File size: 4,930 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
// Copyright 2022 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 <fp16/fp16.h>

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


struct ComputeErrorContext {
  const uint16_t* input;
  const uint16_t* output;
  float* error;
};

static void ComputeError(
  struct ComputeErrorContext* context,
  size_t start,
  size_t range)
{
  const uint16_t* input = context->input;
  const uint16_t* output = context->output;
  float* error = context->error;
  for (size_t i = start; i < start + range; i++) {
    const float output_ref = std::exp(fp16_ieee_to_fp32_value(input[i]));
    const float abs_error = std::abs(output_ref - fp16_ieee_to_fp32_value(output[i]));
    const uint16_t output_abs = fp16_ieee_from_fp32_value(std::abs(output_ref));
    const float output_ulp = fp16_ieee_to_fp32_value(output_abs + 1) - fp16_ieee_to_fp32_value(output_abs);
    error[i] = float(abs_error / output_ulp);
  }
}

static void ExpError(
  benchmark::State& state,
  xnn_f16_unary_math_fn exp,
  benchmark::utils::IsaCheckFunction isa_check = nullptr)
{
  if (!cpuinfo_initialize()) {
    state.SkipWithError("failed cpuinfo init");
    return;
  }
  if (isa_check != nullptr && !isa_check(state)) {
    return;
  }

  // The smallest x for which exph(x) is non-zero (-0x2.2A8p+3h).
  const uint16_t min_input = UINT16_C(0xCC55);
  // The largest x for which exph(x) is finite (0x1.63Cp+3h).
  const uint16_t max_input = UINT16_C(0x498F);

  // Number of elements in one block of inputs/outputs.
  // Combining multiple elements in a block reduce function call overhead.
  const size_t block_size = 16384;
  // 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_cores_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<uint16_t, AlignedAllocator<uint16_t, 64>> x(block_size);
  std::vector<uint16_t, AlignedAllocator<uint16_t, 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 (uint16_t n = min_input; int16_t(n) < 0; n -= block_size) {
      for (uint16_t i = 0; i < block_size; i++) {
        x[i] = std::max<uint16_t>(n - i, UINT16_C(0x8000));
      }
      std::fill(y.begin(), y.end(), UINT16_C(0x7E00) /* NaN */);

      exp(block_size * sizeof(uint16_t), 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>));
    }
    for (uint16_t n = 0; n < max_input; n += block_size) {
      for (uint16_t i = 0; i < block_size; i++) {
        x[i] = std::min<uint16_t>(n + i, max_input);
      }
      std::fill(y.begin(), y.end(), UINT16_C(0x7E00) /* NaN */);

      exp(block_size * sizeof(uint16_t), 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_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
  BENCHMARK_CAPTURE(ExpError, neonfp16arith_rr2_p3,
                    xnn_math_f16_exp__neonfp16arith_rr2_p3,
                    benchmark::utils::CheckNEONFP16ARITH)
    ->Unit(benchmark::kMillisecond)
    ->Iterations(1);
#endif  // XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)

#ifndef XNNPACK_BENCHMARK_NO_MAIN
BENCHMARK_MAIN();
#endif