xref: /aosp_15_r20/external/abseil-cpp/absl/container/internal/layout_benchmark.cc (revision 9356374a3709195abf420251b3e825997ff56c0f)
1 // Copyright 2018 The Abseil Authors.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //      https://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 //
15 // Every benchmark should have the same performance as the corresponding
16 // headroom benchmark.
17 
18 #include <cstddef>
19 #include <cstdint>
20 
21 #include "absl/base/internal/raw_logging.h"
22 #include "absl/container/internal/layout.h"
23 #include "benchmark/benchmark.h"
24 
25 namespace absl {
26 ABSL_NAMESPACE_BEGIN
27 namespace container_internal {
28 namespace {
29 
30 using ::benchmark::DoNotOptimize;
31 
32 using Int128 = int64_t[2];
33 
MyAlign(size_t n,size_t m)34 constexpr size_t MyAlign(size_t n, size_t m) { return (n + m - 1) & ~(m - 1); }
35 
36 // This benchmark provides the upper bound on performance for BM_OffsetConstant.
37 template <size_t Offset, class... Ts>
BM_OffsetConstantHeadroom(benchmark::State & state)38 void BM_OffsetConstantHeadroom(benchmark::State& state) {
39   for (auto _ : state) {
40     DoNotOptimize(Offset);
41   }
42 }
43 
44 template <size_t Offset, class... Ts>
BM_OffsetConstantStatic(benchmark::State & state)45 void BM_OffsetConstantStatic(benchmark::State& state) {
46   using L = typename Layout<Ts...>::template WithStaticSizes<3, 5, 7>;
47   ABSL_RAW_CHECK(L::Partial().template Offset<3>() == Offset, "Invalid offset");
48   for (auto _ : state) {
49     DoNotOptimize(L::Partial().template Offset<3>());
50   }
51 }
52 
53 template <size_t Offset, class... Ts>
BM_OffsetConstant(benchmark::State & state)54 void BM_OffsetConstant(benchmark::State& state) {
55   using L = Layout<Ts...>;
56   ABSL_RAW_CHECK(L::Partial(3, 5, 7).template Offset<3>() == Offset,
57                  "Invalid offset");
58   for (auto _ : state) {
59     DoNotOptimize(L::Partial(3, 5, 7).template Offset<3>());
60   }
61 }
62 
63 template <size_t Offset, class... Ts>
BM_OffsetConstantIndirect(benchmark::State & state)64 void BM_OffsetConstantIndirect(benchmark::State& state) {
65   using L = Layout<Ts...>;
66   auto p = L::Partial(3, 5, 7);
67   ABSL_RAW_CHECK(p.template Offset<3>() == Offset, "Invalid offset");
68   for (auto _ : state) {
69     DoNotOptimize(p);
70     DoNotOptimize(p.template Offset<3>());
71   }
72 }
73 
74 template <class... Ts>
75 size_t PartialOffset(size_t k);
76 
77 template <>
PartialOffset(size_t k)78 size_t PartialOffset<int8_t, int16_t, int32_t, Int128>(size_t k) {
79   constexpr size_t o = MyAlign(MyAlign(3 * 1, 2) + 5 * 2, 4);
80   return MyAlign(o + k * 4, 8);
81 }
82 
83 template <>
PartialOffset(size_t k)84 size_t PartialOffset<Int128, int32_t, int16_t, int8_t>(size_t k) {
85   // No alignment is necessary.
86   return 3 * 16 + 5 * 4 + k * 2;
87 }
88 
89 // This benchmark provides the upper bound on performance for BM_OffsetVariable.
90 template <size_t Offset, class... Ts>
BM_OffsetPartialHeadroom(benchmark::State & state)91 void BM_OffsetPartialHeadroom(benchmark::State& state) {
92   size_t k = 7;
93   ABSL_RAW_CHECK(PartialOffset<Ts...>(k) == Offset, "Invalid offset");
94   for (auto _ : state) {
95     DoNotOptimize(k);
96     DoNotOptimize(PartialOffset<Ts...>(k));
97   }
98 }
99 
100 template <size_t Offset, class... Ts>
BM_OffsetPartialStatic(benchmark::State & state)101 void BM_OffsetPartialStatic(benchmark::State& state) {
102   using L = typename Layout<Ts...>::template WithStaticSizes<3, 5>;
103   size_t k = 7;
104   ABSL_RAW_CHECK(L::Partial(k).template Offset<3>() == Offset,
105                  "Invalid offset");
106   for (auto _ : state) {
107     DoNotOptimize(k);
108     DoNotOptimize(L::Partial(k).template Offset<3>());
109   }
110 }
111 
112 template <size_t Offset, class... Ts>
BM_OffsetPartial(benchmark::State & state)113 void BM_OffsetPartial(benchmark::State& state) {
114   using L = Layout<Ts...>;
115   size_t k = 7;
116   ABSL_RAW_CHECK(L::Partial(3, 5, k).template Offset<3>() == Offset,
117                  "Invalid offset");
118   for (auto _ : state) {
119     DoNotOptimize(k);
120     DoNotOptimize(L::Partial(3, 5, k).template Offset<3>());
121   }
122 }
123 
124 template <class... Ts>
125 size_t VariableOffset(size_t n, size_t m, size_t k);
126 
127 template <>
VariableOffset(size_t n,size_t m,size_t k)128 size_t VariableOffset<int8_t, int16_t, int32_t, Int128>(size_t n, size_t m,
129                                                         size_t k) {
130   return MyAlign(MyAlign(MyAlign(n * 1, 2) + m * 2, 4) + k * 4, 8);
131 }
132 
133 template <>
VariableOffset(size_t n,size_t m,size_t k)134 size_t VariableOffset<Int128, int32_t, int16_t, int8_t>(size_t n, size_t m,
135                                                         size_t k) {
136   // No alignment is necessary.
137   return n * 16 + m * 4 + k * 2;
138 }
139 
140 // This benchmark provides the upper bound on performance for BM_OffsetVariable.
141 template <size_t Offset, class... Ts>
BM_OffsetVariableHeadroom(benchmark::State & state)142 void BM_OffsetVariableHeadroom(benchmark::State& state) {
143   size_t n = 3;
144   size_t m = 5;
145   size_t k = 7;
146   ABSL_RAW_CHECK(VariableOffset<Ts...>(n, m, k) == Offset, "Invalid offset");
147   for (auto _ : state) {
148     DoNotOptimize(n);
149     DoNotOptimize(m);
150     DoNotOptimize(k);
151     DoNotOptimize(VariableOffset<Ts...>(n, m, k));
152   }
153 }
154 
155 template <size_t Offset, class... Ts>
BM_OffsetVariable(benchmark::State & state)156 void BM_OffsetVariable(benchmark::State& state) {
157   using L = Layout<Ts...>;
158   size_t n = 3;
159   size_t m = 5;
160   size_t k = 7;
161   ABSL_RAW_CHECK(L::Partial(n, m, k).template Offset<3>() == Offset,
162                  "Invalid offset");
163   for (auto _ : state) {
164     DoNotOptimize(n);
165     DoNotOptimize(m);
166     DoNotOptimize(k);
167     DoNotOptimize(L::Partial(n, m, k).template Offset<3>());
168   }
169 }
170 
171 template <class... Ts>
172 size_t AllocSize(size_t x);
173 
174 template <>
AllocSize(size_t x)175 size_t AllocSize<int8_t, int16_t, int32_t, Int128>(size_t x) {
176   constexpr size_t o =
177       Layout<int8_t, int16_t, int32_t, Int128>::Partial(3, 5, 7)
178           .template Offset<Int128>();
179   return o + sizeof(Int128) * x;
180 }
181 
182 template <>
AllocSize(size_t x)183 size_t AllocSize<Int128, int32_t, int16_t, int8_t>(size_t x) {
184   constexpr size_t o =
185       Layout<Int128, int32_t, int16_t, int8_t>::Partial(3, 5, 7)
186           .template Offset<int8_t>();
187   return o + sizeof(int8_t) * x;
188 }
189 
190 // This benchmark provides the upper bound on performance for BM_AllocSize
191 template <size_t Size, class... Ts>
BM_AllocSizeHeadroom(benchmark::State & state)192 void BM_AllocSizeHeadroom(benchmark::State& state) {
193   size_t x = 9;
194   ABSL_RAW_CHECK(AllocSize<Ts...>(x) == Size, "Invalid size");
195   for (auto _ : state) {
196     DoNotOptimize(x);
197     DoNotOptimize(AllocSize<Ts...>(x));
198   }
199 }
200 
201 template <size_t Size, class... Ts>
BM_AllocSizeStatic(benchmark::State & state)202 void BM_AllocSizeStatic(benchmark::State& state) {
203   using L = typename Layout<Ts...>::template WithStaticSizes<3, 5, 7>;
204   size_t x = 9;
205   ABSL_RAW_CHECK(L(x).AllocSize() == Size, "Invalid offset");
206   for (auto _ : state) {
207     DoNotOptimize(x);
208     DoNotOptimize(L(x).AllocSize());
209   }
210 }
211 
212 template <size_t Size, class... Ts>
BM_AllocSize(benchmark::State & state)213 void BM_AllocSize(benchmark::State& state) {
214   using L = Layout<Ts...>;
215   size_t n = 3;
216   size_t m = 5;
217   size_t k = 7;
218   size_t x = 9;
219   ABSL_RAW_CHECK(L(n, m, k, x).AllocSize() == Size, "Invalid offset");
220   for (auto _ : state) {
221     DoNotOptimize(n);
222     DoNotOptimize(m);
223     DoNotOptimize(k);
224     DoNotOptimize(x);
225     DoNotOptimize(L(n, m, k, x).AllocSize());
226   }
227 }
228 
229 template <size_t Size, class... Ts>
BM_AllocSizeIndirect(benchmark::State & state)230 void BM_AllocSizeIndirect(benchmark::State& state) {
231   using L = Layout<Ts...>;
232   auto l = L(3, 5, 7, 9);
233   ABSL_RAW_CHECK(l.AllocSize() == Size, "Invalid offset");
234   for (auto _ : state) {
235     DoNotOptimize(l);
236     DoNotOptimize(l.AllocSize());
237   }
238 }
239 
240 // Run all benchmarks in two modes:
241 //
242 //   Layout with padding: int8_t[3], int16_t[5], int32_t[7], Int128[?].
243 //   Layout without padding: Int128[3], int32_t[5], int16_t[7], int8_t[?].
244 
245 #define OFFSET_BENCHMARK(NAME, OFFSET, T1, T2, T3, T4) \
246   auto& NAME##_##OFFSET##_##T1##_##T2##_##T3##_##T4 =  \
247       NAME<OFFSET, T1, T2, T3, T4>;                    \
248   BENCHMARK(NAME##_##OFFSET##_##T1##_##T2##_##T3##_##T4)
249 
250 OFFSET_BENCHMARK(BM_OffsetConstantHeadroom, 48, int8_t, int16_t, int32_t,
251                  Int128);
252 OFFSET_BENCHMARK(BM_OffsetConstantStatic, 48, int8_t, int16_t, int32_t, Int128);
253 OFFSET_BENCHMARK(BM_OffsetConstant, 48, int8_t, int16_t, int32_t, Int128);
254 OFFSET_BENCHMARK(BM_OffsetConstantIndirect, 48, int8_t, int16_t, int32_t,
255                  Int128);
256 
257 OFFSET_BENCHMARK(BM_OffsetConstantHeadroom, 82, Int128, int32_t, int16_t,
258                  int8_t);
259 OFFSET_BENCHMARK(BM_OffsetConstantStatic, 82, Int128, int32_t, int16_t, int8_t);
260 OFFSET_BENCHMARK(BM_OffsetConstant, 82, Int128, int32_t, int16_t, int8_t);
261 OFFSET_BENCHMARK(BM_OffsetConstantIndirect, 82, Int128, int32_t, int16_t,
262                  int8_t);
263 
264 OFFSET_BENCHMARK(BM_OffsetPartialHeadroom, 48, int8_t, int16_t, int32_t,
265                  Int128);
266 OFFSET_BENCHMARK(BM_OffsetPartialStatic, 48, int8_t, int16_t, int32_t, Int128);
267 OFFSET_BENCHMARK(BM_OffsetPartial, 48, int8_t, int16_t, int32_t, Int128);
268 
269 OFFSET_BENCHMARK(BM_OffsetPartialHeadroom, 82, Int128, int32_t, int16_t,
270                  int8_t);
271 OFFSET_BENCHMARK(BM_OffsetPartialStatic, 82, Int128, int32_t, int16_t, int8_t);
272 OFFSET_BENCHMARK(BM_OffsetPartial, 82, Int128, int32_t, int16_t, int8_t);
273 
274 OFFSET_BENCHMARK(BM_OffsetVariableHeadroom, 48, int8_t, int16_t, int32_t,
275                  Int128);
276 OFFSET_BENCHMARK(BM_OffsetVariable, 48, int8_t, int16_t, int32_t, Int128);
277 
278 OFFSET_BENCHMARK(BM_OffsetVariableHeadroom, 82, Int128, int32_t, int16_t,
279                  int8_t);
280 OFFSET_BENCHMARK(BM_OffsetVariable, 82, Int128, int32_t, int16_t, int8_t);
281 
282 OFFSET_BENCHMARK(BM_AllocSizeHeadroom, 192, int8_t, int16_t, int32_t, Int128);
283 OFFSET_BENCHMARK(BM_AllocSizeStatic, 192, int8_t, int16_t, int32_t, Int128);
284 OFFSET_BENCHMARK(BM_AllocSize, 192, int8_t, int16_t, int32_t, Int128);
285 OFFSET_BENCHMARK(BM_AllocSizeIndirect, 192, int8_t, int16_t, int32_t, Int128);
286 
287 OFFSET_BENCHMARK(BM_AllocSizeHeadroom, 91, Int128, int32_t, int16_t, int8_t);
288 OFFSET_BENCHMARK(BM_AllocSizeStatic, 91, Int128, int32_t, int16_t, int8_t);
289 OFFSET_BENCHMARK(BM_AllocSize, 91, Int128, int32_t, int16_t, int8_t);
290 OFFSET_BENCHMARK(BM_AllocSizeIndirect, 91, Int128, int32_t, int16_t, int8_t);
291 
292 }  // namespace
293 }  // namespace container_internal
294 ABSL_NAMESPACE_END
295 }  // namespace absl
296