From db17c67194e32636e3212dd117101d768ac73998 Mon Sep 17 00:00:00 2001
From: Ismael Gomez <ismagom@gmail.com>
Date: Tue, 26 Sep 2017 17:26:29 +0200
Subject: [PATCH] Fixed bug in SSE turbo decoder

---
 lib/include/srslte/phy/fec/turbodecoder.h     | 12 ++++----
 .../srslte/phy/fec/turbodecoder_simd.h        | 30 ++++++++++---------
 .../srslte/phy/fec/turbodecoder_simd_inter.h  | 16 +++++-----
 lib/src/phy/fec/test/turbodecoder_test.c      | 10 +++----
 lib/src/phy/fec/turbodecoder.c                | 30 ++++++++++++-------
 lib/src/phy/fec/turbodecoder_avx.c            |  2 +-
 lib/src/phy/fec/turbodecoder_simd.c           | 30 +++++++++----------
 lib/src/phy/fec/turbodecoder_simd_inter.c     |  8 ++---
 lib/src/phy/phch/pdsch.c                      |  3 +-
 lib/src/phy/phch/sch.c                        | 21 ++++++++-----
 lib/src/phy/phch/test/pusch_test.c            |  2 +-
 11 files changed, 90 insertions(+), 74 deletions(-)

diff --git a/lib/include/srslte/phy/fec/turbodecoder.h b/lib/include/srslte/phy/fec/turbodecoder.h
index 4a24804fa..ac05422bc 100644
--- a/lib/include/srslte/phy/fec/turbodecoder.h
+++ b/lib/include/srslte/phy/fec/turbodecoder.h
@@ -74,7 +74,7 @@ SRSLTE_API int srslte_tdec_reset_cb(srslte_tdec_t * h,
 SRSLTE_API int srslte_tdec_get_nof_iterations_cb(srslte_tdec_t * h, 
                                                  uint32_t cb_idx);
 
-SRSLTE_API int srslte_tdec_get_nof_parallel(srslte_tdec_t * h); 
+SRSLTE_API uint32_t srslte_tdec_get_nof_parallel(srslte_tdec_t * h);
 
 SRSLTE_API void srslte_tdec_iteration(srslte_tdec_t * h, 
                                       int16_t* input, 
@@ -95,15 +95,15 @@ SRSLTE_API int srslte_tdec_run_all(srslte_tdec_t * h,
                                    uint32_t long_cb);
 
 SRSLTE_API void srslte_tdec_iteration_par(srslte_tdec_t * h, 
-                                          int16_t* input[SRSLTE_TDEC_NPAR], 
+                                          int16_t* input[SRSLTE_TDEC_MAX_NPAR],
                                           uint32_t long_cb);
 
 SRSLTE_API void srslte_tdec_decision_par(srslte_tdec_t * h, 
-                                         uint8_t *output[SRSLTE_TDEC_NPAR], 
+                                         uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                          uint32_t long_cb);
 
 SRSLTE_API void srslte_tdec_decision_byte_par(srslte_tdec_t * h, 
-                                              uint8_t *output[SRSLTE_TDEC_NPAR], 
+                                              uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                               uint32_t long_cb); 
 
 SRSLTE_API void srslte_tdec_decision_byte_par_cb(srslte_tdec_t * h, 
@@ -112,8 +112,8 @@ SRSLTE_API void srslte_tdec_decision_byte_par_cb(srslte_tdec_t * h,
                                                  uint32_t long_cb);
 
 SRSLTE_API int srslte_tdec_run_all_par(srslte_tdec_t * h, 
-                                       int16_t * input[SRSLTE_TDEC_NPAR], 
-                                       uint8_t *output[SRSLTE_TDEC_NPAR],
+                                       int16_t * input[SRSLTE_TDEC_MAX_NPAR],
+                                       uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                        uint32_t nof_iterations, 
                                        uint32_t long_cb);
 
diff --git a/lib/include/srslte/phy/fec/turbodecoder_simd.h b/lib/include/srslte/phy/fec/turbodecoder_simd.h
index 8cafbc5f8..a899a5964 100644
--- a/lib/include/srslte/phy/fec/turbodecoder_simd.h
+++ b/lib/include/srslte/phy/fec/turbodecoder_simd.h
@@ -44,7 +44,7 @@
 #include "srslte/phy/fec/cbsegm.h"
 
 // Define maximum number of CB decoded in parallel (2 for AVX2)
-#define SRSLTE_TDEC_NPAR 2
+#define SRSLTE_TDEC_MAX_NPAR 2
 
 #define SRSLTE_TCOD_RATE 3
 #define SRSLTE_TCOD_TOTALTAIL 12
@@ -65,18 +65,18 @@ typedef struct SRSLTE_API {
   
   map_gen_t dec;
 
-  int16_t *app1[SRSLTE_TDEC_NPAR];
-  int16_t *app2[SRSLTE_TDEC_NPAR];
-  int16_t *ext1[SRSLTE_TDEC_NPAR];
-  int16_t *ext2[SRSLTE_TDEC_NPAR];
-  int16_t *syst[SRSLTE_TDEC_NPAR];
-  int16_t *parity0[SRSLTE_TDEC_NPAR];
-  int16_t *parity1[SRSLTE_TDEC_NPAR];
+  int16_t *app1[SRSLTE_TDEC_MAX_NPAR];
+  int16_t *app2[SRSLTE_TDEC_MAX_NPAR];
+  int16_t *ext1[SRSLTE_TDEC_MAX_NPAR];
+  int16_t *ext2[SRSLTE_TDEC_MAX_NPAR];
+  int16_t *syst[SRSLTE_TDEC_MAX_NPAR];
+  int16_t *parity0[SRSLTE_TDEC_MAX_NPAR];
+  int16_t *parity1[SRSLTE_TDEC_MAX_NPAR];
   
   int cb_mask; 
   int current_cbidx; 
   srslte_tc_interl_t interleaver[SRSLTE_NOF_TC_CB_SIZES];
-  int n_iter[SRSLTE_TDEC_NPAR];
+  int n_iter[SRSLTE_TDEC_MAX_NPAR];
 } srslte_tdec_simd_t;
 
 SRSLTE_API int srslte_tdec_simd_init(srslte_tdec_simd_t * h, 
@@ -88,6 +88,8 @@ SRSLTE_API void srslte_tdec_simd_free(srslte_tdec_simd_t * h);
 SRSLTE_API int srslte_tdec_simd_reset(srslte_tdec_simd_t * h, 
                                       uint32_t long_cb);
 
+SRSLTE_API
+
 SRSLTE_API int srslte_tdec_simd_get_nof_iterations_cb(srslte_tdec_simd_t * h, 
                                                       uint32_t cb_idx);
 
@@ -95,15 +97,15 @@ SRSLTE_API int srslte_tdec_simd_reset_cb(srslte_tdec_simd_t * h,
                                          uint32_t cb_idx);
 
 SRSLTE_API void srslte_tdec_simd_iteration(srslte_tdec_simd_t * h, 
-                                           int16_t * input[SRSLTE_TDEC_NPAR], 
+                                           int16_t * input[SRSLTE_TDEC_MAX_NPAR],
                                            uint32_t long_cb);
 
 SRSLTE_API void srslte_tdec_simd_decision(srslte_tdec_simd_t * h, 
-                                          uint8_t *output[SRSLTE_TDEC_NPAR], 
+                                          uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                           uint32_t long_cb);
 
 SRSLTE_API void srslte_tdec_simd_decision_byte(srslte_tdec_simd_t * h, 
-                                               uint8_t *output[SRSLTE_TDEC_NPAR], 
+                                               uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                                uint32_t long_cb); 
 
 SRSLTE_API void srslte_tdec_simd_decision_byte_cb(srslte_tdec_simd_t * h, 
@@ -112,8 +114,8 @@ SRSLTE_API void srslte_tdec_simd_decision_byte_cb(srslte_tdec_simd_t * h,
                                                   uint32_t long_cb); 
 
 SRSLTE_API int srslte_tdec_simd_run_all(srslte_tdec_simd_t * h, 
-                                        int16_t * input[SRSLTE_TDEC_NPAR], 
-                                        uint8_t *output[SRSLTE_TDEC_NPAR],
+                                        int16_t * input[SRSLTE_TDEC_MAX_NPAR],
+                                        uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                         uint32_t nof_iterations, 
                                         uint32_t long_cb);
 
diff --git a/lib/include/srslte/phy/fec/turbodecoder_simd_inter.h b/lib/include/srslte/phy/fec/turbodecoder_simd_inter.h
index 402faa314..c80785809 100644
--- a/lib/include/srslte/phy/fec/turbodecoder_simd_inter.h
+++ b/lib/include/srslte/phy/fec/turbodecoder_simd_inter.h
@@ -49,9 +49,9 @@
 #include "srslte/phy/fec/cbsegm.h"
 
 #if LV_HAVE_AVX2
-  #define SRSLTE_TDEC_NPAR 16
+  #define SRSLTE_TDEC_MAX_NPAR 16
 #else
-  #define SRSLTE_TDEC_NPAR 8
+  #define SRSLTE_TDEC_MAX_NPAR 8
 #endif
 
 
@@ -71,7 +71,7 @@ typedef struct SRSLTE_API {
   int current_cbidx; 
   uint32_t current_long_cb;
   srslte_tc_interl_t interleaver[SRSLTE_NOF_TC_CB_SIZES];
-  int n_iter[SRSLTE_TDEC_NPAR];
+  int n_iter[SRSLTE_TDEC_MAX_NPAR];
 } srslte_tdec_simd_inter_t;
 
 SRSLTE_API int srslte_tdec_simd_inter_init(srslte_tdec_simd_inter_t * h, 
@@ -90,17 +90,17 @@ SRSLTE_API int srslte_tdec_simd_inter_reset_cb(srslte_tdec_simd_inter_t * h,
                                          uint32_t cb_idx);
 
 SRSLTE_API void srslte_tdec_simd_inter_iteration(srslte_tdec_simd_inter_t * h, 
-                                           int16_t * input[SRSLTE_TDEC_NPAR], 
+                                           int16_t * input[SRSLTE_TDEC_MAX_NPAR],
                                            uint32_t nof_cb,
                                            uint32_t long_cb);
 
 SRSLTE_API void srslte_tdec_simd_inter_decision(srslte_tdec_simd_inter_t * h, 
-                                          uint8_t *output[SRSLTE_TDEC_NPAR], 
+                                          uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                           uint32_t nof_cb,
                                           uint32_t long_cb);
 
 SRSLTE_API void srslte_tdec_simd_inter_decision_byte(srslte_tdec_simd_inter_t * h, 
-                                               uint8_t *output[SRSLTE_TDEC_NPAR], 
+                                               uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                                uint32_t nof_cb,
                                                uint32_t long_cb); 
 
@@ -110,8 +110,8 @@ SRSLTE_API void srslte_tdec_simd_inter_decision_byte_cb(srslte_tdec_simd_inter_t
                                                   uint32_t long_cb); 
 
 SRSLTE_API int srslte_tdec_simd_inter_run_all(srslte_tdec_simd_inter_t * h, 
-                                        int16_t *input[SRSLTE_TDEC_NPAR], 
-                                        uint8_t *output[SRSLTE_TDEC_NPAR],
+                                        int16_t *input[SRSLTE_TDEC_MAX_NPAR],
+                                        uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                         uint32_t nof_iterations, 
                                         uint32_t nof_cb,
                                         uint32_t long_cb);
diff --git a/lib/src/phy/fec/test/turbodecoder_test.c b/lib/src/phy/fec/test/turbodecoder_test.c
index df1d4b884..aeac21433 100644
--- a/lib/src/phy/fec/test/turbodecoder_test.c
+++ b/lib/src/phy/fec/test/turbodecoder_test.c
@@ -117,7 +117,7 @@ int main(int argc, char **argv) {
   float *llr;
   short *llr_s;
   uint8_t *llr_c;
-  uint8_t *data_tx, *data_rx, *data_rx_bytes[SRSLTE_TDEC_NPAR], *symbols;
+  uint8_t *data_tx, *data_rx, *data_rx_bytes[SRSLTE_TDEC_MAX_NPAR], *symbols;
   uint32_t i, j;
   float var[SNR_POINTS];
   uint32_t snr_points;
@@ -159,7 +159,7 @@ int main(int argc, char **argv) {
     perror("malloc");
     exit(-1);
   }
-  for (int cb=0;cb<SRSLTE_TDEC_NPAR;cb++) {
+  for (int cb=0;cb<SRSLTE_TDEC_MAX_NPAR;cb++) {
     data_rx_bytes[cb] = srslte_vec_malloc(frame_length * sizeof(uint8_t));
     if (!data_rx_bytes[cb]) {
       perror("malloc");
@@ -254,10 +254,10 @@ int main(int argc, char **argv) {
         t = nof_iterations;
       }
 
-      int16_t *input[SRSLTE_TDEC_NPAR];
-      uint8_t *output[SRSLTE_TDEC_NPAR];        
+      int16_t *input[SRSLTE_TDEC_MAX_NPAR];
+      uint8_t *output[SRSLTE_TDEC_MAX_NPAR];
       
-      for (int n=0;n<SRSLTE_TDEC_NPAR;n++) {
+      for (int n=0;n<SRSLTE_TDEC_MAX_NPAR;n++) {
         if (n < nof_cb) {
           input[n] = llr_s;         
         } else {
diff --git a/lib/src/phy/fec/turbodecoder.c b/lib/src/phy/fec/turbodecoder.c
index b19466a53..7803fde1b 100644
--- a/lib/src/phy/fec/turbodecoder.c
+++ b/lib/src/phy/fec/turbodecoder.c
@@ -43,7 +43,7 @@
 
 int srslte_tdec_init(srslte_tdec_t * h, uint32_t max_long_cb) {
 #ifdef LV_HAVE_SSE
-  return srslte_tdec_simd_init(&h->tdec_simd, SRSLTE_TDEC_NPAR, max_long_cb);
+  return srslte_tdec_simd_init(&h->tdec_simd, SRSLTE_TDEC_MAX_NPAR, max_long_cb);
 #else
   h->input_conv = srslte_vec_malloc(sizeof(float) * (3*max_long_cb+12));
   if (!h->input_conv) {
@@ -91,7 +91,7 @@ int srslte_tdec_get_nof_iterations_cb(srslte_tdec_t * h, uint32_t cb_idx)
 #endif  
 }
 
-void srslte_tdec_iteration_par(srslte_tdec_t * h, int16_t* input[SRSLTE_TDEC_NPAR], uint32_t long_cb) {
+void srslte_tdec_iteration_par(srslte_tdec_t * h, int16_t* input[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb) {
 #ifdef LV_HAVE_SSE
   srslte_tdec_simd_iteration(&h->tdec_simd, input, long_cb);      
 #else
@@ -101,12 +101,12 @@ void srslte_tdec_iteration_par(srslte_tdec_t * h, int16_t* input[SRSLTE_TDEC_NPA
 }
 
 void srslte_tdec_iteration(srslte_tdec_t * h, int16_t* input, uint32_t long_cb) {
-  int16_t *input_par[SRSLTE_TDEC_NPAR]; 
+  int16_t *input_par[SRSLTE_TDEC_MAX_NPAR];
   input_par[0] = input; 
   return srslte_tdec_iteration_par(h, input_par, long_cb);
 }
 
-void srslte_tdec_decision_par(srslte_tdec_t * h, uint8_t *output[SRSLTE_TDEC_NPAR], uint32_t long_cb) {
+void srslte_tdec_decision_par(srslte_tdec_t * h, uint8_t *output[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb) {
 #ifdef LV_HAVE_SSE
   return srslte_tdec_simd_decision(&h->tdec_simd, output, long_cb);
 #else
@@ -114,13 +114,21 @@ void srslte_tdec_decision_par(srslte_tdec_t * h, uint8_t *output[SRSLTE_TDEC_NPA
 #endif
 }
 
+uint32_t srslte_tdec_get_nof_parallel(srslte_tdec_t *h) {
+#ifdef LV_HAVE_AVX2
+  return 2;
+#else
+  return 1;
+#endif
+}
+
 void srslte_tdec_decision(srslte_tdec_t * h, uint8_t *output, uint32_t long_cb) {
-  uint8_t *output_par[SRSLTE_TDEC_NPAR]; 
+  uint8_t *output_par[SRSLTE_TDEC_MAX_NPAR];
   output_par[0] = output; 
   srslte_tdec_decision_par(h, output_par, long_cb);
 }
 
-void srslte_tdec_decision_byte_par(srslte_tdec_t * h, uint8_t *output[SRSLTE_TDEC_NPAR], uint32_t long_cb) {
+void srslte_tdec_decision_byte_par(srslte_tdec_t * h, uint8_t *output[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb) {
 #ifdef LV_HAVE_SSE
   srslte_tdec_simd_decision_byte(&h->tdec_simd, output, long_cb);  
 #else
@@ -137,13 +145,13 @@ void srslte_tdec_decision_byte_par_cb(srslte_tdec_t * h, uint8_t *output, uint32
 }
 
 void srslte_tdec_decision_byte(srslte_tdec_t * h, uint8_t *output, uint32_t long_cb) {
-  uint8_t *output_par[SRSLTE_TDEC_NPAR]; 
+  uint8_t *output_par[SRSLTE_TDEC_MAX_NPAR];
   output_par[0] = output; 
   srslte_tdec_decision_byte_par(h, output_par, long_cb);
 }
 
-int srslte_tdec_run_all_par(srslte_tdec_t * h, int16_t * input[SRSLTE_TDEC_NPAR], 
-                            uint8_t *output[SRSLTE_TDEC_NPAR], 
+int srslte_tdec_run_all_par(srslte_tdec_t * h, int16_t * input[SRSLTE_TDEC_MAX_NPAR],
+                            uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                             uint32_t nof_iterations, uint32_t long_cb) {
 #ifdef LV_HAVE_SSE
   return srslte_tdec_simd_run_all(&h->tdec_simd, input, output, nof_iterations, long_cb);  
@@ -155,9 +163,9 @@ int srslte_tdec_run_all_par(srslte_tdec_t * h, int16_t * input[SRSLTE_TDEC_NPAR]
 
 int srslte_tdec_run_all(srslte_tdec_t * h, int16_t * input, uint8_t *output, uint32_t nof_iterations, uint32_t long_cb)
 {
-  uint8_t *output_par[SRSLTE_TDEC_NPAR]; 
+  uint8_t *output_par[SRSLTE_TDEC_MAX_NPAR];
   output_par[0] = output;   
-  int16_t *input_par[SRSLTE_TDEC_NPAR]; 
+  int16_t *input_par[SRSLTE_TDEC_MAX_NPAR];
   input_par[0] = input; 
  
   return srslte_tdec_run_all_par(h, input_par, output_par, nof_iterations, long_cb);
diff --git a/lib/src/phy/fec/turbodecoder_avx.c b/lib/src/phy/fec/turbodecoder_avx.c
index 2a2f6f925..2e877cbde 100644
--- a/lib/src/phy/fec/turbodecoder_avx.c
+++ b/lib/src/phy/fec/turbodecoder_avx.c
@@ -81,7 +81,7 @@ static inline int16_t hMax1(__m256i masked_value)
 }
 
 /* Computes beta values */
-void map_avx_beta(map_gen_t * s, int16_t * output[SRSLTE_TDEC_NPAR], uint32_t long_cb)
+void map_avx_beta(map_gen_t * s, int16_t * output[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb)
 {
   int k;
   uint32_t end = long_cb + 3;
diff --git a/lib/src/phy/fec/turbodecoder_simd.c b/lib/src/phy/fec/turbodecoder_simd.c
index e245c84a4..a32d52962 100644
--- a/lib/src/phy/fec/turbodecoder_simd.c
+++ b/lib/src/phy/fec/turbodecoder_simd.c
@@ -54,13 +54,13 @@ void map_sse_alpha(map_gen_t * s, uint32_t long_cb);
 void map_sse_gamma(map_gen_t * h, int16_t *input, int16_t *app, int16_t *parity, uint32_t long_cb);
 
 #ifdef LV_HAVE_AVX2
-void map_avx_beta(map_gen_t * s, int16_t * output[SRSLTE_TDEC_NPAR], uint32_t long_cb);
+void map_avx_beta(map_gen_t * s, int16_t * output[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb);
 void map_avx_alpha(map_gen_t * s, uint32_t long_cb);
 void map_avx_gamma(map_gen_t * h, int16_t *input, int16_t *app, int16_t *parity, uint32_t cbidx, uint32_t long_cb);
 #endif
 
 
-void map_simd_beta(map_gen_t * s, int16_t * output[SRSLTE_TDEC_NPAR], uint32_t nof_cb, uint32_t long_cb)
+void map_simd_beta(map_gen_t * s, int16_t * output[SRSLTE_TDEC_MAX_NPAR], uint32_t nof_cb, uint32_t long_cb)
 {
   if (nof_cb == 1) {
     map_sse_beta(s, output[0], long_cb);
@@ -128,12 +128,12 @@ void map_simd_free(map_gen_t * h)
 }
 
 /* Runs one instance of a decoder */
-void map_simd_dec(map_gen_t * h, int16_t * input[SRSLTE_TDEC_NPAR], int16_t *app[SRSLTE_TDEC_NPAR], int16_t * parity[SRSLTE_TDEC_NPAR], 
-                  int16_t *output[SRSLTE_TDEC_NPAR], uint32_t cb_mask, uint32_t long_cb)
+void map_simd_dec(map_gen_t * h, int16_t * input[SRSLTE_TDEC_MAX_NPAR], int16_t *app[SRSLTE_TDEC_MAX_NPAR], int16_t * parity[SRSLTE_TDEC_MAX_NPAR],
+                  int16_t *output[SRSLTE_TDEC_MAX_NPAR], uint32_t cb_mask, uint32_t long_cb)
 {
   
   uint32_t nof_cb = 1;
-  int16_t *outptr[SRSLTE_TDEC_NPAR];
+  int16_t *outptr[SRSLTE_TDEC_MAX_NPAR];
   
   // Compute branch metrics
   switch(cb_mask) {
@@ -354,21 +354,21 @@ void deinterleave_input_simd(srslte_tdec_simd_t *h, int16_t *input, uint32_t cbi
 }
 
 /* Runs 1 turbo decoder iteration */
-void srslte_tdec_simd_iteration(srslte_tdec_simd_t * h, int16_t * input[SRSLTE_TDEC_NPAR], uint32_t long_cb)
+void srslte_tdec_simd_iteration(srslte_tdec_simd_t * h, int16_t * input[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb)
 {
 
-  int16_t *tmp_app[SRSLTE_TDEC_NPAR];
+  int16_t *tmp_app[SRSLTE_TDEC_MAX_NPAR];
 
   if (h->current_cbidx >= 0) {
     uint16_t *inter   = h->interleaver[h->current_cbidx].forward;
     uint16_t *deinter = h->interleaver[h->current_cbidx].reverse;
     
-#if SRSLTE_TDEC_NPAR == 2
-    h->cb_mask = (input[0]?1:0) | (input[1]?2:0);
-#else
-    h->cb_mask = input[0]?1:0;
+#ifndef LV_HAVE_AVX2
+    input[1] = NULL;
 #endif
-    
+
+    h->cb_mask = (input[0]?1:0) | (input[1]?2:0);
+
     for (int i=0;i<h->max_par_cb;i++) {
       if (h->n_iter[i] == 0 && input[i]) {
         //printf("deinterleaveing %d\n",i);
@@ -484,7 +484,7 @@ void tdec_simd_decision(srslte_tdec_simd_t * h, uint8_t *output, uint32_t cbidx,
   }
 }
 
-void srslte_tdec_simd_decision(srslte_tdec_simd_t * h, uint8_t *output[SRSLTE_TDEC_NPAR], uint32_t long_cb)
+void srslte_tdec_simd_decision(srslte_tdec_simd_t * h, uint8_t *output[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb)
 {
   for (int i=0;i<h->max_par_cb;i++) {    
     tdec_simd_decision(h, output[i], i, long_cb);
@@ -510,7 +510,7 @@ void srslte_tdec_simd_decision_byte_cb(srslte_tdec_simd_t * h, uint8_t *output,
   }
 }
 
-void srslte_tdec_simd_decision_byte(srslte_tdec_simd_t * h, uint8_t *output[SRSLTE_TDEC_NPAR], uint32_t long_cb)
+void srslte_tdec_simd_decision_byte(srslte_tdec_simd_t * h, uint8_t *output[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb)
 {
   for (int i=0;i<h->max_par_cb;i++) {
     srslte_tdec_simd_decision_byte_cb(h, output[i], i, long_cb);
@@ -519,7 +519,7 @@ void srslte_tdec_simd_decision_byte(srslte_tdec_simd_t * h, uint8_t *output[SRSL
 
 
 /* Runs nof_iterations iterations and decides the output bits */
-int srslte_tdec_simd_run_all(srslte_tdec_simd_t * h, int16_t * input[SRSLTE_TDEC_NPAR], uint8_t *output[SRSLTE_TDEC_NPAR],
+int srslte_tdec_simd_run_all(srslte_tdec_simd_t * h, int16_t * input[SRSLTE_TDEC_MAX_NPAR], uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                             uint32_t nof_iterations, uint32_t long_cb)
 {
   if (srslte_tdec_simd_reset(h, long_cb)) {
diff --git a/lib/src/phy/fec/turbodecoder_simd_inter.c b/lib/src/phy/fec/turbodecoder_simd_inter.c
index 05d8b2cf5..3c04e2136 100644
--- a/lib/src/phy/fec/turbodecoder_simd_inter.c
+++ b/lib/src/phy/fec/turbodecoder_simd_inter.c
@@ -172,7 +172,7 @@ void extract_input(srslte_tdec_simd_inter_t *h, int16_t *input, uint32_t cbidx,
   }
 }
 
-void srslte_tdec_simd_inter_iteration(srslte_tdec_simd_inter_t * h, int16_t *input[SRSLTE_TDEC_NPAR], uint32_t nof_cb, uint32_t long_cb)
+void srslte_tdec_simd_inter_iteration(srslte_tdec_simd_inter_t * h, int16_t *input[SRSLTE_TDEC_MAX_NPAR], uint32_t nof_cb, uint32_t long_cb)
 {
 
   if (h->current_cbidx >= 0) {
@@ -239,7 +239,7 @@ void srslte_tdec_simd_inter_decision_cb(srslte_tdec_simd_inter_t * h, uint8_t *o
   }
 }
 
-void srslte_tdec_simd_inter_decision(srslte_tdec_simd_inter_t * h, uint8_t *output[SRSLTE_TDEC_NPAR], uint32_t nof_cb, uint32_t long_cb)
+void srslte_tdec_simd_inter_decision(srslte_tdec_simd_inter_t * h, uint8_t *output[SRSLTE_TDEC_MAX_NPAR], uint32_t nof_cb, uint32_t long_cb)
 {
   for (int i=0;i<nof_cb;i++) {
     srslte_tdec_simd_inter_decision_cb(h, output[i], i, long_cb);
@@ -269,7 +269,7 @@ void srslte_tdec_simd_inter_decision_byte_cb(srslte_tdec_simd_inter_t * h, uint8
   }
 }
 
-void srslte_tdec_simd_inter_decision_byte(srslte_tdec_simd_inter_t * h, uint8_t *output[SRSLTE_TDEC_NPAR], uint32_t nof_cb, uint32_t long_cb)
+void srslte_tdec_simd_inter_decision_byte(srslte_tdec_simd_inter_t * h, uint8_t *output[SRSLTE_TDEC_MAX_NPAR], uint32_t nof_cb, uint32_t long_cb)
 {
   for (int i=0;i<nof_cb;i++) {
     srslte_tdec_simd_inter_decision_byte_cb(h, output[i], i, long_cb);
@@ -277,7 +277,7 @@ void srslte_tdec_simd_inter_decision_byte(srslte_tdec_simd_inter_t * h, uint8_t
 }
 
 int srslte_tdec_simd_inter_run_all(srslte_tdec_simd_inter_t * h, 
-                                  int16_t *input[SRSLTE_TDEC_NPAR], uint8_t *output[SRSLTE_TDEC_NPAR],
+                                  int16_t *input[SRSLTE_TDEC_MAX_NPAR], uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
                                   uint32_t nof_iterations, uint32_t nof_cb, uint32_t long_cb)
 {
   uint32_t iter = 0;
diff --git a/lib/src/phy/phch/pdsch.c b/lib/src/phy/phch/pdsch.c
index bdb0a755a..63bfa222d 100644
--- a/lib/src/phy/phch/pdsch.c
+++ b/lib/src/phy/phch/pdsch.c
@@ -600,6 +600,8 @@ static int srslte_pdsch_codeword_decode(srslte_pdsch_t *q, srslte_pdsch_cfg_t *c
     /* Bit scrambling */
     srslte_scrambling_s_offset(seq, q->e[codeword_idx], 0, nbits->nof_bits);
 
+    printf("e: "); srslte_vec_fprint_s(stdout, q->e[codeword_idx], 10);
+
     /* Return  */
     ret = srslte_dlsch_decode2(&q->dl_sch, cfg, softbuffer, q->e[codeword_idx], data, codeword_idx);
 
@@ -685,7 +687,6 @@ int srslte_pdsch_decode(srslte_pdsch_t *q,
         srslte_layerdemap_type(x, q->d, cfg->nof_layers, nof_tb,
                              nof_symbols[0], nof_symbols, cfg->mimo_type);
     }
-
     // Codeword decoding
     for (uint32_t tb = 0; tb < SRSLTE_MAX_CODEWORDS; tb ++) {
       /* Decode only if transport block is enabled and the default ACK is not true */
diff --git a/lib/src/phy/phch/sch.c b/lib/src/phy/phch/sch.c
index b7909192f..3921c8562 100644
--- a/lib/src/phy/phch/sch.c
+++ b/lib/src/phy/phch/sch.c
@@ -310,8 +310,8 @@ bool decode_tb_cb(srslte_sch_t *q,
 
   bool cb_map[SRSLTE_MAX_CODEBLOCKS];
     
-  uint32_t cb_idx[SRSLTE_TDEC_NPAR];
-  int16_t *decoder_input[SRSLTE_TDEC_NPAR];
+  uint32_t cb_idx[SRSLTE_TDEC_MAX_NPAR];
+  int16_t *decoder_input[SRSLTE_TDEC_MAX_NPAR];
   
   uint32_t nof_cb     = cb_size_group?cb_segm->C2:cb_segm->C1;
   uint32_t first_cb   = cb_size_group?cb_segm->C1:0;
@@ -328,7 +328,7 @@ bool decode_tb_cb(srslte_sch_t *q,
     return false; 
   }
   
-  for (int i=0;i<SRSLTE_TDEC_NPAR;i++) {
+  for (int i=0;i<srslte_tdec_get_nof_parallel(&q->decoder);i++) {
     cb_idx[i]        = i+first_cb; 
     decoder_input[i] = NULL;
   }
@@ -346,7 +346,7 @@ bool decode_tb_cb(srslte_sch_t *q,
   while(remaining_cb>0) {
         
     // Unratematch the codeblocks left to decode 
-    for (int i=0;i<SRSLTE_TDEC_NPAR;i++) {
+    for (int i=0;i<srslte_tdec_get_nof_parallel(&q->decoder);i++) {
       
       if (!decoder_input[i] && remaining_cb > 0) {        
         // Find an unprocessed CB 
@@ -364,14 +364,17 @@ bool decode_tb_cb(srslte_sch_t *q,
             n_e2 = n_e+Qm;
             rp   = (cb_segm->C - gamma)*n_e + (cb_idx[i]-(cb_segm->C - gamma))*n_e2;
           }
-                
+
           INFO("CB %d: rp=%d, n_e=%d, i=%d\n", cb_idx[i], rp, n_e2, i);
           if (srslte_rm_turbo_rx_lut(&e_bits[rp], softbuffer->buffer_f[cb_idx[i]], n_e2, cb_len_idx, rv)) {
             fprintf(stderr, "Error in rate matching\n");
             return SRSLTE_ERROR;
           }
 
-          decoder_input[i] = softbuffer->buffer_f[cb_idx[i]];          
+          decoder_input[i] = softbuffer->buffer_f[cb_idx[i]];
+
+          printf("input: "); srslte_vec_fprint_s(stdout, decoder_input[i], 10);
+
         }
       }
     }
@@ -380,7 +383,7 @@ bool decode_tb_cb(srslte_sch_t *q,
     srslte_tdec_iteration_par(&q->decoder, decoder_input, cb_len);
 
     // Decide output bits and compute CRC 
-    for (int i=0;i<SRSLTE_TDEC_NPAR;i++) {
+    for (int i=0;i<srslte_tdec_get_nof_parallel(&q->decoder);i++) {
       if (decoder_input[i]) {        
         srslte_tdec_decision_byte_par_cb(&q->decoder, q->cb_in, i, cb_len);
 
@@ -395,6 +398,8 @@ bool decode_tb_cb(srslte_sch_t *q,
           crc_ptr = &q->crc_tb; 
         }
 
+        printf("output: %d", i); srslte_vec_fprint_b(stdout, q->cb_in, 10);
+
         // CRC is OK
         if (!srslte_crc_checksum_byte(crc_ptr, q->cb_in, len_crc)) {
 
@@ -474,7 +479,7 @@ static int decode_tb(srslte_sch_t *q,
     data[cb_segm->tbs/8+1] = 0; 
     data[cb_segm->tbs/8+2] = 0; 
     
-    // Process Codeblocks in groups of equal CB size to parallelize according to SRSLTE_TDEC_NPAR
+    // Process Codeblocks in groups of equal CB size to parallelize according to SRSLTE_TDEC_MAX_NPAR
     for (uint32_t i=0;i<nof_cb_groups && crc_ok;i++) {
       crc_ok = decode_tb_cb(q, softbuffer, cb_segm, Qm, rv, nof_e_bits, e_bits, data, i);            
     }
diff --git a/lib/src/phy/phch/test/pusch_test.c b/lib/src/phy/phch/test/pusch_test.c
index ece098dcf..9d048e680 100644
--- a/lib/src/phy/phch/test/pusch_test.c
+++ b/lib/src/phy/phch/test/pusch_test.c
@@ -190,7 +190,7 @@ int main(int argc, char **argv) {
   bzero(&uci_data_tx, sizeof(srslte_uci_data_t));
   uci_data_tx.uci_cqi_len = 4; 
   uci_data_tx.uci_ri_len = 0; 
-  uci_data_tx.uci_ack_len = 2;
+  uci_data_tx.uci_ack_len = 1;
   memcpy(&uci_data_rx, &uci_data_tx, sizeof(srslte_uci_data_t));
     
   for (uint32_t i=0;i<20;i++) {