Merge remote branch 'nick/volk_cmake'

4 files changed, 92 insertions, 90 deletions

diff --git a/volk/lib/CMakeLists.txt b/volk/lib/CMakeLists.txt
index 5dd41be0f..33a478265 100644
--- a/volk/lib/CMakeLists.txt
+++ b/volk/lib/CMakeLists.txt
@@ -131,6 +131,7 @@ SET(volk_gen_sources
     ${CMAKE_BINARY_DIR}/include/volk/volk.h
     ${CMAKE_BINARY_DIR}/lib/volk.c
     ${CMAKE_BINARY_DIR}/lib/volk_init.h
+    ${CMAKE_BINARY_DIR}/include/volk/volk_typedefs.h
     ${CMAKE_BINARY_DIR}/include/volk/volk_cpu.h
     ${CMAKE_BINARY_DIR}/lib/volk_cpu.c
     ${CMAKE_BINARY_DIR}/include/volk/volk_config_fixed.h
diff --git a/volk/lib/qa_utils.cc b/volk/lib/qa_utils.cc
index db606a472..7f86dd78b 100644
--- a/volk/lib/qa_utils.cc
+++ b/volk/lib/qa_utils.cc
@@ -233,7 +233,8 @@ bool icompare(t *in1, t *in2, unsigned int vlen, unsigned int tol) {
 
 class volk_qa_aligned_mem_pool{
 public:
-    void *get_new(size_t size, size_t alignment = 32){
+    void *get_new(size_t size){
+        size_t alignment = volk_get_alignment();
         _mems.push_back(std::vector<char>(size + alignment-1, 0));
         size_t ptr = size_t(&_mems.back().front());
         return (void *)((ptr + alignment-1) & ~(alignment-1));
diff --git a/volk/lib/testqa.cc b/volk/lib/testqa.cc
index 349fb0630..62e62c2f4 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_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_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_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_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_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_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_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_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_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_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_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_u, 1e-4, 100, 204600, 2000);
+//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, 20460, 10000);
+VOLK_RUN_TESTS(volk_16ic_deinterleave_16i_x2_a, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_16ic_s32f_deinterleave_32f_x2_a, 1e-4, 32768.0, 20460, 1000);
+VOLK_RUN_TESTS(volk_16ic_deinterleave_real_16i_a, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_16ic_magnitude_16i_a, 1, 0, 20460, 100);
+VOLK_RUN_TESTS(volk_16ic_s32f_magnitude_32f_a, 1e-5, 32768.0, 20460, 1000);
+VOLK_RUN_TESTS(volk_16i_s32f_convert_32f_a, 1e-4, 32768.0, 20460, 10000);
+VOLK_RUN_TESTS(volk_16i_s32f_convert_32f_u, 1e-4, 32768.0, 20460, 10000);
+VOLK_RUN_TESTS(volk_16i_convert_8i_a, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_16i_convert_8i_u, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_16i_max_star_16i_a, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_16i_max_star_horizontal_16i_a, 0, 0, 20460, 10000);
+//VOLK_RUN_TESTS(volk_16i_permute_and_scalar_add_a, 1e-4, 0, 2046, 1000);
+//VOLK_RUN_TESTS(volk_16i_x4_quad_max_star_16i_a, 1e-4, 0, 2046, 1000);
+VOLK_RUN_TESTS(volk_16u_byteswap_a, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_accumulator_s32f_a, 1e-4, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_x2_add_32f_a, 1e-4, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32fc_32f_multiply_32fc_a, 1e-4, 0, 20460, 1000);
+VOLK_RUN_TESTS(volk_32fc_s32f_power_32fc_a, 1e-4, 0, 20460, 50);
+VOLK_RUN_TESTS(volk_32f_s32f_calc_spectral_noise_floor_32f_a, 1e-4, 20.0, 20460, 1000);
+VOLK_RUN_TESTS(volk_32fc_s32f_atan2_32f_a, 1e-4, 10.0, 20460, 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, 20460, 10000);
+VOLK_RUN_TESTS(volk_32fc_deinterleave_32f_x2_a, 1e-4, 0, 20460, 1000);
+VOLK_RUN_TESTS(volk_32fc_deinterleave_64f_x2_a, 1e-4, 0, 20460, 1000);
+VOLK_RUN_TESTS(volk_32fc_s32f_deinterleave_real_16i_a, 0, 32768, 20460, 10000);
+VOLK_RUN_TESTS(volk_32fc_deinterleave_real_32f_a, 1e-4, 0, 20460, 5000);
+VOLK_RUN_TESTS(volk_32fc_deinterleave_real_64f_a, 1e-4, 0, 20460, 1000);
+VOLK_RUN_TESTS(volk_32fc_x2_dot_prod_32fc_a, 1e-4, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32fc_index_max_16u_a, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32fc_s32f_magnitude_16i_a, 1, 32768, 20460, 100);
+VOLK_RUN_TESTS(volk_32fc_magnitude_32f_a, 1e-4, 0, 20460, 1000);
+VOLK_RUN_TESTS(volk_32fc_x2_multiply_32fc_a, 1e-4, 0, 20460, 1000);
+VOLK_RUN_TESTS(volk_32f_s32f_convert_16i_a, 1, 32768, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_s32f_convert_16i_u, 1, 32768, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_s32f_convert_32i_a, 1, 2<<31, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_s32f_convert_32i_u, 1, 2<<31, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_convert_64f_a, 1e-4, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_convert_64f_u, 1e-4, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_s32f_convert_8i_a, 1, 128, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_s32f_convert_8i_u, 1, 128, 20460, 10000);
+//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, 2046, 100);
+VOLK_RUN_TESTS(volk_32fc_x2_square_dist_32f_a, 1e-4, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32fc_x2_s32f_square_dist_scalar_mult_32f_a, 1e-4, 10, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_x2_divide_32f_a, 1e-4, 0, 20460, 2000);
+VOLK_RUN_TESTS(volk_32f_x2_dot_prod_32f_a, 1e-4, 0, 20460, 5000);
+VOLK_RUN_TESTS(volk_32f_x2_dot_prod_32f_u, 1e-4, 0, 20460, 5000);
+//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, 20460, 5000);
+VOLK_RUN_TESTS(volk_32f_x2_s32f_interleave_16ic_a, 1, 32768, 20460, 3000);
+VOLK_RUN_TESTS(volk_32f_x2_interleave_32fc_a, 0, 0, 20460, 5000);
+VOLK_RUN_TESTS(volk_32f_x2_max_32f_a, 1e-4, 0, 20460, 2000);
+VOLK_RUN_TESTS(volk_32f_x2_min_32f_a, 1e-4, 0, 20460, 2000);
+VOLK_RUN_TESTS(volk_32f_x2_multiply_32f_a, 1e-4, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_s32f_normalize_a, 1e-4, 100, 20460, 10000);
+VOLK_RUN_TESTS(volk_32f_s32f_power_32f_a, 1e-4, 4, 20460, 100);
+VOLK_RUN_TESTS(volk_32f_sqrt_32f_a, 1e-4, 0, 20460, 100);
+VOLK_RUN_TESTS(volk_32f_s32f_stddev_32f_a, 1e-4, 100, 20460, 3000);
+VOLK_RUN_TESTS(volk_32f_stddev_and_mean_32f_x2_a, 1e-4, 0, 20460, 3000);
+VOLK_RUN_TESTS(volk_32f_x2_subtract_32f_a, 1e-4, 0, 20460, 5000);
+VOLK_RUN_TESTS(volk_32f_x3_sum_of_poly_32f_a, 1e-4, 0, 20460, 5000);
+VOLK_RUN_TESTS(volk_32i_x2_and_32i_a, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32i_s32f_convert_32f_a, 1e-4, 100, 20460, 10000);
+VOLK_RUN_TESTS(volk_32i_s32f_convert_32f_u, 1e-4, 100, 20460, 10000);
+VOLK_RUN_TESTS(volk_32i_x2_or_32i_a, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_32u_byteswap_a, 0, 0, 20460, 2000);
+//VOLK_RUN_TESTS(volk_32u_popcnt_a, 0, 0, 2046, 10000);
+VOLK_RUN_TESTS(volk_64f_convert_32f_a, 1e-4, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_64f_convert_32f_u, 1e-4, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_64f_x2_max_64f_a, 1e-4, 0, 20460, 1000);
+VOLK_RUN_TESTS(volk_64f_x2_min_64f_a, 1e-4, 0, 20460, 1000);
+VOLK_RUN_TESTS(volk_64u_byteswap_a, 0, 0, 20460, 1000);
+//VOLK_RUN_TESTS(volk_64u_popcnt_a, 0, 0, 2046, 10000);
+VOLK_RUN_TESTS(volk_8ic_deinterleave_16i_x2_a, 0, 0, 20460, 3000);
+VOLK_RUN_TESTS(volk_8ic_s32f_deinterleave_32f_x2_a, 1e-4, 100, 20460, 3000);
+VOLK_RUN_TESTS(volk_8ic_deinterleave_real_16i_a, 0, 256, 20460, 3000);
+VOLK_RUN_TESTS(volk_8ic_s32f_deinterleave_real_32f_a, 1e-4, 100, 20460, 3000);
+VOLK_RUN_TESTS(volk_8ic_deinterleave_real_8i_a, 0, 0, 20460, 10000);
+VOLK_RUN_TESTS(volk_8ic_x2_multiply_conjugate_16ic_a, 0, 0, 20460, 400);
+VOLK_RUN_TESTS(volk_8ic_x2_s32f_multiply_conjugate_32fc_a, 1e-4, 100, 20460, 400);
+VOLK_RUN_TESTS(volk_8i_convert_16i_a, 0, 0, 20460, 20000);
+VOLK_RUN_TESTS(volk_8i_convert_16i_u, 0, 0, 20460, 2000);
+VOLK_RUN_TESTS(volk_8i_s32f_convert_32f_a, 1e-4, 100, 20460, 2000);
+VOLK_RUN_TESTS(volk_8i_s32f_convert_32f_u, 1e-4, 100, 20460, 2000);
 
diff --git a/volk/lib/volk_prefs.c b/volk/lib/volk_prefs.c
index ebfe3bc40..9743c51d9 100644
--- a/volk/lib/volk_prefs.c
+++ b/volk/lib/volk_prefs.c
@@ -13,7 +13,7 @@ void get_config_path(char *path) {
     strcat(path, suffix);
 }
 
-//passing by reference in C can suck my balls
+//passing by reference in C can (***********)
 int load_preferences(struct volk_arch_pref **prefs) {
     FILE *config_file;
     char path[512], line[512], function[128], arch[32];