src/media/base/vector_math_perftest.cc

   1 // Copyright 2013 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 #include "base/cpu.h"
   6 #include "base/memory/aligned_memory.h"
   7 #include "base/memory/scoped_ptr.h"
   8 #include "base/time/time.h"
   9 #include "media/base/vector_math.h"
  10 #include "media/base/vector_math_testing.h"
  11 #include "testing/gtest/include/gtest/gtest.h"
  12 #include "testing/perf/perf_test.h"
  13
  14 using base::TimeTicks;
  15 using std::fill;
  16
  17 namespace media {
  18
  19 static const int kBenchmarkIterations = 200000;
  20 static const int kEWMABenchmarkIterations = 50000;
  21 static const float kScale = 0.5;
  22 static const int kVectorSize = 8192;
  23
  24 class VectorMathPerfTest : public testing::Test {
  25  public:
  26   VectorMathPerfTest() {
  27     // Initialize input and output vectors.
  28     input_vector_.reset(static_cast<float*>(base::AlignedAlloc(
  29         sizeof(float) * kVectorSize, vector_math::kRequiredAlignment)));
  30     output_vector_.reset(static_cast<float*>(base::AlignedAlloc(
  31         sizeof(float) * kVectorSize, vector_math::kRequiredAlignment)));
  32     fill(input_vector_.get(), input_vector_.get() + kVectorSize, 1.0f);
  33     fill(output_vector_.get(), output_vector_.get() + kVectorSize, 0.0f);
  34   }
  35
  36   void RunBenchmark(void (*fn)(const float[], float, int, float[]),
  37                     bool aligned,
  38                     const std::string& test_name,
  39                     const std::string& trace_name) {
  40     TimeTicks start = TimeTicks::HighResNow();
  41     for (int i = 0; i < kBenchmarkIterations; ++i) {
  42       fn(input_vector_.get(),
  43          kScale,
  44          kVectorSize - (aligned ? 0 : 1),
  45          output_vector_.get());
  46     }
  47     double total_time_milliseconds =
  48         (TimeTicks::HighResNow() - start).InMillisecondsF();
  49     perf_test::PrintResult(test_name,
  50                            "",
  51                            trace_name,
  52                            kBenchmarkIterations / total_time_milliseconds,
  53                            "runs/ms",
  54                            true);
  55   }
  56
  57   void RunBenchmark(
  58       std::pair<float, float> (*fn)(float, const float[], int, float),
  59       int len,
  60       const std::string& test_name,
  61       const std::string& trace_name) {
  62     TimeTicks start = TimeTicks::HighResNow();
  63     for (int i = 0; i < kEWMABenchmarkIterations; ++i) {
  64       fn(0.5f, input_vector_.get(), len, 0.1f);
  65     }
  66     double total_time_milliseconds =
  67         (TimeTicks::HighResNow() - start).InMillisecondsF();
  68     perf_test::PrintResult(test_name,
  69                            "",
  70                            trace_name,
  71                            kEWMABenchmarkIterations / total_time_milliseconds,
  72                            "runs/ms",
  73                            true);
  74   }
  75
  76  protected:
  77   scoped_ptr<float, base::AlignedFreeDeleter> input_vector_;
  78   scoped_ptr<float, base::AlignedFreeDeleter> output_vector_;
  79
  80   DISALLOW_COPY_AND_ASSIGN(VectorMathPerfTest);
  81 };
  82
  83 // Define platform independent function name for FMAC* perf tests.
  84 #if defined(ARCH_CPU_X86_FAMILY)
  85 #define FMAC_FUNC FMAC_SSE
  86 #elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
  87 #define FMAC_FUNC FMAC_NEON
  88 #endif
  89
  90 // Benchmark for each optimized vector_math::FMAC() method.
  91 TEST_F(VectorMathPerfTest, FMAC) {
  92   // Benchmark FMAC_C().
  93   RunBenchmark(
  94       vector_math::FMAC_C, true, "vector_math_fmac", "unoptimized");
  95 #if defined(FMAC_FUNC)
  96 #if defined(ARCH_CPU_X86_FAMILY)
  97   ASSERT_TRUE(base::CPU().has_sse());
  98 #endif
  99   // Benchmark FMAC_FUNC() with unaligned size.
 100   ASSERT_NE((kVectorSize - 1) % (vector_math::kRequiredAlignment /
 101                                  sizeof(float)), 0U);
 102   RunBenchmark(
 103       vector_math::FMAC_FUNC, false, "vector_math_fmac", "optimized_unaligned");
 104   // Benchmark FMAC_FUNC() with aligned size.
 105   ASSERT_EQ(kVectorSize % (vector_math::kRequiredAlignment / sizeof(float)),
 106             0U);
 107   RunBenchmark(
 108       vector_math::FMAC_FUNC, true, "vector_math_fmac", "optimized_aligned");
 109 #endif
 110 }
 111
 112 #undef FMAC_FUNC
 113
 114 // Define platform independent function name for FMULBenchmark* tests.
 115 #if defined(ARCH_CPU_X86_FAMILY)
 116 #define FMUL_FUNC FMUL_SSE
 117 #elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
 118 #define FMUL_FUNC FMUL_NEON
 119 #endif
 120
 121 // Benchmark for each optimized vector_math::FMUL() method.
 122 TEST_F(VectorMathPerfTest, FMUL) {
 123   // Benchmark FMUL_C().
 124   RunBenchmark(
 125       vector_math::FMUL_C, true, "vector_math_fmul", "unoptimized");
 126 #if defined(FMUL_FUNC)
 127 #if defined(ARCH_CPU_X86_FAMILY)
 128   ASSERT_TRUE(base::CPU().has_sse());
 129 #endif
 130   // Benchmark FMUL_FUNC() with unaligned size.
 131   ASSERT_NE((kVectorSize - 1) % (vector_math::kRequiredAlignment /
 132                                  sizeof(float)), 0U);
 133   RunBenchmark(
 134       vector_math::FMUL_FUNC, false, "vector_math_fmul", "optimized_unaligned");
 135   // Benchmark FMUL_FUNC() with aligned size.
 136   ASSERT_EQ(kVectorSize % (vector_math::kRequiredAlignment / sizeof(float)),
 137             0U);
 138   RunBenchmark(
 139       vector_math::FMUL_FUNC, true, "vector_math_fmul", "optimized_aligned");
 140 #endif
 141 }
 142
 143 #undef FMUL_FUNC
 144
 145 #if defined(ARCH_CPU_X86_FAMILY)
 146 #define EWMAAndMaxPower_FUNC EWMAAndMaxPower_SSE
 147 #elif defined(ARCH_CPU_ARM_FAMILY) && defined(USE_NEON)
 148 #define EWMAAndMaxPower_FUNC EWMAAndMaxPower_NEON
 149 #endif
 150
 151 // Benchmark for each optimized vector_math::EWMAAndMaxPower() method.
 152 TEST_F(VectorMathPerfTest, EWMAAndMaxPower) {
 153   // Benchmark EWMAAndMaxPower_C().
 154   RunBenchmark(vector_math::EWMAAndMaxPower_C,
 155                kVectorSize,
 156                "vector_math_ewma_and_max_power",
 157                "unoptimized");
 158 #if defined(EWMAAndMaxPower_FUNC)
 159 #if defined(ARCH_CPU_X86_FAMILY)
 160   ASSERT_TRUE(base::CPU().has_sse());
 161 #endif
 162   // Benchmark EWMAAndMaxPower_FUNC() with unaligned size.
 163   ASSERT_NE((kVectorSize - 1) % (vector_math::kRequiredAlignment /
 164                                  sizeof(float)), 0U);
 165   RunBenchmark(vector_math::EWMAAndMaxPower_FUNC,
 166                kVectorSize - 1,
 167                "vector_math_ewma_and_max_power",
 168                "optimized_unaligned");
 169   // Benchmark EWMAAndMaxPower_FUNC() with aligned size.
 170   ASSERT_EQ(kVectorSize % (vector_math::kRequiredAlignment / sizeof(float)),
 171             0U);
 172   RunBenchmark(vector_math::EWMAAndMaxPower_FUNC,
 173                kVectorSize,
 174                "vector_math_ewma_and_max_power",
 175                "optimized_aligned");
 176 #endif
 177 }
 178
 179 #undef EWMAAndMaxPower_FUNC
 180
 181 } // namespace media