File size: 7,186 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
// Copyright 2019 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 <cfloat>
#include <cmath>
#include <functional>
#include <memory>
#include <numeric>
#include <random>
#include <vector>

#include <cpuinfo.h>
#include <pthreadpool.h>

#include <benchmark/benchmark.h>

#include "bench/utils.h"
#include <xnnpack/aligned-allocator.h>
#include <xnnpack/common.h>
#include <xnnpack/math.h>
#include <xnnpack/math-stubs.h>


struct ComputeErrorContext {
  const float* input;
  const float* output;
  float* error;
};

static void ComputeError(
  struct ComputeErrorContext* context,
  size_t start,
  size_t range)
{
  const float* input = context->input;
  const float* output = context->output;
  float* error = context->error;
  for (size_t i = start; i < start + range; i++) {
    const double output_ref = std::sqrt(double(input[i]));
    const double abs_error = std::abs(output_ref - double(output[i]));
    const float output_abs = std::abs(output_ref);
    const float output_ulp = uint32_as_float(float_as_uint32(output_abs) + 1) - output_abs;
    error[i] = float(abs_error / output_ulp);
  }
}

static void SqrtError(benchmark::State& state,
  xnn_f32_unary_math_fn sqrt,
  benchmark::utils::IsaCheckFunction isa_check = nullptr)
{
  if (!cpuinfo_initialize()) {
    state.SkipWithError("failed cpuinfo init");
    return;
  }
  if (isa_check != nullptr && !isa_check(state)) {
    return;
  }

  const uint32_t min_input = 0x3F800000;
  const uint32_t max_input = 0x41800000;
  // Number of elements in one block of inputs/outputs.
  // Combining multiple elements in a block reduce function call overhead.
  const size_t block_size = 1048576;
  // Number of elements in one parallelization tile. Worker threads process this many elements in each task.
  const size_t tile_size = 64;

  uint32_t num_threads = cpuinfo_get_processors_count();
  #if XNN_ARCH_ARM || XNN_ARCH_ARM64
    // Use all cores except for the least performant cluster
    if (cpuinfo_get_clusters_count() > 1) {
      num_threads -= cpuinfo_get_cluster(cpuinfo_get_clusters_count() - 1)->core_count;
    }
  #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64

  std::unique_ptr<pthreadpool, decltype(&pthreadpool_destroy)> threadpool(
    pthreadpool_create(num_threads), pthreadpool_destroy);

  std::vector<float, AlignedAllocator<float, 64>> x(block_size);
  std::vector<float, AlignedAllocator<float, 64>> y(block_size);
  std::vector<float> ulp_error(block_size);
  float max_ulp_error = 0.0f;

  ComputeErrorContext context;
  context.input = x.data();
  context.output = y.data();
  context.error = ulp_error.data();
  for (auto _ : state) {
    for (uint32_t n = min_input; n < max_input; n += block_size) {
      for (uint32_t i = 0; i < block_size; i++) {
        x[i] = uint32_as_float(std::min<uint32_t>(n + i, max_input));
      }
      std::fill(y.begin(), y.end(), std::nanf(""));

      sqrt(block_size * sizeof(float), x.data(), y.data());

      pthreadpool_parallelize_1d_tile_1d(
          threadpool.get(),
          reinterpret_cast<pthreadpool_task_1d_tile_1d_t>(ComputeError),
          static_cast<void*>(&context),
          block_size, tile_size, 0 /* flags */);

      max_ulp_error = std::accumulate(ulp_error.cbegin(), ulp_error.cend(), max_ulp_error,
        static_cast<const float& (*)(const float&, const float&)>(std::max<float>));
    }
  }

  state.counters["ULPERROR"] = benchmark::Counter(max_ulp_error);
}

#if XNN_ARCH_ARM || XNN_ARCH_ARM64
  BENCHMARK_CAPTURE(SqrtError, neonfma_nr1fma,
                    xnn_math_f32_sqrt__neonfma_nr1fma,
                    benchmark::utils::CheckNEONFMA)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, neonfma_nr2fma,
                    xnn_math_f32_sqrt__neonfma_nr2fma,
                    benchmark::utils::CheckNEONFMA)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, neonfma_nr3fma,
                    xnn_math_f32_sqrt__neonfma_nr3fma,
                    benchmark::utils::CheckNEONFMA)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, neonfma_nr2fma1adj,
                    xnn_math_f32_sqrt__neonfma_nr2fma1adj,
                    benchmark::utils::CheckNEONFMA)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, neonfma_nr1rsqrts1fma1adj,
                    xnn_math_f32_sqrt__neonfma_nr1rsqrts1fma1adj,
                    benchmark::utils::CheckNEONFMA)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);

  BENCHMARK_CAPTURE(SqrtError, neon_nr1rsqrts,
                    xnn_math_f32_sqrt__neon_nr1rsqrts,
                    benchmark::utils::CheckNEON)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, neon_nr2rsqrts,
                    xnn_math_f32_sqrt__neon_nr2rsqrts,
                    benchmark::utils::CheckNEON)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, neon_nr3rsqrts,
                    xnn_math_f32_sqrt__neon_nr3rsqrts,
                    benchmark::utils::CheckNEON)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
#endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64

#if XNN_ARCH_X86 || XNN_ARCH_X86_64
  BENCHMARK_CAPTURE(SqrtError, avx512f_nr1fma,
                    xnn_math_f32_sqrt__avx512f_nr1fma,
                    benchmark::utils::CheckAVX512F)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, avx512f_nr2fma,
                    xnn_math_f32_sqrt__avx512f_nr2fma,
                    benchmark::utils::CheckAVX512F)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, avx512f_nr1fma1adj,
                    xnn_math_f32_sqrt__avx512f_nr1fma1adj,
                    benchmark::utils::CheckAVX512F)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);

  BENCHMARK_CAPTURE(SqrtError, fma3_nr1fma,
                    xnn_math_f32_sqrt__fma3_nr1fma,
                    benchmark::utils::CheckFMA3)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, fma3_nr2fma,
                    xnn_math_f32_sqrt__fma3_nr2fma,
                    benchmark::utils::CheckFMA3)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, fma3_nr1fma1adj,
                    xnn_math_f32_sqrt__fma3_nr1fma1adj,
                    benchmark::utils::CheckFMA3)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);

  BENCHMARK_CAPTURE(SqrtError, sse_nr1mac,
                    xnn_math_f32_sqrt__sse_nr1mac)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, sse_nr2mac,
                    xnn_math_f32_sqrt__sse_nr2mac)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
  BENCHMARK_CAPTURE(SqrtError, sse_hh1mac,
                    xnn_math_f32_sqrt__sse_hh1mac)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64

#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif