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/conv.h"
18 #include "bench/utils.h"
19 #include <xnnpack/AlignedAllocator.h>
20 #include <xnnpack/common.h>
21 #include <xnnpack/igemm.h>
22 #include <xnnpack/indirection.h>
23 #include <xnnpack/operator.h>
24 #include <xnnpack/pack.h>
25 #include <xnnpack/params-init.h>
26 #include <xnnpack/params.h>
27
28
IGEMMBenchmark(benchmark::State & state,xnn_f16_igemm_minmax_ukernel_function f16_igemm,uint32_t mr,uint32_t nr,uint32_t kr,uint32_t sr,xnn_init_f16_scaleminmax_params_fn init_params,benchmark::utils::IsaCheckFunction isa_check=nullptr)29 static void IGEMMBenchmark(benchmark::State& state,
30 xnn_f16_igemm_minmax_ukernel_function f16_igemm,
31 uint32_t mr, uint32_t nr, uint32_t kr, uint32_t sr,
32 xnn_init_f16_scaleminmax_params_fn init_params,
33 benchmark::utils::IsaCheckFunction isa_check = nullptr)
34 {
35 if (isa_check && !isa_check(state)) {
36 return;
37 }
38
39 const size_t input_height = state.range(0);
40 const size_t input_width = state.range(1);
41 const size_t kernel_height = state.range(2);
42 const size_t kernel_width = state.range(3);
43 const size_t kernel_size = kernel_height * kernel_width;
44 const size_t padding_height = state.range(4);
45 const size_t padding_width = state.range(5);
46 const size_t subsampling = state.range(6);
47 const size_t dilation = state.range(7);
48 const size_t group_input_channels = state.range(8);
49 const size_t group_output_channels = state.range(9);
50
51 std::random_device random_device;
52 auto rng = std::mt19937(random_device());
53 auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng));
54 auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
55
56 const size_t output_pixel_stride = group_output_channels;
57 const size_t input_pixel_stride = group_input_channels;
58 const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
59 const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
60 const size_t padding_left = padding_width / 2;
61 const size_t padding_top = padding_height / 2;
62 const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
63 const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
64 const size_t output_size = output_height * output_width;
65
66 const size_t mc_stride = benchmark::utils::RoundUp<size_t>(output_size, mr);
67 const size_t nc_stride = benchmark::utils::RoundUp<size_t>(group_output_channels, nr);
68 const size_t kc_stride = benchmark::utils::RoundUp<size_t>(group_input_channels, kr * sr);
69
70 std::vector<uint16_t> a(input_height * input_width * input_pixel_stride);
71 std::generate(a.begin(), a.end(), std::ref(f16rng));
72 std::vector<uint16_t> k(group_output_channels * kernel_height * kernel_width * group_input_channels);
73 std::generate(k.begin(), k.end(), std::ref(f16rng));
74 std::vector<uint16_t> b(group_output_channels);
75 std::generate(b.begin(), b.end(), std::ref(f16rng));
76
77 std::vector<uint16_t> z(group_input_channels);
78
79 const size_t w_elements = (kernel_size * kc_stride + 1) * nc_stride;
80 const size_t i_elements = mc_stride * kernel_size;
81 const size_t c_elements = output_height * output_width * output_pixel_stride;
82 const size_t num_buffers = 1 +
83 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
84 sizeof(uint16_t) * (w_elements + c_elements) + sizeof(void*) * i_elements);
85
86 std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> w(w_elements * num_buffers);
87 std::fill(w.begin(), w.end(), 0);
88 xnn_pack_f16_conv_goki_w(
89 1 /* groups */, group_output_channels, kernel_size, group_input_channels,
90 nr, kr, sr, k.data(), b.data(), w.data(), 0 /* extra bytes */, nullptr);
91 for (size_t n = 1; n < num_buffers; n++) {
92 std::copy(w.cbegin(), w.cbegin() + w_elements, w.begin() + n * w_elements);
93 }
94
95 std::vector<const uint16_t*> i(i_elements * num_buffers);
96 xnn_operator convolution_op = { };
97 convolution_op.indirection_buffer = reinterpret_cast<const void**>(i.data());
98 convolution_op.input = a.data();
99 convolution_op.input_pixel_stride = input_pixel_stride;
100 convolution_op.zero_buffer = z.data();
101 convolution_op.groups = 1;
102 convolution_op.group_input_channels = group_input_channels;
103 convolution_op.batch_size = 1;
104 convolution_op.input_height = input_height;
105 convolution_op.input_width = input_width;
106 convolution_op.output_height = output_height;
107 convolution_op.output_width = output_width;
108 convolution_op.kernel_height = kernel_height;
109 convolution_op.kernel_width = kernel_width;
110 convolution_op.stride_height = subsampling;
111 convolution_op.stride_width = subsampling;
112 convolution_op.dilation_height = dilation;
113 convolution_op.dilation_width = dilation;
114 convolution_op.padding_top = padding_top;
115 convolution_op.padding_left = padding_left;
116 xnn_indirection_init_conv2d(&convolution_op, mr, 1 /* log2(sizeof(uint16_t)) */);
117 for (size_t n = 1; n < num_buffers; n++) {
118 std::copy(i.cbegin(), i.cbegin() + i_elements, i.begin() + n * i_elements);
119 }
120
121 std::vector<uint16_t> c(c_elements * num_buffers);
122 std::fill(c.begin(), c.end(), std::nanf(""));
123
124 // Prepare minmax parameters.
125 xnn_f16_scaleminmax_params params;
126 init_params(¶ms,
127 UINT16_C(0x3C00) /* 1.0 */, UINT16_C(0x7C00) /* inf */, UINT16_C(0xFC00) /* -inf */);
128
129 size_t buffer_index = 0;
130 for (auto _ : state) {
131 state.PauseTiming();
132 benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint16_t));
133 buffer_index = (buffer_index + 1) % num_buffers;
134 state.ResumeTiming();
135
136 for (uint32_t m = 0; m < output_size; m += mr) {
137 const uint32_t mb = min(output_size - m, mr);
138 for (uint32_t n = 0; n < group_output_channels; n += nr) {
139 const uint32_t nb = min(group_output_channels - n, nr);
140 f16_igemm(
141 mb, nb, group_input_channels * sizeof(uint16_t), kernel_size * mr * sizeof(void*),
142 reinterpret_cast<const void**>(i.data()) + buffer_index * i_elements + m,
143 w.data() + buffer_index * w_elements + n * (kc_stride * kernel_size + 1),
144 c.data() + buffer_index * c_elements + m * group_output_channels + n, group_output_channels * sizeof(uint16_t), nr * sizeof(uint16_t),
145 0, z.data(), ¶ms);
146 }
147 }
148 }
149
150 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
151 if (cpu_frequency != 0) {
152 state.counters["cpufreq"] = cpu_frequency;
153 }
154
155 state.counters["FLOPS"] = benchmark::Counter(
156 uint64_t(state.iterations()) * 2 *
157 output_height * output_width *
158 group_input_channels * group_output_channels *
159 kernel_height * kernel_width,
160 benchmark::Counter::kIsRate);
161 }
162
163 #if XNN_ARCH_ARM64
f16_igemm_1x8__neonfp16arith_ld64(benchmark::State & state,const char * net)164 static void f16_igemm_1x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
165 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64, 1, 8, 1, 1,
166 xnn_init_f16_scaleminmax_neon_params, benchmark::utils::CheckNEONFP16ARITH);
167 }
f16_igemm_4x8__neonfp16arith_ld64(benchmark::State & state,const char * net)168 static void f16_igemm_4x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
169 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64, 4, 8, 1, 1,
170 xnn_init_f16_scaleminmax_neon_params, benchmark::utils::CheckNEONFP16ARITH);
171 }
f16_igemm_6x8__neonfp16arith_ld64(benchmark::State & state,const char * net)172 static void f16_igemm_6x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
173 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64, 6, 8, 1, 1,
174 xnn_init_f16_scaleminmax_neon_params, benchmark::utils::CheckNEONFP16ARITH);
175 }
f16_igemm_8x8__neonfp16arith_ld64(benchmark::State & state,const char * net)176 static void f16_igemm_8x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
177 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64, 8, 8, 1, 1,
178 xnn_init_f16_scaleminmax_neon_params, benchmark::utils::CheckNEONFP16ARITH);
179 }
f16_igemm_1x16__neonfp16arith_ld64(benchmark::State & state,const char * net)180 static void f16_igemm_1x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
181 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64, 1, 16, 1, 1,
182 xnn_init_f16_scaleminmax_neon_params, benchmark::utils::CheckNEONFP16ARITH);
183 }
f16_igemm_4x16__neonfp16arith_ld64(benchmark::State & state,const char * net)184 static void f16_igemm_4x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
185 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64, 4, 16, 1, 1,
186 xnn_init_f16_scaleminmax_neon_params, benchmark::utils::CheckNEONFP16ARITH);
187 }
f16_igemm_6x16__neonfp16arith_ld64(benchmark::State & state,const char * net)188 static void f16_igemm_6x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
189 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64, 6, 16, 1, 1,
190 xnn_init_f16_scaleminmax_neon_params, benchmark::utils::CheckNEONFP16ARITH);
191 }
f16_igemm_8x16__neonfp16arith_ld64(benchmark::State & state,const char * net)192 static void f16_igemm_8x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
193 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64, 8, 16, 1, 1,
194 xnn_init_f16_scaleminmax_neon_params, benchmark::utils::CheckNEONFP16ARITH);
195 }
196
197 BENCHMARK_CONV(f16_igemm_1x8__neonfp16arith_ld64)
BENCHMARK_CONV(f16_igemm_4x8__neonfp16arith_ld64)198 BENCHMARK_CONV(f16_igemm_4x8__neonfp16arith_ld64)
199 BENCHMARK_CONV(f16_igemm_6x8__neonfp16arith_ld64)
200 BENCHMARK_CONV(f16_igemm_8x8__neonfp16arith_ld64)
201 BENCHMARK_CONV(f16_igemm_1x16__neonfp16arith_ld64)
202 BENCHMARK_CONV(f16_igemm_4x16__neonfp16arith_ld64)
203 BENCHMARK_CONV(f16_igemm_6x16__neonfp16arith_ld64)
204 BENCHMARK_CONV(f16_igemm_8x16__neonfp16arith_ld64)
205 #endif // XNN_ARCH_ARM64
206
207 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
208 static void f16_igemm_1x8__avx2_broadcast(benchmark::State& state, const char* net) {
209 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_1x8__avx2_broadcast, 1, 8, 1, 1,
210 xnn_init_f16_scaleminmax_avx_params, benchmark::utils::CheckAVX2);
211 }
f16_igemm_4x8__avx2_broadcast(benchmark::State & state,const char * net)212 static void f16_igemm_4x8__avx2_broadcast(benchmark::State& state, const char* net) {
213 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_4x8__avx2_broadcast, 4, 8, 1, 1,
214 xnn_init_f16_scaleminmax_avx_params, benchmark::utils::CheckAVX2);
215 }
f16_igemm_5x8__avx2_broadcast(benchmark::State & state,const char * net)216 static void f16_igemm_5x8__avx2_broadcast(benchmark::State& state, const char* net) {
217 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_5x8__avx2_broadcast, 5, 8, 1, 1,
218 xnn_init_f16_scaleminmax_avx_params, benchmark::utils::CheckAVX2);
219 }
f16_igemm_6x8__avx2_broadcast(benchmark::State & state,const char * net)220 static void f16_igemm_6x8__avx2_broadcast(benchmark::State& state, const char* net) {
221 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_6x8__avx2_broadcast, 6, 8, 1, 1,
222 xnn_init_f16_scaleminmax_avx_params, benchmark::utils::CheckAVX2);
223 }
f16_igemm_7x8__avx2_broadcast(benchmark::State & state,const char * net)224 static void f16_igemm_7x8__avx2_broadcast(benchmark::State& state, const char* net) {
225 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_7x8__avx2_broadcast, 7, 8, 1, 1,
226 xnn_init_f16_scaleminmax_avx_params, benchmark::utils::CheckAVX2);
227 }
f16_igemm_1x16__avx2_broadcast(benchmark::State & state,const char * net)228 static void f16_igemm_1x16__avx2_broadcast(benchmark::State& state, const char* net) {
229 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_1x16__avx2_broadcast, 1, 16, 1, 1,
230 xnn_init_f16_scaleminmax_avx_params, benchmark::utils::CheckAVX2);
231 }
f16_igemm_3x16__avx2_broadcast(benchmark::State & state,const char * net)232 static void f16_igemm_3x16__avx2_broadcast(benchmark::State& state, const char* net) {
233 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_3x16__avx2_broadcast, 3, 16, 1, 1,
234 xnn_init_f16_scaleminmax_avx_params, benchmark::utils::CheckAVX2);
235 }
f16_igemm_4x16__avx2_broadcast(benchmark::State & state,const char * net)236 static void f16_igemm_4x16__avx2_broadcast(benchmark::State& state, const char* net) {
237 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_4x16__avx2_broadcast, 4, 16, 1, 1,
238 xnn_init_f16_scaleminmax_avx_params, benchmark::utils::CheckAVX2);
239 }
f16_igemm_5x16__avx2_broadcast(benchmark::State & state,const char * net)240 static void f16_igemm_5x16__avx2_broadcast(benchmark::State& state, const char* net) {
241 IGEMMBenchmark(state, xnn_f16_igemm_minmax_ukernel_5x16__avx2_broadcast, 5, 16, 1, 1,
242 xnn_init_f16_scaleminmax_avx_params, benchmark::utils::CheckAVX2);
243 }
244
245 BENCHMARK_CONV(f16_igemm_1x8__avx2_broadcast)
246 BENCHMARK_CONV(f16_igemm_4x8__avx2_broadcast)
247 BENCHMARK_CONV(f16_igemm_5x8__avx2_broadcast)
248 BENCHMARK_CONV(f16_igemm_6x8__avx2_broadcast)
249 BENCHMARK_CONV(f16_igemm_7x8__avx2_broadcast)
250 BENCHMARK_CONV(f16_igemm_1x16__avx2_broadcast)
251 BENCHMARK_CONV(f16_igemm_3x16__avx2_broadcast)
252 BENCHMARK_CONV(f16_igemm_4x16__avx2_broadcast)
253 BENCHMARK_CONV(f16_igemm_5x16__avx2_broadcast)
254 #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
255
256 #ifndef XNNPACK_BENCHMARK_NO_MAIN
257 BENCHMARK_MAIN();
258 #endif
259