diff options
Diffstat (limited to 'test/dct16x16_test.cc')
| -rw-r--r-- | test/dct16x16_test.cc | 48 |
1 files changed, 48 insertions, 0 deletions
diff --git a/test/dct16x16_test.cc b/test/dct16x16_test.cc index d4ef7ae13..3c104f3a4 100644 --- a/test/dct16x16_test.cc +++ b/test/dct16x16_test.cc @@ -27,6 +27,7 @@ #include "vpx/vpx_integer.h" #include "vpx_ports/mem.h" #include "vpx_ports/msvc.h" // for round() +#include "vpx_ports/vpx_timer.h" using libvpx_test::ACMRandom; @@ -548,6 +549,44 @@ class Trans16x16TestBase { } } + void RunSpeedTest() { + ACMRandom rnd(ACMRandom::DeterministicSeed()); + const int count_test_block = 10000; + int c_sum_time = 0; + int simd_sum_time = 0; + + DECLARE_ALIGNED(32, int16_t, input_block[kNumCoeffs]); + DECLARE_ALIGNED(32, tran_low_t, output_ref_block[kNumCoeffs]); + DECLARE_ALIGNED(32, tran_low_t, output_block[kNumCoeffs]); + + // Initialize a test block with input range [-mask_, mask_]. + for (int j = 0; j < kNumCoeffs; ++j) { + input_block[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_); + } + + vpx_usec_timer timer_c; + vpx_usec_timer_start(&timer_c); + for (int i = 0; i < count_test_block; ++i) { + vpx_fdct16x16_c(input_block, output_ref_block, pitch_); + } + vpx_usec_timer_mark(&timer_c); + c_sum_time += static_cast<int>(vpx_usec_timer_elapsed(&timer_c)); + + vpx_usec_timer timer_mod; + vpx_usec_timer_start(&timer_mod); + for (int i = 0; i < count_test_block; ++i) { + RunFwdTxfm(input_block, output_block, pitch_); + } + + vpx_usec_timer_mark(&timer_mod); + simd_sum_time += static_cast<int>(vpx_usec_timer_elapsed(&timer_mod)); + + printf( + "c_time = %d \t simd_time = %d \t Gain = %4.2f \n", c_sum_time, + simd_sum_time, + (static_cast<float>(c_sum_time) / static_cast<float>(simd_sum_time))); + } + void CompareInvReference(IdctFunc ref_txfm, int thresh) { ACMRandom rnd(ACMRandom::DeterministicSeed()); const int count_test_block = 10000; @@ -664,6 +703,8 @@ TEST_P(Trans16x16DCT, QuantCheck) { TEST_P(Trans16x16DCT, InvAccuracyCheck) { RunInvAccuracyCheck(); } +TEST_P(Trans16x16DCT, DISABLED_Speed) { RunSpeedTest(); } + class Trans16x16HT : public Trans16x16TestBase, public ::testing::TestWithParam<Ht16x16Param> { public: @@ -823,6 +864,13 @@ INSTANTIATE_TEST_SUITE_P( 3, VPX_BITS_8))); #endif // HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE +#if HAVE_AVX2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE +INSTANTIATE_TEST_SUITE_P( + AVX2, Trans16x16DCT, + ::testing::Values(make_tuple(&vpx_fdct16x16_avx2, + &vpx_idct16x16_256_add_sse2, 0, VPX_BITS_8))); +#endif // HAVE_AVX2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE + #if HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE INSTANTIATE_TEST_SUITE_P( SSE2, Trans16x16DCT, |