debug accelerate cuFFT2048

common/utils/LOG/vcd_signal_dumper.c

@@ -535,6 +535,7 @@ const char* eurecomFunctionsNames[] = {
   "nr_rx_pusch",
   "nr_ulsch_procedures_rx",
   "macxface_gNB_dlsch_ulsch_scheduler",
+  "cufft_wait",
 
   /*NR ue-softmodem signal*/
   "nr_ue_ulsch_encoding",

common/utils/LOG/vcd_signal_dumper.h

@@ -527,6 +527,7 @@ typedef enum {
   VCD_SIGNAL_DUMPER_FUNCTIONS_NR_RX_PUSCH,
   VCD_SIGNAL_DUMPER_FUNCTIONS_NR_ULSCH_PROCEDURES_RX,
   VCD_SIGNAL_DUMPER_FUNCTIONS_gNB_DLSCH_ULSCH_SCHEDULER,
+  VCD_SIGNAL_DUMPER_FUNCTIONS_CUFFT_WAIT,
 
 /* NR ue-softmodem signal*/
   VCD_SIGNAL_DUMPER_FUNCTIONS_NR_UE_ULSCH_ENCODING,

common/utils/T/T_defs.h

@@ -73,7 +73,7 @@ typedef struct {
 } T_cache_t;
 
 /* number of VCD functions (to be kept up to date! see in T_messages.txt) */
-#define VCD_NUM_FUNCTIONS (273)
+#define VCD_NUM_FUNCTIONS (274)
 
 /* number of VCD variables (to be kept up to date! see in T_messages.txt) */
 #define VCD_NUM_VARIABLES (187)

common/utils/T/T_messages.txt

@@ -3513,6 +3513,11 @@ ID = VCD_FUNCTION_gNB_DLSCH_ULSCH_SCHEDULER
     GROUP = ALL:VCD:ENB:VCD_FUNCTION
     FORMAT = int,value
     VCD_NAME = macxface_gNB_dlsch_ulsch_scheduler
+ID = VCD_FUNCTION_CUFFT_WAIT
+    DESC = VCD function CUFFT_WAIT
+    GROUP = ALL:VCD:ENB:VCD_FUNCTION
+    FORMAT = int,value
+    VCD_NAME = cufft_wait
 
   #function for nrUE
 ID = VCD_FUNCTION_NR_UE_ULSCH_ENCODING

openair1/PHY/CODING/nrLDPC_defs.h

@@ -57,4 +57,5 @@ typedef int(*nrLDPC_encoderfunc_t)(unsigned char **,unsigned char **,int,int,sho
 */
 typedef int32_t(*nrLDPC_decoderfunc_t)(t_nrLDPC_dec_params* , int8_t*, int8_t* , t_nrLDPC_procBuf* , t_nrLDPC_time_stats* );
 typedef void(*cudft_EnTx)(int16_t* , int16_t*, unsigned char );
+typedef void(*cudft_load)(void);
 #endif
\ No newline at end of file

openair1/PHY/CODING/nrLDPC_extern.h

@@ -24,7 +24,7 @@
 nrLDPC_decoderfunc_t nrLDPC_decoder;
 nrLDPC_encoderfunc_t nrLDPC_encoder;
 cudft_EnTx cudft2048;
-// cudft_EnTx cudft204;
+cudft_load load_cudft;
 #else
 /* functions to load the LDPC shared lib, implemented in openair1/PHY/CODING/nrLDPC_load.c */
 extern int load_nrLDPClib(void) ;
@@ -34,7 +34,7 @@ extern int load_cuFFT(void) ;
 extern nrLDPC_decoderfunc_t nrLDPC_decoder;
 extern nrLDPC_encoderfunc_t nrLDPC_encoder;
 extern cudft_EnTx cudft2048;
-// extern cudft_EnTx cudft204;
+extern cudft_load load_cudft;
 // inline functions:
 #include "openair1/PHY/CODING/nrLDPC_decoder/nrLDPC_init_mem.h"
 #endif
\ No newline at end of file

openair1/PHY/CODING/nrLDPC_load.c

@@ -84,17 +84,17 @@ int load_cuFFT(void) {
      //根据动态链接库操作句柄与符号，返回符号对应的地址
      cudft2048 = (cudft_EnTx) dlsym(handle, "_Z9cudft2048PsS_h");
      if(!cudft2048){
-          printf("cuFFT.so error!\n");
+          printf("cuFFT.so cudft2048 error!\n");
           dlclose(handle);
           return -1;
      }
-     // cudft204 = (cudft_EnTx) dlsym(handle, "_Z8testtestPsS_h");
-     // if(!cudft204){
-     //      printf("cutest.so error!\n");
-     //      dlclose(handle);
-     //      return -1;
-     // }
-
+     load_cudft = (cudft_load) dlsym(handle, "_Z10load_cuFFTv");
+     if(!load_cudft){
+          printf("cutest.so load_cudft error!\n");
+          dlclose(handle);
+          return -1;
+     }
+     load_cudft();
 return 0;
 }
 

openair1/PHY/MODULATION/slot_fep_nr.c

@@ -27,6 +27,8 @@
 #include "PHY/NR_UE_ESTIMATION/nr_estimation.h"
 #include <common/utils/LOG/log.h>
 #include "PHY/CODING/nrLDPC_extern.h"
+#include "common/utils/LOG/vcd_signal_dumper.h"
+#include "common/utils/LOG/log.h"
 
 //#define DEBUG_FEP
 
@@ -139,16 +141,6 @@ int nr_slot_fep(PHY_VARS_NR_UE *ue,
         (int16_t *)&common_vars->common_vars_rx_data_per_thread[proc->thread_id].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],
         1);
 
-    static int cu_2048 = 1;
-    if(cu_2048==1)
-    {
-      LOG_M("FFT0.m","input",rxdata_ptr,2048,1,3);
-      LOG_M("FFT1.m","output",(int16_t *)&common_vars->common_vars_rx_data_per_thread[proc->thread_id].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],2048,1,1);
-      cudft2048(rxdata_ptr,(int16_t *)&common_vars->common_vars_rx_data_per_thread[proc->thread_id].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],1);
-      LOG_M("cuFFT0.m","input",rxdata_ptr,2048,1,3);
-      LOG_M("cuFFT1.m","output",(int16_t *)&common_vars->common_vars_rx_data_per_thread[proc->thread_id].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],2048,1,1);
-      cu_2048++;
-    }   
 #if UE_TIMING_TRACE
     stop_meas(&ue->rx_dft_stats);
 #endif
@@ -339,13 +331,15 @@ int nr_slot_fep_ul(NR_DL_FRAME_PARMS *frame_parms,
 
   }
 
+  VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_CUFFT_WAIT, 1 );
   cudft2048(rxdata_ptr,
       (int16_t *)&rxdataF[symbol * frame_parms->ofdm_symbol_size],
-      1);
+      0);
   // dft(dftsize,
   //     rxdata_ptr,
   //     (int16_t *)&rxdataF[symbol * frame_parms->ofdm_symbol_size],
   //     1);
+  VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_CUFFT_WAIT, 0 );
 
   // clear DC carrier from OFDM symbols
   rxdataF[symbol * frame_parms->ofdm_symbol_size] = 0;