diff options
-rw-r--r-- | volk/apps/volk_profile.cc | 152 | ||||
-rw-r--r-- | volk/gen/volk_register.py | 4 | ||||
-rw-r--r-- | volk/include/volk/Makefile.am | 152 | ||||
-rw-r--r-- | volk/lib/testqa.cc | 152 | ||||
-rw-r--r-- | volk/orc/Makefile.am | 42 | ||||
-rw-r--r-- | volk/orc/volk_16i_s32f_deinterleave_32f_x2_a_orc_impl.orc (renamed from volk/orc/volk_16i_s32f_deinterleave_32f_x2_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_16ic_deinterleave_16i_x2_a_orc_impl.orc (renamed from volk/orc/volk_16ic_deinterleave_16i_x2_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_16ic_deinterleave_real_8i_a_orc_impl.orc (renamed from volk/orc/volk_16ic_deinterleave_real_8i_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_16ic_magnitude_16i_a_orc_impl.orc (renamed from volk/orc/volk_16ic_magnitude_16i_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_16sc_magnitude_32f_aligned16_orc_impl.orc | 2 | ||||
-rw-r--r-- | volk/orc/volk_16u_byteswap_a16_orc_impl.orc | 3 | ||||
-rw-r--r-- | volk/orc/volk_16u_byteswap_a_orc_impl.orc | 3 | ||||
-rw-r--r-- | volk/orc/volk_32f_s32f_normalize_a_orc_impl.orc (renamed from volk/orc/volk_32f_s32f_normalize_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32f_sqrt_32f_a_orc_impl.orc (renamed from volk/orc/volk_32f_sqrt_32f_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32f_x2_add_32f_a_orc_impl.orc (renamed from volk/orc/volk_32f_x2_add_32f_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32f_x2_divide_32f_a_orc_impl.orc (renamed from volk/orc/volk_32f_x2_divide_32f_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32f_x2_dot_prod_32f_a_orc_impl.orc (renamed from volk/orc/volk_32f_x2_dot_prod_32f_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32f_x2_max_32f_a_orc_impl.orc (renamed from volk/orc/volk_32f_x2_max_32f_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32f_x2_min_32f_a_orc_impl.orc (renamed from volk/orc/volk_32f_x2_min_32f_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32f_x2_multiply_32f_a_orc_impl.orc (renamed from volk/orc/volk_32f_x2_multiply_32f_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32f_x2_subtract_32f_a_orc_impl.orc (renamed from volk/orc/volk_32f_x2_subtract_32f_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32fc_32f_multiply_32fc_a_orc_impl.orc (renamed from volk/orc/volk_32fc_32f_multiply_32fc_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32fc_magnitude_32f_a_orc_impl.orc (renamed from volk/orc/volk_32fc_magnitude_32f_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32fc_s32f_magnitude_16i_a_orc_impl.orc (renamed from volk/orc/volk_32fc_s32f_magnitude_16i_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32fc_x2_multiply_32fc_a_orc_impl.orc (renamed from volk/orc/volk_32fc_x2_multiply_32fc_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32i_x2_and_32i_a_orc_impl.orc (renamed from volk/orc/volk_32i_x2_and_32i_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_32i_x2_or_32i_a_orc_impl.orc (renamed from volk/orc/volk_32i_x2_or_32i_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_8i_convert_16i_a_orc_impl.orc (renamed from volk/orc/volk_8i_convert_16i_a16_orc_impl.orc) | 2 | ||||
-rw-r--r-- | volk/orc/volk_8i_s32f_convert_32f_a_orc_impl.orc (renamed from volk/orc/volk_8i_s32f_convert_32f_a16_orc_impl.orc) | 2 |
29 files changed, 276 insertions, 276 deletions
diff --git a/volk/apps/volk_profile.cc b/volk/apps/volk_profile.cc index fd9507207..b9ac9ecc2 100644 --- a/volk/apps/volk_profile.cc +++ b/volk/apps/volk_profile.cc @@ -14,93 +14,93 @@ int main(int argc, char *argv[]) { std::vector<std::string> results; - //VOLK_PROFILE(volk_16i_x5_add_quad_16i_x4_a16, 1e-4, 2046, 10000, &results); - //VOLK_PROFILE(volk_16i_branch_4_state_8_a16, 1e-4, 2046, 10000, &results); - VOLK_PROFILE(volk_16ic_s32f_deinterleave_real_32f_a16, 1e-5, 32768.0, 204600, 10000, &results); - VOLK_PROFILE(volk_16ic_deinterleave_real_8i_a16, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_16ic_deinterleave_16i_x2_a16, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_16ic_s32f_deinterleave_32f_x2_a16, 1e-4, 32768.0, 204600, 1000, &results); - VOLK_PROFILE(volk_16ic_deinterleave_real_16i_a16, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_16ic_magnitude_16i_a16, 1, 0, 204600, 100, &results); - VOLK_PROFILE(volk_16ic_s32f_magnitude_32f_a16, 1e-5, 32768.0, 204600, 1000, &results); - VOLK_PROFILE(volk_16i_s32f_convert_32f_a16, 1e-4, 32768.0, 204600, 10000, &results); + //VOLK_PROFILE(volk_16i_x5_add_quad_16i_x4_a, 1e-4, 2046, 10000, &results); + //VOLK_PROFILE(volk_16i_branch_4_state_8_a, 1e-4, 2046, 10000, &results); + VOLK_PROFILE(volk_16ic_s32f_deinterleave_real_32f_a, 1e-5, 32768.0, 204600, 10000, &results); + VOLK_PROFILE(volk_16ic_deinterleave_real_8i_a, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_16ic_deinterleave_16i_x2_a, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_16ic_s32f_deinterleave_32f_x2_a, 1e-4, 32768.0, 204600, 1000, &results); + VOLK_PROFILE(volk_16ic_deinterleave_real_16i_a, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_16ic_magnitude_16i_a, 1, 0, 204600, 100, &results); + VOLK_PROFILE(volk_16ic_s32f_magnitude_32f_a, 1e-5, 32768.0, 204600, 1000, &results); + VOLK_PROFILE(volk_16i_s32f_convert_32f_a, 1e-4, 32768.0, 204600, 10000, &results); VOLK_PROFILE(volk_16i_s32f_convert_32f_u, 1e-4, 32768.0, 204600, 10000, &results); - VOLK_PROFILE(volk_16i_convert_8i_a16, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_16i_convert_8i_a, 0, 0, 204600, 10000, &results); VOLK_PROFILE(volk_16i_convert_8i_u, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_16i_max_star_16i_a16, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_16i_max_star_horizontal_16i_a16, 0, 0, 204600, 10000, &results); - //VOLK_PROFILE(volk_16i_permute_and_scalar_add_a16, 1e-4, 0, 2046, 10000, &results); - //VOLK_PROFILE(volk_16i_x4_quad_max_star_16i_a16, 1e-4, 0, 2046, 10000, &results); - VOLK_PROFILE(volk_16u_byteswap_a16, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32f_accumulator_s32f_a16, 1e-4, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32f_x2_add_32f_a16, 1e-4, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32fc_32f_multiply_32fc_a16, 1e-4, 0, 204600, 1000, &results); - VOLK_PROFILE(volk_32fc_s32f_power_32fc_a16, 1e-4, 0, 204600, 50, &results); - VOLK_PROFILE(volk_32f_s32f_calc_spectral_noise_floor_32f_a16, 1e-4, 20.0, 204600, 1000, &results); - VOLK_PROFILE(volk_32fc_s32f_atan2_32f_a16, 1e-4, 10.0, 204600, 100, &results); - //VOLK_PROFILE(volk_32fc_x2_conjugate_dot_prod_32fc_a16, 1e-4, 0, 2046, 10000, &results); + VOLK_PROFILE(volk_16i_max_star_16i_a, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_16i_max_star_horizontal_16i_a, 0, 0, 204600, 10000, &results); + //VOLK_PROFILE(volk_16i_permute_and_scalar_add_a, 1e-4, 0, 2046, 10000, &results); + //VOLK_PROFILE(volk_16i_x4_quad_max_star_16i_a, 1e-4, 0, 2046, 10000, &results); + VOLK_PROFILE(volk_16u_byteswap_a, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_accumulator_s32f_a, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_x2_add_32f_a, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_32f_multiply_32fc_a, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_s32f_power_32fc_a, 1e-4, 0, 204600, 50, &results); + VOLK_PROFILE(volk_32f_s32f_calc_spectral_noise_floor_32f_a, 1e-4, 20.0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_s32f_atan2_32f_a, 1e-4, 10.0, 204600, 100, &results); + //VOLK_PROFILE(volk_32fc_x2_conjugate_dot_prod_32fc_a, 1e-4, 0, 2046, 10000, &results); VOLK_PROFILE(volk_32fc_x2_conjugate_dot_prod_32fc_u, 1e-4, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32fc_deinterleave_32f_x2_a16, 1e-4, 0, 204600, 1000, &results); - VOLK_PROFILE(volk_32fc_deinterleave_64f_x2_a16, 1e-4, 0, 204600, 1000, &results); - VOLK_PROFILE(volk_32fc_s32f_deinterleave_real_16i_a16, 0, 32768, 204600, 10000, &results); - VOLK_PROFILE(volk_32fc_deinterleave_real_32f_a16, 1e-4, 0, 204600, 5000, &results); - VOLK_PROFILE(volk_32fc_deinterleave_real_64f_a16, 1e-4, 0, 204600, 1000, &results); - VOLK_PROFILE(volk_32fc_x2_dot_prod_32fc_a16, 1e-4, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32fc_index_max_16u_a16, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32fc_s32f_magnitude_16i_a16, 1, 32768, 204600, 100, &results); - VOLK_PROFILE(volk_32fc_magnitude_32f_a16, 1e-4, 0, 204600, 1000, &results); - VOLK_PROFILE(volk_32fc_x2_multiply_32fc_a16, 1e-4, 0, 204600, 1000, &results); - VOLK_PROFILE(volk_32f_s32f_convert_16i_a16, 1, 32768, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_deinterleave_32f_x2_a, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_deinterleave_64f_x2_a, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_s32f_deinterleave_real_16i_a, 0, 32768, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_deinterleave_real_32f_a, 1e-4, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32fc_deinterleave_real_64f_a, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_x2_dot_prod_32fc_a, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_index_max_16u_a, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_s32f_magnitude_16i_a, 1, 32768, 204600, 100, &results); + VOLK_PROFILE(volk_32fc_magnitude_32f_a, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32fc_x2_multiply_32fc_a, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_32f_s32f_convert_16i_a, 1, 32768, 204600, 10000, &results); VOLK_PROFILE(volk_32f_s32f_convert_16i_u, 1, 32768, 204600, 10000, &results); - VOLK_PROFILE(volk_32f_s32f_convert_32i_a16, 1, 2<<31, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_s32f_convert_32i_a, 1, 2<<31, 204600, 10000, &results); VOLK_PROFILE(volk_32f_s32f_convert_32i_u, 1, 2<<31, 204600, 10000, &results); - VOLK_PROFILE(volk_32f_convert_64f_a16, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_convert_64f_a, 1e-4, 0, 204600, 10000, &results); VOLK_PROFILE(volk_32f_convert_64f_u, 1e-4, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32f_s32f_convert_8i_a16, 1, 128, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_s32f_convert_8i_a, 1, 128, 204600, 10000, &results); VOLK_PROFILE(volk_32f_s32f_convert_8i_u, 1, 128, 204600, 10000, &results); - //VOLK_PROFILE(volk_32fc_s32f_x2_power_spectral_density_32f_a16, 1e-4, 2046, 10000, &results); - VOLK_PROFILE(volk_32fc_s32f_power_spectrum_32f_a16, 1e-4, 0, 20460, 100, &results); - VOLK_PROFILE(volk_32fc_x2_square_dist_32f_a16, 1e-4, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a16, 1e-4, 10, 204600, 10000, &results); - VOLK_PROFILE(volk_32f_x2_divide_32f_a16, 1e-4, 0, 204600, 2000, &results); - VOLK_PROFILE(volk_32f_x2_dot_prod_32f_a16, 1e-4, 0, 204600, 5000, &results); + //VOLK_PROFILE(volk_32fc_s32f_x2_power_spectral_density_32f_a, 1e-4, 2046, 10000, &results); + VOLK_PROFILE(volk_32fc_s32f_power_spectrum_32f_a, 1e-4, 0, 20460, 100, &results); + VOLK_PROFILE(volk_32fc_x2_square_dist_32f_a, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a, 1e-4, 10, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_x2_divide_32f_a, 1e-4, 0, 204600, 2000, &results); + VOLK_PROFILE(volk_32f_x2_dot_prod_32f_a, 1e-4, 0, 204600, 5000, &results); VOLK_PROFILE(volk_32f_x2_dot_prod_32f_u, 1e-4, 0, 204600, 5000, &results); - //VOLK_PROFILE(volk_32f_s32f_32f_fm_detect_32f_a16, 1e-4, 2046, 10000, &results); - VOLK_PROFILE(volk_32f_index_max_16u_a16, 0, 0, 204600, 5000, &results); - VOLK_PROFILE(volk_32f_x2_s32f_interleave_16ic_a16, 1, 32768, 204600, 3000, &results); - VOLK_PROFILE(volk_32f_x2_interleave_32fc_a16, 0, 0, 204600, 5000, &results); - VOLK_PROFILE(volk_32f_x2_max_32f_a16, 1e-4, 0, 204600, 2000, &results); - VOLK_PROFILE(volk_32f_x2_min_32f_a16, 1e-4, 0, 204600, 2000, &results); - VOLK_PROFILE(volk_32f_x2_multiply_32f_a16, 1e-4, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32f_s32f_normalize_a16, 1e-4, 100, 204600, 10000, &results); - VOLK_PROFILE(volk_32f_s32f_power_32f_a16, 1e-4, 4, 204600, 100, &results); - VOLK_PROFILE(volk_32f_sqrt_32f_a16, 1e-4, 0, 204600, 100, &results); - VOLK_PROFILE(volk_32f_s32f_stddev_32f_a16, 1e-4, 100, 204600, 3000, &results); - VOLK_PROFILE(volk_32f_stddev_and_mean_32f_x2_a16, 1e-4, 0, 204600, 3000, &results); - VOLK_PROFILE(volk_32f_x2_subtract_32f_a16, 1e-4, 0, 204600, 5000, &results); - VOLK_PROFILE(volk_32f_x3_sum_of_poly_32f_a16, 1e-4, 0, 204600, 5000, &results); - VOLK_PROFILE(volk_32i_x2_and_32i_a16, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32i_s32f_convert_32f_a16, 1e-4, 100, 204600, 10000, &results); + //VOLK_PROFILE(volk_32f_s32f_32f_fm_detect_32f_a, 1e-4, 2046, 10000, &results); + VOLK_PROFILE(volk_32f_index_max_16u_a, 0, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32f_x2_s32f_interleave_16ic_a, 1, 32768, 204600, 3000, &results); + VOLK_PROFILE(volk_32f_x2_interleave_32fc_a, 0, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32f_x2_max_32f_a, 1e-4, 0, 204600, 2000, &results); + VOLK_PROFILE(volk_32f_x2_min_32f_a, 1e-4, 0, 204600, 2000, &results); + VOLK_PROFILE(volk_32f_x2_multiply_32f_a, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_s32f_normalize_a, 1e-4, 100, 204600, 10000, &results); + VOLK_PROFILE(volk_32f_s32f_power_32f_a, 1e-4, 4, 204600, 100, &results); + VOLK_PROFILE(volk_32f_sqrt_32f_a, 1e-4, 0, 204600, 100, &results); + VOLK_PROFILE(volk_32f_s32f_stddev_32f_a, 1e-4, 100, 204600, 3000, &results); + VOLK_PROFILE(volk_32f_stddev_and_mean_32f_x2_a, 1e-4, 0, 204600, 3000, &results); + VOLK_PROFILE(volk_32f_x2_subtract_32f_a, 1e-4, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32f_x3_sum_of_poly_32f_a, 1e-4, 0, 204600, 5000, &results); + VOLK_PROFILE(volk_32i_x2_and_32i_a, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32i_s32f_convert_32f_a, 1e-4, 100, 204600, 10000, &results); VOLK_PROFILE(volk_32i_s32f_convert_32f_u, 1e-4, 100, 204600, 10000, &results); - VOLK_PROFILE(volk_32i_x2_or_32i_a16, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_32u_byteswap_a16, 0, 0, 204600, 2000, &results); - //VOLK_PROFILE(volk_32u_popcnt_a16, 0, 0, 2046, 10000, &results); - VOLK_PROFILE(volk_64f_convert_32f_a16, 1e-4, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32i_x2_or_32i_a, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_32u_byteswap_a, 0, 0, 204600, 2000, &results); + //VOLK_PROFILE(volk_32u_popcnt_a, 0, 0, 2046, 10000, &results); + VOLK_PROFILE(volk_64f_convert_32f_a, 1e-4, 0, 204600, 10000, &results); VOLK_PROFILE(volk_64f_convert_32f_u, 1e-4, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_64f_x2_max_64f_a16, 1e-4, 0, 204600, 1000, &results); - VOLK_PROFILE(volk_64f_x2_min_64f_a16, 1e-4, 0, 204600, 1000, &results); - VOLK_PROFILE(volk_64u_byteswap_a16, 0, 0, 204600, 1000, &results); - //VOLK_PROFILE(volk_64u_popcnt_a16, 0, 0, 2046, 10000, &results); - VOLK_PROFILE(volk_8ic_deinterleave_16i_x2_a16, 0, 0, 204600, 3000, &results); - VOLK_PROFILE(volk_8ic_s32f_deinterleave_32f_x2_a16, 1e-4, 100, 204600, 3000, &results); - VOLK_PROFILE(volk_8ic_deinterleave_real_16i_a16, 0, 256, 204600, 3000, &results); - VOLK_PROFILE(volk_8ic_s32f_deinterleave_real_32f_a16, 1e-4, 100, 204600, 3000, &results); - VOLK_PROFILE(volk_8ic_deinterleave_real_8i_a16, 0, 0, 204600, 10000, &results); - VOLK_PROFILE(volk_8ic_x2_multiply_conjugate_16ic_a16, 0, 0, 204600, 400, &results); - VOLK_PROFILE(volk_8ic_x2_s32f_multiply_conjugate_32fc_a16, 1e-4, 100, 204600, 400, &results); - VOLK_PROFILE(volk_8i_convert_16i_a16, 0, 0, 204600, 20000, &results); + VOLK_PROFILE(volk_64f_x2_max_64f_a, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_64f_x2_min_64f_a, 1e-4, 0, 204600, 1000, &results); + VOLK_PROFILE(volk_64u_byteswap_a, 0, 0, 204600, 1000, &results); + //VOLK_PROFILE(volk_64u_popcnt_a, 0, 0, 2046, 10000, &results); + VOLK_PROFILE(volk_8ic_deinterleave_16i_x2_a, 0, 0, 204600, 3000, &results); + VOLK_PROFILE(volk_8ic_s32f_deinterleave_32f_x2_a, 1e-4, 100, 204600, 3000, &results); + VOLK_PROFILE(volk_8ic_deinterleave_real_16i_a, 0, 256, 204600, 3000, &results); + VOLK_PROFILE(volk_8ic_s32f_deinterleave_real_32f_a, 1e-4, 100, 204600, 3000, &results); + VOLK_PROFILE(volk_8ic_deinterleave_real_8i_a, 0, 0, 204600, 10000, &results); + VOLK_PROFILE(volk_8ic_x2_multiply_conjugate_16ic_a, 0, 0, 204600, 400, &results); + VOLK_PROFILE(volk_8ic_x2_s32f_multiply_conjugate_32fc_a, 1e-4, 100, 204600, 400, &results); + VOLK_PROFILE(volk_8i_convert_16i_a, 0, 0, 204600, 20000, &results); VOLK_PROFILE(volk_8i_convert_16i_u, 0, 0, 204600, 2000, &results); - VOLK_PROFILE(volk_8i_s32f_convert_32f_a16, 1e-4, 100, 204600, 2000, &results); + VOLK_PROFILE(volk_8i_s32f_convert_32f_a, 1e-4, 100, 204600, 2000, &results); VOLK_PROFILE(volk_8i_s32f_convert_32f_u, 1e-4, 100, 204600, 2000, &results); char path[256]; diff --git a/volk/gen/volk_register.py b/volk/gen/volk_register.py index 62ebba3c2..e87156fc8 100644 --- a/volk/gen/volk_register.py +++ b/volk/gen/volk_register.py @@ -59,7 +59,7 @@ functions = []; for line in mfile: - subline = re.search(".*_(a16|u)\.h.*", line); + subline = re.search(".*_(a|u)\.h.*", line); if subline: subsubline = re.search("(?<=volk_).*", subline.group(0)); if subsubline: @@ -74,7 +74,7 @@ datatypes = set(datatypes); for line in mfile: for dt in datatypes: if dt in line: - subline = re.search("(volk_" + dt +"_.*(a16|u).*\.h)", line); + subline = re.search("(volk_" + dt +"_.*(a|u).*\.h)", line); if subline: subsubline = re.search(".+(?=\.h)", subline.group(0)); diff --git a/volk/include/volk/Makefile.am b/volk/include/volk/Makefile.am index 5f9e134bc..b7da9b37c 100644 --- a/volk/include/volk/Makefile.am +++ b/volk/include/volk/Makefile.am @@ -33,91 +33,91 @@ volkinclude_HEADERS = \ $(top_gendir)/include/volk/volk_typedefs.h \ $(top_gendir)/include/volk/volk.h \ $(top_gendir)/include/volk/volk_cpu.h \ - volk_16i_x5_add_quad_16i_x4_a16.h \ - volk_16i_branch_4_state_8_a16.h \ - volk_16ic_deinterleave_16i_x2_a16.h \ - volk_16ic_s32f_deinterleave_32f_x2_a16.h \ - volk_16ic_deinterleave_real_16i_a16.h \ - volk_16ic_s32f_deinterleave_real_32f_a16.h \ - volk_16ic_deinterleave_real_8i_a16.h \ - volk_16ic_magnitude_16i_a16.h \ - volk_16ic_s32f_magnitude_32f_a16.h \ - volk_16i_s32f_convert_32f_a16.h \ + volk_16i_x5_add_quad_16i_x4_a.h \ + volk_16i_branch_4_state_8_a.h \ + volk_16ic_deinterleave_16i_x2_a.h \ + volk_16ic_s32f_deinterleave_32f_x2_a.h \ + volk_16ic_deinterleave_real_16i_a.h \ + volk_16ic_s32f_deinterleave_real_32f_a.h \ + volk_16ic_deinterleave_real_8i_a.h \ + volk_16ic_magnitude_16i_a.h \ + volk_16ic_s32f_magnitude_32f_a.h \ + volk_16i_s32f_convert_32f_a.h \ volk_16i_s32f_convert_32f_u.h \ - volk_16i_convert_8i_a16.h \ + volk_16i_convert_8i_a.h \ volk_16i_convert_8i_u.h \ - volk_16i_max_star_16i_a16.h \ - volk_16i_max_star_horizontal_16i_a16.h \ - volk_16i_permute_and_scalar_add_a16.h \ - volk_16i_x4_quad_max_star_16i_a16.h \ - volk_16u_byteswap_a16.h \ - volk_32f_accumulator_s32f_a16.h \ - volk_32f_x2_add_32f_a16.h \ - volk_32fc_32f_multiply_32fc_a16.h \ - volk_32fc_s32f_power_32fc_a16.h \ - volk_32f_s32f_calc_spectral_noise_floor_32f_a16.h \ - volk_32fc_s32f_atan2_32f_a16.h \ - volk_32fc_x2_conjugate_dot_prod_32fc_a16.h \ + volk_16i_max_star_16i_a.h \ + volk_16i_max_star_horizontal_16i_a.h \ + volk_16i_permute_and_scalar_add_a.h \ + volk_16i_x4_quad_max_star_16i_a.h \ + volk_16u_byteswap_a.h \ + volk_32f_accumulator_s32f_a.h \ + volk_32f_x2_add_32f_a.h \ + volk_32fc_32f_multiply_32fc_a.h \ + volk_32fc_s32f_power_32fc_a.h \ + volk_32f_s32f_calc_spectral_noise_floor_32f_a.h \ + volk_32fc_s32f_atan2_32f_a.h \ + volk_32fc_x2_conjugate_dot_prod_32fc_a.h \ volk_32fc_x2_conjugate_dot_prod_32fc_u.h \ - volk_32fc_deinterleave_32f_x2_a16.h \ - volk_32fc_deinterleave_64f_x2_a16.h \ - volk_32fc_s32f_deinterleave_real_16i_a16.h \ - volk_32fc_deinterleave_real_32f_a16.h \ - volk_32fc_deinterleave_real_64f_a16.h \ - volk_32fc_x2_dot_prod_32fc_a16.h \ - volk_32fc_index_max_16u_a16.h \ - volk_32fc_s32f_magnitude_16i_a16.h \ - volk_32fc_magnitude_32f_a16.h \ - volk_32fc_x2_multiply_32fc_a16.h \ - volk_32f_s32f_convert_16i_a16.h \ + volk_32fc_deinterleave_32f_x2_a.h \ + volk_32fc_deinterleave_64f_x2_a.h \ + volk_32fc_s32f_deinterleave_real_16i_a.h \ + volk_32fc_deinterleave_real_32f_a.h \ + volk_32fc_deinterleave_real_64f_a.h \ + volk_32fc_x2_dot_prod_32fc_a.h \ + volk_32fc_index_max_16u_a.h \ + volk_32fc_s32f_magnitude_16i_a.h \ + volk_32fc_magnitude_32f_a.h \ + volk_32fc_x2_multiply_32fc_a.h \ + volk_32f_s32f_convert_16i_a.h \ volk_32f_s32f_convert_16i_u.h \ - volk_32f_s32f_convert_32i_a16.h \ + volk_32f_s32f_convert_32i_a.h \ volk_32f_s32f_convert_32i_u.h \ - volk_32f_convert_64f_a16.h \ + volk_32f_convert_64f_a.h \ volk_32f_convert_64f_u.h \ - volk_32f_s32f_convert_8i_a16.h \ + volk_32f_s32f_convert_8i_a.h \ volk_32f_s32f_convert_8i_u.h \ - volk_32fc_s32f_x2_power_spectral_density_32f_a16.h \ - volk_32fc_s32f_power_spectrum_32f_a16.h \ - volk_32fc_x2_square_dist_32f_a16.h \ - volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a16.h \ - volk_32f_x2_divide_32f_a16.h \ - volk_32f_x2_dot_prod_32f_a16.h \ + volk_32fc_s32f_x2_power_spectral_density_32f_a.h \ + volk_32fc_s32f_power_spectrum_32f_a.h \ + volk_32fc_x2_square_dist_32f_a.h \ + volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a.h \ + volk_32f_x2_divide_32f_a.h \ + volk_32f_x2_dot_prod_32f_a.h \ volk_32f_x2_dot_prod_32f_u.h \ - volk_32f_s32f_32f_fm_detect_32f_a16.h \ - volk_32f_index_max_16u_a16.h \ - volk_32f_x2_s32f_interleave_16ic_a16.h \ - volk_32f_x2_interleave_32fc_a16.h \ - volk_32f_x2_max_32f_a16.h \ - volk_32f_x2_min_32f_a16.h \ - volk_32f_x2_multiply_32f_a16.h \ - volk_32f_s32f_normalize_a16.h \ - volk_32f_s32f_power_32f_a16.h \ - volk_32f_sqrt_32f_a16.h \ - volk_32f_s32f_stddev_32f_a16.h \ - volk_32f_stddev_and_mean_32f_x2_a16.h \ - volk_32f_x2_subtract_32f_a16.h \ - volk_32f_x3_sum_of_poly_32f_a16.h \ - volk_32i_x2_and_32i_a16.h \ - volk_32i_s32f_convert_32f_a16.h \ + volk_32f_s32f_32f_fm_detect_32f_a.h \ + volk_32f_index_max_16u_a.h \ + volk_32f_x2_s32f_interleave_16ic_a.h \ + volk_32f_x2_interleave_32fc_a.h \ + volk_32f_x2_max_32f_a.h \ + volk_32f_x2_min_32f_a.h \ + volk_32f_x2_multiply_32f_a.h \ + volk_32f_s32f_normalize_a.h \ + volk_32f_s32f_power_32f_a.h \ + volk_32f_sqrt_32f_a.h \ + volk_32f_s32f_stddev_32f_a.h \ + volk_32f_stddev_and_mean_32f_x2_a.h \ + volk_32f_x2_subtract_32f_a.h \ + volk_32f_x3_sum_of_poly_32f_a.h \ + volk_32i_x2_and_32i_a.h \ + volk_32i_s32f_convert_32f_a.h \ volk_32i_s32f_convert_32f_u.h \ - volk_32i_x2_or_32i_a16.h \ - volk_32u_byteswap_a16.h \ - volk_32u_popcnt_a16.h \ - volk_64f_convert_32f_a16.h \ + volk_32i_x2_or_32i_a.h \ + volk_32u_byteswap_a.h \ + volk_32u_popcnt_a.h \ + volk_64f_convert_32f_a.h \ volk_64f_convert_32f_u.h \ - volk_64f_x2_max_64f_a16.h \ - volk_64f_x2_min_64f_a16.h \ - volk_64u_byteswap_a16.h \ - volk_64u_popcnt_a16.h \ - volk_8ic_deinterleave_16i_x2_a16.h \ - volk_8ic_s32f_deinterleave_32f_x2_a16.h \ - volk_8ic_deinterleave_real_16i_a16.h \ - volk_8ic_s32f_deinterleave_real_32f_a16.h \ - volk_8ic_deinterleave_real_8i_a16.h \ - volk_8ic_x2_multiply_conjugate_16ic_a16.h \ - volk_8ic_x2_s32f_multiply_conjugate_32fc_a16.h \ - volk_8i_convert_16i_a16.h \ + volk_64f_x2_max_64f_a.h \ + volk_64f_x2_min_64f_a.h \ + volk_64u_byteswap_a.h \ + volk_64u_popcnt_a.h \ + volk_8ic_deinterleave_16i_x2_a.h \ + volk_8ic_s32f_deinterleave_32f_x2_a.h \ + volk_8ic_deinterleave_real_16i_a.h \ + volk_8ic_s32f_deinterleave_real_32f_a.h \ + volk_8ic_deinterleave_real_8i_a.h \ + volk_8ic_x2_multiply_conjugate_16ic_a.h \ + volk_8ic_x2_s32f_multiply_conjugate_32fc_a.h \ + volk_8i_convert_16i_a.h \ volk_8i_convert_16i_u.h \ - volk_8i_s32f_convert_32f_a16.h \ + volk_8i_s32f_convert_32f_a.h \ volk_8i_s32f_convert_32f_u.h diff --git a/volk/lib/testqa.cc b/volk/lib/testqa.cc index 349fb0630..98b09c50f 100644 --- a/volk/lib/testqa.cc +++ b/volk/lib/testqa.cc @@ -2,92 +2,92 @@ #include <volk/volk.h> #include <boost/test/unit_test.hpp> -//VOLK_RUN_TESTS(volk_16i_x5_add_quad_16i_x4_a16, 1e-4, 2046, 10000); -//VOLK_RUN_TESTS(volk_16i_branch_4_state_8_a16, 1e-4, 2046, 10000); -VOLK_RUN_TESTS(volk_16ic_s32f_deinterleave_real_32f_a16, 1e-5, 32768.0, 204600, 10000); -VOLK_RUN_TESTS(volk_16ic_deinterleave_real_8i_a16, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_16ic_deinterleave_16i_x2_a16, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_16ic_s32f_deinterleave_32f_x2_a16, 1e-4, 32768.0, 204600, 1000); -VOLK_RUN_TESTS(volk_16ic_deinterleave_real_16i_a16, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_16ic_magnitude_16i_a16, 1, 0, 204600, 100); -VOLK_RUN_TESTS(volk_16ic_s32f_magnitude_32f_a16, 1e-5, 32768.0, 204600, 1000); -VOLK_RUN_TESTS(volk_16i_s32f_convert_32f_a16, 1e-4, 32768.0, 204600, 10000); +//VOLK_RUN_TESTS(volk_16i_x5_add_quad_16i_x4_a, 1e-4, 2046, 10000); +//VOLK_RUN_TESTS(volk_16i_branch_4_state_8_a, 1e-4, 2046, 10000); +VOLK_RUN_TESTS(volk_16ic_s32f_deinterleave_real_32f_a, 1e-5, 32768.0, 204600, 10000); +VOLK_RUN_TESTS(volk_16ic_deinterleave_real_8i_a, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_16ic_deinterleave_16i_x2_a, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_16ic_s32f_deinterleave_32f_x2_a, 1e-4, 32768.0, 204600, 1000); +VOLK_RUN_TESTS(volk_16ic_deinterleave_real_16i_a, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_16ic_magnitude_16i_a, 1, 0, 204600, 100); +VOLK_RUN_TESTS(volk_16ic_s32f_magnitude_32f_a, 1e-5, 32768.0, 204600, 1000); +VOLK_RUN_TESTS(volk_16i_s32f_convert_32f_a, 1e-4, 32768.0, 204600, 10000); VOLK_RUN_TESTS(volk_16i_s32f_convert_32f_u, 1e-4, 32768.0, 204600, 10000); -VOLK_RUN_TESTS(volk_16i_convert_8i_a16, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_16i_convert_8i_a, 0, 0, 204600, 10000); VOLK_RUN_TESTS(volk_16i_convert_8i_u, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_16i_max_star_16i_a16, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_16i_max_star_horizontal_16i_a16, 0, 0, 204600, 10000); -//VOLK_RUN_TESTS(volk_16i_permute_and_scalar_add_a16, 1e-4, 0, 2046, 10000); -//VOLK_RUN_TESTS(volk_16i_x4_quad_max_star_16i_a16, 1e-4, 0, 2046, 10000); -VOLK_RUN_TESTS(volk_16u_byteswap_a16, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32f_accumulator_s32f_a16, 1e-4, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32f_x2_add_32f_a16, 1e-4, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32fc_32f_multiply_32fc_a16, 1e-4, 0, 204600, 1000); -VOLK_RUN_TESTS(volk_32fc_s32f_power_32fc_a16, 1e-4, 0, 204600, 50); -VOLK_RUN_TESTS(volk_32f_s32f_calc_spectral_noise_floor_32f_a16, 1e-4, 20.0, 204600, 1000); -VOLK_RUN_TESTS(volk_32fc_s32f_atan2_32f_a16, 1e-4, 10.0, 204600, 100); -//VOLK_RUN_TESTS(volk_32fc_x2_conjugate_dot_prod_32fc_a16, 1e-4, 0, 2046, 10000); +VOLK_RUN_TESTS(volk_16i_max_star_16i_a, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_16i_max_star_horizontal_16i_a, 0, 0, 204600, 10000); +//VOLK_RUN_TESTS(volk_16i_permute_and_scalar_add_a, 1e-4, 0, 2046, 10000); +//VOLK_RUN_TESTS(volk_16i_x4_quad_max_star_16i_a, 1e-4, 0, 2046, 10000); +VOLK_RUN_TESTS(volk_16u_byteswap_a, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32f_accumulator_s32f_a, 1e-4, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32f_x2_add_32f_a, 1e-4, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32fc_32f_multiply_32fc_a, 1e-4, 0, 204600, 1000); +VOLK_RUN_TESTS(volk_32fc_s32f_power_32fc_a, 1e-4, 0, 204600, 50); +VOLK_RUN_TESTS(volk_32f_s32f_calc_spectral_noise_floor_32f_a, 1e-4, 20.0, 204600, 1000); +VOLK_RUN_TESTS(volk_32fc_s32f_atan2_32f_a, 1e-4, 10.0, 204600, 100); +//VOLK_RUN_TESTS(volk_32fc_x2_conjugate_dot_prod_32fc_a, 1e-4, 0, 2046, 10000); VOLK_RUN_TESTS(volk_32fc_x2_conjugate_dot_prod_32fc_u, 1e-4, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32fc_deinterleave_32f_x2_a16, 1e-4, 0, 204600, 1000); -VOLK_RUN_TESTS(volk_32fc_deinterleave_64f_x2_a16, 1e-4, 0, 204600, 1000); -VOLK_RUN_TESTS(volk_32fc_s32f_deinterleave_real_16i_a16, 0, 32768, 204600, 10000); -VOLK_RUN_TESTS(volk_32fc_deinterleave_real_32f_a16, 1e-4, 0, 204600, 5000); -VOLK_RUN_TESTS(volk_32fc_deinterleave_real_64f_a16, 1e-4, 0, 204600, 1000); -VOLK_RUN_TESTS(volk_32fc_x2_dot_prod_32fc_a16, 1e-4, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32fc_index_max_16u_a16, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32fc_s32f_magnitude_16i_a16, 1, 32768, 204600, 100); -VOLK_RUN_TESTS(volk_32fc_magnitude_32f_a16, 1e-4, 0, 204600, 1000); -VOLK_RUN_TESTS(volk_32fc_x2_multiply_32fc_a16, 1e-4, 0, 204600, 1000); -VOLK_RUN_TESTS(volk_32f_s32f_convert_16i_a16, 1, 32768, 204600, 10000); +VOLK_RUN_TESTS(volk_32fc_deinterleave_32f_x2_a, 1e-4, 0, 204600, 1000); +VOLK_RUN_TESTS(volk_32fc_deinterleave_64f_x2_a, 1e-4, 0, 204600, 1000); +VOLK_RUN_TESTS(volk_32fc_s32f_deinterleave_real_16i_a, 0, 32768, 204600, 10000); +VOLK_RUN_TESTS(volk_32fc_deinterleave_real_32f_a, 1e-4, 0, 204600, 5000); +VOLK_RUN_TESTS(volk_32fc_deinterleave_real_64f_a, 1e-4, 0, 204600, 1000); +VOLK_RUN_TESTS(volk_32fc_x2_dot_prod_32fc_a, 1e-4, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32fc_index_max_16u_a, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32fc_s32f_magnitude_16i_a, 1, 32768, 204600, 100); +VOLK_RUN_TESTS(volk_32fc_magnitude_32f_a, 1e-4, 0, 204600, 1000); +VOLK_RUN_TESTS(volk_32fc_x2_multiply_32fc_a, 1e-4, 0, 204600, 1000); +VOLK_RUN_TESTS(volk_32f_s32f_convert_16i_a, 1, 32768, 204600, 10000); VOLK_RUN_TESTS(volk_32f_s32f_convert_16i_u, 1, 32768, 204600, 10000); -VOLK_RUN_TESTS(volk_32f_s32f_convert_32i_a16, 1, 2<<31, 204600, 10000); +VOLK_RUN_TESTS(volk_32f_s32f_convert_32i_a, 1, 2<<31, 204600, 10000); VOLK_RUN_TESTS(volk_32f_s32f_convert_32i_u, 1, 2<<31, 204600, 10000); -VOLK_RUN_TESTS(volk_32f_convert_64f_a16, 1e-4, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32f_convert_64f_a, 1e-4, 0, 204600, 10000); VOLK_RUN_TESTS(volk_32f_convert_64f_u, 1e-4, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32f_s32f_convert_8i_a16, 1, 128, 204600, 10000); +VOLK_RUN_TESTS(volk_32f_s32f_convert_8i_a, 1, 128, 204600, 10000); VOLK_RUN_TESTS(volk_32f_s32f_convert_8i_u, 1, 128, 204600, 10000); -//VOLK_RUN_TESTS(volk_32fc_s32f_x2_power_spectral_density_32f_a16, 1e-4, 2046, 10000); -VOLK_RUN_TESTS(volk_32fc_s32f_power_spectrum_32f_a16, 1e-4, 0, 20460, 100); -VOLK_RUN_TESTS(volk_32fc_x2_square_dist_32f_a16, 1e-4, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a16, 1e-4, 10, 204600, 10000); -VOLK_RUN_TESTS(volk_32f_x2_divide_32f_a16, 1e-4, 0, 204600, 2000); -VOLK_RUN_TESTS(volk_32f_x2_dot_prod_32f_a16, 1e-4, 0, 204600, 5000); +//VOLK_RUN_TESTS(volk_32fc_s32f_x2_power_spectral_density_32f_a, 1e-4, 2046, 10000); +VOLK_RUN_TESTS(volk_32fc_s32f_power_spectrum_32f_a, 1e-4, 0, 20460, 100); +VOLK_RUN_TESTS(volk_32fc_x2_square_dist_32f_a, 1e-4, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a, 1e-4, 10, 204600, 10000); +VOLK_RUN_TESTS(volk_32f_x2_divide_32f_a, 1e-4, 0, 204600, 2000); +VOLK_RUN_TESTS(volk_32f_x2_dot_prod_32f_a, 1e-4, 0, 204600, 5000); VOLK_RUN_TESTS(volk_32f_x2_dot_prod_32f_u, 1e-4, 0, 204600, 5000); -//VOLK_RUN_TESTS(volk_32f_s32f_32f_fm_detect_32f_a16, 1e-4, 2046, 10000); -VOLK_RUN_TESTS(volk_32f_index_max_16u_a16, 0, 0, 204600, 5000); -VOLK_RUN_TESTS(volk_32f_x2_s32f_interleave_16ic_a16, 1, 32768, 204600, 3000); -VOLK_RUN_TESTS(volk_32f_x2_interleave_32fc_a16, 0, 0, 204600, 5000); -VOLK_RUN_TESTS(volk_32f_x2_max_32f_a16, 1e-4, 0, 204600, 2000); -VOLK_RUN_TESTS(volk_32f_x2_min_32f_a16, 1e-4, 0, 204600, 2000); -VOLK_RUN_TESTS(volk_32f_x2_multiply_32f_a16, 1e-4, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32f_s32f_normalize_a16, 1e-4, 100, 204600, 10000); -VOLK_RUN_TESTS(volk_32f_s32f_power_32f_a16, 1e-4, 4, 204600, 100); -VOLK_RUN_TESTS(volk_32f_sqrt_32f_a16, 1e-4, 0, 204600, 100); -VOLK_RUN_TESTS(volk_32f_s32f_stddev_32f_a16, 1e-4, 100, 204600, 3000); -VOLK_RUN_TESTS(volk_32f_stddev_and_mean_32f_x2_a16, 1e-4, 0, 204600, 3000); -VOLK_RUN_TESTS(volk_32f_x2_subtract_32f_a16, 1e-4, 0, 204600, 5000); -VOLK_RUN_TESTS(volk_32f_x3_sum_of_poly_32f_a16, 1e-4, 0, 204600, 5000); -VOLK_RUN_TESTS(volk_32i_x2_and_32i_a16, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32i_s32f_convert_32f_a16, 1e-4, 100, 204600, 10000); +//VOLK_RUN_TESTS(volk_32f_s32f_32f_fm_detect_32f_a, 1e-4, 2046, 10000); +VOLK_RUN_TESTS(volk_32f_index_max_16u_a, 0, 0, 204600, 5000); +VOLK_RUN_TESTS(volk_32f_x2_s32f_interleave_16ic_a, 1, 32768, 204600, 3000); +VOLK_RUN_TESTS(volk_32f_x2_interleave_32fc_a, 0, 0, 204600, 5000); +VOLK_RUN_TESTS(volk_32f_x2_max_32f_a, 1e-4, 0, 204600, 2000); +VOLK_RUN_TESTS(volk_32f_x2_min_32f_a, 1e-4, 0, 204600, 2000); +VOLK_RUN_TESTS(volk_32f_x2_multiply_32f_a, 1e-4, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32f_s32f_normalize_a, 1e-4, 100, 204600, 10000); +VOLK_RUN_TESTS(volk_32f_s32f_power_32f_a, 1e-4, 4, 204600, 100); +VOLK_RUN_TESTS(volk_32f_sqrt_32f_a, 1e-4, 0, 204600, 100); +VOLK_RUN_TESTS(volk_32f_s32f_stddev_32f_a, 1e-4, 100, 204600, 3000); +VOLK_RUN_TESTS(volk_32f_stddev_and_mean_32f_x2_a, 1e-4, 0, 204600, 3000); +VOLK_RUN_TESTS(volk_32f_x2_subtract_32f_a, 1e-4, 0, 204600, 5000); +VOLK_RUN_TESTS(volk_32f_x3_sum_of_poly_32f_a, 1e-4, 0, 204600, 5000); +VOLK_RUN_TESTS(volk_32i_x2_and_32i_a, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32i_s32f_convert_32f_a, 1e-4, 100, 204600, 10000); VOLK_RUN_TESTS(volk_32i_s32f_convert_32f_u, 1e-4, 100, 204600, 10000); -VOLK_RUN_TESTS(volk_32i_x2_or_32i_a16, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_32u_byteswap_a16, 0, 0, 204600, 2000); -//VOLK_RUN_TESTS(volk_32u_popcnt_a16, 0, 0, 2046, 10000); -VOLK_RUN_TESTS(volk_64f_convert_32f_a16, 1e-4, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32i_x2_or_32i_a, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_32u_byteswap_a, 0, 0, 204600, 2000); +//VOLK_RUN_TESTS(volk_32u_popcnt_a, 0, 0, 2046, 10000); +VOLK_RUN_TESTS(volk_64f_convert_32f_a, 1e-4, 0, 204600, 10000); VOLK_RUN_TESTS(volk_64f_convert_32f_u, 1e-4, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_64f_x2_max_64f_a16, 1e-4, 0, 204600, 1000); -VOLK_RUN_TESTS(volk_64f_x2_min_64f_a16, 1e-4, 0, 204600, 1000); -VOLK_RUN_TESTS(volk_64u_byteswap_a16, 0, 0, 204600, 1000); -//VOLK_RUN_TESTS(volk_64u_popcnt_a16, 0, 0, 2046, 10000); -VOLK_RUN_TESTS(volk_8ic_deinterleave_16i_x2_a16, 0, 0, 204600, 3000); -VOLK_RUN_TESTS(volk_8ic_s32f_deinterleave_32f_x2_a16, 1e-4, 100, 204600, 3000); -VOLK_RUN_TESTS(volk_8ic_deinterleave_real_16i_a16, 0, 256, 204600, 3000); -VOLK_RUN_TESTS(volk_8ic_s32f_deinterleave_real_32f_a16, 1e-4, 100, 204600, 3000); -VOLK_RUN_TESTS(volk_8ic_deinterleave_real_8i_a16, 0, 0, 204600, 10000); -VOLK_RUN_TESTS(volk_8ic_x2_multiply_conjugate_16ic_a16, 0, 0, 204600, 400); -VOLK_RUN_TESTS(volk_8ic_x2_s32f_multiply_conjugate_32fc_a16, 1e-4, 100, 204600, 400); -VOLK_RUN_TESTS(volk_8i_convert_16i_a16, 0, 0, 204600, 20000); +VOLK_RUN_TESTS(volk_64f_x2_max_64f_a, 1e-4, 0, 204600, 1000); +VOLK_RUN_TESTS(volk_64f_x2_min_64f_a, 1e-4, 0, 204600, 1000); +VOLK_RUN_TESTS(volk_64u_byteswap_a, 0, 0, 204600, 1000); +//VOLK_RUN_TESTS(volk_64u_popcnt_a, 0, 0, 2046, 10000); +VOLK_RUN_TESTS(volk_8ic_deinterleave_16i_x2_a, 0, 0, 204600, 3000); +VOLK_RUN_TESTS(volk_8ic_s32f_deinterleave_32f_x2_a, 1e-4, 100, 204600, 3000); +VOLK_RUN_TESTS(volk_8ic_deinterleave_real_16i_a, 0, 256, 204600, 3000); +VOLK_RUN_TESTS(volk_8ic_s32f_deinterleave_real_32f_a, 1e-4, 100, 204600, 3000); +VOLK_RUN_TESTS(volk_8ic_deinterleave_real_8i_a, 0, 0, 204600, 10000); +VOLK_RUN_TESTS(volk_8ic_x2_multiply_conjugate_16ic_a, 0, 0, 204600, 400); +VOLK_RUN_TESTS(volk_8ic_x2_s32f_multiply_conjugate_32fc_a, 1e-4, 100, 204600, 400); +VOLK_RUN_TESTS(volk_8i_convert_16i_a, 0, 0, 204600, 20000); VOLK_RUN_TESTS(volk_8i_convert_16i_u, 0, 0, 204600, 2000); -VOLK_RUN_TESTS(volk_8i_s32f_convert_32f_a16, 1e-4, 100, 204600, 2000); +VOLK_RUN_TESTS(volk_8i_s32f_convert_32f_a, 1e-4, 100, 204600, 2000); VOLK_RUN_TESTS(volk_8i_s32f_convert_32f_u, 1e-4, 100, 204600, 2000); diff --git a/volk/orc/Makefile.am b/volk/orc/Makefile.am index 960d09b6a..814471bd0 100644 --- a/volk/orc/Makefile.am +++ b/volk/orc/Makefile.am @@ -25,27 +25,27 @@ noinst_LTLIBRARIES = libvolk_orc.la libvolk_orc_la_LDFLAGS = $(ORC_LDFLAGS) libvolk_orc_la_SOURCES = \ -volk_8i_convert_16i_a16_orc_impl.orc \ -volk_8i_s32f_convert_32f_a16_orc_impl.orc \ -volk_16u_byteswap_a16_orc_impl.orc \ -volk_32i_x2_and_32i_a16_orc_impl.orc \ -volk_32i_x2_or_32i_a16_orc_impl.orc \ -volk_32f_x2_add_32f_a16_orc_impl.orc \ -volk_32f_x2_subtract_32f_a16_orc_impl.orc \ -volk_32f_x2_divide_32f_a16_orc_impl.orc \ -volk_32f_x2_multiply_32f_a16_orc_impl.orc \ -volk_32fc_x2_multiply_32fc_a16_orc_impl.orc \ -volk_32fc_32f_multiply_32fc_a16_orc_impl.orc \ -volk_32f_sqrt_32f_a16_orc_impl.orc \ -volk_32f_x2_max_32f_a16_orc_impl.orc \ -volk_32f_x2_min_32f_a16_orc_impl.orc \ -volk_32f_s32f_normalize_a16_orc_impl.orc \ -volk_32fc_magnitude_32f_a16_orc_impl.orc \ -volk_32fc_s32f_magnitude_16i_a16_orc_impl.orc \ -volk_16ic_magnitude_16i_a16_orc_impl.orc \ -volk_16ic_deinterleave_16i_x2_a16_orc_impl.orc \ -volk_16i_s32f_deinterleave_32f_x2_a16_orc_impl.orc \ -volk_16ic_deinterleave_real_8i_a16_orc_impl.orc +volk_8i_convert_16i_a_orc_impl.orc \ +volk_8i_s32f_convert_32f_a_orc_impl.orc \ +volk_16u_byteswap_a_orc_impl.orc \ +volk_32i_x2_and_32i_a_orc_impl.orc \ +volk_32i_x2_or_32i_a_orc_impl.orc \ +volk_32f_x2_add_32f_a_orc_impl.orc \ +volk_32f_x2_subtract_32f_a_orc_impl.orc \ +volk_32f_x2_divide_32f_a_orc_impl.orc \ +volk_32f_x2_multiply_32f_a_orc_impl.orc \ +volk_32fc_x2_multiply_32fc_a_orc_impl.orc \ +volk_32fc_32f_multiply_32fc_a_orc_impl.orc \ +volk_32f_sqrt_32f_a_orc_impl.orc \ +volk_32f_x2_max_32f_a_orc_impl.orc \ +volk_32f_x2_min_32f_a_orc_impl.orc \ +volk_32f_s32f_normalize_a_orc_impl.orc \ +volk_32fc_magnitude_32f_a_orc_impl.orc \ +volk_32fc_s32f_magnitude_16i_a_orc_impl.orc \ +volk_16ic_magnitude_16i_a_orc_impl.orc \ +volk_16ic_deinterleave_16i_x2_a_orc_impl.orc \ +volk_16i_s32f_deinterleave_32f_x2_a_orc_impl.orc \ +volk_16ic_deinterleave_real_8i_a_orc_impl.orc diff --git a/volk/orc/volk_16i_s32f_deinterleave_32f_x2_a16_orc_impl.orc b/volk/orc/volk_16i_s32f_deinterleave_32f_x2_a_orc_impl.orc index 0189fbf5d..fd8915da0 100644 --- a/volk/orc/volk_16i_s32f_deinterleave_32f_x2_a16_orc_impl.orc +++ b/volk/orc/volk_16i_s32f_deinterleave_32f_x2_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_16ic_s32f_deinterleave_32f_x2_a16_orc_impl +.function volk_16ic_s32f_deinterleave_32f_x2_a_orc_impl .dest 4 idst .dest 4 qdst .source 4 src diff --git a/volk/orc/volk_16ic_deinterleave_16i_x2_a16_orc_impl.orc b/volk/orc/volk_16ic_deinterleave_16i_x2_a_orc_impl.orc index 56018edda..76faa936a 100644 --- a/volk/orc/volk_16ic_deinterleave_16i_x2_a16_orc_impl.orc +++ b/volk/orc/volk_16ic_deinterleave_16i_x2_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_16ic_deinterleave_16i_x2_a16_orc_impl +.function volk_16ic_deinterleave_16i_x2_a_orc_impl .dest 2 idst .dest 2 qdst .source 4 src diff --git a/volk/orc/volk_16ic_deinterleave_real_8i_a16_orc_impl.orc b/volk/orc/volk_16ic_deinterleave_real_8i_a_orc_impl.orc index dba9a4c8e..8db49fd7c 100644 --- a/volk/orc/volk_16ic_deinterleave_real_8i_a16_orc_impl.orc +++ b/volk/orc/volk_16ic_deinterleave_real_8i_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_16ic_deinterleave_real_8i_a16_orc_impl +.function volk_16ic_deinterleave_real_8i_a_orc_impl .dest 1 dst .source 4 src .temp 2 iw diff --git a/volk/orc/volk_16ic_magnitude_16i_a16_orc_impl.orc b/volk/orc/volk_16ic_magnitude_16i_a_orc_impl.orc index 37225e9b8..fbaebc46d 100644 --- a/volk/orc/volk_16ic_magnitude_16i_a16_orc_impl.orc +++ b/volk/orc/volk_16ic_magnitude_16i_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_16ic_magnitude_16i_a16_orc_impl +.function volk_16ic_magnitude_16i_a_orc_impl .source 4 src .dest 2 dst .floatparam 4 scalar diff --git a/volk/orc/volk_16sc_magnitude_32f_aligned16_orc_impl.orc b/volk/orc/volk_16sc_magnitude_32f_aligned16_orc_impl.orc index 1e2380837..66fef7d2e 100644 --- a/volk/orc/volk_16sc_magnitude_32f_aligned16_orc_impl.orc +++ b/volk/orc/volk_16sc_magnitude_32f_aligned16_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_16ic_magnitude_32f_a16_orc_impl +.function volk_16ic_magnitude_32f_a_orc_impl .source 4 src .dest 4 dst .floatparam 4 scalar diff --git a/volk/orc/volk_16u_byteswap_a16_orc_impl.orc b/volk/orc/volk_16u_byteswap_a16_orc_impl.orc deleted file mode 100644 index c1c8ee59e..000000000 --- a/volk/orc/volk_16u_byteswap_a16_orc_impl.orc +++ /dev/null @@ -1,3 +0,0 @@ -.function volk_16u_byteswap_a16_orc_impl -.dest 2 dst -swapw dst, dst diff --git a/volk/orc/volk_16u_byteswap_a_orc_impl.orc b/volk/orc/volk_16u_byteswap_a_orc_impl.orc new file mode 100644 index 000000000..b96ba84af --- /dev/null +++ b/volk/orc/volk_16u_byteswap_a_orc_impl.orc @@ -0,0 +1,3 @@ +.function volk_16u_byteswap_a_orc_impl +.dest 2 dst +swapw dst, dst diff --git a/volk/orc/volk_32f_s32f_normalize_a16_orc_impl.orc b/volk/orc/volk_32f_s32f_normalize_a_orc_impl.orc index acd319b16..986fdf665 100644 --- a/volk/orc/volk_32f_s32f_normalize_a16_orc_impl.orc +++ b/volk/orc/volk_32f_s32f_normalize_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32f_s32f_normalize_a16_orc_impl +.function volk_32f_s32f_normalize_a_orc_impl .source 4 src1 .floatparam 4 invscalar .dest 4 dst diff --git a/volk/orc/volk_32f_sqrt_32f_a16_orc_impl.orc b/volk/orc/volk_32f_sqrt_32f_a_orc_impl.orc index ae5680f15..f339b1122 100644 --- a/volk/orc/volk_32f_sqrt_32f_a16_orc_impl.orc +++ b/volk/orc/volk_32f_sqrt_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32f_sqrt_32f_a16_orc_impl +.function volk_32f_sqrt_32f_a_orc_impl .source 4 src .dest 4 dst sqrtf dst, src diff --git a/volk/orc/volk_32f_x2_add_32f_a16_orc_impl.orc b/volk/orc/volk_32f_x2_add_32f_a_orc_impl.orc index 8d095a052..450cc6a9e 100644 --- a/volk/orc/volk_32f_x2_add_32f_a16_orc_impl.orc +++ b/volk/orc/volk_32f_x2_add_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32f_x2_add_32f_a16_orc_impl +.function volk_32f_x2_add_32f_a_orc_impl .dest 4 dst .source 4 src1 .source 4 src2 diff --git a/volk/orc/volk_32f_x2_divide_32f_a16_orc_impl.orc b/volk/orc/volk_32f_x2_divide_32f_a_orc_impl.orc index 0097646cb..ee3b61b82 100644 --- a/volk/orc/volk_32f_x2_divide_32f_a16_orc_impl.orc +++ b/volk/orc/volk_32f_x2_divide_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32f_x2_divide_32f_a16_orc_impl +.function volk_32f_x2_divide_32f_a_orc_impl .dest 4 dst .source 4 src1 .source 4 src2 diff --git a/volk/orc/volk_32f_x2_dot_prod_32f_a16_orc_impl.orc b/volk/orc/volk_32f_x2_dot_prod_32f_a_orc_impl.orc index d92afbe01..b367f3091 100644 --- a/volk/orc/volk_32f_x2_dot_prod_32f_a16_orc_impl.orc +++ b/volk/orc/volk_32f_x2_dot_prod_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32f_x2_dot_prod_32f_a16_orc_impl +.function volk_32f_x2_dot_prod_32f_a_orc_impl .source 4 src1 .source 4 src2 .dest 4 dst diff --git a/volk/orc/volk_32f_x2_max_32f_a16_orc_impl.orc b/volk/orc/volk_32f_x2_max_32f_a_orc_impl.orc index b7f008737..725201633 100644 --- a/volk/orc/volk_32f_x2_max_32f_a16_orc_impl.orc +++ b/volk/orc/volk_32f_x2_max_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32f_x2_max_32f_a16_orc_impl +.function volk_32f_x2_max_32f_a_orc_impl .dest 4 dst .source 4 src1 .source 4 src2 diff --git a/volk/orc/volk_32f_x2_min_32f_a16_orc_impl.orc b/volk/orc/volk_32f_x2_min_32f_a_orc_impl.orc index 78328b576..a71ed8250 100644 --- a/volk/orc/volk_32f_x2_min_32f_a16_orc_impl.orc +++ b/volk/orc/volk_32f_x2_min_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32f_x2_min_32f_a16_orc_impl +.function volk_32f_x2_min_32f_a_orc_impl .dest 4 dst .source 4 src1 .source 4 src2 diff --git a/volk/orc/volk_32f_x2_multiply_32f_a16_orc_impl.orc b/volk/orc/volk_32f_x2_multiply_32f_a_orc_impl.orc index e8fadff19..c17d539fd 100644 --- a/volk/orc/volk_32f_x2_multiply_32f_a16_orc_impl.orc +++ b/volk/orc/volk_32f_x2_multiply_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32f_x2_multiply_32f_a16_orc_impl +.function volk_32f_x2_multiply_32f_a_orc_impl .dest 4 dst .source 4 src1 .source 4 src2 diff --git a/volk/orc/volk_32f_x2_subtract_32f_a16_orc_impl.orc b/volk/orc/volk_32f_x2_subtract_32f_a_orc_impl.orc index 13fbe8c83..b3b0f256e 100644 --- a/volk/orc/volk_32f_x2_subtract_32f_a16_orc_impl.orc +++ b/volk/orc/volk_32f_x2_subtract_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32f_x2_subtract_32f_a16_orc_impl +.function volk_32f_x2_subtract_32f_a_orc_impl .dest 4 dst .source 4 src1 .source 4 src2 diff --git a/volk/orc/volk_32fc_32f_multiply_32fc_a16_orc_impl.orc b/volk/orc/volk_32fc_32f_multiply_32fc_a_orc_impl.orc index 455293cff..aa82699f5 100644 --- a/volk/orc/volk_32fc_32f_multiply_32fc_a16_orc_impl.orc +++ b/volk/orc/volk_32fc_32f_multiply_32fc_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32fc_32f_multiply_32fc_a16_orc_impl +.function volk_32fc_32f_multiply_32fc_a_orc_impl .source 8 src1 .source 4 src2 .dest 8 dst diff --git a/volk/orc/volk_32fc_magnitude_32f_a16_orc_impl.orc b/volk/orc/volk_32fc_magnitude_32f_a_orc_impl.orc index c5e2e57f1..032ab2b1b 100644 --- a/volk/orc/volk_32fc_magnitude_32f_a16_orc_impl.orc +++ b/volk/orc/volk_32fc_magnitude_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32fc_magnitude_32f_a16_orc_impl +.function volk_32fc_magnitude_32f_a_orc_impl .source 8 src .dest 4 dst .temp 8 iqf diff --git a/volk/orc/volk_32fc_s32f_magnitude_16i_a16_orc_impl.orc b/volk/orc/volk_32fc_s32f_magnitude_16i_a_orc_impl.orc index 6116f5e1f..505e73f5d 100644 --- a/volk/orc/volk_32fc_s32f_magnitude_16i_a16_orc_impl.orc +++ b/volk/orc/volk_32fc_s32f_magnitude_16i_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32fc_s32f_magnitude_16i_a16_orc_impl +.function volk_32fc_s32f_magnitude_16i_a_orc_impl .source 8 src .dest 2 dst .floatparam 4 scalar diff --git a/volk/orc/volk_32fc_x2_multiply_32fc_a16_orc_impl.orc b/volk/orc/volk_32fc_x2_multiply_32fc_a_orc_impl.orc index a27d722cd..cb8a12d81 100644 --- a/volk/orc/volk_32fc_x2_multiply_32fc_a16_orc_impl.orc +++ b/volk/orc/volk_32fc_x2_multiply_32fc_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32fc_x2_multiply_32fc_a16_orc_impl +.function volk_32fc_x2_multiply_32fc_a_orc_impl .source 8 src1 .source 8 src2 .dest 8 dst diff --git a/volk/orc/volk_32i_x2_and_32i_a16_orc_impl.orc b/volk/orc/volk_32i_x2_and_32i_a_orc_impl.orc index 7b331f8ed..1845e4654 100644 --- a/volk/orc/volk_32i_x2_and_32i_a16_orc_impl.orc +++ b/volk/orc/volk_32i_x2_and_32i_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32i_x2_and_32i_a16_orc_impl +.function volk_32i_x2_and_32i_a_orc_impl .dest 4 dst .source 4 src1 .source 4 src2 diff --git a/volk/orc/volk_32i_x2_or_32i_a16_orc_impl.orc b/volk/orc/volk_32i_x2_or_32i_a_orc_impl.orc index 4984a9ced..004663f42 100644 --- a/volk/orc/volk_32i_x2_or_32i_a16_orc_impl.orc +++ b/volk/orc/volk_32i_x2_or_32i_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_32i_x2_or_32i_a16_orc_impl +.function volk_32i_x2_or_32i_a_orc_impl .dest 4 dst .source 4 src1 .source 4 src2 diff --git a/volk/orc/volk_8i_convert_16i_a16_orc_impl.orc b/volk/orc/volk_8i_convert_16i_a_orc_impl.orc index f44845c88..d813c6cfa 100644 --- a/volk/orc/volk_8i_convert_16i_a16_orc_impl.orc +++ b/volk/orc/volk_8i_convert_16i_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_8i_convert_16i_a16_orc_impl +.function volk_8i_convert_16i_a_orc_impl .source 1 src .dest 2 dst convsbw dst, src diff --git a/volk/orc/volk_8i_s32f_convert_32f_a16_orc_impl.orc b/volk/orc/volk_8i_s32f_convert_32f_a_orc_impl.orc index 8f6e157e9..ad54fb1e1 100644 --- a/volk/orc/volk_8i_s32f_convert_32f_a16_orc_impl.orc +++ b/volk/orc/volk_8i_s32f_convert_32f_a_orc_impl.orc @@ -1,4 +1,4 @@ -.function volk_8i_s32f_convert_32f_a16_orc_impl +.function volk_8i_s32f_convert_32f_a_orc_impl .source 1 src .dest 4 dst .floatparam 4 scalar |