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/dwconv.h"
18 #include "bench/utils.h"
19 #include <xnnpack/AlignedAllocator.h>
20 #include <xnnpack/common.h>
21 #include <xnnpack/dwconv.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
DWConvBenchmark(benchmark::State & state,xnn_f16_dwconv_minmax_unipass_ukernel_function dwconv,xnn_init_f16_minmax_params_fn init_params,uint32_t cr,uint32_t kr,benchmark::utils::IsaCheckFunction isa_check=nullptr)29 static void DWConvBenchmark(benchmark::State& state,
30 xnn_f16_dwconv_minmax_unipass_ukernel_function dwconv,
31 xnn_init_f16_minmax_params_fn init_params,
32 uint32_t cr, uint32_t kr,
33 benchmark::utils::IsaCheckFunction isa_check = nullptr)
34 {
35 if (!cpuinfo_initialize()) {
36 state.SkipWithError("cpuinfo initialization failed");
37 return;
38 }
39 if (isa_check && !isa_check(state)) {
40 return;
41 }
42
43 const size_t input_height = state.range(0);
44 const size_t input_width = state.range(1);
45 const size_t kernel_height = state.range(2);
46 const size_t kernel_width = state.range(3);
47 const size_t padding_height = state.range(4);
48 const size_t padding_width = state.range(5);
49 const size_t subsampling = state.range(6);
50 const size_t dilation = state.range(7);
51 const size_t channels = state.range(8);
52
53 const size_t kernel_size = kernel_height * kernel_width;
54 if (kernel_size != kr) {
55 state.SkipWithError("kernel size mismatch");
56 return;
57 }
58
59 std::random_device random_device;
60 auto rng = std::mt19937(random_device());
61 auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), std::ref(rng));
62 auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
63
64 const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
65 const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
66 const size_t padding_left = padding_width / 2;
67 const size_t padding_top = padding_height / 2;
68 const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
69 const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
70 const size_t output_size = output_height * output_width;
71 const size_t step_width = dilation == 1 ? subsampling : kernel_width;
72 const size_t step_height = kernel_size + (output_width - 1) * step_width * kernel_height;
73
74 const size_t c_stride = benchmark::utils::RoundUp<size_t>(channels, cr);
75
76 std::vector<uint16_t> a(channels * input_height * input_width + XNN_EXTRA_BYTES / sizeof(uint16_t));
77 std::generate(a.begin(), a.end(), std::ref(f16rng));
78 std::vector<uint16_t> k(channels * kernel_height * kernel_width);
79 std::generate(k.begin(), k.end(), std::ref(f16rng));
80 std::vector<uint16_t> b(channels);
81 std::generate(b.begin(), b.end(), std::ref(f16rng));
82
83 std::vector<uint16_t> z(channels + XNN_EXTRA_BYTES / sizeof(uint16_t));
84
85 const size_t w_elements = (kernel_size + 1) * c_stride;
86 const size_t i_elements = output_height * step_height;
87 const size_t c_elements = output_size * channels;
88 const size_t num_buffers = 1 +
89 benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
90 sizeof(uint16_t) * (w_elements + c_elements) + sizeof(void*) * i_elements);
91
92 std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> w(w_elements * num_buffers);
93 std::fill(w.begin(), w.end(), 0.0f);
94 xnn_pack_f16_dwconv_ghw_w(kernel_height, kernel_width, channels, cr,
95 k.data(), b.data(), w.data(), 0 /* extra bytes */, nullptr);
96 for (size_t n = 1; n < num_buffers; n++) {
97 std::copy(w.cbegin(), w.cbegin() + w_elements, w.begin() + n * w_elements);
98 }
99
100 std::vector<const uint16_t*> i(i_elements * num_buffers);
101 xnn_operator convolution_op = { };
102 convolution_op.indirection_buffer = reinterpret_cast<const void**>(i.data());
103 convolution_op.input = a.data();
104 convolution_op.input_pixel_stride = channels;
105 convolution_op.zero_buffer = z.data();
106 convolution_op.input_height = input_height;
107 convolution_op.input_width = input_width;
108 convolution_op.output_height = output_height;
109 convolution_op.output_width = output_width;
110 convolution_op.kernel_height = kernel_height;
111 convolution_op.kernel_width = kernel_width;
112 convolution_op.stride_height = subsampling;
113 convolution_op.stride_width = subsampling;
114 convolution_op.dilation_height = dilation;
115 convolution_op.dilation_width = dilation;
116 convolution_op.padding_top = padding_top;
117 convolution_op.padding_left = padding_left;
118
119 xnn_indirection_init_dwconv2d(&convolution_op, step_height, step_width, 1 /* log2(sizeof(uint16_t)) */);
120 for (size_t n = 1; n < num_buffers; n++) {
121 std::copy(i.cbegin(), i.cbegin() + i_elements, i.begin() + n * i_elements);
122 }
123
124 std::vector<uint16_t> c(c_elements * num_buffers);
125 std::fill(c.begin(), c.end(), std::nanf(""));
126
127 xnn_f16_minmax_params params;
128 init_params(¶ms, UINT16_C(0xFC00) /* -inf */, UINT16_C(0x7C00) /* inf */);
129
130 size_t buffer_index = 0;
131 for (auto _ : state) {
132 state.PauseTiming();
133 benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(uint16_t));
134 buffer_index = (buffer_index + 1) % num_buffers;
135 state.ResumeTiming();
136
137 for (size_t y = 0; y < output_height; y++) {
138 dwconv(channels, output_width,
139 reinterpret_cast<const void**>(i.data() + buffer_index * i_elements + step_height * y),
140 w.data() + buffer_index * w_elements,
141 c.data() + buffer_index * c_elements + y * output_width * channels,
142 kernel_height * step_width * sizeof(void*), 0,
143 0, z.data(), ¶ms);
144 }
145 }
146
147 const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
148 if (cpu_frequency != 0) {
149 state.counters["cpufreq"] = cpu_frequency;
150 }
151
152 state.counters["FLOPS"] = benchmark::Counter(
153 uint64_t(state.iterations()) * 2 * output_size * channels * kernel_size, benchmark::Counter::kIsRate);
154
155 state.counters["bytes"] = benchmark::Counter(
156 uint64_t(state.iterations()) * (output_size + input_height * input_width + kernel_size + 1 /* bias */) * channels * sizeof(uint16_t),
157 benchmark::Counter::kIsRate);
158 }
159
160 #if XNN_ARCH_ARM64
f16_dwconv_8x4__neonfp16arith_acc2(benchmark::State & state,const char * net)161 static void f16_dwconv_8x4__neonfp16arith_acc2(benchmark::State& state, const char* net) {
162 DWConvBenchmark(state,
163 xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith_acc2,
164 xnn_init_f16_minmax_neon_params,
165 8, 4, benchmark::utils::CheckNEONFP16ARITH);
166 }
167
f16_dwconv_8x4__neonfp16arith(benchmark::State & state,const char * net)168 static void f16_dwconv_8x4__neonfp16arith(benchmark::State& state, const char* net) {
169 DWConvBenchmark(state,
170 xnn_f16_dwconv_minmax_ukernel_up8x4__neonfp16arith,
171 xnn_init_f16_minmax_neon_params,
172 8, 4, benchmark::utils::CheckNEONFP16ARITH);
173 }
174
f16_dwconv_8x9__neonfp16arith_acc2(benchmark::State & state,const char * net)175 static void f16_dwconv_8x9__neonfp16arith_acc2(benchmark::State& state, const char* net) {
176 DWConvBenchmark(state,
177 xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith_acc2,
178 xnn_init_f16_minmax_neon_params,
179 8, 9, benchmark::utils::CheckNEONFP16ARITH);
180 }
181
f16_dwconv_8x9__neonfp16arith(benchmark::State & state,const char * net)182 static void f16_dwconv_8x9__neonfp16arith(benchmark::State& state, const char* net) {
183 DWConvBenchmark(state,
184 xnn_f16_dwconv_minmax_ukernel_up8x9__neonfp16arith,
185 xnn_init_f16_minmax_neon_params,
186 8, 9, benchmark::utils::CheckNEONFP16ARITH);
187 }
188
f16_dwconv_8x25__neonfp16arith_acc2(benchmark::State & state,const char * net)189 static void f16_dwconv_8x25__neonfp16arith_acc2(benchmark::State& state, const char* net) {
190 DWConvBenchmark(state,
191 xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2,
192 xnn_init_f16_minmax_neon_params,
193 8, 25, benchmark::utils::CheckNEONFP16ARITH);
194 }
195
f16_dwconv_8x25__neonfp16arith(benchmark::State & state,const char * net)196 static void f16_dwconv_8x25__neonfp16arith(benchmark::State& state, const char* net) {
197 DWConvBenchmark(state,
198 xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith,
199 xnn_init_f16_minmax_neon_params,
200 8, 25, benchmark::utils::CheckNEONFP16ARITH);
201 }
202
f16_dwconv_16x4__neonfp16arith_acc2(benchmark::State & state,const char * net)203 static void f16_dwconv_16x4__neonfp16arith_acc2(benchmark::State& state, const char* net) {
204 DWConvBenchmark(state,
205 xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith_acc2,
206 xnn_init_f16_minmax_neon_params,
207 16, 4, benchmark::utils::CheckNEONFP16ARITH);
208 }
209
f16_dwconv_16x4__neonfp16arith(benchmark::State & state,const char * net)210 static void f16_dwconv_16x4__neonfp16arith(benchmark::State& state, const char* net) {
211 DWConvBenchmark(state,
212 xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith,
213 xnn_init_f16_minmax_neon_params,
214 16, 4, benchmark::utils::CheckNEONFP16ARITH);
215 }
216
f16_dwconv_16x9__neonfp16arith_acc2(benchmark::State & state,const char * net)217 static void f16_dwconv_16x9__neonfp16arith_acc2(benchmark::State& state, const char* net) {
218 DWConvBenchmark(state,
219 xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith_acc2,
220 xnn_init_f16_minmax_neon_params,
221 16, 9, benchmark::utils::CheckNEONFP16ARITH);
222 }
223
f16_dwconv_16x9__neonfp16arith(benchmark::State & state,const char * net)224 static void f16_dwconv_16x9__neonfp16arith(benchmark::State& state, const char* net) {
225 DWConvBenchmark(state,
226 xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith,
227 xnn_init_f16_minmax_neon_params,
228 16, 9, benchmark::utils::CheckNEONFP16ARITH);
229 }
230
f16_dwconv_16x25__neonfp16arith_acc2(benchmark::State & state,const char * net)231 static void f16_dwconv_16x25__neonfp16arith_acc2(benchmark::State& state, const char* net) {
232 DWConvBenchmark(state,
233 xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2,
234 xnn_init_f16_minmax_neon_params,
235 16, 25, benchmark::utils::CheckNEONFP16ARITH);
236 }
237
f16_dwconv_16x25__neonfp16arith(benchmark::State & state,const char * net)238 static void f16_dwconv_16x25__neonfp16arith(benchmark::State& state, const char* net) {
239 DWConvBenchmark(state,
240 xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith,
241 xnn_init_f16_minmax_neon_params,
242 16, 25, benchmark::utils::CheckNEONFP16ARITH);
243 }
244
f16_dwconv_32x4__neonfp16arith_acc2(benchmark::State & state,const char * net)245 static void f16_dwconv_32x4__neonfp16arith_acc2(benchmark::State& state, const char* net) {
246 DWConvBenchmark(state,
247 xnn_f16_dwconv_minmax_ukernel_up32x4__neonfp16arith_acc2,
248 xnn_init_f16_minmax_neon_params,
249 32, 4, benchmark::utils::CheckNEONFP16ARITH);
250 }
251
f16_dwconv_32x4__neonfp16arith(benchmark::State & state,const char * net)252 static void f16_dwconv_32x4__neonfp16arith(benchmark::State& state, const char* net) {
253 DWConvBenchmark(state,
254 xnn_f16_dwconv_minmax_ukernel_up32x4__neonfp16arith,
255 xnn_init_f16_minmax_neon_params,
256 32, 4, benchmark::utils::CheckNEONFP16ARITH);
257 }
258
f16_dwconv_32x9__neonfp16arith_acc2(benchmark::State & state,const char * net)259 static void f16_dwconv_32x9__neonfp16arith_acc2(benchmark::State& state, const char* net) {
260 DWConvBenchmark(state,
261 xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith_acc2,
262 xnn_init_f16_minmax_neon_params,
263 32, 9, benchmark::utils::CheckNEONFP16ARITH);
264 }
265
f16_dwconv_32x9__neonfp16arith(benchmark::State & state,const char * net)266 static void f16_dwconv_32x9__neonfp16arith(benchmark::State& state, const char* net) {
267 DWConvBenchmark(state,
268 xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith,
269 xnn_init_f16_minmax_neon_params,
270 32, 9, benchmark::utils::CheckNEONFP16ARITH);
271 }
272
f16_dwconv_32x25__neonfp16arith_acc2(benchmark::State & state,const char * net)273 static void f16_dwconv_32x25__neonfp16arith_acc2(benchmark::State& state, const char* net) {
274 DWConvBenchmark(state,
275 xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2,
276 xnn_init_f16_minmax_neon_params,
277 32, 25, benchmark::utils::CheckNEONFP16ARITH);
278 }
279
f16_dwconv_32x25__neonfp16arith(benchmark::State & state,const char * net)280 static void f16_dwconv_32x25__neonfp16arith(benchmark::State& state, const char* net) {
281 DWConvBenchmark(state,
282 xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith,
283 xnn_init_f16_minmax_neon_params,
284 32, 25, benchmark::utils::CheckNEONFP16ARITH);
285 }
286
287 BENCHMARK_DWCONV(f16_dwconv_8x4__neonfp16arith_acc2)
288 BENCHMARK_DWCONV(f16_dwconv_8x4__neonfp16arith)
289 BENCHMARK_DWCONV(f16_dwconv_8x9__neonfp16arith_acc2)
290 BENCHMARK_DWCONV(f16_dwconv_8x9__neonfp16arith)
291 BENCHMARK_DWCONV(f16_dwconv_8x25__neonfp16arith_acc2)
292 BENCHMARK_DWCONV(f16_dwconv_8x25__neonfp16arith)
293 BENCHMARK_DWCONV(f16_dwconv_16x4__neonfp16arith_acc2)
294 BENCHMARK_DWCONV(f16_dwconv_16x4__neonfp16arith)
295 BENCHMARK_DWCONV(f16_dwconv_16x9__neonfp16arith_acc2)
296 BENCHMARK_DWCONV(f16_dwconv_16x9__neonfp16arith)
297 BENCHMARK_DWCONV(f16_dwconv_16x25__neonfp16arith_acc2)
298 BENCHMARK_DWCONV(f16_dwconv_16x25__neonfp16arith)
299 BENCHMARK_DWCONV(f16_dwconv_32x4__neonfp16arith_acc2)
300 BENCHMARK_DWCONV(f16_dwconv_32x4__neonfp16arith)
301 BENCHMARK_DWCONV(f16_dwconv_32x9__neonfp16arith_acc2)
302 BENCHMARK_DWCONV(f16_dwconv_32x9__neonfp16arith)
303 BENCHMARK_DWCONV(f16_dwconv_32x25__neonfp16arith_acc2)
304 BENCHMARK_DWCONV(f16_dwconv_32x25__neonfp16arith)
305 #endif // XNN_ARCH_ARM64
306
307 #ifndef XNNPACK_BENCHMARK_NO_MAIN
308 BENCHMARK_MAIN();
309 #endif
310