xref: /aosp_15_r20/external/XNNPACK/bench/f16-dwconv2d-chw.cc (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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 <benchmark/benchmark.h>
14 #include <fp16/fp16.h>
15 #include "bench/dwconv.h"
16 #include "bench/utils.h"
17 
18 #include <xnnpack.h>
19 #include <xnnpack/aligned-allocator.h>
20 #include <xnnpack/common.h>
21 #include <xnnpack/dwconv.h>
22 #include <xnnpack/indirection.h>
23 #include <xnnpack/microfnptr.h>
24 #include <xnnpack/microparams-init.h>
25 #include <xnnpack/operator.h>
26 #include <xnnpack/pack.h>
27 
28 
f16_dwconv2d_chw(benchmark::State & state,xnn_f16_dwconv2d_chw_ukernel_function dwconv,uint32_t kh,uint32_t kw,uint32_t pw,uint32_t s,benchmark::utils::IsaCheckFunction isa_check=nullptr)29 static void f16_dwconv2d_chw(benchmark::State& state,
30   xnn_f16_dwconv2d_chw_ukernel_function dwconv,
31   uint32_t kh, uint32_t kw, uint32_t pw, uint32_t s,
32   benchmark::utils::IsaCheckFunction isa_check = nullptr)
33 {
34   if (isa_check && !isa_check(state)) {
35     return;
36   }
37 
38   const size_t input_height = state.range(0);
39   const size_t input_width = state.range(1);
40   const size_t kernel_height = state.range(2);
41   const size_t kernel_width = state.range(3);
42   const size_t padding_height = state.range(4);
43   const size_t padding_width = state.range(5);
44   const size_t subsampling = state.range(6);
45   const size_t dilation = state.range(7);
46   const size_t channels = state.range(8);
47 
48   if (kernel_height != kh) {
49     state.SkipWithError("kernel height mismatch");
50     return;
51   }
52 
53   if (kernel_width != kw) {
54     state.SkipWithError("kernel width mismatch");
55     return;
56   }
57 
58   if (subsampling != s) {
59     state.SkipWithError("subsampling mismatch");
60     return;
61   }
62 
63   if (padding_width % 2 != 0 || padding_width / 2 != pw) {
64     state.SkipWithError("padding width mismatch");
65     return;
66   }
67 
68   if (dilation != 1) {
69     state.SkipWithError("unsupported dilation");
70     return;
71   }
72 
73   std::random_device random_device;
74   auto rng = std::mt19937(random_device());
75   auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), std::ref(rng));
76   auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);
77 
78   const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
79   const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
80   const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
81   const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
82 
83   const size_t inputSize = (input_height + padding_height) * input_width;
84   const size_t kernel_size = kernel_height * kernel_width;
85   const size_t output_size = output_height * output_width;
86 
87   std::vector<uint16_t> input(inputSize * channels + 2 * XNN_EXTRA_BYTES);
88   std::generate(input.begin(), input.end(), std::ref(f16rng));
89   std::vector<uint16_t> bias(channels);
90   std::generate(bias.begin(), bias.end(), std::ref(f16rng));
91   std::vector<uint16_t> kernel(channels * kernel_size);
92   std::generate(kernel.begin(), kernel.end(), std::ref(f16rng));
93   std::vector<uint16_t> zero(input_width + padding_width);
94 
95   const size_t w_elements = (kernel_size + 1) * channels;
96   const size_t o_elements = output_size * channels;
97   const size_t num_buffers = 1 +
98     benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
99       sizeof(uint16_t) * (w_elements + o_elements));
100 
101   std::vector<uint16_t, AlignedAllocator<uint16_t, 64>> packed_weights(w_elements * num_buffers);
102   std::fill(packed_weights.begin(), packed_weights.end(), 0.0f);
103   for (size_t c = 0; c < channels; c++) {
104     packed_weights[c * kernel_size + c] = bias[c];
105     for (size_t i = 0; i < kernel_size; i++) {
106       packed_weights[c * kernel_size + c + 1 + i] = kernel[c * kernel_size + i];
107     }
108   }
109   for (size_t n = 1; n < num_buffers; n++) {
110     std::copy(packed_weights.cbegin(), packed_weights.cbegin() + w_elements, packed_weights.begin() + n * w_elements);
111   }
112 
113   std::vector<uint16_t> output(o_elements * num_buffers);
114   std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */);
115 
116   xnn_f16_chw_params chw_params;
117   xnn_init_f16_chw_params(
118     &chw_params, input_width, 0xFC00 /* -inf */, 0x7C00 /* inf */);
119 
120   size_t buffer_index = 0;
121   for (auto _ : state) {
122     state.PauseTiming();
123     benchmark::utils::PrefetchToL1(input.data(), input.size() * sizeof(uint16_t));
124     buffer_index = (buffer_index + 1) % num_buffers;
125     state.ResumeTiming();
126 
127     for (uint32_t channel = 0; channel < channels; channel++) {
128       dwconv(
129         input_height, input_width * sizeof(uint16_t),
130         input.data() + channel * inputSize,
131         packed_weights.data() + channel * (kernel_size + 1) + buffer_index * w_elements,
132         zero.data(),
133         output.data() + channel * output_size + buffer_index * o_elements,
134         padding_height / 2,  // padding_top
135         &chw_params);
136     }
137   }
138 
139   const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
140   if (cpu_frequency != 0) {
141     state.counters["cpufreq"] = cpu_frequency;
142   }
143 
144   state.counters["FLOPS"] = benchmark::Counter(
145     uint64_t(state.iterations()) * 2 * output_size * channels * kernel_size,
146     benchmark::Counter::kIsRate);
147 
148   state.counters["bytes"] = benchmark::Counter(
149     uint64_t(state.iterations()) * (output_size + inputSize + kernel_size + 1 /* bias */) * channels * sizeof(uint16_t),
150     benchmark::Counter::kIsRate);
151 }
152 
153 
154 #if XNN_ENABLE_ARM_FP16 && XNN_ARCH_ARM64
dwconv2d_chw_3x3p1__neonfp16arith_1x8(benchmark::State & state,const char * net)155   static void dwconv2d_chw_3x3p1__neonfp16arith_1x8(benchmark::State& state, const char* net) {
156     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8, 3, 3, 1, 1);
157   }
dwconv2d_chw_3x3p1__neonfp16arith_2x8(benchmark::State & state,const char * net)158   static void dwconv2d_chw_3x3p1__neonfp16arith_2x8(benchmark::State& state, const char* net) {
159     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8, 3, 3, 1, 1);
160   }
dwconv2d_chw_3x3p1__neonfp16arith_3x8(benchmark::State & state,const char * net)161   static void dwconv2d_chw_3x3p1__neonfp16arith_3x8(benchmark::State& state, const char* net) {
162     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_3x8, 3, 3, 1, 1);
163   }
dwconv2d_chw_3x3p1__neonfp16arith_4x8(benchmark::State & state,const char * net)164   static void dwconv2d_chw_3x3p1__neonfp16arith_4x8(benchmark::State& state, const char* net) {
165     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_4x8, 3, 3, 1, 1);
166   }
dwconv2d_chw_3x3p1__neonfp16arith_5x8(benchmark::State & state,const char * net)167   static void dwconv2d_chw_3x3p1__neonfp16arith_5x8(benchmark::State& state, const char* net) {
168     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_5x8, 3, 3, 1, 1);
169   }
dwconv2d_chw_3x3p1__neonfp16arith_6x8(benchmark::State & state,const char * net)170   static void dwconv2d_chw_3x3p1__neonfp16arith_6x8(benchmark::State& state, const char* net) {
171     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_6x8, 3, 3, 1, 1);
172   }
dwconv2d_chw_3x3p1__neonfp16arith_1x8_acc2(benchmark::State & state,const char * net)173   static void dwconv2d_chw_3x3p1__neonfp16arith_1x8_acc2(benchmark::State& state, const char* net) {
174     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc2, 3, 3, 1, 1);
175   }
dwconv2d_chw_3x3p1__neonfp16arith_1x8_acc3(benchmark::State & state,const char * net)176   static void dwconv2d_chw_3x3p1__neonfp16arith_1x8_acc3(benchmark::State& state, const char* net) {
177     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc3, 3, 3, 1, 1);
178   }
dwconv2d_chw_3x3p1__neonfp16arith_1x8_acc4(benchmark::State & state,const char * net)179   static void dwconv2d_chw_3x3p1__neonfp16arith_1x8_acc4(benchmark::State& state, const char* net) {
180     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_1x8_acc4, 3, 3, 1, 1);
181   }
dwconv2d_chw_3x3p1__neonfp16arith_2x8_acc2(benchmark::State & state,const char * net)182   static void dwconv2d_chw_3x3p1__neonfp16arith_2x8_acc2(benchmark::State& state, const char* net) {
183     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8_acc2, 3, 3, 1, 1);
184   }
185 
dwconv2d_chw_3x3s2p1__neonfp16arith_1x4(benchmark::State & state,const char * net)186   static void dwconv2d_chw_3x3s2p1__neonfp16arith_1x4(benchmark::State& state, const char* net) {
187     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_1x4, 3, 3, 1, 2);
188   }
dwconv2d_chw_3x3s2p1__neonfp16arith_2x4(benchmark::State & state,const char * net)189   static void dwconv2d_chw_3x3s2p1__neonfp16arith_2x4(benchmark::State& state, const char* net) {
190     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_2x4, 3, 3, 1, 2);
191   }
dwconv2d_chw_3x3s2p1__neonfp16arith_3x4(benchmark::State & state,const char * net)192   static void dwconv2d_chw_3x3s2p1__neonfp16arith_3x4(benchmark::State& state, const char* net) {
193     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_3x4, 3, 3, 1, 2);
194   }
dwconv2d_chw_3x3s2p1__neonfp16arith_4x4(benchmark::State & state,const char * net)195   static void dwconv2d_chw_3x3s2p1__neonfp16arith_4x4(benchmark::State& state, const char* net) {
196     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_4x4, 3, 3, 1, 2);
197   }
dwconv2d_chw_3x3s2p1__neonfp16arith_1x4_acc2(benchmark::State & state,const char * net)198   static void dwconv2d_chw_3x3s2p1__neonfp16arith_1x4_acc2(benchmark::State& state, const char* net) {
199     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_1x4_acc2, 3, 3, 1, 2);
200   }
dwconv2d_chw_3x3s2p1__neonfp16arith_1x4_acc3(benchmark::State & state,const char * net)201   static void dwconv2d_chw_3x3s2p1__neonfp16arith_1x4_acc3(benchmark::State& state, const char* net) {
202     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_1x4_acc3, 3, 3, 1, 2);
203   }
dwconv2d_chw_3x3s2p1__neonfp16arith_1x4_acc4(benchmark::State & state,const char * net)204   static void dwconv2d_chw_3x3s2p1__neonfp16arith_1x4_acc4(benchmark::State& state, const char* net) {
205     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_1x4_acc4, 3, 3, 1, 2);
206   }
dwconv2d_chw_3x3s2p1__neonfp16arith_2x4_acc2(benchmark::State & state,const char * net)207   static void dwconv2d_chw_3x3s2p1__neonfp16arith_2x4_acc2(benchmark::State& state, const char* net) {
208     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_2x4_acc2, 3, 3, 1, 2);
209   }
210 
dwconv2d_chw_5x5p2__neonfp16arith_1x4(benchmark::State & state,const char * net)211   static void dwconv2d_chw_5x5p2__neonfp16arith_1x4(benchmark::State& state, const char* net) {
212     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4, 5, 5, 2, 1);
213   }
dwconv2d_chw_5x5p2__neonfp16arith_2x4(benchmark::State & state,const char * net)214   static void dwconv2d_chw_5x5p2__neonfp16arith_2x4(benchmark::State& state, const char* net) {
215     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4, 5, 5, 2, 1);
216   }
dwconv2d_chw_5x5p2__neonfp16arith_3x4(benchmark::State & state,const char * net)217   static void dwconv2d_chw_5x5p2__neonfp16arith_3x4(benchmark::State& state, const char* net) {
218     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4, 5, 5, 2, 1);
219   }
dwconv2d_chw_5x5p2__neonfp16arith_4x4(benchmark::State & state,const char * net)220   static void dwconv2d_chw_5x5p2__neonfp16arith_4x4(benchmark::State& state, const char* net) {
221     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_4x4, 5, 5, 2, 1);
222   }
dwconv2d_chw_5x5p2__neonfp16arith_5x4(benchmark::State & state,const char * net)223   static void dwconv2d_chw_5x5p2__neonfp16arith_5x4(benchmark::State& state, const char* net) {
224     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_5x4, 5, 5, 2, 1);
225   }
dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc2(benchmark::State & state,const char * net)226   static void dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc2(benchmark::State& state, const char* net) {
227     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc2, 5, 5, 2, 1);
228   }
dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc3(benchmark::State & state,const char * net)229   static void dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc3(benchmark::State& state, const char* net) {
230     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc3, 5, 5, 2, 1);
231   }
dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc4(benchmark::State & state,const char * net)232   static void dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc4(benchmark::State& state, const char* net) {
233     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc4, 5, 5, 2, 1);
234   }
dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc5(benchmark::State & state,const char * net)235   static void dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc5(benchmark::State& state, const char* net) {
236     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4_acc5, 5, 5, 2, 1);
237   }
dwconv2d_chw_5x5p2__neonfp16arith_2x4_acc2(benchmark::State & state,const char * net)238   static void dwconv2d_chw_5x5p2__neonfp16arith_2x4_acc2(benchmark::State& state, const char* net) {
239     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc2, 5, 5, 2, 1);
240   }
dwconv2d_chw_5x5p2__neonfp16arith_2x4_acc3(benchmark::State & state,const char * net)241   static void dwconv2d_chw_5x5p2__neonfp16arith_2x4_acc3(benchmark::State& state, const char* net) {
242     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_2x4_acc3, 5, 5, 2, 1);
243   }
dwconv2d_chw_5x5p2__neonfp16arith_3x4_acc2(benchmark::State & state,const char * net)244   static void dwconv2d_chw_5x5p2__neonfp16arith_3x4_acc2(benchmark::State& state, const char* net) {
245     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_3x4_acc2, 5, 5, 2, 1);
246   }
dwconv2d_chw_5x5p2__neonfp16arith_4x4_acc2(benchmark::State & state,const char * net)247   static void dwconv2d_chw_5x5p2__neonfp16arith_4x4_acc2(benchmark::State& state, const char* net) {
248     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_4x4_acc2, 5, 5, 2, 1);
249   }
250 
dwconv2d_chw_5x5s2p2__neonfp16arith_1x4(benchmark::State & state,const char * net)251   static void dwconv2d_chw_5x5s2p2__neonfp16arith_1x4(benchmark::State& state, const char* net) {
252     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_1x4, 5, 5, 2, 2);
253   }
dwconv2d_chw_5x5s2p2__neonfp16arith_2x4(benchmark::State & state,const char * net)254   static void dwconv2d_chw_5x5s2p2__neonfp16arith_2x4(benchmark::State& state, const char* net) {
255     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_2x4, 5, 5, 2, 2);
256   }
dwconv2d_chw_5x5s2p2__neonfp16arith_3x4(benchmark::State & state,const char * net)257   static void dwconv2d_chw_5x5s2p2__neonfp16arith_3x4(benchmark::State& state, const char* net) {
258     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_3x4, 5, 5, 2, 2);
259   }
dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc2(benchmark::State & state,const char * net)260   static void dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc2(benchmark::State& state, const char* net) {
261     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_1x4_acc2, 5, 5, 2, 2);
262   }
dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc3(benchmark::State & state,const char * net)263   static void dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc3(benchmark::State& state, const char* net) {
264     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_1x4_acc3, 5, 5, 2, 2);
265   }
dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc4(benchmark::State & state,const char * net)266   static void dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc4(benchmark::State& state, const char* net) {
267     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_1x4_acc4, 5, 5, 2, 2);
268   }
dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc5(benchmark::State & state,const char * net)269   static void dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc5(benchmark::State& state, const char* net) {
270     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_1x4_acc5, 5, 5, 2, 2);
271   }
dwconv2d_chw_5x5s2p2__neonfp16arith_2x4_acc2(benchmark::State & state,const char * net)272   static void dwconv2d_chw_5x5s2p2__neonfp16arith_2x4_acc2(benchmark::State& state, const char* net) {
273     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_2x4_acc2, 5, 5, 2, 2);
274   }
dwconv2d_chw_5x5s2p2__neonfp16arith_2x4_acc3(benchmark::State & state,const char * net)275   static void dwconv2d_chw_5x5s2p2__neonfp16arith_2x4_acc3(benchmark::State& state, const char* net) {
276     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_2x4_acc3, 5, 5, 2, 2);
277   }
dwconv2d_chw_5x5s2p2__neonfp16arith_3x4_acc2(benchmark::State & state,const char * net)278   static void dwconv2d_chw_5x5s2p2__neonfp16arith_3x4_acc2(benchmark::State& state, const char* net) {
279     f16_dwconv2d_chw(state, xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_3x4_acc2, 5, 5, 2, 2);
280   }
281 
282   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_1x8)
283   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_2x8)
284   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_3x8)
285   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_4x8)
286   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_5x8)
287   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_6x8)
288   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_1x8_acc2)
289   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_1x8_acc3)
290   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_1x8_acc4)
291   BENCHMARK_DWCONV(dwconv2d_chw_3x3p1__neonfp16arith_2x8_acc2)
292 
293   BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfp16arith_1x4)
294   BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfp16arith_2x4)
295   BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfp16arith_3x4)
296   BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfp16arith_4x4)
297   BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfp16arith_1x4_acc2)
298   BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfp16arith_1x4_acc3)
299   BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfp16arith_1x4_acc4)
300   BENCHMARK_DWCONV(dwconv2d_chw_3x3s2p1__neonfp16arith_2x4_acc2)
301 
302   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_1x4)
303   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_2x4)
304   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_3x4)
305   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_4x4)
306   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_5x4)
307   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc2)
308   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc3)
309   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc4)
310   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_1x4_acc5)
311   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_2x4_acc2)
312   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_2x4_acc3)
313   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_3x4_acc2)
314   BENCHMARK_DWCONV(dwconv2d_chw_5x5p2__neonfp16arith_4x4_acc2)
315 
316   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_1x4)
317   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_2x4)
318   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_3x4)
319   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc2)
320   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc3)
321   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc4)
322   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_1x4_acc5)
323   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_2x4_acc2)
324   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_2x4_acc3)
325   BENCHMARK_DWCONV(dwconv2d_chw_5x5s2p2__neonfp16arith_3x4_acc2)
326 #endif  // XNN_ENABLE_ARM_FP16 && XNN_ARCH_ARM64
327 
328 
329 #ifndef XNNPACK_BENCHMARK_NO_MAIN
330 BENCHMARK_MAIN();
331 #endif
332