mirrors/gnss-sdr
1
0
Fork 0
mirror of https://github.com/gnss-sdr/gnss-sdr synced 2024-06-26 15:03:14 +00:00
Code Issues Releases Wiki Activity

Exploiting multiply-accumulate in NEON

Browse Source
This commit is contained in:
Carles Fernandez 2016-03-12 13:32:10 +01:00
parent a93a01e9b1
commit c236c3ab67
1 changed files with 51 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

51
src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic.h
View File

@ -294,4 +294,55 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon(lv_16sc_t* out, const
#endif /* LV_HAVE_NEON */
#ifdef LV_HAVE_NEON
#include <arm_neon.h>
static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_fma(lv_16sc_t* out, const lv_16sc_t* in_a, const lv_16sc_t* in_b, unsigned int num_points)
{
unsigned int quarter_points = num_points / 4;
unsigned int number;
lv_16sc_t* a_ptr = (lv_16sc_t*) in_a;
lv_16sc_t* b_ptr = (lv_16sc_t*) in_b;
// for 2-lane vectors, 1st lane holds the real part,
// 2nd lane holds the imaginary part
int16x4x2_t a_val, b_val, accumulator;
int16x4x2_t tmp_imag;
__VOLK_ATTR_ALIGNED(16) lv_16sc_t accum_result[4];
accumulator.val[0] = vdup_n_s16(0);
accumulator.val[1] = vdup_n_s16(0);
for(number = 0; number < quarter_points; ++number)
{
a_val = vld2_s16((int16_t*)a_ptr); // a0r|a1r|a2r|a3r || a0i|a1i|a2i|a3i
b_val = vld2_s16((int16_t*)b_ptr); // b0r|b1r|b2r|b3r || b0i|b1i|b2i|b3i
__builtin_prefetch(a_ptr + 8);
__builtin_prefetch(b_ptr + 8);
tmp_imag.val[0] = vmul_s16(a_val.val[0], b_val.val[0]);
tmp_imag.val[1] = vmul_s16(a_val.val[1], b_val.val[0]);
// use multiply accumulate/subtract to get result
tmp_imag.val[0] = vmls_s16(tmp_imag.val[0], a_val.val[1], b_val.val[1]);
tmp_imag.val[1] = vmla_s16(tmp_imag.val[1], a_val.val[0], b_val.val[1]);
accumulator.val[0] = vadd_s16(accumulator.val[0], tmp_imag.val[0]);
accumulator.val[1] = vadd_s16(accumulator.val[1], tmp_imag.val[1]);
a_ptr += 4;
b_ptr += 4;
}
vst2_s16((int16_t*)accum_result, accumulator);
*out = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3];
// tail case
for(number = quarter_points * 4; number < num_points; ++number)
{
*out += (*a_ptr++) * (*b_ptr++);
}
}
#endif /* LV_HAVE_NEON */
#endif /*INCLUDED_volk_gnsssdr_16ic_x2_dot_prod_16ic_H*/