• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2020 The Android Open Source Project
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  *      http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 #include <array>
18 #include <climits>
19 #include <cstdlib>
20 #include <random>
21 #include <vector>
22 
23 #include <benchmark/benchmark.h>
24 
25 #include <audio_utils/intrinsic_utils.h>
26 #include <audio_utils/format.h>
27 
BM_Intrinsic(benchmark::State & state)28 static void BM_Intrinsic(benchmark::State& state) {
29     using D = float;
30     using namespace android::audio_utils::intrinsics;
31     constexpr size_t SIMD_LENGTH = 4;
32 
33     // Possible testing types:
34     using vec = android::audio_utils::intrinsics::internal_array_t<D, SIMD_LENGTH>;
35     //using vec = float32x4_t;
36     //using vec = float32x4x4_t;
37 
38     constexpr size_t DATA_SIZE = 1024;
39     D a[DATA_SIZE];
40     D b[DATA_SIZE];
41     D c[DATA_SIZE];
42     D d[DATA_SIZE];
43 
44     constexpr std::minstd_rand::result_type SEED = 42; // arbitrary choice.
45     std::minstd_rand gen(SEED);
46     const D amplitude = 1.0f;
47     std::uniform_real_distribution<> dis(-amplitude, amplitude);
48     for (size_t i = 0; i < DATA_SIZE; ++i) {
49         a[i] = dis(gen);
50         b[i] = dis(gen);
51         c[i] = dis(gen);
52     }
53 
54     while (state.KeepRunning()) {
55         for (size_t i = 0; i < DATA_SIZE; i += sizeof(vec) / sizeof(D)) {
56             const vec av = vld1<vec>(a + i);
57             const vec bv = vld1<vec>(b + i);
58             const vec cv = vld1<vec>(c + i);
59             const vec dv = vmla(cv, av, bv);
60             vst1(d + i, dv);
61         }
62         benchmark::DoNotOptimize(d[0]);
63         benchmark::ClobberMemory();
64     }
65     //fprintf(stderr, "%f: %f %f\n %f", d[0], c[0], a[0], b[0]);
66     state.SetComplexityN(state.range(0));
67 }
68 
69 // A simple test using the VMLA intrinsic.
70 // One can alter either the intrinsic code or the compilation flags to see the benefit.
71 // Recommend using objdump to view the assembly.
BM_IntrinsicArgs(benchmark::internal::Benchmark * b)72 static void BM_IntrinsicArgs(benchmark::internal::Benchmark* b) {
73     for (int k = 0; k < 2; k++) // 0 for normal random data, 1 for subnormal random data
74          b->Args({k});
75 }
76 
77 BENCHMARK(BM_Intrinsic)->Apply(BM_IntrinsicArgs);
78 
79 BENCHMARK_MAIN();
80