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/AlignedAllocator.h>
23 #include <xnnpack/common.h>
24 #include <xnnpack/math-stubs.h>
25 
26 
27 struct ComputeErrorContext {
28   const float* input;
29   const float* output;
30   float* error;
31 };
32 
ComputeError(struct ComputeErrorContext * context,size_t start,size_t range)33 static void ComputeError(
34   struct ComputeErrorContext* context,
35   size_t start,
36   size_t range)
37 {
38   const float* input = context->input;
39   const float* output = context->output;
40   float* error = context->error;
41   for (size_t i = start; i < start + range; i++) {
42     const double output_ref = std::sqrt(double(input[i]));
43     const double abs_error = std::abs(output_ref - double(output[i]));
44     const float output_abs = std::abs(output_ref);
45     const float output_ulp = fp32_from_bits(fp32_to_bits(output_abs) + 1) - output_abs;
46     error[i] = float(abs_error / output_ulp);
47   }
48 }
49 
SqrtError(benchmark::State & state,xnn_f32_unary_math_function sqrt,benchmark::utils::IsaCheckFunction isa_check=nullptr)50 static void SqrtError(benchmark::State& state,
51   xnn_f32_unary_math_function sqrt,
52   benchmark::utils::IsaCheckFunction isa_check = nullptr)
53 {
54   if (!cpuinfo_initialize()) {
55     state.SkipWithError("failed cpuinfo init");
56     return;
57   }
58   if (isa_check && !isa_check(state)) {
59     return;
60   }
61 
62   const uint32_t min_input = 0x3F800000;
63   const uint32_t max_input = 0x41800000;
64   // Number of elements in one block of inputs/outputs.
65   // Combining multiple elements in a block reduce function call overhead.
66   const size_t block_size = 16384;
67   // Number of elements in one parallelization tile. Worker threads process this many elements in each task.
68   const size_t tile_size = 64;
69 
70   uint32_t num_threads = cpuinfo_get_cores_count();
71   #if XNN_ARCH_ARM || XNN_ARCH_ARM64
72     // Use all cores except for the least performant cluster
73     if (cpuinfo_get_clusters_count() > 1) {
74       num_threads -= cpuinfo_get_cluster(cpuinfo_get_clusters_count() - 1)->core_count;
75     }
76   #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
77 
78   std::unique_ptr<pthreadpool, decltype(&pthreadpool_destroy)> threadpool(
79     pthreadpool_create(num_threads), pthreadpool_destroy);
80 
81   std::vector<float, AlignedAllocator<float, 64>> x(block_size);
82   std::vector<float, AlignedAllocator<float, 64>> y(block_size);
83   std::vector<float> ulp_error(block_size);
84   float max_ulp_error = 0.0f;
85 
86   ComputeErrorContext context;
87   context.input = x.data();
88   context.output = y.data();
89   context.error = ulp_error.data();
90   for (auto _ : state) {
91     for (uint32_t n = min_input; n < max_input; n += block_size) {
92       for (uint32_t i = 0; i < block_size; i++) {
93         x[i] = fp32_from_bits(std::min<uint32_t>(n + i, max_input));
94       }
95       std::fill(y.begin(), y.end(), std::nanf(""));
96 
97       sqrt(block_size * sizeof(float), x.data(), y.data());
98 
99       pthreadpool_parallelize_1d_tile_1d(
100           threadpool.get(),
101           reinterpret_cast<pthreadpool_task_1d_tile_1d_t>(ComputeError),
102           static_cast<void*>(&context),
103           block_size, tile_size, 0 /* flags */);
104 
105       max_ulp_error = std::accumulate(ulp_error.cbegin(), ulp_error.cend(), max_ulp_error,
106         static_cast<const float& (*)(const float&, const float&)>(std::max<float>));
107     }
108   }
109 
110   state.counters["ULPERROR"] = benchmark::Counter(max_ulp_error);
111 }
112 
113 #if XNN_ARCH_ARM || XNN_ARCH_ARM64
114   BENCHMARK_CAPTURE(SqrtError, neonfma_nr1fma,
115                     xnn_math_f32_sqrt__neonfma_nr1fma,
116                     benchmark::utils::CheckNEONFMA)
117     ->Unit(benchmark::kMillisecond)
118     ->Iterations(1);
119   BENCHMARK_CAPTURE(SqrtError, neonfma_nr2fma,
120                     xnn_math_f32_sqrt__neonfma_nr2fma,
121                     benchmark::utils::CheckNEONFMA)
122     ->Unit(benchmark::kMillisecond)
123     ->Iterations(1);
124   BENCHMARK_CAPTURE(SqrtError, neonfma_nr3fma,
125                     xnn_math_f32_sqrt__neonfma_nr3fma,
126                     benchmark::utils::CheckNEONFMA)
127     ->Unit(benchmark::kMillisecond)
128     ->Iterations(1);
129   BENCHMARK_CAPTURE(SqrtError, neonfma_nr2fma1adj,
130                     xnn_math_f32_sqrt__neonfma_nr2fma1adj,
131                     benchmark::utils::CheckNEONFMA)
132     ->Unit(benchmark::kMillisecond)
133     ->Iterations(1);
134   BENCHMARK_CAPTURE(SqrtError, neonfma_nr1rsqrts1fma1adj,
135                     xnn_math_f32_sqrt__neonfma_nr1rsqrts1fma1adj,
136                     benchmark::utils::CheckNEONFMA)
137     ->Unit(benchmark::kMillisecond)
138     ->Iterations(1);
139 
140   BENCHMARK_CAPTURE(SqrtError, neon_nr1rsqrts,
141                     xnn_math_f32_sqrt__neon_nr1rsqrts,
142                     benchmark::utils::CheckNEON)
143     ->Unit(benchmark::kMillisecond)
144     ->Iterations(1);
145   BENCHMARK_CAPTURE(SqrtError, neon_nr2rsqrts,
146                     xnn_math_f32_sqrt__neon_nr2rsqrts,
147                     benchmark::utils::CheckNEON)
148     ->Unit(benchmark::kMillisecond)
149     ->Iterations(1);
150   BENCHMARK_CAPTURE(SqrtError, neon_nr3rsqrts,
151                     xnn_math_f32_sqrt__neon_nr3rsqrts,
152                     benchmark::utils::CheckNEON)
153     ->Unit(benchmark::kMillisecond)
154     ->Iterations(1);
155 #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
156 
157 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
158   BENCHMARK_CAPTURE(SqrtError, avx512f_nr1fma,
159                     xnn_math_f32_sqrt__avx512f_nr1fma,
160                     benchmark::utils::CheckAVX512F)
161     ->Unit(benchmark::kMillisecond)
162     ->Iterations(1);
163   BENCHMARK_CAPTURE(SqrtError, avx512f_nr2fma,
164                     xnn_math_f32_sqrt__avx512f_nr2fma,
165                     benchmark::utils::CheckAVX512F)
166     ->Unit(benchmark::kMillisecond)
167     ->Iterations(1);
168   BENCHMARK_CAPTURE(SqrtError, avx512f_nr1fma1adj,
169                     xnn_math_f32_sqrt__avx512f_nr1fma1adj,
170                     benchmark::utils::CheckAVX512F)
171     ->Unit(benchmark::kMillisecond)
172     ->Iterations(1);
173 
174   BENCHMARK_CAPTURE(SqrtError, fma3_nr1fma,
175                     xnn_math_f32_sqrt__fma3_nr1fma,
176                     benchmark::utils::CheckFMA3)
177     ->Unit(benchmark::kMillisecond)
178     ->Iterations(1);
179   BENCHMARK_CAPTURE(SqrtError, fma3_nr2fma,
180                     xnn_math_f32_sqrt__fma3_nr2fma,
181                     benchmark::utils::CheckFMA3)
182     ->Unit(benchmark::kMillisecond)
183     ->Iterations(1);
184   BENCHMARK_CAPTURE(SqrtError, fma3_nr1fma1adj,
185                     xnn_math_f32_sqrt__fma3_nr1fma1adj,
186                     benchmark::utils::CheckFMA3)
187     ->Unit(benchmark::kMillisecond)
188     ->Iterations(1);
189 
190   BENCHMARK_CAPTURE(SqrtError, sse_nr1mac,
191                     xnn_math_f32_sqrt__sse_nr1mac)
192     ->Unit(benchmark::kMillisecond)
193     ->Iterations(1);
194   BENCHMARK_CAPTURE(SqrtError, sse_nr2mac,
195                     xnn_math_f32_sqrt__sse_nr2mac)
196     ->Unit(benchmark::kMillisecond)
197     ->Iterations(1);
198   BENCHMARK_CAPTURE(SqrtError, sse_hh1mac,
199                     xnn_math_f32_sqrt__sse_hh1mac)
200     ->Unit(benchmark::kMillisecond)
201     ->Iterations(1);
202 #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
203 
204 #ifndef XNNPACK_BENCHMARK_NO_MAIN
205 BENCHMARK_MAIN();
206 #endif
207