1
0
mirror of https://github.com/gnss-sdr/gnss-sdr synced 2024-12-15 04:30:33 +00:00

Moving two kernels to volk_gnsssdr. Still no testing

This commit is contained in:
Carles Fernandez 2016-01-13 19:38:07 +01:00
parent f88f222ef6
commit ae2b594c3b
8 changed files with 340 additions and 268 deletions

View File

@ -48,7 +48,7 @@
\param bVector One of the vectors to be multiplied and accumulated \param bVector One of the vectors to be multiplied and accumulated
\param num_points The number of complex values in aVector and bVector to be multiplied together, accumulated and stored into cVector \param num_points The number of complex values in aVector and bVector to be multiplied together, accumulated and stored into cVector
*/ */
static inline void volk_gnsssdr_16ic_xn_dot_prod_16ic_xn_generic(lv_16sc_t* result, const lv_16sc_t* in_common, const lv_16sc_t** in_a, unsigned int num_points, int num_a_vectors) static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_generic(lv_16sc_t* result, const lv_16sc_t* in_common, const lv_16sc_t** in_a, unsigned int num_points, int num_a_vectors)
{ {
for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
{ {
@ -68,7 +68,7 @@ static inline void volk_gnsssdr_16ic_xn_dot_prod_16ic_xn_generic(lv_16sc_t* resu
#ifdef LV_HAVE_SSE2 #ifdef LV_HAVE_SSE2
#include <emmintrin.h> #include <emmintrin.h>
static inline void volk_gnsssdr_16ic_xn_dot_prod_16ic_xn_a_sse2(lv_16sc_t* out, const lv_16sc_t* in_common, const lv_16sc_t** in_a, unsigned int num_points, int num_a_vectors) static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_sse2(lv_16sc_t* out, const lv_16sc_t* in_common, const lv_16sc_t** in_a, unsigned int num_points, int num_a_vectors)
{ {
lv_16sc_t dotProduct = lv_cmake(0,0); lv_16sc_t dotProduct = lv_cmake(0,0);
@ -148,8 +148,8 @@ static inline void volk_gnsssdr_16ic_xn_dot_prod_16ic_xn_a_sse2(lv_16sc_t* out,
for (int n_vec = 0; n_vec < num_a_vectors; n_vec++) for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
{ {
for(unsigned int n = sse_iters * 4;n < num_points; n++){ for(unsigned int n = sse_iters * 4; n < num_points; n++)
{
lv_16sc_t tmp = in_common[n] * in_a[n_vec][n]; lv_16sc_t tmp = in_common[n] * in_a[n_vec][n];
_out[n_vec] = lv_cmake(sat_adds16b(lv_creal(_out[n_vec]), lv_creal(tmp)), _out[n_vec] = lv_cmake(sat_adds16b(lv_creal(_out[n_vec]), lv_creal(tmp)),

View File

@ -0,0 +1,171 @@
/*!
* \file volk_gnsssdr_16ic_xn_resampler_16ic_xn.h
* \brief Volk protokernel: resample a 16 bits complex vector
* \authors <ul>
* <li> Javier Arribas, 2015. jarribas(at)cttc.es
* </ul>
*
* Volk protokernel that multiplies two 16 bits vectors (8 bits the real part
* and 8 bits the imaginary part) and accumulates them
*
* -------------------------------------------------------------------------
*
* Copyright (C) 2010-2015 (see AUTHORS file for a list of contributors)
*
* GNSS-SDR is a software defined Global Navigation
* Satellite Systems receiver
*
* This file is part of GNSS-SDR.
*
* GNSS-SDR is free software: you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* GNSS-SDR is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with GNSS-SDR. If not, see <http://www.gnu.org/licenses/>.
*
* -------------------------------------------------------------------------
*/
#ifndef INCLUDED_volk_gnsssdr_16ic_xn_resampler_16ic_xn_a_H
#define INCLUDED_volk_gnsssdr_16ic_xn_resampler_16ic_xn_a_H
#include <math.h>
#include <volk_gnsssdr/volk_gnsssdr_common.h>
#include <volk_gnsssdr/volk_gnsssdr_complex.h>
//#pragma STDC FENV_ACCESS ON
#ifdef LV_HAVE_GENERIC
//int round_int( float r ) {
// return (r > 0.0) ? (r + 0.5) : (r - 0.5);
//}
/*!
\brief Multiplies the two input complex vectors, point-by-point, storing the result in the third vector
\param cVector The vector where the result will be stored
\param aVector One of the vectors to be multiplied
\param bVector One of the vectors to be multiplied
\param num_points The number of complex values in aVector and bVector to be multiplied together, accumulated and stored into cVector
*/
static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_generic(lv_16sc_t** result, const lv_16sc_t* local_code, float* rem_code_phase_chips ,float code_phase_step_chips, unsigned int num_output_samples, unsigned int code_length_chips, int num_out_vectors)
{
int local_code_chip_index;
//fesetround(FE_TONEAREST);
for (int current_vector = 0; current_vector < num_out_vectors; current_vector++)
{
for (unsigned int n = 0; n < num_output_samples; n++)
{
// resample code for current tap
local_code_chip_index = round(code_phase_step_chips * (float)(n) + rem_code_phase_chips[current_vector]-0.5f);
if (local_code_chip_index < 0.0) local_code_chip_index += code_length_chips;
if (local_code_chip_index > (code_length_chips-1)) local_code_chip_index -= code_length_chips;
//std::cout<<"g["<<n<<"]="<<code_phase_step_chips*static_cast<float>(n) + rem_code_phase_chips-0.5f<<","<<local_code_chip_index<<" ";
result[current_vector][n] = local_code[local_code_chip_index];
}
}
//std::cout<<std::endl;
}
#endif /*LV_HAVE_GENERIC*/
#ifdef LV_HAVE_SSE2
#include <emmintrin.h>
static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_sse2(lv_16sc_t** result, const lv_16sc_t* local_code, float* rem_code_phase_chips ,float code_phase_step_chips, unsigned int num_output_samples, unsigned int code_length_chips, int num_out_vectors)
{
_MM_SET_ROUNDING_MODE (_MM_ROUND_NEAREST);//_MM_ROUND_NEAREST, _MM_ROUND_DOWN, _MM_ROUND_UP, _MM_ROUND_TOWARD_ZERO
unsigned int number;
const unsigned int quarterPoints = num_output_samples / 4;
lv_16sc_t** _result = result;
__attribute__((aligned(16))) int local_code_chip_index[4];
float tmp_rem_code_phase_chips;
__m128 _rem_code_phase,_code_phase_step_chips;
__m128i _code_length_chips,_code_length_chips_minus1;
__m128 _code_phase_out,_code_phase_out_with_offset;
_code_phase_step_chips = _mm_load1_ps(&code_phase_step_chips); //load float to all four float values in m128 register
__attribute__((aligned(16))) int four_times_code_length_chips_minus1[4];
four_times_code_length_chips_minus1[0] = code_length_chips - 1;
four_times_code_length_chips_minus1[1] = code_length_chips - 1;
four_times_code_length_chips_minus1[2] = code_length_chips - 1;
four_times_code_length_chips_minus1[3] = code_length_chips - 1;
__attribute__((aligned(16))) int four_times_code_length_chips[4];
four_times_code_length_chips[0] = code_length_chips;
four_times_code_length_chips[1] = code_length_chips;
four_times_code_length_chips[2] = code_length_chips;
four_times_code_length_chips[3] = code_length_chips;
_code_length_chips = _mm_loadu_si128((__m128i*)&four_times_code_length_chips); //load float to all four float values in m128 register
_code_length_chips_minus1 = _mm_loadu_si128((__m128i*)&four_times_code_length_chips_minus1); //load float to all four float values in m128 register
__m128i negative_indexes, overflow_indexes,_code_phase_out_int, _code_phase_out_int_neg,_code_phase_out_int_over;
__m128i zero=_mm_setzero_si128();
__attribute__((aligned(16))) float init_idx_float[4] = { 0.0f, 1.0f, 2.0f, 3.0f };
__m128 _4output_index = _mm_load_ps(init_idx_float);
__attribute__((aligned(16))) float init_4constant_float[4] = { 4.0f, 4.0f, 4.0f, 4.0f };
__m128 _4constant_float = _mm_load_ps(init_4constant_float);
int current_vector = 0;
int sample_idx = 0;
for(number = 0; number < quarterPoints; number++)
{
//common to all outputs
_code_phase_out = _mm_mul_ps(_code_phase_step_chips, _4output_index); //compute the code phase point with the phase step
//output vector dependant (different code phase offset)
for(current_vector = 0; current_vector < num_out_vectors; current_vector++)
{
tmp_rem_code_phase_chips = rem_code_phase_chips[current_vector] - 0.5f; // adjust offset to perform correct rounding (chip transition at 0)
_rem_code_phase = _mm_load1_ps(&tmp_rem_code_phase_chips); //load float to all four float values in m128 register
_code_phase_out_with_offset = _mm_add_ps(_code_phase_out, _rem_code_phase); //add the phase offset
_code_phase_out_int = _mm_cvtps_epi32(_code_phase_out_with_offset); //convert to integer
negative_indexes = _mm_cmplt_epi32 (_code_phase_out_int, zero); //test for negative values
_code_phase_out_int_neg = _mm_add_epi32(_code_phase_out_int, _code_length_chips); //the negative values branch
_code_phase_out_int_neg = _mm_xor_si128(_code_phase_out_int, _mm_and_si128( negative_indexes,_mm_xor_si128( _code_phase_out_int_neg, _code_phase_out_int )));
overflow_indexes = _mm_cmpgt_epi32 (_code_phase_out_int_neg, _code_length_chips_minus1); //test for overflow values
_code_phase_out_int_over = _mm_sub_epi32(_code_phase_out_int_neg, _code_length_chips); //the negative values branch
_code_phase_out_int_over = _mm_xor_si128(_code_phase_out_int_neg, _mm_and_si128( overflow_indexes, _mm_xor_si128( _code_phase_out_int_over, _code_phase_out_int_neg )));
_mm_storeu_si128((__m128i*)local_code_chip_index, _code_phase_out_int_over); // Store the results back
//todo: optimize the local code lookup table with intrinsics, if possible
_result[current_vector][sample_idx] = local_code[local_code_chip_index[0]];
_result[current_vector][sample_idx + 1] = local_code[local_code_chip_index[1]];
_result[current_vector][sample_idx + 2] = local_code[local_code_chip_index[2]];
_result[current_vector][sample_idx + 3] = local_code[local_code_chip_index[3]];
}
_4output_index = _mm_add_ps(_4output_index, _4constant_float);
sample_idx += 4;
}
for(number = quarterPoints * 4; number < num_output_samples; number++)
{
for(current_vector = 0; current_vector < num_out_vectors; current_vector++)
{
local_code_chip_index[0] = (int)(code_phase_step_chips * (float)(number) + rem_code_phase_chips[current_vector]);
if (local_code_chip_index[0] < 0.0) local_code_chip_index[0] += code_length_chips - 1;
if (local_code_chip_index[0] > (code_length_chips - 1)) local_code_chip_index[0] -= code_length_chips;
_result[current_vector][number] = local_code[local_code_chip_index[0]];
}
}
}
#endif /* LV_HAVE_SSE2 */
#endif /*INCLUDED_volk_gnsssdr_16ic_xn_resampler_16ic_xn_a_H*/

View File

@ -57,16 +57,7 @@ std::vector<volk_gnsssdr_test_case_t> init_test_list(volk_gnsssdr_test_params_t
test_params.vlen(), test_params.iter(), test_params.benchmark_mode(), test_params.kernel_regex()); test_params.vlen(), test_params.iter(), test_params.benchmark_mode(), test_params.kernel_regex());
std::vector<volk_gnsssdr_test_case_t> test_cases = boost::assign::list_of std::vector<volk_gnsssdr_test_case_t> test_cases = boost::assign::list_of
// no one uses these, so don't test them
//VOLK_PROFILE(volk_gnsssdr_16i_x5_add_quad_16i_x4, 1e-4, 2046, 10000, &results, benchmark_mode, kernel_regex);
//VOLK_PROFILE(volk_gnsssdr_16i_branch_4_state_8, 1e-4, 2046, 10000, &results, benchmark_mode, kernel_regex);
//VOLK_PROFILE(volk_gnsssdr_16i_max_star_16i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex);
//VOLK_PROFILE(volk_gnsssdr_16i_max_star_horizontal_16i, 0, 0, 204602, 10000, &results, benchmark_mode, kernel_regex);
//VOLK_PROFILE(volk_gnsssdr_16i_permute_and_scalar_add, 1e-4, 0, 2046, 10000, &results, benchmark_mode, kernel_regex);
//VOLK_PROFILE(volk_gnsssdr_16i_x4_quad_max_star_16i, 1e-4, 0, 2046, 10000, &results, benchmark_mode, kernel_regex);
// we need a puppet for this one
//(VOLK_INIT_TEST(volk_gnsssdr_32fc_s32f_x2_power_spectral_density_32f, test_params))
//(VOLK_INIT_TEST(volk_gnsssdr_32f_null_32f, test_params))
(VOLK_INIT_TEST(volk_gnsssdr_8i_accumulator_s8i, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_8i_accumulator_s8i, test_params))
(VOLK_INIT_TEST(volk_gnsssdr_8i_index_max_16u, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_8i_index_max_16u, test_params))
(VOLK_INIT_TEST(volk_gnsssdr_8i_max_s8i, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_8i_max_s8i, test_params))
@ -77,11 +68,13 @@ std::vector<volk_gnsssdr_test_case_t> init_test_list(volk_gnsssdr_test_params_t
(VOLK_INIT_TEST(volk_gnsssdr_8ic_x2_multiply_8ic, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_8ic_x2_multiply_8ic, test_params))
(VOLK_INIT_TEST(volk_gnsssdr_8u_x2_multiply_8u, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_8u_x2_multiply_8u, test_params))
(VOLK_INIT_TEST(volk_gnsssdr_64f_accumulator_64f, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_64f_accumulator_64f, test_params))
(VOLK_INIT_TEST(volk_gnsssdr_32fc_convert_8ic, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_32fc_convert_8ic, test_params_int1))
(VOLK_INIT_TEST(volk_gnsssdr_32fc_convert_16ic, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_32fc_convert_16ic, test_params))
(VOLK_INIT_TEST(volk_gnsssdr_16ic_x2_dot_prod_16ic, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_16ic_x2_dot_prod_16ic, test_params))
(VOLK_INIT_TEST(volk_gnsssdr_16ic_x2_multiply_16ic, test_params)) (VOLK_INIT_TEST(volk_gnsssdr_16ic_x2_multiply_16ic, test_params))
(VOLK_INIT_TEST(volk_gnsssdr_16ic_x2_dot_prod_16ic_xn, volk_gnsssdr_test_params_t(1e-2, test_params.scalar(), test_params.vlen(), test_params.iter(), test_params.benchmark_mode(), test_params.kernel_regex())))
//(VOLK_INIT_TEST(volk_gnsssdr_16ic_resampler_16ic, test_params)) //(VOLK_INIT_TEST(volk_gnsssdr_16ic_resampler_16ic, test_params))
//(VOLK_INIT_TEST(volk_gnsssdr_16ic_xn_resampler_16ic_xn, test_params))
; ;
return test_cases; return test_cases;

View File

@ -49,28 +49,36 @@ void random_floats (t *buf, unsigned n)
buf[i] = uniform (); buf[i] = uniform ();
} }
void load_random_data(void *data, volk_gnsssdr_type_t type, unsigned int n) { void load_random_data(void *data, volk_gnsssdr_type_t type, unsigned int n)
{
if(type.is_complex) n *= 2; if(type.is_complex) n *= 2;
if(type.is_float) { if(type.is_float)
{
if(type.size == 8) random_floats<double>((double *)data, n); if(type.size == 8) random_floats<double>((double *)data, n);
else random_floats<float>((float *)data, n); else random_floats<float>((float *)data, n);
} else { }
else
{
float int_max = float(uint64_t(2) << (type.size*8)); float int_max = float(uint64_t(2) << (type.size*8));
if(type.is_signed) int_max /= 2.0; if(type.is_signed) int_max /= 2.0;
for(unsigned int i=0; i<n; i++) { for(unsigned int i = 0; i < n; i++)
{
float scaled_rand = (((float) (rand() - (RAND_MAX/2))) / static_cast<float>((RAND_MAX/2))) * int_max; float scaled_rand = (((float) (rand() - (RAND_MAX/2))) / static_cast<float>((RAND_MAX/2))) * int_max;
//man i really don't know how to do this in a more clever way, you have to cast down at some point //man i really don't know how to do this in a more clever way, you have to cast down at some point
switch(type.size) { switch(type.size)
{
case 8: case 8:
if(type.is_signed) ((int64_t *)data)[i] = (int64_t) scaled_rand; if(type.is_signed) ((int64_t *)data)[i] = (int64_t) scaled_rand;
else ((uint64_t *)data)[i] = (uint64_t) scaled_rand; else ((uint64_t *)data)[i] = (uint64_t) scaled_rand;
break; break;
case 4: case 4:
if(type.is_signed) ((int32_t *)data)[i] = (int32_t) scaled_rand; if(type.is_signed) ((int32_t *)data)[i] = (int32_t) scaled_rand;
else ((uint32_t *)data)[i] = (uint32_t) scaled_rand; else ((uint32_t *)data)[i] = (uint32_t) scaled_rand;
break; break;
case 2: case 2:
if(type.is_signed) ((int16_t *)data)[i] = (int16_t) scaled_rand; if(type.is_signed) ((int16_t *)data)[i] = (int16_t) scaled_rand / 11; //sqrt(std::abs(scaled_rand / 2.0)); //// std::cout << "222222222222" << std::endl;}
else ((uint16_t *)data)[i] = (uint16_t) scaled_rand; else ((uint16_t *)data)[i] = (uint16_t) scaled_rand;
break; break;
case 1: case 1:
@ -172,15 +180,21 @@ static void get_signatures_from_name(std::vector<volk_gnsssdr_type_t> &inputsig,
if(side == SIDE_INPUT) inputsig.push_back(type); if(side == SIDE_INPUT) inputsig.push_back(type);
else outputsig.push_back(type); else outputsig.push_back(type);
} catch (...){ } catch (...){
if(token[0] == 'x' && (token.size() > 1) && (token[1] > '0' || token[1] < '9')) { //it's a multiplier if(token[0] == 'x' && (token.size() > 1) && (token[1] > '0' || token[1] < '9')) { std::cout << "multiplier normsl" << std::endl;//it's a multiplier
if(side == SIDE_INPUT) assert(inputsig.size() > 0); if(side == SIDE_INPUT) assert(inputsig.size() > 0);
else assert(outputsig.size() > 0); else assert(outputsig.size() > 0);
int multiplier = boost::lexical_cast<int>(token.substr(1, token.size()-1)); //will throw if invalid int multiplier = 1;
try {
multiplier = boost::lexical_cast<int>(token.substr(1, token.size()-1)); //will throw if invalid ///////////
} catch(...) {
multiplier = 1; std::cout << "multiplier 333333333" << std::endl;
}
for(int i=1; i<multiplier; i++) { for(int i=1; i<multiplier; i++) {
if(side == SIDE_INPUT) inputsig.push_back(inputsig.back()); if(side == SIDE_INPUT) inputsig.push_back(inputsig.back());
else outputsig.push_back(outputsig.back()); else outputsig.push_back(outputsig.back());
} }
} }
else if(side == SIDE_INPUT) { //it's the function name, at least it better be else if(side == SIDE_INPUT) { //it's the function name, at least it better be
side = SIDE_NAME; side = SIDE_NAME;
fn_name.append("_"); fn_name.append("_");
@ -268,6 +282,24 @@ inline void run_cast_test3_s8ic(volk_gnsssdr_fn_3arg_s8ic func, std::vector<void
} }
// new
inline void run_cast_test1_s16ic(volk_gnsssdr_fn_1arg_s16ic func, std::vector<void *> &buffs, lv_16sc_t scalar, unsigned int vlen, unsigned int iter, std::string arch)
{
while(iter--) func(buffs[0], scalar, vlen, arch.c_str());
}
inline void run_cast_test2_s16ic(volk_gnsssdr_fn_2arg_s16ic func, std::vector<void *> &buffs, lv_16sc_t scalar, unsigned int vlen, unsigned int iter, std::string arch)
{
while(iter--) func(buffs[0], buffs[1], scalar, vlen, arch.c_str());
}
inline void run_cast_test3_s16ic(volk_gnsssdr_fn_3arg_s16ic func, std::vector<void *> &buffs, lv_16sc_t scalar, unsigned int vlen, unsigned int iter, std::string arch)
{
while(iter--) func(buffs[0], buffs[1], buffs[2], scalar, vlen, arch.c_str());
}
// end new
inline void run_cast_test8(volk_gnsssdr_fn_8arg func, std::vector<void *> &buffs, unsigned int vlen, unsigned int iter, std::string arch) inline void run_cast_test8(volk_gnsssdr_fn_8arg func, std::vector<void *> &buffs, unsigned int vlen, unsigned int iter, std::string arch)
{ {
while(iter--) func(buffs[0], buffs[1], buffs[2], buffs[3], buffs[4], buffs[5], buffs[6], buffs[7], vlen, arch.c_str()); while(iter--) func(buffs[0], buffs[1], buffs[2], buffs[3], buffs[4], buffs[5], buffs[6], buffs[7], vlen, arch.c_str());
@ -439,14 +471,14 @@ bool run_volk_gnsssdr_tests(volk_gnsssdr_func_desc_t desc,
unsigned int iter, unsigned int iter,
std::vector<volk_gnsssdr_test_results_t> *results, std::vector<volk_gnsssdr_test_results_t> *results,
std::string puppet_master_name, std::string puppet_master_name,
bool benchmark_mode bool benchmark_mode)
) { {
// Initialize this entry in results vector // Initialize this entry in results vector
results->push_back(volk_gnsssdr_test_results_t()); results->push_back(volk_gnsssdr_test_results_t());
results->back().name = name; results->back().name = name;
results->back().vlen = vlen; results->back().vlen = vlen;
results->back().iter = iter; results->back().iter = iter;
std::cout << "RUN_VOLK_TESTS: " << name << "(" << vlen << "," << iter << ")" << std::endl; std::cout << "RUN_VOLK_GNSSSDR_TESTS: " << name << "(" << vlen << "," << iter << ")" << std::endl;
// vlen_twiddle will increase vlen for malloc and data generation // vlen_twiddle will increase vlen for malloc and data generation
// but kernels will still be called with the user provided vlen. // but kernels will still be called with the user provided vlen.
@ -545,9 +577,16 @@ bool run_volk_gnsssdr_tests(volk_gnsssdr_func_desc_t desc,
else if(inputsc.size() == 1 && !inputsc[0].is_float) else if(inputsc.size() == 1 && !inputsc[0].is_float)
{ {
if(inputsc[0].is_complex) if(inputsc[0].is_complex)
{
if(inputsc[0].size == 2)
{
run_cast_test1_s16ic((volk_gnsssdr_fn_1arg_s16ic)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]);
}
else
{ {
run_cast_test1_s8ic((volk_gnsssdr_fn_1arg_s8ic)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]); run_cast_test1_s8ic((volk_gnsssdr_fn_1arg_s8ic)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]);
} }
}
else else
{ {
run_cast_test1_s8i((volk_gnsssdr_fn_1arg_s8i)(manual_func), test_data[i], scalar.real(), vlen, iter, arch_list[i]); run_cast_test1_s8i((volk_gnsssdr_fn_1arg_s8i)(manual_func), test_data[i], scalar.real(), vlen, iter, arch_list[i]);
@ -557,6 +596,7 @@ bool run_volk_gnsssdr_tests(volk_gnsssdr_func_desc_t desc,
else throw "unsupported 1 arg function >1 scalars"; else throw "unsupported 1 arg function >1 scalars";
break; break;
case 2: case 2:
//std::cout << "case 2 " << inputsc.size() << std::endl;
if(inputsc.size() == 0) if(inputsc.size() == 0)
{ {
run_cast_test2((volk_gnsssdr_fn_2arg)(manual_func), test_data[i], vlen, iter, arch_list[i]); run_cast_test2((volk_gnsssdr_fn_2arg)(manual_func), test_data[i], vlen, iter, arch_list[i]);
@ -576,9 +616,16 @@ bool run_volk_gnsssdr_tests(volk_gnsssdr_func_desc_t desc,
else if(inputsc.size() == 1 && !inputsc[0].is_float) else if(inputsc.size() == 1 && !inputsc[0].is_float)
{ {
if(inputsc[0].is_complex) if(inputsc[0].is_complex)
{
if(inputsc[0].size == 2)
{
run_cast_test2_s16ic((volk_gnsssdr_fn_2arg_s16ic)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]);
}
else
{ {
run_cast_test2_s8ic((volk_gnsssdr_fn_2arg_s8ic)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]); run_cast_test2_s8ic((volk_gnsssdr_fn_2arg_s8ic)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]);
} }
}
else else
{ {
run_cast_test2_s8i((volk_gnsssdr_fn_2arg_s8i)(manual_func), test_data[i], scalar.real(), vlen, iter, arch_list[i]); run_cast_test2_s8i((volk_gnsssdr_fn_2arg_s8i)(manual_func), test_data[i], scalar.real(), vlen, iter, arch_list[i]);
@ -590,6 +637,7 @@ bool run_volk_gnsssdr_tests(volk_gnsssdr_func_desc_t desc,
case 3: case 3:
if(inputsc.size() == 0) if(inputsc.size() == 0)
{ {
// multipliers are here!
run_cast_test3((volk_gnsssdr_fn_3arg)(manual_func), test_data[i], vlen, iter, arch_list[i]); run_cast_test3((volk_gnsssdr_fn_3arg)(manual_func), test_data[i], vlen, iter, arch_list[i]);
} }
else if(inputsc.size() == 1 && inputsc[0].is_float) else if(inputsc.size() == 1 && inputsc[0].is_float)
@ -607,9 +655,18 @@ bool run_volk_gnsssdr_tests(volk_gnsssdr_func_desc_t desc,
else if(inputsc.size() == 1 && !inputsc[0].is_float) else if(inputsc.size() == 1 && !inputsc[0].is_float)
{ {
if(inputsc[0].is_complex) if(inputsc[0].is_complex)
{
{
if(inputsc[0].size == 4)
{
run_cast_test3_s16ic((volk_gnsssdr_fn_3arg_s16ic)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]);
}
else
{ {
run_cast_test3_s8ic((volk_gnsssdr_fn_3arg_s8ic)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]); run_cast_test3_s8ic((volk_gnsssdr_fn_3arg_s8ic)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]);
} }
}
}
else else
{ {
run_cast_test3_s8i((volk_gnsssdr_fn_3arg_s8i)(manual_func), test_data[i], scalar.real(), vlen, iter, arch_list[i]); run_cast_test3_s8i((volk_gnsssdr_fn_3arg_s8i)(manual_func), test_data[i], scalar.real(), vlen, iter, arch_list[i]);
@ -698,16 +755,8 @@ bool run_volk_gnsssdr_tests(volk_gnsssdr_func_desc_t desc,
} }
break; break;
case 4: case 4:
if(both_sigs[j].is_signed) if(both_sigs[j].is_complex)
{ {
fail = icompare((int32_t *) test_data[generic_offset][j], (int32_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i);
}
else
{
fail = icompare((uint32_t *) test_data[generic_offset][j], (uint32_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i);
}
break;
case 2:
if(both_sigs[j].is_signed) if(both_sigs[j].is_signed)
{ {
fail = icompare((int16_t *) test_data[generic_offset][j], (int16_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i); fail = icompare((int16_t *) test_data[generic_offset][j], (int16_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i);
@ -716,11 +765,47 @@ bool run_volk_gnsssdr_tests(volk_gnsssdr_func_desc_t desc,
{ {
fail = icompare((uint16_t *) test_data[generic_offset][j], (uint16_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i); fail = icompare((uint16_t *) test_data[generic_offset][j], (uint16_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i);
} }
}
else
{
if(both_sigs[j].is_signed)
{
fail = icompare((int32_t *) test_data[generic_offset][j], (int32_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i);
}
else
{
fail = icompare((uint32_t *) test_data[generic_offset][j], (uint32_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i);
}
}
break;
case 2:
if(both_sigs[j].is_complex)
{
if(both_sigs[j].is_signed)
{
fail = icompare((int8_t *) test_data[generic_offset][j], (int8_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i);
}
else
{
fail = icompare((uint16_t *) test_data[generic_offset][j], (uint16_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i);
}
}
else
{
if(both_sigs[j].is_signed)
{
fail = icompare((int16_t *) test_data[generic_offset][j], (int16_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i); //
}
else
{
fail = icompare((uint16_t *) test_data[generic_offset][j], (uint16_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i);
}
}
break; break;
case 1: case 1:
if(both_sigs[j].is_signed) if(both_sigs[j].is_signed)
{ {
fail = icompare((int8_t *) test_data[generic_offset][j], (int8_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol_i); fail = icompare((int8_t *) test_data[generic_offset][j], (int8_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), 3); // check volk_gnsssdr_32fc_convert_8ic !
} }
else else
{ {

View File

@ -173,9 +173,9 @@ typedef void (*volk_gnsssdr_fn_3arg_s8ic)(void *, void *, void *, lv_8sc_t, unsi
//typedef void (*volk_gnsssdr_fn_1arg_s16i)(void *, int16_t, unsigned int, const char*); //one input vector, one scalar int16_t input //typedef void (*volk_gnsssdr_fn_1arg_s16i)(void *, int16_t, unsigned int, const char*); //one input vector, one scalar int16_t input
//typedef void (*volk_gnsssdr_fn_2arg_s16i)(void *, void *, int16_t, unsigned int, const char*); //typedef void (*volk_gnsssdr_fn_2arg_s16i)(void *, void *, int16_t, unsigned int, const char*);
//typedef void (*volk_gnsssdr_fn_3arg_s16i)(void *, void *, void *, int16_t, unsigned int, const char*); //typedef void (*volk_gnsssdr_fn_3arg_s16i)(void *, void *, void *, int16_t, unsigned int, const char*);
//typedef void (*volk_gnsssdr_fn_1arg_s16ic)(void *, lv_16sc_t, unsigned int, const char*); //one input vector, one scalar lv_16sc_t vector input typedef void (*volk_gnsssdr_fn_1arg_s16ic)(void *, lv_16sc_t, unsigned int, const char*); //one input vector, one scalar lv_16sc_t vector input
//typedef void (*volk_gnsssdr_fn_2arg_s16ic)(void *, void *, lv_16sc_t, unsigned int, const char*); typedef void (*volk_gnsssdr_fn_2arg_s16ic)(void *, void *, lv_16sc_t, unsigned int, const char*);
//typedef void (*volk_gnsssdr_fn_3arg_s16ic)(void *, void *, void *, lv_16sc_t, unsigned int, const char*); typedef void (*volk_gnsssdr_fn_3arg_s16ic)(void *, void *, void *, lv_16sc_t, unsigned int, const char*);
typedef void (*volk_gnsssdr_fn_6arg_s16ic)(void *, void *, void *, void *, void *, void *, lv_16sc_t, unsigned int, const char*); typedef void (*volk_gnsssdr_fn_6arg_s16ic)(void *, void *, void *, void *, void *, void *, lv_16sc_t, unsigned int, const char*);
typedef void (*volk_gnsssdr_fn_8arg)(void *, void *, void *, void *, void *, void *, void *, void *, unsigned int, const char*); typedef void (*volk_gnsssdr_fn_8arg)(void *, void *, void *, void *, void *, void *, void *, void *, unsigned int, const char*);

View File

@ -31,18 +31,13 @@
* *
* ------------------------------------------------------------------------- * -------------------------------------------------------------------------
*/ */
#include "cpu_multicorrelator_16sc.h" #include "cpu_multicorrelator_16sc.h"
#include <cmath> #include <cmath>
#include <iostream> #include <iostream>
#include <gnuradio/fxpt.h> // fixed point sine and cosine #include <gnuradio/fxpt.h> // fixed point sine and cosine
#include "volk_gnsssdr/volk_gnsssdr.h" #include "volk_gnsssdr/volk_gnsssdr.h"
#define LV_HAVE_GENERIC
#define LV_HAVE_SSE2
#include "volk_gnsssdr_16ic_xn_resampler_16ic_xn.h"
#include "volk_gnsssdr_16ic_xn_dot_prod_16ic_xn.h"
bool cpu_multicorrelator_16sc::init( bool cpu_multicorrelator_16sc::init(
int max_signal_length_samples, int max_signal_length_samples,
@ -99,7 +94,7 @@ void cpu_multicorrelator_16sc::update_local_code(int correlator_length_samples,f
tmp_code_phases_chips[n] = d_shifts_chips[n] - rem_code_phase_chips; tmp_code_phases_chips[n] = d_shifts_chips[n] - rem_code_phase_chips;
} }
volk_gnsssdr_16ic_xn_resampler_16ic_xn_sse2(d_local_codes_resampled, volk_gnsssdr_16ic_xn_resampler_16ic_xn(d_local_codes_resampled,
d_local_code_in, d_local_code_in,
tmp_code_phases_chips, tmp_code_phases_chips,
code_phase_step_chips, code_phase_step_chips,
@ -153,7 +148,7 @@ bool cpu_multicorrelator_16sc::Carrier_wipeoff_multicorrelator_resampler(
//std::cout<<"d_sig_doppler_wiped 16sc="<<d_sig_doppler_wiped[23]<<std::endl; //std::cout<<"d_sig_doppler_wiped 16sc="<<d_sig_doppler_wiped[23]<<std::endl;
update_local_code(signal_length_samples, rem_code_phase_chips, code_phase_step_chips); update_local_code(signal_length_samples, rem_code_phase_chips, code_phase_step_chips);
volk_gnsssdr_16ic_xn_dot_prod_16ic_xn_a_sse2(d_corr_out, d_sig_doppler_wiped, (const lv_16sc_t**)d_local_codes_resampled, signal_length_samples, d_n_correlators); volk_gnsssdr_16ic_x2_dot_prod_16ic_xn(d_corr_out, d_sig_doppler_wiped, (const lv_16sc_t**)d_local_codes_resampled, signal_length_samples, d_n_correlators);
//for (int current_correlator_tap = 0; current_correlator_tap < d_n_correlators; current_correlator_tap++) //for (int current_correlator_tap = 0; current_correlator_tap < d_n_correlators; current_correlator_tap++)
// { // {

View File

@ -1,172 +0,0 @@
/*!
* \file volk_gnsssdr_16ic_xn_resampler_16ic_xn.h
* \brief Volk protokernel: resample a 16 bits complex vector
* \authors <ul>
* <li> Javier Arribas, 2015. jarribas(at)cttc.es
* </ul>
*
* Volk protokernel that multiplies two 16 bits vectors (8 bits the real part
* and 8 bits the imaginary part) and accumulates them
*
* -------------------------------------------------------------------------
*
* Copyright (C) 2010-2015 (see AUTHORS file for a list of contributors)
*
* GNSS-SDR is a software defined Global Navigation
* Satellite Systems receiver
*
* This file is part of GNSS-SDR.
*
* GNSS-SDR is free software: you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* GNSS-SDR is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with GNSS-SDR. If not, see <http://www.gnu.org/licenses/>.
*
* -------------------------------------------------------------------------
*/
#ifndef INCLUDED_volk_gnsssdr_16ic_xn_resampler_16ic_xn_a_H
#define INCLUDED_volk_gnsssdr_16ic_xn_resampler_16ic_xn_a_H
#include <volk_gnsssdr/volk_gnsssdr_common.h>
#include <volk_gnsssdr/volk_gnsssdr_complex.h>
#include <cmath>
//#pragma STDC FENV_ACCESS ON
#ifdef LV_HAVE_GENERIC
//int round_int( float r ) {
// return (r > 0.0) ? (r + 0.5) : (r - 0.5);
//}
/*!
\brief Multiplies the two input complex vectors, point-by-point, storing the result in the third vector
\param cVector The vector where the result will be stored
\param aVector One of the vectors to be multiplied
\param bVector One of the vectors to be multiplied
\param num_points The number of complex values in aVector and bVector to be multiplied together, accumulated and stored into cVector
*/
static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_generic(lv_16sc_t** result, const lv_16sc_t* local_code, float* rem_code_phase_chips ,float code_phase_step_chips, unsigned int num_output_samples, unsigned int code_length_chips, int num_out_vectors)
{
int local_code_chip_index;
//fesetround(FE_TONEAREST);
for (int current_vector = 0; current_vector < num_out_vectors; current_vector++)
{
for (unsigned int n = 0; n < num_output_samples; n++)
{
// resample code for current tap
local_code_chip_index = round(code_phase_step_chips*static_cast<float>(n) + rem_code_phase_chips[current_vector]-0.5f);
if (local_code_chip_index < 0.0) local_code_chip_index += code_length_chips;
if (local_code_chip_index > (code_length_chips-1)) local_code_chip_index -= code_length_chips;
//std::cout<<"g["<<n<<"]="<<code_phase_step_chips*static_cast<float>(n) + rem_code_phase_chips-0.5f<<","<<local_code_chip_index<<" ";
result[current_vector][n] = local_code[local_code_chip_index];
}
}
//std::cout<<std::endl;
}
#endif /*LV_HAVE_GENERIC*/
#ifdef LV_HAVE_SSE2
#include <emmintrin.h>
static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_sse2(lv_16sc_t** result, const lv_16sc_t* local_code, float* rem_code_phase_chips ,float code_phase_step_chips, unsigned int num_output_samples, unsigned int code_length_chips, int num_out_vectors)
{
_MM_SET_ROUNDING_MODE (_MM_ROUND_NEAREST);//_MM_ROUND_NEAREST, _MM_ROUND_DOWN, _MM_ROUND_UP, _MM_ROUND_TOWARD_ZERO
unsigned int number;
const unsigned int quarterPoints = num_output_samples / 4;
lv_16sc_t** _result = result;
__attribute__((aligned(16))) int local_code_chip_index[4];
float tmp_rem_code_phase_chips;
__m128 _rem_code_phase,_code_phase_step_chips;
__m128i _code_length_chips,_code_length_chips_minus1;
__m128 _code_phase_out,_code_phase_out_with_offset;
_code_phase_step_chips = _mm_load1_ps(&code_phase_step_chips); //load float to all four float values in m128 register
__attribute__((aligned(16))) int four_times_code_length_chips_minus1[4];
four_times_code_length_chips_minus1[0]=code_length_chips-1;
four_times_code_length_chips_minus1[1]=code_length_chips-1;
four_times_code_length_chips_minus1[2]=code_length_chips-1;
four_times_code_length_chips_minus1[3]=code_length_chips-1;
__attribute__((aligned(16))) int four_times_code_length_chips[4];
four_times_code_length_chips[0]=code_length_chips;
four_times_code_length_chips[1]=code_length_chips;
four_times_code_length_chips[2]=code_length_chips;
four_times_code_length_chips[3]=code_length_chips;
_code_length_chips = _mm_loadu_si128((__m128i*)&four_times_code_length_chips); //load float to all four float values in m128 register
_code_length_chips_minus1 = _mm_loadu_si128((__m128i*)&four_times_code_length_chips_minus1); //load float to all four float values in m128 register
__m128i negative_indexes, overflow_indexes,_code_phase_out_int, _code_phase_out_int_neg,_code_phase_out_int_over;
__m128i zero=_mm_setzero_si128();
__attribute__((aligned(16))) float init_idx_float[4] = { 0.0f, 1.0f, 2.0f, 3.0f };
__m128 _4output_index=_mm_load_ps(init_idx_float);
__attribute__((aligned(16))) float init_4constant_float[4] = { 4.0f, 4.0f, 4.0f, 4.0f };
__m128 _4constant_float=_mm_load_ps(init_4constant_float);
int current_vector=0;
int sample_idx=0;
for(number=0;number < quarterPoints; number++){
//common to all outputs
_code_phase_out = _mm_mul_ps(_code_phase_step_chips, _4output_index); //compute the code phase point with the phase step
//output vector dependant (different code phase offset)
for(current_vector=0;current_vector<num_out_vectors;current_vector++)
{
tmp_rem_code_phase_chips=rem_code_phase_chips[current_vector]-0.5f; // adjust offset to perform correct rounding (chip transition at 0)
_rem_code_phase = _mm_load1_ps(&tmp_rem_code_phase_chips); //load float to all four float values in m128 register
_code_phase_out_with_offset = _mm_add_ps(_code_phase_out,_rem_code_phase); //add the phase offset
_code_phase_out_int=_mm_cvtps_epi32(_code_phase_out_with_offset); //convert to integer
negative_indexes=_mm_cmplt_epi32 (_code_phase_out_int, zero); //test for negative values
_code_phase_out_int_neg=_mm_add_epi32(_code_phase_out_int,_code_length_chips); //the negative values branch
_code_phase_out_int_neg=_mm_xor_si128(_code_phase_out_int,_mm_and_si128( negative_indexes,_mm_xor_si128( _code_phase_out_int_neg, _code_phase_out_int )));
overflow_indexes=_mm_cmpgt_epi32 (_code_phase_out_int_neg, _code_length_chips_minus1); //test for overflow values
_code_phase_out_int_over=_mm_sub_epi32(_code_phase_out_int_neg,_code_length_chips); //the negative values branch
_code_phase_out_int_over=_mm_xor_si128(_code_phase_out_int_neg,_mm_and_si128( overflow_indexes,_mm_xor_si128( _code_phase_out_int_over, _code_phase_out_int_neg )));
_mm_storeu_si128((__m128i*)local_code_chip_index,_code_phase_out_int_over); // Store the results back
//todo: optimize the local code lookup table with intrinsics, if possible
_result[current_vector][sample_idx]=local_code[local_code_chip_index[0]];
_result[current_vector][sample_idx+1]=local_code[local_code_chip_index[1]];
_result[current_vector][sample_idx+2]=local_code[local_code_chip_index[2]];
_result[current_vector][sample_idx+3]=local_code[local_code_chip_index[3]];
}
_4output_index = _mm_add_ps(_4output_index,_4constant_float);
sample_idx+=4;
}
for(number = quarterPoints * 4;number < num_output_samples; number++){
for(current_vector=0;current_vector<num_out_vectors;current_vector++)
{
local_code_chip_index[0]=static_cast<int>(code_phase_step_chips*static_cast<float>(number) + rem_code_phase_chips[current_vector]);
if (local_code_chip_index[0] < 0.0) local_code_chip_index[0] += code_length_chips-1;
if (local_code_chip_index[0] > (code_length_chips-1)) local_code_chip_index[0] -= code_length_chips;
_result[current_vector][number]=local_code[local_code_chip_index[0]];
}
}
}
#endif /* LV_HAVE_SSE2 */
#endif /*INCLUDED_volk_gnsssdr_16ic_xn_resampler_16ic_xn_a_H*/