diff options
Diffstat (limited to 'volk/apps/volk_profile.cc')
| -rw-r--r-- | volk/apps/volk_profile.cc | 138 |
1 files changed, 138 insertions, 0 deletions
diff --git a/volk/apps/volk_profile.cc b/volk/apps/volk_profile.cc new file mode 100644 index 000000000..fa89a93bf --- /dev/null +++ b/volk/apps/volk_profile.cc @@ -0,0 +1,138 @@ +#include "qa_utils.h" + +#include <volk/volk.h> +#include <volk/volk_prefs.h> + +#include <ciso646> +#include <vector> +#include <boost/foreach.hpp> +#include <boost/filesystem.hpp> +#include <iostream> +#include <fstream> +#include <sys/stat.h> +#include <sys/types.h> + +namespace fs = boost::filesystem; + +int main(int argc, char *argv[]) { + + std::vector<std::string> results; + + //VOLK_PROFILE(volk_16i_x5_add_quad_16i_x4, 1e-4, 2046, 10000, &results); + //VOLK_PROFILE(volk_16i_branch_4_state_8, 1e-4, 2046, 10000, &results); + VOLK_PUPPET_PROFILE(volk_32fc_s32fc_rotatorpuppet_32fc, volk_32fc_s32fc_x2_rotator_32fc, 1e-2, (lv_32fc_t)lv_cmake(.95393, .3), 20460, 10000, &results); + VOLK_PROFILE(volk_16ic_s32f_deinterleave_real_32f, 1e-5, 32768.0, 204600, 10000, &results); + VOLK_PROFILE(volk_16ic_deinterleave_real_8i, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_16ic_deinterleave_16i_x2, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_16ic_s32f_deinterleave_32f_x2, 1e-4, 32768.0, 204600, 1000, &results); + VOLK_PROFILE(volk_16ic_deinterleave_real_16i, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_16ic_magnitude_16i, 1, 0, 204600, 100, &results); + VOLK_PROFILE(volk_16ic_s32f_magnitude_32f, 1e-5, 32768.0, 204600, 1000, &results); + VOLK_PROFILE(volk_16i_s32f_convert_32f, 1e-4, 32768.0, 204600, 10000, &results); + VOLK_PROFILE(volk_16i_convert_8i, 0, 0, 204600, 10000, &results); + //VOLK_PROFILE(volk_16i_max_star_16i, 0, 0, 204600, 10000, &results); + //VOLK_PROFILE(volk_16i_max_star_horizontal_16i, 0, 0, 204600, 10000, &results); + //VOLK_PROFILE(volk_16i_permute_and_scalar_add, 1e-4, 0, 2046, 10000, &results); + //VOLK_PROFILE(volk_16i_x4_quad_max_star_16i, 1e-4, 0, 2046, 10000, &results); + VOLK_PROFILE(volk_16u_byteswap, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_16i_32fc_dot_prod_32fc, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_accumulator_s32f, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_x2_add_32f, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_32f_multiply_32fc, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_s32f_power_32fc, 1e-4, 0, 204600, 50, &results); + VOLK_PROFILE(volk_32f_s32f_calc_spectral_noise_floor_32f, 1e-4, 20.0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_s32f_atan2_32f, 1e-4, 10.0, 204600, 100, &results); + //VOLK_PROFILE(volk_32fc_x2_conjugate_dot_prod_32fc, 1e-4, 0, 2046, 10000, &results); + VOLK_PROFILE(volk_32fc_x2_conjugate_dot_prod_32fc, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_deinterleave_32f_x2, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_deinterleave_64f_x2, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_s32f_deinterleave_real_16i, 0, 32768, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_deinterleave_imag_32f, 1e-4, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32fc_deinterleave_real_32f, 1e-4, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32fc_deinterleave_real_64f, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_x2_dot_prod_32fc, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_32f_dot_prod_32fc, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_index_max_16u, 3, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_s32f_magnitude_16i, 1, 32768, 204600, 100, &results); + VOLK_PROFILE(volk_32fc_magnitude_32f, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_magnitude_squared_32f, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_x2_multiply_32fc, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_x2_multiply_conjugate_32fc, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_conjugate_32fc, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32f_s32f_convert_16i, 1, 32768, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_s32f_convert_32i, 1, 2<<31, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_convert_64f, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_s32f_convert_8i, 1, 128, 204600, 10000, &results); + //VOLK_PROFILE(volk_32fc_s32f_x2_power_spectral_density_32f, 1e-4, 2046, 10000, &results); + VOLK_PROFILE(volk_32fc_s32f_power_spectrum_32f, 1e-4, 0, 20460, 100, &results); + VOLK_PROFILE(volk_32fc_x2_square_dist_32f, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_x2_s32f_square_dist_scalar_mult_32f, 1e-4, 10, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_x2_divide_32f, 1e-4, 0, 204600, 2000, &results); + VOLK_PROFILE(volk_32f_x2_dot_prod_32f, 1e-4, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32f_x2_dot_prod_16i, 1e-4, 0, 204600, 5000, &results); + //VOLK_PROFILE(volk_32f_s32f_32f_fm_detect_32f, 1e-4, 2046, 10000, &results); + VOLK_PROFILE(volk_32f_index_max_16u, 3, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32f_x2_s32f_interleave_16ic, 1, 32768, 204600, 3000, &results); + VOLK_PROFILE(volk_32f_x2_interleave_32fc, 0, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32f_x2_max_32f, 1e-4, 0, 204600, 2000, &results); + VOLK_PROFILE(volk_32f_x2_min_32f, 1e-4, 0, 204600, 2000, &results); + VOLK_PROFILE(volk_32f_x2_multiply_32f, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_s32f_normalize, 1e-4, 100, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_s32f_power_32f, 1e-4, 4, 204600, 100, &results); + VOLK_PROFILE(volk_32f_sqrt_32f, 1e-4, 0, 204600, 100, &results); + VOLK_PROFILE(volk_32f_s32f_stddev_32f, 1e-4, 100, 204600, 3000, &results); + VOLK_PROFILE(volk_32f_stddev_and_mean_32f_x2, 1e-4, 0, 204600, 3000, &results); + VOLK_PROFILE(volk_32f_x2_subtract_32f, 1e-4, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32f_x3_sum_of_poly_32f, 1e-4, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32i_x2_and_32i, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32i_s32f_convert_32f, 1e-4, 100, 204600, 10000, &results); + VOLK_PROFILE(volk_32i_x2_or_32i, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32u_byteswap, 0, 0, 204600, 2000, &results); + //VOLK_PROFILE(volk_32u_popcnt, 0, 0, 2046, 10000, &results); + VOLK_PROFILE(volk_64f_convert_32f, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_64f_x2_max_64f, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_64f_x2_min_64f, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_64u_byteswap, 0, 0, 204600, 1000, &results); + //VOLK_PROFILE(volk_64u_popcnt, 0, 0, 2046, 10000, &results); + VOLK_PROFILE(volk_8ic_deinterleave_16i_x2, 0, 0, 204600, 3000, &results); + VOLK_PROFILE(volk_8ic_s32f_deinterleave_32f_x2, 1e-4, 100, 204600, 3000, &results); + VOLK_PROFILE(volk_8ic_deinterleave_real_16i, 0, 256, 204600, 3000, &results); + VOLK_PROFILE(volk_8ic_s32f_deinterleave_real_32f, 1e-4, 100, 204600, 3000, &results); + VOLK_PROFILE(volk_8ic_deinterleave_real_8i, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_8ic_x2_multiply_conjugate_16ic, 0, 0, 204600, 400, &results); + VOLK_PROFILE(volk_8ic_x2_s32f_multiply_conjugate_32fc, 1e-4, 100, 204600, 400, &results); + VOLK_PROFILE(volk_8i_convert_16i, 0, 0, 204600, 20000, &results); + VOLK_PROFILE(volk_8i_convert_16i, 0, 0, 204600, 2000, &results); + VOLK_PROFILE(volk_8i_s32f_convert_32f, 1e-4, 100, 204600, 2000, &results); + //VOLK_PROFILE(volk_32fc_s32fc_multiply_32fc, 1e-4, lv_32fc_t(1.0, 0.5), 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_s32fc_multiply_32fc, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32f_s32f_multiply_32f, 1e-4, 1.0, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_s32f_multiply_32f, 1e-4, 0, 204600, 1000, &results); + + + char path[1024]; + volk_get_config_path(path); + const fs::path config_path(path); + + if (not fs::exists(config_path.branch_path())) + { + std::cout << "Creating " << config_path.branch_path() << "..." << std::endl; + fs::create_directories(config_path.branch_path()); + } + + std::cout << "Writing " << config_path << "..." << std::endl; + std::ofstream config(config_path.string().c_str()); + if(!config.is_open()) { //either we don't have write access or we don't have the dir yet + std::cout << "Error opening file " << config_path << std::endl; + } + + config << "\ +#this file is generated by volk_profile.\n\ +#the function name is followed by the preferred architecture.\n\ +"; + + BOOST_FOREACH(std::string result, results) { + config << result << std::endl; + } + config.close(); +} |