New optimization on the PHY for both UE and eNodeB (#251)

* New parallel Turbodecoder implementation in SSE/AVX 16-bit and 8-bit

* Optimised UL Interleaver

* Include TB CRC calculation in FEC encoder

* New threading priorities

CMakeLists.txt

@@ -279,8 +279,8 @@ if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang")
       set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -ggdb -DBUILD_TYPE_RELWITHDEBINFO")
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -ggdb -DBUILD_TYPE_RELWITHDEBINFO")
     else(${CMAKE_BUILD_TYPE} STREQUAL "RelWithDebInfo")
-      set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -O3 -DBUILD_TYPE_RELEASE")
+      set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -O3 -fno-trapping-math -fno-math-errno -DBUILD_TYPE_RELEASE")
-      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -DBUILD_TYPE_RELEASE")
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3 -fno-trapping-math -fno-math-errno -DBUILD_TYPE_RELEASE")
     endif(${CMAKE_BUILD_TYPE} STREQUAL "RelWithDebInfo")
   endif(${CMAKE_BUILD_TYPE} STREQUAL "Debug")
 
@@ -325,9 +325,9 @@ if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang")
   endif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm")
   set(CMAKE_REQUIRED_FLAGS ${CMAKE_C_FLAGS})
 
- if(NOT HAVE_SSE AND NOT HAVE_NEON AND NOT DISABLE_SIMD)
+  if(NOT HAVE_SSE AND NOT HAVE_NEON AND NOT DISABLE_SIMD)
-  message(FATAL_ERROR "no SIMD instructions found")
+    message(FATAL_ERROR "no SIMD instructions found")
- endif(NOT HAVE_SSE AND NOT HAVE_NEON AND NOT DISABLE_SIMD)
+  endif(NOT HAVE_SSE AND NOT HAVE_NEON AND NOT DISABLE_SIMD)
 
   if(NOT WIN32)
       ADD_CXX_COMPILER_FLAG_IF_AVAILABLE(-fvisibility=hidden HAVE_VISIBILITY_HIDDEN)

lib/include/srslte/config.h

@@ -52,6 +52,14 @@
 #endif
 
 
+// Useful macros for templates
+#define CONCAT(a, b) a##b
+#define CONCAT2(a, b) CONCAT(a,b)
+
+#define STRING2(x) #x
+#define STRING(x) STRING2(x)
+
+
 // Common error codes
 #define SRSLTE_SUCCESS                0
 #define SRSLTE_ERROR                  -1

lib/include/srslte/interfaces/ue_interfaces.h

@@ -545,6 +545,7 @@ typedef struct {
   bool sic_pss_enabled;
   float rx_gain_offset;
   bool pdsch_csi_enabled;
+  bool pdsch_8bit_decoder;
   uint32_t intra_freq_meas_len_ms;
   uint32_t intra_freq_meas_period_ms;
 } phy_args_t; 

lib/include/srslte/phy/common/sequence.h

@@ -44,6 +44,7 @@ typedef struct SRSLTE_API {
   uint8_t *c_bytes;
   float *c_float;
   short *c_short;
+  int8_t *c_char;
   uint32_t cur_len;
   uint32_t max_len;
 } srslte_sequence_t;

lib/include/srslte/phy/fec/rm_turbo.h

@@ -36,6 +36,7 @@
 #define SRSLTE_RM_TURBO_H
 
 #include "srslte/config.h"
+#include "srslte/phy/fec/turbodecoder.h"
 
 #ifndef SRSLTE_RX_NULL
 #define SRSLTE_RX_NULL 10000
@@ -47,7 +48,6 @@
 
 #include "srslte/config.h"
 
-
 SRSLTE_API int srslte_rm_turbo_tx(uint8_t *w_buff,
                                   uint32_t buff_len, 
                                   uint8_t *input, 
@@ -84,5 +84,17 @@ SRSLTE_API int srslte_rm_turbo_rx_lut(int16_t *input,
                                       uint32_t cb_idx, 
                                       uint32_t rv_idx);
 
+SRSLTE_API int srslte_rm_turbo_rx_lut_(int16_t *input,
+                                       int16_t *output,
+                                       uint32_t in_len,
+                                       uint32_t cb_idx,
+                                       uint32_t rv_idx,
+                                       bool enable_input_tdec);
+
+SRSLTE_API int srslte_rm_turbo_rx_lut_8bit(int8_t *input,
+                                           int8_t *output,
+                                           uint32_t in_len,
+                                           uint32_t cb_idx,
+                                           uint32_t rv_idx);
 
 #endif // SRSLTE_RM_TURBO_H

lib/include/srslte/phy/fec/tc_interl.h

@@ -47,8 +47,9 @@ typedef struct SRSLTE_API {
 SRSLTE_API int srslte_tc_interl_LTE_gen(srslte_tc_interl_t *h, 
                                         uint32_t long_cb);
 
-SRSLTE_API int srslte_tc_interl_UMTS_gen(srslte_tc_interl_t *h, 
+SRSLTE_API int srslte_tc_interl_LTE_gen_interl(srslte_tc_interl_t *h,
-                                         uint32_t long_cb);
+                                               uint32_t long_cb,
+                                               uint32_t interl_win);
 
 SRSLTE_API int srslte_tc_interl_init(srslte_tc_interl_t *h, 
                                      uint32_t max_long_cb);

lib/include/srslte/phy/fec/turbocoder.h

@@ -70,10 +70,12 @@ SRSLTE_API int srslte_tcod_encode(srslte_tcod_t *h,
                                   uint32_t long_cb);
 
 SRSLTE_API int srslte_tcod_encode_lut(srslte_tcod_t *h,
-                                      srslte_crc_t *crc,
+                                      srslte_crc_t *crc_tb,
+                                      srslte_crc_t *crc_cb,
                                       uint8_t *input,
                                       uint8_t *parity, 
-                                      uint32_t cblen_idx); 
+                                      uint32_t cblen_idx,
+                                      bool last_cb);
 
 SRSLTE_API void srslte_tcod_gentable(); 
 

lib/include/srslte/phy/fec/turbodecoder.h

@@ -47,46 +47,81 @@
 #define SRSLTE_TCOD_TOTALTAIL 12
 
 #define SRSLTE_TCOD_MAX_LEN_CB     6144
-#define SRSLTE_TCOD_MAX_LEN_CODED  (SRSLTE_TCOD_RATE*SRSLTE_TCOD_MAX_LEN_CB+SRSLTE_TCOD_TOTALTAIL)
 
-#include "srslte/phy/fec/turbodecoder_gen.h"
+// Expect the input to be aligned for sub-block window processing.
-#include "srslte/phy/fec/turbodecoder_simd.h"
+#define SRSLTE_TDEC_EXPECT_INPUT_SB 1
+
+// Include interfaces for 8 and 16 bit decoder implementations
+#define LLR_IS_8BIT
+#include "srslte/phy/fec/turbodecoder_impl.h"
+#undef LLR_IS_8BIT
+
+#define LLR_IS_16BIT
+#include "srslte/phy/fec/turbodecoder_impl.h"
+#undef LLR_IS_16BIT
+
+#define SRSLTE_TDEC_NOF_AUTO_MODES_8  2
+#define SRSLTE_TDEC_NOF_AUTO_MODES_16 3
+
+typedef enum {SRSLTE_TDEC_8, SRSLTE_TDEC_16} srslte_tdec_llr_type_t;
 
 typedef struct SRSLTE_API {
-  float *input_conv;
+  uint32_t max_long_cb;
-  union {
+
-    srslte_tdec_simd_t tdec_simd;
+  void *dec8_hdlr[SRSLTE_TDEC_NOF_AUTO_MODES_8];
-    srslte_tdec_gen_t  tdec_gen;
+  void *dec16_hdlr[SRSLTE_TDEC_NOF_AUTO_MODES_16];
-  };
+  srslte_tdec_8bit_impl_t *dec8[SRSLTE_TDEC_NOF_AUTO_MODES_8];
+  srslte_tdec_16bit_impl_t *dec16[SRSLTE_TDEC_NOF_AUTO_MODES_16];
+  int nof_blocks8[SRSLTE_TDEC_NOF_AUTO_MODES_8];
+  int nof_blocks16[SRSLTE_TDEC_NOF_AUTO_MODES_16];
+
+  // Declare as void types as can be int8 or int16
+  void *app1;
+  void *app2;
+  void *ext1;
+  void *ext2;
+  void *syst0;
+  void *parity0;
+  void *parity1;
+
+  void *input_conv;
+
+  bool force_not_sb;
+
+  srslte_tdec_impl_type_t dec_type;
+
+  srslte_tdec_llr_type_t current_llr_type;
+  uint32_t current_dec;
+  uint32_t current_long_cb;
+  uint32_t current_inter_idx;
+  int current_cbidx;
+  srslte_tc_interl_t interleaver[4][SRSLTE_NOF_TC_CB_SIZES];
+  int n_iter;
 } srslte_tdec_t;
 
 SRSLTE_API int srslte_tdec_init(srslte_tdec_t * h, 
                                 uint32_t max_long_cb);
 
+SRSLTE_API int srslte_tdec_init_manual(srslte_tdec_t * h,
+                                      uint32_t max_long_cb,
+                                      srslte_tdec_impl_type_t dec_type);
+
 SRSLTE_API void srslte_tdec_free(srslte_tdec_t * h);
 
-SRSLTE_API int srslte_tdec_reset(srslte_tdec_t * h, 
+SRSLTE_API void srslte_tdec_force_not_sb(srslte_tdec_t *h);
+
+SRSLTE_API int srslte_tdec_new_cb(srslte_tdec_t * h,
                                  uint32_t long_cb);
 
-SRSLTE_API int srslte_tdec_reset_cb(srslte_tdec_t * h, 
+SRSLTE_API int srslte_tdec_get_nof_iterations(srslte_tdec_t * h);
-                                    uint32_t cb_idx);
 
-SRSLTE_API int srslte_tdec_get_nof_iterations_cb(srslte_tdec_t * h, 
+SRSLTE_API uint32_t srslte_tdec_autoimp_get_subblocks(uint32_t long_cb);
-                                                 uint32_t cb_idx);
 
-SRSLTE_API uint32_t srslte_tdec_get_nof_parallel(srslte_tdec_t * h);
+SRSLTE_API uint32_t srslte_tdec_autoimp_get_subblocks_8bit(uint32_t long_cb);
 
 SRSLTE_API void srslte_tdec_iteration(srslte_tdec_t * h, 
                                       int16_t* input,
-                                      uint32_t long_cb);
+                                      uint8_t *output);
-
-SRSLTE_API void srslte_tdec_decision(srslte_tdec_t * h, 
-                                     uint8_t *output, 
-                                     uint32_t long_cb);
-
-SRSLTE_API void srslte_tdec_decision_byte(srslte_tdec_t * h, 
-                                          uint8_t *output, 
-                                          uint32_t long_cb); 
 
 SRSLTE_API int srslte_tdec_run_all(srslte_tdec_t * h,
                                    int16_t * input,
@@ -94,27 +129,15 @@ SRSLTE_API int srslte_tdec_run_all(srslte_tdec_t * h,
                                    uint32_t nof_iterations, 
                                    uint32_t long_cb);
 
-SRSLTE_API void srslte_tdec_iteration_par(srslte_tdec_t * h, 
+SRSLTE_API void srslte_tdec_iteration_8bit(srslte_tdec_t * h,
-                                          int16_t* input[SRSLTE_TDEC_MAX_NPAR],
+                                            int8_t* input,
-                                          uint32_t long_cb);
+                                            uint8_t *output);
 
-SRSLTE_API void srslte_tdec_decision_par(srslte_tdec_t * h, 
+SRSLTE_API int srslte_tdec_run_all_8bit(srslte_tdec_t * h,
-                                         uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
+                                         int8_t * input,
+                                         uint8_t *output,
+                                         uint32_t nof_iterations,
                                          uint32_t long_cb);
 
-SRSLTE_API void srslte_tdec_decision_byte_par(srslte_tdec_t * h, 
-                                              uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
-                                              uint32_t long_cb); 
-
-SRSLTE_API void srslte_tdec_decision_byte_par_cb(srslte_tdec_t * h, 
-                                                 uint8_t *output, 
-                                                 uint32_t cb_idx, 
-                                                 uint32_t long_cb);
-
-SRSLTE_API 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);
 
 #endif // SRSLTE_TURBODECODER_H

lib/include/srslte/phy/fec/turbodecoder_gen.h

@@ -47,53 +47,16 @@
 #define SRSLTE_TCOD_TOTALTAIL 12
 
 #define SRSLTE_TCOD_MAX_LEN_CB     6144
-#define SRSLTE_TCOD_MAX_LEN_CODED  (SRSLTE_TCOD_RATE*SRSLTE_TCOD_MAX_LEN_CB+SRSLTE_TCOD_TOTALTAIL)
 
 typedef struct SRSLTE_API {
-  int max_long_cb;
+  uint32_t max_long_cb;
-  float *beta;
+  int16_t *beta;
-} srslte_map_gen_vl_t;
+} tdec_gen_t;
-
+
-typedef struct SRSLTE_API {
+int tdec_gen_init(void **h, uint32_t max_long_cb);
-  int max_long_cb;
+void tdec_gen_free(void *h);
-
+void tdec_gen_dec(void *h, int16_t * input, int16_t *app, int16_t * parity, int16_t *output, uint32_t long_cb);
-  srslte_map_gen_vl_t dec;
+void tdec_gen_extract_input(int16_t *input, int16_t *syst, int16_t *parity0, int16_t *parity1, int16_t *app2, uint32_t long_cb);
-
+void tdec_gen_decision_byte(int16_t *app1, uint8_t *output, uint32_t long_cb);
-  float *llr1;
-  float *llr2;
-  float *w;
-  float *syst;
-  float *parity;
-
-  int current_cbidx; 
-  uint32_t current_cb_len;
-  uint32_t n_iter; 
-  srslte_tc_interl_t interleaver[SRSLTE_NOF_TC_CB_SIZES];
-} srslte_tdec_gen_t;
-
-SRSLTE_API int srslte_tdec_gen_init(srslte_tdec_gen_t * h, 
-                                uint32_t max_long_cb);
-
-SRSLTE_API void srslte_tdec_gen_free(srslte_tdec_gen_t * h);
-
-SRSLTE_API int srslte_tdec_gen_reset(srslte_tdec_gen_t * h, uint32_t long_cb);
-
-SRSLTE_API void srslte_tdec_gen_iteration(srslte_tdec_gen_t * h, 
-                                      float * input, 
-                                      uint32_t long_cb);
-
-SRSLTE_API void srslte_tdec_gen_decision(srslte_tdec_gen_t * h, 
-                                     uint8_t *output, 
-                                     uint32_t long_cb);
-
-SRSLTE_API void srslte_tdec_gen_decision_byte(srslte_tdec_gen_t * h, 
-                                          uint8_t *output, 
-                                          uint32_t long_cb); 
-
-SRSLTE_API int srslte_tdec_gen_run_all(srslte_tdec_gen_t * h, 
-                                   float * input, 
-                                   uint8_t *output,
-                                   uint32_t nof_iterations, 
-                                   uint32_t long_cb);
 
 #endif // SRSLTE_TURBODECODER_GEN_H

lib/include/srslte/phy/fec/turbodecoder_impl.h

@@ -0,0 +1,68 @@
+/**
+ *
+ * \section COPYRIGHT
+ *
+ * Copyright 2013-2015 Software Radio Systems Limited
+ *
+ * \section LICENSE
+ *
+ * This file is part of the srsLTE library.
+ *
+ * srsLTE is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as
+ * published by the Free Software Foundation, either version 3 of
+ * the License, or (at your option) any later version.
+ *
+ * srsLTE is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * A copy of the GNU Affero General Public License can be found in
+ * the LICENSE file in the top-level directory of this distribution
+ * and at http://www.gnu.org/licenses/.
+ *
+ */
+
+#ifndef SRSLTE_TURBODECODER_IMPL_H
+#define SRSLTE_TURBODECODER_IMPL_H
+
+#include "srslte/config.h"
+
+/* Interface for internal decoder implementation */
+typedef enum SRSLTE_API {
+  SRSLTE_TDEC_AUTO = 0,
+  SRSLTE_TDEC_GENERIC,
+  SRSLTE_TDEC_SSE,
+  SRSLTE_TDEC_SSE_WINDOW,
+  SRSLTE_TDEC_AVX_WINDOW,
+  SRSLTE_TDEC_SSE8_WINDOW,
+  SRSLTE_TDEC_AVX8_WINDOW,
+  SRSLTE_TDEC_NOF_IMP
+} srslte_tdec_impl_type_t;
+
+#endif
+
+#ifdef LLR_IS_8BIT
+#define llr_t int8_t
+#define type_name srslte_tdec_8bit_impl_t
+#else
+  #ifdef LLR_IS_16BIT
+  #define llr_t int16_t
+#define type_name srslte_tdec_16bit_impl_t
+  #else
+  #error "Unsupported LLR mode"
+  #endif
+#endif
+
+
+typedef struct SRSLTE_API {
+  int (*tdec_init)(void **h, uint32_t max_long_cb);
+  void (*tdec_free)(void *h);
+  void (*tdec_dec)(void *h, llr_t * input, llr_t *app, llr_t * parity, llr_t *output, uint32_t long_cb);
+  void (*tdec_extract_input)(llr_t *input, llr_t *syst, llr_t *parity0, llr_t *parity1, llr_t *app2, uint32_t long_cb);
+  void (*tdec_decision_byte)(llr_t *app1, uint8_t *output, uint32_t long_cb);
+} type_name;
+
+#undef llr_t
+#undef type_name

lib/include/srslte/phy/fec/turbodecoder_iter.h

@@ -0,0 +1,158 @@
+/**
+ *
+ * \section COPYRIGHT
+ *
+ * Copyright 2013-2015 Software Radio Systems Limited
+ *
+ * \section LICENSE
+ *
+ * This file is part of the srsLTE library.
+ *
+ * srsLTE is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as
+ * published by the Free Software Foundation, either version 3 of
+ * the License, or (at your option) any later version.
+ *
+ * srsLTE is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * A copy of the GNU Affero General Public License can be found in
+ * the LICENSE file in the top-level directory of this distribution
+ * and at http://www.gnu.org/licenses/.
+ *
+ */
+
+#include "srslte/config.h"
+
+#define MAKE_CALL(a)  CONCAT2(a,type_name)
+#define MAKE_VEC(a)   CONCAT2(a,vec_suffix)
+#define PRINT         CONCAT2(srslte_vec_fprint,print_suffix)
+
+#ifdef LLR_IS_8BIT
+#define llr_t int8_t
+#define type_name _8bit
+#define vec_suffix _bbb
+#define print_suffix _bs
+#define decptr h->dec8[h->current_dec]
+#define dechdlr h->dec8_hdlr[h->current_dec]
+#define input_is_interleaved 1
+#else
+#ifdef LLR_IS_16BIT
+  #define llr_t int16_t
+  #define vec_suffix _sss
+  #define print_suffix _s
+  #define decptr h->dec16[h->current_dec]
+  #define dechdlr h->dec16_hdlr[h->current_dec]
+  #define input_is_interleaved (h->current_dec > 0)
+#define type_name _16bit
+  #else
+  #warning "Unsupported LLR mode"
+  #endif
+#endif
+
+#define debug_enabled_iter 0
+#define debug_len     20 
+
+#define debug_vec(a) if (debug_enabled_iter) {printf("%s it=%d: ", STRING(a), n_iter);PRINT(stdout, a, debug_len);}
+
+
+static void MAKE_CALL(extract_input_tail_sb)(llr_t *input, llr_t *syst, llr_t *app2, llr_t *parity0, llr_t *parity1, uint32_t long_cb)
+{
+  for (int i = long_cb; i < long_cb + 3; i++) {
+    syst[i]    = input[3*(long_cb+32) + 2*(i - long_cb)];
+    parity0[i] = input[3*(long_cb+32)+ 2*(i - long_cb) + 1];
+
+    app2[i]    = input[3*(long_cb+32) + 6 + 2*(i - long_cb)];
+    parity1[i] = input[3*(long_cb+32) + 6 + 2*(i - long_cb) + 1];
+  }
+}
+
+/* Runs 1 turbo decoder iteration */
+void MAKE_CALL(run_tdec_iteration)(srslte_tdec_t * h, llr_t * input)
+{
+
+  if (h->current_cbidx >= 0) {
+    uint16_t *inter   = h->interleaver[h->current_inter_idx][h->current_cbidx].forward;
+    uint16_t *deinter = h->interleaver[h->current_inter_idx][h->current_cbidx].reverse;
+    llr_t *syst    = (llr_t*) h->syst0;
+    llr_t *parity0 = (llr_t*) h->parity0;
+    llr_t *parity1 = (llr_t*) h->parity1;
+
+    llr_t *app1 = (llr_t*) h->app1;
+    llr_t *app2 = (llr_t*) h->app2;
+    llr_t *ext1 = (llr_t*) h->ext1;
+    llr_t *ext2 = (llr_t*) h->ext2;
+
+    uint32_t long_cb = h->current_long_cb;
+    uint32_t n_iter  = h->n_iter;
+    
+    if (SRSLTE_TDEC_EXPECT_INPUT_SB && !h->force_not_sb && input_is_interleaved) {
+      syst = input;
+      // align to 32 bytes (warning: must be same alignment as in rm_turbo.c)
+      parity0 = &input[long_cb+32];
+      parity1 = &input[2*(long_cb+32)];
+      if (n_iter == 0) {
+        MAKE_CALL(extract_input_tail_sb)(input, syst, app2, parity0, parity1, long_cb);
+      }
+    } else {
+      if (n_iter == 0) {
+        decptr->tdec_extract_input(input, syst, app2, parity0, parity1, long_cb);
+      }
+    }
+    
+    if ((n_iter%2) == 0) {
+
+      // Add apriori information to decoder 1
+      if (n_iter) {
+        MAKE_VEC(srslte_vec_sub)(app1, ext1, app1, long_cb);
+      }
+
+      // Run MAP DEC #1
+      decptr->tdec_dec(dechdlr, syst, n_iter ? app1 : NULL, parity0, ext1, long_cb);
+
+    }
+    // Interleave extrinsic output of DEC1 to form apriori info for decoder 2
+    if (n_iter%2) {
+      // Convert aposteriori information into extrinsic information
+      if (n_iter > 1) {
+        MAKE_VEC(srslte_vec_sub)(ext1, app1, ext1, long_cb);
+      }
+
+      MAKE_VEC(srslte_vec_lut)(ext1, deinter, app2, long_cb);
+
+      // Run MAP DEC #2. 2nd decoder uses apriori information as systematic bits
+      decptr->tdec_dec(dechdlr, app2, NULL, parity1, ext2, long_cb);
+
+      // Deinterleaved extrinsic bits become apriori info for decoder 1
+     MAKE_VEC(srslte_vec_lut)(ext2, inter, app1, long_cb);
+
+    }
+
+    if (h->n_iter == 0) {
+      debug_vec(syst);
+      debug_vec(parity0);
+      debug_vec(parity1);
+    }
+    debug_vec(ext1);
+    debug_vec(ext2);
+    debug_vec(app1);
+    debug_vec(app2);
+
+    h->n_iter++;
+  } else {
+    fprintf(stderr, "Error CB index not set (call srslte_tdec_new_cb() first\n");
+  }
+}
+
+#undef debug_enabled
+#undef debug_len
+#undef debug_vec
+#undef llr_t
+#undef vec_suffix
+#undef print_suffix
+#undef decptr
+#undef dechdlr
+#undef type_name
+#undef input_is_interleaved

lib/include/srslte/phy/fec/turbodecoder_simd.h

@@ -1,122 +0,0 @@
-/**
- *
- * \section COPYRIGHT
- *
- * Copyright 2013-2015 Software Radio Systems Limited
- *
- * \section LICENSE
- *
- * This file is part of the srsLTE library.
- *
- * srsLTE is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as
- * published by the Free Software Foundation, either version 3 of
- * the License, or (at your option) any later version.
- *
- * srsLTE is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * A copy of the GNU Affero General Public License can be found in
- * the LICENSE file in the top-level directory of this distribution
- * and at http://www.gnu.org/licenses/.
- *
- */
-
-/**********************************************************************************************
- *  File:         turbodecoder.h
- *
- *  Description:  Turbo Decoder.
- *                Parallel Concatenated Convolutional Code (PCCC) with two 8-state constituent
- *                encoders and one turbo code internal interleaver. The coding rate of turbo
- *                encoder is 1/3.
- *                MAP_GEN is the MAX-LOG-MAP generic implementation of the decoder.
- *
- *  Reference:    3GPP TS 36.212 version 10.0.0 Release 10 Sec. 5.1.3.2
- *********************************************************************************************/
-
-#ifndef SRSLTE_TURBODECODER_SIMD_H
-#define SRSLTE_TURBODECODER_SIMD_H
-
-#include "srslte/config.h"
-#include "srslte/phy/fec/tc_interl.h"
-#include "srslte/phy/fec/cbsegm.h"
-
-// Define maximum number of CB decoded in parallel (2 for AVX2)
-#define SRSLTE_TDEC_MAX_NPAR 2
-
-#define SRSLTE_TCOD_RATE 3
-#define SRSLTE_TCOD_TOTALTAIL 12
-
-#define SRSLTE_TCOD_MAX_LEN_CB     6144
-#define SRSLTE_TCOD_MAX_LEN_CODED  (SRSLTE_TCOD_RATE*SRSLTE_TCOD_MAX_LEN_CB+SRSLTE_TCOD_TOTALTAIL)
-
-typedef struct SRSLTE_API {
-  uint32_t max_long_cb;
-  uint32_t max_par_cb; 
-  int16_t *alpha;
-  int16_t *branch;
-} map_gen_t;
-
-typedef struct SRSLTE_API {
-  uint32_t max_long_cb;
-  uint32_t max_par_cb; 
-  
-  map_gen_t dec;
-
-  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_MAX_NPAR];
-} srslte_tdec_simd_t;
-
-SRSLTE_API int srslte_tdec_simd_init(srslte_tdec_simd_t * h, 
-                                     uint32_t max_par_cb, 
-                                     uint32_t max_long_cb);
-
-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);
-
-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_MAX_NPAR],
-                                           uint32_t long_cb);
-
-SRSLTE_API void srslte_tdec_simd_decision(srslte_tdec_simd_t * h, 
-                                          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_MAX_NPAR],
-                                               uint32_t long_cb); 
-
-SRSLTE_API void srslte_tdec_simd_decision_byte_cb(srslte_tdec_simd_t * h, 
-                                                  uint8_t *output, 
-                                                  uint32_t cbidx, 
-                                                  uint32_t long_cb); 
-
-SRSLTE_API 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);
-
-#endif // SRSLTE_TURBODECODER_SIMD_H

lib/include/srslte/phy/fec/turbodecoder_simd_inter.h

@@ -1,119 +0,0 @@
-/**
- *
- * \section COPYRIGHT
- *
- * Copyright 2013-2015 Software Radio Systems Limited
- *
- * \section LICENSE
- *
- * This file is part of the srsLTE library.
- *
- * srsLTE is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as
- * published by the Free Software Foundation, either version 3 of
- * the License, or (at your option) any later version.
- *
- * srsLTE is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * A copy of the GNU Affero General Public License can be found in
- * the LICENSE file in the top-level directory of this distribution
- * and at http://www.gnu.org/licenses/.
- *
- */
-
-/**********************************************************************************************
- *  File:         turbodecoder.h
- *
- *  Description:  Turbo Decoder.
- *                Parallel Concatenated Convolutional Code (PCCC) with two 8-state constituent
- *                encoders and one turbo code internal interleaver. The coding rate of turbo
- *                encoder is 1/3.
- *                MAP_GEN is the MAX-LOG-MAP generic implementation of the decoder.
- *
- *  Reference:    3GPP TS 36.212 version 10.0.0 Release 10 Sec. 5.1.3.2
- *********************************************************************************************/
-
-#ifndef SRSLTE_TURBODECODER_SIMD_INTER_H
-#define SRSLTE_TURBODECODER_SIMD_INTER_H
-
-
-/** This is an simd inter-frame parallel turbo decoder. Parallizes 8 code-blocks using SSE
- *  This implementation is currently not functional and not used by the rest of the code 
- */
-
-#include "srslte/config.h"
-#include "srslte/phy/fec/tc_interl.h"
-#include "srslte/phy/fec/cbsegm.h"
-
-#if LV_HAVE_AVX2
-  #define SRSLTE_TDEC_MAX_NPAR 16
-#else
-  #define SRSLTE_TDEC_MAX_NPAR 8
-#endif
-
-
-typedef struct SRSLTE_API {
-  int max_long_cb;
-
-  int16_t *syst0;
-  int16_t *parity0;
-  int16_t *syst1;
-  int16_t *parity1;
-  int16_t *llr1;
-  int16_t *llr2;
-  int16_t *w;
-  int16_t *alpha;
-
-  uint32_t max_par_cb;   
-  int current_cbidx; 
-  uint32_t current_long_cb;
-  srslte_tc_interl_t interleaver[SRSLTE_NOF_TC_CB_SIZES];
-  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, 
-                                           uint32_t max_par_cb, 
-                                           uint32_t max_long_cb);
-
-SRSLTE_API void srslte_tdec_simd_inter_free(srslte_tdec_simd_inter_t * h);
-
-SRSLTE_API int srslte_tdec_simd_inter_reset(srslte_tdec_simd_inter_t * h, 
-                                      uint32_t long_cb);
-
-SRSLTE_API int srslte_tdec_simd_inter_get_nof_iterations_cb(srslte_tdec_simd_inter_t * h, 
-                                                      uint32_t cb_idx);
-
-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_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_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_MAX_NPAR],
-                                               uint32_t nof_cb,
-                                               uint32_t long_cb); 
-
-SRSLTE_API void srslte_tdec_simd_inter_decision_byte_cb(srslte_tdec_simd_inter_t * h, 
-                                                  uint8_t *output, 
-                                                  uint32_t cbidx, 
-                                                  uint32_t long_cb); 
-
-SRSLTE_API int srslte_tdec_simd_inter_run_all(srslte_tdec_simd_inter_t * h, 
-                                        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);
-
-#endif // SRSLTE_TURBODECODER_SIMD_INTER_H

lib/include/srslte/phy/fec/turbodecoder_sse.h

@@ -24,78 +24,22 @@
  *
  */
 
-/**********************************************************************************************
+#ifndef SRSLTE_TURBODECODER_SSE_H
- *  File:         turbodecoder.h
+#define SRSLTE_TURBODECODER_SSE_H
- *
- *  Description:  Turbo Decoder.
- *                Parallel Concatenated Convolutional Code (PCCC) with two 8-state constituent
- *                encoders and one turbo code internal interleaver. The coding rate of turbo
- *                encoder is 1/3.
- *                MAP_GEN is the MAX-LOG-MAP generic implementation of the decoder.
- *
- *  Reference:    3GPP TS 36.212 version 10.0.0 Release 10 Sec. 5.1.3.2
- *********************************************************************************************/
-
-#ifndef SRSLTE_TURBODECODER_SSE_
-#define SRSLTE_TURBODECODER_SSE_
 
 #include "srslte/config.h"
-#include "srslte/phy/fec/tc_interl.h"
-#include "srslte/phy/fec/cbsegm.h"
-
-#define SRSLTE_TCOD_RATE 3
-#define SRSLTE_TCOD_TOTALTAIL 12
-
-#define SRSLTE_TCOD_MAX_LEN_CB     6144
-#define SRSLTE_TCOD_MAX_LEN_CODED  (SRSLTE_TCOD_RATE*SRSLTE_TCOD_MAX_LEN_CB+SRSLTE_TCOD_TOTALTAIL)
 
 typedef struct SRSLTE_API {
-  int max_long_cb;
+  uint32_t max_long_cb;
   int16_t *alpha;
   int16_t *branch;
-} map_gen_t;
+} tdec_sse_t;
-
-typedef struct SRSLTE_API {
-  int max_long_cb;
-
-  map_gen_t dec;
-
-  int16_t *app1;
-  int16_t *app2;
-  int16_t *ext1;
-  int16_t *ext2;
-  int16_t *syst;
-  int16_t *parity0;
-  int16_t *parity1;
-  
-  int current_cbidx; 
-  srslte_tc_interl_t interleaver[SRSLTE_NOF_TC_CB_SIZES];
-  int n_iter;
-} srslte_tdec_sse_t;
-
-SRSLTE_API int srslte_tdec_sse_init(srslte_tdec_sse_t * h, 
-                                uint32_t max_long_cb);
-
-SRSLTE_API void srslte_tdec_sse_free(srslte_tdec_sse_t * h);
-
-SRSLTE_API int srslte_tdec_sse_reset(srslte_tdec_sse_t * h, uint32_t long_cb);
-
-SRSLTE_API void srslte_tdec_sse_iteration(srslte_tdec_sse_t * h, 
-                                      int16_t * input, 
-                                      uint32_t long_cb);
-
-SRSLTE_API void srslte_tdec_sse_decision(srslte_tdec_sse_t * h, 
-                                     uint8_t *output, 
-                                     uint32_t long_cb);
-
-SRSLTE_API void srslte_tdec_sse_decision_byte(srslte_tdec_sse_t * h, 
-                                          uint8_t *output, 
-                                          uint32_t long_cb); 
 
-SRSLTE_API int srslte_tdec_sse_run_all(srslte_tdec_sse_t * h, 
+int tdec_sse_init(void **h, uint32_t max_long_cb);
-                                   int16_t * input, 
+void tdec_sse_free(void *h);
-                                   uint8_t *output,
+void tdec_sse_dec(void *h, int16_t * input, int16_t *app, int16_t * parity,
-                                   uint32_t nof_iterations, 
+                 int16_t *output, uint32_t long_cb);
-                                   uint32_t long_cb);
+void tdec_sse_extract_input(int16_t *input, int16_t *syst, int16_t *parity0, int16_t *parity1, int16_t *app2, uint32_t long_cb);
+void tdec_sse_decision_byte(int16_t *app1, uint8_t *output, uint32_t long_cb);
 
-#endif // SRSLTE_TURBODECODER_SSE_
+#endif // SRSLTE_TURBODECODER_SSE_H

lib/include/srslte/phy/fec/turbodecoder_win.h

@@ -0,0 +1,750 @@
+/**
+ *
+ * \section COPYRIGHT
+ *
+ * Copyright 2013-2015 Software Radio Systems Limited
+ *
+ * \section LICENSE
+ *
+ * This file is part of the srsLTE library.
+ *
+ * srsLTE is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as
+ * published by the Free Software Foundation, either version 3 of
+ * the License, or (at your option) any later version.
+ *
+ * srsLTE is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * A copy of the GNU Affero General Public License can be found in
+ * the LICENSE file in the top-level directory of this distribution
+ * and at http://www.gnu.org/licenses/.
+ *
+ */
+
+#include "srslte/config.h"
+
+#define MAKE_FUNC(a) CONCAT2(CONCAT2(tdec_win,WINIMP),CONCAT2(_,a))
+#define MAKE_TYPE CONCAT2(CONCAT2(tdec_win_,WINIMP),_t)
+
+#ifdef WINIMP_IS_SSE16
+
+  #ifndef LV_HAVE_SSE
+  #error "Selected SSE window decoder but instruction set not supported"
+  #endif
+
+  #include <nmmintrin.h>
+
+  #define WINIMP sse16
+  #define nof_blocks 8
+
+  #define llr_t int16_t
+
+  #define simd_type_t  __m128i
+  #define simd_load    _mm_load_si128
+  #define simd_store   _mm_store_si128
+  #define simd_add     _mm_adds_epi16
+  #define simd_sub     _mm_subs_epi16
+  #define simd_max     _mm_max_epi16
+  #define simd_set1    _mm_set1_epi16
+  #define simd_insert  _mm_insert_epi16
+  #define simd_shuffle _mm_shuffle_epi8
+  #define move_right   _mm_set_epi8(15,14,15,14,13,12,11,10,9,8,7,6,5,4,3,2)
+  #define move_left    _mm_set_epi8(13,12,11,10,9,8,7,6,5,4,3,2,1,0,1,0)
+  #define simd_rb_shift _mm_srai_epi16
+
+  #define normalize_period 2
+  #define win_overlap_len  40
+
+#define divide_output 1
+
+#define INF 10000
+
+#else
+#ifdef WINIMP_IS_AVX16
+
+  #ifndef LV_HAVE_AVX
+  #error "Selected AVX window decoder but instruction set not supported"
+  #endif
+
+  #include <immintrin.h>
+
+  #define WINIMP avx16
+  #define nof_blocks 16
+
+  #define llr_t int16_t
+
+  #define simd_type_t  __m256i
+  #define simd_load    _mm256_load_si256
+  #define simd_store   _mm256_store_si256
+  #define simd_add     _mm256_adds_epi16
+  #define simd_sub     _mm256_subs_epi16
+  #define simd_max     _mm256_max_epi16
+  #define simd_set1    _mm256_set1_epi16
+  #define simd_insert  _mm256_insert_epi16
+  #define simd_shuffle _mm256_shuffle_epi8
+  #define move_right   _mm256_set_epi8(31,30,31,30,29,28,27,26,25,24,23,22,21,20,19,18,17,16,15,14,13,12,11,10,9,8,7,6,5,4,3,2)
+  #define move_left    _mm256_set_epi8(29,28,27,26,25,24,23,22,21,20,19,18,17,16,15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0,1,0)
+
+  #define normalize_period 2
+  #define win_overlap_len  40
+
+  #define INF 10000
+#else
+
+#ifdef WINIMP_IS_SSE8
+
+  #ifndef LV_HAVE_SSE
+  #error "Selected SSE window decoder but instruction set not supported"
+  #endif
+
+  #include <nmmintrin.h>
+
+  #define WINIMP sse8
+  #define nof_blocks 16
+
+  #define llr_t int8_t
+
+  #define simd_type_t  __m128i
+  #define simd_load    _mm_load_si128
+  #define simd_store   _mm_store_si128
+  #define simd_add     _mm_adds_epi8
+  #define simd_sub     _mm_subs_epi8
+  #define simd_max     _mm_max_epi8
+  #define simd_set1    _mm_set1_epi8
+  #define simd_insert  _mm_insert_epi8
+  #define simd_shuffle _mm_shuffle_epi8
+  #define move_right   _mm_set_epi8(15,15,14,13,12,11,10,9,8,7,6,5,4,3,2,1)
+  #define move_left    _mm_set_epi8(14,13,12,11,10,9,8,7,6,5,4,3,2,1,0,0)
+  #define simd_rb_shift simd_rb_shift_128
+
+  #define normalize_max
+  #define normalize_period 1
+  #define win_overlap_len  40
+  #define use_saturated_add
+  #define divide_output 1
+
+  #define INF 0
+
+  inline static simd_type_t simd_rb_shift_128(simd_type_t v, const int l) {
+    __m128i low = _mm_srai_epi16(_mm_slli_epi16(v,8), l+8);
+    __m128i hi  = _mm_srai_epi16(v,l);
+    return _mm_blendv_epi8(hi, low, _mm_set1_epi32(0x00FF00FF));
+  }
+
+
+#else
+
+#ifdef WINIMP_IS_AVX8
+
+  #ifndef LV_HAVE_AVX
+  #error "Selected AVX window decoder but instruction set not supported"
+  #endif
+
+  #include <immintrin.h>
+
+  #define WINIMP avx8
+  #define nof_blocks 32
+
+  #define llr_t int8_t
+
+  #define simd_type_t  __m256i
+  #define simd_load    _mm256_load_si256
+  #define simd_store   _mm256_store_si256
+  #define simd_add     _mm256_adds_epi8
+  #define simd_sub     _mm256_subs_epi8
+  #define simd_max     _mm256_max_epi8
+  #define simd_set1    _mm256_set1_epi8
+  #define simd_insert  _mm256_insert_epi8
+  #define simd_shuffle _mm256_shuffle_epi8
+  #define move_right   _mm256_set_epi8(31,31,30,29,28,27,26,25,24,23,22,21,20,19,18,17,16,15,14,13,12,11,10,9,8,7,6,5,4,3,2,1)
+  #define move_left    _mm256_set_epi8(30,29,28,27,26,25,24,23,22,21,20,19,18,17,16,15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0,0)
+  #define simd_rb_shift simd_rb_shift_256
+
+  #define INF 0
+
+  #define normalize_max
+  #define normalize_period 1
+  #define win_overlap_len  40
+  #define use_saturated_add
+  #define divide_output 1
+
+  inline static simd_type_t simd_rb_shift_256(simd_type_t v, const int l) {
+    __m256i low = _mm256_srai_epi16(_mm256_slli_epi16(v,8), l+8);
+    __m256i hi  = _mm256_srai_epi16(v,l);
+    return _mm256_blendv_epi8(hi, low, _mm256_set1_epi32(0x00FF00FF));
+  }
+
+
+#else
+  #error "Unknown WINIMP value"
+#endif
+#endif
+#endif
+#endif
+
+typedef struct SRSLTE_API {
+  uint32_t max_long_cb;
+  llr_t *beta;
+} MAKE_TYPE;
+
+
+#define long_sb (long_cb/nof_blocks)
+
+
+
+#define debug_enabled_win 0
+
+#if debug_enabled_win
+#define debug_state(d) printf("k=%5d, in=%5d, pa=%3d, out=%5d, alpha=[", d*long_sb+k+1, MAKE_FUNC(get_simd)(x,d), MAKE_FUNC(get_simd)(y,d), MAKE_FUNC(get_simd)(out,d)); \
+                        for (int j=0;j<8;j++) printf("%5d, ", MAKE_FUNC(get_simd)(old[j],d)); \
+                        printf("], beta=["); \
+                        for (int j=0;j<8;j++) printf("%5d, ", MAKE_FUNC(get_simd)(beta_save[j], d));printf("\n");
+
+#define debug_state_pre(d) printf("pre-window k=%5d, in=%5d, pa=%3d, alpha=[", (d+1)*long_sb-loop_len+k+1, MAKE_FUNC(get_simd)(x,d), MAKE_FUNC(get_simd)(y,d)); \
+                        for (int j=0;j<8;j++) printf("%5d, ", MAKE_FUNC(get_simd)(old[j],d)); \
+                        printf("]\n");
+
+#define debug_state_beta(d) printf("k=%5d, in=%5d, pa=%3d, beta=[", d*long_sb+k, MAKE_FUNC(get_simd)(x,d), MAKE_FUNC(get_simd)(y,d)); \
+                        for (int j=0;j<8;j++) printf("%5d, ", MAKE_FUNC(get_simd)(old[j],d));\
+                        printf("\n");
+
+static llr_t MAKE_FUNC(get_simd)(simd_type_t x, uint32_t pos) {
+  llr_t *s = (llr_t*) &x;
+  return s[pos];
+}
+
+
+#else
+#define debug_state(a)
+#define debug_state_pre(a)
+#define debug_state_beta(a)
+#endif
+/*
+static void MAKE_FUNC(print_simd)(simd_type_t x) {
+  llr_t *s = (llr_t*) &x;
+  printf("[");
+  for (int i=0;i<nof_blocks;i++) {
+    printf("%4d, ", s[i]);
+  }
+  printf("]\n");
+}*/
+
+inline static llr_t MAKE_FUNC(sadd)(llr_t x, llr_t y) {
+#ifndef use_saturated_add
+  return x+y;
+#else
+  int16_t z = (int16_t) x+y;
+  return z>127?127:(int8_t) z;
+#endif
+}
+
+inline static void MAKE_FUNC(normalize)(uint32_t k, simd_type_t old[8]) {
+  if ((k % normalize_period) == 0 && k != 0) {
+#ifdef normalize_max
+    simd_type_t m = simd_max(old[0],old[1]);
+    for (int i=2;i<8;i++) {
+      m = simd_max(m,old[i]);
+    }
+    for (int i=0;i<8;i++) {
+      old[i] = simd_sub(old[i], m);
+    }
+#else
+    for (int i = 1; i < 8; i++) {
+      old[i] = simd_sub(old[i], old[0]);
+    }
+    old[0] = simd_set1(0);
+#endif
+  }
+}
+
+static void MAKE_FUNC(beta_trellis)(llr_t *input, llr_t *parity, uint32_t long_cb, llr_t old[8])
+{
+  llr_t m_b[8], new[8];
+  llr_t x, y, xy;
+
+  /* Calculate last state using Tail. No need to use SIMD here */
+  old[0] = 0;
+  for (int i = 1; i < 8; i++) {
+    old[i] = -INF;
+  }
+  for (int k=long_cb+2;k >= long_cb; k--) {
+    x = input[k];
+    y = parity[k];
+
+    xy = MAKE_FUNC(sadd)(x, y);
+
+    m_b[0] = MAKE_FUNC(sadd)(old[4],xy);
+    m_b[1] = old[4];
+    m_b[2] = MAKE_FUNC(sadd)(old[5], y);
+    m_b[3] = MAKE_FUNC(sadd)(old[5], x);
+    m_b[4] = MAKE_FUNC(sadd)(old[6], x);
+    m_b[5] = MAKE_FUNC(sadd)(old[6], y);
+    m_b[6] = old[7];
+    m_b[7] = MAKE_FUNC(sadd)(old[7], xy);
+
+    new[0] = old[0];
+    new[1] = MAKE_FUNC(sadd)(old[0], xy);
+    new[2] = MAKE_FUNC(sadd)(old[1], x);
+    new[3] = MAKE_FUNC(sadd)(old[1], y);
+    new[4] = MAKE_FUNC(sadd)(old[2], y);
+    new[5] = MAKE_FUNC(sadd)(old[2], x);
+    new[6] = MAKE_FUNC(sadd)(old[3], xy);
+    new[7] = old[3];
+
+#if debug_enabled_win
+    printf("trellis: k=%d, in=%d, pa=%d, beta: ", k, x, y); for (int i=0;i<8;i++) {printf("%d,", old[i]);} printf("\n");
+#endif
+
+    for (int i = 0; i < 8; i++) {
+      if (m_b[i] > new[i])
+        new[i] = m_b[i];
+      old[i] = new[i];
+    }
+  }
+}
+
+/* Computes beta values */
+static void MAKE_FUNC(beta)(MAKE_TYPE * s, llr_t *input, llr_t *app, llr_t *parity, uint32_t long_cb)
+{
+  simd_type_t m_b[8], new[8], old[8];
+  simd_type_t x, y, xy, ap;
+
+  simd_type_t *inputPtr;
+  simd_type_t *appPtr;
+  simd_type_t *parityPtr;
+  simd_type_t *betaPtr   = (simd_type_t*) s->beta;
+
+  uint32_t loop_len;
+  for (int j=0;j<2;j++) {
+
+    // First run L states to find initial state for all sub-blocks after first
+    if (j==0) {
+      loop_len = win_overlap_len;
+    } else {
+      loop_len = long_sb;
+    }
+
+    // When passing through all window pick estimated initial states (known state for sb=0)
+    if (loop_len == long_sb) {
+
+      // shuffle across 128-bit boundary manually
+#ifdef WINIMP_IS_AVX16
+      llr_t tmp[8];
+      for (int i = 0; i < 8; i++) {
+        tmp[i] = _mm256_extract_epi16(old[i], 8);
+      }
+#endif
+#ifdef WINIMP_IS_AVX8
+      llr_t tmp[8];
+      for (int i = 0; i < 8; i++) {
+        tmp[i] = _mm256_extract_epi8(old[i], 16);
+      }
+#endif
+
+      for (int i = 0; i < 8; i++) {
+        old[i] = simd_shuffle(old[i], move_right);
+      }
+      // last sub-block state is calculated from the trellis
+      llr_t trellis_old[8];
+      MAKE_FUNC(beta_trellis)(input, parity, long_cb, trellis_old);
+      for (int i = 0; i < 8; i++) {
+        old[i] = simd_insert(old[i], trellis_old[i], nof_blocks-1);
+      }
+
+#ifdef WINIMP_IS_AVX16
+      for (int i = 0; i < 8; i++) {
+        old[i] = _mm256_insert_epi16(old[i], tmp[i], 7);
+      }
+#endif
+#ifdef WINIMP_IS_AVX8
+      for (int i = 0; i < 8; i++) {
+        old[i] = _mm256_insert_epi8(old[i], tmp[i], 15);
+      }
+#endif
+
+      inputPtr  = (simd_type_t*) &input[long_cb-nof_blocks];
+      appPtr    = (simd_type_t*) &app[long_cb-nof_blocks];
+      parityPtr = (simd_type_t*) &parity[long_cb-nof_blocks];
+
+      for (int i = 0; i < 8; i++) {
+        simd_store(&betaPtr[8*long_sb + i], old[i]);
+      }
+
+    } else {
+      // when estimating states, just set all to unknown
+      for (int i = 0; i < 8; i++) {
+        old[i] = simd_set1(-INF);
+      }
+      inputPtr  = (simd_type_t*) &input[nof_blocks*(loop_len-1)];
+      appPtr    = (simd_type_t*) &app[nof_blocks*(loop_len-1)];
+      parityPtr = (simd_type_t*) &parity[nof_blocks*(loop_len-1)];
+    }
+
+    for (int k = loop_len - 1; k >= 0; k--) {
+      x = simd_load(inputPtr--);
+      y = simd_load(parityPtr--);
+
+      if (app) {
+        ap = simd_load(appPtr--);
+        x = simd_add(ap, x);
+      }
+
+      xy = simd_add(x, y);
+
+      m_b[0] = simd_add(old[4], xy);
+      m_b[1] = old[4];
+      m_b[2] = simd_add(old[5], y);
+      m_b[3] = simd_add(old[5], x);
+      m_b[4] = simd_add(old[6], x);
+      m_b[5] = simd_add(old[6], y);
+      m_b[6] = old[7];
+      m_b[7] = simd_add(old[7], xy);
+
+      new[0] = old[0];
+      new[1] = simd_add(old[0], xy);
+      new[2] = simd_add(old[1], x);
+      new[3] = simd_add(old[1], y);
+      new[4] = simd_add(old[2], y);
+      new[5] = simd_add(old[2], x);
+      new[6] = simd_add(old[3], xy);
+      new[7] = old[3];
+
+      // Calculate maximum metric
+      for (int i = 0; i < 8; i++) {
+        old[i] = simd_max(m_b[i], new[i]);
+      }
+      // Store metric only when doing the final pass
+      if (loop_len == long_sb) {
+        for (int i = 0; i < 8; i++) {
+          simd_store(&betaPtr[8*k + i], old[i]);
+        }
+      }
+      if (loop_len!=long_sb) {
+        debug_state_beta(0);
+      } else {
+        debug_state_beta(0);
+      }
+
+      // normalize
+      MAKE_FUNC(normalize)(k, old);
+    }
+  }
+}
+
+/* Computes alpha metrics */
+static void MAKE_FUNC(alpha)(MAKE_TYPE * s, llr_t *input, llr_t *app, llr_t *parity, llr_t * output, uint32_t long_cb)
+{
+  simd_type_t m_b[8], new[8], old[8], max1[8], max0[8];
+  simd_type_t x, y, xy, ap;
+  simd_type_t m1, m0;
+
+  simd_type_t *inputPtr;
+  simd_type_t *appPtr;
+  simd_type_t *parityPtr;
+  simd_type_t *betaPtr   = (simd_type_t*) s->beta;
+  simd_type_t *outputPtr = (simd_type_t*) output;
+
+#if debug_enabled_win
+  simd_type_t beta_save[8];
+#endif
+
+  // Skip state 0
+  betaPtr+=8;
+
+  uint32_t loop_len;
+
+  for (int j=0;j<2;j++) {
+
+    // First run L states to find initial state for all sub-blocks after first
+    if (j==0) {
+      loop_len = win_overlap_len;
+    } else {
+      loop_len = long_sb;
+    }
+
+    // When passing through all window pick estimated initial states (known state for sb=0)
+    if (loop_len == long_sb) {
+
+#ifdef WINIMP_IS_AVX16
+      llr_t tmp[8];
+      for (int i=0;i<8;i++) {
+        tmp[i] = _mm256_extract_epi16(old[i], 7);
+      }
+#endif
+#ifdef WINIMP_IS_AVX8
+      llr_t tmp[8];
+      for (int i=0;i<8;i++) {
+        tmp[i] = _mm256_extract_epi8(old[i], 15);
+      }
+#endif
+      for (int i = 0; i < 8; i++) {
+        old[i] = simd_shuffle(old[i], move_left);
+      }
+#ifdef WINIMP_IS_AVX16
+      for (int i=0;i<8;i++) {
+        old[i] = _mm256_insert_epi16(old[i], tmp[i], 8);
+      }
+#endif
+#ifdef WINIMP_IS_AVX8
+      for (int i=0;i<8;i++) {
+        old[i] = _mm256_insert_epi8(old[i], tmp[i], 16);
+      }
+#endif
+      // 1st sub-block state is known
+      old[0] = simd_insert(old[0], 0, 0);
+      for (int i = 1; i < 8; i++) {
+        old[i] = simd_insert(old[i], -INF, 0);
+      }
+    } else {
+      // when estimating states, just set all to unknown
+      for (int i = 0; i < 8; i++) {
+        old[i] = simd_set1(-INF);
+      }
+    }
+
+    inputPtr  = (simd_type_t*) &input[nof_blocks*(long_sb-loop_len)];
+    appPtr    = (simd_type_t*) &app[nof_blocks*(long_sb-loop_len)];
+    parityPtr = (simd_type_t*) &parity[nof_blocks*(long_sb-loop_len)];
+
+    for (int k = 0; k < loop_len; k++) {
+      x = simd_load(inputPtr++);
+      y = simd_load(parityPtr++);
+
+      if (app) {
+        ap = simd_load(appPtr++);
+        x = simd_add(ap, x);
+      }
+
+      xy = simd_add(x,y);
+
+      m_b[0] = old[0];
+      m_b[1] = simd_add(old[3], y);
+      m_b[2] = simd_add(old[4], y);
+      m_b[3] = old[7];
+      m_b[4] = old[1];
+      m_b[5] = simd_add(old[2], y);
+      m_b[6] = simd_add(old[5], y);
+      m_b[7] = old[6];
+
+      new[0] = simd_add(old[1], xy);
+      new[1] = simd_add(old[2], x);
+      new[2] = simd_add(old[5], x);
+      new[3] = simd_add(old[6], xy);
+      new[4] = simd_add(old[0], xy);
+      new[5] = simd_add(old[3], x);
+      new[6] = simd_add(old[4], x);
+      new[7] = simd_add(old[7], xy);
+
+      // Load beta and compute output only when passing through all window
+      if (loop_len == long_sb) {
+        simd_type_t beta;
+        for (int i = 0; i < 8; i++) {
+          beta = simd_load(betaPtr++);
+          max0[i] = simd_add(beta, m_b[i]);
+          max1[i] = simd_add(beta, new[i]);
+
+#if debug_enabled_win
+          beta_save[i] = beta;
+#endif
+        }
+
+        m1 = simd_max(max1[0], max1[1]);
+        m0 = simd_max(max0[0], max0[1]);
+
+        for (int i = 2; i < 8; i++) {
+          m1 = simd_max(m1, max1[i]);
+          m0 = simd_max(m0, max0[i]);
+        }
+
+        simd_type_t out = simd_sub(m1, m0);
+
+        // Divide output when using 8-bit arithmetic
+#ifdef divide_output
+        out = simd_rb_shift(out, divide_output);
+#endif
+
+        simd_store(outputPtr++, out);
+
+        debug_state(0);
+      }
+
+      for (int i = 0; i < 8; i++) {
+        old[i] = simd_max(m_b[i], new[i]);
+      }
+
+      // normalize
+      MAKE_FUNC(normalize)(k, old);
+
+      if (loop_len != long_sb) {
+        debug_state_pre(0);
+      }
+    }
+  }
+}
+
+int MAKE_FUNC(init)(void **hh, uint32_t max_long_cb)
+{
+  *hh = calloc(1, sizeof(MAKE_TYPE));
+
+  MAKE_TYPE *h = (MAKE_TYPE*) *hh;
+
+  h->beta = srslte_vec_malloc(sizeof(llr_t) * 8 * max_long_cb * nof_blocks);
+  if (!h->beta) {
+    perror("srslte_vec_malloc");
+    return -1;
+  }
+  h->max_long_cb = max_long_cb;
+  return nof_blocks;
+}
+
+void MAKE_FUNC(free)(void *hh)
+{
+  MAKE_TYPE *h = (MAKE_TYPE*) hh;
+  if (h->beta) {
+    free(h->beta);
+  }
+  bzero(h, sizeof(MAKE_TYPE));
+}
+
+void MAKE_FUNC(dec)(void *hh, llr_t *input, llr_t *app, llr_t *parity, llr_t *output, uint32_t long_cb)
+{
+  MAKE_TYPE *h = (MAKE_TYPE*) hh;
+  MAKE_FUNC(beta)(h, input, app, parity, long_cb);
+  MAKE_FUNC(alpha)(h, input, app, parity, output, long_cb);
+#if debug_enabled_win
+  printf("running win decoder: %s\n", STRING(WINIMP));
+#endif
+}
+
+#define INSERT8_INPUT(reg, st, off)     reg = simd_insert(reg, input[3*(i+(st+0)*long_sb)+off], st+0);\
+                                        reg = simd_insert(reg, input[3*(i+(st+1)*long_sb)+off], st+1);\
+                                        reg = simd_insert(reg, input[3*(i+(st+2)*long_sb)+off], st+2);\
+                                        reg = simd_insert(reg, input[3*(i+(st+3)*long_sb)+off], st+3);\
+                                        reg = simd_insert(reg, input[3*(i+(st+4)*long_sb)+off], st+4);\
+                                        reg = simd_insert(reg, input[3*(i+(st+5)*long_sb)+off], st+5);\
+                                        reg = simd_insert(reg, input[3*(i+(st+6)*long_sb)+off], st+6);\
+                                        reg = simd_insert(reg, input[3*(i+(st+7)*long_sb)+off], st+7);
+
+
+void MAKE_FUNC(extract_input)(llr_t *input, llr_t *systematic, llr_t *app2, llr_t *parity_0, llr_t *parity_1, uint32_t long_cb)
+{
+  simd_type_t *systPtr    = (simd_type_t*) systematic;
+  simd_type_t *parity0Ptr = (simd_type_t*) parity_0;
+  simd_type_t *parity1Ptr = (simd_type_t*) parity_1;
+
+  simd_type_t syst, parity0, parity1;
+
+  for (int i=0;i<long_sb;i++) {
+    INSERT8_INPUT(syst,    0, 0);
+    INSERT8_INPUT(parity0, 0, 1);
+    INSERT8_INPUT(parity1, 0, 2);
+
+#if nof_blocks >= 16
+    INSERT8_INPUT(syst,    8, 0);
+    INSERT8_INPUT(parity0, 8, 1);
+    INSERT8_INPUT(parity1, 8, 2);
+#endif
+
+#if nof_blocks >= 32
+    INSERT8_INPUT(syst,    16, 0);
+    INSERT8_INPUT(parity0, 16, 1);
+    INSERT8_INPUT(parity1, 16, 2);
+    INSERT8_INPUT(syst,    24, 0);
+    INSERT8_INPUT(parity0, 24, 1);
+    INSERT8_INPUT(parity1, 24, 2);
+#endif
+
+    simd_store(systPtr++,    syst);
+    simd_store(parity0Ptr++, parity0);
+    simd_store(parity1Ptr++, parity1);
+  }
+
+  for (int i = long_cb; i < long_cb + 3; i++) {
+    systematic[i] = input[3*long_cb + 2*(i - long_cb)];
+    parity_0[i]   = input[3*long_cb + 2*(i - long_cb) + 1];
+
+    app2[i]       = input[3*long_cb + 6 + 2*(i - long_cb)];
+    parity_1[i]   = input[3*long_cb + 6 + 2*(i - long_cb) + 1];
+  }
+}
+
+#define deinter(x,win) ((x%(long_cb/win))*(win)+x/(long_cb/win))
+
+#define reset_cnt(a,b)      if(!((a+1)%b)) { \
+                              k+=b*nof_blocks; \
+                              if (k >= long_cb) { \
+                                k -= (long_cb-1);\
+                              }\
+                            }
+#define insert_bit(a,b)     ap = _mm_insert_epi16(ap, app1[k+(a%b)*nof_blocks], 7-a); \
+                            reset_cnt(a,b); \
+
+
+#define decide_for(b)     for (uint32_t i = 0; i < long_cb/8; i++) { \
+                            insert_bit(0,b);\
+                            insert_bit(1,b);\
+                            insert_bit(2,b);\
+                            insert_bit(3,b);\
+                            insert_bit(4,b);\
+                            insert_bit(5,b);\
+                            insert_bit(6,b);\
+                            insert_bit(7,b);\
+                            output[i] = (uint8_t) _mm_movemask_epi8(_mm_cmpgt_epi8(_mm_packs_epi16(ap,zeros),zeros));\
+                          }
+
+/* No improvement to use AVX here */
+void MAKE_FUNC(decision_byte)(llr_t *app1, uint8_t *output, uint32_t long_cb)
+{
+  uint32_t k=0;
+  __m128i zeros = _mm_setzero_si128();
+  __m128i ap;
+
+  if ((long_cb%(nof_blocks*8)) == 0) {
+    decide_for(8);
+  } else if ((long_cb%(nof_blocks*4)) == 0) {
+    decide_for(4);
+  } else if ((long_cb%(nof_blocks*2)) == 0) {
+    decide_for(2);
+  } else {
+    decide_for(1);
+  }
+}
+
+
+#undef WINIMP
+#undef nof_blocks
+#undef llr_t
+#undef normalize_period
+#undef INF
+#undef win_overlap_len
+#undef simd_type_t
+#undef simd_load
+#undef simd_store
+#undef simd_add
+#undef simd_sub
+#undef simd_max
+#undef simd_set1
+#undef simd_insert
+#undef simd_shuffle
+#undef move_right
+#undef move_left
+#undef debug_enabled_win
+
+#ifdef normalize_max
+#undef normalize_max
+#endif
+
+#ifdef use_saturated_add
+#undef use_saturated_add
+#endif
+
+#ifdef simd_rb_shift
+#undef simd_rb_shift
+#endif
+
+#ifdef divide_output
+#undef divide_output
+#endif

lib/include/srslte/phy/modem/demod_soft.h

@@ -53,4 +53,9 @@ SRSLTE_API int srslte_demod_soft_demodulate_s(srslte_mod_t modulation,
                                               short* llr, 
                                               int nsymbols); 
 
+SRSLTE_API int srslte_demod_soft_demodulate_b(srslte_mod_t modulation,
+                                              const cf_t* symbols,
+                                              int8_t* llr,
+                                              int nsymbols);
+
 #endif // SRSLTE_DEMOD_SOFT_H

lib/include/srslte/phy/phch/pdsch.h

@@ -65,6 +65,8 @@ typedef struct SRSLTE_API {
   uint16_t ue_rnti;
   bool is_ue;
 
+  bool llr_is_8bit;
+
   /* Power allocation parameter 3GPP 36.213 Clause 5.2 Rho_b */
   float rho_a;
 

lib/include/srslte/phy/phch/pusch.h

@@ -74,6 +74,8 @@ typedef struct SRSLTE_API {
   uint16_t ue_rnti;
   uint32_t max_re;
 
+  bool llr_is_8bit;
+
   srslte_dft_precoding_t dft_precoding;  
   
   /* buffers */

lib/include/srslte/phy/phch/sch.h

@@ -59,6 +59,8 @@ typedef struct SRSLTE_API {
   uint32_t max_iterations;
   uint32_t nof_iterations;
 
+  bool llr_is_8bit;
+
   /* buffers */
   uint8_t *cb_in; 
   uint8_t *parity_bits;  

lib/include/srslte/phy/scrambling/scrambling.h

@@ -68,6 +68,11 @@ SRSLTE_API void srslte_scrambling_s_offset(srslte_sequence_t *s,
                                            int offset, 
                                            int len);
 
+SRSLTE_API void srslte_scrambling_sb_offset(srslte_sequence_t *s,
+                                            int8_t *data,
+                                            int offset,
+                                            int len);
+
 SRSLTE_API void srslte_scrambling_c(srslte_sequence_t *s, 
                                     cf_t *data);
 

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

@@ -100,6 +100,7 @@
 
 #define SRSLTE_SIMD_I_SIZE    16
 
+#define SRSLTE_SIMD_B_SIZE    64
 #define SRSLTE_SIMD_S_SIZE    32
 #define SRSLTE_SIMD_C16_SIZE  0
 
@@ -111,6 +112,7 @@
 
 #define SRSLTE_SIMD_I_SIZE    8
 
+#define SRSLTE_SIMD_B_SIZE    32
 #define SRSLTE_SIMD_S_SIZE    16
 #define SRSLTE_SIMD_C16_SIZE  16
 
@@ -122,6 +124,7 @@
 
 #define SRSLTE_SIMD_I_SIZE    4
 
+#define SRSLTE_SIMD_B_SIZE    16
 #define SRSLTE_SIMD_S_SIZE    8
 #define SRSLTE_SIMD_C16_SIZE  8
 
@@ -132,7 +135,7 @@
 #define SRSLTE_SIMD_CF_SIZE   4
 
 #define SRSLTE_SIMD_I_SIZE    4
-
+#define SRSLTE_SIMD_B_SIZE    16
 #define SRSLTE_SIMD_S_SIZE    8
 #define SRSLTE_SIMD_C16_SIZE  8
 
@@ -141,7 +144,7 @@
 #define SRSLTE_SIMD_CF_SIZE   0
 
 #define SRSLTE_SIMD_I_SIZE    0
-
+#define SRSLTE_SIMD_B_SIZE    0
 #define SRSLTE_SIMD_S_SIZE    0
 #define SRSLTE_SIMD_C16_SIZE  0
 
@@ -1336,6 +1339,24 @@ static inline simd_s_t srslte_simd_s_mul(simd_s_t a, simd_s_t b) {
 #endif /* LV_HAVE_AVX512 */
 }
 
+static inline simd_s_t srslte_simd_s_neg(simd_s_t a, simd_s_t b) {
+#ifdef LV_HAVE_AVX512
+#error sign instruction not available in avx512
+#else /* LV_HAVE_AVX512 */
+#ifdef LV_HAVE_AVX2
+  return _mm256_sign_epi16(a, b);
+#else /* LV_HAVE_AVX2 */
+  #ifdef LV_HAVE_SSE
+  return _mm_sign_epi16(a, b);
+#else /* LV_HAVE_SSE */
+#ifdef HAVE_NEON
+  #error sign instruction not available in Neon
+#endif /* HAVE_NEON */
+#endif /* LV_HAVE_SSE */
+#endif /* LV_HAVE_AVX2 */
+#endif /* LV_HAVE_AVX512 */
+}
+
 static inline simd_s_t srslte_simd_s_add(simd_s_t a, simd_s_t b) {
 #ifdef LV_HAVE_AVX512
   return _mm512_add_epi16(a, b);
@@ -1681,7 +1702,7 @@ typedef int8x16_t simd_b_t;
 
 
 
-static inline simd_b_t srslte_simd_b_load(int8_t *ptr){
+static inline simd_b_t srslte_simd_b_load(const int8_t *ptr){
 #ifdef LV_HAVE_AVX512
   return _mm512_load_si512(ptr);
 #else /* LV_HAVE_AVX512 */
@@ -1699,7 +1720,7 @@ static inline simd_b_t srslte_simd_b_load(int8_t *ptr){
 #endif /* LV_HAVE_AVX512 */
 }  
   
-static inline simd_b_t srslte_simd_b_loadu(int8_t *ptr){
+static inline simd_b_t srslte_simd_b_loadu(const int8_t *ptr){
 #ifdef LV_HAVE_AVX512
   return _mm512_loadu_si512(ptr);
 #else /* LV_HAVE_AVX512 */
@@ -1773,6 +1794,44 @@ static inline simd_b_t srslte_simd_b_xor(simd_b_t a, simd_b_t b) {
 #endif /* LV_HAVE_AVX512 */
 }
 
+static inline simd_s_t srslte_simd_b_sub(simd_s_t a, simd_s_t b) {
+#ifdef LV_HAVE_AVX512
+  return _mm512_subs_epi8(a, b);
+#else /* LV_HAVE_AVX512 */
+#ifdef LV_HAVE_AVX2
+  return _mm256_subs_epi8(a, b);
+#else /* LV_HAVE_AVX2 */
+  #ifdef LV_HAVE_SSE
+  return _mm_subs_epi8(a, b);
+#else /* LV_HAVE_SSE */
+#ifdef HAVE_NEON
+  return vsubqs_s8(a, b);
+#endif /* HAVE_NEON */
+#endif /* LV_HAVE_SSE */
+#endif /* LV_HAVE_AVX2 */
+#endif /* LV_HAVE_AVX512 */
+}
+
+static inline simd_s_t srslte_simd_b_neg(simd_b_t a, simd_b_t b) {
+#ifdef LV_HAVE_AVX512
+#error sign instruction not available in avx512
+#else /* LV_HAVE_AVX512 */
+#ifdef LV_HAVE_AVX2
+  return _mm256_sign_epi8(a, b);
+#else /* LV_HAVE_AVX2 */
+  #ifdef LV_HAVE_SSE
+  return _mm_sign_epi8(a, b);
+#else /* LV_HAVE_SSE */
+#ifdef HAVE_NEON
+  #error sign instruction not available in Neon
+#endif /* HAVE_NEON */
+#endif /* LV_HAVE_SSE */
+#endif /* LV_HAVE_AVX2 */
+#endif /* LV_HAVE_AVX512 */
+}
+
+
+
 #endif /*SRSLTE_SIMD_B_SIZE */
 
 

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

@@ -69,6 +69,7 @@ SRSLTE_API void *srslte_vec_realloc(void *ptr, uint32_t old_size, uint32_t new_s
 SRSLTE_API void srslte_vec_fprint_c(FILE *stream, cf_t *x, const uint32_t len);
 SRSLTE_API void srslte_vec_fprint_f(FILE *stream, float *x, const uint32_t len);
 SRSLTE_API void srslte_vec_fprint_b(FILE *stream, uint8_t *x, const uint32_t len);
+SRSLTE_API void srslte_vec_fprint_bs(FILE *stream, int8_t *x, const uint32_t len);
 SRSLTE_API void srslte_vec_fprint_byte(FILE *stream, uint8_t *x, const uint32_t len);
 SRSLTE_API void srslte_vec_fprint_i(FILE *stream, int *x, const uint32_t len);
 SRSLTE_API void srslte_vec_fprint_s(FILE *stream, short *x, const uint32_t len); 
@@ -82,12 +83,13 @@ SRSLTE_API void srslte_vec_load_file(char *filename, void *buffer, const uint32_
 /* sum two vectors */
 SRSLTE_API void srslte_vec_sum_fff(const float *x, const float *y, float *z, const uint32_t len);
 SRSLTE_API void srslte_vec_sum_ccc(const cf_t *x, const cf_t *y, cf_t *z, const uint32_t len);
-SRSLTE_API void srslte_vec_sub_sss(const int16_t *x, const int16_t *y, int16_t *z, const uint32_t len);
 SRSLTE_API void srslte_vec_sum_sss(const int16_t *x, const int16_t *y, int16_t *z, const uint32_t len);
 
 /* substract two vectors z=x-y */
 SRSLTE_API void srslte_vec_sub_fff(const float *x, const float *y, float *z, const uint32_t len);
 SRSLTE_API void srslte_vec_sub_ccc(const cf_t *x, const cf_t *y, cf_t *z, const uint32_t len);
+SRSLTE_API void srslte_vec_sub_sss(const int16_t *x, const int16_t *y, int16_t *z, const uint32_t len);
+SRSLTE_API void srslte_vec_sub_bbb(const int8_t *x, const int8_t *y, int8_t *z, const uint32_t len);
 
 /* scalar product */
 SRSLTE_API void srslte_vec_sc_prod_cfc(const cf_t *x, const float h, cf_t *z, const uint32_t len);
@@ -97,8 +99,10 @@ SRSLTE_API void srslte_vec_sc_prod_fff(const float *x, const float h, float *z,
 
 SRSLTE_API void srslte_vec_convert_fi(const float *x, const float scale, int16_t *z, const uint32_t len);
 SRSLTE_API void srslte_vec_convert_if(const int16_t *x, const float scale, float *z, const uint32_t len);
+SRSLTE_API void srslte_vec_convert_fb(const float *x, const float scale, int8_t *z, const uint32_t len);
 
 SRSLTE_API void srslte_vec_lut_sss(const short *x, const unsigned short *lut, short *y, const uint32_t len);
+SRSLTE_API void srslte_vec_lut_bbb(const int8_t *x, const unsigned short *lut, int8_t *y, const uint32_t len);
 SRSLTE_API void srslte_vec_lut_sis(const short *x, const unsigned int *lut, short *y, const uint32_t len);
 
 /* vector product (element-wise) */
@@ -115,6 +119,10 @@ SRSLTE_API void srslte_vec_prod_conj_ccc(const cf_t *x, const cf_t *y, cf_t *z,
 SRSLTE_API void srslte_vec_prod_fff(const float *x, const float *y, float *z, const uint32_t len);
 SRSLTE_API void srslte_vec_prod_sss(const int16_t *x, const int16_t *y, int16_t *z, const uint32_t len);
 
+// Negate sign (scrambling)
+SRSLTE_API void srslte_vec_neg_sss(const int16_t *x, const int16_t *y, int16_t *z, const uint32_t len);
+SRSLTE_API void srslte_vec_neg_bbb(const int8_t *x, const int8_t *y, int8_t *z, const uint32_t len);
+
 /* Dot-product */
 SRSLTE_API cf_t srslte_vec_dot_prod_cfc(const cf_t *x, const float *y, const uint32_t len);
 SRSLTE_API cf_t srslte_vec_dot_prod_ccc(const cf_t *x, const cf_t *y, const uint32_t len);

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

@@ -62,6 +62,8 @@ SRSLTE_API void srslte_vec_sum_sss_simd(const int16_t *x, const int16_t *y, int1
 
 SRSLTE_API void srslte_vec_sub_sss_simd(const int16_t *x, const int16_t *y, int16_t *z, int len);
 
+SRSLTE_API void srslte_vec_sub_bbb_simd(const int8_t *x, const int8_t *y, int8_t *z, int len);
+
 SRSLTE_API float srslte_vec_acc_ff_simd(const float *x, int len);
 
 SRSLTE_API cf_t srslte_vec_acc_cc_simd(const cf_t *x, int len);
@@ -86,6 +88,10 @@ SRSLTE_API void srslte_vec_prod_ccc_c16_simd(const int16_t *a_re, const int16_t
 
 SRSLTE_API void srslte_vec_prod_sss_simd(const int16_t *x, const int16_t *y, int16_t *z, const int len);
 
+SRSLTE_API void srslte_vec_neg_sss_simd(const int16_t *x, const int16_t *y, int16_t *z, const int len);
+
+SRSLTE_API void srslte_vec_neg_bbb_simd(const int8_t *x, const int8_t *y, int8_t *z, const int len);
+
 SRSLTE_API void srslte_vec_prod_cfc_simd(const cf_t *x, const float *y, cf_t *z, const int len);
 
 SRSLTE_API void srslte_vec_prod_fff_simd(const float *x, const float *y, float *z, const int len);
@@ -120,10 +126,14 @@ SRSLTE_API void srslte_vec_abs_square_cf_simd(const cf_t *x, float *z, const int
 /* Other Functions */
 SRSLTE_API void srslte_vec_lut_sss_simd(const short *x, const unsigned short *lut, short *y, const int len);
 
+SRSLTE_API void srslte_vec_lut_bbb_simd(const int8_t *x, const unsigned short *lut, int8_t *y, const int len);
+
 SRSLTE_API void srslte_vec_convert_if_simd(const int16_t *x, float *z, const float scale, const int len);
 
 SRSLTE_API void srslte_vec_convert_fi_simd(const float *x, int16_t *z, const float scale, const int len);
 
+SRSLTE_API void srslte_vec_convert_fb_simd(const float *x, int8_t *z, const float scale, const int len);
+
 SRSLTE_API void srslte_vec_cp_simd(const cf_t *src, cf_t *dst, int len);
 
 SRSLTE_API void srslte_vec_interleave_simd(const cf_t *x, const cf_t *y, cf_t *z, const int len);

lib/src/common/threads.c

@@ -53,6 +53,7 @@ bool threads_new_rt_cpu(pthread_t *thread, void *(*start_routine) (void*), void
   cpu_set_t cpuset;
   bool attr_enable = false;
 
+#ifdef PER_THREAD_PRIO
   if (prio_offset >= 0) {
     param.sched_priority = sched_get_priority_max(SCHED_FIFO) - prio_offset;
     pthread_attr_init(&attr);
@@ -82,6 +83,25 @@ bool threads_new_rt_cpu(pthread_t *thread, void *(*start_routine) (void*), void
     }
     attr_enable = true;
   } else if (prio_offset == -2) {
+#else
+  // All threads have normal priority except prio_offset=0,1,2,3,4
+  if (prio_offset >= 0 && prio_offset < 5) {
+    param.sched_priority = 50;
+    pthread_attr_init(&attr);
+    if (pthread_attr_setinheritsched(&attr, PTHREAD_EXPLICIT_SCHED)) {
+      perror("pthread_attr_setinheritsched");
+    }
+    if (pthread_attr_setschedpolicy(&attr, SCHED_FIFO)) {
+      perror("pthread_attr_setschedpolicy");
+    }
+    if (pthread_attr_setschedparam(&attr, &param)) {
+      perror("pthread_attr_setschedparam");
+      fprintf(stderr, "Error not enough privileges to set Scheduling priority\n");
+    }
+    attr_enable = true;
+
+  } else {
+#endif
     param.sched_priority = 0;
     pthread_attr_init(&attr);
     if (pthread_attr_setinheritsched(&attr, PTHREAD_EXPLICIT_SCHED)) {

lib/src/phy/common/sequence.c

@@ -139,6 +139,7 @@ int srslte_sequence_LTE_pr(srslte_sequence_t *q, uint32_t len, uint32_t seed) {
   for (int i=0;i<len;i++) {
     q->c_float[i] = (1-2*q->c[i]);
     q->c_short[i] = (int16_t) q->c_float[i];
+    q->c_char[i]  = (int8_t) q->c_float[i];;
   }
   return SRSLTE_SUCCESS;
 }
@@ -164,6 +165,10 @@ int srslte_sequence_init(srslte_sequence_t *q, uint32_t len) {
     if (!q->c_short) {
       return SRSLTE_ERROR;
     }
+    q->c_char = srslte_vec_malloc(len * sizeof(int8_t));
+    if (!q->c_char) {
+      return SRSLTE_ERROR;
+    }
     q->max_len = len;
   }
   return SRSLTE_SUCCESS;
@@ -182,6 +187,9 @@ void srslte_sequence_free(srslte_sequence_t *q) {
   if (q->c_short) {
     free(q->c_short);
   }
+  if (q->c_char) {
+    free(q->c_char);
+  }
   bzero(q, sizeof(srslte_sequence_t));
 }
 

lib/src/phy/fec/rm_turbo.c

@@ -43,13 +43,15 @@
 #endif
 
 #ifdef LV_HAVE_SSE
-#include <smmintrin.h>
+#include <x86intrin.h>
-int srslte_rm_turbo_rx_lut_sse(int16_t *input, int16_t *output, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx);
+int srslte_rm_turbo_rx_lut_sse(int16_t *input, int16_t *output, uint16_t *deinter, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx);
+int srslte_rm_turbo_rx_lut_sse_8bit(int8_t *input, int8_t *output, uint16_t *deinter, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx);
 #endif
 
 #ifdef LV_HAVE_AVX
-#include <immintrin.h>
+#include <x86intrin.h>
-int srslte_rm_turbo_rx_lut_avx(int16_t *input, int16_t *output, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx);
+int srslte_rm_turbo_rx_lut_avx(int16_t *input, int16_t *output, uint16_t *deinter, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx);
+int srslte_rm_turbo_rx_lut_avx_8bit(int8_t *input, int8_t *output, uint16_t *deinter, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx);
 #endif
 
 #define NCOLS 32
@@ -69,9 +71,28 @@ static int k0_vec[SRSLTE_NOF_TC_CB_SIZES][4][2];
 static bool rm_turbo_tables_generated = false;
 
 
+// Store deinterleaver version for sub-block turbo decoder
+#if SRSLTE_TDEC_EXPECT_INPUT_SB == 1
+// Prepare bit for sub-block decoder processing. These are the nof subblock sizes
+#define NOF_DEINTER_TABLE_SB_IDX 3
+const static int deinter_table_sb_idx[NOF_DEINTER_TABLE_SB_IDX] = {8, 16, 32};
+int deinter_table_idx_from_sb_len(uint32_t nof_subblocks) {
+  for (int i=0;i<NOF_DEINTER_TABLE_SB_IDX;i++) {
+    if (deinter_table_sb_idx[i] == nof_subblocks) {
+      return i;
+    }
+  }
+  if (nof_subblocks != 0) {
+    fprintf(stderr, "Error number of sub-blocks %d not supported in RM\n", nof_subblocks);
+  }
+  return -1;
+}
+static uint16_t deinterleaver_sb[NOF_DEINTER_TABLE_SB_IDX][192][4][18448];
+#endif
+
 static uint16_t temp_table1[3*6176], temp_table2[3*6176];
 
-void srslte_rm_turbo_gentable_systematic(uint16_t *table_bits, int k0_vec[4][2], uint32_t nrows, int ndummy) {
+static void srslte_rm_turbo_gentable_systematic(uint16_t *table_bits, int k0_vec[4][2], uint32_t nrows, int ndummy) {
 
   bool last_is_null=true;
   int k_b=0, buff_idx=0;
@@ -96,7 +117,7 @@ void srslte_rm_turbo_gentable_systematic(uint16_t *table_bits, int k0_vec[4][2],
   }
 }
 
-void srslte_rm_turbo_gentable_parity(uint16_t *table_parity, int k0_vec[4][2], int offset, uint16_t nrows, int ndummy) {
+static void srslte_rm_turbo_gentable_parity(uint16_t *table_parity, int k0_vec[4][2], int offset, uint16_t nrows, int ndummy) {
   
   bool last_is_null=true;
   int k_b=0, buff_idx0=0;
@@ -140,9 +161,7 @@ void srslte_rm_turbo_gentable_parity(uint16_t *table_parity, int k0_vec[4][2], i
   }
 }
 
-
+static void srslte_rm_turbo_gentable_receive(uint16_t *table, uint32_t cb_len, uint32_t rv_idx)
-
-void srslte_rm_turbo_gentable_receive(uint16_t *table, uint32_t cb_len, uint32_t rv_idx) 
 {
   
   int nrows = (uint32_t) (cb_len / 3 - 1) / NCOLS + 1;
@@ -217,13 +236,40 @@ void srslte_rm_turbo_gentable_receive(uint16_t *table, uint32_t cb_len, uint32_t
     table[i] = temp_table2[temp_table1[i]];
   }
 }
+#if SRSLTE_TDEC_EXPECT_INPUT_SB==1
+#define inter(x,win) ((x%(long_cb/win))*(win)+x/(long_cb/win))
+
+/* Prepare output for sliding window decoder:
+ * (0..long_cb-1) bits are systematic
+ * (long_cb..2*long_cb-1) are parity0
+ * (2*long_cb..3*long_cb-1) are parity1
+ * then tail bits
+ *
+ * Within each block, bits are interleaved every nof_sb
+ */
+static void interleave_table_sb(uint16_t *in, uint16_t *out, uint32_t cb_idx, uint32_t nof_sb)
+{
+  int long_cb = srslte_cbsegm_cbsize(cb_idx);
+  int out_len = 3*long_cb+12;
+  for (int i=0;i<out_len;i++) {
+    // Do not change tail bit order
+    if (in[i] < 3*long_cb) {
+
+      // align to 32 bytes (warning: must be same alignment as in rm_turbo.c)
+      out[i] = (in[i]%3)*(long_cb+32)+inter(in[i]/3,nof_sb);
+    } else {
+      out[i] = (in[i]-3*long_cb)+3*(long_cb+32);
+    }
+  }
+}
+#endif
 
 void srslte_rm_turbo_gentables() {
   if (!rm_turbo_tables_generated) {
     rm_turbo_tables_generated = true; 
     for (int cb_idx=0;cb_idx<SRSLTE_NOF_TC_CB_SIZES;cb_idx++) {
-      int cb_len=srslte_cbsegm_cbsize(cb_idx);
+      int cb_len = srslte_cbsegm_cbsize(cb_idx);
-      int in_len=3*cb_len+12;
+      int in_len = 3 * cb_len + 12;
 
       int nrows = (in_len / 3 - 1) / NCOLS + 1;
       int K_p = nrows * NCOLS;
@@ -232,20 +278,27 @@ void srslte_rm_turbo_gentables() {
         ndummy = 0;
       }
 
-      for (int i=0;i<4;i++) {
+      for (int i = 0; i < 4; i++) {
-        k0_vec[cb_idx][i][0] = nrows * (2 * (uint16_t) ceilf((float) (3*K_p) / (float) (8 * nrows)) * i + 2);
+        k0_vec[cb_idx][i][0] = nrows * (2 * (uint16_t) ceilf((float) (3 * K_p) / (float) (8 * nrows)) * i + 2);
         k0_vec[cb_idx][i][1] = -1;
       }
       srslte_rm_turbo_gentable_systematic(interleaver_systematic_bits[cb_idx], k0_vec[cb_idx], nrows, ndummy);
       srslte_bit_interleaver_init(&bit_interleavers_systematic_bits[cb_idx], interleaver_systematic_bits[cb_idx],
-                                  (uint32_t) srslte_cbsegm_cbsize(cb_idx)+4);
+                                  (uint32_t) srslte_cbsegm_cbsize(cb_idx) + 4);
 
-      srslte_rm_turbo_gentable_parity(interleaver_parity_bits[cb_idx], k0_vec[cb_idx], in_len/3, nrows, ndummy);
+      srslte_rm_turbo_gentable_parity(interleaver_parity_bits[cb_idx], k0_vec[cb_idx], in_len / 3, nrows, ndummy);
       srslte_bit_interleaver_init(&bit_interleavers_parity_bits[cb_idx], interleaver_parity_bits[cb_idx],
-                                  (uint32_t) (srslte_cbsegm_cbsize(cb_idx)+4)*2);
+                                  (uint32_t) (srslte_cbsegm_cbsize(cb_idx) + 4) * 2);
 
-      for (int i=0;i<4;i++) {
+      for (int i = 0; i < 4; i++) {
         srslte_rm_turbo_gentable_receive(deinterleaver[cb_idx][i], in_len, i);
+
+#if SRSLTE_TDEC_EXPECT_INPUT_SB == 1
+        for (uint32_t s = 0; s < NOF_DEINTER_TABLE_SB_IDX; s++) {
+          interleave_table_sb(deinterleaver[cb_idx][i], deinterleaver_sb[s][cb_idx][i], cb_idx,
+                              deinter_table_sb_idx[s]);
+        }
+#endif
       }
     }
   }
@@ -322,6 +375,10 @@ int srslte_rm_turbo_tx_lut(uint8_t *w_buff, uint8_t *systematic, uint8_t *parity
   }
 }
 
+int srslte_rm_turbo_rx_lut(int16_t *input, int16_t *output, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx)
+{
+  return srslte_rm_turbo_rx_lut_(input, output, in_len, cb_idx, rv_idx, true);
+}
 /**
  * Undoes rate matching for LTE Turbo Coder. Expands rate matched buffer to full size buffer.
  *
@@ -331,37 +388,93 @@ int srslte_rm_turbo_tx_lut(uint8_t *w_buff, uint8_t *systematic, uint8_t *parity
  * @param[in] rv_idx Redundancy Version from DCI control message
  * @return Error code
  */
-int srslte_rm_turbo_rx_lut(int16_t *input, int16_t *output, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx) 
+int srslte_rm_turbo_rx_lut_(int16_t *input, int16_t *output, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx, bool enable_input_tdec)
 {
 
-#ifdef LV_HAVE_AVX
+  if (rv_idx < 4 && cb_idx < SRSLTE_NOF_TC_CB_SIZES) {
-  return srslte_rm_turbo_rx_lut_avx(input, output, in_len, cb_idx, rv_idx);
+
+#if SRSLTE_TDEC_EXPECT_INPUT_SB == 1
+  int cb_len=srslte_cbsegm_cbsize(cb_idx);
+  int idx = deinter_table_idx_from_sb_len(srslte_tdec_autoimp_get_subblocks(cb_len));
+  uint16_t *deinter = NULL;
+  if (idx < 0 || !enable_input_tdec) {
+    deinter = deinterleaver[cb_idx][rv_idx];
+  } else if (idx < NOF_DEINTER_TABLE_SB_IDX) {
+    deinter = deinterleaver_sb[idx][cb_idx][rv_idx];
+  } else {
+    fprintf(stderr, "Sub-block size index %d not supported in srslte_rm_turbo_rx_lut()\n", idx);
+    return -1;
+  }
 #else
+  uint16_t *deinter = deinterleaver[cb_idx][rv_idx];
+#endif
+
+  #ifdef LV_HAVE_AVX
+    return srslte_rm_turbo_rx_lut_avx(input, output, deinter, in_len, cb_idx, rv_idx);
+  #else
   #ifdef LV_HAVE_SSE
-    return srslte_rm_turbo_rx_lut_sse(input, output, in_len, cb_idx, rv_idx);
+    return srslte_rm_turbo_rx_lut_sse(input, output, deinter, in_len, cb_idx, rv_idx);
   #else
-    if (rv_idx < 4 && cb_idx < SRSLTE_NOF_TC_CB_SIZES) {
       uint32_t out_len = 3*srslte_cbsegm_cbsize(cb_idx)+12;
-      uint16_t *deinter = deinterleaver[cb_idx][rv_idx];
+
       for (int i=0;i<in_len;i++) {
         output[deinter[i%out_len]] += input[i];
       }
       return 0;    
+  #endif
+#endif
+  } else {
+    printf("Invalid inputs rv_idx=%d, cb_idx=%d\n", rv_idx, cb_idx);
+    return SRSLTE_ERROR_INVALID_INPUTS;
+  }
+}
+
+int srslte_rm_turbo_rx_lut_8bit(int8_t *input, int8_t *output, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx)
+{
+  if (rv_idx < 4 && cb_idx < SRSLTE_NOF_TC_CB_SIZES) {
+
+#if SRSLTE_TDEC_EXPECT_INPUT_SB == 1
+    int cb_len=srslte_cbsegm_cbsize(cb_idx);
+    int idx = deinter_table_idx_from_sb_len(srslte_tdec_autoimp_get_subblocks_8bit(cb_len));
+    uint16_t *deinter = NULL;
+    if (idx < 0) {
+      deinter = deinterleaver[cb_idx][rv_idx];
+    } else if (idx < NOF_DEINTER_TABLE_SB_IDX) {
+      deinter = deinterleaver_sb[idx][cb_idx][rv_idx];
     } else {
-      printf("Invalid inputs rv_idx=%d, cb_idx=%d\n", rv_idx, cb_idx);
+      fprintf(stderr, "Sub-block size index %d not supported in srslte_rm_turbo_rx_lut()\n", idx);
-      return SRSLTE_ERROR_INVALID_INPUTS; 
+      return -1;
     }
-  #endif
+#else
+    uint16_t *deinter = deinterleaver[cb_idx][rv_idx];
 #endif
+
+    // FIXME: AVX version of rm_turbo_rx_lut not working
+    // Warning: Need to check if 8-bit sse version is correct
+
+#ifdef LV_HAVE_SSE
+    return srslte_rm_turbo_rx_lut_sse_8bit(input, output, deinter, in_len, cb_idx, rv_idx);
+#else
+    uint32_t out_len = 3*srslte_cbsegm_cbsize(cb_idx)+12;
+
+    for (int i=0;i<in_len;i++) {
+      output[deinter[i%out_len]] += input[i];
+    }
+    return 0;
+#endif
+  } else {
+    printf("Invalid inputs rv_idx=%d, cb_idx=%d\n", rv_idx, cb_idx);
+    return SRSLTE_ERROR_INVALID_INPUTS;
+  }
 }
 
+
 #ifdef LV_HAVE_SSE
 
-int srslte_rm_turbo_rx_lut_sse(int16_t *input, int16_t *output, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx) 
+int srslte_rm_turbo_rx_lut_sse(int16_t *input, int16_t *output, uint16_t *deinter, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx)
 {
   if (rv_idx < 4 && cb_idx < SRSLTE_NOF_TC_CB_SIZES) {
     uint32_t out_len = 3*srslte_cbsegm_cbsize(cb_idx)+12;
-    uint16_t *deinter = deinterleaver[cb_idx][rv_idx];
 
     const __m128i* xPtr   = (const __m128i*) input;
     const __m128i* lutPtr = (const __m128i*) deinter;
@@ -427,6 +540,97 @@ int srslte_rm_turbo_rx_lut_sse(int16_t *input, int16_t *output, uint32_t in_len,
   }
 }
 
+int srslte_rm_turbo_rx_lut_sse_8bit(int8_t *input, int8_t *output, uint16_t *deinter, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx)
+{
+  if (rv_idx < 4 && cb_idx < SRSLTE_NOF_TC_CB_SIZES) {
+    uint32_t out_len = 3*srslte_cbsegm_cbsize(cb_idx)+12;
+
+    const __m128i* xPtr   = (const __m128i*) input;
+    const __m128i* lutPtr = (const __m128i*) deinter;
+    __m128i xVal, lutVal1, lutVal2;
+
+    /* Simplify load if we do not need to wrap (ie high rates) */
+    if (in_len <= out_len) {
+      for (int i=0;i<in_len/16;i++) {
+        xVal   = _mm_loadu_si128(xPtr);
+        xPtr ++;
+        lutVal1 = _mm_loadu_si128(lutPtr);
+        lutPtr++;
+        lutVal2 = _mm_loadu_si128(lutPtr);
+        lutPtr ++;
+
+        for (int j=0;j<8;j++) {
+          int8_t x  = (int8_t)  _mm_extract_epi8(xVal,   j);
+          uint16_t l = (uint16_t) _mm_extract_epi16(lutVal1, j);
+          output[l] += x;
+        }
+        for (int j=0;j<8;j++) {
+          int8_t x  = (int8_t)  _mm_extract_epi8(xVal,   j+8);
+          uint16_t l = (uint16_t) _mm_extract_epi16(lutVal2, j);
+          output[l] += x;
+        }
+      }
+      for (int i=16*(in_len/16);i<in_len;i++) {
+        output[deinter[i%out_len]] += input[i];
+      }
+    } else {
+      int intCnt = 16;
+      int inputCnt = 0;
+      int nwrapps = 0;
+      while(inputCnt < in_len - 16) {
+        xVal   = _mm_loadu_si128(xPtr);
+        xPtr ++;
+        lutVal1 = _mm_loadu_si128(lutPtr);
+        lutPtr++;
+        lutVal2 = _mm_loadu_si128(lutPtr);
+        lutPtr ++;
+
+        for (int j=0;j<8;j++) {
+          int8_t x  = (int8_t)  _mm_extract_epi8(xVal,   j);
+          uint16_t l = (uint16_t) _mm_extract_epi16(lutVal1, j);
+          output[l] += x;
+        }
+        for (int j=0;j<8;j++) {
+          int8_t x  = (int8_t)  _mm_extract_epi8(xVal,   j+8);
+          uint16_t l = (uint16_t) _mm_extract_epi16(lutVal2, j);
+          output[l] += x;
+        }
+        intCnt   += 16;
+        inputCnt += 16;
+        if (intCnt >= out_len && inputCnt < in_len - 16) {
+          /* Copy last elements */
+          if ((out_len%16) == 12) {
+            for (int j=(nwrapps+1)*out_len-12;j<(nwrapps+1)*out_len;j++) {
+              output[deinter[j%out_len]] += input[j];
+              inputCnt++;
+            }
+          } else {
+            for (int j=(nwrapps+1)*out_len-4;j<(nwrapps+1)*out_len;j++) {
+              output[deinter[j%out_len]] += input[j];
+              inputCnt++;
+            }
+          }
+          /* And wrap pointers */
+          nwrapps++;
+          intCnt = 16;
+          xPtr   = (const __m128i*) &input[nwrapps*out_len];
+          lutPtr = (const __m128i*) deinter;
+        }
+      }
+      for (int i=inputCnt;i<in_len;i++) {
+        output[deinter[i%out_len]] += input[i];
+      }
+    }
+
+
+    return 0;
+  } else {
+    printf("Invalid inputs rv_idx=%d, cb_idx=%d\n", rv_idx, cb_idx);
+    return SRSLTE_ERROR_INVALID_INPUTS;
+  }
+}
+
+
 #endif
 
 
@@ -436,12 +640,10 @@ int srslte_rm_turbo_rx_lut_sse(int16_t *input, int16_t *output, uint32_t in_len,
                        l = (uint16_t) _mm256_extract_epi16(lutVal, j);\
                        output[l] += x;
 
-
+int srslte_rm_turbo_rx_lut_avx(int16_t *input, int16_t *output, uint16_t *deinter, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx)
-int srslte_rm_turbo_rx_lut_avx(int16_t *input, int16_t *output, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx) 
 {
   if (rv_idx < 4 && cb_idx < SRSLTE_NOF_TC_CB_SIZES) {
     uint32_t out_len = 3*srslte_cbsegm_cbsize(cb_idx)+12;
-    uint16_t *deinter = deinterleaver[cb_idx][rv_idx];
 
     const __m256i* xPtr   = (const __m256i*) input;
     const __m256i* lutPtr = (const __m256i*) deinter;
@@ -539,6 +741,155 @@ int srslte_rm_turbo_rx_lut_avx(int16_t *input, int16_t *output, uint32_t in_len,
   }
 }
 
+#define SAVE_OUTPUT8(j) x = (int8_t) _mm256_extract_epi8(xVal,   j);\
+                        l = (uint16_t) _mm256_extract_epi16(lutVal1, j);\
+                        output[l] += x;
+
+#define SAVE_OUTPUT8_2(j) x = (int8_t) _mm256_extract_epi8(xVal,   j+8);\
+                          l = (uint16_t) _mm256_extract_epi16(lutVal2, j);\
+                          output[l] += x;
+
+
+int srslte_rm_turbo_rx_lut_avx_8bit(int8_t *input, int8_t *output, uint16_t *deinter, uint32_t in_len, uint32_t cb_idx, uint32_t rv_idx)
+{
+  if (rv_idx < 4 && cb_idx < SRSLTE_NOF_TC_CB_SIZES) {
+    uint32_t out_len = 3*srslte_cbsegm_cbsize(cb_idx)+12;
+
+    const __m256i* xPtr   = (const __m256i*) input;
+    const __m256i* lutPtr = (const __m256i*) deinter;
+    __m256i xVal, lutVal1, lutVal2;
+
+    int8_t x;
+    uint16_t l;
+
+    /* Simplify load if we do not need to wrap (ie high rates) */
+    if (in_len <= out_len) {
+      for (int i=0;i<in_len/32;i++) {
+        xVal   = _mm256_loadu_si256(xPtr);
+        xPtr ++;
+        lutVal1 = _mm256_loadu_si256(lutPtr);
+        lutPtr ++;
+        lutVal2 = _mm256_loadu_si256(lutPtr);
+        lutPtr ++;
+
+        SAVE_OUTPUT8(0);
+        SAVE_OUTPUT8(1);
+        SAVE_OUTPUT8(2);
+        SAVE_OUTPUT8(3);
+        SAVE_OUTPUT8(4);
+        SAVE_OUTPUT8(5);
+        SAVE_OUTPUT8(6);
+        SAVE_OUTPUT8(7);
+
+        SAVE_OUTPUT8(8);
+        SAVE_OUTPUT8(9);
+        SAVE_OUTPUT8(10);
+        SAVE_OUTPUT8(11);
+        SAVE_OUTPUT8(12);
+        SAVE_OUTPUT8(13);
+        SAVE_OUTPUT8(14);
+        SAVE_OUTPUT8(15);
+
+        SAVE_OUTPUT8_2(0);
+        SAVE_OUTPUT8_2(1);
+        SAVE_OUTPUT8_2(2);
+        SAVE_OUTPUT8_2(3);
+        SAVE_OUTPUT8_2(4);
+        SAVE_OUTPUT8_2(5);
+        SAVE_OUTPUT8_2(6);
+        SAVE_OUTPUT8_2(7);
+
+        SAVE_OUTPUT8_2(8);
+        SAVE_OUTPUT8_2(9);
+        SAVE_OUTPUT8_2(10);
+        SAVE_OUTPUT8_2(11);
+        SAVE_OUTPUT8_2(12);
+        SAVE_OUTPUT8_2(13);
+        SAVE_OUTPUT8_2(14);
+        SAVE_OUTPUT8_2(15);
+
+      }
+      for (int i=32*(in_len/32);i<in_len;i++) {
+        output[deinter[i%out_len]] += input[i];
+      }
+    } else {
+      printf("wraps not implemented!\n");
+#ifdef kk
+      int intCnt = 32;
+      int inputCnt = 0;
+      int nwrapps = 0;
+      while(inputCnt < in_len - 32) {
+        xVal   = _mm256_loadu_si256(xPtr);
+        xPtr ++;
+        lutVal1 = _mm256_loadu_si256(lutPtr);
+        lutPtr ++;
+        lutVal2 = _mm256_loadu_si256(lutPtr);
+        lutPtr ++;
+
+        SAVE_OUTPUT8(0);
+        SAVE_OUTPUT8(1);
+        SAVE_OUTPUT8(2);
+        SAVE_OUTPUT8(3);
+        SAVE_OUTPUT8(4);
+        SAVE_OUTPUT8(5);
+        SAVE_OUTPUT8(6);
+        SAVE_OUTPUT8(7);
+
+        SAVE_OUTPUT8(8);
+        SAVE_OUTPUT8(9);
+        SAVE_OUTPUT8(10);
+        SAVE_OUTPUT8(11);
+        SAVE_OUTPUT8(12);
+        SAVE_OUTPUT8(13);
+        SAVE_OUTPUT8(14);
+        SAVE_OUTPUT8(15);
+
+        SAVE_OUTPUT8_2(0);
+        SAVE_OUTPUT8_2(1);
+        SAVE_OUTPUT8_2(2);
+        SAVE_OUTPUT8_2(3);
+        SAVE_OUTPUT8_2(4);
+        SAVE_OUTPUT8_2(5);
+        SAVE_OUTPUT8_2(6);
+        SAVE_OUTPUT8_2(7);
+
+        SAVE_OUTPUT8_2(8);
+        SAVE_OUTPUT8_2(9);
+        SAVE_OUTPUT8_2(10);
+        SAVE_OUTPUT8_2(11);
+        SAVE_OUTPUT8_2(12);
+        SAVE_OUTPUT8_2(13);
+        SAVE_OUTPUT8_2(14);
+        SAVE_OUTPUT8_2(15);
+        
+        intCnt   += 32;
+        inputCnt += 32;
+        if (intCnt >= out_len && inputCnt < in_len - 32) {
+          printf("warning rate matching wrapping remainder %d\n", out_len%32);
+          /* Copy last elements */
+          for (int j=(nwrapps+1)*out_len-(out_len%32) ;j<(nwrapps+1)*out_len;j++) {
+            output[deinter[j%out_len]] += input[j];
+            inputCnt++;
+          }
+          /* And wrap pointers */
+          nwrapps++;
+          intCnt = 32;
+          xPtr   = (const __m256i*) &input[nwrapps*out_len];
+          lutPtr = (const __m256i*) deinter;
+        }
+      }
+      for (int i=inputCnt;i<in_len;i++) {
+        output[deinter[i%out_len]] += input[i];
+      }
+#endif
+    }
+    return 0;
+  } else {
+    printf("Invalid inputs rv_idx=%d, cb_idx=%d\n", rv_idx, cb_idx);
+    return SRSLTE_ERROR_INVALID_INPUTS;
+  }
+}
+
 #endif
 
 

lib/src/phy/fec/tc_interl_lte.c

@@ -66,7 +66,16 @@ const uint32_t f2_list[SRSLTE_NOF_TC_CB_SIZES] = { 10, 12, 42, 16, 18, 20, 22, 2
     280, 142, 480, 146, 444, 120, 152, 462, 234, 158, 80, 96, 902, 166, 336,
     170, 86, 174, 176, 178, 120, 182, 184, 186, 94, 190, 480 };
 
-int srslte_tc_interl_LTE_gen(srslte_tc_interl_t *h, uint32_t long_cb) {
+int srslte_tc_interl_LTE_gen(srslte_tc_interl_t *h, uint32_t long_cb)
+{
+  return srslte_tc_interl_LTE_gen_interl(h, long_cb, 1);
+}
+
+#define deinter(x,win) ((x%(long_cb/win))*(win)+x/(long_cb/win))
+#define inter(x,win) ((x%win)*(long_cb/win)+x/win)
+
+
+int srslte_tc_interl_LTE_gen_interl(srslte_tc_interl_t *h, uint32_t long_cb, uint32_t interl_win) {
   uint32_t cb_table_idx, f1, f2;
   uint64_t i, j;
 
@@ -92,6 +101,19 @@ int srslte_tc_interl_LTE_gen(srslte_tc_interl_t *h, uint32_t long_cb) {
     h->forward[i] = (uint32_t) j;
     h->reverse[j] = (uint32_t) i;
   }
+  if (interl_win != 1) {
+    uint16_t *f = malloc(long_cb*sizeof(uint16_t));
+    uint16_t *r = malloc(long_cb*sizeof(uint16_t));
+    memcpy(f, h->forward, long_cb*sizeof(uint16_t));
+    memcpy(r, h->reverse, long_cb*sizeof(uint16_t));
+    for (i = 0; i < long_cb; i++) {
+      h->forward[i] = deinter(f[inter(i,interl_win)],interl_win);
+      h->reverse[i] = deinter(r[inter(i,interl_win)],interl_win);
+    }
+    free(f);
+    free(r);
+  }
+
   return 0;
 
 }

lib/src/phy/fec/tc_interl_umts.c

@@ -30,7 +30,6 @@
 #include <stdint.h>
 
 #include "srslte/phy/fec/tc_interl.h"
-#include "srslte/phy/fec/turbocoder.h"
 
 #define TURBO_SRSLTE_TCOD_RATE 	3
 

lib/src/phy/fec/test/rm_turbo_test.c

@@ -183,7 +183,7 @@ int main(int argc, char **argv) {
       srslte_rm_turbo_rx(buff_f, BUFFSZ, rm_bits_f, nof_e_bits, bits_f, long_cb_enc, rv_idx, 0);
 
       bzero(bits2_s, long_cb_enc*sizeof(short));
-      srslte_rm_turbo_rx_lut(rm_bits_s, bits2_s, nof_e_bits, cb_idx, rv_idx);
+      srslte_rm_turbo_rx_lut_(rm_bits_s, bits2_s, nof_e_bits, cb_idx, rv_idx, false);
 
       for (int i=0;i<long_cb_enc;i++) {
         if (bits_f[i] != bits2_s[i]) {

lib/src/phy/fec/test/turbocoder_test.c

@@ -96,8 +96,15 @@ int main(int argc, char **argv) {
       }
     }
 
+    /* Create CRC for Transport Block, it is not currently used but it is required */
+    srslte_crc_t crc_tb = {};
+    if (srslte_crc_init(&crc_tb, SRSLTE_LTE_CRC24A, 24)) {
+      printf("error initialising CRC\n");
+      exit(-1);
+    }
+
     srslte_tcod_encode(&tcod, input_bits, output_bits, long_cb);
-    srslte_tcod_encode_lut(&tcod, NULL, input_bytes, parity, len);
+    srslte_tcod_encode_lut(&tcod, &crc_tb, NULL, input_bytes, parity, len, false);
 
     srslte_bit_unpack_vector(parity, parity_bits, 2*(long_cb+4));
     

lib/src/phy/fec/test/turbodecoder_test.c

@@ -52,12 +52,14 @@ int test_known_data = 0;
 int test_errors = 0;
 int nof_repetitions = 1;
 
+srslte_tdec_impl_type_t tdec_type;
+
 #define SNR_POINTS      4
 #define SNR_MIN         1.0
 #define SNR_MAX         8.0
 
 void usage(char *prog) {
-  printf("Usage: %s [nlesv]\n", prog);
+  printf("Usage: %s [kcinNledts]\n", prog);
   printf(
       "\t-k Test with known data (ignores frame_length) [Default disabled]\n");
   printf("\t-c nof_cb in parallel [Default %d]\n", nof_cb);
@@ -66,13 +68,14 @@ void usage(char *prog) {
   printf("\t-N nof_repetitions [Default %d]\n", nof_repetitions);
   printf("\t-l frame_length [Default %d]\n", frame_length);
   printf("\t-e ebno in dB [Default scan]\n");
+  printf("\t-d Decoder implementation type: 0: Generic, 1: SSE, 2: SSE-window\n");
   printf("\t-t test: check errors on exit [Default disabled]\n");
   printf("\t-s seed [Default 0=time]\n");
 }
 
 void parse_args(int argc, char **argv) {
   int opt;
-  while ((opt = getopt(argc, argv, "cinNlstvekt")) != -1) {
+  while ((opt = getopt(argc, argv, "kcinNledts")) != -1) {
     switch (opt) {
     case 'c':
       nof_cb = atoi(argv[optind]);
@@ -95,6 +98,9 @@ void parse_args(int argc, char **argv) {
     case 'l':
       frame_length = atoi(argv[optind]);
       break;
+    case 'd':
+      tdec_type = (srslte_tdec_impl_type_t) atoi(argv[optind]);
+      break;
     case 'e':
       ebno_db = atof(argv[optind]);
       break;
@@ -117,7 +123,7 @@ 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_MAX_NPAR], *symbols;
+  uint8_t *data_tx, *data_rx, *data_rx_bytes, *symbols;
   uint32_t i, j;
   float var[SNR_POINTS];
   uint32_t snr_points;
@@ -159,12 +165,10 @@ int main(int argc, char **argv) {
     perror("malloc");
     exit(-1);
   }
-  for (int cb=0;cb<SRSLTE_TDEC_MAX_NPAR;cb++) {
+  data_rx_bytes = srslte_vec_malloc(frame_length * sizeof(uint8_t));
-    data_rx_bytes[cb] = srslte_vec_malloc(frame_length * sizeof(uint8_t));
+  if (!data_rx_bytes) {
-    if (!data_rx_bytes[cb]) {
+    perror("malloc");
-      perror("malloc");
+    exit(-1);
-      exit(-1);
-    }    
   }
 
   symbols = srslte_vec_malloc(coded_length * sizeof(uint8_t));
@@ -193,11 +197,13 @@ int main(int argc, char **argv) {
     exit(-1);
   }
 
-  if (srslte_tdec_init(&tdec, frame_length)) {
+  if (srslte_tdec_init_manual(&tdec, frame_length, tdec_type)) {
     fprintf(stderr, "Error initiating Turbo decoder\n");
     exit(-1);
   }
 
+  srslte_tdec_force_not_sb(&tdec);
+
   float ebno_inc, esno_db;
   ebno_inc = (SNR_MAX - SNR_MIN) / SNR_POINTS;
   if (ebno_db == 100.0) {
@@ -246,7 +252,7 @@ int main(int argc, char **argv) {
       }
 
       /* decoder */
-      srslte_tdec_reset(&tdec, frame_length);
+      srslte_tdec_new_cb(&tdec, frame_length);
 
       uint32_t t;
       if (nof_iterations == -1) {
@@ -255,36 +261,21 @@ int main(int argc, char **argv) {
         t = nof_iterations;
       }
 
-      int16_t *input[SRSLTE_TDEC_MAX_NPAR];
-      uint8_t *output[SRSLTE_TDEC_MAX_NPAR];
-      
-      for (int n=0;n<SRSLTE_TDEC_MAX_NPAR;n++) {
-        if (n < nof_cb) {
-          input[n] = llr_s;
-          output[n] = data_rx_bytes[n];
-        } else {
-          input[n] = NULL;
-          output[n] = NULL;
-        }
-      }
-
       gettimeofday(&tdata[1], NULL); 
       for (int k=0;k<nof_repetitions;k++) { 
-        srslte_tdec_run_all_par(&tdec, input, output, t, frame_length);        
+        srslte_tdec_run_all(&tdec, llr_s, data_rx_bytes, t, frame_length);
       }
       gettimeofday(&tdata[2], NULL);
       get_time_interval(tdata);
-      mean_usec = (float) mean_usec * 0.9 + (float) (tdata[0].tv_usec/nof_repetitions) * 0.1;
+      mean_usec = (tdata[0].tv_sec*1e6+tdata[0].tv_usec)/nof_repetitions;
       
       frame_cnt++;
       uint32_t errors_this = 0; 
-      for (int cb=0;cb<nof_cb;cb++) {
+      srslte_bit_unpack_vector(data_rx_bytes, data_rx, frame_length);
-        srslte_bit_unpack_vector(data_rx_bytes[cb], data_rx, frame_length);
 
-        errors_this=srslte_bit_diff(data_tx, data_rx, frame_length);
+      errors_this=srslte_bit_diff(data_tx, data_rx, frame_length);
-        //printf("error[%d]=%d\n", cb, errors_this);
+      //printf("error[%d]=%d\n", cb, errors_this);
-        errors += errors_this;
+      errors += errors_this;
-      }
       printf("Eb/No: %2.2f %10d/%d   ", SNR_MIN + i * ebno_inc, frame_cnt, nof_frames);
       printf("BER: %.2e  ", (float) errors / (nof_cb*frame_cnt * frame_length));
       printf("%3.1f Mbps (%6.2f usec)", (float) (nof_cb*frame_length) / mean_usec, mean_usec);
@@ -301,10 +292,8 @@ int main(int argc, char **argv) {
     }
   }
 
-  for (int cb=0;cb<SRSLTE_TDEC_MAX_NPAR;cb++) {
+  if (data_rx_bytes) {
-    if (data_rx_bytes[cb]) {
+    free(data_rx_bytes);
-      free(data_rx_bytes[cb]);
-    }
   }
   free(data_tx);
   free(symbols);

lib/src/phy/fec/turbocoder.c

@@ -193,7 +193,13 @@ int srslte_tcod_encode(srslte_tcod_t *h, uint8_t *input, uint8_t *output, uint32
 }
 
 /* Expects bytes and produces bytes. The systematic and parity bits are interlaced in the output */
-int srslte_tcod_encode_lut(srslte_tcod_t *h, srslte_crc_t *crc, uint8_t *input, uint8_t *parity, uint32_t cblen_idx)
+int srslte_tcod_encode_lut(srslte_tcod_t *h,
+                           srslte_crc_t *crc_tb,
+                           srslte_crc_t *crc_cb,
+                           uint8_t *input,
+                           uint8_t *parity,
+                           uint32_t cblen_idx,
+                           bool last_cb)
 {
   if (cblen_idx < 188) {
     uint32_t long_cb = (uint32_t) srslte_cbsegm_cbsize(cblen_idx);
@@ -204,20 +210,24 @@ int srslte_tcod_encode_lut(srslte_tcod_t *h, srslte_crc_t *crc, uint8_t *input,
     }
 
     /* Reset CRC */
-    if (crc) {
+    if (crc_cb) {
-      srslte_crc_set_init(crc, 0);
+      srslte_crc_set_init(crc_cb, 0);
     }
 
     /* Parity bits for the 1st constituent encoders */
     uint8_t state0 = 0;
-    if (crc) {
+    if (crc_cb) {
+      int block_size_nocrc = (long_cb - crc_cb->order - ((last_cb) ? crc_tb->order : 0)) / 8;
 
       /* if CRC pointer is given */
-      for (int i = 0; i < (long_cb - crc->order) / 8; i++) {
+      for (int i = 0; i < block_size_nocrc; i++) {
         uint8_t in = input[i];
 
-        /* Put byte in CRC and save latest checksum */
+        /* Put byte in TB CRC and save latest checksum */
-        srslte_crc_checksum_put_byte(crc, in);
+        srslte_crc_checksum_put_byte(crc_tb, in);
+
+        /* Put byte in CB CRC and save latest checksum */
+        srslte_crc_checksum_put_byte(crc_cb, in);
 
         /* Run actual encoder */
         tcod_lut_t l = tcod_lut[state0][in];
@@ -225,10 +235,27 @@ int srslte_tcod_encode_lut(srslte_tcod_t *h, srslte_crc_t *crc, uint8_t *input,
         state0 = l.next_state;
       }
 
-      uint32_t checksum = (uint32_t) srslte_crc_checksum_get(crc);
+      if (last_cb) {
-      for (int i = 0; i < crc->order / 8; i++) {
+        uint32_t checksum = (uint32_t) srslte_crc_checksum_get(crc_tb);
-        int mask_shift = 8 * (crc->order / 8 - i - 1);
+        for (int i = 0; i < crc_tb->order / 8; i++) {
-        int idx = (long_cb - crc->order) / 8 + i;
+          int mask_shift = 8 * (crc_tb->order / 8 - i - 1);
+          int idx = block_size_nocrc + i;
+          uint8_t in = (uint8_t) ((checksum >> mask_shift) & 0xff);
+
+          /* Put byte in CB CRC and save latest checksum */
+          srslte_crc_checksum_put_byte(crc_cb, in);
+
+          input[idx] = in;
+          tcod_lut_t l = tcod_lut[state0][in];
+          parity[idx] = l.output;
+          state0 = l.next_state;
+        }
+      }
+
+      uint32_t checksum = (uint32_t) srslte_crc_checksum_get(crc_cb);
+      for (int i = 0; i < crc_cb->order / 8; i++) {
+        int mask_shift = 8 * (crc_cb->order / 8 - i - 1);
+        int idx = (long_cb - crc_cb->order) / 8 + i;
         uint8_t in = (uint8_t) ((checksum >> mask_shift) & 0xff);
 
         input[idx] = in;
@@ -239,11 +266,31 @@ int srslte_tcod_encode_lut(srslte_tcod_t *h, srslte_crc_t *crc, uint8_t *input,
 
     } else {
       /* No CRC given */
-      for (uint32_t i = 0; i < long_cb / 8; i++) {
+      int block_size_nocrc = (long_cb - ((last_cb) ? crc_tb->order : 0)) / 8;
-        tcod_lut_t l = tcod_lut[state0][input[i]];
+
+      for (uint32_t i = 0; i < block_size_nocrc; i++) {
+        uint8_t in = input[i];
+
+        srslte_crc_checksum_put_byte(crc_tb, in);
+
+        tcod_lut_t l = tcod_lut[state0][in];
         parity[i] = l.output;
         state0 = l.next_state;
       }
+
+      if (last_cb) {
+        uint32_t checksum = (uint32_t) srslte_crc_checksum_get(crc_tb);
+        for (int i = 0; i < crc_tb->order / 8; i++) {
+          int mask_shift = 8 * (crc_tb->order / 8 - i - 1);
+          int idx = block_size_nocrc + i;
+          uint8_t in = (uint8_t) ((checksum >> mask_shift) & 0xff);
+
+          input[idx] = in;
+          tcod_lut_t l = tcod_lut[state0][in];
+          parity[idx] = l.output;
+          state0 = l.next_state;
+        }
+      }
     }
     parity[long_cb / 8] = 0;  // will put tail here later
     

lib/src/phy/fec/turbodecoder.c

@@ -24,151 +24,543 @@
  *
  */
 
-
-#include <stdio.h>
 #include <stdint.h>
-#include <string.h>
 #include <stdlib.h>
+#include <stdio.h>
+#include <strings.h>
+#include <srslte/srslte.h>
 
+#include "srslte/phy/utils/vector.h"
 #include "srslte/phy/fec/turbodecoder.h"
+
+#define debug_enabled 0
+
+/* Generic (no SSE) implementation */
 #include "srslte/phy/fec/turbodecoder_gen.h"
+srslte_tdec_16bit_impl_t gen_impl = {
+    tdec_gen_init,
+    tdec_gen_free,
+    tdec_gen_dec,
+    tdec_gen_extract_input,
+    tdec_gen_decision_byte
+};
 
+/* SSE no-window implementation */
+#include "srslte/phy/fec/turbodecoder_sse.h"
+srslte_tdec_16bit_impl_t sse_impl = {
+    tdec_sse_init,
+    tdec_sse_free,
+    tdec_sse_dec,
+    tdec_sse_extract_input,
+    tdec_sse_decision_byte
+};
 
+/* SSE window implementation */
 #ifdef LV_HAVE_SSE
-#include "srslte/phy/fec/turbodecoder_simd.h"
+#define WINIMP_IS_SSE16
+#include "srslte/phy/fec/turbodecoder_win.h"
+#undef WINIMP_IS_SSE16
+
+srslte_tdec_16bit_impl_t sse16_win_impl = {
+    tdec_winsse16_init,
+    tdec_winsse16_free,
+    tdec_winsse16_dec,
+    tdec_winsse16_extract_input,
+    tdec_winsse16_decision_byte
+};
 #endif
 
-#include "srslte/phy/utils/vector.h"
+/* AVX window implementation */
+#ifdef LV_HAVE_AVX
+#define WINIMP_IS_AVX16
+#include "srslte/phy/fec/turbodecoder_win.h"
+#undef WINIMP_IS_AVX16
+srslte_tdec_16bit_impl_t avx16_win_impl = {
+    tdec_winavx16_init,
+    tdec_winavx16_free,
+    tdec_winavx16_dec,
+    tdec_winavx16_extract_input,
+    tdec_winavx16_decision_byte
+};
+#endif
+
+/* SSE window implementation */
+#ifdef LV_HAVE_SSE
+#define WINIMP_IS_SSE8
+#include "srslte/phy/fec/turbodecoder_win.h"
+#undef WINIMP_IS_SSE8
+
+srslte_tdec_8bit_impl_t sse8_win_impl = {
+    tdec_winsse8_init,
+    tdec_winsse8_free,
+    tdec_winsse8_dec,
+    tdec_winsse8_extract_input,
+    tdec_winsse8_decision_byte
+};
+#endif
+
+/* AVX window implementation */
+#ifdef LV_HAVE_AVX
+#define WINIMP_IS_AVX8
+#include "srslte/phy/fec/turbodecoder_win.h"
+#undef WINIMP_IS_AVX8
+srslte_tdec_8bit_impl_t avx8_win_impl = {
+    tdec_winavx8_init,
+    tdec_winavx8_free,
+    tdec_winavx8_dec,
+    tdec_winavx8_extract_input,
+    tdec_winavx8_decision_byte
+};
+#endif
+
+#define AUTO_16_SSE    0
+#define AUTO_16_SSEWIN 1
+#define AUTO_16_AVXWIN 2
+#define AUTO_8_SSEWIN  0
+#define AUTO_8_AVXWIN  1
+
+
+// Include interfaces for 8 and 16 bit decoder implementations
+#define LLR_IS_8BIT
+#include "srslte/phy/fec/turbodecoder_iter.h"
+#undef LLR_IS_8BIT
+
+#define LLR_IS_16BIT
+#include "srslte/phy/fec/turbodecoder_iter.h"
+#undef LLR_IS_16BIT
 
 
 int srslte_tdec_init(srslte_tdec_t * h, uint32_t max_long_cb) {
-#ifdef LV_HAVE_SSE
+  return srslte_tdec_init_manual(h, max_long_cb, SRSLTE_TDEC_AUTO);
-  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));
+uint32_t interleaver_idx(uint32_t nof_subblocks) {
+  switch (nof_subblocks) {
+    case 32:
+      return 3;
+    case 16:
+      return 2;
+    case 8:
+      return 1;
+    case 1:
+      return 0;
+    default:
+      return 0;
+  }
+}
+
+/* Initializes the turbo decoder object */
+int srslte_tdec_init_manual(srslte_tdec_t * h, uint32_t max_long_cb, srslte_tdec_impl_type_t dec_type)
+{
+  int ret = -1;
+  bzero(h, sizeof(srslte_tdec_t));
+  uint32_t len = max_long_cb + SRSLTE_TCOD_TOTALTAIL;
+
+  h->dec_type = dec_type;
+
+  // Set manual
+  switch(dec_type) {
+    case SRSLTE_TDEC_AUTO:
+      break;
+    case SRSLTE_TDEC_SSE:
+      h->dec16[0] = &sse_impl;
+      h->current_llr_type = SRSLTE_TDEC_16;
+      break;
+    case SRSLTE_TDEC_SSE_WINDOW:
+      h->dec16[0] = &sse16_win_impl;
+      h->current_llr_type = SRSLTE_TDEC_16;
+      break;
+    case SRSLTE_TDEC_GENERIC:
+      h->dec16[0] = &gen_impl;
+      h->current_llr_type = SRSLTE_TDEC_16;
+      break;
+    case SRSLTE_TDEC_SSE8_WINDOW:
+      h->dec8[0] = &sse8_win_impl;
+      h->current_llr_type = SRSLTE_TDEC_8;
+      break;
+#ifdef LV_HAVE_AVX
+    case SRSLTE_TDEC_AVX_WINDOW:
+      h->dec16[0] = &avx16_win_impl;
+      h->current_llr_type = SRSLTE_TDEC_16;
+      break;
+    case SRSLTE_TDEC_AVX8_WINDOW:
+      h->dec8[0] = &avx8_win_impl;
+      h->current_llr_type = SRSLTE_TDEC_8;
+      break;
+#endif
+    default:
+      fprintf(stderr, "Error decoder %d not supported\n", dec_type);
+      goto clean_and_exit;
+  }
+  
+  h->max_long_cb = max_long_cb;
+
+  h->app1 = srslte_vec_malloc(sizeof(int16_t) * len);
+  if (!h->app1) {
+    perror("srslte_vec_malloc");
+    goto clean_and_exit;
+  }
+  h->app2 = srslte_vec_malloc(sizeof(int16_t) * len);
+  if (!h->app2) {
+    perror("srslte_vec_malloc");
+    goto clean_and_exit;
+  }
+  h->ext1 = srslte_vec_malloc(sizeof(int16_t) * len);
+  if (!h->ext1) {
+    perror("srslte_vec_malloc");
+    goto clean_and_exit;
+  }
+  h->ext2 = srslte_vec_malloc(sizeof(int16_t) * len);
+  if (!h->ext2) {
+    perror("srslte_vec_malloc");
+    goto clean_and_exit;
+  }
+  h->syst0 = srslte_vec_malloc(sizeof(int16_t) * len);
+  if (!h->syst0) {
+    perror("srslte_vec_malloc");
+    goto clean_and_exit;
+  }
+  h->parity0 = srslte_vec_malloc(sizeof(int16_t) * len);
+  if (!h->parity0) {
+    perror("srslte_vec_malloc");
+    goto clean_and_exit;
+  }
+  h->parity1 = srslte_vec_malloc(sizeof(int16_t) * len);
+  if (!h->parity1) {
+    perror("srslte_vec_malloc");
+    goto clean_and_exit;
+  }
+  h->input_conv = srslte_vec_malloc(sizeof(int16_t) * (len * 3+32*3));
   if (!h->input_conv) {
-    perror("malloc");
+    perror("srslte_vec_malloc");
-    return -1;
+    goto clean_and_exit;
   }
-  return srslte_tdec_gen_init(&h->tdec_gen, max_long_cb);
+
+  if (dec_type == SRSLTE_TDEC_AUTO) {
+    h->dec16[AUTO_16_SSE] = &sse_impl;
+    h->dec16[AUTO_16_SSEWIN] = &sse16_win_impl;
+    h->dec8[AUTO_8_SSEWIN]  = &sse8_win_impl;
+#ifdef LV_HAVE_AVX
+    h->dec16[AUTO_16_AVXWIN] = &avx16_win_impl;
+    h->dec8[AUTO_8_AVXWIN]  = &avx8_win_impl;
 #endif
+
+    for (int td=0;td<SRSLTE_TDEC_NOF_AUTO_MODES_16;td++) {
+      if (h->dec16[td]) {
+        if ((h->nof_blocks16[td] = h->dec16[td]->tdec_init(&h->dec16_hdlr[td], h->max_long_cb))<0) {
+          goto clean_and_exit;
+        }
+      }
+    }
+    for (int td=0;td<SRSLTE_TDEC_NOF_AUTO_MODES_8;td++) {
+      if (h->dec8[td]) {
+        if ((h->nof_blocks8[td] = h->dec8[td]->tdec_init(&h->dec8_hdlr[td], h->max_long_cb))<0) {
+          goto clean_and_exit;
+        }
+      }
+    }
+
+    // Compute 1 interleaver for each possible nof_subblocks (1, 8, 16 or 32)
+    for (int s=0;s<4;s++) {
+      for (int i=0;i<SRSLTE_NOF_TC_CB_SIZES;i++) {
+        if (srslte_tc_interl_init(&h->interleaver[s][i], srslte_cbsegm_cbsize(i)) < 0) {
+          goto clean_and_exit;
+        }
+        srslte_tc_interl_LTE_gen_interl(&h->interleaver[s][i], srslte_cbsegm_cbsize(i), s?(8<<(s-1)):1);
+      }
+    }
+  } else {
+    uint32_t nof_subblocks;
+    if (dec_type < SRSLTE_TDEC_SSE8_WINDOW) {
+      if ((h->nof_blocks16[0] = h->dec16[0]->tdec_init(&h->dec16_hdlr[0], h->max_long_cb))<0) {
+        goto clean_and_exit;
+      }
+      nof_subblocks = h->nof_blocks16[0];
+    } else {
+      if ((h->nof_blocks8[0] = h->dec8[0]->tdec_init(&h->dec8_hdlr[0], h->max_long_cb))<0) {
+        goto clean_and_exit;
+      }
+      nof_subblocks = h->nof_blocks8[0];
+    }
+    for (int i=0;i<SRSLTE_NOF_TC_CB_SIZES;i++) {
+      if (srslte_tc_interl_init(&h->interleaver[interleaver_idx(nof_subblocks)][i], srslte_cbsegm_cbsize(i)) < 0) {
+        goto clean_and_exit;
+      }
+      srslte_tc_interl_LTE_gen_interl(&h->interleaver[interleaver_idx(nof_subblocks)][i], srslte_cbsegm_cbsize(i), nof_subblocks);
+    }
+  }
+
+  h->current_cbidx = -1;
+  ret = 0;
+
+clean_and_exit:
+  if (ret == -1) {
+    srslte_tdec_free(h);
+  }
+  return ret;
 }
 
-void srslte_tdec_free(srslte_tdec_t * h) {
+void srslte_tdec_free(srslte_tdec_t * h)
-#ifdef LV_HAVE_SSE
+{
-  srslte_tdec_simd_free(&h->tdec_simd);  
+  if (h->app1) {
-#else
+    free(h->app1);
+  }
+  if (h->app2) {
+    free(h->app2);
+  }
+  if (h->ext1) {
+    free(h->ext1);
+  }
+  if (h->ext2) {
+    free(h->ext2);
+  }
+  if (h->syst0) {
+    free(h->syst0);
+  }
+  if (h->parity0) {
+    free(h->parity0);
+  }
+  if (h->parity1) {
+    free(h->parity1);
+  }
   if (h->input_conv) {
     free(h->input_conv);
   }
-  srslte_tdec_gen_free(&h->tdec_gen);
-#endif
 
+  for (int td=0;td<SRSLTE_TDEC_NOF_AUTO_MODES_8;td++) {
+    if (h->dec8[td] && h->dec8_hdlr[td]) {
+      h->dec8[td]->tdec_free(h->dec8_hdlr[td]);
+    }
+  }
+  for (int td=0;td<SRSLTE_TDEC_NOF_AUTO_MODES_16;td++) {
+    if (h->dec16[td] && h->dec16_hdlr[td]) {
+      h->dec16[td]->tdec_free(h->dec16_hdlr[td]);
+    }
+  }
+  for (int s=0;s<4;s++) {
+    for (int i=0;i<SRSLTE_NOF_TC_CB_SIZES;i++) {
+      srslte_tc_interl_free(&h->interleaver[s][i]);
+    }
+  }
+
+  bzero(h, sizeof(srslte_tdec_t));
 }
 
-int srslte_tdec_reset(srslte_tdec_t * h, uint32_t long_cb) {
+void srslte_tdec_force_not_sb(srslte_tdec_t *h) {
-#ifdef LV_HAVE_SSE
+  h->force_not_sb = true;
-  return srslte_tdec_simd_reset(&h->tdec_simd, long_cb);
-#else
-  return srslte_tdec_gen_reset(&h->tdec_gen, long_cb);
-#endif
 }
 
-int srslte_tdec_reset_cb(srslte_tdec_t * h, uint32_t cb_idx) {
+static void tdec_decision_byte(srslte_tdec_t * h, uint8_t *output)
-#ifdef LV_HAVE_SSE
+{
-  return srslte_tdec_simd_reset_cb(&h->tdec_simd, cb_idx);      
+  if (h->current_llr_type == SRSLTE_TDEC_16) {
-#else
+    h->dec16[h->current_dec]->tdec_decision_byte(!(h->n_iter%2)?h->app1:h->ext1, output, h->current_long_cb);
-  return srslte_tdec_gen_reset(&h->tdec_gen, h->tdec_gen.current_cb_len);
+  } else {
-#endif
+    h->dec8[h->current_dec]->tdec_decision_byte(!(h->n_iter%2)?(int8_t*)h->app1:(int8_t*)h->ext1, output, h->current_long_cb);
+  }
 }
 
-int srslte_tdec_get_nof_iterations_cb(srslte_tdec_t * h, uint32_t cb_idx)
+
+/* Returns number of subblocks in automatic mode for this long_cb */
+uint32_t srslte_tdec_autoimp_get_subblocks(uint32_t long_cb)
 {
-#ifdef LV_HAVE_SSE
+#ifdef LV_HAVE_AVX
-  return srslte_tdec_simd_get_nof_iterations_cb(&h->tdec_simd, cb_idx);
+  if (!(long_cb%16) && long_cb > 800) {
-#else
+    return 16;
-  return h->tdec_gen.n_iter;
+  } else
 #endif
+  if (!(long_cb%8) && long_cb > 400) {
+    return 8;
+  } else {
+    return 0;
+  }
 }
 
-void srslte_tdec_iteration_par(srslte_tdec_t * h, int16_t* input[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb) {
+static int tdec_sb_idx(uint32_t long_cb) {
-#ifdef LV_HAVE_SSE
+  uint32_t nof_sb = srslte_tdec_autoimp_get_subblocks(long_cb);
-  srslte_tdec_simd_iteration(&h->tdec_simd, input, long_cb);      
+  switch(nof_sb) {
-#else
+    case 16:
-  srslte_vec_convert_if(input[0], 0.01, h->input_conv, 3*long_cb+12);
+      return AUTO_16_AVXWIN;
-  srslte_tdec_gen_iteration(&h->tdec_gen, h->input_conv, long_cb);
+    case 8:
+      return AUTO_16_SSEWIN;
+    case 0:
+      return AUTO_16_SSE;
+  }
+  fprintf(stderr, "Error in tdec_sb_idx() invalid nof_sb=%d\n", nof_sb);
+  return 0;
+}
+
+uint32_t srslte_tdec_autoimp_get_subblocks_8bit(uint32_t long_cb)
+{
+#ifdef LV_HAVE_AVX
+  if (!(long_cb%32) && long_cb > 2048) {
+    return 32;
+  } else
 #endif
+  if (!(long_cb%16) && long_cb > 800) {
+    return 16;
+  } else if (!(long_cb%8) && long_cb > 400) {
+    return 8;
+  } else {
+    return 0;
+  }
 }
 
-void srslte_tdec_iteration(srslte_tdec_t * h, int16_t* input, uint32_t long_cb) {
+static int tdec_sb_idx_8(uint32_t long_cb) {
-  int16_t *input_par[SRSLTE_TDEC_MAX_NPAR];
+  uint32_t nof_sb = srslte_tdec_autoimp_get_subblocks_8bit(long_cb);
-  input_par[0] = input; 
+  switch(nof_sb) {
-  return srslte_tdec_iteration_par(h, input_par, long_cb);
+    case 32:
+      return AUTO_8_AVXWIN;
+    case 16:
+      return AUTO_8_SSEWIN;
+    case 8:
+      return 10+AUTO_16_SSEWIN;
+    case 0:
+      return 10+AUTO_16_SSE;
+  }
+  fprintf(stderr, "Error in tdec_sb_idx_8() invalid nof_sb=%d\n", nof_sb);
+  return 0;
 }
 
-void srslte_tdec_decision_par(srslte_tdec_t * h, uint8_t *output[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb) {
+// TODO: Implement SSE version. Don't really a problem since this only called at very low rates
-#ifdef LV_HAVE_SSE
+static void convert_8_to_16(int8_t *in, int16_t *out, uint32_t len)
-  return srslte_tdec_simd_decision(&h->tdec_simd, output, long_cb);
+{
-#else
+  for (int i=0;i<len;i++) {
-  return srslte_tdec_gen_decision(&h->tdec_gen, output[0], long_cb);
+    out[i] = (int16_t) in[i];
-#endif
+  }
 }
 
-uint32_t srslte_tdec_get_nof_parallel(srslte_tdec_t *h) {
+static void convert_16_to_8(int16_t *in, int8_t *out, uint32_t len)
-#ifdef LV_HAVE_AVX2
+{
-  return 2;
+  for (int i=0;i<len;i++) {
-#else
+    out[i] = (int8_t) in[i];
-  return 1;
+  }
-#endif
 }
 
-void srslte_tdec_decision(srslte_tdec_t * h, uint8_t *output, uint32_t long_cb) {
+static void tdec_iteration_8(srslte_tdec_t * h, int8_t * input)
-  uint8_t *output_par[SRSLTE_TDEC_MAX_NPAR];
+{
-  output_par[0] = output; 
+  // Select decoder if in auto mode
-  srslte_tdec_decision_par(h, output_par, long_cb);
+  if (h->dec_type == SRSLTE_TDEC_AUTO) {
+    h->current_llr_type = SRSLTE_TDEC_8;
+    h->current_dec = tdec_sb_idx_8(h->current_long_cb);
+    h->current_inter_idx = interleaver_idx(h->nof_blocks8[h->current_dec]);
+
+    // If long_cb is not multiple of any 8-bit decoder, use a 16-bit decoder and do type conversion
+    if (h->current_dec >= 10) {
+      h->current_llr_type = SRSLTE_TDEC_16;
+      h->current_dec -= 10;
+      h->current_inter_idx = interleaver_idx(h->nof_blocks16[h->current_dec]);
+    }
+  } else {
+    h->current_dec = 0;
+  }
+
+  if (h->current_llr_type == SRSLTE_TDEC_16) {
+    if (!h->n_iter) {
+      convert_8_to_16(input, h->input_conv, 3*h->current_long_cb+12);
+    }
+    run_tdec_iteration_16bit(h, h->input_conv);
+  } else {
+    run_tdec_iteration_8bit(h, input);
+  }
 }
 
-void srslte_tdec_decision_byte_par(srslte_tdec_t * h, uint8_t *output[SRSLTE_TDEC_MAX_NPAR], uint32_t long_cb) {
+static void tdec_iteration_16(srslte_tdec_t * h, int16_t * input)
-#ifdef LV_HAVE_SSE
+{
-  srslte_tdec_simd_decision_byte(&h->tdec_simd, output, long_cb);  
+  // Select decoder if in auto mode
-#else
+  if (h->dec_type == SRSLTE_TDEC_AUTO) {
-  srslte_tdec_gen_decision_byte(&h->tdec_gen, output[0], long_cb);
+    h->current_llr_type = SRSLTE_TDEC_16;
-#endif  
+    h->current_dec = tdec_sb_idx(h->current_long_cb);
+  } else {
+    h->current_dec = 0;
+  }
+  h->current_inter_idx = interleaver_idx(h->nof_blocks16[h->current_dec]);
+
+  if (h->current_llr_type == SRSLTE_TDEC_8) {
+
+    h->current_inter_idx = interleaver_idx(h->nof_blocks8[h->current_dec]);
+
+    if (!h->n_iter) {
+      convert_16_to_8(input, h->input_conv, 3*h->current_long_cb+12);
+    }
+    run_tdec_iteration_8bit(h, h->input_conv);
+  } else {
+    run_tdec_iteration_16bit(h, input);
+  }
 }
 
-void srslte_tdec_decision_byte_par_cb(srslte_tdec_t * h, uint8_t *output, uint32_t cb_idx, uint32_t long_cb) {
+/* Resets the decoder and sets the codeblock length */
-#ifdef LV_HAVE_SSE
+int srslte_tdec_new_cb(srslte_tdec_t * h, uint32_t long_cb)
-  srslte_tdec_simd_decision_byte_cb(&h->tdec_simd, output, cb_idx, long_cb);
+{
-#else
+  if (long_cb > h->max_long_cb) {
-  srslte_tdec_gen_decision_byte(&h->tdec_gen, output, long_cb);
+    fprintf(stderr, "TDEC was initialized for max_long_cb=%d\n",
-#endif  
+            h->max_long_cb);
+    return -1;
+  }
+
+  h->n_iter = 0;
+  h->current_long_cb = long_cb;
+  h->current_cbidx   = srslte_cbsegm_cbindex(long_cb);
+  if (h->current_cbidx < 0) {
+    fprintf(stderr, "Invalid CB length %d\n", long_cb);
+    return -1;
+  }
+  return 0;
 }
 
-void srslte_tdec_decision_byte(srslte_tdec_t * h, uint8_t *output, uint32_t long_cb) {
+void srslte_tdec_iteration(srslte_tdec_t * h, int16_t * input, uint8_t *output)
-  uint8_t *output_par[SRSLTE_TDEC_MAX_NPAR];
+{
-  output_par[0] = output; 
+  if (h->current_cbidx >= 0) {
-  srslte_tdec_decision_byte_par(h, output_par, long_cb);
+    tdec_iteration_16(h, input);
+    tdec_decision_byte(h, output);
+  }
 }
 
-int srslte_tdec_run_all_par(srslte_tdec_t * h, int16_t * input[SRSLTE_TDEC_MAX_NPAR],
+/* Runs nof_iterations iterations and decides the output bits */
-                            uint8_t *output[SRSLTE_TDEC_MAX_NPAR],
+int srslte_tdec_run_all(srslte_tdec_t * h, int16_t * input, uint8_t *output,
-                            uint32_t nof_iterations, uint32_t long_cb) {
+                        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);  
+  if (srslte_tdec_new_cb(h, long_cb)) {
-#else
+    return SRSLTE_ERROR;
-  srslte_vec_convert_if(input[0], 0.01, h->input_conv, 3*long_cb+12);
+  }
-  return srslte_tdec_gen_run_all(&h->tdec_gen, h->input_conv, output[0], nof_iterations, long_cb);
+
-#endif
+  do {
+    tdec_iteration_16(h, input);
+  } while (h->n_iter < nof_iterations);
+
+  tdec_decision_byte(h, output);
+
+  return SRSLTE_SUCCESS;
 }
 
-int srslte_tdec_run_all(srslte_tdec_t * h, int16_t * input, uint8_t *output, uint32_t nof_iterations, uint32_t long_cb)
+
+void srslte_tdec_iteration_8bit(srslte_tdec_t * h, int8_t * input, uint8_t *output)
 {
-  uint8_t *output_par[SRSLTE_TDEC_MAX_NPAR];
+  if (h->current_cbidx >= 0) {
-  output_par[0] = output;   
+    tdec_iteration_8(h, input);
-  int16_t *input_par[SRSLTE_TDEC_MAX_NPAR];
+    tdec_decision_byte(h, output);
-  input_par[0] = input; 
+  }
+}
 
-  return srslte_tdec_run_all_par(h, input_par, output_par, nof_iterations, long_cb);
+/* Runs nof_iterations iterations and decides the output bits */
+int srslte_tdec_run_all_8bit(srslte_tdec_t * h, int8_t * input, uint8_t *output,
+                             uint32_t nof_iterations, uint32_t long_cb)
+{
+  if (srslte_tdec_new_cb(h, long_cb)) {
+    return SRSLTE_ERROR;
+  }
+
+  do {
+    tdec_iteration_8(h, input);
+  } while (h->n_iter < nof_iterations);
+
+  tdec_decision_byte(h, output);
+
+  return SRSLTE_SUCCESS;
+}
+
+int srslte_tdec_get_nof_iterations(srslte_tdec_t * h)
+{
+  return h->n_iter;
 }
 
 

lib/src/phy/fec/turbodecoder_avx.c

@@ -1,475 +0,0 @@
-/**
- *
- * \section COPYRIGHT
- *
- * Copyright 2013-2015 Software Radio Systems Limited
- *
- * \section LICENSE
- *
- * This file is part of the srsLTE library.
- *
- * srsLTE is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as
- * published by the Free Software Foundation, either version 3 of
- * the License, or (at your option) any later version.
- *
- * srsLTE is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * A copy of the GNU Affero General Public License can be found in
- * the LICENSE file in the top-level directory of this distribution
- * and at http://www.gnu.org/licenses/.
- *
- */
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <stdint.h>
-#include <string.h>
-#include <strings.h>
-#include <math.h>
-
-#include "srslte/phy/fec/turbodecoder_simd.h"
-#include "srslte/phy/utils/vector.h"
-
-#include <inttypes.h>
-
-#define NUMSTATES       8
-#define NINPUTS         2
-#define TAIL            3
-#define TOTALTAIL       12
-
-#define INF 10000
-#define ZERO 0
-
-
-#ifdef LV_HAVE_AVX2
-
-#include <smmintrin.h>
-#include <immintrin.h>
-
-
-// Number of CB processed in parllel in AVX
-#define NCB 2 
-
-/*
-static void print_256i(__m256i x) {
-  int16_t *s = (int16_t*) &x; 
-  printf("[%d", s[0]);
-  for (int i=1;i<16;i++) {
-    printf(",%d", s[i]);
-  }
-  printf("]\n");
-}
-*/
-
-/* Computes the horizontal MAX from 8 16-bit integers using the minpos_epu16 SSE4.1 instruction */
-static inline int16_t hMax0(__m256i masked_value)
-{
-  __m128i tmp1 = _mm256_extractf128_si256(masked_value, 0); 
-  __m128i tmp3 = _mm_minpos_epu16(tmp1);
-  return (int16_t)(_mm_cvtsi128_si32(tmp3));
-}
-
-static inline int16_t hMax1(__m256i masked_value)
-{
-  __m128i tmp1 = _mm256_extractf128_si256(masked_value, 1); 
-  __m128i tmp3 = _mm_minpos_epu16(tmp1);
-  return (int16_t)(_mm_cvtsi128_si32(tmp3));
-}
-
-/* Computes beta values */
-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;
-  const __m256i *alphaPtr = (const __m256i*) s->alpha;
- 
-  __m256i beta_k = _mm256_set_epi16(-INF, -INF, -INF, -INF, -INF, -INF, -INF, 0, -INF, -INF, -INF, -INF, -INF, -INF, -INF, 0);
-  __m256i g, bp, bn, alpha_k; 
-  
-  /* Define the shuffle constant for the positive beta */
-  __m256i shuf_bp = _mm256_set_epi8(
-    // 1st CB
-    15+16, 14+16, // 7
-    7+16,  6+16,  // 3
-    5+16,  4+16,  // 2
-    13+16, 12+16, // 6
-    11+16, 10+16, // 5
-    3+16,  2+16,  // 1
-    1+16,  0+16,  // 0
-    9+16,  8+16,  // 4
-
-    // 2nd CB
-    15, 14, // 7
-    7,  6,  // 3
-    5,  4,  // 2
-    13, 12, // 6
-    11, 10, // 5
-    3,  2,  // 1
-    1,  0,  // 0
-    9,  8   // 4
-  );
-
-  /* Define the shuffle constant for the negative beta */
-  __m256i shuf_bn = _mm256_set_epi8(
-    7+16, 6+16,   // 3
-    15+16, 14+16, // 7
-    13+16, 12+16, // 6
-    5+16,  4+16,  // 2
-    3+16,  2+16,  // 1
-    11+16, 10+16, // 5
-    9+16,  8+16,  // 4
-    1+16,  0+16,  // 0
-    
-    7,   6, // 3
-    15, 14, // 7
-    13, 12, // 6
-    5,  4,  // 2
-    3,  2,  // 1
-    11, 10, // 5
-    9,  8,  // 4
-    1,  0   // 0
-  );
- 
-  alphaPtr += long_cb-1;
-
-  /* Define shuffle for branch costs */
-  __m256i shuf_g[4];
-  shuf_g[3] = _mm256_set_epi8(3+16,2+16,1+16,0+16,1+16,0+16,3+16,2+16,3+16,2+16,1+16,0+16,1+16,0+16,3+16,2+16, 
-                              3,2,1,0,1,0,3,2,3,2,1,0,1,0,3,2);
-  shuf_g[2] = _mm256_set_epi8(7+16,6+16,5+16,4+16,5+16,4+16,7+16,6+16,7+16,6+16,5+16,4+16,5+16,4+16,7+16,6+16,
-                              7,6,5,4,5,4,7,6,7,6,5,4,5,4,7,6);
-  shuf_g[1] = _mm256_set_epi8(11+16,10+16,9+16,8+16,9+16,8+16,11+16,10+16,11+16,10+16,9+16,8+16,9+16,8+16,11+16,10+16,
-                              11,10,9,8,9,8,11,10,11,10,9,8,9,8,11,10);
-  shuf_g[0] = _mm256_set_epi8(15+16,14+16,13+16,12+16,13+16,12+16,15+16,14+16,15+16,14+16,13+16,12+16,13+16,12+16,15+16,14+16,
-                              15,14,13,12,13,12,15,14,15,14,13,12,13,12,15,14);
-
-  /* Define shuffle for beta normalization */
-  __m256i shuf_norm = _mm256_set_epi8(17,16,17,16,17,16,17,16,17,16,17,16,17,16,17,16,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0);
-
-  __m256i gv;
-  int16_t *b = &s->branch[2*NCB*long_cb-16];
-  __m256i *gPtr = (__m256i*) b;
-  
-  /* This defines a beta computation step: 
-   * Adds and substracts the branch metrics to the previous beta step, 
-   * shuffles the states according to the trellis path and selects maximum state 
-   */
-#define BETA_STEP(g)     bp = _mm256_add_epi16(beta_k, g);\
-    bn = _mm256_sub_epi16(beta_k, g);\
-    bp = _mm256_shuffle_epi8(bp, shuf_bp);\
-    bn = _mm256_shuffle_epi8(bn, shuf_bn);\
-    beta_k = _mm256_max_epi16(bp, bn);    
-
-    /* Loads the alpha metrics from memory and adds them to the temporal bn and bp 
-     * metrics. Then computes horizontal maximum of both metrics and computes difference
-     */
-#define BETA_STEP_CNT(c,d) g = _mm256_shuffle_epi8(gv, shuf_g[c]);\
-    BETA_STEP(g)\
-    alpha_k = _mm256_load_si256(alphaPtr);\
-    alphaPtr--;\
-    bp = _mm256_add_epi16(bp, alpha_k);\
-    bn = _mm256_add_epi16(bn, alpha_k);\
-    bn = _mm256_sub_epi16(_mm256_set1_epi16(0x7FFF), bn);\
-    bp = _mm256_sub_epi16(_mm256_set1_epi16(0x7FFF), bp);\
-    output[0][k-d] = hMax0(bn) - hMax0(bp);\
-    output[1][k-d] = hMax1(bn) - hMax1(bp);
-
-  /* The tail does not require to load alpha or produce outputs. Only update 
-   * beta metrics accordingly */
-  for (k=end-1; k>=long_cb; k--) {
-    int16_t g0_1 = s->branch[2*NCB*k];
-    int16_t g1_1 = s->branch[2*NCB*k+1];
-    int16_t g0_2 = s->branch[2*NCB*k+6];
-    int16_t g1_2 = s->branch[2*NCB*k+6+1];
-    g = _mm256_set_epi16(g1_2, g0_2, g0_2, g1_2, g1_2, g0_2, g0_2, g1_2, g1_1, g0_1, g0_1, g1_1, g1_1, g0_1, g0_1, g1_1);
-    BETA_STEP(g);
-  }  
-  
-  /* We inline 2 trelis steps for each normalization */
-  __m256i norm;
-  for (; k >= 0; k-=8) {    
-    gv = _mm256_load_si256(gPtr);
-    gPtr--;
-    BETA_STEP_CNT(0,0);
-    BETA_STEP_CNT(1,1);
-    BETA_STEP_CNT(2,2);
-    BETA_STEP_CNT(3,3);
-    norm = _mm256_shuffle_epi8(beta_k, shuf_norm); 
-    beta_k = _mm256_sub_epi16(beta_k, norm);
-    gv = _mm256_load_si256(gPtr);
-    gPtr--;
-    BETA_STEP_CNT(0,4);
-    BETA_STEP_CNT(1,5);
-    BETA_STEP_CNT(2,6);
-    BETA_STEP_CNT(3,7);
-    norm = _mm256_shuffle_epi8(beta_k, shuf_norm); 
-    beta_k = _mm256_sub_epi16(beta_k, norm);
-  }  
-}
-
-/* Computes alpha metrics */
-void map_avx_alpha(map_gen_t * s, uint32_t long_cb)
-{
-  uint32_t k;
-  int16_t *alpha1 = s->alpha;
-  int16_t *alpha2 = &s->alpha[8];
-  uint32_t i;
-
-  alpha1[0] = 0; 
-  alpha2[0] = 0; 
-  for (i = 1; i < 8; i++) {
-    alpha1[i] = -INF;
-    alpha2[i] = -INF;
-  }
-    
-  /* Define the shuffle constant for the positive alpha */
-  __m256i shuf_ap = _mm256_set_epi8(
-
-    // 1st CB 
-    31, 30, // 7
-    25, 24, // 4
-    23, 22, // 3
-    17, 16, // 0
-    29, 28, // 6
-    27, 26, // 5
-    21, 20, // 2
-    19, 18, // 1
-
-    // 2nd CB
-    15, 14, // 7
-    9,  8,  // 4
-    7,  6,  // 3
-    1,  0,  // 0
-    13, 12, // 6
-    11, 10, // 5
-    5,  4,  // 2
-    3,  2   // 1
-  );
-
-  /* Define the shuffle constant for the negative alpha */
-  __m256i shuf_an = _mm256_set_epi8(
-
-    // 1nd CB 
-    29, 28, // 6
-    27, 26, // 5
-    21, 20, // 2
-    19, 18, // 1
-    31, 30, // 7
-    25, 24, // 4
-    23, 22, // 3
-    17, 16, // 0
-
-    // 2nd CB 
-    13, 12, // 6
-    11, 10, // 5
-    5,  4,  // 2
-    3,  2,  // 1
-    15, 14, // 7
-    9,  8,  // 4
-    7,  6,  // 3
-    1,  0   // 0
-  );
-  
-  /* Define shuffle for branch costs */
-  __m256i shuf_g[4];
-  shuf_g[0] = _mm256_set_epi8(3+16,2+16,3+16,2+16,1+16,0+16,1+16,0+16,1+16,0+16,1+16,0+16,3+16,2+16,3+16,2+16,       
-                              3,2,3,2,1,0,1,0,1,0,1,0,3,2,3,2);
-  shuf_g[1] = _mm256_set_epi8(7+16,6+16,7+16,6+16,5+16,4+16,5+16,4+16,5+16,4+16,5+16,4+16,7+16,6+16,7+16,6+16,    
-                              7,6,7,6,5,4,5,4,5,4,5,4,7,6,7,6);
-  shuf_g[2] = _mm256_set_epi8(11+16,10+16,11+16,10+16,9+16,8+16,9+16,8+16,9+16,8+16,9+16,8+16,11+16,10+16,11+16,10+16,       
-                              11,10,11,10,9,8,9,8,9,8,9,8,11,10,11,10);
-  shuf_g[3] = _mm256_set_epi8(15+16,14+16,15+16,14+16,13+16,12+16,13+16,12+16,13+16,12+16,13+16,12+16,15+16,14+16,15+16,14+16,
-                              15,14,15,14,13,12,13,12,13,12,13,12,15,14,15,14);
-
-  __m256i shuf_norm = _mm256_set_epi8(17,16,17,16,17,16,17,16,17,16,17,16,17,16,17,16,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1,0);
-  
-  __m256i* alphaPtr = (__m256i*) s->alpha;
-  alphaPtr++;
-
-  __m256i gv; 
-  __m256i *gPtr = (__m256i*) s->branch;
-  __m256i g, ap, an; 
-    
-  __m256i alpha_k = _mm256_set_epi16(-INF, -INF, -INF, -INF, -INF, -INF, -INF, 0, -INF, -INF, -INF, -INF, -INF, -INF, -INF, 0);
-  
-  /* This defines a alpha computation step: 
-   * Adds and substracts the branch metrics to the previous alpha step, 
-   * shuffles the states according to the trellis path and selects maximum state 
-   */
-#define ALPHA_STEP(c)  g = _mm256_shuffle_epi8(gv, shuf_g[c]); \
-  ap = _mm256_add_epi16(alpha_k, g);\
-  an = _mm256_sub_epi16(alpha_k, g);\
-  ap = _mm256_shuffle_epi8(ap, shuf_ap);\
-  an = _mm256_shuffle_epi8(an, shuf_an);\
-  alpha_k = _mm256_max_epi16(ap, an);\
-  _mm256_store_si256(alphaPtr, alpha_k);\
-  alphaPtr++;\
-
-
-  /* In this loop, we compute 8 steps and normalize twice for each branch metrics memory load */
-  __m256i norm;
-  for (k = 0; k < long_cb/8; k++) {
-    gv = _mm256_load_si256(gPtr);
-    
-    gPtr++;
-    ALPHA_STEP(0);
-    ALPHA_STEP(1);
-    ALPHA_STEP(2);
-    ALPHA_STEP(3);
-    norm = _mm256_shuffle_epi8(alpha_k, shuf_norm); 
-    alpha_k = _mm256_sub_epi16(alpha_k, norm);
-    gv = _mm256_load_si256(gPtr);
-    gPtr++;
-    ALPHA_STEP(0);
-    ALPHA_STEP(1);
-    ALPHA_STEP(2);
-    ALPHA_STEP(3);
-    norm = _mm256_shuffle_epi8(alpha_k, shuf_norm); 
-    alpha_k = _mm256_sub_epi16(alpha_k, norm);
-  }  
-}
-
-void map_sse_gamma_single(int16_t *output, int16_t *input, int16_t *app, int16_t *parity) 
-{
-  __m128i res00, res10, res01, res11, res0, res1; 
-  __m128i in, ap, pa, g1, g0;
-
-  __m128i *inPtr  = (__m128i*) input;
-  __m128i *appPtr = (__m128i*) app;
-  __m128i *paPtr  = (__m128i*) parity;
-  __m128i *resPtr = (__m128i*) output;
-  
-  __m128i res00_mask = _mm_set_epi8(0xff,0xff,7,6,0xff,0xff,5,4,0xff,0xff,3,2,0xff,0xff,1,0);
-  __m128i res10_mask = _mm_set_epi8(0xff,0xff,15,14,0xff,0xff,13,12,0xff,0xff,11,10,0xff,0xff,9,8);
-  __m128i res01_mask = _mm_set_epi8(7,6,0xff,0xff,5,4,0xff,0xff,3,2,0xff,0xff,1,0,0xff,0xff);
-  __m128i res11_mask = _mm_set_epi8(15,14,0xff,0xff,13,12,0xff,0xff,11,10,0xff,0xff,9,8,0xff,0xff);
-  
-  in = _mm_load_si128(inPtr);
-  inPtr++;
-  pa = _mm_load_si128(paPtr);
-  paPtr++;
-  
-  if (appPtr) {
-    ap = _mm_load_si128(appPtr);
-    appPtr++;
-    in = _mm_add_epi16(ap, in);
-  }
-  
-  g1 = _mm_add_epi16(in, pa);
-  g0 = _mm_sub_epi16(in, pa);
-
-  g1 = _mm_srai_epi16(g1, 1);
-  g0 = _mm_srai_epi16(g0, 1);
-  
-  res00 = _mm_shuffle_epi8(g0, res00_mask);
-  res10 = _mm_shuffle_epi8(g0, res10_mask);
-  res01 = _mm_shuffle_epi8(g1, res01_mask);
-  res11 = _mm_shuffle_epi8(g1, res11_mask);
-
-  res0  = _mm_or_si128(res00, res01);
-  res1  = _mm_or_si128(res10, res11);
-
-  _mm_store_si128(resPtr, res0);
-  resPtr++;
-  _mm_store_si128(resPtr, res1);    
-  resPtr++;
-}
-
-
-/* Compute branch metrics (gamma) */
-void map_avx_gamma(map_gen_t * h, int16_t *input, int16_t *app, int16_t *parity, uint32_t cbidx, uint32_t long_cb) 
-{
-  __m128i res10, res20, res11, res21, res1, res2; 
-  __m256i in, ap, pa, g1, g0;
-
-  __m256i *inPtr  = (__m256i*) input;
-  __m256i *appPtr = (__m256i*) app;
-  __m256i *paPtr  = (__m256i*) parity;
-  __m128i *resPtr = (__m128i*) h->branch;
-  
-  if (cbidx) {
-    resPtr++;
-  }
-  
-  __m128i res10_mask = _mm_set_epi8(0xff,0xff,7,6,0xff,0xff,5,4,0xff,0xff,3,2,0xff,0xff,1,0);
-  __m128i res11_mask = _mm_set_epi8(7,6,0xff,0xff,5,4,0xff,0xff,3,2,0xff,0xff,1,0,0xff,0xff);
-
-  __m128i res20_mask = _mm_set_epi8(0xff,0xff,15,14,0xff,0xff,13,12,0xff,0xff,11,10,0xff,0xff,9,8);
-  __m128i res21_mask = _mm_set_epi8(15,14,0xff,0xff,13,12,0xff,0xff,11,10,0xff,0xff,9,8,0xff,0xff);
-
-  for (int i=0;i<long_cb/16;i++) {
-    in = _mm256_load_si256(inPtr);
-    inPtr++;
-    pa = _mm256_load_si256(paPtr);
-    paPtr++;
-    
-    if (appPtr) {
-      ap = _mm256_load_si256(appPtr);
-      appPtr++;
-      in = _mm256_add_epi16(ap, in);
-    }
-    
-    g0 = _mm256_sub_epi16(in, pa);
-    g1 = _mm256_add_epi16(in, pa);
- 
-    g0 = _mm256_srai_epi16(g0, 1);
-    g1 = _mm256_srai_epi16(g1, 1);
-    
-    __m128i g0_t = _mm256_extractf128_si256(g0, 0);
-    __m128i g1_t = _mm256_extractf128_si256(g1, 0);
-    
-    res10 = _mm_shuffle_epi8(g0_t, res10_mask);
-    res11 = _mm_shuffle_epi8(g1_t, res11_mask);
-
-    res20 = _mm_shuffle_epi8(g0_t, res20_mask);
-    res21 = _mm_shuffle_epi8(g1_t, res21_mask);
-
-    res1  = _mm_or_si128(res10, res11);
-    res2  = _mm_or_si128(res20, res21);
-
-    _mm_store_si128(resPtr, res1);
-    resPtr++;
-    resPtr++;
-    _mm_store_si128(resPtr, res2);    
-    resPtr++;
-    resPtr++;          
-    
-    g0_t = _mm256_extractf128_si256(g0, 1);
-    g1_t = _mm256_extractf128_si256(g1, 1);
-    
-    res10 = _mm_shuffle_epi8(g0_t, res10_mask);
-    res11 = _mm_shuffle_epi8(g1_t, res11_mask);
-
-    res20 = _mm_shuffle_epi8(g0_t, res20_mask);
-    res21 = _mm_shuffle_epi8(g1_t, res21_mask);
-
-    res1  = _mm_or_si128(res10, res11);
-    res2  = _mm_or_si128(res20, res21);
-
-    _mm_store_si128(resPtr, res1);
-    resPtr++;
-    resPtr++;
-    _mm_store_si128(resPtr, res2);    
-    resPtr++;
-    resPtr++;          
-
-  }
-  
-  if (long_cb%16) {
-    map_sse_gamma_single((int16_t*) resPtr, (int16_t*) inPtr, (int16_t*) appPtr, (int16_t*) paPtr);
-  }
-    
-  for (int i=long_cb;i<long_cb+3;i++) {
-    h->branch[2*i*NCB+cbidx*6]   = (input[i] - parity[i])/2;
-    h->branch[2*i*NCB+cbidx*6+1] = (input[i] + parity[i])/2;
-  }
-}
-
-
-#endif
-
-

lib/src/phy/fec/turbodecoder_gen.c

@@ -39,8 +39,16 @@
 #define TAIL            3
 #define TOTALTAIL       12
 
-#define INF 9e4
+#define INF 10000
-#define ZERO 9e-4
+
+#define debug_enabled 0
+
+#if debug_enabled
+#define debug_state       printf("k=%5d, in=%5d, pa=%3d, out=%5d, alpha=", k, x, parity[k-1], out); srslte_vec_fprint_s(stdout, alpha, 8); \
+                          printf(", beta="); srslte_vec_fprint_s(stdout, &beta[8*(k)], 8); printf("\n");
+#else
+#define debug_state
+#endif
 
 /************************************************
  *
@@ -48,14 +56,13 @@
  *  Decoder
  *
  ************************************************/
-static void map_gen_beta(srslte_map_gen_vl_t * s, float * input, float * parity,
+static void map_gen_beta(tdec_gen_t *s, int16_t *input, int16_t *app, int16_t *parity, uint32_t long_cb)
-                  uint32_t long_cb)
 {
-  float m_b[8], new[8], old[8];
+  int16_t m_b[8], new[8], old[8];
-  float x, y, xy;
+  int16_t x, y, xy;
   int k;
   uint32_t end = long_cb + SRSLTE_TCOD_RATE;
-  float *beta = s->beta;
+  int16_t *beta = s->beta;
   uint32_t i;
 
   for (i = 0; i < 8; i++) {
@@ -64,6 +71,9 @@ static void map_gen_beta(srslte_map_gen_vl_t * s, float * input, float * parity,
 
   for (k = end - 1; k >= 0; k--) {
     x = input[k];
+    if (app && k<long_cb) {
+      x += app[k];
+    }
     y = parity[k];
 
     xy = x + y;
@@ -92,19 +102,25 @@ static void map_gen_beta(srslte_map_gen_vl_t * s, float * input, float * parity,
       old[i] = new[i];
       beta[8 * k + i] = old[i];
     }
+
+    if ((k%4)==0 && k < long_cb) {
+      for (i = 1; i < 8; i++) {
+        old[i] -= old[0];
+      }
+      old[0] = 0;
+    }
   }
 }
 
-static void map_gen_alpha(srslte_map_gen_vl_t * s, float * input, float * parity, float * output,
+static void map_gen_alpha(tdec_gen_t *s, int16_t *input, int16_t *app, int16_t *parity, int16_t *output, uint32_t long_cb)
-                   uint32_t long_cb)
 {
-  float m_b[8], new[8], old[8], max1[8], max0[8];
+  int16_t m_b[8], new[8], old[8], max1[8], max0[8];
-  float m1, m0;
+  int16_t m1, m0;
-  float x, y, xy;
+  int16_t x, y, xy;
-  float out;
+  int16_t out;
   uint32_t k;
   uint32_t end = long_cb;
-  float *beta = s->beta;
+  int16_t *beta = s->beta;
   uint32_t i;
 
   old[0] = 0;
@@ -112,12 +128,23 @@ static void map_gen_alpha(srslte_map_gen_vl_t * s, float * input, float * parity
     old[i] = -INF;
   }
 
+#if debug_enabled
+  int16_t alpha[8];
+#endif
+
   for (k = 1; k < end + 1; k++) {
     x = input[k - 1];
+    if (app) {
+      x += app[k - 1];
+    }
     y = parity[k - 1];
 
     xy = x + y;
 
+#if debug_enabled
+    memcpy(alpha, old, sizeof(int16_t)*8);
+#endif
+
     m_b[0] = old[0];
     m_b[1] = old[3] + y;
     m_b[2] = old[4] + y;
@@ -150,251 +177,97 @@ static void map_gen_alpha(srslte_map_gen_vl_t * s, float * input, float * parity
       if (max0[i] > m0)
         m0 = max0[i];
     }
-
     for (i = 0; i < 8; i++) {
       if (m_b[i] > new[i])
         new[i] = m_b[i];
-      old[i] = new[i];
+      old[i] = new[i]; }
+
+    if ((k%4)==0) {
+      for (i = 1; i < 8; i++) {
+        old[i] -= old[0];
+      }
+      old[0] = 0;
     }
 
     out = m1 - m0;
     output[k - 1] = out;
+
+    debug_state;
+
   }
 }
 
-static int map_gen_init(srslte_map_gen_vl_t * h, int max_long_cb)
+int tdec_gen_init(void **hh, uint32_t max_long_cb)
 {
-  bzero(h, sizeof(srslte_map_gen_vl_t));
+  *hh = calloc(1, sizeof(tdec_gen_t));
-  h->beta = srslte_vec_malloc(sizeof(float) * (max_long_cb + SRSLTE_TCOD_TOTALTAIL + 1) * NUMSTATES);
+
+  tdec_gen_t *h = (tdec_gen_t*) *hh;
+
+  h->beta = srslte_vec_malloc(sizeof(int16_t) * (max_long_cb + SRSLTE_TCOD_TOTALTAIL + 1) * NUMSTATES);
   if (!h->beta) {
     perror("srslte_vec_malloc");
     return -1;
   }
   h->max_long_cb = max_long_cb;
-  return 0;
+  return 1;
 }
 
-static void map_gen_free(srslte_map_gen_vl_t * h)
+void tdec_gen_free(void *hh)
 {
+  tdec_gen_t *h = (tdec_gen_t*) hh;
   if (h->beta) {
     free(h->beta);
   }
-  bzero(h, sizeof(srslte_map_gen_vl_t));
+  bzero(h, sizeof(tdec_gen_t));
 }
 
-static void map_gen_dec(srslte_map_gen_vl_t * h, float * input, float * parity, float * output,
+void tdec_gen_dec(void *hh, int16_t *input, int16_t *app, int16_t *parity, int16_t *output, uint32_t long_cb)
-                 uint32_t long_cb)
 {
-  uint32_t k;
+  tdec_gen_t *h = (tdec_gen_t*) hh;
 
   h->beta[(long_cb + TAIL) * NUMSTATES] = 0;
-  for (k = 1; k < NUMSTATES; k++)
+  for (uint32_t k = 1; k < NUMSTATES; k++)
     h->beta[(long_cb + TAIL) * NUMSTATES + k] = -INF;
 
-  map_gen_beta(h, input, parity, long_cb);
+  map_gen_beta(h, input, app, parity, long_cb);
-  map_gen_alpha(h, input, parity, output, long_cb);
+  map_gen_alpha(h, input, app, parity, output, long_cb);
 }
 
-/************************************************
+void tdec_gen_extract_input(int16_t *input, int16_t *syst, int16_t *app2, int16_t *parity0, int16_t *parity1, uint32_t long_cb)
- *
- *  TURBO DECODER INTERFACE
- *
- ************************************************/
-int srslte_tdec_gen_init(srslte_tdec_gen_t * h, uint32_t max_long_cb)
 {
-  int ret = -1;
+  // Prepare systematic and parity bits for MAP DEC #1
-  bzero(h, sizeof(srslte_tdec_gen_t));
+  for (uint32_t i = 0; i < long_cb; i++) {
-  uint32_t len = max_long_cb + SRSLTE_TCOD_TOTALTAIL;
+    syst[i]    = input[SRSLTE_TCOD_RATE * i];
-
+    parity0[i] = input[SRSLTE_TCOD_RATE * i + 1];
-  h->max_long_cb = max_long_cb;
+    parity1[i] = input[SRSLTE_TCOD_RATE * i + 2];
-
-  h->llr1 = srslte_vec_malloc(sizeof(float) * len);
-  if (!h->llr1) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->llr2 = srslte_vec_malloc(sizeof(float) * len);
-  if (!h->llr2) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->w = srslte_vec_malloc(sizeof(float) * len);
-  if (!h->w) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->syst = srslte_vec_malloc(sizeof(float) * len);
-  if (!h->syst) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->parity = srslte_vec_malloc(sizeof(float) * len);
-  if (!h->parity) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
   }
+  for (uint32_t i = long_cb; i < long_cb + SRSLTE_TCOD_RATE; i++) {
+    syst[i]    = input[SRSLTE_TCOD_RATE * long_cb + NINPUTS * (i - long_cb)];
+    parity0[i] = input[SRSLTE_TCOD_RATE * long_cb + NINPUTS * (i - long_cb) + 1];
 
-  if (map_gen_init(&h->dec, h->max_long_cb)) {
+    app2[i]    = input[SRSLTE_TCOD_RATE * long_cb + NINPUTS * SRSLTE_TCOD_RATE + NINPUTS * (i - long_cb)];
-    goto clean_and_exit;
+    parity1[i] = input[SRSLTE_TCOD_RATE * long_cb + NINPUTS * SRSLTE_TCOD_RATE
+                         + NINPUTS * (i - long_cb) + 1];
   }
 
-  for (int i=0;i<SRSLTE_NOF_TC_CB_SIZES;i++) {
-    if (srslte_tc_interl_init(&h->interleaver[i], srslte_cbsegm_cbsize(i)) < 0) {
-      goto clean_and_exit;
-    }
-    srslte_tc_interl_LTE_gen(&h->interleaver[i], srslte_cbsegm_cbsize(i));
-  }
-  h->current_cbidx = -1; 
-  ret = 0;
-clean_and_exit:if (ret == -1) {
-    srslte_tdec_gen_free(h);
-  }
-  return ret;
 }
 
-void srslte_tdec_gen_free(srslte_tdec_gen_t * h)
+void tdec_gen_decision_byte(int16_t *app1, uint8_t *output, uint32_t long_cb)
 {
-  if (h->llr1) {
-    free(h->llr1);
-  }
-  if (h->llr2) {
-    free(h->llr2);
-  }
-  if (h->w) {
-    free(h->w);
-  }
-  if (h->syst) {
-    free(h->syst);
-  }
-  if (h->parity) {
-    free(h->parity);
-  }
-
-  map_gen_free(&h->dec);
-
-  for (int i=0;i<SRSLTE_NOF_TC_CB_SIZES;i++) {
-    srslte_tc_interl_free(&h->interleaver[i]);    
-  }
-
-  bzero(h, sizeof(srslte_tdec_gen_t));
-}
-
-void srslte_tdec_gen_iteration(srslte_tdec_gen_t * h, float * input, uint32_t long_cb)
-{
-  uint32_t i;
-
-  if (h->current_cbidx >= 0) {
-
-    uint16_t *inter = h->interleaver[h->current_cbidx].forward;
-    uint16_t *deinter = h->interleaver[h->current_cbidx].reverse;
-    
-    // Prepare systematic and parity bits for MAP DEC #1
-    for (i = 0; i < long_cb; i++) {
-      h->syst[i] = input[SRSLTE_TCOD_RATE * i] + h->w[i];
-      h->parity[i] = input[SRSLTE_TCOD_RATE * i + 1];
-    }
-    for (i = long_cb; i < long_cb + SRSLTE_TCOD_RATE; i++) {
-      h->syst[i] = input[SRSLTE_TCOD_RATE * long_cb + NINPUTS * (i - long_cb)];
-      h->parity[i] = input[SRSLTE_TCOD_RATE * long_cb + NINPUTS * (i - long_cb) + 1];
-    }
-
-    // Run MAP DEC #1
-    map_gen_dec(&h->dec, h->syst, h->parity, h->llr1, long_cb);
-
-    // Prepare systematic and parity bits for MAP DEC #1
-    for (i = 0; i < long_cb; i++) {
-      h->syst[i] = h->llr1[inter[i]]
-        - h->w[inter[i]];
-      h->parity[i] = input[SRSLTE_TCOD_RATE * i + 2];
-    }
-    for (i = long_cb; i < long_cb + SRSLTE_TCOD_RATE; i++) {
-      h->syst[i] =
-        input[SRSLTE_TCOD_RATE * long_cb + NINPUTS * SRSLTE_TCOD_RATE + NINPUTS * (i - long_cb)];
-      h->parity[i] = input[SRSLTE_TCOD_RATE * long_cb + NINPUTS * SRSLTE_TCOD_RATE
-                          + NINPUTS * (i - long_cb) + 1];
-    }
-
-    // Run MAP DEC #2
-    map_gen_dec(&h->dec, h->syst, h->parity, h->llr2, long_cb);
-
-    //printf("llr2=");
-    //srslte_vec_fprint_f(stdout, h->llr2, long_cb);
-
-    
-    // Update a-priori LLR from the last iteration
-    for (i = 0; i < long_cb; i++) {
-      h->w[i] += h->llr2[deinter[i]] - h->llr1[i];
-    }
-  } else {
-    fprintf(stderr, "Error CB index not set (call srslte_tdec_gen_reset() first\n");    
-  }
-
-  // Increase number of iterations
-  h->n_iter++;
-}
-
-int srslte_tdec_gen_reset(srslte_tdec_gen_t * h, uint32_t long_cb)
-{
-  if (long_cb > h->max_long_cb) {
-    fprintf(stderr, "TDEC was initialized for max_long_cb=%d\n",
-            h->max_long_cb);
-    return -1;
-  }
-  memset(h->w, 0, sizeof(float) * long_cb);
-  h->current_cbidx = srslte_cbsegm_cbindex(long_cb);
-  h->current_cb_len = long_cb;
-  if (h->current_cbidx < 0) {
-    fprintf(stderr, "Invalid CB length %d\n", long_cb);
-    return -1; 
-  }
-  return 0;
-}
-
-void srslte_tdec_gen_decision(srslte_tdec_gen_t * h, uint8_t *output, uint32_t long_cb)
-{
-  uint16_t *deinter = h->interleaver[h->current_cbidx].reverse;
-  uint32_t i;
-  for (i = 0; i < long_cb; i++) {
-    output[i] = (h->llr2[deinter[i]] > 0) ? 1 : 0;    
-  }
-}
-
-void srslte_tdec_gen_decision_byte(srslte_tdec_gen_t * h, uint8_t *output, uint32_t long_cb)
-{
-  uint32_t i;
   uint8_t mask[8] = {0x80, 0x40, 0x20, 0x10, 0x8, 0x4, 0x2, 0x1};
-  uint16_t *deinter = h->interleaver[h->current_cbidx].reverse;
 
   // long_cb is always byte aligned
-  for (i = 0; i < long_cb/8; i++) {
+  for (uint32_t i = 0; i < long_cb/8; i++) {
-    uint8_t out0 = h->llr2[deinter[8*i+0]]>0?mask[0]:0;
+    uint8_t out0 = app1[8*i+0]>0?mask[0]:0;
-    uint8_t out1 = h->llr2[deinter[8*i+1]]>0?mask[1]:0;
+    uint8_t out1 = app1[8*i+1]>0?mask[1]:0;
-    uint8_t out2 = h->llr2[deinter[8*i+2]]>0?mask[2]:0;
+    uint8_t out2 = app1[8*i+2]>0?mask[2]:0;
-    uint8_t out3 = h->llr2[deinter[8*i+3]]>0?mask[3]:0;
+    uint8_t out3 = app1[8*i+3]>0?mask[3]:0;
-    uint8_t out4 = h->llr2[deinter[8*i+4]]>0?mask[4]:0;
+    uint8_t out4 = app1[8*i+4]>0?mask[4]:0;
-    uint8_t out5 = h->llr2[deinter[8*i+5]]>0?mask[5]:0;
+    uint8_t out5 = app1[8*i+5]>0?mask[5]:0;
-    uint8_t out6 = h->llr2[deinter[8*i+6]]>0?mask[6]:0;
+    uint8_t out6 = app1[8*i+6]>0?mask[6]:0;
-    uint8_t out7 = h->llr2[deinter[8*i+7]]>0?mask[7]:0;
+    uint8_t out7 = app1[8*i+7]>0?mask[7]:0;
 
     output[i] = out0 | out1 | out2 | out3 | out4 | out5 | out6 | out7;
   }
 }
 
-int srslte_tdec_gen_run_all(srslte_tdec_gen_t * h, float * input, uint8_t *output,
-                  uint32_t nof_iterations, uint32_t long_cb)
-{
-  uint32_t iter = 0;
-
-  if (srslte_tdec_gen_reset(h, long_cb)) {
-    return SRSLTE_ERROR; 
-  }
-
-  do {
-    srslte_tdec_gen_iteration(h, input, long_cb);
-    iter++;
-  } while (iter < nof_iterations);
-
-  srslte_tdec_gen_decision_byte(h, output, long_cb);
-  
-  return SRSLTE_SUCCESS;
-}

lib/src/phy/fec/turbodecoder_simd.c

@@ -1,542 +0,0 @@
-/**
- *
- * \section COPYRIGHT
- *
- * Copyright 2013-2015 Software Radio Systems Limited
- *
- * \section LICENSE
- *
- * This file is part of the srsLTE library.
- *
- * srsLTE is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as
- * published by the Free Software Foundation, either version 3 of
- * the License, or (at your option) any later version.
- *
- * srsLTE is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * A copy of the GNU Affero General Public License can be found in
- * the LICENSE file in the top-level directory of this distribution
- * and at http://www.gnu.org/licenses/.
- *
- */
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <stdint.h>
-#include <string.h>
-#include <strings.h>
-#include <math.h>
-
-#include "srslte/phy/fec/turbodecoder_simd.h"
-#include "srslte/phy/utils/vector.h"
-
-#include <inttypes.h>
-
-#define NUMSTATES       8
-#define NINPUTS         2
-#define TAIL            3
-#define TOTALTAIL       12
-
-#define INF 10000
-#define ZERO 0
-
-
-#ifdef LV_HAVE_SSE
-#include <smmintrin.h>
-
-// Define SSE/AVX implementations 
-void map_sse_beta(map_gen_t * s, int16_t * output, uint32_t long_cb);
-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_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_MAX_NPAR], uint32_t nof_cb, uint32_t long_cb)
-{
-  if (nof_cb == 1) {
-    map_sse_beta(s, output[0], long_cb);
-  } 
-#ifdef LV_HAVE_AVX2
-  else if (nof_cb == 2) {
-    map_avx_beta(s, output, long_cb);
-  }
-#endif
-}
-
-void map_simd_alpha(map_gen_t * s, uint32_t nof_cb, uint32_t long_cb)
-{
-  if (nof_cb == 1) {
-    map_sse_alpha(s, long_cb);
-  }
-#ifdef LV_HAVE_AVX2
-  else if (nof_cb == 2) {
-    map_avx_alpha(s, long_cb);
-  }
-#endif
-}
-void map_simd_gamma(map_gen_t * s, int16_t *input, int16_t *app, int16_t *parity, uint32_t cbidx, uint32_t nof_cb, uint32_t long_cb) 
-{
-  if (nof_cb == 1) {
-    map_sse_gamma(s, input, app, parity, long_cb);
-  } 
-#ifdef LV_HAVE_AVX2
-  else if (nof_cb == 2) {
-    map_avx_gamma(s, input, app, parity, cbidx, long_cb);
-  }    
-#endif
-}
-
-/* Inititalizes constituent decoder object */
-int map_simd_init(map_gen_t * h, uint32_t max_par_cb, uint32_t max_long_cb)
-{
-  bzero(h, sizeof(map_gen_t));
-
-  h->max_par_cb  = max_par_cb;
-  h->max_long_cb = max_long_cb;
-
-  h->alpha = srslte_vec_malloc(sizeof(int16_t) * (max_long_cb + SRSLTE_TCOD_TOTALTAIL + 1) * NUMSTATES * h->max_par_cb);
-  if (!h->alpha) {
-    perror("srslte_vec_malloc");
-    return -1;
-  }
-  h->branch = srslte_vec_malloc(sizeof(int16_t) * (max_long_cb + SRSLTE_TCOD_TOTALTAIL + 1) * NUMSTATES * h->max_par_cb);
-  if (!h->branch) {
-    perror("srslte_vec_malloc");
-    return -1;
-  }
-  return 0;
-}
-
-void map_simd_free(map_gen_t * h)
-{
-  if (h->alpha) {
-    free(h->alpha);
-  }
-  if (h->branch) {
-    free(h->branch);
-  }
-  bzero(h, sizeof(map_gen_t));
-}
-
-/* Runs one instance of a decoder */
-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_MAX_NPAR] = { NULL, NULL };
-  
-  // Compute branch metrics
-  switch(cb_mask) {
-    case 1:
-      nof_cb = 1; 
-      outptr[0] = output[0];
-      map_simd_gamma(h, input[0], app?app[0]:NULL, parity[0], 0, 1, long_cb);    
-      break;
-    case 2:
-      nof_cb = 1; 
-      outptr[0] = output[1];
-      map_simd_gamma(h, input[1], app?app[1]:NULL, parity[1], 0, 1, long_cb);    
-      break;
-    case 3:
-      nof_cb = 2; 
-      for (int i=0;i<2;i++) {    
-        outptr[i] = output[i];
-        map_simd_gamma(h, input[i], app?app[i]:NULL, parity[i], i, 2, long_cb);    
-      }
-      break;
-  }
-
-  // Forward recursion
-  map_simd_alpha(h, nof_cb, long_cb);
-
-  // Backwards recursion + LLR computation
-  map_simd_beta(h, outptr, nof_cb, long_cb);
-}
-
-/* Initializes the turbo decoder object */
-int srslte_tdec_simd_init(srslte_tdec_simd_t * h, uint32_t max_par_cb, uint32_t max_long_cb)
-{
-  int ret = -1;
-  bzero(h, sizeof(srslte_tdec_simd_t));
-  uint32_t len = max_long_cb + SRSLTE_TCOD_TOTALTAIL;
-
-  h->max_long_cb = max_long_cb;
-  h->max_par_cb  = max_par_cb; 
-  
-  for (int i=0;i<h->max_par_cb;i++) {
-    h->app1[i] = srslte_vec_malloc(sizeof(int16_t) * len);
-    if (!h->app1[i]) {
-      perror("srslte_vec_malloc");
-      goto clean_and_exit;
-    }
-    h->app2[i] = srslte_vec_malloc(sizeof(int16_t) * len);
-    if (!h->app2[i]) {
-      perror("srslte_vec_malloc");
-      goto clean_and_exit;
-    }
-    h->ext1[i] = srslte_vec_malloc(sizeof(int16_t) * len);
-    if (!h->ext1[i]) {
-      perror("srslte_vec_malloc");
-      goto clean_and_exit;
-    }
-    h->ext2[i] = srslte_vec_malloc(sizeof(int16_t) * len);
-    if (!h->ext2[i]) {
-      perror("srslte_vec_malloc");
-      goto clean_and_exit;
-    }
-    h->syst[i] = srslte_vec_malloc(sizeof(int16_t) * len);
-    if (!h->syst[i]) {
-      perror("srslte_vec_malloc");
-      goto clean_and_exit;
-    }
-    h->parity0[i] = srslte_vec_malloc(sizeof(int16_t) * len);
-    if (!h->parity0[i]) {
-      perror("srslte_vec_malloc");
-      goto clean_and_exit;
-    }
-    h->parity1[i] = srslte_vec_malloc(sizeof(int16_t) * len);
-    if (!h->parity1[i]) {
-      perror("srslte_vec_malloc");
-      goto clean_and_exit;
-    }
-    
-  }
-
-  if (map_simd_init(&h->dec, h->max_par_cb, h->max_long_cb)) {
-    goto clean_and_exit;
-  }
-
-  for (int i=0;i<SRSLTE_NOF_TC_CB_SIZES;i++) {
-    if (srslte_tc_interl_init(&h->interleaver[i], srslte_cbsegm_cbsize(i)) < 0) {
-      goto clean_and_exit;
-    }
-    srslte_tc_interl_LTE_gen(&h->interleaver[i], srslte_cbsegm_cbsize(i));
-  }
-  h->current_cbidx = -1; 
-  h->cb_mask = 0; 
-  ret = 0;
-clean_and_exit:if (ret == -1) {
-    srslte_tdec_simd_free(h);
-  }
-  return ret;
-}
-
-void srslte_tdec_simd_free(srslte_tdec_simd_t * h)
-{
-  for (int i=0;i<h->max_par_cb;i++) {
-    if (h->app1[i]) {
-      free(h->app1[i]);
-    }
-    if (h->app2[i]) {
-      free(h->app2[i]);
-    }
-    if (h->ext1[i]) {
-      free(h->ext1[i]);
-    }
-    if (h->ext2[i]) {
-      free(h->ext2[i]);
-    }
-    if (h->syst[i]) {
-      free(h->syst[i]);
-    }
-    if (h->parity0[i]) {
-      free(h->parity0[i]);
-    }
-    if (h->parity1[i]) {
-      free(h->parity1[i]);
-    }    
-  }
-
-  map_simd_free(&h->dec);
-
-  for (int i=0;i<SRSLTE_NOF_TC_CB_SIZES;i++) {
-    srslte_tc_interl_free(&h->interleaver[i]);    
-  }
-
-  bzero(h, sizeof(srslte_tdec_simd_t));
-}
-
-/* Deinterleaves the 3 streams from the input (systematic and 2 parity bits) into 
- * 3 buffers ready to be used by compute_gamma() 
- */
-void deinterleave_input_simd(srslte_tdec_simd_t *h, int16_t *input, uint32_t cbidx, uint32_t long_cb) {
-  uint32_t i;
- 
-  __m128i *inputPtr = (__m128i*) input; 
-  __m128i in0, in1, in2;
-  __m128i s0, s1, s2, s;
-  __m128i p00, p01, p02, p0;
-  __m128i p10, p11, p12, p1;
-  
-  __m128i *sysPtr = (__m128i*) h->syst[cbidx]; 
-  __m128i *pa0Ptr = (__m128i*) h->parity0[cbidx]; 
-  __m128i *pa1Ptr = (__m128i*) h->parity1[cbidx]; 
-  
-  // pick bits 0, 3, 6 from 1st word
-  __m128i s0_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,13,12,7,6,1,0);
-  // pick bits 1, 4, 7 from 2st word
-  __m128i s1_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,15,14,9,8,3,2,0xff,0xff,0xff,0xff,0xff,0xff);
-  // pick bits 2, 5 from 3rd word
-  __m128i s2_mask = _mm_set_epi8(11,10,5,4,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-
-  // pick bits 1, 4, 7 from 1st word
-  __m128i p00_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,15,14,9,8,3,2);
-  // pick bits 2, 5, from 2st word
-  __m128i p01_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,11,10,5,4,0xff,0xff,0xff,0xff,0xff,0xff);
-  // pick bits 0, 3, 6 from 3rd word
-  __m128i p02_mask = _mm_set_epi8(13,12,7,6,1,0,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-  
-  // pick bits 2, 5 from 1st word
-  __m128i p10_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,11,10,5,4);
-  // pick bits 0, 3, 6, from 2st word
-  __m128i p11_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,13,12,7,6,1,0,0xff,0xff,0xff,0xff);
-  // pick bits 1, 4, 7 from 3rd word
-  __m128i p12_mask = _mm_set_epi8(15,14,9,8,3,2,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
-      
-  // Split systematic and parity bits
-  for (i = 0; i < long_cb/8; i++) {
-        
-    in0 = _mm_load_si128(inputPtr); inputPtr++; 
-    in1 = _mm_load_si128(inputPtr); inputPtr++;   
-    in2 = _mm_load_si128(inputPtr); inputPtr++;
-    
-    /* Deinterleave Systematic bits */
-    s0 = _mm_shuffle_epi8(in0, s0_mask);
-    s1 = _mm_shuffle_epi8(in1, s1_mask);
-    s2 = _mm_shuffle_epi8(in2, s2_mask);    
-    s = _mm_or_si128(s0, s1);
-    s = _mm_or_si128(s, s2);
-
-    _mm_store_si128(sysPtr, s);
-    sysPtr++;
-
-    /* Deinterleave parity 0 bits */
-    p00 = _mm_shuffle_epi8(in0, p00_mask);
-    p01 = _mm_shuffle_epi8(in1, p01_mask);
-    p02 = _mm_shuffle_epi8(in2, p02_mask);    
-    p0 = _mm_or_si128(p00, p01);
-    p0 = _mm_or_si128(p0, p02);
-    
-    _mm_store_si128(pa0Ptr, p0);
-    pa0Ptr++;
-
-    /* Deinterleave parity 1 bits */
-    p10 = _mm_shuffle_epi8(in0, p10_mask);
-    p11 = _mm_shuffle_epi8(in1, p11_mask);
-    p12 = _mm_shuffle_epi8(in2, p12_mask);    
-    p1 = _mm_or_si128(p10, p11);
-    p1 = _mm_or_si128(p1, p12);
-
-    _mm_store_si128(pa1Ptr, p1);
-    pa1Ptr++;    
-
-  }
-  
-  for (i = 0; i < 3; i++) {
-    h->syst[cbidx][i+long_cb]    = input[3*long_cb + 2*i];
-    h->parity0[cbidx][i+long_cb] = input[3*long_cb + 2*i + 1];
-  }
-  for (i = 0; i < 3; i++) {
-    h->app2[cbidx][i+long_cb]    = input[3*long_cb + 6 + 2*i];
-    h->parity1[cbidx][i+long_cb] = input[3*long_cb + 6 + 2*i + 1];
-  }
-
-}
-
-/* Runs 1 turbo decoder iteration */
-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_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;
-    
-#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);
-        deinterleave_input_simd(h, input[i], i, long_cb);        
-      }        
-    }
-    
-    // Add apriori information to decoder 1 
-    for (int i=0;i<h->max_par_cb;i++) {
-      if (h->n_iter[i] > 0 && input[i]) {
-        srslte_vec_sub_sss(h->app1[i], h->ext1[i], h->app1[i], long_cb);
-      }
-    }
-
-    // Run MAP DEC #1
-    for (int i=0;i<h->max_par_cb;i++) {
-      if (input[i]) {
-        tmp_app[i] = h->n_iter[i]?h->app1[i]:NULL;
-      } else {
-        tmp_app[i] = NULL; 
-      }
-    }
-    map_simd_dec(&h->dec, h->syst, tmp_app, h->parity0, h->ext1, h->cb_mask, long_cb);            
-    
-    // Convert aposteriori information into extrinsic information    
-    for (int i=0;i<h->max_par_cb;i++) {
-      if (h->n_iter[i] > 0 && input[i]) {
-        srslte_vec_sub_sss(h->ext1[i], h->app1[i], h->ext1[i], long_cb);
-      }
-    }
-    
-    // Interleave extrinsic output of DEC1 to form apriori info for decoder 2
-    for (int i=0;i<h->max_par_cb;i++) {
-      if (input[i]) {
-        srslte_vec_lut_sss(h->ext1[i], deinter, h->app2[i], long_cb);
-      }
-    }
-
-    // Run MAP DEC #2. 2nd decoder uses apriori information as systematic bits
-    map_simd_dec(&h->dec, h->app2, NULL, h->parity1, h->ext2, h->cb_mask, long_cb);
-
-    // Deinterleaved extrinsic bits become apriori info for decoder 1 
-    for (int i=0;i<h->max_par_cb;i++) {
-      if (input[i]) {
-        srslte_vec_lut_sss(h->ext2[i], inter, h->app1[i], long_cb);
-      }
-    }
-
-    for (int i=0;i<h->max_par_cb;i++) {
-      if (input[i]) {
-        h->n_iter[i]++;     
-      }
-    }
-  } else {
-    fprintf(stderr, "Error CB index not set (call srslte_tdec_simd_reset() first\n");    
-  }
-}
-
-/* Resets the decoder and sets the codeblock length */
-int srslte_tdec_simd_reset(srslte_tdec_simd_t * h, uint32_t long_cb)
-{
-  if (long_cb > h->max_long_cb) {
-    fprintf(stderr, "TDEC was initialized for max_long_cb=%d\n",
-            h->max_long_cb);
-    return -1;
-  }
-  for (int i=0;i<h->max_par_cb;i++) {
-    h->n_iter[i] = 0;     
-  }
-  h->cb_mask = 0; 
-  h->current_cbidx = srslte_cbsegm_cbindex(long_cb);
-  if (h->current_cbidx < 0) {
-    fprintf(stderr, "Invalid CB length %d\n", long_cb);
-    return -1; 
-  }
-  return 0;
-}
-
-int srslte_tdec_simd_reset_cb(srslte_tdec_simd_t * h, uint32_t cb_idx)
-{
-  h->n_iter[cb_idx] = 0; 
-  return 0;
-}
-
-int srslte_tdec_simd_get_nof_iterations_cb(srslte_tdec_simd_t * h, uint32_t cb_idx)
-{
-  return h->n_iter[cb_idx];
-}
-
-void tdec_simd_decision(srslte_tdec_simd_t * h, uint8_t *output, uint32_t cbidx, uint32_t long_cb)
-{
-  __m128i zero     = _mm_set1_epi16(0);
-  __m128i lsb_mask = _mm_set1_epi16(1);
-  
-  __m128i *appPtr = (__m128i*) h->app1[cbidx];
-  __m128i *outPtr = (__m128i*) output;
-  __m128i ap, out, out0, out1; 
-    
-  for (uint32_t i = 0; i < long_cb/16; i++) {
-    ap   = _mm_load_si128(appPtr); appPtr++;    
-    out0 = _mm_and_si128(_mm_cmpgt_epi16(ap, zero), lsb_mask);
-    ap   = _mm_load_si128(appPtr); appPtr++;
-    out1 = _mm_and_si128(_mm_cmpgt_epi16(ap, zero), lsb_mask);
-    
-    out  = _mm_packs_epi16(out0, out1);
-    _mm_store_si128(outPtr, out);
-    outPtr++;
-  }
-  if (long_cb%16) {
-    for (int i=0;i<8;i++) {
-      output[long_cb-8+i] = h->app1[cbidx][long_cb-8+i]>0?1:0;
-    }
-  }
-}
-
-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);
-  }
-}
-
-void srslte_tdec_simd_decision_byte_cb(srslte_tdec_simd_t * h, uint8_t *output, uint32_t cbidx, uint32_t long_cb)
-{
-  uint8_t mask[8] = {0x80, 0x40, 0x20, 0x10, 0x8, 0x4, 0x2, 0x1};
-  
-  // long_cb is always byte aligned
-  for (uint32_t i = 0; i < long_cb/8; i++) {
-    uint8_t out0 = h->app1[cbidx][8*i+0]>0?mask[0]:0;
-    uint8_t out1 = h->app1[cbidx][8*i+1]>0?mask[1]:0;
-    uint8_t out2 = h->app1[cbidx][8*i+2]>0?mask[2]:0;
-    uint8_t out3 = h->app1[cbidx][8*i+3]>0?mask[3]:0;
-    uint8_t out4 = h->app1[cbidx][8*i+4]>0?mask[4]:0;
-    uint8_t out5 = h->app1[cbidx][8*i+5]>0?mask[5]:0;
-    uint8_t out6 = h->app1[cbidx][8*i+6]>0?mask[6]:0;
-    uint8_t out7 = h->app1[cbidx][8*i+7]>0?mask[7]:0;
-    
-    output[i] = out0 | out1 | out2 | out3 | out4 | out5 | out6 | out7;
-  }
-}
-
-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++) {
-    if (output[i]) {
-      srslte_tdec_simd_decision_byte_cb(h, output[i], i, long_cb);
-    }
-  }
-}
-
-
-/* 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_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)) {
-    return SRSLTE_ERROR; 
-  }
-
-  do {
-    srslte_tdec_simd_iteration(h, input, long_cb);
-  } while (h->n_iter[0] < nof_iterations);
-
-  srslte_tdec_simd_decision_byte(h, output, long_cb);
-  
-  return SRSLTE_SUCCESS;
-}
-
-#endif
-
-

lib/src/phy/fec/turbodecoder_simd_inter.c

@@ -1,299 +0,0 @@
-/**
- *
- * \section COPYRIGHT
- *
- * Copyright 2013-2015 Software Radio Systems Limited
- *
- * \section LICENSE
- *
- * This file is part of the srsLTE library.
- *
- * srsLTE is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as
- * published by the Free Software Foundation, either version 3 of
- * the License, or (at your option) any later version.
- *
- * srsLTE is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * A copy of the GNU Affero General Public License can be found in
- * the LICENSE file in the top-level directory of this distribution
- * and at http://www.gnu.org/licenses/.
- *
- */
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <stdint.h>
-#include <string.h>
-#include <strings.h>
-#include <math.h>
-
-#include "srslte/phy/fec/turbodecoder_simd_inter.h"
-#include "srslte/phy/utils/vector.h"
-
-#define TOTALTAIL       12
-
-#ifdef LV_HAVE_SSE
-#include <smmintrin.h>
-
-void map_see_inter_alpha(srslte_tdec_simd_inter_t * s, int16_t *input, int16_t *parity, uint32_t long_cb);
-void map_sse_inter_beta(srslte_tdec_simd_inter_t * s, int16_t *input, int16_t *parity, int16_t * output, uint32_t long_cb);
-void sse_inter_update_w(srslte_tdec_simd_inter_t *h, uint16_t *deinter, uint32_t long_cb); 
-void sse_inter_extract_syst1(srslte_tdec_simd_inter_t *h, uint16_t *inter, uint32_t long_cb);
-
-
-static void map_sse_inter_dec(srslte_tdec_simd_inter_t * h, int16_t * input, int16_t * parity, int16_t * output,
-                 uint32_t long_cb)
-{
-  map_see_inter_alpha(h, input, parity, long_cb);
-  map_sse_inter_beta(h, input, parity, output, long_cb);
-}
-
-/************************************************
- *
- *  TURBO DECODER INTERFACE
- *
- ************************************************/
-int srslte_tdec_simd_inter_init(srslte_tdec_simd_inter_t * h, uint32_t max_par_cb, uint32_t max_long_cb)
-{
-  int ret = -1;
-  bzero(h, sizeof(srslte_tdec_simd_inter_t));
-  uint32_t len = max_long_cb + 12;
-
-  h->max_long_cb = max_long_cb;
-  h->max_par_cb  = max_par_cb;
-
-  h->llr1 = srslte_vec_malloc(sizeof(int16_t) * len * h->max_par_cb);
-  if (!h->llr1) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->llr2 = srslte_vec_malloc(sizeof(int16_t) * len * h->max_par_cb);
-  if (!h->llr2) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->w = srslte_vec_malloc(sizeof(int16_t) * len * h->max_par_cb);
-  if (!h->w) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->syst0 = srslte_vec_malloc(sizeof(int16_t) * len * h->max_par_cb);
-  if (!h->syst0) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->syst1 = srslte_vec_malloc(sizeof(int16_t) * len * h->max_par_cb);
-  if (!h->syst1) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->parity0 = srslte_vec_malloc(sizeof(int16_t) * len * h->max_par_cb);
-  if (!h->parity0) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->parity1 = srslte_vec_malloc(sizeof(int16_t) * len * h->max_par_cb);
-  if (!h->parity1) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-  h->alpha = srslte_vec_malloc(sizeof(int16_t) * 8*(len+12) * h->max_par_cb);
-  if (!h->alpha) {
-    perror("srslte_vec_malloc");
-    goto clean_and_exit;
-  }
-
-  for (int i=0;i<SRSLTE_NOF_TC_CB_SIZES;i++) {
-    if (srslte_tc_interl_init(&h->interleaver[i], srslte_cbsegm_cbsize(i)) < 0) {
-      goto clean_and_exit;
-    }
-    srslte_tc_interl_LTE_gen(&h->interleaver[i], srslte_cbsegm_cbsize(i));
-  }
-  h->current_cbidx = -1; 
-  ret = 0;
-clean_and_exit:if (ret == -1) {
-    srslte_tdec_simd_inter_free(h);
-  }
-  return ret;
-}
-
-void srslte_tdec_simd_inter_free(srslte_tdec_simd_inter_t * h)
-{
-  if (h->llr1) {
-    free(h->llr1);
-  }
-  if (h->llr2) {
-    free(h->llr2);
-  }
-  if (h->w) {
-    free(h->w);
-  }
-  if (h->syst0) {
-    free(h->syst0);
-  }
-  if (h->syst1) {
-    free(h->syst1);
-  }
-  if (h->parity0) {
-    free(h->parity0);
-  }
-  if (h->parity1) {
-    free(h->parity1);
-  }
-  if (h->alpha) {
-    free(h->alpha);
-  }
-
-  for (int i=0;i<SRSLTE_NOF_TC_CB_SIZES;i++) {
-    srslte_tc_interl_free(&h->interleaver[i]);    
-  }
-
-  bzero(h, sizeof(srslte_tdec_simd_inter_t));
-}
-
-
-/* Deinterleave for inter-frame parallelization */
-void extract_input(srslte_tdec_simd_inter_t *h, int16_t *input, uint32_t cbidx, uint32_t long_cb) 
-{  
-  for (int i=0;i<long_cb;i++) {
-    h->syst0[h->max_par_cb*i+cbidx]    = input[3*i+0];
-    h->parity0[h->max_par_cb*i+cbidx] = input[3*i+1];
-    h->parity1[h->max_par_cb*i+cbidx] = input[3*i+2];
-  }
-  for (int i = long_cb; i < long_cb + 3; i++) {
-    h->syst0[h->max_par_cb*i+cbidx] = input[3*long_cb + 2*(i - long_cb)];
-    h->syst1[h->max_par_cb*i+cbidx] = input[3*long_cb + 2*(i - long_cb)];
-    h->parity0[h->max_par_cb*i+cbidx] = input[3*long_cb + 2*(i - long_cb) + 1];
-    h->parity0[h->max_par_cb*i+cbidx] = input[3*long_cb + 2*(i - long_cb) + 2];
-  }
-}
-
-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) {
-
-    uint16_t *inter = h->interleaver[h->current_cbidx].forward;
-    uint16_t *deinter = h->interleaver[h->current_cbidx].reverse;
-    
-    // Prepare systematic and parity bits for MAP DEC #1
-    for (int i=0;i<nof_cb;i++) {
-      if (h->n_iter[i] == 0) {
-        extract_input(h, input[i], i, long_cb);
-      }
-      srslte_vec_sum_sss(h->syst0, h->w, h->syst0, long_cb*h->max_par_cb);      
-    }
-    
-    // Run MAP DEC #1
-    map_sse_inter_dec(h, h->syst0, h->parity0, h->llr1, long_cb);
-
-    // Prepare systematic and parity bits for MAP DEC #1
-    sse_inter_extract_syst1(h, inter, long_cb);
-
-    // Run MAP DEC #2
-    map_sse_inter_dec(h, h->syst1, h->parity1, h->llr2, long_cb);
-    
-    // Update a-priori LLR from the last iteration
-    sse_inter_update_w(h, deinter, long_cb);
-    
-  } else {
-    fprintf(stderr, "Error CB index not set (call srslte_tdec_simd_inter_reset() first\n");    
-  }
-}
-
-int srslte_tdec_simd_inter_reset_cb(srslte_tdec_simd_inter_t * h, uint32_t cb_idx)
-{
-  for (int i=0;i<h->current_long_cb;i++) {
-    h->w[h->max_par_cb*i+cb_idx] = 0; 
-  }
-  return 0;
-}
-
-int srslte_tdec_simd_inter_reset(srslte_tdec_simd_inter_t * h, uint32_t long_cb)
-{
-  if (long_cb > h->max_long_cb) {
-    fprintf(stderr, "TDEC was initialized for max_long_cb=%d\n",
-            h->max_long_cb);
-    return -1;
-  }
-  h->current_long_cb = long_cb; 
-  h->current_cbidx   = srslte_cbsegm_cbindex(long_cb);
-  if (h->current_cbidx < 0) {
-    fprintf(stderr, "Invalid CB length %d\n", long_cb);
-    return -1; 
-  }
-  memset(h->w, 0, sizeof(int16_t) * long_cb * h->max_par_cb);
-  return 0; 
-}
-
-void srslte_tdec_simd_inter_decision_cb(srslte_tdec_simd_inter_t * h, uint8_t *output, uint32_t cb_idx, uint32_t long_cb)
-{
-  uint16_t *deinter = h->interleaver[h->current_cbidx].reverse;
-  uint32_t i;
-  for (i = 0; i < long_cb; i++) {
-    output[i] = (h->llr2[h->max_par_cb*deinter[i]+cb_idx] > 0) ? 1 : 0;    
-  }
-}
-
-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);
-  }
-}
-
-void srslte_tdec_simd_inter_decision_byte_cb(srslte_tdec_simd_inter_t * h, uint8_t *output, uint32_t cb_idx, uint32_t long_cb)
-{
-  uint32_t i;
-  uint8_t mask[8] = {0x80, 0x40, 0x20, 0x10, 0x8, 0x4, 0x2, 0x1};
-  uint16_t *deinter = h->interleaver[h->current_cbidx].reverse;
-  
-#define indexOf_cb(idx, cb) (h->max_par_cb*(deinter[8*i+idx])+cb)
-  
-  // long_cb is always byte aligned
-  for (i = 0; i < long_cb/8; i++) {
-    uint8_t out0 = h->llr2[indexOf_cb(0, cb_idx)]>0?mask[0]:0;
-    uint8_t out1 = h->llr2[indexOf_cb(1, cb_idx)]>0?mask[1]:0;
-    uint8_t out2 = h->llr2[indexOf_cb(2, cb_idx)]>0?mask[2]:0;
-    uint8_t out3 = h->llr2[indexOf_cb(3, cb_idx)]>0?mask[3]:0;
-    uint8_t out4 = h->llr2[indexOf_cb(4, cb_idx)]>0?mask[4]:0;
-    uint8_t out5 = h->llr2[indexOf_cb(5, cb_idx)]>0?mask[5]:0;
-    uint8_t out6 = h->llr2[indexOf_cb(6, cb_idx)]>0?mask[6]:0;
-    uint8_t out7 = h->llr2[indexOf_cb(7, cb_idx)]>0?mask[7]:0;
-    
-    output[i] = out0 | out1 | out2 | out3 | out4 | out5 | out6 | out7; 
-  }
-}
-
-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);
-  }
-}
-
-int srslte_tdec_simd_inter_run_all(srslte_tdec_simd_inter_t * h, 
-                                  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;
-
-  if (srslte_tdec_simd_inter_reset(h, long_cb)) {
-    return SRSLTE_ERROR; 
-  }
-
-  do {
-    srslte_tdec_simd_inter_iteration(h, input, nof_cb, long_cb);
-    iter++;
-  } while (iter < nof_iterations);
-
-  srslte_tdec_simd_inter_decision_byte(h, output, nof_cb, long_cb);
-  
-  return SRSLTE_SUCCESS;
-}
-
-#endif

lib/src/phy/fec/turbodecoder_sse.c

@@ -31,13 +31,15 @@
 #include <strings.h>
 #include <math.h>
 
-#include "srslte/phy/fec/turbodecoder_simd.h"
+#include "srslte/phy/fec/turbodecoder_sse.h"
 #include "srslte/phy/utils/vector.h"
 
 #include <inttypes.h>
 
 #ifdef LV_HAVE_SSE
 #include <smmintrin.h>
+#include <srslte/phy/fec/turbodecoder_sse.h>
+
 #endif
 
 
@@ -47,21 +49,46 @@
 #define TOTALTAIL       12
 
 #define INF 10000
-#define ZERO 0
 
 
 #ifdef LV_HAVE_SSE
 
-/*
+
+#define debug_enabled 0
+
+#if debug_enabled
+#define debug_state(c,d)       printf("k=%5d, in=%5d, pa=%5d, out=%5d, alpha=", k-d,\
+                          s->branch[2*(k-d)] + s->branch[2*(k-d)+1], \
+                          -s->branch[2*(k-d)] + s->branch[2*(k-d)+1], output[k-d]);print_128i(alpha_k);\
+                          printf(", beta=");print_128i(beta_k);printf("\n");
+
 static void print_128i(__m128i x) {
   int16_t *s = (int16_t*) &x;
-  printf("[%d", s[0]);
+  printf("[%5d", s[0]);
   for (int i=1;i<8;i++) {
-    printf(",%d", s[i]);
+    printf(",%5d", s[i]);
   }
-  printf("]\n");
+  printf("]");
 }
-*/
+
+static uint32_t max_128i(__m128i x) {
+  int16_t *s = (int16_t*) &x;
+  int16_t m = -INF;
+  uint32_t max = 0;
+  for (int i=1;i<8;i++) {
+    if (s[i] > m) {
+      max = i;
+      m = s[i];
+    }
+  }
+  return max;
+}
+
+#else
+#define debug_state(c,d)
+#endif
+
+
 //#define use_beta_transposed_max
 
 #ifndef use_beta_transposed_max
@@ -75,7 +102,7 @@ static inline int16_t hMax(__m128i buffer)
 }
 
 /* Computes beta values */
-void map_sse_beta(map_gen_t * s, int16_t * output, uint32_t long_cb)
+void tdec_sse_beta(tdec_sse_t * s, int16_t * output, uint32_t long_cb)
 {
   int k;
   uint32_t end = long_cb + 3;
@@ -141,7 +168,9 @@ void map_sse_beta(map_gen_t * s, int16_t * output, uint32_t long_cb)
     alphaPtr--;\
     bp = _mm_add_epi16(bp, alpha_k);\
     bn = _mm_add_epi16(bn, alpha_k);\
-    output[k-d] = hMax(bn)-hMax(bp);
+    output[k-d] = hMax(bn)-hMax(bp);\
+    debug_state(c,d);
+
 
   /* The tail does not require to load alpha or produce outputs. Only update 
    * beta metrics accordingly */
@@ -179,7 +208,7 @@ void map_sse_beta(map_gen_t * s, int16_t * output, uint32_t long_cb)
 #endif
 
 /* Computes alpha metrics */
-void map_sse_alpha(map_gen_t * s, uint32_t long_cb)
+void tdec_sse_alpha(tdec_sse_t * s, uint32_t long_cb)
 {
   uint32_t k;
   int16_t *alpha = s->alpha;
@@ -268,7 +297,7 @@ void map_sse_alpha(map_gen_t * s, uint32_t long_cb)
 }
 
 /* Compute branch metrics (gamma) */
-void map_sse_gamma(map_gen_t * h, int16_t *input, int16_t *app, int16_t *parity, uint32_t long_cb) 
+void tdec_sse_gamma(tdec_sse_t * h, int16_t *input, int16_t *app, int16_t *parity, uint32_t long_cb) 
 {
   __m128i res00, res10, res01, res11, res0, res1; 
   __m128i in, ap, pa, g1, g0;
@@ -313,6 +342,8 @@ void map_sse_gamma(map_gen_t * h, int16_t *input, int16_t *app, int16_t *parity,
     resPtr++;
     _mm_store_si128(resPtr, res1);    
     resPtr++;
+
+    //printf("k=%d, in=%d, pa=%d, g0=%d, g1=%d\n", i, input[i], parity[i], h->branch[2*i], h->branch[2*i+1]);
   }
 
   for (int i=long_cb;i<long_cb+3;i++) {
@@ -322,6 +353,161 @@ void map_sse_gamma(map_gen_t * h, int16_t *input, int16_t *app, int16_t *parity,
 }
 
 
+/* Inititalizes constituent decoder object */
+int tdec_sse_init(void **hh, uint32_t max_long_cb)
+{
+  *hh = calloc(1, sizeof(tdec_sse_t));
+  
+  tdec_sse_t *h = (tdec_sse_t*) *hh; 
+  
+  h->max_long_cb = max_long_cb;
+
+  h->alpha = srslte_vec_malloc(sizeof(int16_t) * (max_long_cb + TOTALTAIL + 1) * NUMSTATES);
+  if (!h->alpha) {
+    perror("srslte_vec_malloc");
+    return -1;
+  }
+  h->branch = srslte_vec_malloc(sizeof(int16_t) * (max_long_cb + TOTALTAIL + 1) * NUMSTATES);
+  if (!h->branch) {
+    perror("srslte_vec_malloc");
+    return -1;
+  }
+  return 1;
+}
+
+void tdec_sse_free(void *hh)
+{
+  tdec_sse_t *h = (tdec_sse_t*) hh;
+
+  if (h->alpha) {
+    free(h->alpha);
+  }
+  if (h->branch) {
+    free(h->branch);
+  }
+  bzero(h, sizeof(tdec_sse_t));
+}
+
+/* Runs one instance of a decoder */
+void tdec_sse_dec(void *hh, int16_t * input, int16_t *app, int16_t * parity,
+                 int16_t *output, uint32_t long_cb)
+{
+  tdec_sse_t *h = (tdec_sse_t*) hh;
+
+  // Compute branch metrics
+  tdec_sse_gamma(h, input, app, parity, long_cb);
+
+  // Forward recursion
+  tdec_sse_alpha(h, long_cb);
+
+  // Backwards recursion + LLR computation
+  tdec_sse_beta(h, output, long_cb);
+}
+
+/* Deinterleaves the 3 streams from the input (systematic and 2 parity bits) into
+ * 3 buffers ready to be used by compute_gamma()
+ */
+void tdec_sse_extract_input(int16_t *input, int16_t *syst0, int16_t *app2, int16_t *parity0, int16_t *parity1, uint32_t long_cb) {
+  uint32_t i;
+
+  __m128i *inputPtr = (__m128i*) input;
+  __m128i in0, in1, in2;
+  __m128i s0, s1, s2, s;
+  __m128i p00, p01, p02, p0;
+  __m128i p10, p11, p12, p1;
+
+  __m128i *sysPtr = (__m128i*) syst0;
+  __m128i *pa0Ptr = (__m128i*) parity0;
+  __m128i *pa1Ptr = (__m128i*) parity1;
+
+  // pick bits 0, 3, 6 from 1st word
+  __m128i s0_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,13,12,7,6,1,0);
+  // pick bits 1, 4, 7 from 2st word
+  __m128i s1_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,15,14,9,8,3,2,0xff,0xff,0xff,0xff,0xff,0xff);
+  // pick bits 2, 5 from 3rd word
+  __m128i s2_mask = _mm_set_epi8(11,10,5,4,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
+
+  // pick bits 1, 4, 7 from 1st word
+  __m128i p00_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,15,14,9,8,3,2);
+  // pick bits 2, 5, from 2st word
+  __m128i p01_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,11,10,5,4,0xff,0xff,0xff,0xff,0xff,0xff);
+  // pick bits 0, 3, 6 from 3rd word
+  __m128i p02_mask = _mm_set_epi8(13,12,7,6,1,0,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
+
+  // pick bits 2, 5 from 1st word
+  __m128i p10_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,11,10,5,4);
+  // pick bits 0, 3, 6, from 2st word
+  __m128i p11_mask = _mm_set_epi8(0xff,0xff,0xff,0xff,0xff,0xff,13,12,7,6,1,0,0xff,0xff,0xff,0xff);
+  // pick bits 1, 4, 7 from 3rd word
+  __m128i p12_mask = _mm_set_epi8(15,14,9,8,3,2,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff);
+
+  // Split systematic and parity bits
+  for (i = 0; i < long_cb/8; i++) {
+
+    in0 = _mm_load_si128(inputPtr); inputPtr++;
+    in1 = _mm_load_si128(inputPtr); inputPtr++;
+    in2 = _mm_load_si128(inputPtr); inputPtr++;
+
+    /* Deinterleave Systematic bits */
+    s0 = _mm_shuffle_epi8(in0, s0_mask);
+    s1 = _mm_shuffle_epi8(in1, s1_mask);
+    s2 = _mm_shuffle_epi8(in2, s2_mask);
+    s = _mm_or_si128(s0, s1);
+    s = _mm_or_si128(s, s2);
+
+    _mm_store_si128(sysPtr, s);
+    sysPtr++;
+
+    /* Deinterleave parity 0 bits */
+    p00 = _mm_shuffle_epi8(in0, p00_mask);
+    p01 = _mm_shuffle_epi8(in1, p01_mask);
+    p02 = _mm_shuffle_epi8(in2, p02_mask);
+    p0 = _mm_or_si128(p00, p01);
+    p0 = _mm_or_si128(p0, p02);
+
+    _mm_store_si128(pa0Ptr, p0);
+    pa0Ptr++;
+
+    /* Deinterleave parity 1 bits */
+    p10 = _mm_shuffle_epi8(in0, p10_mask);
+    p11 = _mm_shuffle_epi8(in1, p11_mask);
+    p12 = _mm_shuffle_epi8(in2, p12_mask);
+    p1 = _mm_or_si128(p10, p11);
+    p1 = _mm_or_si128(p1, p12);
+
+    _mm_store_si128(pa1Ptr, p1);
+    pa1Ptr++;
+
+  }
+
+  for (i = 0; i < 3; i++) {
+    syst0[i+long_cb]    = input[3*long_cb + 2*i];
+    parity0[i+long_cb] = input[3*long_cb + 2*i + 1];
+  }
+  for (i = 0; i < 3; i++) {
+    app2[i+long_cb]    = input[3*long_cb + 6 + 2*i];
+    parity1[i+long_cb] = input[3*long_cb + 6 + 2*i + 1];
+  }
+}
+
+void tdec_sse_decision_byte(int16_t *app1, uint8_t *output, uint32_t long_cb)
+{
+  uint8_t mask[8] = {0x80, 0x40, 0x20, 0x10, 0x8, 0x4, 0x2, 0x1};
+
+  // long_cb is always byte aligned
+  for (uint32_t i = 0; i < long_cb/8; i++) {
+    uint8_t out0 = app1[8*i+0]>0?mask[0]:0;
+    uint8_t out1 = app1[8*i+1]>0?mask[1]:0;
+    uint8_t out2 = app1[8*i+2]>0?mask[2]:0;
+    uint8_t out3 = app1[8*i+3]>0?mask[3]:0;
+    uint8_t out4 = app1[8*i+4]>0?mask[4]:0;
+    uint8_t out5 = app1[8*i+5]>0?mask[5]:0;
+    uint8_t out6 = app1[8*i+6]>0?mask[6]:0;
+    uint8_t out7 = app1[8*i+7]>0?mask[7]:0;
+
+    output[i] = out0 | out1 | out2 | out3 | out4 | out5 | out6 | out7;
+  }
+}
 
 
 
@@ -381,7 +567,7 @@ static inline __m128i transposed_max(__m128i a, __m128i b, __m128i c, __m128i d,
   return res;   
 }
 
-void map_sse_beta(map_gen_t * s, int16_t * output, uint32_t long_cb)
+void tdec_sse_beta(tdec_sse_t * s, int16_t * output, uint32_t long_cb)
 {
   int k;
   uint32_t end = long_cb + 3;

lib/src/phy/fec/turbodecoder_sse_inter.c

@@ -1,202 +0,0 @@
-/**
- *
- * \section COPYRIGHT
- *
- * Copyright 2013-2015 Software Radio Systems Limited
- *
- * \section LICENSE
- *
- * This file is part of the srsLTE library.
- *
- * srsLTE is free software: you can redistribute it and/or modify
- * it under the terms of the GNU Affero General Public License as
- * published by the Free Software Foundation, either version 3 of
- * the License, or (at your option) any later version.
- *
- * srsLTE is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- * GNU Affero General Public License for more details.
- *
- * A copy of the GNU Affero General Public License can be found in
- * the LICENSE file in the top-level directory of this distribution
- * and at http://www.gnu.org/licenses/.
- *
- */
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <stdint.h>
-#include <string.h>
-#include <strings.h>
-#include <math.h>
-
-#include "srslte/phy/fec/turbodecoder_simd_inter.h"
-#include "srslte/phy/utils/vector.h"
-
-
-#define NCB 8
-
-#define INF 10000
-
-#ifdef LV_HAVE_SSE
-#include <smmintrin.h>
-
-void sse_inter_extract_syst1(srslte_tdec_simd_inter_t *h, uint16_t *inter, uint32_t long_cb) 
-{
-  __m128i *llr1Ptr  = (__m128i*) h->llr1; 
-  __m128i *wPtr     = (__m128i*) h->w; 
-  __m128i *syst1Ptr = (__m128i*) h->syst1; 
-  
-  for (int i = 0; i < long_cb; i++) {    
-    __m128i llr1 = _mm_load_si128(&llr1Ptr[inter[i]]);
-    __m128i w    = _mm_load_si128(&wPtr[inter[i]]);
-    _mm_store_si128(syst1Ptr++, _mm_sub_epi16(llr1, w));
-  }
-}
-
-void sse_inter_update_w(srslte_tdec_simd_inter_t *h, uint16_t *deinter, uint32_t long_cb) 
-{
-  __m128i *llr1Ptr  = (__m128i*) h->llr1; 
-  __m128i *llr2Ptr  = (__m128i*) h->llr2; 
-  __m128i *wPtr     = (__m128i*) h->w; 
-  __m128i *syst1Ptr = (__m128i*) h->syst1; 
-  
-  for (int i = 0; i < long_cb; i++) {    
-    __m128i llr1 = _mm_load_si128(llr1Ptr++);
-    __m128i w    = _mm_load_si128(wPtr++);
-    __m128i llr2 = _mm_load_si128(&llr2Ptr[deinter[i]]);
-    
-    _mm_store_si128(syst1Ptr++, _mm_add_epi16(w, _mm_sub_epi16(llr2, llr1)));
-  }
-}
-
-/* Computes beta values */
-void map_sse_inter_beta(srslte_tdec_simd_inter_t * s, int16_t *input, int16_t *parity, int16_t * output, uint32_t long_cb)
-{
-  __m128i m_b[8], new[8], old[8], max1[8], max0[8];
-  __m128i x, y, xy;
-  __m128i m1, m0;
-  uint32_t end = long_cb + 3;
-  uint32_t i;
-
-  __m128i *inputPtr  = (__m128i*) input;
-  __m128i *parityPtr = (__m128i*) parity;
-  __m128i *outputPtr = (__m128i*) output;
-  __m128i *alphaPtr  = (__m128i*) s->alpha;
-
-  for (int i = 0; i < 8; i++) {
-    old[i] = _mm_set1_epi16(0);
-  }
-
-  for (int k = end - 1; k >= 0; k--) {
-    x = _mm_load_si128(inputPtr++);
-    y = _mm_load_si128(parityPtr++);
-
-    xy = _mm_add_epi16(x,y);
-
-    m_b[0] = _mm_add_epi16(old[4], xy);
-    m_b[1] = old[4];
-    m_b[2] = _mm_add_epi16(old[5], y);
-    m_b[3] = _mm_add_epi16(old[5], x);
-    m_b[4] = _mm_add_epi16(old[6], x);
-    m_b[5] = _mm_add_epi16(old[6], y);
-    m_b[6] = old[7];
-    m_b[7] = _mm_add_epi16(old[7], xy);
-
-    new[0] = old[0];
-    new[1] = _mm_add_epi16(old[0], xy);
-    new[2] = _mm_add_epi16(old[1], x);
-    new[3] = _mm_add_epi16(old[1], y);
-    new[4] = _mm_add_epi16(old[2], y);
-    new[5] = _mm_add_epi16(old[2], x);
-    new[6] = _mm_add_epi16(old[3], xy);
-    new[7] = old[3];
-
-    for (i = 0; i < 8; i++) {
-      __m128i alpha = _mm_load_si128(alphaPtr++);
-      max0[i] = _mm_add_epi16(alpha, m_b[i]);
-      max1[i] = _mm_add_epi16(alpha, new[i]);
-    }
-
-    m1 = _mm_max_epi16(max1[0], max1[1]);
-    m0 = _mm_max_epi16(max0[0], max0[1]);
-
-    for (i = 2; i < 8; i++) {
-      m1 = _mm_max_epi16(m1, max1[i]);
-      m0 = _mm_max_epi16(m0, max0[i]);      
-    }
-
-    for (i = 0; i < 8; i++) {
-      new[i] = _mm_max_epi16(m_b[i], new[i]);
-      old[i] = new[i];
-    }
-
-    __m128i out = _mm_sub_epi16(m1, m0);
-    _mm_store_si128(outputPtr++, out);
-
-    // normalize 
-    if ((k%4)==0) {
-      for (int i=1;i<8;i++) {
-        _mm_sub_epi16(old[i], old[0]);
-      }
-    }
-  }
-}
-
-/* Computes alpha metrics */
-void map_see_inter_alpha(srslte_tdec_simd_inter_t * s, int16_t *input, int16_t *parity, uint32_t long_cb)
-{
-  __m128i m_b[8], new[8], old[8];
-  __m128i x, y, xy;
-  uint32_t k;
-  
-  __m128i *inputPtr  = (__m128i*) input;
-  __m128i *parityPtr = (__m128i*) parity;
-  __m128i *alphaPtr  = (__m128i*) s->alpha;
-  
-  old[0] = _mm_set1_epi16(0);
-  for (int i = 1; i < 8; i++) {
-    old[i] = _mm_set1_epi16(-INF);
-  }
-
-  for (k = 0; k < long_cb; k++) {
-    x = _mm_load_si128(inputPtr++);
-    y = _mm_load_si128(parityPtr++);
-
-    xy = _mm_add_epi16(x,y);
-
-    m_b[0] = old[0];
-    m_b[1] = _mm_add_epi16(old[3], y);
-    m_b[2] = _mm_add_epi16(old[4], y);
-    m_b[3] = old[7];
-    m_b[4] = old[1];
-    m_b[5] = _mm_add_epi16(old[2], y);
-    m_b[6] = _mm_add_epi16(old[5], y);
-    m_b[7] = old[6];
-
-    new[0] = _mm_add_epi16(old[1], xy);
-    new[1] = _mm_add_epi16(old[2], x);
-    new[2] = _mm_add_epi16(old[5], x);
-    new[3] = _mm_add_epi16(old[6], xy);
-    new[4] = _mm_add_epi16(old[0], xy);
-    new[5] = _mm_add_epi16(old[3], x);
-    new[6] = _mm_add_epi16(old[4], x);
-    new[7] = _mm_add_epi16(old[7], xy);    
-    
-    for (int i = 0; i < 8; i++) {
-      new[i] = _mm_max_epi16(m_b[i], new[i]);
-      old[i] = new[i];
-      _mm_store_si128(alphaPtr++, old[i]);
-    }
-
-    // normalize 
-    if ((k%4)==0) {
-      for (int i=1;i<8;i++) {
-        _mm_sub_epi16(old[i], old[0]);
-      }
-    }
-  }
-}
-
-#endif

lib/src/phy/modem/demod_soft.c

@@ -32,18 +32,26 @@
 #include "srslte/phy/utils/bit.h"
 #include "srslte/phy/modem/demod_soft.h"
 
-// AVX implementation not useful for integers. Wait for AVX2
 
 #ifdef LV_HAVE_SSE
 #include <smmintrin.h>
 void demod_16qam_lte_s_sse(const cf_t *symbols, short *llr, int nsymbols);
 #endif
 
-
 #define SCALE_SHORT_CONV_QPSK  100
 #define SCALE_SHORT_CONV_QAM16 400
 #define SCALE_SHORT_CONV_QAM64 700
 
+#define SCALE_BYTE_CONV_QPSK  20
+#define SCALE_BYTE_CONV_QAM16 30
+#define SCALE_BYTE_CONV_QAM64 40
+
+void demod_bpsk_lte_b(const cf_t *symbols, int8_t *llr, int nsymbols) {
+  for (int i=0;i<nsymbols;i++) {
+    llr[i] = (int8_t) -SCALE_BYTE_CONV_QPSK*(crealf(symbols[i]) + cimagf(symbols[i]))/sqrt(2);
+  }
+}
+
 void demod_bpsk_lte_s(const cf_t *symbols, short *llr, int nsymbols) {
   for (int i=0;i<nsymbols;i++) {
     llr[i] = (short) -SCALE_SHORT_CONV_QPSK*(crealf(symbols[i]) + cimagf(symbols[i]))/sqrt(2);
@@ -56,6 +64,10 @@ void demod_bpsk_lte(const cf_t *symbols, float *llr, int nsymbols) {
   }
 }
 
+void demod_qpsk_lte_b(const cf_t *symbols, int8_t *llr, int nsymbols) {
+  srslte_vec_convert_fb((const float*) symbols, -SCALE_BYTE_CONV_QPSK*sqrt(2), llr, nsymbols*2);
+}
+
 void demod_qpsk_lte_s(const cf_t *symbols, short *llr, int nsymbols) {
   srslte_vec_convert_fi((const float*) symbols, -SCALE_SHORT_CONV_QPSK*sqrt(2), llr, nsymbols*2);
 }
@@ -87,9 +99,11 @@ void demod_16qam_lte_s_sse(const cf_t *symbols, short *llr, int nsymbols) {
   __m128i result11, result12, result22, result21; 
   __m128 scale_v = _mm_set1_ps(-SCALE_SHORT_CONV_QAM16);
   __m128i shuffle_negated_1 = _mm_set_epi8(0xff,0xff,0xff,0xff,7,6,5,4,0xff,0xff,0xff,0xff,3,2,1,0);
-  __m128i shuffle_negated_2 = _mm_set_epi8(0xff,0xff,0xff,0xff,15,14,13,12,0xff,0xff,0xff,0xff,11,10,9,8);
   __m128i shuffle_abs_1 = _mm_set_epi8(7,6,5,4,0xff,0xff,0xff,0xff,3,2,1,0,0xff,0xff,0xff,0xff);
+
+  __m128i shuffle_negated_2 = _mm_set_epi8(0xff,0xff,0xff,0xff,15,14,13,12,0xff,0xff,0xff,0xff,11,10,9,8);
   __m128i shuffle_abs_2 = _mm_set_epi8(15,14,13,12,0xff,0xff,0xff,0xff,11,10,9,8,0xff,0xff,0xff,0xff);
+
   for (int i=0;i<nsymbols/4;i++) {
     symbol1   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
     symbol2   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
@@ -120,6 +134,60 @@ void demod_16qam_lte_s_sse(const cf_t *symbols, short *llr, int nsymbols) {
     llr[4*i+3] = abs(yim)-2*SCALE_SHORT_CONV_QAM16/sqrt(10);    
   }
 }
+
+void demod_16qam_lte_b_sse(const cf_t *symbols, int8_t *llr, int nsymbols) {
+  float *symbolsPtr = (float*) symbols;
+  __m128i *resultPtr = (__m128i*) llr;
+  __m128 symbol1, symbol2, symbol3, symbol4;
+  __m128i symbol_i1, symbol_i2, symbol_i3, symbol_i4, symbol_i, symbol_abs, symbol_12, symbol_34;
+  __m128i offset = _mm_set1_epi8(2*SCALE_BYTE_CONV_QAM16/sqrt(10));
+  __m128i result1n, result1a, result2n, result2a;
+  __m128 scale_v = _mm_set1_ps(-SCALE_BYTE_CONV_QAM16);
+
+  __m128i shuffle_negated_1 = _mm_set_epi8(0xff,0xff,7,6,0xff,0xff,5,4,0xff,0xff,3,2,0xff,0xff,1,0);
+  __m128i shuffle_abs_1     = _mm_set_epi8(7,6,0xff,0xff,5,4,0xff,0xff,3,2,0xff,0xff,1,0,0xff,0xff);
+
+  __m128i shuffle_negated_2 = _mm_set_epi8(0xff,0xff,15,14,0xff,0xff,13,12,0xff,0xff,11,10,0xff,0xff,9,8);
+  __m128i shuffle_abs_2     = _mm_set_epi8(15,14,0xff,0xff,13,12,0xff,0xff,11,10,0xff,0xff,9,8,0xff,0xff);
+
+  for (int i=0;i<nsymbols/8;i++) {
+    symbol1   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
+    symbol2   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
+    symbol3   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
+    symbol4   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
+    symbol_i1 = _mm_cvtps_epi32(_mm_mul_ps(symbol1, scale_v));
+    symbol_i2 = _mm_cvtps_epi32(_mm_mul_ps(symbol2, scale_v));
+    symbol_i3 = _mm_cvtps_epi32(_mm_mul_ps(symbol3, scale_v));
+    symbol_i4 = _mm_cvtps_epi32(_mm_mul_ps(symbol4, scale_v));
+    symbol_12  = _mm_packs_epi32(symbol_i1, symbol_i2);
+    symbol_34  = _mm_packs_epi32(symbol_i3, symbol_i4);
+    symbol_i   = _mm_packs_epi16(symbol_12, symbol_34);
+
+    symbol_abs  = _mm_abs_epi8(symbol_i);
+    symbol_abs  = _mm_sub_epi8(symbol_abs, offset);
+
+    result1n = _mm_shuffle_epi8(symbol_i, shuffle_negated_1);
+    result1a = _mm_shuffle_epi8(symbol_abs, shuffle_abs_1);
+
+    result2n = _mm_shuffle_epi8(symbol_i, shuffle_negated_2);
+    result2a = _mm_shuffle_epi8(symbol_abs, shuffle_abs_2);
+
+    _mm_store_si128(resultPtr, _mm_or_si128(result1n, result1a)); resultPtr++;
+    _mm_store_si128(resultPtr, _mm_or_si128(result2n, result2a)); resultPtr++;
+
+  }
+  // Demodulate last symbols
+  for (int i=8*(nsymbols/8);i<nsymbols;i++) {
+    short yre = (int8_t) (SCALE_BYTE_CONV_QAM16*crealf(symbols[i]));
+    short yim = (int8_t) (SCALE_BYTE_CONV_QAM16*cimagf(symbols[i]));
+
+    llr[4*i+0] = -yre;
+    llr[4*i+1] = -yim;
+    llr[4*i+2] = abs(yre)-2*SCALE_BYTE_CONV_QAM16/sqrt(10);
+    llr[4*i+3] = abs(yim)-2*SCALE_BYTE_CONV_QAM16/sqrt(10);
+  }
+}
+
 #endif
 
 void demod_16qam_lte_s(const cf_t *symbols, short *llr, int nsymbols) {
@@ -138,6 +206,22 @@ void demod_16qam_lte_s(const cf_t *symbols, short *llr, int nsymbols) {
 #endif
 }
 
+void demod_16qam_lte_b(const cf_t *symbols, int8_t *llr, int nsymbols) {
+#ifdef LV_HAVE_SSE
+  demod_16qam_lte_b_sse(symbols, llr, nsymbols);
+#else
+  for (int i=0;i<nsymbols;i++) {
+    int8_t yre = (int8_t) (SCALE_BYTE_CONV_QAM16*crealf(symbols[i]));
+    int8_t yim = (int8_t) (SCALE_BYTE_CONV_QAM16*cimagf(symbols[i]));
+
+    llr[4*i+0] = -yre;
+    llr[4*i+1] = -yim;
+    llr[4*i+2] = abs(yre)-2*SCALE_BYTE_CONV_QAM16/sqrt(10);
+    llr[4*i+3] = abs(yim)-2*SCALE_BYTE_CONV_QAM16/sqrt(10);
+  }
+#endif
+}
+
 void demod_64qam_lte(const cf_t *symbols, float *llr, int nsymbols) 
 {
   for (int i=0;i<nsymbols;i++) {
@@ -219,6 +303,76 @@ void demod_64qam_lte_s_sse(const cf_t *symbols, short *llr, int nsymbols)
   }
 }
 
+void demod_64qam_lte_b_sse(const cf_t *symbols, int8_t *llr, int nsymbols)
+{
+  float *symbolsPtr = (float*) symbols;
+  __m128i *resultPtr = (__m128i*) llr;
+  __m128 symbol1, symbol2, symbol3, symbol4;
+  __m128i symbol_i1, symbol_i2, symbol_i3, symbol_i4, symbol_i, symbol_abs, symbol_abs2,symbol_12, symbol_34;
+  __m128i offset1 = _mm_set1_epi8(4*SCALE_BYTE_CONV_QAM64/sqrt(42));
+  __m128i offset2 = _mm_set1_epi8(2*SCALE_BYTE_CONV_QAM64/sqrt(42));
+  __m128 scale_v = _mm_set1_ps(-SCALE_BYTE_CONV_QAM64);
+  __m128i result11, result12, result13, result22, result21,result23, result31, result32, result33;
+
+  __m128i shuffle_negated_1 = _mm_set_epi8(0xff,0xff,5,4,0xff,0xff,0xff,0xff,3,2,0xff,0xff,0xff,0xff,1,0);
+  __m128i shuffle_negated_2 = _mm_set_epi8(11,10,0xff,0xff,0xff,0xff,9,8,0xff,0xff,0xff,0xff,7,6,0xff,0xff);
+  __m128i shuffle_negated_3 = _mm_set_epi8(0xff,0xff,0xff,0xff,15,14,0xff,0xff,0xff,0xff,13,12,0xff,0xff,0xff,0xff);
+
+  __m128i shuffle_abs_1 = _mm_set_epi8(5,4,0xff,0xff,0xff,0xff,3,2,0xff,0xff,0xff,0xff,1,0,0xff,0xff);
+  __m128i shuffle_abs_2 = _mm_set_epi8(0xff,0xff,0xff,0xff,9,8,0xff,0xff,0xff,0xff,7,6,0xff,0xff,0xff,0xff);
+  __m128i shuffle_abs_3 = _mm_set_epi8(0xff,0xff,15,14,0xff,0xff,0xff,0xff,13,12,0xff,0xff,0xff,0xff,11,10);
+
+  __m128i shuffle_abs2_1 = _mm_set_epi8(0xff,0xff,0xff,0xff,3,2,0xff,0xff,0xff,0xff,1,0,0xff,0xff,0xff,0xff);
+  __m128i shuffle_abs2_2 = _mm_set_epi8(0xff,0xff,9,8,0xff,0xff,0xff,0xff,7,6,0xff,0xff,0xff,0xff,5,4);
+  __m128i shuffle_abs2_3 = _mm_set_epi8(15,14,0xff,0xff,0xff,0xff,13,12,0xff,0xff,0xff,0xff,11,10,0xff,0xff);
+
+  for (int i=0;i<nsymbols/8;i++) {
+    symbol1   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
+    symbol2   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
+    symbol3   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
+    symbol4   = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
+    symbol_i1 = _mm_cvtps_epi32(_mm_mul_ps(symbol1, scale_v));
+    symbol_i2 = _mm_cvtps_epi32(_mm_mul_ps(symbol2, scale_v));
+    symbol_i3 = _mm_cvtps_epi32(_mm_mul_ps(symbol3, scale_v));
+    symbol_i4 = _mm_cvtps_epi32(_mm_mul_ps(symbol4, scale_v));
+    symbol_12  = _mm_packs_epi32(symbol_i1, symbol_i2);
+    symbol_34  = _mm_packs_epi32(symbol_i3, symbol_i4);
+    symbol_i   = _mm_packs_epi16(symbol_12, symbol_34);
+
+    symbol_abs  = _mm_abs_epi8(symbol_i);
+    symbol_abs  = _mm_sub_epi8(symbol_abs, offset1);
+    symbol_abs2 = _mm_sub_epi8(_mm_abs_epi8(symbol_abs), offset2);
+
+    result11 = _mm_shuffle_epi8(symbol_i, shuffle_negated_1);
+    result12 = _mm_shuffle_epi8(symbol_abs, shuffle_abs_1);
+    result13 = _mm_shuffle_epi8(symbol_abs2, shuffle_abs2_1);
+
+    result21 = _mm_shuffle_epi8(symbol_i, shuffle_negated_2);
+    result22 = _mm_shuffle_epi8(symbol_abs, shuffle_abs_2);
+    result23 = _mm_shuffle_epi8(symbol_abs2, shuffle_abs2_2);
+
+    result31 = _mm_shuffle_epi8(symbol_i, shuffle_negated_3);
+    result32 = _mm_shuffle_epi8(symbol_abs, shuffle_abs_3);
+    result33 = _mm_shuffle_epi8(symbol_abs2, shuffle_abs2_3);
+
+    _mm_store_si128(resultPtr, _mm_or_si128(_mm_or_si128(result11, result12),result13)); resultPtr++;
+    _mm_store_si128(resultPtr, _mm_or_si128(_mm_or_si128(result21, result22),result23)); resultPtr++;
+    _mm_store_si128(resultPtr, _mm_or_si128(_mm_or_si128(result31, result32),result33)); resultPtr++;
+
+  }
+  for (int i=8*(nsymbols/8);i<nsymbols;i++) {
+    float yre = (int8_t) (SCALE_BYTE_CONV_QAM64*crealf(symbols[i]));
+    float yim = (int8_t) (SCALE_BYTE_CONV_QAM64*cimagf(symbols[i]));
+
+    llr[6*i+0] = -yre;
+    llr[6*i+1] = -yim;
+    llr[6*i+2] = abs(yre)-4*SCALE_BYTE_CONV_QAM64/sqrt(42);
+    llr[6*i+3] = abs(yim)-4*SCALE_BYTE_CONV_QAM64/sqrt(42);
+    llr[6*i+4] = abs(llr[6*i+2])-2*SCALE_BYTE_CONV_QAM64/sqrt(42);
+    llr[6*i+5] = abs(llr[6*i+3])-2*SCALE_BYTE_CONV_QAM64/sqrt(42);
+  }
+}
+  
 #endif
 
 void demod_64qam_lte_s(const cf_t *symbols, short *llr, int nsymbols) 
@@ -240,6 +394,25 @@ void demod_64qam_lte_s(const cf_t *symbols, short *llr, int nsymbols)
 #endif
 }
 
+void demod_64qam_lte_b(const cf_t *symbols, int8_t *llr, int nsymbols)
+{
+#ifdef LV_HAVE_SSE
+  demod_64qam_lte_b_sse(symbols, llr, nsymbols);
+#else
+  for (int i=0;i<nsymbols;i++) {
+    float yre = (int8_t) (SCALE_BYTE_CONV_QAM64*crealf(symbols[i]));
+    float yim = (int8_t) (SCALE_BYTE_CONV_QAM64*cimagf(symbols[i]));
+
+    llr[6*i+0] = -yre;
+    llr[6*i+1] = -yim;
+    llr[6*i+2] = abs(yre)-4*SCALE_BYTE_CONV_QAM64/sqrt(42);
+    llr[6*i+3] = abs(yim)-4*SCALE_BYTE_CONV_QAM64/sqrt(42);
+    llr[6*i+4] = abs(llr[6*i+2])-2*SCALE_BYTE_CONV_QAM64/sqrt(42);
+    llr[6*i+5] = abs(llr[6*i+3])-2*SCALE_BYTE_CONV_QAM64/sqrt(42);
+  }
+#endif
+}
+
 int srslte_demod_soft_demodulate(srslte_mod_t modulation, const cf_t* symbols, float* llr, int nsymbols) {
   switch(modulation) {
     case SRSLTE_MOD_BPSK:
@@ -281,3 +454,24 @@ int srslte_demod_soft_demodulate_s(srslte_mod_t modulation, const cf_t* symbols,
   } 
   return 0; 
 }
+
+int srslte_demod_soft_demodulate_b(srslte_mod_t modulation, const cf_t* symbols, int8_t* llr, int nsymbols) {
+  switch(modulation) {
+    case SRSLTE_MOD_BPSK:
+      demod_bpsk_lte_b(symbols, llr, nsymbols);
+      break;
+    case SRSLTE_MOD_QPSK:
+      demod_qpsk_lte_b(symbols, llr, nsymbols);
+      break;
+    case SRSLTE_MOD_16QAM:
+      demod_16qam_lte_b(symbols, llr, nsymbols);
+      break;
+    case SRSLTE_MOD_64QAM:
+      demod_64qam_lte_b(symbols, llr, nsymbols);
+      break;
+    default:
+      fprintf(stderr, "Invalid modulation %d\n", modulation);
+      return -1;
+  }
+  return 0;
+}

lib/src/phy/modem/test/soft_demod_test.c

@@ -112,6 +112,7 @@ int main(int argc, char **argv) {
   cf_t *symbols;
   float *llr;
   short *llr_s;
+  int8_t *llr_b;
 
   parse_args(argc, argv);
 
@@ -153,6 +154,12 @@ int main(int argc, char **argv) {
     exit(-1);
   }
 
+  llr_b = srslte_vec_malloc(sizeof(int8_t) * num_bits);
+  if (!llr_b) {
+    perror("malloc");
+    exit(-1);
+  }
+
   /* generate random data */
   srand(0);
   
@@ -160,6 +167,7 @@ int main(int argc, char **argv) {
   struct timeval t[3]; 
   float mean_texec = 0.0; 
   float mean_texec_s = 0.0;
+  float mean_texec_b = 0.0;
   for (int n=0;n<nof_frames;n++) {
     for (i=0;i<num_bits;i++) {
       input[i] = rand()%2;
@@ -187,6 +195,15 @@ int main(int argc, char **argv) {
       mean_texec_s = SRSLTE_VEC_CMA((float) t[0].tv_usec, mean_texec_s, n-1);      
     }
 
+    gettimeofday(&t[1], NULL);
+    srslte_demod_soft_demodulate_b(modulation, symbols, llr_b, num_bits / mod.nbits_x_symbol);
+    gettimeofday(&t[2], NULL);
+    get_time_interval(t);
+
+    if (n > 0) {
+      mean_texec_b = SRSLTE_VEC_CMA((float) t[0].tv_usec, mean_texec_b, n-1);
+    }
+
     if (SRSLTE_VERBOSE_ISDEBUG()) {
       printf("bits=");
       srslte_vec_fprint_b(stdout, input, num_bits);
@@ -200,6 +217,9 @@ int main(int argc, char **argv) {
       printf("llr_s=");
       srslte_vec_fprint_s(stdout, llr_s, num_bits);
 
+      printf("llr_b=");
+      srslte_vec_fprint_bs(stdout, llr_b, num_bits);
+
     }
 
     // Check demodulation errors
@@ -213,6 +233,8 @@ int main(int argc, char **argv) {
   ret = 0; 
 
 clean_exit:
+  free(llr_b);
+  free(llr_s);
   free(llr);
   free(symbols);
   free(output);
@@ -220,7 +242,7 @@ clean_exit:
 
   srslte_modem_table_free(&mod);
 
-  printf("Mean Throughput: %.2f/%.2f. Mbps ExTime: %.2f/%.2f us\n", 
+  printf("Mean Throughput: %.2f/%.2f/%.2f. Mbps ExTime: %.2f/%.2f/%.2f us\n",
-         num_bits/mean_texec, num_bits/mean_texec_s, mean_texec, mean_texec_s);    
+         num_bits/mean_texec, num_bits/mean_texec_s, num_bits/mean_texec_b, mean_texec, mean_texec_s, mean_texec_b);
   exit(ret);
 }

lib/src/phy/phch/pdsch.c

@@ -669,6 +669,50 @@ static int srslte_pdsch_codeword_encode(srslte_pdsch_t *q, srslte_pdsch_cfg_t *c
   return SRSLTE_SUCCESS;
 }
 
+static void csi_correction(srslte_pdsch_t *q, srslte_pdsch_cfg_t *cfg, uint32_t codeword_idx, uint32_t tb_idx, void *e)
+{
+
+  srslte_ra_nbits_t *nbits = &cfg->nbits[tb_idx];
+  uint32_t qm = 0;
+  switch(cfg->grant.mcs[tb_idx].mod) {
+
+    case SRSLTE_MOD_BPSK:
+      qm = 1;
+      break;
+    case SRSLTE_MOD_QPSK:
+      qm = 2;
+      break;
+    case SRSLTE_MOD_16QAM:
+      qm = 4;
+      break;
+    case SRSLTE_MOD_64QAM:
+      qm = 6;
+      break;
+    default:
+      ERROR("No modulation");
+  }
+
+  const uint32_t csi_max_idx = srslte_vec_max_fi(q->csi[codeword_idx], nbits->nof_bits / qm);
+  float csi_max = 1.0f;
+  if (csi_max_idx < nbits->nof_bits / qm) {
+    csi_max = q->csi[codeword_idx][csi_max_idx];
+  }
+  int8_t *e_b   = e;
+  int16_t  *e_s = e;
+  for (int i = 0; i < nbits->nof_bits / qm; i++) {
+    const float csi = q->csi[codeword_idx][i] / csi_max;
+    if (q->llr_is_8bit) {
+      for (int k = 0; k < qm; k++) {
+        e_b[qm * i + k] = (int8_t) ((float) e_b[qm * i + k] * csi);
+      }
+    } else {
+      for (int k = 0; k < qm; k++) {
+        e_s[qm * i + k] = (int16_t) ((float) e_s[qm * i + k] * csi);
+      }
+    }
+  }
+}
+
 static int srslte_pdsch_codeword_decode(srslte_pdsch_t *q, srslte_pdsch_cfg_t *cfg, srslte_sch_t *dl_sch,
                                         srslte_softbuffer_rx_t *softbuffer, uint16_t rnti, uint8_t *data,
                                         uint32_t codeword_idx, uint32_t tb_idx, bool *ack) {
@@ -686,47 +730,24 @@ static int srslte_pdsch_codeword_decode(srslte_pdsch_t *q, srslte_pdsch_cfg_t *c
      * The MAX-log-MAP algorithm used in turbo decoding is unsensitive to SNR estimation,
      * thus we don't need tot set it in the LLRs normalization
      */
-    srslte_demod_soft_demodulate_s(mcs->mod, q->d[codeword_idx], q->e[codeword_idx], nbits->nof_re);
+    if (q->llr_is_8bit) {
+      srslte_demod_soft_demodulate_b(mcs->mod, q->d[codeword_idx], q->e[codeword_idx], nbits->nof_re);
+    } else {
+      srslte_demod_soft_demodulate_s(mcs->mod, q->d[codeword_idx], q->e[codeword_idx], nbits->nof_re);
+    }
 
     /* Select scrambling sequence */
     srslte_sequence_t *seq = get_user_sequence(q, rnti, codeword_idx, cfg->sf_idx, nbits->nof_bits);
 
     /* Bit scrambling */
-    srslte_scrambling_s_offset(seq, q->e[codeword_idx], 0, nbits->nof_bits);
+    if (q->llr_is_8bit) {
-
+      srslte_scrambling_sb_offset(seq, q->e[codeword_idx], 0, nbits->nof_bits);
-    uint32_t qm = 0;
+    } else {
-    switch(cfg->grant.mcs[tb_idx].mod) {
+      srslte_scrambling_s_offset(seq, q->e[codeword_idx], 0, nbits->nof_bits);
-
-      case SRSLTE_MOD_BPSK:
-        qm = 1;
-        break;
-      case SRSLTE_MOD_QPSK:
-        qm = 2;
-        break;
-      case SRSLTE_MOD_16QAM:
-        qm = 4;
-        break;
-      case SRSLTE_MOD_64QAM:
-        qm = 6;
-        break;
-      default:
-        ERROR("No modulation");
     }
 
-    int16_t *e = q->e[codeword_idx];
-
     if (q->csi_enabled) {
-      const uint32_t csi_max_idx = srslte_vec_max_fi(q->csi[codeword_idx], nbits->nof_bits / qm);
+      csi_correction(q, cfg, codeword_idx, tb_idx, q->e[codeword_idx]);
-      float csi_max = 1.0f;
-      if (csi_max_idx < nbits->nof_bits / qm) {
-        csi_max = q->csi[codeword_idx][csi_max_idx];
-      }
-      for (int i = 0; i < nbits->nof_bits / qm; i++) {
-        const float csi = q->csi[codeword_idx][i] / csi_max;
-        for (int k = 0; k < qm; k++) {
-          e[qm * i + k] = (int16_t) ((float) e[qm * i + k] * csi);
-        }
-      }
     }
 
     /* Return  */

lib/src/phy/phch/pusch.c

@@ -603,7 +603,11 @@ int srslte_pusch_decode(srslte_pusch_t *q,
     srslte_dft_precoding(&q->dft_precoding, q->z, q->d, cfg->grant.L_prb, cfg->nbits.nof_symb);
     
     // Soft demodulation
-    srslte_demod_soft_demodulate_s(cfg->grant.mcs.mod, q->d, q->q, cfg->nbits.nof_re);
+    if (q->llr_is_8bit) {
+      srslte_demod_soft_demodulate_b(cfg->grant.mcs.mod, q->d, q->q, cfg->nbits.nof_re);
+    } else {
+      srslte_demod_soft_demodulate_s(cfg->grant.mcs.mod, q->d, q->q, cfg->nbits.nof_re);
+    }
 
     // Generate scrambling sequence if not pre-generated
     srslte_sequence_t *seq = get_user_sequence(q, rnti, cfg->sf_idx, cfg->nbits.nof_bits);
@@ -632,7 +636,11 @@ int srslte_pusch_decode(srslte_pusch_t *q,
     }
     
     // Descrambling
-    srslte_scrambling_s_offset(seq, q->q, 0, cfg->nbits.nof_bits);
+    if (q->llr_is_8bit) {
+      srslte_scrambling_sb_offset(seq, q->q, 0, cfg->nbits.nof_bits);
+    } else {
+      srslte_scrambling_s_offset(seq, q->q, 0, cfg->nbits.nof_bits);
+    }
 
     // Decode
     ret = srslte_ulsch_uci_decode(&q->ul_sch, cfg, softbuffer, q->q, q->g, data, uci_data);

lib/src/phy/phch/sch.c

@@ -32,12 +32,18 @@
 #include <stdbool.h>
 #include <assert.h>
 #include <math.h>
+#include <srslte/phy/phch/sch.h>
 #include "srslte/phy/phch/pdsch.h"
 #include "srslte/phy/utils/bit.h"
 #include "srslte/phy/utils/debug.h"
 #include "srslte/phy/utils/vector.h"
 
-#define SRSLTE_PDSCH_MAX_TDEC_ITERS         4
+#define SRSLTE_PDSCH_MAX_TDEC_ITERS         10
+
+#ifdef LV_HAVE_SSE
+#include <immintrin.h>
+#endif /* LV_HAVE_SSE */
+
 
 /* 36.213 Table 8.6.3-1: Mapping of HARQ-ACK offset values and the index signalled by higher layers */
 float beta_harq_offset[16] = {2.0, 2.5, 3.125, 4.0, 5.0, 6.250, 8.0, 10.0, 
@@ -184,8 +190,6 @@ static int encode_tb_off(srslte_sch_t *q,
                      uint32_t Qm, uint32_t rv, uint32_t nof_e_bits,  
                      uint8_t *data, uint8_t *e_bits, uint32_t w_offset) 
 {
-  uint8_t parity[3] = {0, 0, 0};
-  uint32_t par;
   uint32_t i;
   uint32_t cb_len=0, rp=0, wp=0, rlen=0, n_e=0;
   int ret = SRSLTE_ERROR_INVALID_INPUTS; 
@@ -213,16 +217,8 @@ static int encode_tb_off(srslte_sch_t *q,
       gamma = Gp%cb_segm->C;
     }
 
-    if (data) {
+    /* Reset TB CRC */
-
+    srslte_crc_set_init(&q->crc_tb, 0);
-      /* Compute transport block CRC */
-      par = srslte_crc_checksum_byte(&q->crc_tb, data, cb_segm->tbs);
-
-      /* parity bits will be appended later */
-      parity[0] = (par&(0xff<<16))>>16;
-      parity[1] = (par&(0xff<<8))>>8;
-      parity[2] = par&0xff;
-    }
 
     wp = 0;
     rp = 0;
@@ -252,6 +248,7 @@ static int encode_tb_off(srslte_sch_t *q,
           cb_len, rlen, wp, rp, n_e);
 
       if (data) {
+        bool last_cb = false;
 
         /* Copy data to another buffer, making space for the Codeblock CRC */
         if (i < cb_segm->C - 1) {
@@ -263,13 +260,19 @@ static int encode_tb_off(srslte_sch_t *q,
           
           /* Append Transport Block parity bits to the last CB */
           memcpy(q->cb_in, &data[rp/8], (rlen - 24) * sizeof(uint8_t)/8);
-          memcpy(&q->cb_in[(rlen - 24)/8], parity, 3 * sizeof(uint8_t));
+          last_cb = true;
         }
 
         /* Turbo Encoding
          * If Codeblock CRC is required it is given the CRC instance pointer, otherwise CRC pointer shall be NULL
          */
-        srslte_tcod_encode_lut(&q->encoder, (cb_segm->C > 1) ? &q->crc_cb : NULL, q->cb_in, q->parity_bits, cblen_idx);
+        srslte_tcod_encode_lut(&q->encoder,
+                               &q->crc_tb,
+                               (cb_segm->C > 1) ? &q->crc_cb : NULL,
+                               q->cb_in,
+                               q->parity_bits,
+                               cblen_idx,
+                               last_cb);
       }
       DEBUG("RM cblen_idx=%d, n_e=%d, wp=%d, nof_e_bits=%d\n",cblen_idx, n_e, wp, nof_e_bits);
       
@@ -304,88 +307,65 @@ static int encode_tb(srslte_sch_t *q,
 bool decode_tb_cb(srslte_sch_t *q, 
                      srslte_softbuffer_rx_t *softbuffer, srslte_cbsegm_t *cb_segm, 
                      uint32_t Qm, uint32_t rv, uint32_t nof_e_bits, 
-                     int16_t *e_bits, uint8_t *data,
+                     void *e_bits, uint8_t *data)
-                     uint32_t cb_size_group) 
 {
 
-  bool cb_map[SRSLTE_MAX_CODEBLOCKS];
+  int8_t *e_bits_b  = e_bits;
-    
+  int16_t *e_bits_s = e_bits;
-  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;
+  if (cb_segm->C > SRSLTE_MAX_CODEBLOCKS) {
-  uint32_t first_cb   = cb_size_group?cb_segm->C1:0;
-  uint32_t cb_len     = cb_size_group?cb_segm->K2:cb_segm->K1;
-  uint32_t cb_len_idx = cb_size_group?cb_segm->K2_idx:cb_segm->K1_idx;
-
-  uint32_t rlen       = cb_segm->C==1?cb_len:(cb_len-24);
-  uint32_t Gp         = nof_e_bits / Qm;
-  uint32_t gamma      = cb_segm->C>0?Gp%cb_segm->C:Gp;
-  uint32_t n_e        = Qm * (Gp/cb_segm->C);
-  
-  if (nof_cb > SRSLTE_MAX_CODEBLOCKS) {
     fprintf(stderr, "Error SRSLTE_MAX_CODEBLOCKS=%d\n", SRSLTE_MAX_CODEBLOCKS);
     return false;
   }
 
-  for (int i=0;i<srslte_tdec_get_nof_parallel(&q->decoder);i++) {
+  q->nof_iterations = 0;
-    cb_idx[i]        = i+first_cb; 
-    decoder_input[i] = NULL;
-  }
 
-  uint32_t remaining_cb = 0;
+  for (int cb_idx=0;cb_idx<cb_segm->C;cb_idx++)
-  for (int i=0;i<nof_cb;i++) {
+  {
     /* Do not process blocks with CRC Ok */
-    cb_map[i] = softbuffer->cb_crc[i];
+    if (softbuffer->cb_crc[cb_idx] == false) {
-    if (softbuffer->cb_crc[i] == false) {
-      remaining_cb ++;
-    }
-  }
 
-  srslte_tdec_reset(&q->decoder, cb_len);
+      uint32_t cb_len     = cb_idx<cb_segm->C1?cb_segm->K1:cb_segm->K2;
+      uint32_t cb_len_idx = cb_idx<cb_segm->C1?cb_segm->K1_idx:cb_segm->K2_idx;
 
-  q->nof_iterations = 0;
+      uint32_t rlen       = cb_segm->C==1?cb_len:(cb_len-24);
+      uint32_t Gp         = nof_e_bits / Qm;
+      uint32_t gamma      = cb_segm->C>0?Gp%cb_segm->C:Gp;
+      uint32_t n_e        = Qm * (Gp/cb_segm->C);
 
-  while(remaining_cb>0) {
+      uint32_t rp   = cb_idx*n_e;
+      uint32_t n_e2 = n_e;
 
-    // Unratematch the codeblocks left to decode 
+      if (cb_idx > cb_segm->C - gamma) {
-    for (int i=0;i<srslte_tdec_get_nof_parallel(&q->decoder);i++) {
+        n_e2 = n_e+Qm;
+        rp   = (cb_segm->C - gamma)*n_e + (cb_idx-(cb_segm->C - gamma))*n_e2;
+      }
 
-      if (!decoder_input[i] && remaining_cb > 0) {        
+      if (q->llr_is_8bit) {
-        // Find an unprocessed CB 
+        if (srslte_rm_turbo_rx_lut_8bit(&e_bits_b[rp], (int8_t*) softbuffer->buffer_f[cb_idx], n_e2, cb_len_idx, rv)) {
-        cb_idx[i]=first_cb;
+          fprintf(stderr, "Error in rate matching\n");
-        while(cb_idx[i]<first_cb+nof_cb-1 && cb_map[cb_idx[i]]) {
+          return SRSLTE_ERROR;
-          cb_idx[i]++;
         }
-        if (cb_map[cb_idx[i]] == false) {
+      } else {
-          cb_map[cb_idx[i]] = true; 
+        if (srslte_rm_turbo_rx_lut(&e_bits_s[rp], softbuffer->buffer_f[cb_idx], n_e2, cb_len_idx, rv)) {
-          
+          fprintf(stderr, "Error in rate matching\n");
-          uint32_t rp   = cb_idx[i]*n_e;  
+          return SRSLTE_ERROR;
-          uint32_t n_e2 = n_e;
-          
-          if (cb_idx[i] > cb_segm->C - gamma) {
-            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]];
         }
       }
-    }
 
-    // Run 1 iteration for the codeblocks in queue
+      srslte_tdec_new_cb(&q->decoder, cb_len);
-    srslte_tdec_iteration_par(&q->decoder, decoder_input, cb_len);
 
-    // Decide output bits and compute CRC 
+      // Run iterations and use CRC for early stopping
-    for (int i=0;i<srslte_tdec_get_nof_parallel(&q->decoder);i++) {
+      bool early_stop = false;
-      if (decoder_input[i]) {        
+      uint32_t cb_noi = 0;
-        srslte_tdec_decision_byte_par_cb(&q->decoder, q->cb_in, i, cb_len);
+      do {
+        if (q->llr_is_8bit) {
+          srslte_tdec_iteration_8bit(&q->decoder, (int8_t*) softbuffer->buffer_f[cb_idx], &data[cb_idx*rlen/8]);
+        } else {
+          srslte_tdec_iteration(&q->decoder, softbuffer->buffer_f[cb_idx], &data[cb_idx*rlen/8]);
+        }
+        q->nof_iterations++;
+        cb_noi++;
 
         uint32_t len_crc;
         srslte_crc_t *crc_ptr;
@@ -399,47 +379,45 @@ bool decode_tb_cb(srslte_sch_t *q,
         }
 
         // CRC is OK
-        if (!srslte_crc_checksum_byte(crc_ptr, q->cb_in, len_crc)) {
+        if (!srslte_crc_checksum_byte(crc_ptr, &data[cb_idx*rlen/8], len_crc)) {
-
+
-          memcpy(softbuffer->data[cb_idx[i]], q->cb_in, rlen/8 * sizeof(uint8_t));
+          softbuffer->cb_crc[cb_idx] = true;
-          softbuffer->cb_crc[cb_idx[i]] = true;
+          early_stop = true;
-
+
-          q->nof_iterations += srslte_tdec_get_nof_iterations_cb(&q->decoder, i);
+          // CRC is error and exceeded maximum iterations for this CB.
-
+          // Early stop the whole transport block.
-          // Reset number of iterations for that CB in the decoder 
-          srslte_tdec_reset_cb(&q->decoder, i);
-          remaining_cb--;        
-          decoder_input[i] = NULL; 
-          cb_idx[i] = 0; 
-
-        // CRC is error and exceeded maximum iterations for this CB.
-        // Early stop the whole transport block.
-        } else if (srslte_tdec_get_nof_iterations_cb(&q->decoder, i) >= q->max_iterations) {
-          INFO("CB %d: Error. CB is erroneous. remaining_cb=%d, i=%d, first_cb=%d, nof_cb=%d\n", 
-                cb_idx[i], remaining_cb, i, first_cb, nof_cb);
-
-          q->nof_iterations += q->max_iterations;
-          srslte_tdec_reset_cb(&q->decoder, i);
-          remaining_cb--;
-          decoder_input[i] = NULL;
-          cb_idx[i] = 0;
         }
-      }
+
+      } while (cb_noi < q->max_iterations && !early_stop);
+
+      INFO("CB %d: rp=%d, n_e=%d, cb_len=%d, CRC=%s, rlen=%d, iterations=%d/%d\n",
+           cb_idx, rp, n_e2, cb_len, early_stop?"OK":"KO", rlen, cb_noi, q->max_iterations);
+
+    } else {
+      // Copy decoded data from previous transmissions
+      uint32_t cb_len     = cb_idx<cb_segm->C1?cb_segm->K1:cb_segm->K2;
+      uint32_t rlen       = cb_segm->C==1?cb_len:(cb_len-24);
+      memcpy(&data[cb_idx*rlen/8], softbuffer->data[cb_idx], rlen/8 * sizeof(uint8_t));
     }
   }
 
   softbuffer->tb_crc = true;
-  for (int i = 0; i < nof_cb && softbuffer->tb_crc; i++) {
+  for (int i = 0; i < cb_segm->C && softbuffer->tb_crc; i++) {
     /* If one CB failed return false */
     softbuffer->tb_crc = softbuffer->cb_crc[i];
   }
-  if (softbuffer->tb_crc) {
+  // If TB CRC failed, save correct CB for next retransmission
-    for (int i = 0; i < nof_cb; i++) {
+  if (!softbuffer->tb_crc) {
-      memcpy(&data[i * rlen / 8], softbuffer->data[i], rlen/8 * sizeof(uint8_t));
+    for (int i = 0; i < cb_segm->C; i++) {
+      if (softbuffer->cb_crc[i]) {
+        uint32_t cb_len     = i<cb_segm->C1?cb_segm->K1:cb_segm->K2;
+        uint32_t rlen       = cb_segm->C==1?cb_len:(cb_len-24);
+        memcpy(softbuffer->data[i], &data[i * rlen / 8], rlen/8 * sizeof(uint8_t));
+      }
     }
   }
 
-  q->nof_iterations /= nof_cb;
+  q->nof_iterations /= cb_segm->C;
   return softbuffer->tb_crc;
 }
 
@@ -485,16 +463,12 @@ static int decode_tb(srslte_sch_t *q,
         
     bool crc_ok = true; 
 
-    uint32_t nof_cb_groups = cb_segm->C2>0?2:1; 
-    
     data[cb_segm->tbs/8+0] = 0; 
     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_MAX_NPAR
+    // Process Codeblocks
-    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);
-      crc_ok = decode_tb_cb(q, softbuffer, cb_segm, Qm, rv, nof_e_bits, e_bits, data, i);            
-    }
 
     if (crc_ok) {
 
@@ -593,22 +567,286 @@ static void ulsch_interleave_gen(uint32_t H_prime_total, uint32_t N_pusch_symbs,
   }
 }
 
+static void ulsch_interleave_qm2(const uint8_t *g_bits, uint32_t rows, uint32_t cols, uint8_t *q_bits, uint32_t ri_min_row, const uint8_t *ri_present) {
+  uint32_t bit_read_idx = 0;
+
+  for (uint32_t j = 0; j < ri_min_row; j++) {
+    for (uint32_t i = 0; i < cols; i++) {
+      uint32_t k = (i * rows + j) * 2;
+
+      uint32_t read_byte_idx = bit_read_idx / 8;
+      uint32_t read_bit_idx = bit_read_idx % 8;
+      uint32_t write_byte_idx = k / 8;
+      uint32_t write_bit_idx = k % 8;
+      uint8_t w = (g_bits[read_byte_idx] >> (6 - read_bit_idx)) & (uint8_t) 0x03;
+      q_bits[write_byte_idx] |= w << (6 - write_bit_idx);
+
+      bit_read_idx += 2;
+    }
+  }
+
+  for (uint32_t j = ri_min_row; j < rows; j++) {
+    for (uint32_t i = 0; i < cols; i++) {
+      uint32_t k = (i * rows + j) * 2;
+
+      if (ri_present[k]) {
+        /* do nothing */
+      } else {
+        uint32_t read_byte_idx = bit_read_idx / 8;
+        uint32_t read_bit_idx = bit_read_idx % 8;
+        uint32_t write_byte_idx = k / 8;
+        uint32_t write_bit_idx = k % 8;
+        uint8_t w = (g_bits[read_byte_idx] >> (6 - read_bit_idx)) & (uint8_t) 0x03;
+        q_bits[write_byte_idx] |= w << (6 - write_bit_idx);
+
+        bit_read_idx += 2;
+      }
+    }
+  }
+}
+
+static void ulsch_interleave_qm4(uint8_t *g_bits, uint32_t rows, uint32_t cols, uint8_t *q_bits, uint32_t ri_min_row, const uint8_t *ri_present) {
+  uint32_t bit_read_idx = 0;
+  
+  for (uint32_t j = 0; j < ri_min_row; j++) {
+    int32_t i = 0;
+
+#ifndef LV_HAVE_SSE
+    __m128i _counter = _mm_slli_epi32(_mm_add_epi32(_mm_mullo_epi32(_counter0,_rows),_mm_set1_epi32(j)), 2);
+    uint8_t *_g_bits = &g_bits[bit_read_idx/8];
+
+    /* First bits are aligned to byte */
+    if (0 == (bit_read_idx & 0x3)) {
+      for (; i < (cols - 3); i += 4) {
+
+        uint8_t w1 = *(_g_bits++);
+        uint8_t w2 = *(_g_bits++);
+
+        __m128i _write_byte_idx = _mm_srli_epi32(_counter, 3);
+        __m128i _write_bit_idx = _mm_and_si128(_counter, _7);
+        __m128i _write_shift = _mm_sub_epi32(_4, _write_bit_idx);
+
+        q_bits[_mm_extract_epi32(_write_byte_idx, 0)] |= (w1 >> 0x4) << _mm_extract_epi32(_write_shift, 0);
+        q_bits[_mm_extract_epi32(_write_byte_idx, 1)] |= (w1 & 0xf) << _mm_extract_epi32(_write_shift, 1);
+        q_bits[_mm_extract_epi32(_write_byte_idx, 2)] |= (w2 >> 0x4) << _mm_extract_epi32(_write_shift, 2);
+        q_bits[_mm_extract_epi32(_write_byte_idx, 3)] |= (w2 & 0xf) << _mm_extract_epi32(_write_shift, 3);
+        _counter = _mm_add_epi32(_counter, _inc);
+      }
+    } else {
+      for (; i < (cols - 3); i += 4) {
+        __m128i _write_byte_idx = _mm_srli_epi32(_counter, 3);
+        __m128i _write_bit_idx = _mm_and_si128(_counter, _7);
+        __m128i _write_shift = _mm_sub_epi32(_4, _write_bit_idx);
+
+        uint8_t w1 = *(_g_bits);
+        uint8_t w2 = *(_g_bits++);
+        uint8_t w3 = *(_g_bits++);
+        q_bits[_mm_extract_epi32(_write_byte_idx, 0)] |= (w1 & 0xf) << _mm_extract_epi32(_write_shift, 0);
+        q_bits[_mm_extract_epi32(_write_byte_idx, 1)] |= (w2 >> 0x4) << _mm_extract_epi32(_write_shift, 1);
+        q_bits[_mm_extract_epi32(_write_byte_idx, 2)] |= (w2 & 0xf) << _mm_extract_epi32(_write_shift, 2);
+        q_bits[_mm_extract_epi32(_write_byte_idx, 3)] |= (w3 >> 0x4) << _mm_extract_epi32(_write_shift, 3);
+
+        _counter = _mm_add_epi32(_counter, _inc);
+      }
+    }
+    bit_read_idx += i * 4;
+#endif /* LV_HAVE_SSE */
+
+    /* Spare bits */
+    for (; i < cols; i++) {
+      uint32_t k = (i * rows + j) * 4;
+
+      uint32_t read_byte_idx = bit_read_idx / 8;
+      uint32_t read_bit_idx = bit_read_idx % 8;
+      uint32_t write_byte_idx = k / 8;
+      uint32_t write_bit_idx = k % 8;
+      uint8_t w = (g_bits[read_byte_idx] >> (4 - read_bit_idx)) & (uint8_t) 0x0f;
+      q_bits[write_byte_idx] |= w << (4 - write_bit_idx);
+
+      bit_read_idx += 4;
+    }
+  }
+
+  /* Do rows containing RI */
+  for (uint32_t j = ri_min_row; j < rows; j++) {
+    for (uint32_t i = 0; i < cols; i++) {
+      uint32_t k = (i * rows + j) * 4;
+
+      if (ri_present[k]) {
+        /* do nothing */
+      } else {
+        uint32_t read_byte_idx = bit_read_idx / 8;
+        uint32_t read_bit_idx = bit_read_idx % 8;
+        uint32_t write_byte_idx = k / 8;
+        uint32_t write_bit_idx = k % 8;
+        uint8_t w = (g_bits[read_byte_idx] >> (4 - read_bit_idx)) & (uint8_t) 0x0f;
+        q_bits[write_byte_idx] |= w << (4 - write_bit_idx);
+
+        bit_read_idx += 4;
+      }
+    }
+
+  }
+}
+
+static void ulsch_interleave_qm6(const uint8_t *g_bits,
+                                 uint32_t rows,
+                                 uint32_t cols,
+                                 uint8_t *q_bits,
+                                 uint32_t ri_min_row,
+                                 const uint8_t *ri_present) {
+  uint32_t bit_read_idx = 0;
+
+  for (uint32_t j = 0; j < ri_min_row; j++) {
+    for (uint32_t i = 0; i < cols; i++) {
+      uint32_t k = (i * rows + j) * 6;
+
+      uint32_t read_byte_idx = bit_read_idx / 8;
+      uint32_t read_bit_idx = bit_read_idx % 8;
+      uint32_t write_byte_idx = k / 8;
+      uint32_t write_bit_idx = k % 8;
+      uint8_t w;
+
+      switch (read_bit_idx) {
+        case 0:
+          w = g_bits[read_byte_idx] >> 2;
+          break;
+        case 2:
+          w = g_bits[read_byte_idx] & (uint8_t) 0x3f;
+          break;
+        case 4:
+          w = ((g_bits[read_byte_idx] << 2) | (g_bits[read_byte_idx + 1] >> 6)) & (uint8_t) 0x3f;
+          break;
+        case 6:
+          w = ((g_bits[read_byte_idx] << 4) | (g_bits[read_byte_idx + 1] >> 4)) & (uint8_t) 0x3f;
+          break;
+        default:
+          w = 0;
+      }
+
+      switch (write_bit_idx) {
+        case 0:
+          q_bits[write_byte_idx] |= w << 2;
+          break;
+        case 2:
+          q_bits[write_byte_idx] |= w;
+          break;
+        case 4:
+          q_bits[write_byte_idx] |= w >> 2;
+          q_bits[write_byte_idx + 1] |= w << 6;
+          break;
+        case 6:
+          q_bits[write_byte_idx] |= w >> 4;
+          q_bits[write_byte_idx + 1] |= w << 4;
+          break;
+        default:
+          /* Do nothing */;
+      }
+
+      bit_read_idx += 6;
+    }
+  }
+
+  for (uint32_t j = ri_min_row; j < rows; j++) {
+    for (uint32_t i = 0; i < cols; i++) {
+      uint32_t k = (i * rows + j) * 6;
+
+      if (ri_present[k]) {
+        /* do nothing */
+      } else {
+        uint32_t read_byte_idx = bit_read_idx / 8;
+        uint32_t read_bit_idx = bit_read_idx % 8;
+        uint32_t write_byte_idx = k / 8;
+        uint32_t write_bit_idx = k % 8;
+        uint8_t w;
+
+        switch (read_bit_idx) {
+          case 0:
+            w = g_bits[read_byte_idx] >> 2;
+            break;
+          case 2:
+            w = g_bits[read_byte_idx] & (uint8_t) 0x3f;
+            break;
+          case 4:
+            w = ((g_bits[read_byte_idx] << 2) | (g_bits[read_byte_idx + 1] >> 6)) & (uint8_t) 0x3f;
+            break;
+          case 6:
+            w = ((g_bits[read_byte_idx] << 4) | (g_bits[read_byte_idx + 1] >> 4)) & (uint8_t) 0x3f;
+            break;
+          default:
+            w = 0;
+        }
+
+        switch (write_bit_idx) {
+          case 0:
+            q_bits[write_byte_idx] |= w << 2;
+            break;
+          case 2:
+            q_bits[write_byte_idx] |= w;
+            break;
+          case 4:
+            q_bits[write_byte_idx] |= w >> 2;
+            q_bits[write_byte_idx + 1] |= w << 6;
+            break;
+          case 6:
+            q_bits[write_byte_idx] |= w >> 4;
+            q_bits[write_byte_idx + 1] |= w << 4;
+            break;
+          default:
+            /* Do nothing */;
+        }
+
+        bit_read_idx += 6;
+      }
+    }
+  }
+}
+
 /* UL-SCH channel interleaver according to 5.2.2.8 of 36.212 */
 void ulsch_interleave(uint8_t *g_bits, uint32_t Qm, uint32_t H_prime_total, 
                       uint32_t N_pusch_symbs, uint8_t *q_bits, srslte_uci_bit_t *ri_bits, uint32_t nof_ri_bits, 
                       uint8_t *ri_present, uint32_t *inteleaver_lut)
 {
 
+  const uint32_t nof_bits = H_prime_total * Qm;
+  uint32_t rows = H_prime_total / N_pusch_symbs;
+  uint32_t cols = N_pusch_symbs;
+  uint32_t ri_min_row = rows;
+
   // Prepare ri_bits for fast search using temp_buffer
   if (nof_ri_bits > 0) {
     for (uint32_t i=0;i<nof_ri_bits;i++) {
+      uint32_t ri_row = (ri_bits[i].position / Qm) % rows;
+
+      if (ri_row < ri_min_row) {
+        ri_min_row = ri_row;
+      }
+
       ri_present[ri_bits[i].position] = 1;
     }
   }
 
+#if 1
+  bzero(q_bits, nof_bits / 8);
+  switch (Qm) {
+    case 2:
+      ulsch_interleave_qm2(g_bits, rows, cols, q_bits, ri_min_row, ri_present);
+      break;
+    case 4:
+      ulsch_interleave_qm4(g_bits, rows, cols, q_bits, ri_min_row, ri_present);
+      break;
+    case 6:
+      ulsch_interleave_qm6(g_bits, rows, cols, q_bits, ri_min_row, ri_present);
+      break;
+    default:
+      /* This line should never be reached */
+      fprintf(stderr, "Wrong Qm (%d)\n", Qm);
+  }
+#else
   // Genearate interleaver table and interleave bits
   ulsch_interleave_gen(H_prime_total, N_pusch_symbs, Qm, ri_present, inteleaver_lut); 
   srslte_bit_interleave_i(g_bits, q_bits, inteleaver_lut, H_prime_total*Qm);
+#endif
 
   // Reset temp_buffer because will be reused next time
   if (nof_ri_bits > 0) {

lib/src/phy/phch/test/pdsch_test.c

@@ -63,18 +63,23 @@ bool tb_cw_swap = false;
 bool enable_coworker = false;
 uint32_t pmi = 0;
 char *input_file = NULL;
+int M=1;
+
+bool use_8_bit = false;
 
 void usage(char *prog) {
-  printf("Usage: %s [fmMcsrtRFpnwav] \n", prog);
+  printf("Usage: %s [fmMbcsrtRFpnwav] \n", prog);
   printf("\t-f read signal from file [Default generate it with pdsch_encode()]\n");
   printf("\t-m MCS [Default %d]\n", mcs[0]);
   printf("\t-M MCS2 [Default %d]\n", mcs[1]);
   printf("\t-c cell id [Default %d]\n", cell.id);
+  printf("\t-b Use 8-bit LLR [Default 16-bit]\n");
   printf("\t-s subframe [Default %d]\n", subframe);
   printf("\t-r rv_idx [Default %d]\n", rv_idx[0]);
   printf("\t-t rv_idx2 [Default %d]\n", rv_idx[1]);
   printf("\t-R rnti [Default %d]\n", rnti);
   printf("\t-F cfi [Default %d]\n", cfi);
+  printf("\t-X Number of repetitions for time measurement [Default %d]\n", M);
   printf("\t-x Transmission mode [single|diversity|cdd|multiplex] [Default %s]\n", mimo_type_str);
   printf("\t-n cell.nof_prb [Default %d]\n", cell.nof_prb);
   printf("\t-a nof_rx_antennas [Default %d]\n", nof_rx_antennas);
@@ -86,7 +91,7 @@ void usage(char *prog) {
 
 void parse_args(int argc, char **argv) {
   int opt;
-  while ((opt = getopt(argc, argv, "fmMcsrtRFpnawvxj")) != -1) {
+  while ((opt = getopt(argc, argv, "fmMcsbrtRFpnawvXxj")) != -1) {
     switch(opt) {
     case 'f':
       input_file = argv[optind];
@@ -94,12 +99,18 @@ void parse_args(int argc, char **argv) {
     case 'm':
       mcs[0] = (uint32_t) atoi(argv[optind]);
       break;
+    case 'b':
+      use_8_bit = true;
+      break;
     case 'M':
       mcs[1] = (uint32_t) atoi(argv[optind]);
       break;
     case 's':
       subframe = atoi(argv[optind]);
       break;
+    case 'X':
+      M = (uint32_t) atoi(argv[optind]);
+      break;
     case 'r':
       rv_idx[0] = (uint32_t) atoi(argv[optind]);
       break;
@@ -166,7 +177,6 @@ int main(int argc, char **argv) {
   int ret = -1;
   struct timeval t[3];
   srslte_softbuffer_tx_t *softbuffers_tx[SRSLTE_MAX_CODEWORDS];
-  int M=1;
   bool acks[SRSLTE_MAX_CODEWORDS] = {false};
 
   parse_args(argc,argv);
@@ -319,6 +329,9 @@ int main(int argc, char **argv) {
     goto quit;
   }
 
+  pdsch_rx.llr_is_8bit = use_8_bit;
+  pdsch_rx.dl_sch.llr_is_8bit = use_8_bit;
+
   srslte_pdsch_set_rnti(&pdsch_rx, rnti);
 
   for (i = 0; i < SRSLTE_MAX_CODEWORDS; i++) {
@@ -517,8 +530,8 @@ int main(int argc, char **argv) {
       for (int byte = 0; byte < grant.mcs[tb].tbs / 8; byte++) {
         if (data_tx[tb][byte] != data_rx[tb][byte]) {
           ERROR("Found BYTE (%d) error in TB %d (%02X != %02X), quiting...", byte, tb, data_tx[tb][byte], data_rx[tb][byte]);
-          printf("Tx: "); srslte_vec_fprint_byte(stdout, data_tx[tb], grant.mcs[tb].tbs / 8);
+          //printf("Tx: "); srslte_vec_fprint_byte(stdout, data_tx[tb], grant.mcs[tb].tbs / 8);
-          printf("Rx: "); srslte_vec_fprint_byte(stdout, data_rx[tb], grant.mcs[tb].tbs / 8);
+          //printf("Rx: "); srslte_vec_fprint_byte(stdout, data_rx[tb], grant.mcs[tb].tbs / 8);
           ret = SRSLTE_ERROR;
           goto quit;
         }

lib/src/phy/scrambling/scrambling.c

@@ -47,7 +47,12 @@ void srslte_scrambling_s(srslte_sequence_t *s, short *data) {
 
 void srslte_scrambling_s_offset(srslte_sequence_t *s, short *data, int offset, int len) {
   assert (len + offset <= s->cur_len);
-  srslte_vec_prod_sss(data, &s->c_short[offset], data, len);
+  srslte_vec_neg_sss(data, &s->c_short[offset], data, len);
+}
+
+void srslte_scrambling_sb_offset(srslte_sequence_t *s, int8_t *data, int offset, int len) {
+  assert (len + offset <= s->cur_len);
+  srslte_vec_neg_bbb(data, &s->c_char[offset], data, len);
 }
 
 void srslte_scrambling_c(srslte_sequence_t *s, cf_t *data) {

lib/src/phy/scrambling/test/scrambling_test.c

@@ -177,6 +177,83 @@ int main(int argc, char **argv) {
 
     free(input_f);
     free(scrambled_f);
+
+    int16_t *input_s, *scrambled_s;
+
+    // Scramble also shorts
+    input_s= malloc(sizeof(int16_t) * seq.cur_len);
+    if (!input_s) {
+      perror("malloc");
+      exit(-1);
+    }
+    scrambled_s = malloc(sizeof(int16_t) * seq.cur_len);
+    if (!scrambled_s) {
+      perror("malloc");
+      exit(-1);
+    }
+
+    for (i=0;i<seq.cur_len;i++) {
+      input_s[i] = 100*(rand()/RAND_MAX)-50;
+      scrambled_s[i] = input_s[i];
+    }
+
+    gettimeofday(&t[1], NULL);
+    srslte_scrambling_s(&seq, scrambled_s);
+    gettimeofday(&t[2], NULL);
+    srslte_scrambling_s(&seq, scrambled_s);
+
+
+    get_time_interval(t);
+    printf("Texec short=%ld us for %d bits\n", t[0].tv_usec, seq.cur_len);
+
+    for (i=0;i<seq.cur_len;i++) {
+      if (scrambled_s[i] != input_s[i]) {
+        printf("Error in %d\n", i);
+        exit(-1);
+      }
+    }
+
+    free(input_s);
+    free(scrambled_s);
+
+
+    int8_t *input_b, *scrambled_b;
+
+    // Scramble also bytes
+    input_b= malloc(sizeof(int8_t) * seq.cur_len);
+    if (!input_b) {
+      perror("malloc");
+      exit(-1);
+    }
+    scrambled_b = malloc(sizeof(int8_t) * seq.cur_len);
+    if (!scrambled_b) {
+      perror("malloc");
+      exit(-1);
+    }
+
+    for (i=0;i<seq.cur_len;i++) {
+      input_b[i] = 100*(rand()/RAND_MAX)-50;
+      scrambled_b[i] = input_b[i];
+    }
+
+    gettimeofday(&t[1], NULL);
+    srslte_scrambling_sb_offset(&seq, scrambled_b, 0, seq.cur_len);
+    gettimeofday(&t[2], NULL);
+    srslte_scrambling_sb_offset(&seq, scrambled_b, 0, seq.cur_len);
+
+
+    get_time_interval(t);
+    printf("Texec char=%ld us for %d bits\n", t[0].tv_usec, seq.cur_len);
+
+    for (i=0;i<seq.cur_len;i++) {
+      if (scrambled_b[i] != input_b[i]) {
+        printf("Error in %d\n", i);
+        exit(-1);
+      }
+    }
+
+    free(input_b);
+    free(scrambled_b);
   }
   printf("Ok\n");
   srslte_sequence_free(&seq);

lib/src/phy/utils/vector.c

@@ -58,6 +58,10 @@ void srslte_vec_sub_sss(const int16_t *x, const int16_t *y, int16_t *z, const ui
   srslte_vec_sub_sss_simd(x, y, z, len);
 }
 
+void srslte_vec_sub_bbb(const int8_t *x, const int8_t *y, int8_t *z, const uint32_t len) {
+  srslte_vec_sub_bbb_simd(x, y, z, len);
+}
+
 // Noise estimation in chest_dl, interpolation
 void srslte_vec_sub_ccc(const cf_t *x, const cf_t *y, cf_t *z, const uint32_t len) {
   return srslte_vec_sub_fff((const float*) x,(const float*) y,(float*) z, 2*len);
@@ -100,10 +104,18 @@ void srslte_vec_convert_fi(const float *x, const float scale, int16_t *z, const
   srslte_vec_convert_fi_simd(x, z, scale, len);
 }
 
+void srslte_vec_convert_fb(const float *x, const float scale, int8_t *z, const uint32_t len) {
+  srslte_vec_convert_fb_simd(x, z, scale, len);
+}
+
 void srslte_vec_lut_sss(const short *x, const unsigned short *lut, short *y, const uint32_t len) {
   srslte_vec_lut_sss_simd(x, lut, y, len);
 }
 
+void srslte_vec_lut_bbb(const int8_t *x, const unsigned short *lut, int8_t *y, const uint32_t len) {
+  srslte_vec_lut_bbb_simd(x, lut, y, len);
+}
+
 void srslte_vec_lut_sis(const short *x, const unsigned int *lut, short *y, const uint32_t len) {
   for (int i=0; i < len; i++) {
     y[lut[i]] = x[i];
@@ -163,6 +175,15 @@ void srslte_vec_fprint_b(FILE *stream, uint8_t *x, const uint32_t len) {
   fprintf(stream, "];\n");
 }
 
+void srslte_vec_fprint_bs(FILE *stream, int8_t *x, const uint32_t len) {
+  int i;
+  fprintf(stream, "[");
+  for (i=0;i<len;i++) {
+    fprintf(stream, "%4d, ", x[i]);
+  }
+  fprintf(stream, "];\n");
+}
+
 void srslte_vec_fprint_byte(FILE *stream, uint8_t *x, const uint32_t len) {
   int i;
   fprintf(stream, "[");
@@ -185,7 +206,7 @@ void srslte_vec_fprint_s(FILE *stream, short *x, const uint32_t len) {
   int i;
   fprintf(stream, "[");
   for (i=0;i<len;i++) {
-    fprintf(stream, "%d, ", x[i]);
+    fprintf(stream, "%4d, ", x[i]);
   }
   fprintf(stream, "];\n");
 }
@@ -271,11 +292,18 @@ void srslte_vec_prod_fff(const float *x, const float *y, float *z, const uint32_
   srslte_vec_prod_fff_simd(x, y, z, len);
 }
 
-// Scrambling Short
 void srslte_vec_prod_sss(const int16_t *x, const int16_t *y, int16_t *z, const uint32_t len) {
   srslte_vec_prod_sss_simd(x,y,z,len);
 }
 
+// Scrambling
+void srslte_vec_neg_sss(const int16_t *x, const int16_t *y, int16_t *z, const uint32_t len) {
+  srslte_vec_neg_sss_simd(x,y,z,len);
+}
+void srslte_vec_neg_bbb(const int8_t *x, const int8_t *y, int8_t *z, const uint32_t len) {
+  srslte_vec_neg_bbb_simd(x,y,z,len);
+}
+
 // CFO and OFDM processing
 void srslte_vec_prod_ccc(const cf_t *x, const cf_t *y, cf_t *z, const uint32_t len) {
   srslte_vec_prod_ccc_simd(x,y,z,len);

lib/src/phy/utils/vector_simd.c

@@ -162,6 +162,35 @@ void srslte_vec_sub_sss_simd(const int16_t *x, const int16_t *y, int16_t *z, con
   }
 }
 
+void srslte_vec_sub_bbb_simd(const int8_t *x, const int8_t *y, int8_t *z, const int len) {
+  int i = 0;
+#if SRSLTE_SIMD_B_SIZE
+  if (SRSLTE_IS_ALIGNED(x) && SRSLTE_IS_ALIGNED(y) && SRSLTE_IS_ALIGNED(z)) {
+    for (; i < len - SRSLTE_SIMD_B_SIZE + 1; i += SRSLTE_SIMD_B_SIZE) {
+      simd_s_t a = srslte_simd_b_load(&x[i]);
+      simd_s_t b = srslte_simd_b_load(&y[i]);
+
+      simd_s_t r = srslte_simd_b_sub(a, b);
+
+      srslte_simd_b_store(&z[i], r);
+    }
+  } else {
+    for (; i < len - SRSLTE_SIMD_S_SIZE + 1; i += SRSLTE_SIMD_S_SIZE) {
+      simd_s_t a = srslte_simd_b_loadu(&x[i]);
+      simd_s_t b = srslte_simd_b_loadu(&y[i]);
+
+      simd_s_t r = srslte_simd_b_sub(a, b);
+
+      srslte_simd_b_storeu(&z[i], r);
+    }
+  }
+#endif /* SRSLTE_SIMD_S_SIZE */
+
+  for(; i < len; i++){
+    z[i] = x[i] - y[i];
+  }
+}
+
 void srslte_vec_prod_sss_simd(const int16_t *x, const int16_t *y, int16_t *z, const int len) {
   int i = 0;
 #if SRSLTE_SIMD_S_SIZE
@@ -191,12 +220,69 @@ void srslte_vec_prod_sss_simd(const int16_t *x, const int16_t *y, int16_t *z, co
   }
 }
 
+void srslte_vec_neg_sss_simd(const int16_t *x, const int16_t *y, int16_t *z, const int len) {
+  int i = 0;
+#if SRSLTE_SIMD_S_SIZE
+  if (SRSLTE_IS_ALIGNED(x) && SRSLTE_IS_ALIGNED(y) && SRSLTE_IS_ALIGNED(z)) {
+    for (; i < len - SRSLTE_SIMD_S_SIZE + 1; i += SRSLTE_SIMD_S_SIZE) {
+      simd_s_t a = srslte_simd_s_load(&x[i]);
+      simd_s_t b = srslte_simd_s_load(&y[i]);
+
+      simd_s_t r = srslte_simd_s_neg(a, b);
+
+      srslte_simd_s_store(&z[i], r);
+    }
+  } else {
+    for (; i < len - SRSLTE_SIMD_S_SIZE + 1; i += SRSLTE_SIMD_S_SIZE) {
+      simd_s_t a = srslte_simd_s_loadu(&x[i]);
+      simd_s_t b = srslte_simd_s_loadu(&y[i]);
+
+      simd_s_t r = srslte_simd_s_neg(a, b);
+
+      srslte_simd_s_storeu(&z[i], r);
+    }
+  }
+#endif /* SRSLTE_SIMD_S_SIZE */
+
+  for(; i < len; i++){
+    z[i] = y[i]<0?-x[i]:x[i];
+  }
+}
+
+void srslte_vec_neg_bbb_simd(const int8_t *x, const int8_t *y, int8_t *z, const int len) {
+  int i = 0;
+#if SRSLTE_SIMD_B_SIZE
+  if (SRSLTE_IS_ALIGNED(x) && SRSLTE_IS_ALIGNED(y) && SRSLTE_IS_ALIGNED(z)) {
+    for (; i < len - SRSLTE_SIMD_B_SIZE + 1; i += SRSLTE_SIMD_B_SIZE) {
+      simd_s_t a = srslte_simd_b_load(&x[i]);
+      simd_s_t b = srslte_simd_b_load(&y[i]);
+
+      simd_s_t r = srslte_simd_b_neg(a, b);
+
+      srslte_simd_b_store(&z[i], r);
+    }
+  } else {
+    for (; i < len - SRSLTE_SIMD_B_SIZE + 1; i += SRSLTE_SIMD_B_SIZE) {
+      simd_s_t a = srslte_simd_b_loadu(&x[i]);
+      simd_s_t b = srslte_simd_b_loadu(&y[i]);
+
+      simd_s_t r = srslte_simd_b_neg(a, b);
+
+      srslte_simd_b_storeu(&z[i], r);
+    }
+  }
+#endif /* SRSLTE_SIMD_S_SIZE */
+
+  for(; i < len; i++){
+    z[i] = y[i]<0?-x[i]:x[i];
+  }
+}
+
 /* No improvement with AVX */
 void srslte_vec_lut_sss_simd(const short *x, const unsigned short *lut, short *y, const int len) {
   int i = 0;
 #ifdef LV_HAVE_SSE
-#if CMAKE_BUILD_TYPE!=Debug
+#ifndef DEBUG_MODE
-
   if (SRSLTE_IS_ALIGNED(x) && SRSLTE_IS_ALIGNED(lut)) {
     for (; i < len - 7; i += 8) {
       __m128i xVal = _mm_load_si128((__m128i *) &x[i]);
@@ -228,6 +314,53 @@ void srslte_vec_lut_sss_simd(const short *x, const unsigned short *lut, short *y
   }
 }
 
+void srslte_vec_lut_bbb_simd(const int8_t *x, const unsigned short *lut, int8_t *y, const int len) {
+  int i = 0;
+#ifdef LV_HAVE_SSE
+#ifndef DEBUG_MODE
+  if (SRSLTE_IS_ALIGNED(x) && SRSLTE_IS_ALIGNED(lut)) {
+    for (; i < len - 15; i += 16) {
+      __m128i xVal = _mm_load_si128((__m128i *) &x[i]);
+      __m128i lutVal1 = _mm_load_si128((__m128i *) &lut[i]);
+      __m128i lutVal2 = _mm_load_si128((__m128i *) &lut[i+8]);
+
+      for (int k = 0; k < 8; k++) {
+        int8_t x = (int8_t) _mm_extract_epi8(xVal, k);
+        uint16_t l = (uint16_t) _mm_extract_epi16(lutVal1, k);
+        y[l] = (char) x;
+      }
+      for (int k = 0; k < 8; k++) {
+        int8_t x = (int8_t) _mm_extract_epi8(xVal, k+8);
+        uint16_t l = (uint16_t) _mm_extract_epi16(lutVal2, k);
+        y[l] = (char) x;
+      }
+    }
+  } else {
+    for (; i < len - 15; i += 16) {
+      __m128i xVal = _mm_loadu_si128((__m128i *) &x[i]);
+      __m128i lutVal1 = _mm_loadu_si128((__m128i *) &lut[i]);
+      __m128i lutVal2 = _mm_loadu_si128((__m128i *) &lut[i+8]);
+
+      for (int k = 0; k < 8; k++) {
+        int8_t x = (int8_t) _mm_extract_epi8(xVal, k);
+        uint16_t l = (uint16_t) _mm_extract_epi16(lutVal1, k);
+        y[l] = (char) x;
+      }
+      for (int k = 0; k < 8; k++) {
+        int8_t x = (int8_t) _mm_extract_epi8(xVal, k+8);
+        uint16_t l = (uint16_t) _mm_extract_epi16(lutVal2, k);
+        y[l] = (char) x;
+      }
+    }
+  }
+#endif
+#endif
+
+  for (; i < len; i++) {
+    y[lut[i]] = x[i];
+  }
+}
+
 void srslte_vec_convert_if_simd(const int16_t *x, float *z, const float scale, const int len) {
   int i = 0;
   const float gain = 1.0f / scale;
@@ -295,6 +428,73 @@ void srslte_vec_convert_fi_simd(const float *x, int16_t *z, const float scale, c
   }
 }
 
+#define SRSLTE_IS_ALIGNED_SSE(PTR) (((size_t)(PTR) & 0x0F) == 0)
+
+void srslte_vec_convert_fb_simd(const float *x, int8_t *z, const float scale, const int len) {
+  int i = 0;
+
+  // Force the use of SSE here instead of AVX since the implementations requires too many permutes across 128-bit boundaries
+
+#ifdef LV_HAVE_SSE
+  __m128 s = _mm_set1_ps(scale);
+  if (SRSLTE_IS_ALIGNED_SSE(x) && SRSLTE_IS_ALIGNED_SSE(z)) {
+    for (; i < len - 16 + 1; i += 16) {
+      __m128 a = _mm_load_ps(&x[i]);
+      __m128 b = _mm_load_ps(&x[i + 1*4]);
+      __m128 c = _mm_load_ps(&x[i + 2*4]);
+      __m128 d = _mm_load_ps(&x[i + 3*4]);
+
+      __m128  sa = _mm_mul_ps(a, s);
+      __m128  sb = _mm_mul_ps(b, s);
+      __m128  sc = _mm_mul_ps(c, s);
+      __m128  sd = _mm_mul_ps(d, s);
+
+      __m128i ai = _mm_cvttps_epi32(sa);
+      __m128i bi = _mm_cvttps_epi32(sb);
+      __m128i ci = _mm_cvttps_epi32(sc);
+      __m128i di = _mm_cvttps_epi32(sd);
+      __m128i ab = _mm_packs_epi32(ai, bi);
+      __m128i cd = _mm_packs_epi32(ci, di);
+
+      __m128i i8 =_mm_packs_epi16(ab, cd);
+
+      _mm_store_si128((__m128i*)&z[i], i8);
+    }
+  } else {
+    for (; i < len - 16 + 1; i += 16) {
+      __m128 a = _mm_load_ps(&x[i]);
+      __m128 b = _mm_load_ps(&x[i + 1*4]);
+      __m128 c = _mm_load_ps(&x[i + 2*4]);
+      __m128 d = _mm_load_ps(&x[i + 3*4]);
+
+      __m128  sa = _mm_mul_ps(a, s);
+      __m128  sb = _mm_mul_ps(b, s);
+      __m128  sc = _mm_mul_ps(c, s);
+      __m128  sd = _mm_mul_ps(d, s);
+
+      __m128i ai = _mm_cvttps_epi32(sa);
+      __m128i bi = _mm_cvttps_epi32(sb);
+      __m128i ci = _mm_cvttps_epi32(sc);
+      __m128i di = _mm_cvttps_epi32(sd);
+      __m128i ab = _mm_packs_epi32(ai, bi);
+      __m128i cd = _mm_packs_epi32(ci, di);
+
+      __m128i i8 =_mm_packs_epi16(ab, cd);
+
+      _mm_storeu_si128((__m128i*)&z[i], i8);
+    }
+  }
+#endif
+
+#ifdef HAVE_NEON
+#warning srslte_vec_convert_fb_simd not implemented in neon
+#endif /* HAVE_NEON */
+
+  for(; i < len; i++){
+    z[i] = (int8_t) (x[i] * scale);
+  }
+}
+
 float srslte_vec_acc_ff_simd(const float *x, const int len) {
   int i = 0;
   float acc_sum = 0.0f;

srsenb/enb.conf.example

@@ -141,7 +141,8 @@ nof_ctrl_symbols = 3
 #####################################################################
 # Expert configuration options
 #
-# pdsch_max_its:        Maximum number of turbo decoder iterations (Default 4)
+# pusch_max_its:        Maximum number of turbo decoder iterations (Default 4)
+# pusch_8bit_decoder:   Use 8-bit for LLR representation and turbo decoder trellis computation (Experimental)
 # nof_phy_threads:      Selects the number of PHY threads (maximum 4, minimum 1, default 2)
 # metrics_period_secs:  Sets the period at which metrics are requested from the UE. 
 # pregenerate_signals:  Pregenerate uplink signals after attach. Improves CPU performance.
@@ -152,7 +153,8 @@ nof_ctrl_symbols = 3
 #
 #####################################################################
 [expert]
-#pdsch_max_its        = 4
+#pusch_max_its        = 8 # These are half iterations
+#pusch_8bit_decoder   = false
 #nof_phy_threads      = 2
 #pregenerate_signals  = false
 #tx_amplitude         = 0.6

srsenb/hdr/phy/phch_common.h

@@ -41,6 +41,7 @@ namespace srsenb {
 typedef struct {
   float max_prach_offset_us; 
   int pusch_max_its;
+  bool pusch_8bit_decoder;
   float tx_amplitude; 
   int nof_phy_threads;  
   std::string equalizer_mode; 

srsenb/src/main.cc

@@ -157,10 +157,14 @@ void parse_args(all_args_t *args, int argc, char* argv[]) {
         "Pregenerate uplink signals after attach. Improves CPU performance.")
 
     ("expert.pusch_max_its",
-        bpo::value<int>(&args->expert.phy.pusch_max_its)->default_value(4),
+        bpo::value<int>(&args->expert.phy.pusch_max_its)->default_value(8),
         "Maximum number of turbo decoder iterations")
 
-    ("expert.tx_amplitude",
+      ("expert.pusch_8bit_decoder",
+       bpo::value<bool>(&args->expert.phy.pusch_8bit_decoder)->default_value(false),
+       "Use 8-bit for LLR representation and turbo decoder trellis computation (Experimental)")
+
+      ("expert.tx_amplitude",
         bpo::value<float>(&args->expert.phy.tx_amplitude)->default_value(0.6),
         "Transmit amplitude factor")
 

srsenb/src/phy/phch_worker.cc

@@ -154,6 +154,10 @@ void phch_worker::init(phch_common* phy_, srslte::log *log_h_)
   
   Info("Worker %d configured cell %d PRB\n", get_id(), phy->cell.nof_prb);
 
+  if (phy->params.pusch_8bit_decoder) {
+    enb_ul.pusch.llr_is_8bit = true;
+    enb_ul.pusch.ul_sch.llr_is_8bit = true;
+  }
   initiated = true; 
   running   = true;
 

srsue/src/main.cc

@@ -214,7 +214,7 @@ void parse_args(all_args_t *args, int argc, char *argv[]) {
      "Sets the noise estimation algorithm. (Default refs)")
 
     ("expert.pdsch_max_its",
-     bpo::value<int>(&args->expert.phy.pdsch_max_its)->default_value(4),
+     bpo::value<int>(&args->expert.phy.pdsch_max_its)->default_value(8),
      "Maximum number of turbo decoder iterations")
 
     ("expert.attach_enable_64qam",
@@ -307,7 +307,11 @@ void parse_args(all_args_t *args, int argc, char *argv[]) {
      bpo::value<bool>(&args->expert.phy.pdsch_csi_enabled)->default_value(true),
      "Stores the Channel State Information and uses it for weightening the softbits. It is only used in TM1.")
 
-    ("rf_calibration.tx_corr_dc_gain", bpo::value<float>(&args->rf_cal.tx_corr_dc_gain)->default_value(0.0),
+    ("expert.pdsch_8bit_decoder",
+       bpo::value<bool>(&args->expert.phy.pdsch_8bit_decoder)->default_value(false),
+       "Use 8-bit for LLR representation and turbo decoder trellis computation (Experimental)")
+
+      ("rf_calibration.tx_corr_dc_gain", bpo::value<float>(&args->rf_cal.tx_corr_dc_gain)->default_value(0.0),
      "TX DC offset gain correction")
     ("rf_calibration.tx_corr_dc_phase", bpo::value<float>(&args->rf_cal.tx_corr_dc_phase)->default_value(0.0),
      "TX DC offset phase correction")

srsue/src/phy/phch_worker.cc

@@ -139,6 +139,10 @@ bool phch_worker::init(uint32_t max_prb, srslte::log *log_h, srslte::log *log_ph
     return false;
   }
 
+  if (phy->args->pdsch_8bit_decoder) {
+    ue_dl.pdsch.llr_is_8bit = true;
+    ue_dl.pdsch.dl_sch.llr_is_8bit = true;
+  }
 
   srslte_chest_dl_set_rsrp_neighbour(&ue_dl.chest, true);
   srslte_chest_dl_average_subframe(&ue_dl.chest, phy->args->average_subframe_enabled);

srsue/ue.conf.example

@@ -204,6 +204,8 @@ enable = false
 # pdsch_csi_enabled:     Stores the Channel State Information and uses it for weightening the softbits. It is only
 #                        used in TM1. It is True by default.
 #
+# pdsch_8bit_decoder:    Use 8-bit for LLR representation and turbo decoder trellis computation (Experimental)
+#
 #####################################################################
 [expert]
 #ip_netmask          = 255.255.255.0
@@ -215,7 +217,7 @@ enable = false
 #cqi_fixed           = 10
 #snr_ema_coeff       = 0.1
 #snr_estim_alg       = refs
-#pdsch_max_its       = 4
+#pdsch_max_its       = 8    # These are half iterations
 #attach_enable_64qam = false
 #nof_phy_threads     = 2
 #equalizer_mode      = mmse
@@ -234,6 +236,7 @@ enable = false
 #metrics_period_secs = 1
 #metrics_csv_filename = /tmp/ue_metrics.csv
 #pdsch_csi_enabled  = true
+#pdsch_8bit_decoder = false
 
 # CFO related values
 #cfo_is_doppler      = false