN_ARCH_ARM64
f16_conv_hwc2chw_3x3s2p1c3x4__neonfp16arith_2x2(benchmark::State & state,const char * net)122 static void f16_conv_hwc2chw_3x3s2p1c3x4__neonfp16arith_2x2(benchmark::State& state, const char* net) {
123 f16_conv_hwc2chw(state, xnn_f16_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfp16arith_2x2, 4,
124 xnn_init_f16_minmax_neon_params, benchmark::utils::CheckNEONFP16ARITH);
125 }
126
127 BENCHMARK_DCONV(f16_conv_hwc2chw_3x3s2p1c3x4__neonfp16arith_2x2);
128 #endif // XNN_ENABLE_ARM_FP16 && XNN_ARCH_ARM64
129
130 #ifndef XNNPACK_BENCHMARK_NO_MAIN
131 BENCHMARK_MAIN();
132 #endif
133