xref: /aosp_15_r20/external/XNNPACK/eval/f32-sqrt-ulp.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 <memory>
11 #include <numeric>
12 #include <random>
13 #include <vector>
14 
15 #include <cpuinfo.h>
16 #include <pthreadpool.h>
17 
18 #include <benchmark/benchmark.h>
19 #include <fp16/fp16.h>
20 
21 #include "bench/utils.h"
22 #include <xnnpack/aligned-allocator.h>
23 #include <xnnpack/common.h>
24 #include <xnnpack/math.h>
25 #include <xnnpack/math-stubs.h>
26 
27 
28 struct ComputeErrorContext {
29   const float* input;
30   const float* output;
31   float* error;
32 };
33 
ComputeError(struct ComputeErrorContext * context,size_t start,size_t range)34 static void ComputeError(
35   struct ComputeErrorContext* context,
36   size_t start,
37   size_t range)
38 {
39   const float* input = context->input;
40   const float* output = context->output;
41   float* error = context->error;
42   for (size_t i = start; i < start + range; i++) {
43     const double output_ref = std::sqrt(double(input[i]));
44     const double abs_error = std::abs(output_ref - double(output[i]));
45     const float output_abs = std::abs(output_ref);
46     const float output_ulp = uint32_as_float(float_as_uint32(output_abs) + 1) - output_abs;
47     error[i] = float(abs_error / output_ulp);
48   }
49 }
50 
SqrtError(benchmark::State & state,xnn_f32_unary_math_function sqrt,benchmark::utils::IsaCheckFunction isa_check=nullptr)51 static void SqrtError(benchmark::State& state,
52   xnn_f32_unary_math_function sqrt,
53   benchmark::utils::IsaCheckFunction isa_check = nullptr)
54 {
55   if (!cpuinfo_initialize()) {
56     state.SkipWithError("failed cpuinfo init");
57     return;
58   }
59   if (isa_check && !isa_check(state)) {
60     return;
61   }
62 
63   const uint32_t min_input = 0x3F800000;
64   const uint32_t max_input = 0x41800000;
65   // Number of elements in one block of inputs/outputs.
66   // Combining multiple elements in a block reduce function call overhead.
67   const size_t block_size = 16384;
68   // Number of elements in one parallelization tile. Worker threads process this many elements in each task.
69   const size_t tile_size = 64;
70 
71   uint32_t num_threads = cpuinfo_get_cores_count();
72   #if XNN_ARCH_ARM || XNN_ARCH_ARM64
73     // Use all cores except for the least performant cluster
74     if (cpuinfo_get_clusters_count() > 1) {
75       num_threads -= cpuinfo_get_cluster(cpuinfo_get_clusters_count() - 1)->core_count;
76     }
77   #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
78 
79   std::unique_ptr<pthreadpool, decltype(&pthreadpool_destroy)> threadpool(
80     pthreadpool_create(num_threads), pthreadpool_destroy);
81 
82   std::vector<float, AlignedAllocator<float, 64>> x(block_size);
83   std::vector<float, AlignedAllocator<float, 64>> y(block_size);
84   std::vector<float> ulp_error(block_size);
85   float max_ulp_error = 0.0f;
86 
87   ComputeErrorContext context;
88   context.input = x.data();
89   context.output = y.data();
90   context.error = ulp_error.data();
91   for (auto _ : state) {
92     for (uint32_t n = min_input; n < max_input; n += block_size) {
93       for (uint32_t i = 0; i < block_size; i++) {
94         x[i] = uint32_as_float(std::min<uint32_t>(n + i, max_input));
95       }
96       std::fill(y.begin(), y.end(), std::nanf(""));
97 
98       sqrt(block_size * sizeof(float), x.data(), y.data());
99 
100       pthreadpool_parallelize_1d_tile_1d(
101           threadpool.get(),
102           reinterpret_cast<pthreadpool_task_1d_tile_1d_t>(ComputeError),
103           static_cast<void*>(&context),
104           block_size, tile_size, 0 /* flags */);
105 
106       max_ulp_error = std::accumulate(ulp_error.cbegin(), ulp_error.cend(), max_ulp_error,
107         static_cast<const float& (*)(const float&, const float&)>(std::max<float>));
108     }
109   }
110 
111   state.counters["ULPERROR"] = benchmark::Counter(max_ulp_error);
112 }
113 
114 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
115   BENCHMARK_CAPTURE(SqrtError, neonfma_nr1fma,
116                     xnn_math_f32_sqrt__neonfma_nr1fma,
117                     benchmark::utils::CheckNEONFMA)
118     ->Unit(benchmark::kMillisecond)
119     ->Iterations(1);
120   BENCHMARK_CAPTURE(SqrtError, neonfma_nr2fma,
121                     xnn_math_f32_sqrt__neonfma_nr2fma,
122                     benchmark::utils::CheckNEONFMA)
123     ->Unit(benchmark::kMillisecond)
124     ->Iterations(1);
125   BENCHMARK_CAPTURE(SqrtError, neonfma_nr3fma,
126                     xnn_math_f32_sqrt__neonfma_nr3fma,
127                     benchmark::utils::CheckNEONFMA)
128     ->Unit(benchmark::kMillisecond)
129     ->Iterations(1);
130   BENCHMARK_CAPTURE(SqrtError, neonfma_nr2fma1adj,
131                     xnn_math_f32_sqrt__neonfma_nr2fma1adj,
132                     benchmark::utils::CheckNEONFMA)
133     ->Unit(benchmark::kMillisecond)
134     ->Iterations(1);
135   BENCHMARK_CAPTURE(SqrtError, neonfma_nr1rsqrts1fma1adj,
136                     xnn_math_f32_sqrt__neonfma_nr1rsqrts1fma1adj,
137                     benchmark::utils::CheckNEONFMA)
138     ->Unit(benchmark::kMillisecond)
139     ->Iterations(1);
140 
141   BENCHMARK_CAPTURE(SqrtError, neon_nr1rsqrts,
142                     xnn_math_f32_sqrt__neon_nr1rsqrts,
143                     benchmark::utils::CheckNEON)
144     ->Unit(benchmark::kMillisecond)
145     ->Iterations(1);
146   BENCHMARK_CAPTURE(SqrtError, neon_nr2rsqrts,
147                     xnn_math_f32_sqrt__neon_nr2rsqrts,
148                     benchmark::utils::CheckNEON)
149     ->Unit(benchmark::kMillisecond)
150     ->Iterations(1);
151   BENCHMARK_CAPTURE(SqrtError, neon_nr3rsqrts,
152                     xnn_math_f32_sqrt__neon_nr3rsqrts,
153                     benchmark::utils::CheckNEON)
154     ->Unit(benchmark::kMillisecond)
155     ->Iterations(1);
156 #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
157 
158 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
159   BENCHMARK_CAPTURE(SqrtError, avx512f_nr1fma,
160                     xnn_math_f32_sqrt__avx512f_nr1fma,
161                     benchmark::utils::CheckAVX512F)
162     ->Unit(benchmark::kMillisecond)
163     ->Iterations(1);
164   BENCHMARK_CAPTURE(SqrtError, avx512f_nr2fma,
165                     xnn_math_f32_sqrt__avx512f_nr2fma,
166                     benchmark::utils::CheckAVX512F)
167     ->Unit(benchmark::kMillisecond)
168     ->Iterations(1);
169   BENCHMARK_CAPTURE(SqrtError, avx512f_nr1fma1adj,
170                     xnn_math_f32_sqrt__avx512f_nr1fma1adj,
171                     benchmark::utils::CheckAVX512F)
172     ->Unit(benchmark::kMillisecond)
173     ->Iterations(1);
174 
175   BENCHMARK_CAPTURE(SqrtError, fma3_nr1fma,
176                     xnn_math_f32_sqrt__fma3_nr1fma,
177                     benchmark::utils::CheckFMA3)
178     ->Unit(benchmark::kMillisecond)
179     ->Iterations(1);
180   BENCHMARK_CAPTURE(SqrtError, fma3_nr2fma,
181                     xnn_math_f32_sqrt__fma3_nr2fma,
182                     benchmark::utils::CheckFMA3)
183     ->Unit(benchmark::kMillisecond)
184     ->Iterations(1);
185   BENCHMARK_CAPTURE(SqrtError, fma3_nr1fma1adj,
186                     xnn_math_f32_sqrt__fma3_nr1fma1adj,
187                     benchmark::utils::CheckFMA3)
188     ->Unit(benchmark::kMillisecond)
189     ->Iterations(1);
190 
191   BENCHMARK_CAPTURE(SqrtError, sse_nr1mac,
192                     xnn_math_f32_sqrt__sse_nr1mac)
193     ->Unit(benchmark::kMillisecond)
194     ->Iterations(1);
195   BENCHMARK_CAPTURE(SqrtError, sse_nr2mac,
196                     xnn_math_f32_sqrt__sse_nr2mac)
197     ->Unit(benchmark::kMillisecond)
198     ->Iterations(1);
199   BENCHMARK_CAPTURE(SqrtError, sse_hh1mac,
200                     xnn_math_f32_sqrt__sse_hh1mac)
201     ->Unit(benchmark::kMillisecond)
202     ->Iterations(1);
203 #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
204 
205 #ifndef XNNPACK_BENCHMARK_NO_MAIN
206 BENCHMARK_MAIN();
207 #endif
208