xref: /aosp_15_r20/external/XNNPACK/eval/f32-expm1minus-ulp.cc (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Copyright 2020 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::expm1(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 
ExpM1Error(benchmark::State & state,xnn_f32_unary_math_function expm1,benchmark::utils::IsaCheckFunction isa_check=nullptr)51 static void ExpM1Error(benchmark::State& state,
52   xnn_f32_unary_math_function expm1,
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   // The smallest x for which expm1f(x) is not saturated at -1 (-0x1.154244p+4f).
64   const uint32_t min_input = UINT32_C(0xC18AA122);
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; int32_t(n) < 0; n -= block_size) {
93       for (uint32_t i = 0; i < block_size; i++) {
94         x[i] = uint32_as_float(std::max<uint32_t>(n - i, 0x80000000));
95       }
96       std::fill(y.begin(), y.end(), std::nanf(""));
97 
98       expm1(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(ExpM1Error, neon_rr2_lut16_p3,
116                     xnn_math_f32_expm1minus__neon_rr2_lut16_p3,
117                     benchmark::utils::CheckNEON)
118     ->Unit(benchmark::kMillisecond)
119     ->Iterations(1);
120   BENCHMARK_CAPTURE(ExpM1Error, neon_rr2_p6,
121                     xnn_math_f32_expm1minus__neon_rr2_p6,
122                     benchmark::utils::CheckNEON)
123     ->Unit(benchmark::kMillisecond)
124     ->Iterations(1);
125 
126   BENCHMARK_CAPTURE(ExpM1Error, neonfma_rr1_lut16_p3,
127                     xnn_math_f32_expm1minus__neonfma_rr1_lut16_p3,
128                     benchmark::utils::CheckNEONFMA)
129     ->Unit(benchmark::kMillisecond)
130     ->Iterations(1);
131   BENCHMARK_CAPTURE(ExpM1Error, neonfma_rr1_p6,
132                     xnn_math_f32_expm1minus__neonfma_rr1_p6,
133                     benchmark::utils::CheckNEONFMA)
134     ->Unit(benchmark::kMillisecond)
135     ->Iterations(1);
136 #endif  // XNN_ARCH_ARM || XNN_ARCH_ARM64
137 
138 #if XNN_ARCH_X86 || XNN_ARCH_X86_64
139   BENCHMARK_CAPTURE(ExpM1Error, avx512f_rr1_lut16_p3_perm,
140                     xnn_math_f32_expm1minus__avx512f_rr1_lut16_p3_perm,
141                     benchmark::utils::CheckAVX512F)
142     ->Unit(benchmark::kMillisecond)
143     ->Iterations(1);
144   BENCHMARK_CAPTURE(ExpM1Error, avx512f_rr1_p6,
145                     xnn_math_f32_expm1minus__avx512f_rr1_p6,
146                     benchmark::utils::CheckAVX512F)
147     ->Unit(benchmark::kMillisecond)
148     ->Iterations(1);
149 
150   BENCHMARK_CAPTURE(ExpM1Error, avx2_rr1_lut4_p4_perm,
151                     xnn_math_f32_expm1minus__avx2_rr1_lut4_p4_perm,
152                     benchmark::utils::CheckAVX2)
153     ->Unit(benchmark::kMillisecond)
154     ->Iterations(1);
155   BENCHMARK_CAPTURE(ExpM1Error, avx2_rr1_lut8_p4_perm,
156                     xnn_math_f32_expm1minus__avx2_rr1_lut8_p4_perm,
157                     benchmark::utils::CheckAVX2)
158     ->Unit(benchmark::kMillisecond)
159     ->Iterations(1);
160   BENCHMARK_CAPTURE(ExpM1Error, avx2_rr1_lut16_p3_gather,
161                     xnn_math_f32_expm1minus__avx2_rr1_lut16_p3_gather,
162                     benchmark::utils::CheckAVX2)
163     ->Unit(benchmark::kMillisecond)
164     ->Iterations(1);
165   BENCHMARK_CAPTURE(ExpM1Error, avx2_rr1_p6,
166                     xnn_math_f32_expm1minus__avx2_rr1_p6,
167                     benchmark::utils::CheckAVX2)
168     ->Unit(benchmark::kMillisecond)
169     ->Iterations(1);
170 
171   BENCHMARK_CAPTURE(ExpM1Error, avx_rr2_lut4_p4_perm,
172                     xnn_math_f32_expm1minus__avx_rr2_lut4_p4_perm,
173                     benchmark::utils::CheckAVX)
174     ->Unit(benchmark::kMillisecond)
175     ->Iterations(1);
176   BENCHMARK_CAPTURE(ExpM1Error, avx_rr2_lut16_p3,
177                     xnn_math_f32_expm1minus__avx_rr2_lut16_p3,
178                     benchmark::utils::CheckAVX)
179     ->Unit(benchmark::kMillisecond)
180     ->Iterations(1);
181   BENCHMARK_CAPTURE(ExpM1Error, avx_rr2_p6,
182                     xnn_math_f32_expm1minus__avx_rr2_p6,
183                     benchmark::utils::CheckAVX)
184     ->Unit(benchmark::kMillisecond)
185     ->Iterations(1);
186 
187   BENCHMARK_CAPTURE(ExpM1Error, sse2_rr2_lut16_p3,
188                     xnn_math_f32_expm1minus__sse2_rr2_lut16_p3)
189     ->Unit(benchmark::kMillisecond)
190     ->Iterations(1);
191   BENCHMARK_CAPTURE(ExpM1Error, sse2_rr2_p6,
192                     xnn_math_f32_expm1minus__sse2_rr2_p6)
193     ->Unit(benchmark::kMillisecond)
194     ->Iterations(1);
195 #endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
196 
197 #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
198   BENCHMARK_CAPTURE(ExpM1Error, wasmsimd_rr2_lut16_p3_andnot,
199                     xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_andnot)
200     ->Unit(benchmark::kMillisecond)
201     ->Iterations(1);
202   BENCHMARK_CAPTURE(ExpM1Error, wasmsimd_rr2_lut16_p3_max,
203                     xnn_math_f32_expm1minus__wasmsimd_rr2_lut16_p3_max)
204     ->Unit(benchmark::kMillisecond)
205     ->Iterations(1);
206   BENCHMARK_CAPTURE(ExpM1Error, wasmsimd_rr2_p6_andnot,
207                     xnn_math_f32_expm1minus__wasmsimd_rr2_p6_andnot)
208     ->Unit(benchmark::kMillisecond)
209     ->Iterations(1);
210   BENCHMARK_CAPTURE(ExpM1Error, wasmsimd_rr2_p6_max,
211                     xnn_math_f32_expm1minus__wasmsimd_rr2_p6_max)
212     ->Unit(benchmark::kMillisecond)
213     ->Iterations(1);
214 #endif  // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD
215 
216 BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut4_p4,
217                   xnn_math_f32_expm1minus__scalar_rr2_lut4_p4)
218   ->Unit(benchmark::kMillisecond)
219   ->Iterations(1);
220 BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut8_p3,
221                   xnn_math_f32_expm1minus__scalar_rr2_lut8_p3)
222   ->Unit(benchmark::kMillisecond)
223   ->Iterations(1);
224 BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut8_p4,
225                   xnn_math_f32_expm1minus__scalar_rr2_lut8_p4)
226   ->Unit(benchmark::kMillisecond)
227   ->Iterations(1);
228 BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut16_p3,
229                   xnn_math_f32_expm1minus__scalar_rr2_lut16_p3)
230   ->Unit(benchmark::kMillisecond)
231   ->Iterations(1);
232 BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_lut16_p4,
233                   xnn_math_f32_expm1minus__scalar_rr2_lut16_p4)
234   ->Unit(benchmark::kMillisecond)
235   ->Iterations(1);
236 BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_p5,
237                   xnn_math_f32_expm1minus__scalar_rr2_p5)
238   ->Unit(benchmark::kMillisecond)
239   ->Iterations(1);
240 BENCHMARK_CAPTURE(ExpM1Error, scalar_rr2_p6,
241                   xnn_math_f32_expm1minus__scalar_rr2_p6)
242   ->Unit(benchmark::kMillisecond)
243   ->Iterations(1);
244 
245 #ifndef XNNPACK_BENCHMARK_NO_MAIN
246 BENCHMARK_MAIN();
247 #endif
248