add neon implementation

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

@@ -255,4 +255,94 @@ static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_u_sse2(lv_16sc_t* out,
 }
 #endif /* LV_HAVE_SSE2 */
 
+#ifdef LV_HAVE_NEON
+#include <arm_neon.h>
+
+static inline void volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon(lv_16sc_t* out, 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 = out;
+
+    if (neon_iters > 0)
+        {
+            __VOLK_ATTR_ALIGNED(16) lv_16sc_t dotProductVector[4];
+
+            int16x4x2_t a_val, b_val, c_val;
+
+            //todo dyn mem reg
+            int16x4x2_t* accumulator;
+            accumulator = (int16x4x2_t*)calloc(num_a_vectors, sizeof(int16x4x2_t));
+
+            int16x4x2_t tmp_real, tmp_imag;
+            lv_16sc_t accum_result[4];
+
+            for(int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                {
+                    accumulator[n_vec].val[0] = vdup_n_s16(0);
+                    accumulator[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])); //load (2 byte imag, 2 byte real) x 4 into 128 bits reg
+                            //__builtin_prefetch(_in_a[n_vec] + 8);
+
+                            // multiply the real*real and imag*imag to get real result
+                            // a0r*b0r|a1r*b1r|a2r*b2r|a3r*b3r
+                            tmp_real.val[0] = vmul_s16(a_val.val[0], b_val.val[0]);
+                            // a0i*b0i|a1i*b1i|a2i*b2i|a3i*b3i
+                            tmp_real.val[1] = vmul_s16(a_val.val[1], b_val.val[1]);
+
+                            // Multiply cross terms to get the imaginary result
+                            // a0r*b0i|a1r*b1i|a2r*b2i|a3r*b3i
+                            tmp_imag.val[0] = vmul_s16(a_val.val[0], b_val.val[1]);
+                            // a0i*b0r|a1i*b1r|a2i*b2r|a3i*b3r
+                            tmp_imag.val[1] = vmul_s16(a_val.val[1], b_val.val[0]);
+
+                            c_val.val[0] = vsub_s16(tmp_real.val[0], tmp_real.val[1]);
+                            c_val.val[1] = vadd_s16(tmp_imag.val[0], tmp_imag.val[1]);
+
+                            accumulator[n_vec].val[0] = vadd_s16(accumulator[n_vec].val[0], c_val.val[0]);
+                            accumulator[n_vec].val[1] = vadd_s16(accumulator[n_vec].val[1], c_val.val[1]);
+                        }
+                    _in_common += 4;
+                }
+
+            for (int n_vec = 0; n_vec < num_a_vectors; n_vec++)
+                {
+                    vst2_s16((int16_t*)dotProductVector, accumulator[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;
+                }
+            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 */
+
 #endif /*INCLUDED_volk_gnsssdr_16ic_xn_dot_prod_16ic_xn_H*/

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

@@ -105,6 +105,28 @@ static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_u_sse2(lv_16sc_t* r
 
 #endif // SSE2
 
+#ifdef LV_HAVE_NEON
+
+static inline void volk_gnsssdr_16ic_x2_dotprodxnpuppet_16ic_neon(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(in_a[n], in, sizeof(lv_16sc_t)*num_points);
+    }
+    volk_gnsssdr_16ic_x2_dot_prod_16ic_xn_neon(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