Added DL CSI decoding to TM2 and TM3

lib/include/srslte/phy/mimo/precoding.h

@@ -87,7 +87,7 @@ SRSLTE_API int srslte_predecoding_single(cf_t *y,
 SRSLTE_API int srslte_predecoding_single_multi(cf_t *y[SRSLTE_MAX_PORTS], 
                                                cf_t *h[SRSLTE_MAX_PORTS], 
                                                cf_t *x, 
-                                               float *csi,
+                                               float *csi[SRSLTE_MAX_CODEWORDS],
                                                int nof_rxant,
                                                int nof_symbols, 
                                                float scaling,
@@ -102,7 +102,8 @@ SRSLTE_API int srslte_predecoding_diversity(cf_t *y,
 
 SRSLTE_API int srslte_predecoding_diversity_multi(cf_t *y[SRSLTE_MAX_PORTS], 
                                                   cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS], 
-                                                  cf_t *x[SRSLTE_MAX_LAYERS],    
+                                                  cf_t *x[SRSLTE_MAX_LAYERS],
+                                                  float *csi[SRSLTE_MAX_LAYERS],
                                                   int nof_rxant,
                                                   int nof_ports, 
                                                   int nof_symbols,
@@ -113,7 +114,7 @@ SRSLTE_API void srslte_predecoding_set_mimo_decoder (srslte_mimo_decoder_t _mimo
 SRSLTE_API int srslte_predecoding_type(cf_t *y[SRSLTE_MAX_PORTS],
                                        cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
                                        cf_t *x[SRSLTE_MAX_LAYERS],
-                                       float *csi,
+                                       float *csi[SRSLTE_MAX_CODEWORDS],
                                        int nof_rxant,
                                        int nof_ports,
                                        int nof_layers,

lib/include/srslte/phy/utils/mat.h

@@ -53,12 +53,17 @@ SRSLTE_API void srslte_mat_2x2_mmse_gen(cf_t y0, cf_t y1,
                                         float noise_estimate,
                                         float norm);
 
+SRSLTE_API void srslte_mat_2x2_mmse_csi_gen(cf_t y0, cf_t y1,
+                                            cf_t h00, cf_t h01, cf_t h10, cf_t h11,
+                                            cf_t *x0, cf_t *x1, float *csi0, float *csi1,
+                                            float noise_estimate,
+                                            float norm);
+
 SRSLTE_API float srslte_mat_2x2_cn(cf_t h00,
                                    cf_t h01,
                                    cf_t h10,
                                    cf_t h11);
 
-
 #ifdef LV_HAVE_SSE
 
 /* SSE implementation for complex reciprocal */
@@ -103,4 +108,114 @@ SRSLTE_API void srslte_mat_2x2_mmse_avx(__m256 y0, __m256 y1,
 
 #endif /* LV_HAVE_AVX */
 
-#endif // SRSLTE_MAT_H 
+#if SRSLTE_SIMD_CF_SIZE != 0
+
+/* Generic SIMD implementation for 2x2 determinant */
+static inline simd_cf_t srslte_mat_2x2_det_simd(simd_cf_t a00, simd_cf_t a01, simd_cf_t a10, simd_cf_t a11) {
+  return srslte_simd_cf_sub(srslte_simd_cf_prod(a00, a11), srslte_simd_cf_prod(a01, a10));
+}
+
+/* Generic SIMD implementation for Zero Forcing (ZF) solver */
+static inline void srslte_mat_2x2_zf_csi_simd(simd_cf_t y0,
+                                              simd_cf_t y1,
+                                              simd_cf_t h00,
+                                              simd_cf_t h01,
+                                              simd_cf_t h10,
+                                              simd_cf_t h11,
+                                              simd_cf_t *x0,
+                                              simd_cf_t *x1,
+                                              simd_f_t *csi0,
+                                              simd_f_t *csi1,
+                                              float norm) {
+  simd_cf_t det = srslte_mat_2x2_det_simd(h00, h01, h10, h11);
+  simd_cf_t detrec = srslte_simd_cf_mul(srslte_simd_cf_rcp(det), srslte_simd_f_set1(norm));
+
+  *x0 = srslte_simd_cf_prod(srslte_simd_cf_sub(srslte_simd_cf_prod(h11, y0), srslte_simd_cf_prod(h01, y1)), detrec);
+  *x1 = srslte_simd_cf_prod(srslte_simd_cf_sub(srslte_simd_cf_prod(h00, y1), srslte_simd_cf_prod(h10, y0)), detrec);
+
+  *csi0 = srslte_simd_f_set1(1.0f);
+  *csi1 = srslte_simd_f_set1(1.0f);
+}
+
+static inline void srslte_mat_2x2_zf_simd(simd_cf_t y0,
+                                          simd_cf_t y1,
+                                          simd_cf_t h00,
+                                          simd_cf_t h01,
+                                          simd_cf_t h10,
+                                          simd_cf_t h11,
+                                          simd_cf_t *x0,
+                                          simd_cf_t *x1,
+                                          float norm) {
+  simd_f_t csi1, csi2;
+  srslte_mat_2x2_zf_csi_simd(y0, y1, h00, h01, h10, h11, x0, x1, &csi1, &csi2, norm);
+}
+
+/* Generic SIMD implementation for Minimum Mean Squared Error (MMSE) solver */
+static inline void srslte_mat_2x2_mmse_csi_simd(simd_cf_t y0,
+                                                simd_cf_t y1,
+                                                simd_cf_t h00,
+                                                simd_cf_t h01,
+                                                simd_cf_t h10,
+                                                simd_cf_t h11,
+                                                simd_cf_t *x0,
+                                                simd_cf_t *x1,
+                                                simd_f_t *csi0,
+                                                simd_f_t *csi1,
+                                                float noise_estimate,
+                                                float norm) {
+  simd_cf_t _noise_estimate;
+  simd_f_t _norm = srslte_simd_f_set1(norm);
+
+  _noise_estimate.re = srslte_simd_f_set1(noise_estimate);
+  _noise_estimate.im = srslte_simd_f_zero();
+
+  /* 1. A = H' x H + No*/
+  simd_cf_t a00 =
+      srslte_simd_cf_add(srslte_simd_cf_add(srslte_simd_cf_conjprod(h00, h00), srslte_simd_cf_conjprod(h10, h10)),
+                         _noise_estimate);
+  simd_cf_t a01 = srslte_simd_cf_add(srslte_simd_cf_conjprod(h01, h00), srslte_simd_cf_conjprod(h11, h10));
+  simd_cf_t a10 = srslte_simd_cf_add(srslte_simd_cf_conjprod(h00, h01), srslte_simd_cf_conjprod(h10, h11));
+  simd_cf_t a11 =
+      srslte_simd_cf_add(srslte_simd_cf_add(srslte_simd_cf_conjprod(h01, h01), srslte_simd_cf_conjprod(h11, h11)),
+                         _noise_estimate);
+  simd_cf_t a_det_rcp = srslte_simd_cf_rcp(srslte_mat_2x2_det_simd(a00, a01, a10, a11));
+
+  /* 2. B = inv(H' x H + No) = inv(A) */
+  simd_cf_t _norm2 = srslte_simd_cf_mul(a_det_rcp, _norm);
+  simd_cf_t b00 = srslte_simd_cf_prod(a11, _norm2);
+  simd_cf_t b01 = srslte_simd_cf_prod(srslte_simd_cf_neg(a01), _norm2);
+  simd_cf_t b10 = srslte_simd_cf_prod(srslte_simd_cf_neg(a10), _norm2);
+  simd_cf_t b11 = srslte_simd_cf_prod(a00, _norm2);
+
+
+  /* 3. W = inv(H' x H + No) x H' = B x H' */
+  simd_cf_t w00 = srslte_simd_cf_add(srslte_simd_cf_conjprod(b00, h00), srslte_simd_cf_conjprod(b01, h01));
+  simd_cf_t w01 = srslte_simd_cf_add(srslte_simd_cf_conjprod(b00, h10), srslte_simd_cf_conjprod(b01, h11));
+  simd_cf_t w10 = srslte_simd_cf_add(srslte_simd_cf_conjprod(b10, h00), srslte_simd_cf_conjprod(b11, h01));
+  simd_cf_t w11 = srslte_simd_cf_add(srslte_simd_cf_conjprod(b10, h10), srslte_simd_cf_conjprod(b11, h11));
+
+  /* 4. X = W x Y */
+  *x0 = srslte_simd_cf_add(srslte_simd_cf_prod(y0, w00), srslte_simd_cf_prod(y1, w01));
+  *x1 = srslte_simd_cf_add(srslte_simd_cf_prod(y0, w10), srslte_simd_cf_prod(y1, w11));
+
+  /* 5. Extract CSI */
+  *csi0 = srslte_simd_f_rcp(srslte_simd_cf_re(b00));
+  *csi1 = srslte_simd_f_rcp(srslte_simd_cf_re(b11));
+}
+
+static inline void srslte_mat_2x2_mmse_simd(simd_cf_t y0,
+                                            simd_cf_t y1,
+                                            simd_cf_t h00,
+                                            simd_cf_t h01,
+                                            simd_cf_t h10,
+                                            simd_cf_t h11,
+                                            simd_cf_t *x0,
+                                            simd_cf_t *x1,
+                                            float noise_estimate,
+                                            float norm) {
+  simd_f_t csi0, csi1;
+  srslte_mat_2x2_mmse_csi_simd(y0, y1, h00, h01, h10, h11, x0, x1, &csi0, &csi1, noise_estimate, norm);
+}
+
+#endif /* SRSLTE_SIMD_CF_SIZE != 0 */
+#endif /* SRSLTE_MAT_H */

lib/include/srslte/phy/utils/simd.h

@@ -197,7 +197,7 @@ static inline simd_f_t srslte_simd_f_loadu(const float *ptr) {
 #ifdef LV_HAVE_AVX512
   return _mm512_loadu_ps(ptr);
 #else /* LV_HAVE_AVX512 */
- #ifdef LV_HAVE_AVX2
+#ifdef LV_HAVE_AVX2
   return _mm256_loadu_ps(ptr);
 #else /* LV_HAVE_AVX2 */
 #ifdef LV_HAVE_SSE
@@ -233,7 +233,7 @@ static inline void srslte_simd_f_storeu(float *ptr, simd_f_t simdreg) {
 #ifdef LV_HAVE_AVX512
   _mm512_storeu_ps(ptr, simdreg);
 #else /* LV_HAVE_AVX512 */
-  #ifdef LV_HAVE_AVX2
+#ifdef LV_HAVE_AVX2
   _mm256_storeu_ps(ptr, simdreg);
 #else /* LV_HAVE_AVX2 */
 #ifdef LV_HAVE_SSE
@@ -360,7 +360,7 @@ static inline simd_f_t srslte_simd_f_add(simd_f_t a, simd_f_t b) {
 #ifdef LV_HAVE_AVX2
   return _mm256_add_ps(a, b);
 #else /* LV_HAVE_AVX2 */
-#ifdef LV_HAVE_SSE 
+#ifdef LV_HAVE_SSE
   return _mm_add_ps(a, b);
 #else /* LV_HAVE_SSE */
 #ifdef HAVE_NEON
@@ -376,9 +376,9 @@ static inline simd_f_t srslte_simd_f_zero (void) {
   return _mm512_setzero_ps();
 #else /* LV_HAVE_AVX512 */
 #ifdef LV_HAVE_AVX2
- return _mm256_setzero_ps();
+  return _mm256_setzero_ps();
 #else /* LV_HAVE_AVX2 */
-#ifdef LV_HAVE_SSE 
+#ifdef LV_HAVE_SSE
  return _mm_setzero_ps();
 #else /* LV_HAVE_SSE */
 #ifdef HAVE_NEON
@@ -401,7 +401,7 @@ static inline simd_f_t srslte_simd_f_swap(simd_f_t a) {
 #else /* LV_HAVE_SSE */
 #ifdef HAVE_NEON
   return vcombine_f32(vrev64_f32(vget_low_f32(a)), vrev64_f32(vget_high_f32(a)));
-#endif /* HAVE_NEON */  
+#endif /* HAVE_NEON */
 #endif /* LV_HAVE_SSE */
 #endif /* LV_HAVE_AVX2 */
 #endif /* LV_HAVE_AVX512 */
@@ -443,7 +443,7 @@ static inline simd_f_t srslte_simd_f_sqrt(simd_f_t a) {
 #ifdef LV_HAVE_AVX512
   return _mm512_sqrt_ps(a);
 #else /* LV_HAVE_AVX512 */
-  #ifdef LV_HAVE_AVX2
+#ifdef LV_HAVE_AVX2
   return _mm256_sqrt_ps(a);
 #else /* LV_HAVE_AVX2 */
   #ifdef LV_HAVE_SSE
@@ -458,7 +458,43 @@ static inline simd_f_t srslte_simd_f_sqrt(simd_f_t a) {
   float32x4_t zeros = vmovq_n_f32(0);    /* Zero vector */
   uint32x4_t mask = vceqq_f32(a, zeros); /* Zero vector mask */
   return vbslq_f32(mask, zeros, result); /* Force zero results and return */
-#endif /* HAVE_NEON */  
+#endif /* HAVE_NEON */
+#endif /* LV_HAVE_SSE */
+#endif /* LV_HAVE_AVX2 */
+#endif /* LV_HAVE_AVX512 */
+}
+
+static inline simd_f_t srslte_simd_f_neg(simd_f_t a) {
+#ifdef LV_HAVE_AVX512
+  return _mm512_xor_ps(_mm512_set1_ps(-0.0f), a);
+#else /* LV_HAVE_AVX512 */
+#ifdef LV_HAVE_AVX2
+  return _mm256_xor_ps(_mm256_set1_ps(-0.0f), a);
+#else /* LV_HAVE_AVX2 */
+  #ifdef LV_HAVE_SSE
+  return _mm_xor_ps(_mm_set1_ps(-0.0f), a);
+#else /* LV_HAVE_SSE */
+#ifdef HAVE_NEON
+  return vnegq_f32(a);
+#endif /* HAVE_NEON */
+#endif /* LV_HAVE_SSE */
+#endif /* LV_HAVE_AVX2 */
+#endif /* LV_HAVE_AVX512 */
+}
+
+static inline simd_f_t srslte_simd_f_neg_mask(simd_f_t a, simd_f_t mask) {
+#ifdef LV_HAVE_AVX512
+  return _mm512_xor_ps(mask, a);
+#else /* LV_HAVE_AVX512 */
+#ifdef LV_HAVE_AVX2
+  return _mm256_xor_ps(mask, a);
+#else /* LV_HAVE_AVX2 */
+  #ifdef LV_HAVE_SSE
+  return _mm_xor_ps(mask, a);
+#else /* LV_HAVE_SSE */
+#ifdef HAVE_NEON
+  return (float32x4_t) veorq_s32((int32x4_t) a, (int32x4_t) mask);
+#endif /* HAVE_NEON */
 #endif /* LV_HAVE_SSE */
 #endif /* LV_HAVE_AVX2 */
 #endif /* LV_HAVE_AVX512 */
@@ -470,12 +506,11 @@ static inline simd_f_t srslte_simd_f_sqrt(simd_f_t a) {
 #if SRSLTE_SIMD_CF_SIZE
 
 #ifdef HAVE_NEON
-   typedef float32x4x2_t simd_cf_t;
+typedef float32x4x2_t simd_cf_t;
 #else
 typedef struct {
   simd_f_t re;
   simd_f_t im;
-  
 } simd_cf_t;
 #endif
 
@@ -667,8 +702,8 @@ static inline void srslte_simd_cf_store(float *re, float *im, simd_cf_t simdreg)
   _mm512_store_ps(im, simdreg.im);
 #else /* LV_HAVE_AVX512 */
 #ifdef LV_HAVE_AVX2
-  _mm256_store_ps((float *) re, simdreg.re);
+  _mm256_store_ps(re, simdreg.re);
-  _mm256_store_ps((float *) im, simdreg.im);
+  _mm256_store_ps(im, simdreg.im);
 #else /* LV_HAVE_AVX512 */
 #ifdef LV_HAVE_SSE
   _mm_store_ps((float *) re, simdreg.re);
@@ -689,8 +724,8 @@ static inline void srslte_simd_cf_storeu(float *re, float *im, simd_cf_t simdreg
   _mm512_storeu_ps(im, simdreg.im);
 #else /* LV_HAVE_AVX512 */
 #ifdef LV_HAVE_AVX2
-  _mm256_storeu_ps((float *) re, simdreg.re);
+  _mm256_storeu_ps(re, simdreg.re);
-  _mm256_storeu_ps((float *) im, simdreg.im);
+  _mm256_storeu_ps(im, simdreg.im);
 #else /* LV_HAVE_AVX512 */
 #ifdef LV_HAVE_SSE
   _mm_storeu_ps((float *) re, simdreg.re);
@@ -833,8 +868,32 @@ static inline simd_cf_t srslte_simd_cf_add (simd_cf_t a, simd_cf_t b) {
   return ret;
 }
 
+static inline simd_cf_t srslte_simd_cf_sub (simd_cf_t a, simd_cf_t b) {
+  simd_cf_t ret;
+#ifdef LV_HAVE_AVX512
+  ret.re = _mm512_sub_ps(a.re, b.re);
+  ret.im = _mm512_sub_ps(a.im, b.im);
+#else /* LV_HAVE_AVX512 */
+#ifdef LV_HAVE_AVX2
+  ret.re = _mm256_sub_ps(a.re, b.re);
+  ret.im = _mm256_sub_ps(a.im, b.im);
+#else /* LV_HAVE_AVX2 */
+  #ifdef LV_HAVE_SSE
+  ret.re = _mm_sub_ps(a.re, b.re);
+  ret.im = _mm_sub_ps(a.im, b.im);
+#else /* LV_HAVE_SSE */
+#ifdef HAVE_NEON
+  ret.val[0] = vsubq_f32(a.val[0],b.val[0]);
+  ret.val[1] = vsubq_f32(a.val[1],b.val[1]);
+#endif /* HAVE_NEON */
+#endif /* LV_HAVE_SSE */
+#endif /* LV_HAVE_AVX2 */
+#endif /* LV_HAVE_AVX512 */
+  return ret;
+}
+
 static inline simd_cf_t srslte_simd_cf_mul (simd_cf_t a, simd_f_t b) {
-    simd_cf_t ret;
+  simd_cf_t ret;
 #ifdef LV_HAVE_AVX512
   ret.re = _mm512_mul_ps(a.re, b);
   ret.im = _mm512_mul_ps(a.im, b);
@@ -855,7 +914,7 @@ static inline simd_cf_t srslte_simd_cf_mul (simd_cf_t a, simd_f_t b) {
 #endif /* LV_HAVE_SSE */
 #endif /* LV_HAVE_AVX2 */
 #endif /* LV_HAVE_AVX512 */
-    return ret;
+  return ret;
 }
 
 static inline simd_cf_t srslte_simd_cf_rcp (simd_cf_t a) {
@@ -902,6 +961,59 @@ static inline simd_cf_t srslte_simd_cf_rcp (simd_cf_t a) {
   return ret;
 }
 
+static inline simd_cf_t srslte_simd_cf_neg (simd_cf_t a) {
+  simd_cf_t ret;
+#if LV_HAVE_NEON
+  ret.val[0] = srslte_simd_f_neg(a.val[0]);
+  ret.val[1] = srslte_simd_f_neg(a.val[1]);
+#else  /* LV_HAVE_NEON */
+  ret.re = srslte_simd_f_neg(a.re);
+  ret.im = srslte_simd_f_neg(a.im);
+#endif /* LV_HAVE_NEON */
+  return ret;
+}
+
+static inline simd_cf_t srslte_simd_cf_neg_mask (simd_cf_t a, simd_f_t mask) {
+  simd_cf_t ret;
+#ifndef LV_HAVE_AVX512
+#ifdef LV_HAVE_AVX2
+  mask = _mm256_permutevar8x32_ps(mask, _mm256_setr_epi32(0,4,1,5,2,6,3,7));
+#endif /* LV_HAVE_AVX2 */
+#endif /* LV_HAVE_AVX512 */
+#if LV_HAVE_NEON
+  ret.val[0] = srslte_simd_f_neg_mask(a.val[0], mask);
+  ret.val[1] = srslte_simd_f_neg_mask(a.val[1], mask);
+#else /* LV_HAVE_NEON */
+  ret.re = srslte_simd_f_neg_mask(a.re, mask);
+  ret.im = srslte_simd_f_neg_mask(a.im, mask);
+#endif /* LV_HAVE_NEON */
+  return ret;
+}
+
+static inline simd_cf_t srslte_simd_cf_conj (simd_cf_t a) {
+  simd_cf_t ret;
+#if LV_HAVE_NEON
+  ret.val[0] = a.val[0];
+  ret.val[1] = srslte_simd_f_neg(a.val[1]);
+#else /* LV_HAVE_NEON */
+  ret.re = a.re;
+  ret.im = srslte_simd_f_neg(a.im);
+#endif /* LV_HAVE_NEON */
+  return ret;
+}
+
+static inline simd_cf_t srslte_simd_cf_mulj (simd_cf_t a) {
+  simd_cf_t ret;
+#if LV_HAVE_NEON
+  ret.val[0] = srslte_simd_f_neg(a.val[1]);
+  ret.val[1] = a.val[0];
+#else /* LV_HAVE_NEON */
+  ret.re = srslte_simd_f_neg(a.im);
+  ret.im = a.re;
+#endif /* LV_HAVE_NEON */
+  return ret;
+}
+
 static inline simd_cf_t srslte_simd_cf_zero (void) {
   simd_cf_t ret;
 #ifdef LV_HAVE_AVX512
@@ -1057,7 +1169,7 @@ static inline simd_i_t srslte_simd_i_select(simd_i_t a, simd_i_t b, simd_sel_t s
  int* sel = (int*) &selector;
  int* c_ptr = (int*) &ret;
  for(int i = 0;i<4;i++)
- { 
+ {
    if(sel[i] == -1){
      c_ptr[i] = b_ptr[i];
    }else{
@@ -1115,7 +1227,7 @@ static inline simd_s_t srslte_simd_s_loadu(const int16_t *ptr) {
 #ifdef LV_HAVE_AVX512
   return _mm512_loadu_si512(ptr);
 #else /* LV_HAVE_AVX512 */
-  #ifdef LV_HAVE_AVX2
+#ifdef LV_HAVE_AVX2
   return _mm256_loadu_si256((__m256i*) ptr);
 #else /* LV_HAVE_AVX2 */
 #ifdef LV_HAVE_SSE

lib/src/phy/mimo/layermap.c

@@ -54,7 +54,7 @@ int srslte_layermap_multiplex(cf_t *d[SRSLTE_MAX_CODEWORDS], cf_t *x[SRSLTE_MAX_
     int nof_symbols[SRSLTE_MAX_CODEWORDS]) {
   if (nof_cw == nof_layers) {
     for (int i = 0; i < nof_cw; i++) {
-      srs_vec_cf_cpy(x[i], d[i], (uint32_t) nof_symbols[0]);
+      srs_vec_cf_cpy(d[i], x[i], (uint32_t) nof_symbols[0]);
     }
     return nof_symbols[0];
   } else if (nof_cw == 1) {

lib/src/phy/mimo/precoding.c

@@ -283,13 +283,13 @@ int srslte_predecoding_single_csi(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_
   for (; i < nof_symbols; i++) {
     cf_t r = 0;
     float hh = 0;
-    float _scaling = 1.0f / scaling;
+    float norm = 1.0f / scaling;
     for (int p = 0; p < nof_rxant; p++) {
       r += y[p][i] * conj(h[p][i]);
       hh += (__real__ h[p][i] * __real__ h[p][i]) + (__imag__ h[p][i] * __imag__ h[p][i]);
     }
     csi[i] = hh + noise_estimate;
-    x[i] = r * _scaling / csi[i];
+    x[i] = r * norm / csi[i];
   }
   return nof_symbols;
 }
@@ -327,10 +327,10 @@ int srslte_predecoding_single(cf_t *y_, cf_t *h_, cf_t *x, float *csi, int nof_s
 }
 
 /* ZF/MMSE SISO equalizer x=y(h'h+no)^(-1)h' (ZF if n0=0.0)*/
-int srslte_predecoding_single_multi(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS], cf_t *x, float *csi,
+int srslte_predecoding_single_multi(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS], cf_t *x, float *csi[SRSLTE_MAX_CODEWORDS],
                                     int nof_rxant, int nof_symbols, float scaling, float noise_estimate) {
-  if (csi) {
+  if (csi && csi[0]) {
-    return srslte_predecoding_single_csi(y, h, x, csi, nof_rxant, nof_symbols, scaling, noise_estimate);
+    return srslte_predecoding_single_csi(y, h, x, csi[0], nof_rxant, nof_symbols, scaling, noise_estimate);
   }
 
 #ifdef LV_HAVE_AVX
@@ -566,18 +566,123 @@ int srslte_predecoding_diversity(cf_t *y_, cf_t *h_[SRSLTE_MAX_PORTS], cf_t *x[S
 #endif   
 }
 
+int srslte_predecoding_diversity_csi(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
+                                      cf_t *x[SRSLTE_MAX_LAYERS], float *csi[SRSLTE_MAX_CODEWORDS],
+                                      int nof_rxant, int nof_ports, int nof_symbols, float scaling) {
+  int i;
+  if (nof_ports == 2) {
+    cf_t h00, h01, h10, h11, r0, r1;
+
+    for (i = 0; i < nof_symbols / 2; i++) {
+      float hh = 0;
+      cf_t x0 = 0;
+      cf_t x1 = 0;
+      for (int p=0;p<nof_rxant;p++) {
+        h00 = h[0][p][2 * i];
+        h01 = h[0][p][2 * i+1];
+        h10 = h[1][p][2 * i];
+        h11 = h[1][p][2 * i+1];
+        hh += crealf(h00) * crealf(h00) + cimagf(h00) * cimagf(h00)
+            + crealf(h11) * crealf(h11) + cimagf(h11) * cimagf(h11);
+        r0 = y[p][2 * i];
+        r1 = y[p][2 * i + 1];
+        if (hh == 0) {
+          hh = 1e-4;
+        }
+        x0 += (conjf(h00) * r0 + h11 * conjf(r1));
+        x1 += (-h10 * conj(r0) + conj(h01) * r1);
+      }
+
+      csi[0][2*i + 0] = hh;
+      csi[0][2*i + 1] = hh;
+
+      hh *= scaling;
+      x[0][i] = x0 / hh * sqrt(2);
+      x[1][i] = x1 / hh * sqrt(2);
+    }
+    return i;
+  } else if (nof_ports == 4) {
+    int m_ap = (nof_symbols % 4) ? ((nof_symbols - 2) / 4) : nof_symbols / 4;
+    for (i = 0; i < m_ap; i++) {
+      cf_t x0 = 0, x1 = 0, x2 = 0, x3 = 0;
+      float a0 = 0, a1 = 0, a2 = 0, a3 = 0;
+      cf_t r0, r1, r2, r3;
+      cf_t h00, h01, h10, h11;
+
+      for (int p=0;p<nof_rxant;p++) {
+        h00 = h[0][p][4 * i + 0];
+        h01 = h[2][p][4 * i + 0];
+        h10 = h[0][p][4 * i + 1];
+        h11 = h[2][p][4 * i + 1];
+
+        a0 += __real__ h00 * __real__ h00 + __imag__ h00 * __imag__ h00
+            + __real__ h11 * __real__ h11 + __imag__ h11 * __imag__ h11;
+
+        a1 += __real__ h10 * __real__ h10 + __imag__ h10 * __imag__ h10
+            + __real__ h01 * __real__ h01 + __imag__ h01 * __imag__ h01;
+
+        r0 = y[p][4 * i];
+        r1 = y[p][4 * i + 1];
+
+        x0 += (conjf(h00) * r0 + h11 * conjf(r1));
+        x1 += (-h01 * conjf(r0) + conjf(h10) * r1);
+
+        h00 = h[1][p][4 * i + 2];
+        h01 = h[3][p][4 * i + 2];
+        h10 = h[1][p][4 * i + 3];
+        h11 = h[3][p][4 * i + 3];
+
+        a2 += __real__ h00 * __real__ h00 + __imag__ h00 * __imag__ h00
+            + __real__ h11 * __real__ h11 + __imag__ h11 * __imag__ h11;
+
+        a3 += __real__ h10 * __real__ h10 + __imag__ h10 * __imag__ h10
+            + __real__ h01 * __real__ h01 + __imag__ h01 * __imag__ h01;
+
+        r2 = y[p][4 * i + 2];
+        r3 = y[p][4 * i + 3];
+
+        x2 += (conjf(h00) * r2 + h11 * conjf(r3));
+        x3 += (-h01 * conjf(r2) + conjf(h10) * r3);
+      }
+
+      a0 *= scaling;
+      a1 *= scaling;
+      a2 *= scaling;
+      a3 *= scaling;
+
+      csi[0][4 * i + 0] = a0 / nof_rxant;
+      csi[0][4 * i + 1] = a1 / nof_rxant;
+      csi[0][4 * i + 2] = a2 / nof_rxant;
+      csi[0][4 * i + 3] = a3 / nof_rxant;
+
+      x[0][i] = x0 / a0 * sqrtf(2.0f);
+      x[1][i] = x1 / a1 * sqrtf(2.0f);
+      x[2][i] = x2 / a2 * sqrtf(2.0f);
+      x[3][i] = x3 / a3 * sqrtf(2.0f);
+    }
+    return i;
+  } else {
+    fprintf(stderr, "Number of ports must be 2 or 4 for transmit diversity (nof_ports=%d)\n", nof_ports);
+    return -1;
+  }
+}
+
 int srslte_predecoding_diversity_multi(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS], cf_t *x[SRSLTE_MAX_LAYERS], 
-                          int nof_rxant, int nof_ports, int nof_symbols, float scaling)
+                          float *csi[SRSLTE_MAX_CODEWORDS], int nof_rxant, int nof_ports, int nof_symbols, float scaling)
 {
-#ifdef LV_HAVE_SSE
+  if (csi && csi[0]) {
-  if (nof_symbols > 32 && nof_ports == 2) {
+    return srslte_predecoding_diversity_csi(y, h, x, csi, nof_rxant, nof_ports, nof_symbols, scaling);
-    return srslte_predecoding_diversity2_sse(y, h, x, nof_rxant, nof_symbols, scaling);
   } else {
+#ifdef LV_HAVE_SSE
+    if (nof_symbols > 32 && nof_ports == 2) {
+      return srslte_predecoding_diversity2_sse(y, h, x, nof_rxant, nof_symbols, scaling);
+    } else {
+      return srslte_predecoding_diversity_gen(y, h, x, nof_rxant, nof_ports, nof_symbols, scaling);
+    }
+#else
     return srslte_predecoding_diversity_gen(y, h, x, nof_rxant, nof_ports, nof_symbols, scaling);
+#endif
   }
-#else
-  return srslte_predecoding_diversity_gen(y, h, x, nof_rxant, nof_ports, nof_symbols, scaling);
-#endif   
 }
 
 int srslte_precoding_mimo_2x2_gen(cf_t W[2][2], cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS], cf_t *x[SRSLTE_MAX_LAYERS], 
@@ -614,102 +719,164 @@ int srslte_precoding_mimo_2x2_gen(cf_t W[2][2], cf_t *y[SRSLTE_MAX_PORTS], cf_t
   return SRSLTE_SUCCESS; 
 }
 
-// AVX implementation of ZF 2x2 CCD equalizer
+static int srslte_predecoding_ccd_2x2_zf_csi(cf_t *y[SRSLTE_MAX_PORTS],
-#ifdef LV_HAVE_AVX
+                                             cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
-
+                                             cf_t *x[SRSLTE_MAX_LAYERS],
-int srslte_predecoding_ccd_2x2_zf_avx(cf_t *y[SRSLTE_MAX_PORTS],
+                                             float *csi[SRSLTE_MAX_CODEWORDS],
-                                      cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
+                                             int nof_symbols,
-                                      cf_t *x[SRSLTE_MAX_LAYERS],
+                                             float scaling) {
-                                      uint32_t nof_symbols,
-                                      float scaling) {
   uint32_t i = 0;
+  float norm = 2.0f / scaling;
+
+#if SRSLTE_SIMD_CF_SIZE != 0
+
+#if SRSLTE_SIMD_CF_SIZE == 16
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f,
+                                       +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f
+                                       -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f};
+#elif SRSLTE_SIMD_CF_SIZE == 8
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f};
+#elif SRSLTE_SIMD_CF_SIZE == 4
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, +0.0f, -0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, -0.0f, +0.0f, +0.0f};
+#endif
 
-  __m256 mask0 = _mm256_setr_ps(+0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f);
+  simd_f_t mask1 = srslte_simd_f_loadu(_mask1);
-  __m256 mask1 = _mm256_setr_ps(-0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f);
+  simd_f_t mask2 = srslte_simd_f_loadu(_mask2);
 
-  for (i = 0; i < nof_symbols - 3; i += 4) {
+  for (; i < nof_symbols - SRSLTE_SIMD_CF_SIZE + 1; i += SRSLTE_SIMD_CF_SIZE) {
     /* Load channel */
-    __m256 h00i = _mm256_load_ps((float *) &h[0][0][i]);
+    simd_cf_t h00i = srslte_simd_cfi_load(&h[0][0][i]);
-    __m256 h01i = _mm256_load_ps((float *) &h[0][1][i]);
+    simd_cf_t h01i = srslte_simd_cfi_load(&h[0][1][i]);
-    __m256 h10i = _mm256_load_ps((float *) &h[1][0][i]);
+    simd_cf_t h10i = srslte_simd_cfi_load(&h[1][0][i]);
-    __m256 h11i = _mm256_load_ps((float *) &h[1][1][i]);
+    simd_cf_t h11i = srslte_simd_cfi_load(&h[1][1][i]);
 
     /* Apply precoding */
-    __m256 h00 = _mm256_add_ps(h00i, _mm256_xor_ps(h10i, mask0));
+    simd_cf_t h00, h01, h10, h11;
-    __m256 h10 = _mm256_add_ps(h01i, _mm256_xor_ps(h11i, mask0));
+    h00 = srslte_simd_cf_add(h00i, srslte_simd_cf_neg_mask(h10i, mask1));
-    __m256 h01 = _mm256_add_ps(h00i, _mm256_xor_ps(h10i, mask1));
+    h10 = srslte_simd_cf_add(h01i, srslte_simd_cf_neg_mask(h11i, mask1));
-    __m256 h11 = _mm256_add_ps(h01i, _mm256_xor_ps(h11i, mask1));
+    h01 = srslte_simd_cf_add(h00i, srslte_simd_cf_neg_mask(h10i, mask2));
+    h11 = srslte_simd_cf_add(h01i, srslte_simd_cf_neg_mask(h11i, mask2));
 
-    __m256 y0 = _mm256_load_ps((float *) &y[0][i]);
+    simd_cf_t y0 = srslte_simd_cfi_load(&y[0][i]);
-    __m256 y1 = _mm256_load_ps((float *) &y[1][i]);
+    simd_cf_t y1 = srslte_simd_cfi_load(&y[1][i]);
 
-    __m256 x0, x1;
+    simd_cf_t x0, x1;
+    simd_f_t csi0, csi1;
 
-    srslte_mat_2x2_zf_avx(y0, y1, h00, h01, h10, h11, &x0, &x1, 2.0f / scaling);
+    srslte_mat_2x2_zf_csi_simd(y0, y1, h00, h01, h10, h11, &x0, &x1, &csi0, &csi1, norm);
 
-    _mm256_store_ps((float *) &x[0][i], x0);
+    srslte_simd_cfi_store(&x[0][i], x0);
-    _mm256_store_ps((float *) &x[1][i], x1);
+    srslte_simd_cfi_store(&x[1][i], x1);
+
+    srslte_simd_f_store(&csi[0][i], csi0);
+    srslte_simd_f_store(&csi[1][i], csi1);
   }
+#endif /* SRSLTE_SIMD_CF_SIZE != 0 */
 
-  return nof_symbols;
+  cf_t h00, h01, h10, h11, det;
-}
+  for (; i < nof_symbols; i++) {
-#endif /* LV_HAVE_AVX */
 
-// SSE implementation of ZF 2x2 CCD equalizer
+    // Even precoder
-#ifdef LV_HAVE_SSE
+    h00 = +h[0][0][i] + h[1][0][i];
+    h10 = +h[0][1][i] + h[1][1][i];
+    h01 = +h[0][0][i] - h[1][0][i];
+    h11 = +h[0][1][i] - h[1][1][i];
+    det = (h00 * h11 - h01 * h10);
+    det = conjf(det) * (norm / (crealf(det) * crealf(det) + cimagf(det) * cimagf(det)));
+
+    x[0][i] = (+h11 * y[0][i] - h01 * y[1][i]) * det;
+    x[1][i] = (-h10 * y[0][i] + h00 * y[1][i]) * det;
+
+    csi[0][i] = 1.0f;
+    csi[1][i] = 1.0f;
+
+    i++;
+
+    // Odd precoder
+    h00 = h[0][0][i] - h[1][0][i];
+    h10 = h[0][1][i] - h[1][1][i];
+    h01 = h[0][0][i] + h[1][0][i];
+    h11 = h[0][1][i] + h[1][1][i];
+    det = (h00 * h11 - h01 * h10);
+    det = conjf(det) * (norm / (crealf(det) * crealf(det) + cimagf(det) * cimagf(det)));
+
+    x[0][i] = (+h11 * y[0][i] - h01 * y[1][i]) * det;
+    x[1][i] = (-h10 * y[0][i] + h00 * y[1][i]) * det;
 
-int srslte_predecoding_ccd_2x2_zf_sse(cf_t *y[SRSLTE_MAX_PORTS],
+    csi[0][i] = 1.0f;
-                                      cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
+    csi[1][i] = 1.0f;
-                                      cf_t *x[SRSLTE_MAX_LAYERS],
+  }
-                                      uint32_t nof_symbols,
+  return SRSLTE_SUCCESS;
-                                      float scaling) {
+}
+
+static int srslte_predecoding_ccd_2x2_zf(cf_t *y[SRSLTE_MAX_PORTS],
+                                         cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
+                                         cf_t *x[SRSLTE_MAX_LAYERS],
+                                         int nof_symbols,
+                                         float scaling) {
   uint32_t i = 0;
+  float norm = 2.0f / scaling;
+
+#if SRSLTE_SIMD_CF_SIZE != 0
+
+#if SRSLTE_SIMD_CF_SIZE == 16
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f,
+                                       +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f
+                                       -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f};
+#elif SRSLTE_SIMD_CF_SIZE == 8
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f};
+#elif SRSLTE_SIMD_CF_SIZE == 4
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, +0.0f, -0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, -0.0f, +0.0f, +0.0f};
+#endif
+
+  simd_f_t mask1 = srslte_simd_f_loadu(_mask1);
+  simd_f_t mask2 = srslte_simd_f_loadu(_mask2);
 
-  for (i = 0; i < nof_symbols - 1; i += 2) {
+  for (; i < nof_symbols - SRSLTE_SIMD_CF_SIZE + 1; i += SRSLTE_SIMD_CF_SIZE) {
     /* Load channel */
-    __m128 h00i = _mm_load_ps((float *) &h[0][0][i]);
+    simd_cf_t h00i = srslte_simd_cfi_load(&h[0][0][i]);
-    __m128 h01i = _mm_load_ps((float *) &h[0][1][i]);
+    simd_cf_t h01i = srslte_simd_cfi_load(&h[0][1][i]);
-    __m128 h10i = _mm_load_ps((float *) &h[1][0][i]);
+    simd_cf_t h10i = srslte_simd_cfi_load(&h[1][0][i]);
-    __m128 h11i = _mm_load_ps((float *) &h[1][1][i]);
+    simd_cf_t h11i = srslte_simd_cfi_load(&h[1][1][i]);
 
     /* Apply precoding */
-    __m128 h00 = _mm_add_ps(h00i, _mm_xor_ps(h10i, _mm_setr_ps(+0.0f, +0.0f, -0.0f, -0.0f)));
+    simd_cf_t h00, h01, h10, h11;
-    __m128 h10 = _mm_add_ps(h01i, _mm_xor_ps(h11i, _mm_setr_ps(+0.0f, +0.0f, -0.0f, -0.0f)));
+    h00 = srslte_simd_cf_add(h00i, srslte_simd_cf_neg_mask(h10i, mask1));
-    __m128 h01 = _mm_add_ps(h00i, _mm_xor_ps(h10i, _mm_setr_ps(-0.0f, -0.0f, +0.0f, +0.0f)));
+    h10 = srslte_simd_cf_add(h01i, srslte_simd_cf_neg_mask(h11i, mask1));
-    __m128 h11 = _mm_add_ps(h01i, _mm_xor_ps(h11i, _mm_setr_ps(-0.0f, -0.0f, +0.0f, +0.0f)));
+    h01 = srslte_simd_cf_add(h00i, srslte_simd_cf_neg_mask(h10i, mask2));
+    h11 = srslte_simd_cf_add(h01i, srslte_simd_cf_neg_mask(h11i, mask2));
 
-    __m128 y0 = _mm_load_ps((float *) &y[0][i]);
+    simd_cf_t y0 = srslte_simd_cfi_load(&y[0][i]);
-    __m128 y1 = _mm_load_ps((float *) &y[1][i]);
+    simd_cf_t y1 = srslte_simd_cfi_load(&y[1][i]);
 
-    __m128 x0, x1;
+    simd_cf_t x0, x1;
 
-    srslte_mat_2x2_zf_sse(y0, y1, h00, h01, h10, h11, &x0, &x1, 2.0f / scaling);
+    srslte_mat_2x2_zf_simd(y0, y1, h00, h01, h10, h11, &x0, &x1, norm);
 
-    _mm_store_ps((float *) &x[0][i], x0);
+    srslte_simd_cfi_store(&x[0][i], x0);
-    _mm_store_ps((float *) &x[1][i], x1);
+    srslte_simd_cfi_store(&x[1][i], x1);
   }
+#endif /* SRSLTE_SIMD_CF_SIZE != 0 */
 
-  return nof_symbols;
+  cf_t h00, h01, h10, h11, det;
-}
+  for (; i < nof_symbols; i++) {
-#endif
-
-// Generic implementation of ZF 2x2 CCD equalizer
-int srslte_predecoding_ccd_2x2_zf_gen(cf_t *y[SRSLTE_MAX_PORTS],
-                                      cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
-                                      cf_t *x[SRSLTE_MAX_LAYERS],
-                                      int nof_symbols,
-                                      float scaling) {
-  cf_t h00, h01, h10, h11;
-
-  for (int i = 0; i < nof_symbols; i++) {
 
     // Even precoder
     h00 = +h[0][0][i] + h[1][0][i];
     h10 = +h[0][1][i] + h[1][1][i];
     h01 = +h[0][0][i] - h[1][0][i];
     h11 = +h[0][1][i] - h[1][1][i];
+    det = (h00 * h11 - h01 * h10);
+    det = conjf(det) * (norm / (crealf(det) * crealf(det) + cimagf(det) * cimagf(det)));
 
-    srslte_mat_2x2_zf_gen(y[0][i], y[1][i], h00, h01, h10, h11, &x[0][i], &x[1][i], 2.0f / scaling);
+    x[0][i] = (+h11 * y[0][i] - h01 * y[1][i]) * det;
+    x[1][i] = (-h10 * y[0][i] + h00 * y[1][i]) * det;
 
     i++;
 
@@ -718,30 +885,35 @@ int srslte_predecoding_ccd_2x2_zf_gen(cf_t *y[SRSLTE_MAX_PORTS],
     h10 = h[0][1][i] - h[1][1][i];
     h01 = h[0][0][i] + h[1][0][i];
     h11 = h[0][1][i] + h[1][1][i];
+    det = (h00 * h11 - h01 * h10);
+    det = conjf(det) * (norm / (crealf(det) * crealf(det) + cimagf(det) * cimagf(det)));
 
-    srslte_mat_2x2_zf_gen(y[0][i], y[1][i], h00, h01, h10, h11, &x[0][i], &x[1][i], 2.0f / scaling);
+    x[0][i] = (+h11 * y[0][i] - h01 * y[1][i]) * det;
+    x[1][i] = (-h10 * y[0][i] + h00 * y[1][i]) * det;
   }
   return SRSLTE_SUCCESS;
 }
 
-int srslte_predecoding_ccd_zf(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS], cf_t *x[SRSLTE_MAX_LAYERS],
+static int srslte_predecoding_ccd_zf(cf_t *y[SRSLTE_MAX_PORTS],
-                              int nof_rxant, int nof_ports, int nof_layers, int nof_symbols, float scaling)
+                                     cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
-{
+                                     cf_t *x[SRSLTE_MAX_LAYERS],
+                                     float *csi[SRSLTE_MAX_CODEWORDS],
+                                     int nof_rxant,
+                                     int nof_ports,
+                                     int nof_layers,
+                                     int nof_symbols,
+                                     float scaling) {
   if (nof_ports == 2 && nof_rxant == 2) {
     if (nof_layers == 2) {
-#ifdef LV_HAVE_AVX
+      if (csi && csi[0]) {
-      return srslte_predecoding_ccd_2x2_zf_avx(y, h, x, nof_symbols, scaling);
+        return srslte_predecoding_ccd_2x2_zf_csi(y, h, x, csi, nof_symbols, scaling);
-#else
+      } else {
-#ifdef LV_HAVE_SSE
+        return srslte_predecoding_ccd_2x2_zf(y, h, x, nof_symbols, scaling);
-      return srslte_predecoding_ccd_2x2_zf_sse(y, h, x, nof_symbols, scaling);
+      }
-#else
-      return srslte_predecoding_ccd_2x2_zf_gen(y, h, x, nof_symbols, scaling);
-#endif /* LV_HAVE_SSE */
-#endif /* LV_HAVE_AVX */
     } else {
       DEBUG("Error predecoding CCD: Invalid number of layers %d\n", nof_layers);
-      return -1;       
+      return -1;
-    }          
+    }
   } else if (nof_ports == 4) {
     DEBUG("Error predecoding CCD: Only 2 ports supported\n");
   } else {
@@ -750,86 +922,155 @@ int srslte_predecoding_ccd_zf(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORT
   return SRSLTE_ERROR;
 }
 
-// AVX implementation of MMSE 2x2 CCD equalizer
+static int srslte_predecoding_ccd_2x2_mmse_csi(cf_t *y[SRSLTE_MAX_PORTS],
-#ifdef LV_HAVE_AVX
+                                               cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
+                                               cf_t *x[SRSLTE_MAX_LAYERS],
+                                               float *csi[SRSLTE_MAX_CODEWORDS],
+                                               int nof_symbols, float scaling, float noise_estimate) {
+  int i = 0;
+  float norm = 2.0f / scaling;
+
+#if SRSLTE_SIMD_CF_SIZE != 0
+#if SRSLTE_SIMD_CF_SIZE == 16
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f,
+                                       +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f,
+                                       -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f};
+#elif SRSLTE_SIMD_CF_SIZE == 8
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f};
+#elif SRSLTE_SIMD_CF_SIZE == 4
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, -0.0f, +0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, +0.0f, -0.0f, +0.0f};
+#endif
 
-int srslte_predecoding_ccd_2x2_mmse_avx(cf_t *y[SRSLTE_MAX_PORTS],
+  simd_f_t mask1 = srslte_simd_f_loadu(_mask1);
-                                        cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
+  simd_f_t mask2 = srslte_simd_f_loadu(_mask2);
-                                        cf_t *x[SRSLTE_MAX_LAYERS],
-                                        uint32_t nof_symbols, float scaling, float noise_estimate) {
-  uint32_t i = 0;
 
-  for (i = 0; i < nof_symbols - 3; i += 4) {
+  for (; i < nof_symbols - SRSLTE_SIMD_CF_SIZE + 1; i += SRSLTE_SIMD_CF_SIZE) {
     /* Load channel */
-    __m256 h00i = _mm256_load_ps((float *) &h[0][0][i]);
+    simd_cf_t h00i = srslte_simd_cfi_load(&h[0][0][i]);
-    __m256 h01i = _mm256_load_ps((float *) &h[0][1][i]);
+    simd_cf_t h01i = srslte_simd_cfi_load(&h[0][1][i]);
-    __m256 h10i = _mm256_load_ps((float *) &h[1][0][i]);
+    simd_cf_t h10i = srslte_simd_cfi_load(&h[1][0][i]);
-    __m256 h11i = _mm256_load_ps((float *) &h[1][1][i]);
+    simd_cf_t h11i = srslte_simd_cfi_load(&h[1][1][i]);
 
     /* Apply precoding */
-    __m256 h00 = _mm256_add_ps(h00i, _mm256_xor_ps(h10i, _mm256_setr_ps(+0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f)));
+    simd_cf_t h00, h01, h10, h11;
-    __m256 h10 = _mm256_add_ps(h01i, _mm256_xor_ps(h11i, _mm256_setr_ps(+0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f)));
+    h00 = srslte_simd_cf_add(h00i, srslte_simd_cf_neg_mask(h10i, mask1));
-    __m256 h01 = _mm256_add_ps(h00i, _mm256_xor_ps(h10i, _mm256_setr_ps(-0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f)));
+    h10 = srslte_simd_cf_add(h01i, srslte_simd_cf_neg_mask(h11i, mask1));
-    __m256 h11 = _mm256_add_ps(h01i, _mm256_xor_ps(h11i, _mm256_setr_ps(-0.0f, -0.0f, +0.0f, +0.0f, -0.0f, -0.0f, +0.0f, +0.0f)));
+    h01 = srslte_simd_cf_add(h00i, srslte_simd_cf_neg_mask(h10i, mask2));
+    h11 = srslte_simd_cf_add(h01i, srslte_simd_cf_neg_mask(h11i, mask2));
 
-    __m256 y0 = _mm256_load_ps((float *) &y[0][i]);
+    simd_cf_t y0 = srslte_simd_cfi_load(&y[0][i]);
-    __m256 y1 = _mm256_load_ps((float *) &y[1][i]);
+    simd_cf_t y1 = srslte_simd_cfi_load(&y[1][i]);
 
-    __m256 x0, x1;
+    simd_cf_t x0, x1;
+    simd_f_t csi0, csi1;
 
-    srslte_mat_2x2_mmse_avx(y0, y1, h00, h01, h10, h11, &x0, &x1, noise_estimate, 2.0f / scaling);
+    srslte_mat_2x2_mmse_csi_simd(y0, y1, h00, h01, h10, h11, &x0, &x1, &csi0, &csi1, noise_estimate, norm);
 
-    _mm256_store_ps((float *) &x[0][i], x0);
+    srslte_simd_cfi_store(&x[0][i], x0);
-    _mm256_store_ps((float *) &x[1][i], x1);
+    srslte_simd_cfi_store(&x[1][i], x1);
+
+    srslte_simd_f_store(&csi[0][i], csi0);
+    srslte_simd_f_store(&csi[1][i], csi1);
   }
+#endif /* SRSLTE_SIMD_CF_SIZE != 0 */
 
-  return nof_symbols;
+  cf_t h00, h01, h10, h11;
+  for (; i < nof_symbols; i++) {
+
+    // Even precoder
+    h00 = +h[0][0][i] + h[1][0][i];
+    h10 = +h[0][1][i] + h[1][1][i];
+    h01 = +h[0][0][i] - h[1][0][i];
+    h11 = +h[0][1][i] - h[1][1][i];
+    srslte_mat_2x2_mmse_csi_gen(y[0][i],
+                                y[1][i],
+                                h00,
+                                h01,
+                                h10,
+                                h11,
+                                &x[0][i],
+                                &x[1][i],
+                                &csi[0][i],
+                                &csi[1][i],
+                                noise_estimate,
+                                norm);
+    i++;
+
+    // Odd precoder
+    h00 = h[0][0][i] - h[1][0][i];
+    h10 = h[0][1][i] - h[1][1][i];
+    h01 = h[0][0][i] + h[1][0][i];
+    h11 = h[0][1][i] + h[1][1][i];
+    srslte_mat_2x2_mmse_csi_gen(y[0][i],
+                                y[1][i],
+                                h00,
+                                h01,
+                                h10,
+                                h11,
+                                &x[0][i],
+                                &x[1][i],
+                                &csi[0][i],
+                                &csi[1][i],
+                                noise_estimate,
+                                norm);
+  }
+  return SRSLTE_SUCCESS;
 }
-#endif /* LV_HAVE_AVX */
 
-// SSE implementation of ZF 2x2 CCD equalizer
+static int srslte_predecoding_ccd_2x2_mmse(cf_t *y[SRSLTE_MAX_PORTS],
-#ifdef LV_HAVE_SSE
+                                           cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
+                                           cf_t *x[SRSLTE_MAX_LAYERS],
+                                           int nof_symbols, float scaling, float noise_estimate) {
+  int i = 0;
+  float norm = 2.0f / scaling;
+
+#if SRSLTE_SIMD_CF_SIZE != 0
+#if SRSLTE_SIMD_CF_SIZE == 16
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f,
+                                       +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f,
+                                       -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f};
+#elif SRSLTE_SIMD_CF_SIZE == 8
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f, -0.0f, +0.0f};
+#elif SRSLTE_SIMD_CF_SIZE == 4
+  float _mask1[SRSLTE_SIMD_CF_SIZE] = {+0.0f, -0.0f, +0.0f, -0.0f};
+  float _mask2[SRSLTE_SIMD_CF_SIZE] = {-0.0f, +0.0f, -0.0f, +0.0f};
+#endif
 
-int srslte_predecoding_ccd_2x2_mmse_sse(cf_t *y[SRSLTE_MAX_PORTS],
+  simd_f_t mask1 = srslte_simd_f_loadu(_mask1);
-                                      cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
+  simd_f_t mask2 = srslte_simd_f_loadu(_mask2);
-                                      cf_t *x[SRSLTE_MAX_LAYERS],
-                                      uint32_t nof_symbols, float scaling, float noise_estimate) {
-  uint32_t i = 0;
 
-  for (i = 0; i < nof_symbols - 1; i += 2) {
+  for (; i < nof_symbols - SRSLTE_SIMD_CF_SIZE + 1; i += SRSLTE_SIMD_CF_SIZE) {
     /* Load channel */
-    __m128 h00i = _mm_load_ps((float *) &h[0][0][i]);
+    simd_cf_t h00i = srslte_simd_cfi_load(&h[0][0][i]);
-    __m128 h01i = _mm_load_ps((float *) &h[0][1][i]);
+    simd_cf_t h01i = srslte_simd_cfi_load(&h[0][1][i]);
-    __m128 h10i = _mm_load_ps((float *) &h[1][0][i]);
+    simd_cf_t h10i = srslte_simd_cfi_load(&h[1][0][i]);
-    __m128 h11i = _mm_load_ps((float *) &h[1][1][i]);
+    simd_cf_t h11i = srslte_simd_cfi_load(&h[1][1][i]);
 
     /* Apply precoding */
-    __m128 h00 = _mm_add_ps(h00i, _mm_xor_ps(h10i, _mm_setr_ps(+0.0f, +0.0f, -0.0f, -0.0f)));
+    simd_cf_t h00, h01, h10, h11;
-    __m128 h10 = _mm_add_ps(h01i, _mm_xor_ps(h11i, _mm_setr_ps(+0.0f, +0.0f, -0.0f, -0.0f)));
+    h00 = srslte_simd_cf_add(h00i, srslte_simd_cf_neg_mask(h10i, mask1));
-    __m128 h01 = _mm_add_ps(h00i, _mm_xor_ps(h10i, _mm_setr_ps(-0.0f, -0.0f, +0.0f, +0.0f)));
+    h10 = srslte_simd_cf_add(h01i, srslte_simd_cf_neg_mask(h11i, mask1));
-    __m128 h11 = _mm_add_ps(h01i, _mm_xor_ps(h11i, _mm_setr_ps(-0.0f, -0.0f, +0.0f, +0.0f)));
+    h01 = srslte_simd_cf_add(h00i, srslte_simd_cf_neg_mask(h10i, mask2));
+    h11 = srslte_simd_cf_add(h01i, srslte_simd_cf_neg_mask(h11i, mask2));
 
-    __m128 y0 = _mm_load_ps((float *) &y[0][i]);
+    simd_cf_t y0 = srslte_simd_cfi_load(&y[0][i]);
-    __m128 y1 = _mm_load_ps((float *) &y[1][i]);
+    simd_cf_t y1 = srslte_simd_cfi_load(&y[1][i]);
 
-    __m128 x0, x1;
+    simd_cf_t x0, x1;
+    srslte_mat_2x2_mmse_simd(y0, y1, h00, h01, h10, h11, &x0, &x1, noise_estimate, norm);
 
-    srslte_mat_2x2_mmse_sse(y0, y1, h00, h01, h10, h11, &x0, &x1, noise_estimate, 2.0f / scaling);
+    srslte_simd_cfi_store(&x[0][i], x0);
-
+    srslte_simd_cfi_store(&x[1][i], x1);
-    _mm_store_ps((float *) &x[0][i], x0);
-    _mm_store_ps((float *) &x[1][i], x1);
   }
+#endif /* SRSLTE_SIMD_CF_SIZE != 0 */
 
-  return nof_symbols;
-}
-#endif /* LV_HAVE_SSE */
-
-// Generic implementation of ZF 2x2 CCD equalizer
-int srslte_predecoding_ccd_2x2_mmse_gen(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS], cf_t *x[SRSLTE_MAX_LAYERS],
-                                      int nof_symbols, float scaling, float noise_estimate) {
   cf_t h00, h01, h10, h11;
-
+  for (i = 0; i < nof_symbols; i++) {
-  for (int i = 0; i < nof_symbols; i++) {
 
     // Even precoder
     h00 = +h[0][0][i] + h[1][0][i];
@@ -850,21 +1091,23 @@ int srslte_predecoding_ccd_2x2_mmse_gen(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLT
   return SRSLTE_SUCCESS;
 }
 
-
+int srslte_predecoding_ccd_mmse(cf_t *y[SRSLTE_MAX_PORTS],
-int srslte_predecoding_ccd_mmse(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS], cf_t *x[SRSLTE_MAX_LAYERS],
+                                cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
-                              int nof_rxant, int nof_ports, int nof_layers, int nof_symbols, float scaling, float noise_estimate)
+                                cf_t *x[SRSLTE_MAX_LAYERS],
-{
+                                float *csi[SRSLTE_MAX_CODEWORDS],
+                                int nof_rxant,
+                                int nof_ports,
+                                int nof_layers,
+                                int nof_symbols,
+                                float scaling,
+                                float noise_estimate) {
   if (nof_ports == 2 && nof_rxant == 2) {
     if (nof_layers == 2) {
-#ifdef LV_HAVE_AVX
+      if (csi && csi[0])
-      return srslte_predecoding_ccd_2x2_mmse_avx(y, h, x, nof_symbols, scaling, noise_estimate);
+        return srslte_predecoding_ccd_2x2_mmse_csi(y, h, x, csi, nof_symbols, scaling, noise_estimate);
-#else
+      else {
-#ifdef LV_HAVE_SSE
+        return srslte_predecoding_ccd_2x2_mmse(y, h, x, nof_symbols, scaling, noise_estimate);
-      return srslte_predecoding_ccd_2x2_mmse_sse(y, h, x, nof_symbols, scaling, noise_estimate);
+      }
-#else
-      return srslte_predecoding_ccd_2x2_mmse_gen(y, h, x, nof_symbols, scaling, noise_estimate);
-#endif /* LV_HAVE_SSE */
-#endif /* LV_HAVE_AVX */
     } else {
       DEBUG("Error predecoding CCD: Invalid number of layers %d\n", nof_layers);
       return -1;
@@ -1468,7 +1711,7 @@ void srslte_predecoding_set_mimo_decoder (srslte_mimo_decoder_t _mimo_decoder) {
 
 /* 36.211 v10.3.0 Section 6.3.4 */
 int srslte_predecoding_type(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS][SRSLTE_MAX_PORTS],
-                            cf_t *x[SRSLTE_MAX_LAYERS], float *csi, int nof_rxant, int nof_ports, int nof_layers,
+                            cf_t *x[SRSLTE_MAX_LAYERS], float *csi[SRSLTE_MAX_CODEWORDS], int nof_rxant, int nof_ports, int nof_layers,
                             int codebook_idx, int nof_symbols, srslte_mimo_type_t type, float scaling,
                             float noise_estimate) {
 
@@ -1488,10 +1731,10 @@ int srslte_predecoding_type(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS]
     if (nof_layers >= 2 && nof_layers <= 4) {
       switch (mimo_decoder) {
         case SRSLTE_MIMO_DECODER_ZF:
-          return srslte_predecoding_ccd_zf(y, h, x, nof_rxant, nof_ports, nof_layers, nof_symbols, scaling);
+          return srslte_predecoding_ccd_zf(y, h, x, csi, nof_rxant, nof_ports, nof_layers, nof_symbols, scaling);
           break;
         case SRSLTE_MIMO_DECODER_MMSE:
-          return srslte_predecoding_ccd_mmse(y, h, x, nof_rxant, nof_ports, nof_layers, nof_symbols, scaling, noise_estimate);
+          return srslte_predecoding_ccd_mmse(y, h, x, csi, nof_rxant, nof_ports, nof_layers, nof_symbols, scaling, noise_estimate);
           break;
       }
     } else {
@@ -1510,7 +1753,7 @@ int srslte_predecoding_type(cf_t *y[SRSLTE_MAX_PORTS], cf_t *h[SRSLTE_MAX_PORTS]
     break;
   case SRSLTE_MIMO_TYPE_TX_DIVERSITY:
     if (nof_ports == nof_layers) {
-      return srslte_predecoding_diversity_multi(y, h, x, nof_rxant, nof_ports, nof_symbols, scaling);
+      return srslte_predecoding_diversity_multi(y, h, x, csi, nof_rxant, nof_ports, nof_symbols, scaling);
     } else {
       fprintf(stderr,
           "Error number of layers must equal number of ports in transmit diversity\n");

lib/src/phy/phch/pcfich.c

@@ -221,7 +221,7 @@ int srslte_pcfich_decode_multi(srslte_pcfich_t *q, cf_t *sf_symbols[SRSLTE_MAX_P
       /* no need for layer demapping */
       srslte_predecoding_single_multi(q_symbols, q_ce[0], q->d, NULL, q->nof_rx_antennas, q->nof_symbols, 1.0f, noise_estimate);
     } else {
-      srslte_predecoding_diversity_multi(q_symbols, q_ce, x, q->nof_rx_antennas, q->cell.nof_ports, q->nof_symbols, 1.0f);
+      srslte_predecoding_diversity_multi(q_symbols, q_ce, x, NULL, q->nof_rx_antennas, q->cell.nof_ports, q->nof_symbols, 1.0f);
       srslte_layerdemap_diversity(x, q->d, q->cell.nof_ports, q->nof_symbols / q->cell.nof_ports);
     }
 

lib/src/phy/phch/pdcch.c

@@ -492,7 +492,7 @@ int srslte_pdcch_extract_llr_multi(srslte_pdcch_t *q, cf_t *sf_symbols[SRSLTE_MA
       /* no need for layer demapping */
       srslte_predecoding_single_multi(q->symbols, q->ce[0], q->d, NULL, q->nof_rx_antennas, nof_symbols, 1.0f, noise_estimate/2);
     } else {
-      srslte_predecoding_diversity_multi(q->symbols, q->ce, x, q->nof_rx_antennas, q->cell.nof_ports, nof_symbols, 1.0f);
+      srslte_predecoding_diversity_multi(q->symbols, q->ce, x, NULL, q->nof_rx_antennas, q->cell.nof_ports, nof_symbols, 1.0f);
       srslte_layerdemap_diversity(x, q->d, q->cell.nof_ports, nof_symbols / q->cell.nof_ports);
     }
 

lib/src/phy/phch/pdsch.c

@@ -402,7 +402,7 @@ int srslte_pdsch_enable_csi(srslte_pdsch_t *q, bool enable) {
   if (enable) {
     for (int i = 0; i < SRSLTE_MAX_CODEWORDS; i++) {
       if (!q->csi[i]) {
-        q->csi[i] = srslte_vec_malloc(sizeof(float) * q->max_re);
+        q->csi[i] = srslte_vec_malloc(sizeof(float) * q->max_re * 2);
         if (!q->csi[i]) {
           return SRSLTE_ERROR;
         }
@@ -757,7 +757,7 @@ int srslte_pdsch_decode(srslte_pdsch_t *q,
     }
 
     // Pre-decoder
-    if (srslte_predecoding_type(q->symbols, q->ce, x, q->csi[0], q->nof_rx_antennas, q->cell.nof_ports, cfg->nof_layers,
+    if (srslte_predecoding_type(q->symbols, q->ce, x, q->csi, q->nof_rx_antennas, q->cell.nof_ports, cfg->nof_layers,
                                       cfg->codebook_idx, cfg->nbits[0].nof_re, cfg->mimo_type, pdsch_scaling, noise_estimate)<0) {
       DEBUG("Error predecoding\n");
       return SRSLTE_ERROR;

lib/src/phy/phch/phich.c

@@ -241,7 +241,7 @@ int srslte_phich_decode(srslte_phich_t *q, cf_t *sf_symbols[SRSLTE_MAX_PORTS],
     /* no need for layer demapping */
     srslte_predecoding_single_multi(q_sf_symbols, q_ce[0], q->d0, NULL, q->nof_rx_antennas, SRSLTE_PHICH_MAX_NSYMB, 1.0f, noise_estimate);
   } else {
-    srslte_predecoding_diversity_multi(q_sf_symbols, q_ce, x, q->nof_rx_antennas, q->cell.nof_ports, SRSLTE_PHICH_MAX_NSYMB, 1.0f);
+    srslte_predecoding_diversity_multi(q_sf_symbols, q_ce, x, NULL, q->nof_rx_antennas, q->cell.nof_ports, SRSLTE_PHICH_MAX_NSYMB, 1.0f);
     srslte_layerdemap_diversity(x, q->d0, q->cell.nof_ports, SRSLTE_PHICH_MAX_NSYMB / q->cell.nof_ports);
   }
   DEBUG("Recv!!: \n");

lib/src/phy/utils/mat.c

@@ -60,8 +60,8 @@ inline void srslte_mat_2x2_zf_gen(cf_t y0, cf_t y1, cf_t h00, cf_t h01, cf_t h10
 }
 
 /* Generic implementation for Minimum Mean Squared Error (MMSE) solver */
-inline void srslte_mat_2x2_mmse_gen(cf_t y0, cf_t y1, cf_t h00, cf_t h01, cf_t h10, cf_t h11,
+inline void srslte_mat_2x2_mmse_csi_gen(cf_t y0, cf_t y1, cf_t h00, cf_t h01, cf_t h10, cf_t h11,
-                                        cf_t *x0, cf_t *x1, float noise_estimate, float norm) {
+                                        cf_t *x0, cf_t *x1, float *csi0, float *csi1, float noise_estimate, float norm) {
   /* Create conjugated matrix */
   cf_t _h00 = conjf(h00);
   cf_t _h01 = conjf(h01);
@@ -73,14 +73,14 @@ inline void srslte_mat_2x2_mmse_gen(cf_t y0, cf_t y1, cf_t h00, cf_t h01, cf_t h
   cf_t a01 = _h00 * h01 + _h10 * h11;
   cf_t a10 = _h01 * h00 + _h11 * h10;
   cf_t a11 = _h01 * h01 + _h11 * h11 + noise_estimate;
+  cf_t a_det_rcp = srslte_mat_cf_recip_gen(srslte_mat_2x2_det_gen(a00, a01, a10, a11));
 
   /* 2. B = inv(H' x H + No) = inv(A) */
-  cf_t b00 = a11;
+  cf_t _norm = norm * a_det_rcp;
-  cf_t b01 = -a01;
+  cf_t b00 = a11 * _norm;
-  cf_t b10 = -a10;
+  cf_t b01 = -a01 * _norm;
-  cf_t b11 = a00;
+  cf_t b10 = -a10 * _norm;
-  cf_t _norm = norm * srslte_mat_cf_recip_gen(srslte_mat_2x2_det_gen(a00, a01, a10, a11));
+  cf_t b11 = a00 * _norm;
-
 
   /* 3. W = inv(H' x H + No) x H' = B x H' */
   cf_t w00 = b00 * _h00 + b01 * _h01;
@@ -89,8 +89,19 @@ inline void srslte_mat_2x2_mmse_gen(cf_t y0, cf_t y1, cf_t h00, cf_t h01, cf_t h
   cf_t w11 = b10 * _h10 + b11 * _h11;
 
   /* 4. X = W x Y */
-  *x0 = (y0 * w00 + y1 * w01) * _norm;
+  *x0 = (y0 * w00 + y1 * w01);
-  *x1 = (y0 * w10 + y1 * w11) * _norm;
+  *x1 = (y0 * w10 + y1 * w11);
+
+  /* 5. Set CSI */
+  *csi0 = 1.0f / crealf(b00);
+  *csi1 = 1.0f / crealf(b11);
+}
+
+/* Generic implementation for Minimum Mean Squared Error (MMSE) solver */
+void srslte_mat_2x2_mmse_gen(cf_t y0, cf_t y1, cf_t h00, cf_t h01, cf_t h10, cf_t h11,
+                                        cf_t *x0, cf_t *x1, float noise_estimate, float norm) {
+  float csi0, csi1;
+  srslte_mat_2x2_mmse_csi_gen(y0, y1, h00, h01, h10, h11, x0, x1, &csi0, &csi1, noise_estimate, norm);
 }
 
 inline float srslte_mat_2x2_cn(cf_t h00, cf_t h01, cf_t h10, cf_t h11) {

lib/src/phy/utils/test/mat_test.c

@@ -32,8 +32,8 @@
 #include <sys/time.h>
 
 #include "srslte/phy/utils/mat.h"
-#include "srslte/phy/utils/simd.h"
 #include "srslte/phy/utils/vector.h"
+#include "srslte/phy/utils/vector_simd.h"
 
 
 bool zf_solver = false;
@@ -378,6 +378,98 @@ bool test_mmse_solver_avx(void) {
 
 #endif /* LV_HAVE_AVX */
 
+#if SRSLTE_SIMD_CF_SIZE != 0
+
+bool test_zf_solver_simd(void) {
+  cf_t cf_error0, cf_error1;
+  float error = 0.0f;
+
+  cf_t x0_gold_1 = RANDOM_CF();
+  cf_t x1_gold_1 = RANDOM_CF();
+  cf_t h00_1 = RANDOM_CF();
+  cf_t h01_1 = RANDOM_CF();
+  cf_t h10_1 = RANDOM_CF();
+  cf_t h11_1 = (1 - h01_1 * h10_1) / h00_1;
+  cf_t y0_1 = x0_gold_1 * h00_1 + x1_gold_1 * h01_1;
+  cf_t y1_1 = x0_gold_1 * h10_1 + x1_gold_1 * h11_1;
+
+  simd_cf_t _y0 = srslte_simd_cf_set1(y0_1);
+  simd_cf_t _y1 = srslte_simd_cf_set1(y1_1);
+
+  simd_cf_t _h00 = srslte_simd_cf_set1(h00_1);
+  simd_cf_t _h01 = srslte_simd_cf_set1(h01_1);
+  simd_cf_t _h10 = srslte_simd_cf_set1(h10_1);
+  simd_cf_t _h11 = srslte_simd_cf_set1(h11_1);
+
+  simd_cf_t _x0, _x1;
+
+  srslte_mat_2x2_zf_simd(_y0, _y1, _h00, _h01, _h10, _h11, &_x0, &_x1, 1.0f);
+
+  __attribute__((aligned(SRSLTE_SIMD_BIT_ALIGN))) cf_t x0[SRSLTE_SIMD_CF_SIZE];
+  __attribute__((aligned(SRSLTE_SIMD_BIT_ALIGN))) cf_t x1[SRSLTE_SIMD_CF_SIZE];
+
+  srslte_simd_cfi_store(x0, _x0);
+  srslte_simd_cfi_store(x1, _x1);
+
+  cf_error0 = x0[1] - x0_gold_1;
+  cf_error1 = x1[1] - x1_gold_1;
+  error += crealf(cf_error0) * crealf(cf_error0) + cimagf(cf_error0) * cimagf(cf_error0) +
+      crealf(cf_error1) * crealf(cf_error1) + cimagf(cf_error1) * cimagf(cf_error1);
+
+  return (error < 1e-3);
+}
+
+bool test_mmse_solver_simd(void) {
+  cf_t cf_error0, cf_error1;
+  float error = 0.0f;
+
+  cf_t x0_gold[SRSLTE_SIMD_CF_SIZE];
+  cf_t x1_gold[SRSLTE_SIMD_CF_SIZE];
+  cf_t h00[SRSLTE_SIMD_CF_SIZE];
+  cf_t h01[SRSLTE_SIMD_CF_SIZE];
+  cf_t h10[SRSLTE_SIMD_CF_SIZE];
+  cf_t h11[SRSLTE_SIMD_CF_SIZE];
+  cf_t y0[SRSLTE_SIMD_CF_SIZE];
+  cf_t y1[SRSLTE_SIMD_CF_SIZE];
+  for (int i = 0; i < SRSLTE_SIMD_CF_SIZE; i++) {
+    x0_gold[i] = RANDOM_CF();
+    x1_gold[i] = RANDOM_CF();
+    h00[i] = RANDOM_CF();
+    h01[i] = RANDOM_CF();
+    h10[i] = RANDOM_CF();
+    h11[i] = (1 - h01[i] * h10[i]) / h00[i];
+    y0[i] = x0_gold[i] * h00[i]+ x1_gold[i] * h01[i];
+    y1[i] = x0_gold[i] * h10[i] + x1_gold[i] * h11[i];
+  }
+
+  simd_cf_t _y0 = srslte_simd_cfi_loadu(y0);
+  simd_cf_t _y1 = srslte_simd_cfi_loadu(y1);
+
+  simd_cf_t _h00 = srslte_simd_cfi_loadu(h00);
+  simd_cf_t _h01 = srslte_simd_cfi_loadu(h01);
+  simd_cf_t _h10 = srslte_simd_cfi_loadu(h10);
+  simd_cf_t _h11 = srslte_simd_cfi_loadu(h11);
+
+  simd_cf_t _x0, _x1;
+
+  srslte_mat_2x2_mmse_simd(_y0, _y1, _h00, _h01, _h10, _h11, &_x0, &_x1, 0.0f, 1.0f);
+
+  __attribute__((aligned(SRSLTE_SIMD_BIT_ALIGN))) cf_t x0[SRSLTE_SIMD_CF_SIZE];
+  __attribute__((aligned(SRSLTE_SIMD_BIT_ALIGN))) cf_t x1[SRSLTE_SIMD_CF_SIZE];
+
+  srslte_simd_cfi_store(x0, _x0);
+  srslte_simd_cfi_store(x1, _x1);
+
+  cf_error0 = x0[1] - x0_gold[1];
+  cf_error1 = x1[1] - x1_gold[1];
+  error += crealf(cf_error0) * crealf(cf_error0) + cimagf(cf_error0) * cimagf(cf_error0) +
+      crealf(cf_error1) * crealf(cf_error1) + cimagf(cf_error1) * cimagf(cf_error1);
+
+  return (error < 1e-3);
+}
+
+#endif /* SRSLTE_SIMD_CF_SIZE != 0 */
+
 bool test_vec_dot_prod_ccc(void) {
   __attribute__((aligned(256))) cf_t a[14];
   __attribute__((aligned(256))) cf_t b[14];
@@ -413,6 +505,10 @@ int main(int argc, char **argv) {
 #ifdef LV_HAVE_AVX
     RUN_TEST(test_zf_solver_avx);
 #endif /* LV_HAVE_AVX */
+
+#if SRSLTE_SIMD_CF_SIZE != 0
+    RUN_TEST(test_zf_solver_simd);
+#endif /* SRSLTE_SIMD_CF_SIZE != 0*/
   }
 
   if (mmse_solver) {
@@ -426,6 +522,10 @@ int main(int argc, char **argv) {
 #ifdef LV_HAVE_AVX
     RUN_TEST(test_mmse_solver_avx);
 #endif /* LV_HAVE_AVX */
+
+#if SRSLTE_SIMD_CF_SIZE != 0
+    RUN_TEST(test_mmse_solver_simd);
+#endif /* SRSLTE_SIMD_CF_SIZE != 0*/
   }
 
   RUN_TEST(test_vec_dot_prod_ccc);

lib/src/phy/utils/vector.c

@@ -427,8 +427,8 @@ void srslte_vec_quant_sus(const int16_t *in, uint16_t *out, const float gain, co
   }
 }
 
-void srs_vec_cf_cpy(const cf_t *dst, cf_t *src, int len) {
+void srs_vec_cf_cpy(const cf_t *src, cf_t *dst, int len) {
-  srslte_vec_cp_simd(dst, src, len);
+  srslte_vec_cp_simd(src, dst, len);
 }
 
 void srslte_vec_interleave(const cf_t *x, const cf_t *y, cf_t *z, const int len) {