Fixing NEON proto-kernels

src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_xn_resampler2_16ic_xn.h

@@ -1,12 +1,12 @@
 /*!
- * \file volk_gnsssdr_16ic_xn_resampler_16ic_xn.h
+ * \file volk_gnsssdr_16ic_xn_resampler2_16ic_xn.h
  * \brief VOLK_GNSSSDR kernel: Resamples N 16 bits integer short complex vectors using zero hold resample algorithm.
  * \authors <ul>
  *          <li> Javier Arribas, 2015. jarribas(at)cttc.es
  *          </ul>
  *
- * VOLK_GNSSSDR kernel that esamples N 16 bits integer short complex vectors using zero hold resample algorithm.
+ * VOLK_GNSSSDR kernel that resamples N 16 bits integer short complex vectors using zero hold resample algorithm.
- * It is optimized to resample a sigle GNSS local code signal replica into N vectors fractional-resampled and fractional-delayed
+ * It resamples a single GNSS local code signal replica into N vectors fractional-resampled and fractional-delayed
  * (i.e. it creates the Early, Prompt, and Late code replicas)
  *
  * -------------------------------------------------------------------------
@@ -35,7 +35,7 @@
  */
 
 /*!
- * \page volk_gnsssdr_16ic_xn_resampler_16ic_xn
+ * \page volk_gnsssdr_16ic_xn_resampler2_16ic_xn
  *
  * \b Overview
  *
@@ -43,13 +43,14 @@
  *
  * <b>Dispatcher Prototype</b>
  * \code
- * void volk_gnsssdr_16ic_xn_resampler_16ic_xn(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)
+ * void volk_gnsssdr_16ic_xn_resampler2_16ic_xn(lv_16sc_t** result, const lv_16sc_t* local_code, float* rem_code_phase_chips, float code_phase_step_chips, float* shifts_chips, unsigned int code_length_chips, int num_out_vectors, unsigned int num_output_samples)
  * \endcode
  *
  * \b Inputs
  * \li local_code:            One of the vectors to be multiplied.
  * \li rem_code_phase_chips:  Remnant code phase [chips].
  * \li code_phase_step_chips: Phase increment per sample [chips/sample].
+ * \li shifts_chips:          Vector of floats that defines the spacing (in chips) between the replicas of \p local_code
  * \li code_length_chips:     Code length in chips.
  * \li num_out_vectors        Number of output vectors.
  * \li num_output_samples:    The number of data values to be in the resampled vector.
@@ -149,27 +150,17 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_sse4_1(lv_16sc_t**
                 }
         }
 }
+
 #endif 
 
 
 #ifdef LV_HAVE_SSE4_1
 #include <smmintrin.h>
 static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_u_sse4_1(lv_16sc_t** result, const lv_16sc_t* local_code, float rem_code_phase_chips, float code_phase_step_chips, float* shifts_chips, unsigned int code_length_chips, int num_out_vectors, unsigned int num_output_samples)
-{
-    volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_sse4_1(result, local_code, rem_code_phase_chips, code_phase_step_chips, shifts_chips, code_length_chips, num_out_vectors, num_output_samples);
-}
-
-#endif
-
-
-#ifdef LV_HAVE_SSE3
-#include <pmmintrin.h>
-static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_sse3(lv_16sc_t** result, const lv_16sc_t* local_code, float rem_code_phase_chips, float code_phase_step_chips, float* shifts_chips, unsigned int code_length_chips, int num_out_vectors, unsigned int num_output_samples)
 {
     lv_16sc_t** _result = result;
     const unsigned int quarterPoints = num_output_samples / 4;
 
-    const __m128 ones = _mm_set1_ps(1.0f);
     const __m128 fours = _mm_set1_ps(4.0f);
     const __m128 rem_code_phase_chips_reg = _mm_set_ps1(rem_code_phase_chips);
     const __m128 code_phase_step_chips_reg = _mm_set_ps1(code_phase_step_chips);
@@ -181,7 +172,7 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_sse3(lv_16sc_t** re
     const __m128 code_length_chips_reg_f = _mm_set_ps1((float)code_length_chips);
     const __m128i code_length_chips_reg_i = _mm_set1_epi32((int)code_length_chips);
     __m128i local_code_chip_index_reg, aux_i, negatives, i;
-    __m128 aux, aux2, shifts_chips_reg, fi, igx, j, c, cTrunc, base;
+    __m128 aux, aux2, shifts_chips_reg, c, cTrunc, base;
 
     for (int current_correlator_tap = 0; current_correlator_tap < num_out_vectors; current_correlator_tap++)
         {
@@ -193,11 +184,8 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_sse3(lv_16sc_t** re
                     aux = _mm_mul_ps(code_phase_step_chips_reg, indexn);
                     aux = _mm_add_ps(aux, aux2);
                     // floor
-                    i = _mm_cvttps_epi32(aux);
+                    aux = _mm_floor_ps(aux);
-                    fi = _mm_cvtepi32_ps(i);
+
-                    igx = _mm_cmpgt_ps(fi, aux);
-                    j = _mm_and_ps(igx, ones);
-                    aux = _mm_sub_ps(fi, j);
                     // fmod
                     c = _mm_div_ps(aux, code_length_chips_reg_f);
                     i = _mm_cvttps_epi32(c);
@@ -230,11 +218,142 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_sse3(lv_16sc_t** re
 #endif
 
 
+#ifdef LV_HAVE_SSE3
+#include <pmmintrin.h>
+static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_sse3(lv_16sc_t** result, const lv_16sc_t* local_code, float rem_code_phase_chips, float code_phase_step_chips, float* shifts_chips, unsigned int code_length_chips, int num_out_vectors, unsigned int num_output_samples)
+{
+    lv_16sc_t** _result = result;
+    const unsigned int quarterPoints = num_output_samples / 4;
+
+    const __m128 ones = _mm_set1_ps(1.0f);
+    const __m128 fours = _mm_set1_ps(4.0f);
+    const __m128 rem_code_phase_chips_reg = _mm_set_ps1(rem_code_phase_chips);
+    const __m128 code_phase_step_chips_reg = _mm_set_ps1(code_phase_step_chips);
+
+    __VOLK_ATTR_ALIGNED(16) int local_code_chip_index[4];
+    int local_code_chip_index_;
+
+    const __m128i zeros = _mm_setzero_si128();
+    const __m128 code_length_chips_reg_f = _mm_set_ps1((float)code_length_chips);
+    const __m128i code_length_chips_reg_i = _mm_set1_epi32((int)code_length_chips);
+    __m128i local_code_chip_index_reg, aux_i, negatives, i;
+    __m128 aux, aux2, shifts_chips_reg, fi, igx, j, c, cTrunc, base;
+
+    for (int current_correlator_tap = 0; current_correlator_tap < num_out_vectors; current_correlator_tap++)
+        {
+            shifts_chips_reg = _mm_set_ps1((float)shifts_chips[current_correlator_tap]);
+            aux2 = _mm_sub_ps(shifts_chips_reg, rem_code_phase_chips_reg);
+            __m128 indexn = _mm_set_ps(3.0f, 2.0f, 1.0f, 0.0f);
+            for(unsigned int n = 0; n < quarterPoints; n++)
+                {
+                    //__builtin_prefetch(&_result[current_correlator_tap][4 * n] + 8, 1, 0);
+                    aux = _mm_mul_ps(code_phase_step_chips_reg, indexn);
+                    aux = _mm_add_ps(aux, aux2);
+                    // floor
+                    i = _mm_cvttps_epi32(aux);
+                    fi = _mm_cvtepi32_ps(i);
+                    igx = _mm_cmpgt_ps(fi, aux);
+                    j = _mm_and_ps(igx, ones);
+                    aux = _mm_sub_ps(fi, j);
+                    // fmod
+                    c = _mm_div_ps(aux, code_length_chips_reg_f);
+                    i = _mm_cvttps_epi32(c);
+                    cTrunc = _mm_cvtepi32_ps(i);
+                    base = _mm_mul_ps(cTrunc, code_length_chips_reg_f);
+                    local_code_chip_index_reg = _mm_cvtps_epi32(_mm_sub_ps(aux, base));
+
+                    negatives = _mm_cmplt_epi32(local_code_chip_index_reg, zeros);
+                    aux_i = _mm_and_si128(code_length_chips_reg_i, negatives);
+                    local_code_chip_index_reg = _mm_add_epi32(local_code_chip_index_reg, aux_i);
+                    _mm_store_si128((__m128i*)local_code_chip_index, local_code_chip_index_reg);
+                    for(unsigned int k = 0; k < 4; ++k)
+                        {
+                            _result[current_correlator_tap][n * 4 + k] = local_code[local_code_chip_index[k]];
+                        }
+                    indexn = _mm_add_ps(indexn, fours);
+                }
+            for(unsigned int n = quarterPoints * 4; n < num_output_samples; n++)
+                {
+                    //__builtin_prefetch(&_result[current_correlator_tap][n] + 2, 1, 0);
+                    // resample code for current tap
+                    local_code_chip_index_ = (int)floor(code_phase_step_chips * (float)n + shifts_chips[current_correlator_tap] - rem_code_phase_chips);
+                    local_code_chip_index_ = local_code_chip_index_ % code_length_chips;
+                    //Take into account that in multitap correlators, the shifts can be negative!
+                    if (local_code_chip_index_ < 0) local_code_chip_index_ += code_length_chips;
+                    _result[current_correlator_tap][n] = local_code[local_code_chip_index_];
+                }
+        }
+}
+
+#endif
+
+
 #ifdef LV_HAVE_SSE3
 #include <pmmintrin.h>
 static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_u_sse3(lv_16sc_t** result, const lv_16sc_t* local_code, float rem_code_phase_chips, float code_phase_step_chips, float* shifts_chips, unsigned int code_length_chips, int num_out_vectors, unsigned int num_output_samples)
 {
-    volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_sse3(result, local_code, rem_code_phase_chips, code_phase_step_chips, shifts_chips, code_length_chips, num_out_vectors, num_output_samples);
+    lv_16sc_t** _result = result;
+    const unsigned int quarterPoints = num_output_samples / 4;
+
+    const __m128 ones = _mm_set1_ps(1.0f);
+    const __m128 fours = _mm_set1_ps(4.0f);
+    const __m128 rem_code_phase_chips_reg = _mm_set_ps1(rem_code_phase_chips);
+    const __m128 code_phase_step_chips_reg = _mm_set_ps1(code_phase_step_chips);
+
+    __VOLK_ATTR_ALIGNED(16) int local_code_chip_index[4];
+    int local_code_chip_index_;
+
+    const __m128i zeros = _mm_setzero_si128();
+    const __m128 code_length_chips_reg_f = _mm_set_ps1((float)code_length_chips);
+    const __m128i code_length_chips_reg_i = _mm_set1_epi32((int)code_length_chips);
+    __m128i local_code_chip_index_reg, aux_i, negatives, i;
+    __m128 aux, aux2, shifts_chips_reg, fi, igx, j, c, cTrunc, base;
+
+    for (int current_correlator_tap = 0; current_correlator_tap < num_out_vectors; current_correlator_tap++)
+        {
+            shifts_chips_reg = _mm_set_ps1((float)shifts_chips[current_correlator_tap]);
+            aux2 = _mm_sub_ps(shifts_chips_reg, rem_code_phase_chips_reg);
+            __m128 indexn = _mm_set_ps(3.0f, 2.0f, 1.0f, 0.0f);
+            for(unsigned int n = 0; n < quarterPoints; n++)
+                {
+                    //__builtin_prefetch(&_result[current_correlator_tap][4 * n + 3], 1, 0);
+                    //__builtin_prefetch(&local_code_chip_index[4]);
+                    aux = _mm_mul_ps(code_phase_step_chips_reg, indexn);
+                    aux = _mm_add_ps(aux, aux2);
+                    // floor
+                    i = _mm_cvttps_epi32(aux);
+                    fi = _mm_cvtepi32_ps(i);
+                    igx = _mm_cmpgt_ps(fi, aux);
+                    j = _mm_and_ps(igx, ones);
+                    aux = _mm_sub_ps(fi, j);
+                    // fmod
+                    c = _mm_div_ps(aux, code_length_chips_reg_f);
+                    i = _mm_cvttps_epi32(c);
+                    cTrunc = _mm_cvtepi32_ps(i);
+                    base = _mm_mul_ps(cTrunc, code_length_chips_reg_f);
+                    local_code_chip_index_reg = _mm_cvtps_epi32(_mm_sub_ps(aux, base));
+
+                    negatives = _mm_cmplt_epi32(local_code_chip_index_reg, zeros);
+                    aux_i = _mm_and_si128(code_length_chips_reg_i, negatives);
+                    local_code_chip_index_reg = _mm_add_epi32(local_code_chip_index_reg, aux_i);
+                    _mm_store_si128((__m128i*)local_code_chip_index, local_code_chip_index_reg);
+                    for(unsigned int k = 0; k < 4; ++k)
+                        {
+                            _result[current_correlator_tap][n * 4 + k] = local_code[local_code_chip_index[k]];
+                        }
+                    indexn = _mm_add_ps(indexn, fours);
+                }
+            for(unsigned int n = quarterPoints * 4; n < num_output_samples; n++)
+                {
+                    //__builtin_prefetch(&_result[current_correlator_tap][n], 1, 0);
+                    // resample code for current tap
+                    local_code_chip_index_ = (int)floor(code_phase_step_chips * (float)n + shifts_chips[current_correlator_tap] - rem_code_phase_chips);
+                    local_code_chip_index_ = local_code_chip_index_ % code_length_chips;
+                    //Take into account that in multitap correlators, the shifts can be negative!
+                    if (local_code_chip_index_ < 0) local_code_chip_index_ += code_length_chips;
+                    _result[current_correlator_tap][n] = local_code[local_code_chip_index_];
+                }
+        }
 }
 
 #endif
@@ -260,6 +379,78 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_avx(lv_16sc_t** res
     __m256i local_code_chip_index_reg, i;
     __m256 aux, aux2, shifts_chips_reg, c, cTrunc, base, negatives;
 
+    for (int current_correlator_tap = 0; current_correlator_tap < num_out_vectors; current_correlator_tap++)
+        {
+            shifts_chips_reg = _mm256_set1_ps((float)shifts_chips[current_correlator_tap]);
+            aux2 = _mm256_sub_ps(shifts_chips_reg, rem_code_phase_chips_reg);
+            __m256 indexn = _mm256_set_ps(7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f, 0.0f);
+            for(unsigned int n = 0; n < avx_iters; n++)
+                {
+                    __builtin_prefetch(&_result[current_correlator_tap][8 * n + 7], 1, 0);
+                    __builtin_prefetch(&local_code_chip_index[8]);
+                    aux = _mm256_mul_ps(code_phase_step_chips_reg, indexn);
+                    aux = _mm256_add_ps(aux, aux2);
+                    // floor
+                    aux = _mm256_floor_ps(aux);
+
+                    // fmod
+                    c = _mm256_div_ps(aux, code_length_chips_reg_f);
+                    i = _mm256_cvttps_epi32(c);
+                    cTrunc = _mm256_cvtepi32_ps(i);
+                    base = _mm256_mul_ps(cTrunc, code_length_chips_reg_f);
+                    aux = _mm256_sub_ps(aux, base);
+
+                    negatives = _mm256_cmp_ps(aux, zeros, 0x01);
+                    aux2 = _mm256_and_ps(code_length_chips_reg_f, negatives);
+                    local_code_chip_index_reg = _mm256_cvtps_epi32(_mm256_add_ps(aux, aux2));
+                    _mm256_store_si256((__m256i*)local_code_chip_index, local_code_chip_index_reg);
+                    for(unsigned int k = 0; k < 8; ++k)
+                        {
+                            _result[current_correlator_tap][n * 8 + k] = local_code[local_code_chip_index[k]];
+                        }
+                    indexn = _mm256_add_ps(indexn, eights);
+                }
+        }
+    _mm256_zeroupper();
+
+    for (int current_correlator_tap = 0; current_correlator_tap < num_out_vectors; current_correlator_tap++)
+        {
+            for(unsigned int n = avx_iters * 8; n < num_output_samples; n++)
+                {
+                    // resample code for current tap
+                    __builtin_prefetch(&_result[current_correlator_tap][n], 1, 0);
+                    local_code_chip_index_ = (int)floor(code_phase_step_chips * (float)n + shifts_chips[current_correlator_tap] - rem_code_phase_chips);
+                    local_code_chip_index_ = local_code_chip_index_ % code_length_chips;
+                    //Take into account that in multitap correlators, the shifts can be negative!
+                    if (local_code_chip_index_ < 0) local_code_chip_index_ += code_length_chips;
+                    _result[current_correlator_tap][n] = local_code[local_code_chip_index_];
+                }
+        }
+}
+
+#endif
+
+
+#ifdef LV_HAVE_AVX
+#include <immintrin.h>
+static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_u_avx(lv_16sc_t** result, const lv_16sc_t* local_code, float rem_code_phase_chips, float code_phase_step_chips, float* shifts_chips, unsigned int code_length_chips, int num_out_vectors, unsigned int num_output_samples)
+{
+    lv_16sc_t** _result = result;
+    const unsigned int avx_iters = num_output_samples / 8;
+
+    const __m256 eights = _mm256_set1_ps(8.0f);
+    const __m256 rem_code_phase_chips_reg = _mm256_set1_ps(rem_code_phase_chips);
+    const __m256 code_phase_step_chips_reg = _mm256_set1_ps(code_phase_step_chips);
+
+    __VOLK_ATTR_ALIGNED(32) int local_code_chip_index[8];
+    int local_code_chip_index_;
+
+    const __m256 zeros = _mm256_setzero_ps();
+    const __m256 code_length_chips_reg_f = _mm256_set1_ps((float)code_length_chips);
+
+    __m256i local_code_chip_index_reg, i;
+    __m256 aux, aux2, shifts_chips_reg, c, cTrunc, base, negatives;
+
     for (int current_correlator_tap = 0; current_correlator_tap < num_out_vectors; current_correlator_tap++)
         {
             shifts_chips_reg = _mm256_set1_ps((float)shifts_chips[current_correlator_tap]);
@@ -282,6 +473,7 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_avx(lv_16sc_t** res
                     negatives = _mm256_cmp_ps(aux, zeros, 0x01);
                     aux2 = _mm256_and_ps(code_length_chips_reg_f, negatives);
                     local_code_chip_index_reg = _mm256_cvtps_epi32(_mm256_add_ps(aux, aux2));
+
                     _mm256_store_si256((__m256i*)local_code_chip_index, local_code_chip_index_reg);
                     for(unsigned int k = 0; k < 8; ++k)
                         {
@@ -289,7 +481,10 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_avx(lv_16sc_t** res
                         }
                     indexn = _mm256_add_ps(indexn, eights);
                 }
-
+        }
+    _mm256_zeroupper();
+    for (int current_correlator_tap = 0; current_correlator_tap < num_out_vectors; current_correlator_tap++)
+        {
             for(unsigned int n = avx_iters * 8; n < num_output_samples; n++)
                 {
                     // resample code for current tap
@@ -300,17 +495,6 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_avx(lv_16sc_t** res
                     _result[current_correlator_tap][n] = local_code[local_code_chip_index_];
                 }
         }
-    _mm256_zeroupper();
-}
-
-#endif
-
-
-#ifdef LV_HAVE_AVX
-#include <immintrin.h>
-static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_u_avx(lv_16sc_t** result, const lv_16sc_t* local_code, float rem_code_phase_chips, float code_phase_step_chips, float* shifts_chips, unsigned int code_length_chips, int num_out_vectors, unsigned int num_output_samples)
-{
-    volk_gnsssdr_16ic_xn_resampler2_16ic_xn_a_avx(result, local_code, rem_code_phase_chips, code_phase_step_chips, shifts_chips, code_length_chips, num_out_vectors, num_output_samples);
 }
 
 #endif
@@ -318,7 +502,6 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_u_avx(lv_16sc_t** res
 
 #ifdef LV_HAVE_NEON
 #include <arm_neon.h>
-#include <volk_gnsssdr/volk_gnsssdr_neon_intrinsics.h>
 static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_neon(lv_16sc_t** result, const lv_16sc_t* local_code, float rem_code_phase_chips, float code_phase_step_chips, float* shifts_chips, unsigned int code_length_chips, int num_out_vectors, unsigned int num_output_samples)
 {
     lv_16sc_t** _result = result;
@@ -328,26 +511,26 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_neon(lv_16sc_t** resu
     const float32x4_t rem_code_phase_chips_reg = vdupq_n_f32(rem_code_phase_chips);
     const float32x4_t code_phase_step_chips_reg = vdupq_n_f32(code_phase_step_chips);
 
-    __attribute__((aligned(16))) int local_code_chip_index[4];
+    __VOLK_ATTR_ALIGNED(16) int32_t local_code_chip_index[4];
-    int local_code_chip_index_;
+    int32_t local_code_chip_index_;
-
     const int32x4_t zeros = vdupq_n_s32(0);
     const float32x4_t code_length_chips_reg_f = vdupq_n_f32((float)code_length_chips);
-    const int32x4_t code_length_chips_reg_i = vdupq_n_s32((int)code_length_chips);
+
+    const int32x4_t code_length_chips_reg_i = vdupq_n_s32((int32_t)code_length_chips);
     int32x4_t local_code_chip_index_reg, aux_i, negatives, i;
-    float32x4_t aux, aux2, shifts_chips_reg, fi, c, j, cTrunc, base;
+    float32x4_t aux, aux2, shifts_chips_reg, fi, c, j, cTrunc, base, indexn, reciprocal;
-    __attribute__((aligned(16))) float vec[4] = { 3.0f, 2.0f, 1.0f, 0.0f };
+    __VOLK_ATTR_ALIGNED(16) const float vec[4] = { 0.0f, 1.0f, 2.0f, 3.0f };
     uint32x4_t igx;
-    float32x4_t half = vdupq_n_f32(0.5f);
+
-    float32x4_t sign, PlusHalf, Round;
     for (int current_correlator_tap = 0; current_correlator_tap < num_out_vectors; current_correlator_tap++)
         {
             shifts_chips_reg = vdupq_n_f32((float)shifts_chips[current_correlator_tap]);
             aux2 = vsubq_f32(shifts_chips_reg, rem_code_phase_chips_reg);
-            float32x4_t indexn = vld1q_f32((float*)vec);
+            indexn = vld1q_f32((float*)vec);
-
             for(unsigned int n = 0; n < neon_iters; n++)
                 {
+                    __builtin_prefetch(&_result[current_correlator_tap][4 * n + 3], 1, 0);
+                    __builtin_prefetch(&local_code_chip_index[4]);
                     aux = vmulq_f32(code_phase_step_chips_reg, indexn);
                     aux = vaddq_f32(aux, aux2);
                     // floor
@@ -358,20 +541,21 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_neon(lv_16sc_t** resu
                     aux = vsubq_f32(fi, j);
 
                     // fmod
-                    c = vdivq_f32(aux, code_length_chips_reg_f);
+                    reciprocal = vrecpeq_f32(code_length_chips_reg_f);
+                    reciprocal = vmulq_f32(vrecpsq_f32(code_length_chips_reg_f, reciprocal), reciprocal);
+                    reciprocal = vmulq_f32(vrecpsq_f32(code_length_chips_reg_f, reciprocal), reciprocal); // this refinement is required!
+                    c = vmulq_f32(aux, reciprocal);
                     i =  vcvtq_s32_f32(c);
                     cTrunc = vcvtq_f32_s32(i);
                     base = vmulq_f32(cTrunc, code_length_chips_reg_f);
                     aux = vsubq_f32(aux, base);
-                    sign = vcvtq_f32_u32((vshrq_n_u32(vreinterpretq_u32_f32(aux), 31)));
+                    local_code_chip_index_reg = vcvtq_s32_f32(aux);
-                    PlusHalf = vaddq_f32(aux, half);
-                    Round = vsubq_f32(PlusHalf, sign);
-                    local_code_chip_index_reg = vcvtq_s32_f32(Round);
 
                     negatives = vreinterpretq_s32_u32(vcltq_s32(local_code_chip_index_reg, zeros));
                     aux_i = vandq_s32(code_length_chips_reg_i, negatives);
                     local_code_chip_index_reg = vaddq_s32(local_code_chip_index_reg, aux_i);
-                    vst1q_s32((int32_t*)local_code_chip_index, local_code_chip_index_reg);
+
+                    vst1q_s32((int*)local_code_chip_index, local_code_chip_index_reg);
                     for(unsigned int k = 0; k < 4; ++k)
                         {
                             _result[current_correlator_tap][n * 4 + k] = local_code[local_code_chip_index[k]];
@@ -380,8 +564,9 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_neon(lv_16sc_t** resu
                 }
             for(unsigned int n = neon_iters * 4; n < num_output_samples; n++)
                 {
+                    __builtin_prefetch(&_result[current_correlator_tap][n], 1, 0);
                     // resample code for current tap
-                    local_code_chip_index_ = (int)floor(code_phase_step_chips * (float)n + shifts_chips[current_correlator_tap] - rem_code_phase_chips);
+                    local_code_chip_index_ = (int32_t)floor(code_phase_step_chips * (float)n + shifts_chips[current_correlator_tap] - rem_code_phase_chips);
                     local_code_chip_index_ = local_code_chip_index_ % code_length_chips;
                     //Take into account that in multitap correlators, the shifts can be negative!
                     if (local_code_chip_index_ < 0) local_code_chip_index_ += code_length_chips;
@@ -390,6 +575,7 @@ static inline void volk_gnsssdr_16ic_xn_resampler2_16ic_xn_neon(lv_16sc_t** resu
         }
 }
 
+
 #endif
 
 

src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_32fc_xn_resampler_32fc_xn.h

@@ -335,7 +335,7 @@ static inline void volk_gnsssdr_32fc_xn_resampler_32fc_xn_neon(lv_32fc_t** resul
     const float32x4_t code_length_chips_reg_f = vdupq_n_f32((float)code_length_chips);
     const int32x4_t code_length_chips_reg_i = vdupq_n_s32((int32_t)code_length_chips);
     int32x4_t local_code_chip_index_reg, aux_i, negatives, i;
-    float32x4_t aux, aux2, shifts_chips_reg, fi, c, j, cTrunc, base, indexn;
+    float32x4_t aux, aux2, shifts_chips_reg, fi, c, j, cTrunc, base, indexn, reciprocal;
     __VOLK_ATTR_ALIGNED(16) const float vec[4] = { 0.0f, 1.0f, 2.0f, 3.0f };
     uint32x4_t igx;
     for (int current_correlator_tap = 0; current_correlator_tap < num_out_vectors; current_correlator_tap++)
@@ -345,6 +345,8 @@ static inline void volk_gnsssdr_32fc_xn_resampler_32fc_xn_neon(lv_32fc_t** resul
             indexn = vld1q_f32((float*)vec);
             for(unsigned int n = 0; n < neon_iters; n++)
                 {
+                    __builtin_prefetch(&_result[current_correlator_tap][4 * n + 3], 1, 0);
+                    __builtin_prefetch(&local_code_chip_index[4]);
                     aux = vmulq_f32(code_phase_step_chips_reg, indexn);
                     aux = vaddq_f32(aux, aux2);
                     // floor
@@ -354,7 +356,10 @@ static inline void volk_gnsssdr_32fc_xn_resampler_32fc_xn_neon(lv_32fc_t** resul
                     j = vcvtq_f32_s32(vandq_s32(vreinterpretq_s32_u32(igx), ones));
                     aux = vsubq_f32(fi, j);
                     // fmod
-                    c = vdivq_f32(aux, code_length_chips_reg_f);
+                    reciprocal = vrecpeq_f32(code_length_chips_reg_f);
+                    reciprocal = vmulq_f32(vrecpsq_f32(code_length_chips_reg_f, reciprocal), reciprocal);
+                    reciprocal = vmulq_f32(vrecpsq_f32(code_length_chips_reg_f, reciprocal), reciprocal); // this refinement is required!
+                    c = vmulq_f32(aux, reciprocal);
                     i =  vcvtq_s32_f32(c);
                     cTrunc = vcvtq_f32_s32(i);
                     base = vmulq_f32(cTrunc, code_length_chips_reg_f);
@@ -364,6 +369,7 @@ static inline void volk_gnsssdr_32fc_xn_resampler_32fc_xn_neon(lv_32fc_t** resul
                     negatives = vreinterpretq_s32_u32(vcltq_s32(local_code_chip_index_reg, zeros));
                     aux_i = vandq_s32(code_length_chips_reg_i, negatives);
                     local_code_chip_index_reg = vaddq_s32(local_code_chip_index_reg, aux_i);
+
                     vst1q_s32((int32_t*)local_code_chip_index, local_code_chip_index_reg);
 
                     for(unsigned int k = 0; k < 4; ++k)
@@ -374,6 +380,7 @@ static inline void volk_gnsssdr_32fc_xn_resampler_32fc_xn_neon(lv_32fc_t** resul
                 }
             for(unsigned int n = neon_iters * 4; n < num_points; n++)
                 {
+                    __builtin_prefetch(&_result[current_correlator_tap][n], 1, 0);
                     // resample code for current tap
                     local_code_chip_index_ = (int32_t)floor(code_phase_step_chips * (float)n + shifts_chips[current_correlator_tap] - rem_code_phase_chips);
                     local_code_chip_index_ = local_code_chip_index_ % code_length_chips;