File size: 9,979 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <algorithm>
#include <array>
#include <cmath>
#include <functional>
#include <limits>
#include <memory>
#include <random>
#include <vector>

#include <fp16/fp16.h>

#include <xnnpack.h>

#include <benchmark/benchmark.h>
#include "bench/utils.h"
#ifdef BENCHMARK_TENSORFLOW_LITE
#include "flatbuffers/include/flatbuffers/flatbuffers.h"
#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/schema/schema_generated.h"
#include "tensorflow/lite/version.h"
#endif  // BENCHMARK_TENSORFLOW_LITE


static void xnnpack_bankers_rounding_f16(benchmark::State& state) {
  const size_t batch_size = state.range(0);

  std::random_device random_device;
  auto rng = std::mt19937(random_device());
  auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng));
  auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng);

  std::vector<uint16_t> input(batch_size + XNN_EXTRA_BYTES / sizeof(uint16_t));
  std::vector<uint16_t> output(batch_size);
  std::generate(input.begin(), input.end(), std::ref(f16rng));
  std::fill(output.begin(), output.end(), UINT16_C(0x7E00) /* NaN */);

  xnn_status status = xnn_initialize(nullptr /* allocator */);
  if (status != xnn_status_success) {
    state.SkipWithError("failed to initialize XNNPACK");
    return;
  }

  xnn_operator_t bankers_rounding_op = nullptr;
  status = xnn_create_bankers_rounding_nc_f16(
    1 /* channels */, 1 /* input stride */, 1 /* output stride */,
    0 /* flags */, &bankers_rounding_op);
  if (status != xnn_status_success || bankers_rounding_op == nullptr) {
    state.SkipWithError("failed to create Bankers' Rounding operator");
    return;
  }

  status = xnn_reshape_bankers_rounding_nc_f16(bankers_rounding_op, batch_size, /*threadpool=*/nullptr);
  if (status != xnn_status_success) {
    state.SkipWithError("failed to reshape Bankers' Rounding operator");
    return;
  }

  status = xnn_setup_bankers_rounding_nc_f16(bankers_rounding_op, input.data(), output.data());
  if (status != xnn_status_success) {
    state.SkipWithError("failed to setup Bankers' Rounding operator");
    return;
  }

  for (auto _ : state) {
    status = xnn_run_operator(bankers_rounding_op, nullptr /* thread pool */);
    if (status != xnn_status_success) {
      state.SkipWithError("failed to run Bankers' Rounding operator");
      return;
    }
  }

  status = xnn_delete_operator(bankers_rounding_op);
  if (status != xnn_status_success) {
    state.SkipWithError("failed to delete Bankers' Rounding operator");
    return;
  }

  const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
  if (cpu_frequency != 0) {
    state.counters["cpufreq"] = cpu_frequency;
  }

  state.counters["elements"] =
    benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate);

  const size_t bytes_per_iteration = 2 * batch_size * sizeof(uint16_t);
  state.counters["bytes"] =
    benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
}

static void xnnpack_bankers_rounding_f32(benchmark::State& state) {
  const size_t batch_size = state.range(0);

  std::random_device random_device;
  auto rng = std::mt19937(random_device());
  auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng));

  std::vector<float> input(batch_size + XNN_EXTRA_BYTES / sizeof(float));
  std::vector<float> output(batch_size);
  std::generate(input.begin(), input.end(), std::ref(f32rng));
  std::fill(output.begin(), output.end(), std::nanf(""));

  xnn_status status = xnn_initialize(nullptr /* allocator */);
  if (status != xnn_status_success) {
    state.SkipWithError("failed to initialize XNNPACK");
    return;
  }

  xnn_operator_t bankers_rounding_op = nullptr;
  status = xnn_create_bankers_rounding_nc_f32(
    1 /* channels */, 1 /* input stride */, 1 /* output stride */,
    0 /* flags */, &bankers_rounding_op);
  if (status != xnn_status_success || bankers_rounding_op == nullptr) {
    state.SkipWithError("failed to create Bankers' Rounding operator");
    return;
  }

  status = xnn_reshape_bankers_rounding_nc_f32(bankers_rounding_op, batch_size, /*threadpool=*/nullptr);
  if (status != xnn_status_success) {
    state.SkipWithError("failed to reshape Bankers' Rounding operator");
    return;
  }

  status = xnn_setup_bankers_rounding_nc_f32(bankers_rounding_op, input.data(), output.data());
  if (status != xnn_status_success) {
    state.SkipWithError("failed to setup Bankers' Rounding operator");
    return;
  }

  for (auto _ : state) {
    status = xnn_run_operator(bankers_rounding_op, nullptr /* thread pool */);
    if (status != xnn_status_success) {
      state.SkipWithError("failed to run Bankers' Rounding operator");
      return;
    }
  }

  status = xnn_delete_operator(bankers_rounding_op);
  if (status != xnn_status_success) {
    state.SkipWithError("failed to delete Bankers' Rounding operator");
    return;
  }

  const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
  if (cpu_frequency != 0) {
    state.counters["cpufreq"] = cpu_frequency;
  }

  state.counters["elements"] =
    benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate);

  const size_t bytes_per_iteration = 2 * batch_size * sizeof(float);
  state.counters["bytes"] =
    benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);
}

#ifdef BENCHMARK_TENSORFLOW_LITE
static void tflite_bankers_rounding_f32(benchmark::State& state) {
  const size_t batch_size = state.range(0);

  std::random_device random_device;
  auto rng = std::mt19937(random_device());
  auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng));

  flatbuffers::FlatBufferBuilder builder;
  const flatbuffers::Offset<tflite::OperatorCode> operator_code =
      CreateOperatorCode(builder, tflite::BuiltinOperator_ROUND);

  const std::array<flatbuffers::Offset<tflite::Buffer>, 1> buffers{{
    tflite::CreateBuffer(builder, builder.CreateVector({})),
  }};

  const std::array<int32_t, 1> shape{{
    static_cast<int32_t>(batch_size)
  }};

  const std::array<flatbuffers::Offset<tflite::Tensor>, 2> tensors{{
    tflite::CreateTensor(builder,
                         builder.CreateVector<int32_t>(shape.data(), shape.size()),
                         tflite::TensorType_FLOAT32),
    tflite::CreateTensor(builder,
                         builder.CreateVector<int32_t>(shape.data(), shape.size()),
                         tflite::TensorType_FLOAT32),
  }};

  const std::array<int32_t, 1> op_inputs{{ 0 }};
  const std::array<int32_t, 1> op_outputs{{ 1 }};
  flatbuffers::Offset<tflite::Operator> op = tflite::CreateOperator(
      builder,
      0 /* opcode_index */,
      builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()),
      builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size()));

  const std::array<int32_t, 1> graph_inputs{{ 0 }};
  const std::array<int32_t, 1> graph_outputs{{ 1 }};
  const flatbuffers::Offset<tflite::SubGraph> subgraph = tflite::CreateSubGraph(
      builder,
      builder.CreateVector(tensors.data(), tensors.size()),
      builder.CreateVector<int32_t>(graph_inputs.data(), graph_inputs.size()),
      builder.CreateVector<int32_t>(graph_outputs.data(), graph_outputs.size()),
      builder.CreateVector(&op, 1));

  const flatbuffers::Offset<tflite::Model> model_buffer = tflite::CreateModel(builder,
      TFLITE_SCHEMA_VERSION,
      builder.CreateVector(&operator_code, 1),
      builder.CreateVector(&subgraph, 1),
      builder.CreateString("Round model"),
      builder.CreateVector(buffers.data(), buffers.size()));

  builder.Finish(model_buffer);

  const tflite::Model* model = tflite::GetModel(builder.GetBufferPointer());
  tflite::ops::builtin::BuiltinOpResolverWithoutDefaultDelegates resolver;
  tflite::InterpreterBuilder interpreterBuilder(model, resolver);
  std::unique_ptr<tflite::Interpreter> interpreter;
  if (interpreterBuilder(&interpreter) != kTfLiteOk || interpreter == nullptr) {
    state.SkipWithError("failed to create TFLite interpreter");
    return;
  }
  interpreter->SetNumThreads(1);

  if (interpreter->AllocateTensors() != kTfLiteOk) {
    state.SkipWithError("failed to allocate tensors");
    return;
  }

  std::generate(
    interpreter->typed_tensor<float>(0),
    interpreter->typed_tensor<float>(0) + batch_size,
    std::ref(f32rng));

  for (auto _ : state) {
    if (interpreter->Invoke() != kTfLiteOk) {
      state.SkipWithError("failed to invoke TFLite interpreter");
      return;
    }
  }

  const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
  if (cpu_frequency != 0) {
    state.counters["cpufreq"] = cpu_frequency;
  }

  state.counters["elements"] =
    benchmark::Counter(uint64_t(state.iterations()) * batch_size, benchmark::Counter::kIsRate);

  const size_t bytes_per_iteration = 2 * batch_size * sizeof(float);
  state.counters["bytes"] =
    benchmark::Counter(uint64_t(state.iterations()) * bytes_per_iteration, benchmark::Counter::kIsRate);

  interpreter.reset();
}
#endif  // BENCHMARK_TENSORFLOW_LITE

BENCHMARK(xnnpack_bankers_rounding_f16)
  ->Apply(benchmark::utils::UnaryElementwiseParameters<uint16_t, uint16_t>)
  ->UseRealTime();
BENCHMARK(xnnpack_bankers_rounding_f32)
  ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
  ->UseRealTime();

#ifdef BENCHMARK_TENSORFLOW_LITE
  BENCHMARK(tflite_bankers_rounding_f32)
    ->Apply(benchmark::utils::UnaryElementwiseParameters<float, float>)
    ->UseRealTime();
#endif  // BENCHMARK_TENSORFLOW_LITE

#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif