diff options
| -rw-r--r-- | volk/apps/volk_profile.cc | 212 | ||||
| -rw-r--r-- | volk/lib/qa_utils.cc | 9 | ||||
| -rw-r--r-- | volk/lib/qa_utils.h | 6 |
3 files changed, 123 insertions, 104 deletions
diff --git a/volk/apps/volk_profile.cc b/volk/apps/volk_profile.cc index 3986cb6e62..103dfdcf6a 100644 --- a/volk/apps/volk_profile.cc +++ b/volk/apps/volk_profile.cc @@ -19,119 +19,131 @@ int main(int argc, char *argv[]) { boost::program_options::options_description desc("Options"); desc.add_options() ("help,h", "Print help messages") - ("benchmark,b", boost::program_options::value<bool>()->default_value( false ), "run all kernels (benchmark mode)"); - //("benchmark,b", boost::program_options::value<bool>(&benchmark_mode)->default_value( false ), "run all kernels (benchmark mode)"); + ("benchmark,b", + boost::program_options::value<bool>()->default_value( false ), + "Run all kernels (benchmark mode)") + ("tests-regex,R", + boost::program_options::value<std::string>(), + "Run tests matching regular expression.") + ; boost::program_options::variables_map vm; - bool benchmark_mode = false; + bool benchmark_mode; + std::string kernel_regex; try { - boost::program_options::store(boost::program_options::parse_command_line(argc, argv, desc), vm); + boost::program_options::store(boost::program_options::parse_command_line(argc, argv, desc), vm); boost::program_options::notify(vm); - benchmark_mode = vm.count("benchmark")?vm["benchmark"].as<bool>():false; + benchmark_mode, kernel_regex = vm.count("benchmark")?vm["benchmark"].as<bool>():false; + if ( vm.count("tests-regex" ) ) { + kernel_regex = vm["tests-regex"].as<std::string>(); + } + else { + kernel_regex = ".*"; + } } catch (boost::program_options::error& error) { std::cerr << "Error: " << error.what() << std::endl << std::endl; std::cerr << desc << std::endl; return 1; } /** --help option - */ - if ( vm.count("help") ) - { - std::cout << "Basic Command Line Parameter App" << std::endl - << desc << std::endl; - return 0; - } +*/ + if ( vm.count("help") ) + { + std::cout << "The VOLK profiler." << std::endl + << desc << std::endl; + return 0; + } std::vector<std::string> results; - //VOLK_PROFILE(volk_16i_x5_add_quad_16i_x4, 1e-4, 2046, 10000, &results, benchmark_mode); - //VOLK_PROFILE(volk_16i_branch_4_state_8, 1e-4, 2046, 10000, &results, benchmark_mode); - VOLK_PUPPET_PROFILE(volk_32fc_s32fc_rotatorpuppet_32fc, volk_32fc_s32fc_x2_rotator_32fc, 1e-2, (lv_32fc_t)lv_cmake(0.953939201, 0.3), 20462, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_16ic_s32f_deinterleave_real_32f, 1e-5, 32768.0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_16ic_deinterleave_real_8i, 0, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_16ic_deinterleave_16i_x2, 0, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_16ic_s32f_deinterleave_32f_x2, 1e-4, 32768.0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_16ic_deinterleave_real_16i, 0, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_16ic_magnitude_16i, 1, 0, 204602, 100, &results, benchmark_mode); - VOLK_PROFILE(volk_16ic_s32f_magnitude_32f, 1e-5, 32768.0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_16i_s32f_convert_32f, 1e-4, 32768.0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_16i_convert_8i, 0, 0, 204602, 10000, &results, benchmark_mode); - //VOLK_PROFILE(volk_16i_max_star_16i, 0, 0, 204602, 10000, &results, benchmark_mode); - //VOLK_PROFILE(volk_16i_max_star_horizontal_16i, 0, 0, 204602, 10000, &results, benchmark_mode); - //VOLK_PROFILE(volk_16i_permute_and_scalar_add, 1e-4, 0, 2046, 10000, &results, benchmark_mode); - //VOLK_PROFILE(volk_16i_x4_quad_max_star_16i, 1e-4, 0, 2046, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_16u_byteswap, 0, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_16i_32fc_dot_prod_32fc, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_accumulator_s32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_add_32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_32f_multiply_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_s32f_power_32fc, 1e-4, 0, 204602, 50, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_s32f_calc_spectral_noise_floor_32f, 1e-4, 20.0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_s32f_atan2_32f, 1e-4, 10.0, 204602, 100, &results, benchmark_mode); - //VOLK_PROFILE(volk_32fc_x2_conjugate_dot_prod_32fc, 1e-4, 0, 2046, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_x2_conjugate_dot_prod_32fc, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_deinterleave_32f_x2, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_deinterleave_64f_x2, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_s32f_deinterleave_real_16i, 0, 32768, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_deinterleave_imag_32f, 1e-4, 0, 204602, 5000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_deinterleave_real_32f, 1e-4, 0, 204602, 5000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_deinterleave_real_64f, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_x2_dot_prod_32fc, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_32f_dot_prod_32fc, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_index_max_16u, 3, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_s32f_magnitude_16i, 1, 32768, 204602, 100, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_magnitude_32f, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_magnitude_squared_32f, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_x2_multiply_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_x2_multiply_conjugate_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_conjugate_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_s32f_convert_16i, 1, 32768, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_s32f_convert_32i, 1, 1<<31, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_convert_64f, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_s32f_convert_8i, 1, 128, 204602, 10000, &results, benchmark_mode); - //VOLK_PROFILE(volk_32fc_s32f_x2_power_spectral_density_32f, 1e-4, 2046, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_s32f_power_spectrum_32f, 1e-4, 0, 20462, 100, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_x2_square_dist_32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_x2_s32f_square_dist_scalar_mult_32f, 1e-4, 10, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_divide_32f, 1e-4, 0, 204602, 2000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_dot_prod_32f, 1e-4, 0, 204602, 5000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_dot_prod_16i, 1e-4, 0, 204602, 5000, &results, benchmark_mode); - //VOLK_PROFILE(volk_32f_s32f_32f_fm_detect_32f, 1e-4, 2046, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_index_max_16u, 3, 0, 204602, 5000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_s32f_interleave_16ic, 1, 32768, 204602, 3000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_interleave_32fc, 0, 0, 204602, 5000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_max_32f, 1e-4, 0, 204602, 2000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_min_32f, 1e-4, 0, 204602, 2000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_multiply_32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_s32f_normalize, 1e-4, 100, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_s32f_power_32f, 1e-4, 4, 204602, 100, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_sqrt_32f, 1e-4, 0, 204602, 100, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_s32f_stddev_32f, 1e-4, 100, 204602, 3000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_stddev_and_mean_32f_x2, 1e-4, 0, 204602, 3000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x2_subtract_32f, 1e-4, 0, 204602, 5000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_x3_sum_of_poly_32f, 1e-2, 0, 204602, 5000, &results, benchmark_mode); - VOLK_PROFILE(volk_32i_x2_and_32i, 0, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32i_s32f_convert_32f, 1e-4, 100, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32i_x2_or_32i, 0, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_32u_byteswap, 0, 0, 204602, 2000, &results, benchmark_mode); - //VOLK_PROFILE(volk_32u_popcnt, 0, 0, 2046, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_64f_convert_32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_64f_x2_max_64f, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_64f_x2_min_64f, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_64u_byteswap, 0, 0, 204602, 1000, &results, benchmark_mode); - //VOLK_PROFILE(volk_64u_popcnt, 0, 0, 2046, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_8ic_deinterleave_16i_x2, 0, 0, 204602, 3000, &results, benchmark_mode); - VOLK_PROFILE(volk_8ic_s32f_deinterleave_32f_x2, 1e-4, 100, 204602, 3000, &results, benchmark_mode); - VOLK_PROFILE(volk_8ic_deinterleave_real_16i, 0, 256, 204602, 3000, &results, benchmark_mode); - VOLK_PROFILE(volk_8ic_s32f_deinterleave_real_32f, 1e-4, 100, 204602, 3000, &results, benchmark_mode); - VOLK_PROFILE(volk_8ic_deinterleave_real_8i, 0, 0, 204602, 10000, &results, benchmark_mode); - VOLK_PROFILE(volk_8ic_x2_multiply_conjugate_16ic, 0, 0, 204602, 400, &results, benchmark_mode); - VOLK_PROFILE(volk_8ic_x2_s32f_multiply_conjugate_32fc, 1e-4, 100, 204602, 400, &results, benchmark_mode); - VOLK_PROFILE(volk_8i_convert_16i, 0, 0, 204602, 20000, &results, benchmark_mode); - VOLK_PROFILE(volk_8i_s32f_convert_32f, 1e-4, 100, 204602, 2000, &results, benchmark_mode); - //VOLK_PROFILE(volk_32fc_s32fc_multiply_32fc, 1e-4, lv_32fc_t(1.0, 0.5), 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32fc_s32fc_multiply_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode); - VOLK_PROFILE(volk_32f_s32f_multiply_32f, 1e-4, 1.0, 204602, 10000, &results, benchmark_mode); + //VOLK_PROFILE(volk_16i_x5_add_quad_16i_x4, 1e-4, 2046, 10000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_16i_branch_4_state_8, 1e-4, 2046, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PUPPET_PROFILE(volk_32fc_s32fc_rotatorpuppet_32fc, volk_32fc_s32fc_x2_rotator_32fc, 1e-2, (lv_32fc_t)lv_cmake(0.953939201, 0.3), 20462, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16ic_s32f_deinterleave_real_32f, 1e-5, 32768.0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16ic_deinterleave_real_8i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16ic_deinterleave_16i_x2, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16ic_s32f_deinterleave_32f_x2, 1e-4, 32768.0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16ic_deinterleave_real_16i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16ic_magnitude_16i, 1, 0, 204602, 100, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16ic_s32f_magnitude_32f, 1e-5, 32768.0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16i_s32f_convert_32f, 1e-4, 32768.0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16i_convert_8i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_16i_max_star_16i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_16i_max_star_horizontal_16i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_16i_permute_and_scalar_add, 1e-4, 0, 2046, 10000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_16i_x4_quad_max_star_16i, 1e-4, 0, 2046, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16u_byteswap, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_16i_32fc_dot_prod_32fc, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_accumulator_s32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_add_32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_32f_multiply_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_s32f_power_32fc, 1e-4, 0, 204602, 50, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_s32f_calc_spectral_noise_floor_32f, 1e-4, 20.0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_s32f_atan2_32f, 1e-4, 10.0, 204602, 100, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_32fc_x2_conjugate_dot_prod_32fc, 1e-4, 0, 2046, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_x2_conjugate_dot_prod_32fc, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_deinterleave_32f_x2, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_deinterleave_64f_x2, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_s32f_deinterleave_real_16i, 0, 32768, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_deinterleave_imag_32f, 1e-4, 0, 204602, 5000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_deinterleave_real_32f, 1e-4, 0, 204602, 5000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_deinterleave_real_64f, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_x2_dot_prod_32fc, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_32f_dot_prod_32fc, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_index_max_16u, 3, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_s32f_magnitude_16i, 1, 32768, 204602, 100, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_magnitude_32f, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_magnitude_squared_32f, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_x2_multiply_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_x2_multiply_conjugate_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_conjugate_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_s32f_convert_16i, 1, 32768, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_s32f_convert_32i, 1, 1<<31, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_convert_64f, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_s32f_convert_8i, 1, 128, 204602, 10000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_32fc_s32f_x2_power_spectral_density_32f, 1e-4, 2046, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_s32f_power_spectrum_32f, 1e-4, 0, 20462, 100, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_x2_square_dist_32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_x2_s32f_square_dist_scalar_mult_32f, 1e-4, 10, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_divide_32f, 1e-4, 0, 204602, 2000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_dot_prod_32f, 1e-4, 0, 204602, 5000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_dot_prod_16i, 1e-4, 0, 204602, 5000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_32f_s32f_32f_fm_detect_32f, 1e-4, 2046, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_index_max_16u, 3, 0, 204602, 5000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_s32f_interleave_16ic, 1, 32768, 204602, 3000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_interleave_32fc, 0, 0, 204602, 5000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_max_32f, 1e-4, 0, 204602, 2000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_min_32f, 1e-4, 0, 204602, 2000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_multiply_32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_s32f_normalize, 1e-4, 100, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_s32f_power_32f, 1e-4, 4, 204602, 100, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_sqrt_32f, 1e-4, 0, 204602, 100, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_s32f_stddev_32f, 1e-4, 100, 204602, 3000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_stddev_and_mean_32f_x2, 1e-4, 0, 204602, 3000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x2_subtract_32f, 1e-4, 0, 204602, 5000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_x3_sum_of_poly_32f, 1e-2, 0, 204602, 5000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32i_x2_and_32i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32i_s32f_convert_32f, 1e-4, 100, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32i_x2_or_32i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32u_byteswap, 0, 0, 204602, 2000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_32u_popcnt, 0, 0, 2046, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_64f_convert_32f, 1e-4, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_64f_x2_max_64f, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_64f_x2_min_64f, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_64u_byteswap, 0, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_64u_popcnt, 0, 0, 2046, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_8ic_deinterleave_16i_x2, 0, 0, 204602, 3000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_8ic_s32f_deinterleave_32f_x2, 1e-4, 100, 204602, 3000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_8ic_deinterleave_real_16i, 0, 256, 204602, 3000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_8ic_s32f_deinterleave_real_32f, 1e-4, 100, 204602, 3000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_8ic_deinterleave_real_8i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_8ic_x2_multiply_conjugate_16ic, 0, 0, 204602, 400, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_8ic_x2_s32f_multiply_conjugate_32fc, 1e-4, 100, 204602, 400, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_8i_convert_16i, 0, 0, 204602, 20000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_8i_s32f_convert_32f, 1e-4, 100, 204602, 2000, &results, benchmark_mode, kernel_regex); + //VOLK_PROFILE(volk_32fc_s32fc_multiply_32fc, 1e-4, lv_32fc_t(1.0, 0.5), 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32fc_s32fc_multiply_32fc, 1e-4, 0, 204602, 1000, &results, benchmark_mode, kernel_regex); + VOLK_PROFILE(volk_32f_s32f_multiply_32f, 1e-4, 1.0, 204602, 10000, &results, benchmark_mode, kernel_regex); char path[1024]; volk_get_config_path(path); diff --git a/volk/lib/qa_utils.cc b/volk/lib/qa_utils.cc index 8b95d7d6b0..8007fe75a6 100644 --- a/volk/lib/qa_utils.cc +++ b/volk/lib/qa_utils.cc @@ -3,6 +3,7 @@ #include <boost/foreach.hpp> #include <boost/assign/list_of.hpp> #include <boost/tokenizer.hpp> +#include <boost/xpressive/xpressive.hpp> #include <iostream> #include <vector> #include <list> @@ -322,8 +323,14 @@ bool run_volk_tests(volk_func_desc_t desc, int iter, std::vector<std::string> *best_arch_vector = 0, std::string puppet_master_name = "NULL", - bool benchmark_mode + bool benchmark_mode, + std::string kernel_regex ) { + boost::xpressive::sregex kernel_expression = boost::xpressive::sregex::compile(kernel_regex); + if( !boost::xpressive::regex_search(name, kernel_expression) ) { + // in this case we have a regex and are only looking to test one kernel + return false; + } std::cout << "RUN_VOLK_TESTS: " << name << "(" << vlen << "," << iter << ")" << std::endl; // The multiply and lv_force_cast_hf are work arounds for GNU Radio bugs 582 and 583 diff --git a/volk/lib/qa_utils.h b/volk/lib/qa_utils.h index 0ede962e41..fc1a0239eb 100644 --- a/volk/lib/qa_utils.h +++ b/volk/lib/qa_utils.h @@ -21,12 +21,12 @@ volk_type_t volk_type_from_string(std::string); float uniform(void); void random_floats(float *buf, unsigned n); -bool run_volk_tests(volk_func_desc_t, void(*)(), std::string, float, lv_32fc_t, int, int, std::vector<std::string> *, std::string, bool benchmark_mode=false); +bool run_volk_tests(volk_func_desc_t, void(*)(), std::string, float, lv_32fc_t, int, int, std::vector<std::string> *, std::string, bool benchmark_mode=false, std::string kernel_regex=""); #define VOLK_RUN_TESTS(func, tol, scalar, len, iter) BOOST_AUTO_TEST_CASE(func##_test) { BOOST_CHECK_EQUAL(run_volk_tests(func##_get_func_desc(), (void (*)())func##_manual, std::string(#func), tol, scalar, len, iter, 0, "NULL"), 0); } -#define VOLK_PROFILE(func, tol, scalar, len, iter, results, bnmode) run_volk_tests(func##_get_func_desc(), (void (*)())func##_manual, std::string(#func), tol, scalar, len, iter, results, "NULL", bnmode) -#define VOLK_PUPPET_PROFILE(func, puppet_master_func, tol, scalar, len, iter, results, bnmode) run_volk_tests(func##_get_func_desc(), (void (*)())func##_manual, std::string(#func), tol, scalar, len, iter, results, std::string(#puppet_master_func), bnmode) +#define VOLK_PROFILE(func, tol, scalar, len, iter, results, bnmode, kernel_regex) run_volk_tests(func##_get_func_desc(), (void (*)())func##_manual, std::string(#func), tol, scalar, len, iter, results, "NULL", bnmode, kernel_regex) +#define VOLK_PUPPET_PROFILE(func, puppet_master_func, tol, scalar, len, iter, results, bnmode, kernel_regex) run_volk_tests(func##_get_func_desc(), (void (*)())func##_manual, std::string(#func), tol, scalar, len, iter, results, std::string(#puppet_master_func), bnmode, kernel_regex) typedef void (*volk_fn_1arg)(void *, unsigned int, const char*); //one input, operate in place typedef void (*volk_fn_2arg)(void *, void *, unsigned int, const char*); typedef void (*volk_fn_3arg)(void *, void *, void *, unsigned int, const char*); |