Fix rebasing problems

cmake_targets/CMakeLists.txt

@@ -166,7 +166,7 @@ if (CUDA_FOUND)
     add_definitions("-L/usr/local/cuda/lib64")
 
     SET(CUDA_NVCC_FLAGS
-      "${CUDA_NVCC_FLAGS};-arch=sm_60;")
+      "${CUDA_NVCC_FLAGS};--gpu-architecture=compute_60")
 
     # Disable warnings for CUDA
     SET(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS};-lpthread;-w;-O3;--default-stream;per-thread;-I/usr/local/cuda/inc;-L/usr/local/cuda/lib -lcutil;-rdc=true;-lcudadevrt")
@@ -1600,9 +1600,11 @@ add_library(ldpc_orig MODULE ${PHY_LDPC_ORIG_SRC} )
 add_library(ldpc_optim MODULE ${PHY_LDPC_OPTIM_SRC} )
 add_library(ldpc_optim8seg MODULE ${PHY_LDPC_OPTIM8SEG_SRC} )
 if (CUDA_FOUND)
-  add_library(ldpc_cuda MODULE ${PHY_LDPC_CUDA_SRC} )
-  CUDA_ADD_CUFFT_TO_TARGET(ldpc_cuda)
+  cuda_add_library(ldpc_cuda MODULE ${PHY_LDPC_CUDA_SRC} )
+  set_target_properties(ldpc_cuda PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
+#  CUDA_ADD_CUFFT_TO_TARGET(ldpc_cuda)
 endif (CUDA_FOUND)
+
 add_library(ldpc MODULE ${PHY_LDPC_OPTIM8SEGMULTI_SRC} )
 
 
@@ -3193,7 +3195,9 @@ add_executable(ldpctest
    )
 
 add_dependencies( ldpctest ldpc_orig ldpc_optim ldpc_optim8seg ldpc )
-
+if (CUDA_FOUND)
+   add_dependencies( ldpctest ldpc_cuda)
+endif (CUDA_FOUND)
 target_link_libraries(ldpctest
   -Wl,--start-group UTIL SIMU_COMMON  SIMU PHY_NR PHY_COMMON PHY_NR_COMMON CONFIG_LIB -Wl,--end-group
   m pthread ${ATLAS_LIBRARIES} dl

executables/nr-uesoftmodem.c

@@ -436,7 +436,7 @@ int main( int argc, char **argv ) {
   itti_init(TASK_MAX, tasks_info);
 
   init_opt() ;
-  load_nrLDPClib();
+  load_nrLDPClib(0);
 
   if (ouput_vcd) {
     vcd_signal_dumper_init("/tmp/openair_dump_nrUE.vcd");

openair1/PHY/CODING/TESTBENCH/ldpctest.c

@@ -101,8 +101,8 @@ int test_ldpc(short No_iteration,
               unsigned int *crc_misses,
               time_stats_t *time_optim,
               time_stats_t *time_decoder,
-              n_iter_stats_t *dec_iter,
-              short run_cuda)
+              n_iter_stats_t *dec_iter
+              )
 {
   //clock initiate
   //time_stats_t time,time_optim,tinput,tprep,tparity,toutput, time_decoder;
@@ -393,27 +393,12 @@ int test_ldpc(short No_iteration,
       decParams.R=code_rate_vec[R_ind];//13;
       decParams.numMaxIter=No_iteration;
       decParams.outMode = nrLDPC_outMode_BIT;
+      decParams.block_length=block_length;
       //decParams.outMode =nrLDPC_outMode_LLRINT8;
-#ifdef CUDA_FLAG
-	  set_compact_BG(Zc,BG);
-	  init_LLR_DMA_for_CUDA(&decParams, (int8_t*)channel_output_fixed[j], (int8_t*)estimated_output[j], block_length);
-#endif
+	  nrLDPC_initcall(&decParams, (int8_t*)channel_output_fixed[j], (int8_t*)estimated_output[j]);
 	  for(j=0;j<n_segments;j++) {
     	  start_meas(time_decoder);
-#ifdef CUDA_FLAG
-        if(run_cuda){
-          n_iter = nrLDPC_decoder_LYC(&decParams, (int8_t*)channel_output_fixed[j], (int8_t*)estimated_output[j], block_length, time_decoder);
-        }  
-        else{ 
-        // decode the sequence
-        // decoder supports BG2, Z=128 & 256
-        //esimated_output=ldpc_decoder(channel_output_fixed, block_length, No_iteration, (double)((float)nom_rate/(float)denom_rate));
-        ///nrLDPC_decoder(&decParams, channel_output_fixed, estimated_output, NULL);
-          n_iter = nrLDPC_decoder(&decParams, (int8_t*)channel_output_fixed[j], (int8_t*)estimated_output[j], p_nrLDPC_procBuf, p_decoder_profiler);
-        }
-#else
           n_iter = nrLDPC_decoder(&decParams, (int8_t*)channel_output_fixed[j], (int8_t*)estimated_output[j], p_nrLDPC_procBuf, p_decoder_profiler);
-#endif
 	      stop_meas(time_decoder);
       }
 
@@ -514,9 +499,7 @@ int test_ldpc(short No_iteration,
 
 int main(int argc, char *argv[])
 {
-#ifdef CUDA_FLAG	
-  warmup_for_GPU();
-#endif
+
   unsigned int errors, errors_bit, crc_misses;
   double errors_bit_uncoded;
   short block_length=8448; // decoder supports length: 1201 -> 1280, 2401 -> 2560
@@ -619,7 +602,7 @@ int main(int argc, char *argv[])
   printf("SNR0 %f: \n", SNR0);
 
 
-  load_nrLDPClib();
+  load_nrLDPClib(run_cuda);
   load_nrLDPClib_ref("_orig", &encoder_orig);
   //for (block_length=8;block_length<=MAX_BLOCK_LENGTH;block_length+=8)
 
@@ -691,8 +674,7 @@ int main(int argc, char *argv[])
                                 &crc_misses,
                                 time_optim,
                                 time_decoder,
-                                dec_iter,
-								run_cuda);
+                                dec_iter);
 
     printf("SNR %f, BLER %f (%u/%d)\n", SNR, (float)decoded_errors[i]/(float)n_trials, decoded_errors[i], n_trials);
     printf("SNR %f, BER %f (%u/%d)\n", SNR, (float)errors_bit/(float)n_trials/(float)block_length/(double)n_segments, decoded_errors[i], n_trials);

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_decoder.c

@@ -46,7 +46,8 @@
 #endif
 
 static inline uint32_t nrLDPC_decoder_core(int8_t* p_llr, int8_t* p_out, t_nrLDPC_procBuf* p_procBuf, uint32_t numLLR, t_nrLDPC_lut* p_lut, t_nrLDPC_dec_params* p_decParams, t_nrLDPC_time_stats* p_profiler);
-
+void nrLDPC_initcall(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out) {
+}
 int32_t nrLDPC_decod(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out, t_nrLDPC_procBuf* p_procBuf, t_nrLDPC_time_stats* p_profiler)
 {
     uint32_t numLLR;

openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_types.h

@@ -70,6 +70,7 @@ typedef struct nrLDPC_dec_params {
     uint16_t Z; /**< Lifting size */
     uint8_t R; /**< Decoding rate: Format 15,13,... for code rates 1/5, 1/3,... */
     uint8_t numMaxIter; /**< Maximum number of iterations */
+    int block_length;
     e_nrLDPC_outMode outMode; /**< Output format */
 } t_nrLDPC_dec_params;
 

openair1/PHY/CODING/nrLDPC_decoder_LYC/nrLDPC_decoder_LYC.cu

@@ -7,14 +7,14 @@
  * \note
  * \warning
  */
-
+#include <iostream>
 #include <stdio.h>
 #include <unistd.h>
 #include <cuda_runtime.h>
 #include <cuda.h>
 #include "PHY/CODING/nrLDPC_decoder/nrLDPC_types.h"
 #include "PHY/CODING/nrLDPC_decoder/nrLDPCdecoder_defs.h"
-
+#include "assertions.h"
 #include "bgs/BG1_I0"
 #include "bgs/BG1_I1"
 #include "bgs/BG1_I2"
@@ -462,10 +462,11 @@ void read_BG(int BG, int *h, int row, int col)
 }
 
 extern "C"
-void init_LLR_DMA_for_CUDA(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out, int block_length){
+void init_LLR_DMA(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out){
 	
 	uint16_t Zc          = p_decParams->Z;
     uint8_t  BG         = p_decParams->BG;
+    int block_length     = p_decParams->block_length;
 	uint8_t row,col;
 	if(BG == 1){
 		row = 46;
@@ -530,13 +531,12 @@ void nrLDPC_initcall(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_
 
 
 extern "C"
-int32_t nrLDPC_decoder_LYC(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out, int block_length, time_stats_t *time_decoder)
+int32_t nrLDPC_decod(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out,t_nrLDPC_procBuf* p_procBuf, t_nrLDPC_time_stats *time_decoder)
 {
-
-
     uint16_t Zc          = p_decParams->Z;
     uint8_t  BG         = p_decParams->BG;
     uint8_t  numMaxIter = p_decParams->numMaxIter;
+    int block_length    = p_decParams->block_length;
     e_nrLDPC_outMode outMode = p_decParams->outMode;
 	cudaError_t cudaStatus;
 	uint8_t row,col;

openair1/PHY/CODING/nrLDPC_defs.h

@@ -45,6 +45,7 @@ typedef struct {
 	time_stats_t *toutput;
 }encoder_implemparams_t;
 #define INIT0_LDPCIMPLEMPARAMS {0,0,0,NULL,NULL,NULL,NULL}
+typedef void(*nrLDPC_initcallfunc_t)(t_nrLDPC_dec_params* p_decParams, int8_t* p_llr, int8_t* p_out);
 typedef int(*nrLDPC_encoderfunc_t)(unsigned char **,unsigned char **,int,int,short, short, encoder_implemparams_t*);
 //============================================================================================================================
 // decoder interface

openair1/PHY/CODING/nrLDPC_extern.h

@@ -23,11 +23,14 @@
 #ifdef LDPC_LOADER
 nrLDPC_decoderfunc_t nrLDPC_decoder;
 nrLDPC_encoderfunc_t nrLDPC_encoder;
+nrLDPC_initcallfunc_t nrLDPC_initcall;
 #else
 /* functions to load the LDPC shared lib, implemented in openair1/PHY/CODING/nrLDPC_load.c */
-extern int load_nrLDPClib(void) ;
+extern int load_nrLDPClib(int) ;
 extern int load_nrLDPClib_ref(char *libversion, nrLDPC_encoderfunc_t * nrLDPC_encoder_ptr); // for ldpctest
 /* ldpc coder/decoder functions, as loaded by load_nrLDPClib(). */
+extern nrLDPC_initcallfunc_t nrLDPC_initcall;
+
 extern nrLDPC_decoderfunc_t nrLDPC_decoder;
 extern nrLDPC_encoderfunc_t nrLDPC_encoder;
 // inline functions:

openair1/PHY/CODING/nrLDPC_load.c

@@ -42,14 +42,13 @@
 
 
 /* function description array, to be used when loading the encoding/decoding shared lib */
-
 static loader_shlibfunc_t shlib_fdesc[3];
 
 /* arguments used when called from phy simulators exec's which do not use the config module */
 /* arg is used to initialize the config module so that the loader works as expected */
 char *arg[64]={"ldpctest","-O","cmdlineonly::dbgl0",NULL,NULL};
 
-int load_nrLDPClib(void) {
+int load_nrLDPClib(int run_cuda) {
 	 char *ptr = (char*)config_get_if();
 	 char libname[64]="ldpc";
 	 int argc=3;
@@ -65,10 +64,12 @@ int load_nrLDPClib(void) {
      }	 
      shlib_fdesc[0].fname = "nrLDPC_decod";
      shlib_fdesc[1].fname = "nrLDPC_encod";
-     int ret=load_module_shlib("ldpc",shlib_fdesc,sizeof(shlib_fdesc)/sizeof(loader_shlibfunc_t),NULL);
+     shlib_fdesc[2].fname = "nrLDPC_initcall";
+     int ret=load_module_shlib(libname,shlib_fdesc,sizeof(shlib_fdesc)/sizeof(loader_shlibfunc_t),NULL);
      AssertFatal( (ret >= 0),"Error loading ldpc decoder");
      nrLDPC_decoder = (nrLDPC_decoderfunc_t)shlib_fdesc[0].fptr;
      nrLDPC_encoder = (nrLDPC_encoderfunc_t)shlib_fdesc[1].fptr;
+     nrLDPC_initcall = (nrLDPC_initcallfunc_t)shlib_fdesc[2].fptr;
 return 0;
 }
 

openair1/PHY/INIT/nr_init.c

@@ -108,7 +108,7 @@ int phy_init_nr_gNB(PHY_VARS_gNB *gNB,
   crcTableInit();
   init_scrambling_luts();
   init_pucch2_luts();
-  load_nrLDPClib();
+  load_nrLDPClib(0);
   // PBCH DMRS gold sequences generation
   nr_init_pbch_dmrs(gNB);
   //PDCCH DMRS init