diff options
| author | Nathan West <nathan.west@okstate.edu> | 2014-10-21 17:32:30 -0500 |
|---|---|---|
| committer | Nathan West <nathan.west@okstate.edu> | 2014-10-21 17:32:30 -0500 |
| commit | 398f96a94e2b0b0df8870ad4ccc956bce4c9633b (patch) | |
| tree | 12f960e64c009ec56a2f6c96c855fdeec6715a82 /volk | |
| parent | 150bb12cffd1e2a07028aabd2cc395ea3502c17a (diff) | |
volk: add neon version of complex<float> dot product
Diffstat (limited to 'volk')
| -rw-r--r-- | volk/kernels/volk/volk_32fc_x2_conjugate_dot_prod_32fc.h | 49 | ||||
| -rw-r--r-- | volk/kernels/volk/volk_32fc_x2_dot_prod_32fc.h | 2 |
2 files changed, 49 insertions, 2 deletions
diff --git a/volk/kernels/volk/volk_32fc_x2_conjugate_dot_prod_32fc.h b/volk/kernels/volk/volk_32fc_x2_conjugate_dot_prod_32fc.h index 750f508b7e..8964434bef 100644 --- a/volk/kernels/volk/volk_32fc_x2_conjugate_dot_prod_32fc.h +++ b/volk/kernels/volk/volk_32fc_x2_conjugate_dot_prod_32fc.h @@ -164,10 +164,57 @@ static inline void volk_32fc_x2_conjugate_dot_prod_32fc_u_sse3(lv_32fc_t* result #endif /*LV_HAVE_SSE3*/ +#ifdef LV_HAVE_NEON +#include <arm_neon.h> +static inline void volk_32fc_x2_conjugate_dot_prod_32fc_neon(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { + + unsigned int quarter_points = num_points / 4; + unsigned int number; + + lv_32fc_t* a_ptr = (lv_32fc_t*) taps; + lv_32fc_t* b_ptr = (lv_32fc_t*) input; + // for 2-lane vectors, 1st lane holds the real part, + // 2nd lane holds the imaginary part + float32x4x2_t a_val, b_val, accumulator; + float32x4x2_t tmp_imag; + accumulator.val[0] = vdupq_n_f32(0); + accumulator.val[1] = vdupq_n_f32(0); + + for(number = 0; number < quarter_points; ++number) { + a_val = vld2q_f32((float*)a_ptr); // a0r|a1r|a2r|a3r || a0i|a1i|a2i|a3i + b_val = vld2q_f32((float*)b_ptr); // b0r|b1r|b2r|b3r || b0i|b1i|b2i|b3i + __builtin_prefetch(a_ptr+8); + __builtin_prefetch(b_ptr+8); + + // do the first multiply + tmp_imag.val[1] = vmulq_f32(a_val.val[1], b_val.val[0]); + tmp_imag.val[0] = vmulq_f32(a_val.val[0], b_val.val[0]); + + // use multiply accumulate/subtract to get result + tmp_imag.val[1] = vmlsq_f32(tmp_imag.val[1], a_val.val[0], b_val.val[1]); + tmp_imag.val[0] = vmlaq_f32(tmp_imag.val[0], a_val.val[1], b_val.val[1]); + + accumulator.val[0] = vaddq_f32(accumulator.val[0], tmp_imag.val[0]); + accumulator.val[1] = vaddq_f32(accumulator.val[1], tmp_imag.val[1]); + + // increment pointers + a_ptr += 4; + b_ptr += 4; + } + lv_32fc_t accum_result[4]; + vst2q_f32((float*)accum_result, accumulator); + *result = accum_result[0] + accum_result[1] + accum_result[2] + accum_result[3]; -#endif /*INCLUDED_volk_32fc_x2_conjugate_dot_prod_32fc_u_H*/ + // tail case + for(number = quarter_points*4; number < num_points; ++number) { + *result += (*a_ptr++) * lv_conj(*b_ptr++); + } + *result = lv_conj(*result); +} +#endif /*LV_HAVE_NEON*/ +#endif /*INCLUDED_volk_32fc_x2_conjugate_dot_prod_32fc_u_H*/ #ifndef INCLUDED_volk_32fc_x2_conjugate_dot_prod_32fc_a_H #define INCLUDED_volk_32fc_x2_conjugate_dot_prod_32fc_a_H diff --git a/volk/kernels/volk/volk_32fc_x2_dot_prod_32fc.h b/volk/kernels/volk/volk_32fc_x2_dot_prod_32fc.h index d0a09f989f..c65d0984c5 100644 --- a/volk/kernels/volk/volk_32fc_x2_dot_prod_32fc.h +++ b/volk/kernels/volk/volk_32fc_x2_dot_prod_32fc.h @@ -896,7 +896,7 @@ static inline void volk_32fc_x2_dot_prod_32fc_neon(lv_32fc_t* result, const lv_3 #endif /*LV_HAVE_NEON*/ #ifdef LV_HAVE_NEON - +#include <arm_neon.h> static inline void volk_32fc_x2_dot_prod_32fc_neon_opttests(lv_32fc_t* result, const lv_32fc_t* input, const lv_32fc_t* taps, unsigned int num_points) { unsigned int quarter_points = num_points / 4; |