[hamradio-commits] [gnss-sdr] 115/149: add neon implementation

[Carles Fernandez]

This is an automated email from the git hooks/post-receive script.

carles_fernandez-guest pushed a commit to branch next
in repository gnss-sdr.

commit 812a4df93f116751b785379c9c5480e0e3b6944c
Author: Carles Fernandez <carles.fernandez at gmail.com>
Date:   Sun Jan 24 12:01:40 2016 +0100

    add neon implementation
---
 .../volk_gnsssdr_16ic_resampler_16ic.h             | 84 ++++++++++++++++++++++
 .../volk_gnsssdr_16ic_resamplerpuppet_16ic.h       | 12 ++++
 2 files changed, 96 insertions(+)

diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_resampler_16ic.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_resampler_16ic.h
index 0f3a816..5afd62f 100644
--- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_resampler_16ic.h
+++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_resampler_16ic.h
@@ -233,4 +233,88 @@ static inline void volk_gnsssdr_16ic_resampler_16ic_u_sse2(lv_16sc_t* result, co
 
 #endif /* LV_HAVE_SSE2 */
 
+#ifdef LV_HAVE_NEON
+#include <arm_neon.h>
+static inline void volk_gnsssdr_16ic_resampler_16ic_neon(lv_16sc_t* result, const lv_16sc_t* local_code, float rem_code_phase_chips, float code_phase_step_chips, int code_length_chips, unsigned int num_output_samples)//, int* scratch_buffer, float* scratch_buffer_float)
+{
+    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];
+    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;
+    rem_code_phase_chips = rem_code_phase_chips - 0.5f;
+    float32x4_t sign, PlusHalf, Round;
+
+    _rem_code_phase =  vld1q_dup_f32(&rem_code_phase_chips); //load float to all four float values in m128 register
+    _code_phase_step_chips = vld1q_dup_f32(&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 = vld1q_s32((int32_t*)&four_times_code_length_chips); //load float to all four float values in m128 register
+    _code_length_chips_minus1 = vld1q_s32((int32_t*)&four_times_code_length_chips_minus1); //load float to all four float values in m128 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);
+
+
+    for(number = 0; number < quarterPoints; number++)
+        {
+            _code_phase_out = vmulq_f32(_code_phase_step_chips, _4output_index); //compute the code phase point with the phase step
+            _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  int32x4_t = f(float32x4_t)
+            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++ = local_code[local_code_chip_index[0]];
+            *_result++ = local_code[local_code_chip_index[1]];
+            *_result++ = local_code[local_code_chip_index[2]];
+            *_result++ = local_code[local_code_chip_index[3]];
+
+            _4output_index = vaddq_f32(_4output_index, _4constant_float);
+        }
+
+    for(number = quarterPoints * 4; number < num_output_samples; number++)
+        {
+            local_code_chip_index[0] = (int)(code_phase_step_chips * (float)number + rem_code_phase_chips + 0.5f);
+            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++ = local_code[local_code_chip_index[0]];
+        }
+}
+
+#endif /* LV_HAVE_NEON */
+
 #endif /*INCLUDED_volk_gnsssdr_16ic_resampler_16ic_H*/
diff --git a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_resamplerpuppet_16ic.h b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_resamplerpuppet_16ic.h
index 03d45e4..9e74960 100644
--- a/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_resamplerpuppet_16ic.h
+++ b/src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_resamplerpuppet_16ic.h
@@ -72,4 +72,16 @@ static inline void volk_gnsssdr_16ic_resamplerpuppet_16ic_u_sse2(lv_16sc_t* resu
 
 #endif /* LV_HAVE_SSE2 */
 
+#ifdef LV_HAVE_NEON
+
+static inline void volk_gnsssdr_16ic_resamplerpuppet_16ic_neon(lv_16sc_t* result, const lv_16sc_t* local_code, unsigned int num_points)
+{
+    float rem_code_phase_chips = -0.123;
+    float code_phase_step_chips = 0.1;
+    int code_length_chips = 1023;
+    volk_gnsssdr_16ic_resampler_16ic_neon(result, local_code, rem_code_phase_chips, code_phase_step_chips, code_length_chips, num_points );
+}
+
+#endif /* LV_HAVE_NEON */
+
 #endif // INCLUDED_volk_gnsssdr_16ic_resamplerpuppet_16ic_H

-- 
Alioth's /usr/local/bin/git-commit-notice on /srv/git.debian.org/git/pkg-hamradio/gnss-sdr.git