sync

src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/CMakeLists.txt

@@ -176,6 +176,7 @@ install(FILES
     ${PROJECT_BINARY_DIR}/include/volk_gnsssdr/volk_gnsssdr_config_fixed.h
     ${PROJECT_BINARY_DIR}/include/volk_gnsssdr/volk_gnsssdr_typedefs.h
     ${PROJECT_SOURCE_DIR}/include/volk_gnsssdr/volk_gnsssdr_malloc.h
+    ${PROJECT_SOURCE_DIR}/include/volk_gnsssdr/volk_gnsssdr_sine_table.h
     DESTINATION include/volk_gnsssdr
     COMPONENT "volk_gnsssdr_devel"
 )

src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic.h

@@ -336,8 +336,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_vma(lv_16sc_t* out, c
             tmp.val[0] = vmls_s16(tmp.val[0], a_val.val[1], b_val.val[1]);
             tmp.val[1] = vmla_s16(tmp.val[1], a_val.val[0], b_val.val[1]);
 
-            accumulator.val[0] = vadd_s16(accumulator.val[0], tmp.val[0]);
+            accumulator.val[0] = vqadd_s16(accumulator.val[0], tmp.val[0]);
-            accumulator.val[1] = vadd_s16(accumulator.val[1], tmp.val[1]);
+            accumulator.val[1] = vqadd_s16(accumulator.val[1], tmp.val[1]);
 
             a_ptr += 4;
             b_ptr += 4;
@@ -355,4 +355,57 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_vma(lv_16sc_t* out, c
 
 #endif /* LV_HAVE_NEON */
 
+
+#ifdef LV_HAVE_NEON
+#include <arm_neon.h>
+
+static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_neon_optvma(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, accumulator1, accumulator2;
+
+    __VOLK_ATTR_ALIGNED(16) lv_16sc_t accum_result[4];
+    accumulator1.val[0] = vdup_n_s16(0);
+    accumulator1.val[1] = vdup_n_s16(0);
+    accumulator2.val[0] = vdup_n_s16(0);
+    accumulator2.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);
+
+            // use 2 accumulators to remove inter-instruction data dependencies
+            accumulator1.val[0] = vmla_s16(accumulator1.val[0], a_val.val[0], b_val.val[0]);
+            accumulator1.val[1] = vmla_s16(accumulator1.val[1], a_val.val[0], b_val.val[1]);
+            accumulator2.val[0] = vmls_s16(accumulator2.val[0], a_val.val[1], b_val.val[1]);
+            accumulator2.val[1] = vmla_s16(accumulator2.val[1], a_val.val[1], b_val.val[0]);
+
+            a_ptr += 4;
+            b_ptr += 4;
+        }
+
+    accumulator1.val[0] = vqadd_s16(accumulator1.val[0], accumulator2.val[0]);
+    accumulator1.val[1] = vqadd_s16(accumulator1.val[1], accumulator2.val[1]);
+
+    vst2_s16((int16_t*)accum_result, accumulator1);
+    *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*/

src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dot_prod_16ic_xn.h

@@ -62,6 +62,8 @@
 
 
 #include <volk_gnsssdr/volk_gnsssdr_complex.h>
+#include <volk_gnsssdr/volk_gnsssdr_malloc.h>
+#include <volk_gnsssdr/volk_gnsssdr.h>
 #include <volk_gnsssdr/saturation_arithmetic.h>
 
 #ifdef LV_HAVE_GENERIC
@@ -120,11 +122,14 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_sse2(lv_16sc_t* resul
         {
             __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
 
-            __m128i* realcacc;
+            __m128i* realcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
-            __m128i* imagcacc;
+            __m128i* imagcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
 
-            realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+            for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
-            imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+                {
+                    realcacc[n_vec] = _mm_setzero_si128();
+                    imagcacc[n_vec] = _mm_setzero_si128();
+                }
 
             __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag;
 
@@ -176,8 +181,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_a_sse2(lv_16sc_t* resul
                         }
                     _out[n_vec] = dotProduct;
                 }
-            free(realcacc);
+            volk_gnsssdr_free(realcacc);
-            free(imagcacc);
+            volk_gnsssdr_free(imagcacc);
         }
 
     for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
@@ -211,11 +216,14 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul
         {
             __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
 
-            __m128i* realcacc;
+            __m128i* realcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
-            __m128i* imagcacc;
+            __m128i* imagcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
 
-            realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+            for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
-            imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+                {
+                    realcacc[n_vec] = _mm_setzero_si128();
+                    imagcacc[n_vec] = _mm_setzero_si128();
+                }
 
             __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag;
 
@@ -246,7 +254,6 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul
 
                             realcacc[n_vec] = _mm_adds_epi16(realcacc[n_vec], real);
                             imagcacc[n_vec] = _mm_adds_epi16(imagcacc[n_vec], imag);
-
                         }
                     _in_common += 4;
                 }
@@ -267,8 +274,8 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* resul
                         }
                     _out[n_vec] = dotProduct;
                 }
-            free(realcacc);
+            volk_gnsssdr_free(realcacc);
-            free(imagcacc);
+            volk_gnsssdr_free(imagcacc);
         }
 
     for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
@@ -304,9 +311,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon(lv_16sc_t* result,
 
             int16x4x2_t a_val, b_val, c_val;
 
-            //todo dyn mem reg
+            int16x4x2_t* accumulator = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment());
-            int16x4x2_t* accumulator;
-            accumulator = (int16x4x2_t*)malloc(num_a_vectors * sizeof(int16x4x2_t));
 
             int16x4x2_t tmp_real, tmp_imag;
 
@@ -357,7 +362,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon(lv_16sc_t* result,
                         }
                     _out[n_vec] = dotProduct;
                 }
-            free(accumulator);
+            volk_gnsssdr_free(accumulator);
         }
 
     for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
@@ -393,8 +398,7 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon_vma(lv_16sc_t* res
 
             int16x4x2_t a_val, b_val, tmp;
 
-            int16x4x2_t* accumulator;
+            int16x4x2_t* accumulator = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment());
-            accumulator = (int16x4x2_t*)malloc(num_a_vectors * sizeof(int16x4x2_t));
 
             for(int n_vec = 0; n_vec < num_a_vectors; n_vec++)
                 {
@@ -434,7 +438,88 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon_vma(lv_16sc_t* res
                         }
                     _out[n_vec] = dotProduct;
                 }
-            free(accumulator);
+            volk_gnsssdr_free(accumulator);
+        }
+
+    for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+        {
+            for(unsigned int n  = neon_iters * 4; n < num_points; n++)
+                {
+                    lv_16sc_t tmp = in_common[n] * in_a[n_vec][n];
+
+                    _out[n_vec] = lv_cmake(sat_adds16i(lv_creal(_out[n_vec]), lv_creal(tmp)),
+                            sat_adds16i(lv_cimag(_out[n_vec]), lv_cimag(tmp)));
+                }
+        }
+}
+#endif /* LV_HAVE_NEON */
+
+
+#ifdef LV_HAVE_NEON
+#include <arm_neon.h>
+
+static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon_optvma(lv_16sc_t* result, const lv_16sc_t* in_common, const lv_16sc_t** in_a,  int num_a_vectors, unsigned int num_points)
+{
+    lv_16sc_t dotProduct = lv_cmake(0,0);
+
+    const unsigned int neon_iters = num_points / 4;
+
+    const lv_16sc_t** _in_a = in_a;
+    const lv_16sc_t* _in_common = in_common;
+    lv_16sc_t* _out = result;
+
+    if (neon_iters > 0)
+        {
+            __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
+
+            int16x4x2_t a_val, b_val;
+
+            int16x4x2_t* accumulator1 = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment());
+            int16x4x2_t* accumulator2 = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment());
+
+            for(int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                {
+                    accumulator1[n_vec].val[0] = vdup_n_s16(0);
+                    accumulator1[n_vec].val[1] = vdup_n_s16(0);
+                    accumulator2[n_vec].val[0] = vdup_n_s16(0);
+                    accumulator2[n_vec].val[1] = vdup_n_s16(0);
+                }
+
+            for(unsigned int number = 0; number < neon_iters; number++)
+                {
+                    b_val = vld2_s16((int16_t*)_in_common); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
+                    __builtin_prefetch(_in_common + 8);
+                    for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                        {
+                            a_val = vld2_s16((int16_t*)&(_in_a[n_vec][number*4]));
+
+                            accumulator1[n_vec].val[0] = vmla_s16(accumulator1[n_vec].val[0], a_val.val[0], b_val.val[0]);
+                            accumulator1[n_vec].val[1] = vmla_s16(accumulator1[n_vec].val[1], a_val.val[0], b_val.val[1]);
+                            accumulator2[n_vec].val[0] = vmls_s16(accumulator2[n_vec].val[0], a_val.val[1], b_val.val[1]);
+                            accumulator2[n_vec].val[1] = vmla_s16(accumulator2[n_vec].val[1], a_val.val[1], b_val.val[0]);
+                        }
+                    _in_common += 4;
+                }
+
+            for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                {
+                    accumulator1[n_vec].val[0] = vqadd_s16(accumulator1[n_vec].val[0], accumulator2[n_vec].val[0]);
+                    accumulator1[n_vec].val[1] = vqadd_s16(accumulator1[n_vec].val[1], accumulator2[n_vec].val[1]);
+                }
+
+            for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                {
+                    vst2_s16((int16_t*)dotProductVector, accumulator1[n_vec]); // Store the results back into the dot product vector
+                    dotProduct = lv_cmake(0,0);
+                    for (int i = 0; i < 4; ++i)
+                        {
+                            dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[i])),
+                                    sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[i])));
+                        }
+                    _out[n_vec] = dotProduct;
+                }
+            volk_gnsssdr_free(accumulator1);
+            volk_gnsssdr_free(accumulator2);
         }
 
     for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)

src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic.h

@@ -110,7 +110,7 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_a_sse2(lv_16sc_t* r
 #endif /*  SSE2   */
 
 
-#if LV_HAVE_SSE2 && LV_HAVE_64
+#if LV_HAVE_SSE2
 
 static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_u_sse2(lv_16sc_t* result, const lv_16sc_t* local_code, const lv_16sc_t* in, unsigned int num_points)
 {
@@ -131,7 +131,7 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_u_sse2(lv_16sc_t* r
     volk_gnsssdr_free(in_a);
 }
 
-#endif /* LV_HAVE_SSE2 && LV_HAVE_64 */
+#endif /* LV_HAVE_SSE2  */
 
 
 #ifdef LV_HAVE_NEON
@@ -180,6 +180,30 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_neon_vma(lv_16sc_t*
 }
 
 #endif // NEON
+
+#ifdef LV_HAVE_NEON
+
+static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_neon_optvma(lv_16sc_t* result, const lv_16sc_t* local_code, const lv_16sc_t* in, unsigned int num_points)
+{
+    int num_a_vectors = 3;
+    lv_16sc_t** in_a = (lv_16sc_t**)volk_gnsssdr_malloc(sizeof(lv_16sc_t*) * num_a_vectors, volk_gnsssdr_get_alignment());
+    for(unsigned int n = 0; n < num_a_vectors; n++)
+    {
+       in_a[n] = (lv_16sc_t*)volk_gnsssdr_malloc(sizeof(lv_16sc_t)*num_points, volk_gnsssdr_get_alignment());
+       memcpy((lv_16sc_t*)in_a[n], (lv_16sc_t*)in, sizeof(lv_16sc_t)*num_points);
+    }
+
+    volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon_optvma(result, local_code, (const lv_16sc_t**) in_a, num_a_vectors, num_points);
+
+    for(unsigned int n = 0; n < num_a_vectors; n++)
+    {
+        volk_gnsssdr_free(in_a[n]);
+    }
+    volk_gnsssdr_free(in_a);
+}
+
+#endif // NEON
+
 #endif  // INCLUDED_volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_H
 
 

src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn.h

@@ -69,6 +69,9 @@
 #define INCLUDED_volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_H
 
 
+#include <volk_gnsssdr/volk_gnsssdr.h>
+#include <volk_gnsssdr/volk_gnsssdr_malloc.h>
+
 #include <volk_gnsssdr/volk_gnsssdr_complex.h>
 #include <volk_gnsssdr/saturation_arithmetic.h>
 #include <math.h>
@@ -184,13 +187,14 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3(lv_16sc_
 
     __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
 
-    //todo dyn mem reg
+    __m128i* realcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
+    __m128i* imagcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
 
-    __m128i* realcacc;
+    for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
-    __m128i* imagcacc;
+        {
-
+            realcacc[n_vec] = _mm_setzero_si128();
-    realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+            imagcacc[n_vec] = _mm_setzero_si128();
-    imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+        }
 
     __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl;
 
@@ -308,8 +312,8 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3(lv_16sc_
                 }
             _out[n_vec] = dotProduct;
         }
-    free(realcacc);
+    volk_gnsssdr_free(realcacc);
-    free(imagcacc);
+    volk_gnsssdr_free(imagcacc);
 
     tmp1 = _mm_mul_ps(two_phase_acc_reg, two_phase_acc_reg);
     tmp2 = _mm_hadd_ps(tmp1, tmp1);
@@ -356,13 +360,14 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3_reload(l
 
     __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
 
-    //todo dyn mem reg
+    __m128i* realcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
+    __m128i* imagcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
 
-    __m128i* realcacc;
+    for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
-    __m128i* imagcacc;
+        {
-
+            realcacc[n_vec] = _mm_setzero_si128();
-    realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+            imagcacc[n_vec] = _mm_setzero_si128();
-    imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+        }
 
     __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl;
 
@@ -550,8 +555,8 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_a_sse3_reload(l
             _out[n_vec] = dotProduct;
         }
 
-    free(realcacc);
+    volk_gnsssdr_free(realcacc);
-    free(imagcacc);
+    volk_gnsssdr_free(imagcacc);
 
     tmp1 = _mm_mul_ps(two_phase_acc_reg, two_phase_acc_reg);
     tmp2 = _mm_hadd_ps(tmp1, tmp1);
@@ -598,13 +603,14 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
     lv_16sc_t* _out = result;
     __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
 
-    //todo dyn mem reg
+    __m128i* realcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
+    __m128i* imagcacc = (__m128i*)volk_gnsssdr_malloc(num_a_vectors * sizeof(__m128i), volk_gnsssdr_get_alignment());
 
-    __m128i* realcacc;
+    for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
-    __m128i* imagcacc;
+        {
-
+            realcacc[n_vec] = _mm_setzero_si128();
-    realcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+            imagcacc[n_vec] = _mm_setzero_si128();
-    imagcacc = (__m128i*)calloc(num_a_vectors, sizeof(__m128i)); //calloc also sets memory to 0
+        }
 
     __m128i a, b, c, c_sr, mask_imag, mask_real, real, imag, imag1, imag2, b_sl, a_sl;
 
@@ -722,8 +728,8 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_u_sse3(lv_16sc_
                 }
             _out[n_vec] = dotProduct;
         }
-    free(realcacc);
+    volk_gnsssdr_free(realcacc);
-    free(imagcacc);
+    volk_gnsssdr_free(imagcacc);
 
     _mm_storeu_ps((float*)two_phase_acc, two_phase_acc_reg);
     (*phase) = two_phase_acc[0];
@@ -792,8 +798,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon(lv_16sc_t*
             float32x4x2_t tmp32f, tmp32_real, tmp32_imag;
             float32x4_t sign, PlusHalf, Round;
 
-            int16x4x2_t* accumulator;
+            int16x4x2_t* accumulator = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment());
-            accumulator = (int16x4x2_t*)calloc(num_a_vectors, sizeof(int16x4x2_t));
 
             for(int n_vec = 0; n_vec < num_a_vectors; n_vec++)
                 {
@@ -904,7 +909,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon(lv_16sc_t*
                         }
                     _out[n_vec] = dotProduct;
                 }
-            free(accumulator);
+            volk_gnsssdr_free(accumulator);
             vst1q_f32((float32_t*)__phase_real, _phase_real);
             vst1q_f32((float32_t*)__phase_imag, _phase_imag);
 
@@ -976,8 +981,7 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon_vma(lv_16s
             float32x4x2_t tmp32f, tmp32_real, tmp32_imag;
             float32x4_t sign, PlusHalf, Round;
 
-            int16x4x2_t* accumulator;
+            int16x4x2_t* accumulator = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment());
-            accumulator = (int16x4x2_t*)calloc(num_a_vectors, sizeof(int16x4x2_t));
 
             for(int n_vec = 0; n_vec < num_a_vectors; n_vec++)
                 {
@@ -1095,7 +1099,189 @@ static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon_vma(lv_16s
                         }
                     _out[n_vec] = dotProduct;
                 }
-            free(accumulator);
+            volk_gnsssdr_free(accumulator);
+
+            vst1q_f32((float32_t*)__phase_real, _phase_real);
+            vst1q_f32((float32_t*)__phase_imag, _phase_imag);
+
+            (*phase) = lv_cmake((float32_t)__phase_real[0], (float32_t)__phase_imag[0]);
+        }
+
+    for (unsigned int n = neon_iters * 4; n < num_points; n++)
+        {
+            tmp16_ = in_common[n];  //printf("neon phase %i: %f,%f\n", n,lv_creal(*phase),lv_cimag(*phase));
+            tmp32_ = lv_cmake((float32_t)lv_creal(tmp16_), (float32_t)lv_cimag(tmp16_)) * (*phase);
+            tmp16_ = lv_cmake((int16_t)rintf(lv_creal(tmp32_)), (int16_t)rintf(lv_cimag(tmp32_)));
+            (*phase) *= phase_inc;
+            for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                {
+                    tmp = tmp16_ * in_a[n_vec][n];
+                    _out[n_vec] = lv_cmake(sat_adds16i(lv_creal(_out[n_vec]), lv_creal(tmp)), sat_adds16i(lv_cimag(_out[n_vec]), lv_cimag(tmp)));
+                }
+        }
+}
+
+#endif /* LV_HAVE_NEON */
+
+
+#ifdef LV_HAVE_NEON
+#include <arm_neon.h>
+#include <volk_gnsssdr/volk_gnsssdr_neon_intrinsics.h>
+
+static inline void volk_gnsssdr_16ic_x2_rotator_dot_prod_16ic_xn_neon_optvma(lv_16sc_t* result, const lv_16sc_t* in_common, const lv_32fc_t phase_inc, lv_32fc_t* phase, const lv_16sc_t** in_a,  int num_a_vectors, unsigned int num_points)
+{
+    const unsigned int neon_iters = num_points / 4;
+
+    const lv_16sc_t** _in_a = in_a;
+    const lv_16sc_t* _in_common = in_common;
+    lv_16sc_t* _out = result;
+
+    lv_16sc_t tmp16_, tmp;
+    lv_32fc_t tmp32_;
+
+    if (neon_iters > 0)
+        {
+            lv_16sc_t dotProduct = lv_cmake(0,0);
+            float arg_phase0 = cargf(*phase);
+            float arg_phase_inc = cargf(phase_inc);
+            float phase_est;
+
+            lv_32fc_t ___phase4 = phase_inc * phase_inc * phase_inc * phase_inc;
+            __VOLK_ATTR_ALIGNED(16) float32_t __phase4_real[4] = { lv_creal(___phase4), lv_creal(___phase4), lv_creal(___phase4), lv_creal(___phase4) };
+            __VOLK_ATTR_ALIGNED(16) float32_t __phase4_imag[4] = { lv_cimag(___phase4), lv_cimag(___phase4), lv_cimag(___phase4), lv_cimag(___phase4) };
+
+            float32x4_t _phase4_real = vld1q_f32(__phase4_real);
+            float32x4_t _phase4_imag = vld1q_f32(__phase4_imag);
+
+            lv_32fc_t phase2 = (lv_32fc_t)(*phase) * phase_inc;
+            lv_32fc_t phase3 = phase2 * phase_inc;
+            lv_32fc_t phase4 = phase3 * phase_inc;
+
+            __VOLK_ATTR_ALIGNED(16) float32_t __phase_real[4] = { lv_creal((*phase)), lv_creal(phase2), lv_creal(phase3), lv_creal(phase4) };
+            __VOLK_ATTR_ALIGNED(16) float32_t __phase_imag[4] = { lv_cimag((*phase)), lv_cimag(phase2), lv_cimag(phase3), lv_cimag(phase4) };
+
+            float32x4_t _phase_real = vld1q_f32(__phase_real);
+            float32x4_t _phase_imag = vld1q_f32(__phase_imag);
+
+            int16x4x2_t a_val, b_val;
+            __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
+            float32x4_t half = vdupq_n_f32(0.5f);
+            int32x4x2_t tmp32i;
+
+            float32x4x2_t tmp32f, tmp32_real, tmp32_imag;
+            float32x4_t sign, PlusHalf, Round;
+
+            int16x4x2_t* accumulator1 = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment());
+            int16x4x2_t* accumulator2 = (int16x4x2_t*)volk_gnsssdr_malloc(num_a_vectors * sizeof(int16x4x2_t), volk_gnsssdr_get_alignment());
+
+            for(int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                {
+                    accumulator1[n_vec].val[0] = vdup_n_s16(0);
+                    accumulator1[n_vec].val[1] = vdup_n_s16(0);
+                    accumulator2[n_vec].val[0] = vdup_n_s16(0);
+                    accumulator2[n_vec].val[1] = vdup_n_s16(0);
+                }
+
+            for(unsigned int number = 0; number < neon_iters; number++)
+                {
+                    /* load 4 complex numbers (int 16 bits each component) */
+                    b_val = vld2_s16((int16_t*)_in_common);
+                    __builtin_prefetch(_in_common + 8);
+                    _in_common += 4;
+
+                    /* promote them to int 32 bits */
+                    tmp32i.val[0] = vmovl_s16(b_val.val[0]);
+                    tmp32i.val[1] = vmovl_s16(b_val.val[1]);
+
+                    /* promote them to float 32 bits */
+                    tmp32f.val[0] = vcvtq_f32_s32(tmp32i.val[0]);
+                    tmp32f.val[1] = vcvtq_f32_s32(tmp32i.val[1]);
+
+                    /* complex multiplication of four complex samples (float 32 bits each component) */
+                    tmp32_real.val[0] = vmulq_f32(tmp32f.val[0], _phase_real);
+                    tmp32_real.val[1] = vmulq_f32(tmp32f.val[1], _phase_imag);
+                    tmp32_imag.val[0] = vmulq_f32(tmp32f.val[0], _phase_imag);
+                    tmp32_imag.val[1] = vmulq_f32(tmp32f.val[1], _phase_real);
+
+                    tmp32f.val[0] = vsubq_f32(tmp32_real.val[0], tmp32_real.val[1]);
+                    tmp32f.val[1] = vaddq_f32(tmp32_imag.val[0], tmp32_imag.val[1]);
+
+                    /* downcast results to int32 */
+                    /* in __aarch64__ we can do that with vcvtaq_s32_f32(ret1); vcvtaq_s32_f32(ret2); */
+                    sign = vcvtq_f32_u32((vshrq_n_u32(vreinterpretq_u32_f32(tmp32f.val[0]), 31)));
+                    PlusHalf = vaddq_f32(tmp32f.val[0], half);
+                    Round = vsubq_f32(PlusHalf, sign);
+                    tmp32i.val[0] = vcvtq_s32_f32(Round);
+
+                    sign = vcvtq_f32_u32((vshrq_n_u32(vreinterpretq_u32_f32(tmp32f.val[1]), 31)));
+                    PlusHalf = vaddq_f32(tmp32f.val[1], half);
+                    Round = vsubq_f32(PlusHalf, sign);
+                    tmp32i.val[1] = vcvtq_s32_f32(Round);
+
+                    /* downcast results to int16 */
+                    b_val.val[0] = vqmovn_s32(tmp32i.val[0]);
+                    b_val.val[1] = vqmovn_s32(tmp32i.val[1]);
+
+                    /* compute next four phases */
+                    tmp32_real.val[0] = vmulq_f32(_phase_real, _phase4_real);
+                    tmp32_real.val[1] = vmulq_f32(_phase_imag, _phase4_imag);
+                    tmp32_imag.val[0] = vmulq_f32(_phase_real, _phase4_imag);
+                    tmp32_imag.val[1] = vmulq_f32(_phase_imag, _phase4_real);
+
+                    _phase_real = vsubq_f32(tmp32_real.val[0], tmp32_real.val[1]);
+                    _phase_imag = vaddq_f32(tmp32_imag.val[0], tmp32_imag.val[1]);
+
+                    // Regenerate phase
+                    if ((number % 256) == 0)
+                        {
+                            //printf("computed phase: %f\n", cos(cargf(lv_cmake(_phase_real[0],_phase_imag[0]))));
+                            phase_est = arg_phase0 + (number + 1) * 4 * arg_phase_inc;
+                            //printf("Estimated phase: %f\n\n", cos(phase_est));
+
+                            *phase = lv_cmake(cos(phase_est), sin(phase_est));
+                            phase2 = (lv_32fc_t)(*phase) * phase_inc;
+                            phase3 = phase2 * phase_inc;
+                            phase4 = phase3 * phase_inc;
+
+                            __VOLK_ATTR_ALIGNED(16) float32_t ____phase_real[4] = { lv_creal((*phase)), lv_creal(phase2), lv_creal(phase3), lv_creal(phase4) };
+                            __VOLK_ATTR_ALIGNED(16) float32_t ____phase_imag[4] = { lv_cimag((*phase)), lv_cimag(phase2), lv_cimag(phase3), lv_cimag(phase4) };
+
+                            _phase_real = vld1q_f32(____phase_real);
+                            _phase_imag = vld1q_f32(____phase_imag);
+                        }
+
+                    vst1q_f32((float32_t*)__phase_real, _phase_real);
+                    vst1q_f32((float32_t*)__phase_imag, _phase_imag);
+
+                    for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                        {
+                            a_val = vld2_s16((int16_t*)&(_in_a[n_vec][number*4]));
+
+                            // use 2 accumulators to remove inter-instruction data dependencies
+                            accumulator1[n_vec].val[0] = vmla_s16(accumulator1[n_vec].val[0], a_val.val[0], b_val.val[0]);
+                            accumulator1[n_vec].val[1] = vmla_s16(accumulator1[n_vec].val[1], a_val.val[0], b_val.val[1]);
+                            accumulator2[n_vec].val[0] = vmls_s16(accumulator2[n_vec].val[0], a_val.val[1], b_val.val[1]);
+                            accumulator2[n_vec].val[1] = vmla_s16(accumulator2[n_vec].val[1], a_val.val[1], b_val.val[0]);
+                        }
+                }
+            for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                {
+                    accumulator1[n_vec].val[0] = vqadd_s16(accumulator1[n_vec].val[0], accumulator2[n_vec].val[0]);
+                    accumulator1[n_vec].val[1] = vqadd_s16(accumulator1[n_vec].val[1], accumulator2[n_vec].val[1]);
+                }
+            for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                {
+                    vst2_s16((int16_t*)dotProductVector, accumulator1[n_vec]); // Store the results back into the dot product vector
+                    dotProduct = lv_cmake(0,0);
+                    for (int i = 0; i < 4; ++i)
+                        {
+                            dotProduct = lv_cmake(sat_adds16i(lv_creal(dotProduct), lv_creal(dotProductVector[i])),
+                                    sat_adds16i(lv_cimag(dotProduct), lv_cimag(dotProductVector[i])));
+                        }
+                    _out[n_vec] = dotProduct;
+                }
+            volk_gnsssdr_free(accumulator1);
+            volk_gnsssdr_free(accumulator2);
 
             vst1q_f32((float32_t*)__phase_real, _phase_real);
             vst1q_f32((float32_t*)__phase_imag, _phase_imag);

src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/kernels/volk_gnsssdr/volk_gnsssdr_s32f_sincos_32fc.h

@@ -42,15 +42,17 @@
  *
  * <b>Dispatcher Prototype</b>
  * \code
- * void volk_gnsssdr_s32f_sincos_32fc(lv_32fc_t* out, const float phase_inc, unsigned int num_points)
+ * void volk_gnsssdr_s32f_sincos_32fc(lv_32fc_t* out, const float phase_inc, float* phase, unsigned int num_points)
  * \endcode
  *
  * \b Inputs
  * \li phase_inc:      Phase increment per sample, in radians.
+ * \li phase:          Pointer to a float containing the initial phase, in radians.
  * \li num_points:     Number of components in \p in to be computed.
  *
  * \b Outputs
  * \li out:            Vector of the form lv_32fc_t out[n] = lv_cmake(cos(in[n]), sin(in[n]))
+ * \li phase:          Pointer to a float containing the final phase, in radians.
  *
  */
 
@@ -67,13 +69,13 @@
 #include <emmintrin.h>
 /* Adapted from http://gruntthepeon.free.fr/ssemath/sse_mathfun.h, original code from Julien Pommier  */
 /* Based on algorithms from the cephes library http://www.netlib.org/cephes/   */
-static inline void volk_gnsssdr_s32f_sincos_32fc_a_sse2(lv_32fc_t* out, const float phase_inc, unsigned int num_points)
+static inline void volk_gnsssdr_s32f_sincos_32fc_a_sse2(lv_32fc_t* out, const float phase_inc, float* phase, unsigned int num_points)
 {
     lv_32fc_t* bPtr = out;
 
     const unsigned int sse_iters = num_points / 4;
     unsigned int number = 0;
-    float _phase;
+    float _phase = (*phase);
 
     __m128 sine, cosine, aux, x, four_phases_reg;
     __m128 xmm1, xmm2, xmm3 = _mm_setzero_ps(), sign_bit_sin, y;
@@ -101,7 +103,7 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_a_sse2(lv_32fc_t* out, const fl
     static const float _ps_0p5[4] __attribute__((aligned(16))) = { 0.5f, 0.5f, 0.5f, 0.5f };
     static const float _ps_1[4] __attribute__((aligned(16))) = { 1.0f, 1.0f, 1.0f, 1.0f };
 
-    float four_phases[4] __attribute__((aligned(16))) = { 0.0f, phase_inc, 2 * phase_inc, 3 * phase_inc };
+    float four_phases[4] __attribute__((aligned(16))) = { _phase, _phase + phase_inc, _phase + 2 * phase_inc, _phase + 3 * phase_inc };
     float four_phases_inc[4] __attribute__((aligned(16))) = { 4 * phase_inc, 4 * phase_inc, 4 * phase_inc, 4 * phase_inc };
     four_phases_reg = _mm_load_ps(four_phases);
     const __m128 four_phases_inc_reg = _mm_load_ps(four_phases_inc);
@@ -207,12 +209,13 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_a_sse2(lv_32fc_t* out, const fl
             four_phases_reg = _mm_add_ps(four_phases_reg, four_phases_inc_reg);
         }
 
-    _phase = phase_inc * (sse_iters * 4);
+    _phase = _phase + phase_inc * (sse_iters * 4);
     for(number = sse_iters * 4; number < num_points; number++)
         {
-            *bPtr++ = lv_cmake((float)cos(_phase), (float)sin(_phase) );
+            *bPtr++ = lv_cmake((float)cos((_phase)), (float)sin((_phase)) );
             _phase += phase_inc;
         }
+    (*phase) = _phase;
 }
 
 #endif /* LV_HAVE_SSE2  */
@@ -222,13 +225,14 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_a_sse2(lv_32fc_t* out, const fl
 #include <emmintrin.h>
 /* Adapted from http://gruntthepeon.free.fr/ssemath/sse_mathfun.h, original code from Julien Pommier  */
 /* Based on algorithms from the cephes library http://www.netlib.org/cephes/   */
-static inline void volk_gnsssdr_s32f_sincos_32fc_u_sse2(lv_32fc_t* out, const float phase_inc, unsigned int num_points)
+static inline void volk_gnsssdr_s32f_sincos_32fc_u_sse2(lv_32fc_t* out, const float phase_inc, float* phase, unsigned int num_points)
 {
     lv_32fc_t* bPtr = out;
 
     const unsigned int sse_iters = num_points / 4;
     unsigned int number = 0;
-    float _phase;
+
+    float _phase = (*phase);
 
     __m128 sine, cosine, aux, x, four_phases_reg;
     __m128 xmm1, xmm2, xmm3 = _mm_setzero_ps(), sign_bit_sin, y;
@@ -256,7 +260,7 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_u_sse2(lv_32fc_t* out, const fl
     static const float _ps_0p5[4] __attribute__((aligned(16))) = { 0.5f, 0.5f, 0.5f, 0.5f };
     static const float _ps_1[4] __attribute__((aligned(16))) = { 1.0f, 1.0f, 1.0f, 1.0f };
 
-    float four_phases[4] __attribute__((aligned(16))) = { 0.0f, phase_inc, 2 * phase_inc, 3 * phase_inc };
+    float four_phases[4] __attribute__((aligned(16))) = { _phase, _phase + phase_inc, _phase + 2 * phase_inc, _phase + 3 * phase_inc };
     float four_phases_inc[4] __attribute__((aligned(16))) = { 4 * phase_inc, 4 * phase_inc, 4 * phase_inc, 4 * phase_inc };
     four_phases_reg = _mm_load_ps(four_phases);
     const __m128 four_phases_inc_reg = _mm_load_ps(four_phases_inc);
@@ -362,26 +366,29 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_u_sse2(lv_32fc_t* out, const fl
             four_phases_reg = _mm_add_ps(four_phases_reg, four_phases_inc_reg);
         }
 
-    _phase = phase_inc * (sse_iters * 4);
+    _phase = _phase + phase_inc * (sse_iters * 4);
     for(number = sse_iters * 4; number < num_points; number++)
         {
             *bPtr++ = lv_cmake((float)cos(_phase), (float)sin(_phase) );
             _phase += phase_inc;
         }
+    (*phase) = _phase;
 }
 
 #endif /* LV_HAVE_SSE2  */
 
+
 #ifdef LV_HAVE_GENERIC
 
-static inline void volk_gnsssdr_s32f_sincos_32fc_generic(lv_32fc_t* out, const float phase_inc, unsigned int num_points)
+static inline void volk_gnsssdr_s32f_sincos_32fc_generic(lv_32fc_t* out, const float phase_inc, float* phase, unsigned int num_points)
 {
-    float _phase = 0.0;
+    float _phase = (*phase);
     for(unsigned int i = 0; i < num_points; i++)
         {
             *out++ = lv_cmake((float)cos(_phase), (float)sin(_phase) );
             _phase += phase_inc;
         }
+    (*phase) = _phase;
 }
 
 #endif /* LV_HAVE_GENERIC  */
@@ -390,7 +397,7 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_generic(lv_32fc_t* out, const f
 #ifdef LV_HAVE_GENERIC
 #include <volk_gnsssdr/volk_gnsssdr_sine_table.h>
 #include <stdint.h>
-static inline void volk_gnsssdr_s32f_sincos_32fc_generic_fxpt(lv_32fc_t* out, const float phase_inc, unsigned int num_points)
+static inline void volk_gnsssdr_s32f_sincos_32fc_generic_fxpt(lv_32fc_t* out, const float phase_inc, float* phase, unsigned int num_points)
 {
     float _in, s, c;
     int32_t x, sin_index, cos_index, d;
@@ -401,7 +408,7 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_generic_fxpt(lv_32fc_t* out, co
     const int32_t Nbits = 10;
     const int32_t diffbits = bitlength - Nbits;
     uint32_t ux;
-    float _phase = 0.0;
+    float _phase = (*phase);
     for(unsigned int i = 0; i < num_points; i++)
         {
             _in = _phase;
@@ -420,6 +427,7 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_generic_fxpt(lv_32fc_t* out, co
             *out++ = lv_cmake((float)c, (float)s );
             _phase += phase_inc;
         }
+    (*phase) = _phase;
 }
 
 #endif /* LV_HAVE_GENERIC  */
@@ -429,12 +437,13 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_generic_fxpt(lv_32fc_t* out, co
 #include <arm_neon.h>
 /* Adapted from http://gruntthepeon.free.fr/ssemath/neon_mathfun.h, original code from Julien Pommier  */
 /* Based on algorithms from the cephes library http://www.netlib.org/cephes/   */
-static inline void volk_gnsssdr_s32f_sincos_32fc_neon(lv_32fc_t* out, const float phase_inc, unsigned int num_points)
+static inline void volk_gnsssdr_s32f_sincos_32fc_neon(lv_32fc_t* out, const float phase_inc, float* phase, unsigned int num_points)
 {
     lv_32fc_t* bPtr = out;
     const unsigned int neon_iters = num_points / 4;
+    float _phase = (*phase);
 
-    __VOLK_ATTR_ALIGNED(16) float32_t four_phases[4] = { 0.0f , phase_inc, 2 * phase_inc,  3 * phase_inc };
+    __VOLK_ATTR_ALIGNED(16) float32_t four_phases[4] = {  _phase, _phase + phase_inc, _phase + 2 * phase_inc, _phase + 3 * phase_inc };
     float four_inc = 4 * phase_inc;
     __VOLK_ATTR_ALIGNED(16) float32_t four_phases_inc[4] = { four_inc, four_inc, four_inc, four_inc };
 
@@ -453,7 +462,6 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_neon(lv_32fc_t* out, const floa
     const float32_t c_cephes_FOPI = 1.27323954473516;
 
     unsigned int number = 0;
-    float _phase;
 
     float32x4_t x, xmm1, xmm2, xmm3, y, y1, y2, ys, yc, z;
     float32x4x2_t result;
@@ -529,12 +537,13 @@ static inline void volk_gnsssdr_s32f_sincos_32fc_neon(lv_32fc_t* out, const floa
             four_phases_reg = vaddq_f32(four_phases_reg, four_phases_inc_reg);
         }
 
-    _phase = phase_inc * (neon_iters * 4);
+    _phase = _phase + phase_inc * (neon_iters * 4);
     for(number = neon_iters * 4; number < num_points; number++)
         {
             *bPtr++ = lv_cmake((float)cos(_phase), (float)sin(_phase) );
             _phase += phase_inc;
         }
+    (*phase) = _phase;
 }
 
 #endif /* LV_HAVE_NEON  */

src/algorithms/libs/volk_gnsssdr_module/volk_gnsssdr/lib/kernel_tests.h

@@ -61,6 +61,8 @@ std::vector<volk_gnsssdr_test_case_t> init_test_list(volk_gnsssdr_test_params_t
     // ... or more tolerance *****  ADDED BY GNSS-SDR
     volk_gnsssdr_test_params_t test_params_int16 = volk_gnsssdr_test_params_t(16, test_params.scalar(),
                 test_params.vlen(), test_params.iter(), test_params.benchmark_mode(), test_params.kernel_regex());
+    volk_gnsssdr_test_params_t test_params_inacc2 = volk_gnsssdr_test_params_t(2e-1, test_params.scalar(),
+                test_params.vlen(), test_params.iter(), test_params.benchmark_mode(), test_params.kernel_regex());
 
     std::vector<volk_gnsssdr_test_case_t> test_cases = boost::assign::list_of
 
@@ -75,11 +77,13 @@ std::vector<volk_gnsssdr_test_case_t> init_test_list(volk_gnsssdr_test_params_t
         (VOLK_INIT_TEST(volk_gnsssdr_8ic_s8ic_multiply_8ic, test_params))
         (VOLK_INIT_TEST(volk_gnsssdr_8u_x2_multiply_8u, test_params_more_iters))
         (VOLK_INIT_TEST(volk_gnsssdr_64f_accumulator_64f, test_params))
+        (VOLK_INIT_TEST(volk_gnsssdr_32f_sincos_32fc, test_params_inacc))
         (VOLK_INIT_TEST(volk_gnsssdr_32fc_convert_8ic, test_params))
         (VOLK_INIT_TEST(volk_gnsssdr_32fc_convert_16ic, test_params_more_iters))
         (VOLK_INIT_TEST(volk_gnsssdr_16ic_x2_dot_prod_16ic, test_params))
         (VOLK_INIT_TEST(volk_gnsssdr_16ic_x2_multiply_16ic, test_params_more_iters))
         (VOLK_INIT_TEST(volk_gnsssdr_16ic_convert_32fc, test_params_more_iters))
+        (VOLK_INIT_PUPP(volk_gnsssdr_s32f_sincospuppet_32fc, volk_gnsssdr_s32f_sincos_32fc, test_params_inacc2))
         (VOLK_INIT_PUPP(volk_gnsssdr_16ic_rotatorpuppet_16ic, volk_gnsssdr_16ic_s32fc_x2_rotator_16ic, test_params_int1))
         (VOLK_INIT_PUPP(volk_gnsssdr_16ic_resamplerpuppet_16ic, volk_gnsssdr_16ic_resampler_16ic, test_params))
         (VOLK_INIT_PUPP(volk_gnsssdr_16ic_resamplerxnpuppet_16ic, volk_gnsssdr_16ic_xn_resampler_16ic_xn, test_params))