1 | //===-- Performance test for nearest integer functions --------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #include "src/__support/FPUtil/FPBits.h" |
10 | #include "src/math/ceilf.h" |
11 | #include "src/math/ceilf16.h" |
12 | #include "src/math/floorf.h" |
13 | #include "src/math/floorf16.h" |
14 | #include "src/math/rintf.h" |
15 | #include "src/math/rintf16.h" |
16 | #include "src/math/roundevenf.h" |
17 | #include "src/math/roundevenf16.h" |
18 | #include "src/math/roundf.h" |
19 | #include "src/math/roundf16.h" |
20 | #include "src/math/truncf.h" |
21 | #include "src/math/truncf16.h" |
22 | #include "test/UnitTest/RoundingModeUtils.h" |
23 | #include "test/src/math/performance_testing/Timer.h" |
24 | |
25 | #include <fstream> |
26 | #include <math.h> |
27 | |
28 | using LIBC_NAMESPACE::fputil::testing::ForceRoundingMode; |
29 | using LIBC_NAMESPACE::fputil::testing::RoundingMode; |
30 | |
31 | namespace LIBC_NAMESPACE::testing { |
32 | |
33 | template <typename T> class NearestIntegerPerf { |
34 | using FPBits = fputil::FPBits<T>; |
35 | using StorageType = typename FPBits::StorageType; |
36 | |
37 | public: |
38 | typedef T Func(T); |
39 | |
40 | static void run_perf_in_range(Func my_func, Func other_func, |
41 | StorageType starting_bit, |
42 | StorageType ending_bit, StorageType step, |
43 | size_t rounds, const char *name_a, |
44 | const char *name_b, std::ofstream &log) { |
45 | auto runner = [=](Func func) { |
46 | [[maybe_unused]] volatile T result; |
47 | for (size_t i = 0; i < rounds; i++) { |
48 | for (StorageType bits = starting_bit; bits <= ending_bit; |
49 | bits += step) { |
50 | T x = FPBits(bits).get_val(); |
51 | result = func(x); |
52 | } |
53 | } |
54 | }; |
55 | |
56 | Timer timer; |
57 | timer.start(); |
58 | runner(my_func); |
59 | timer.stop(); |
60 | |
61 | size_t number_of_runs = (ending_bit - starting_bit) / step + 1; |
62 | double my_average = |
63 | static_cast<double>(timer.nanoseconds()) / number_of_runs / rounds; |
64 | log << "-- Function A: " << name_a << " --\n" ; |
65 | log << " Total time : " << timer.nanoseconds() << " ns \n" ; |
66 | log << " Average runtime : " << my_average << " ns/op \n" ; |
67 | log << " Ops per second : " |
68 | << static_cast<uint64_t>(1'000'000'000.0 / my_average) << " op/s \n" ; |
69 | |
70 | timer.start(); |
71 | runner(other_func); |
72 | timer.stop(); |
73 | |
74 | double other_average = |
75 | static_cast<double>(timer.nanoseconds()) / number_of_runs / rounds; |
76 | log << "-- Function B: " << name_b << " --\n" ; |
77 | log << " Total time : " << timer.nanoseconds() << " ns \n" ; |
78 | log << " Average runtime : " << other_average << " ns/op \n" ; |
79 | log << " Ops per second : " |
80 | << static_cast<uint64_t>(1'000'000'000.0 / other_average) << " op/s \n" ; |
81 | |
82 | log << "-- Average ops per second ratio --\n" ; |
83 | log << " A / B : " << other_average / my_average << " \n" ; |
84 | } |
85 | |
86 | static void run_perf(Func my_func, Func other_func, size_t rounds, |
87 | const char *name_a, const char *name_b, |
88 | const char *log_file) { |
89 | std::ofstream log(log_file); |
90 | log << "Performance tests with inputs in normal integral range:\n" ; |
91 | run_perf_in_range( |
92 | my_func, other_func, |
93 | /*starting_bit=*/StorageType((FPBits::EXP_BIAS + 1) << FPBits::SIG_LEN), |
94 | /*ending_bit=*/ |
95 | StorageType((FPBits::EXP_BIAS + FPBits::FRACTION_LEN - 1) |
96 | << FPBits::SIG_LEN), |
97 | /*step=*/StorageType(1 << FPBits::SIG_LEN), |
98 | rounds * FPBits::EXP_BIAS * FPBits::EXP_BIAS * 2, name_a, name_b, log); |
99 | log << "\n Performance tests with inputs in low integral range:\n" ; |
100 | run_perf_in_range( |
101 | my_func, other_func, |
102 | /*starting_bit=*/StorageType(1 << FPBits::SIG_LEN), |
103 | /*ending_bit=*/StorageType((FPBits::EXP_BIAS - 1) << FPBits::SIG_LEN), |
104 | /*step_bit=*/StorageType(1 << FPBits::SIG_LEN), |
105 | rounds * FPBits::EXP_BIAS * FPBits::EXP_BIAS * 2, name_a, name_b, log); |
106 | log << "\n Performance tests with inputs in high integral range:\n" ; |
107 | run_perf_in_range( |
108 | my_func, other_func, |
109 | /*starting_bit=*/ |
110 | StorageType((FPBits::EXP_BIAS + FPBits::FRACTION_LEN) |
111 | << FPBits::SIG_LEN), |
112 | /*ending_bit=*/ |
113 | StorageType(FPBits::MAX_BIASED_EXPONENT << FPBits::SIG_LEN), |
114 | /*step=*/StorageType(1 << FPBits::SIG_LEN), |
115 | rounds * FPBits::EXP_BIAS * FPBits::EXP_BIAS * 2, name_a, name_b, log); |
116 | log << "\n Performance tests with inputs in normal fractional range:\n" ; |
117 | run_perf_in_range( |
118 | my_func, other_func, |
119 | /*starting_bit=*/ |
120 | StorageType(((FPBits::EXP_BIAS + 1) << FPBits::SIG_LEN) + 1), |
121 | /*ending_bit=*/ |
122 | StorageType(((FPBits::EXP_BIAS + 2) << FPBits::SIG_LEN) - 1), |
123 | /*step=*/StorageType(1), rounds * 2, name_a, name_b, log); |
124 | log << "\n Performance tests with inputs in subnormal fractional range:\n" ; |
125 | run_perf_in_range(my_func, other_func, /*starting_bit=*/StorageType(1), |
126 | /*ending_bit=*/StorageType(FPBits::SIG_MASK), |
127 | /*step=*/StorageType(1), rounds, name_a, name_b, log); |
128 | } |
129 | }; |
130 | |
131 | } // namespace LIBC_NAMESPACE::testing |
132 | |
133 | #define NEAREST_INTEGER_PERF(T, my_func, other_func, rounds, filename) \ |
134 | { \ |
135 | LIBC_NAMESPACE::testing::NearestIntegerPerf<T>::run_perf( \ |
136 | &my_func, &other_func, rounds, #my_func, #other_func, filename); \ |
137 | } |
138 | |
139 | static constexpr size_t FLOAT16_ROUNDS = 20'000; |
140 | static constexpr size_t FLOAT_ROUNDS = 40; |
141 | |
142 | // LLVM libc might be the only libc implementation with support for float16 math |
143 | // functions currently. We can't compare our float16 functions against the |
144 | // system libc, so we compare them against this placeholder function. |
145 | float16 placeholderf16(float16 x) { return x; } |
146 | |
147 | // The system libc might not provide the roundeven* C23 math functions either. |
148 | float placeholderf(float x) { return x; } |
149 | |
150 | int main() { |
151 | NEAREST_INTEGER_PERF(float16, LIBC_NAMESPACE::ceilf16, ::placeholderf16, |
152 | FLOAT16_ROUNDS, "ceilf16_perf.log" ) |
153 | NEAREST_INTEGER_PERF(float16, LIBC_NAMESPACE::floorf16, ::placeholderf16, |
154 | FLOAT16_ROUNDS, "floorf16_perf.log" ) |
155 | NEAREST_INTEGER_PERF(float16, LIBC_NAMESPACE::roundf16, ::placeholderf16, |
156 | FLOAT16_ROUNDS, "roundf16_perf.log" ) |
157 | NEAREST_INTEGER_PERF(float16, LIBC_NAMESPACE::roundevenf16, ::placeholderf16, |
158 | FLOAT16_ROUNDS, "roundevenf16_perf.log" ) |
159 | NEAREST_INTEGER_PERF(float16, LIBC_NAMESPACE::truncf16, ::placeholderf16, |
160 | FLOAT16_ROUNDS, "truncf16_perf.log" ) |
161 | |
162 | NEAREST_INTEGER_PERF(float, LIBC_NAMESPACE::ceilf, ::ceilf, FLOAT_ROUNDS, |
163 | "ceilf_perf.log" ) |
164 | NEAREST_INTEGER_PERF(float, LIBC_NAMESPACE::floorf, ::floorf, FLOAT_ROUNDS, |
165 | "floorf_perf.log" ) |
166 | NEAREST_INTEGER_PERF(float, LIBC_NAMESPACE::roundf, ::roundf, FLOAT_ROUNDS, |
167 | "roundf_perf.log" ) |
168 | NEAREST_INTEGER_PERF(float, LIBC_NAMESPACE::roundevenf, ::placeholderf, |
169 | FLOAT_ROUNDS, "roundevenf_perf.log" ) |
170 | NEAREST_INTEGER_PERF(float, LIBC_NAMESPACE::truncf, ::truncf, FLOAT_ROUNDS, |
171 | "truncf_perf.log" ) |
172 | |
173 | if (ForceRoundingMode r(RoundingMode::Upward); r.success) { |
174 | NEAREST_INTEGER_PERF(float16, LIBC_NAMESPACE::rintf16, ::placeholderf16, |
175 | FLOAT16_ROUNDS, "rintf16_upward_perf.log" ) |
176 | NEAREST_INTEGER_PERF(float, LIBC_NAMESPACE::rintf, ::rintf, FLOAT_ROUNDS, |
177 | "rintf_upward_perf.log" ) |
178 | } |
179 | if (ForceRoundingMode r(RoundingMode::Downward); r.success) { |
180 | NEAREST_INTEGER_PERF(float16, LIBC_NAMESPACE::rintf16, ::placeholderf16, |
181 | FLOAT16_ROUNDS, "rintf16_downward_perf.log" ) |
182 | NEAREST_INTEGER_PERF(float, LIBC_NAMESPACE::rintf, ::rintf, FLOAT_ROUNDS, |
183 | "rintf_downward_perf.log" ) |
184 | } |
185 | if (ForceRoundingMode r(RoundingMode::TowardZero); r.success) { |
186 | NEAREST_INTEGER_PERF(float16, LIBC_NAMESPACE::rintf16, ::placeholderf16, |
187 | FLOAT16_ROUNDS, "rintf16_towardzero_perf.log" ) |
188 | NEAREST_INTEGER_PERF(float, LIBC_NAMESPACE::rintf, ::rintf, FLOAT_ROUNDS, |
189 | "rintf_towardzero_perf.log" ) |
190 | } |
191 | if (ForceRoundingMode r(RoundingMode::Nearest); r.success) { |
192 | NEAREST_INTEGER_PERF(float16, LIBC_NAMESPACE::rintf16, ::placeholderf16, |
193 | FLOAT16_ROUNDS, "rintf16_nearest_perf.log" ) |
194 | NEAREST_INTEGER_PERF(float, LIBC_NAMESPACE::rintf, ::rintf, FLOAT_ROUNDS, |
195 | "rintf_nearest_perf.log" ) |
196 | } |
197 | |
198 | return 0; |
199 | } |
200 | |