mirrors/gnss-sdr
1
0
Fork 0
mirror of https://github.com/gnss-sdr/gnss-sdr synced 2024-12-14 04:00:34 +00:00
Code Issues Releases Wiki Activity

add neon implementation

Browse Source
This commit is contained in:
Carles Fernandez 2016-01-24 13:02:02 +01:00
parent 812a4df93f
commit 2d21706041
2 changed files with 123 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_resamplerxnpuppet_16ic.h
View File

@ -119,7 +119,34 @@ static inline void volk_gnsssdr_16ic_resamplerxnpuppet_16ic_u_sse2(lv_16sc_t* re
volk_gnsssdr_free(result_aux); volk_gnsssdr_free(result_aux);
} }
#endif
#ifdef LV_HAVE_NEON
static inline void volk_gnsssdr_16ic_resamplerxnpuppet_16ic_neon(lv_16sc_t* result, const lv_16sc_t* local_code, unsigned int num_points)
{
float code_phase_step_chips = 0.1;
int code_length_chips = 1023;
int num_out_vectors = 3;
float * rem_code_phase_chips = (float*)volk_gnsssdr_malloc(sizeof(float) * num_out_vectors, volk_gnsssdr_get_alignment());
lv_16sc_t** result_aux = (lv_16sc_t**)volk_gnsssdr_malloc(sizeof(lv_16sc_t*) * num_out_vectors, volk_gnsssdr_get_alignment());
for(unsigned int n = 0; n < num_out_vectors; n++)
{
rem_code_phase_chips[n] = -0.234;
result_aux[n] = (lv_16sc_t*)volk_gnsssdr_malloc(sizeof(lv_16sc_t) * num_points, volk_gnsssdr_get_alignment());
}
volk_gnsssdr_16ic_xn_resampler_16ic_xn_neon(result_aux, local_code, rem_code_phase_chips, code_phase_step_chips, code_length_chips, num_out_vectors, num_points);
memcpy(result, result_aux[0], sizeof(lv_16sc_t) * num_points);
volk_gnsssdr_free(rem_code_phase_chips);
for(unsigned int n = 0; n < num_out_vectors; n++)
{
volk_gnsssdr_free(result_aux[n]);
}
volk_gnsssdr_free(result_aux);
}
#endif #endif
#endif // INCLUDED_volk_gnsssdr_16ic_resamplerpuppet_16ic_H #endif // INCLUDED_volk_gnsssdr_16ic_resamplerpuppet_16ic_H

96
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_xn_resampler_16ic_xn.h
View File

@ -262,4 +262,100 @@ static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_u_sse2(lv_16sc_t** res
#endif /* LV_HAVE_SSE2 */ #endif /* LV_HAVE_SSE2 */
#ifdef LV_HAVE_NEON
#include <arm_neon.h>
static inline void volk_gnsssdr_16ic_xn_resampler_16ic_xn_neon(lv_16sc_t** result, const lv_16sc_t* local_code, float* rem_code_phase_chips ,float code_phase_step_chips, unsigned int code_length_chips, int num_out_vectors, unsigned int num_output_samples)
{
unsigned int number;
const unsigned int quarterPoints = num_output_samples / 4;
float32x4_t half = vdupq_n_f32(0.5f);
lv_16sc_t** _result = result;
__attribute__((aligned(16))) int local_code_chip_index[4];
float tmp_rem_code_phase_chips;
float32x4_t _rem_code_phase, _code_phase_step_chips;
int32x4_t _code_length_chips, _code_length_chips_minus1;
float32x4_t _code_phase_out, _code_phase_out_with_offset;
float32x4_t sign, PlusHalf, Round;
_code_phase_step_chips = vld1q_dup_f32(&code_phase_step_chips); //load float to all four float values in float32x4_t 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 = vld1q_s32((int32_t*)&four_times_code_length_chips); //load float to all four float values in float32x4_t register
_code_length_chips_minus1 = vld1q_s32((int32_t*)&four_times_code_length_chips_minus1); //load float to all four float values in float32x4_t register
int32x4_t _code_phase_out_int, _code_phase_out_int_neg, _code_phase_out_int_over;
uint32x4_t negative_indexes, overflow_indexes;
int32x4_t zero = vmovq_n_s32(0);
__attribute__((aligned(16))) float init_idx_float[4] = { 0.0f, 1.0f, 2.0f, 3.0f };
float32x4_t _4output_index = vld1q_f32(init_idx_float);
__attribute__((aligned(16))) float init_4constant_float[4] = { 4.0f, 4.0f, 4.0f, 4.0f };
float32x4_t _4constant_float = vld1q_f32(init_4constant_float);
int current_vector = 0;
int sample_idx = 0;
for(number = 0; number < quarterPoints; number++)
{
//common to all outputs
_code_phase_out = vmulq_f32(_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 = vld1q_dup_f32(&tmp_rem_code_phase_chips); //load float to all four float values in float32x4_t register
_code_phase_out_with_offset = vaddq_f32(_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
sign = vcvtq_f32_u32((vshrq_n_u32(vreinterpretq_u32_f32(_code_phase_out_with_offset), 31)));
PlusHalf = vaddq_f32(_code_phase_out_with_offset, half);
Round = vsubq_f32(PlusHalf, sign);
_code_phase_out_int = vcvtq_s32_f32(Round);
negative_indexes = vcltq_s32(_code_phase_out_int, zero); //test for negative values
_code_phase_out_int_neg = vaddq_s32(_code_phase_out_int, _code_length_chips); //the negative values branch
_code_phase_out_int_neg = veorq_s32(_code_phase_out_int, vandq_s32( (int32x4_t)negative_indexes, veorq_s32( _code_phase_out_int_neg, _code_phase_out_int )));
overflow_indexes = vcgtq_s32(_code_phase_out_int_neg, _code_length_chips_minus1); //test for overflow values
_code_phase_out_int_over = vsubq_s32(_code_phase_out_int_neg, _code_length_chips); //the negative values branch
_code_phase_out_int_over = veorq_s32(_code_phase_out_int_neg, vandq_s32( (int32x4_t)overflow_indexes, veorq_s32( _code_phase_out_int_over, _code_phase_out_int_neg )));
vst1q_s32((int32_t*)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 = vaddq_f32(_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_NEON */
#endif /*INCLUDED_volk_gnsssdr_16ic_xn_resampler_16ic_xn_H*/ #endif /*INCLUDED_volk_gnsssdr_16ic_xn_resampler_16ic_xn_H*/