From e1b9881c7815f6e4de929b1cb56da939878877bb Mon Sep 17 00:00:00 2001
From: Laurent THOMAS <laurent.thomas@open-cells.com>
Date: Tue, 5 Apr 2022 19:26:05 +0200
Subject: [PATCH] fix memory errors

---
 openair1/PHY/INIT/nr_init_ue.c        |  32 ++--
 openair1/PHY/NR_UE_TRANSPORT/sss_nr.c |   3 +
 openair1/PHY/TOOLS/oai_dfts.c         |  27 ++-
 openair1/PHY/TOOLS/tools_defs.h       | 245 +++++++++++++-------------
 4 files changed, 158 insertions(+), 149 deletions(-)

diff --git a/openair1/PHY/INIT/nr_init_ue.c b/openair1/PHY/INIT/nr_init_ue.c
index 142ef9b1d9..1417a16819 100644
--- a/openair1/PHY/INIT/nr_init_ue.c
+++ b/openair1/PHY/INIT/nr_init_ue.c
@@ -311,11 +311,26 @@ int init_nr_ue_signal(PHY_VARS_NR_UE *ue, int nb_connected_gNB)
   }
 
   // DLSCH
-  for (gNB_id = 0; gNB_id < ue->n_connected_gNB; gNB_id++) {
+  for (gNB_id = 0; gNB_id < ue->n_connected_gNB+1; gNB_id++) {
     for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
       ue->pdsch_vars[th_id][gNB_id] = (NR_UE_PDSCH *)malloc16_clear(sizeof(NR_UE_PDSCH));
     }
+    for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
+      phy_init_nr_ue_PDSCH( ue->pdsch_vars[th_id][gNB_id], fp );
+    }
 
+    int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
+    for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
+      for (i=0; i<nb_codewords; i++) {
+        ue->pdsch_vars[th_id][gNB_id]->llr[i] = (int16_t *)malloc16_clear( (8*(3*8*8448))*sizeof(int16_t) );//Q_m = 8 bits/Sym, Code_Rate=3, Number of Segments =8, Circular Buffer K_cb = 8448
+      }
+      for (i=0; i<NR_MAX_NB_LAYERS; i++) {
+        ue->pdsch_vars[th_id][gNB_id]->layer_llr[i] = (int16_t *)malloc16_clear( (8*(3*8*8448))*sizeof(int16_t) );//Q_m = 8 bits/Sym, Code_Rate=3, Number of Segments =8, Circular Buffer K_cb = 8448
+      }
+    }
+  }
+
+  for (gNB_id = 0; gNB_id < ue->n_connected_gNB; gNB_id++) {
     for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
       ue->pdcch_vars[th_id][gNB_id] = (NR_UE_PDCCH *)malloc16_clear(sizeof(NR_UE_PDCCH));
     }
@@ -342,21 +357,6 @@ int init_nr_ue_signal(PHY_VARS_NR_UE *ue, int nb_connected_gNB)
       ue->nr_srs_info->srs_estimated_channel_time_shifted[i] = (int32_t *) malloc16_clear(fp->ofdm_symbol_size*MAX_NUM_NR_SRS_SYMBOLS*sizeof(int32_t));
     }
 
-
-    for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
-      phy_init_nr_ue_PDSCH( ue->pdsch_vars[th_id][gNB_id], fp );
-    }
-
-    int nb_codewords = NR_MAX_NB_LAYERS > 4 ? 2 : 1;
-    for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
-      for (i=0; i<nb_codewords; i++) {
-        ue->pdsch_vars[th_id][gNB_id]->llr[i] = (int16_t *)malloc16_clear( (8*(3*8*8448))*sizeof(int16_t) );//Q_m = 8 bits/Sym, Code_Rate=3, Number of Segments =8, Circular Buffer K_cb = 8448
-      }
-      for (i=0; i<NR_MAX_NB_LAYERS; i++) {
-        ue->pdsch_vars[th_id][gNB_id]->layer_llr[i] = (int16_t *)malloc16_clear( (8*(3*8*8448))*sizeof(int16_t) );//Q_m = 8 bits/Sym, Code_Rate=3, Number of Segments =8, Circular Buffer K_cb = 8448
-      }
-    }
-
     // 100 PRBs * 12 REs/PRB * 4 PDCCH SYMBOLS * 2 LLRs/RE
     for (th_id=0; th_id<RX_NB_TH_MAX; th_id++) {
       ue->pdcch_vars[th_id][gNB_id]->llr                 = (int16_t *)malloc16_clear( 2*4*100*12*sizeof(uint16_t) );
diff --git a/openair1/PHY/NR_UE_TRANSPORT/sss_nr.c b/openair1/PHY/NR_UE_TRANSPORT/sss_nr.c
index dc73fe0add..88585e6094 100644
--- a/openair1/PHY/NR_UE_TRANSPORT/sss_nr.c
+++ b/openair1/PHY/NR_UE_TRANSPORT/sss_nr.c
@@ -558,6 +558,9 @@ int rx_sss_nr(PHY_VARS_NR_UE *ue, UE_nr_rxtx_proc_t *proc, int32_t *tot_metric,
   }
   //#endif
 
+  if (Nid1==N_ID_1_NUMBER)
+    return -1;
+  
   int re = 0;
   int im = 0;
   d = (int16_t *)&d_sss[Nid2][Nid1];
diff --git a/openair1/PHY/TOOLS/oai_dfts.c b/openair1/PHY/TOOLS/oai_dfts.c
index 83207426fd..c266a7f82b 100644
--- a/openair1/PHY/TOOLS/oai_dfts.c
+++ b/openair1/PHY/TOOLS/oai_dfts.c
@@ -10606,15 +10606,30 @@ int dfts_autoinit(void)
 
 #ifndef MR_MAIN
 
-void dft(uint8_t sizeidx, int16_t *sigF,int16_t *sig,unsigned char scale_flag){
-	AssertFatal((sizeidx>=0 && sizeidx<(int)DFT_SIZE_IDXTABLESIZE),"Invalid dft size index %i\n",sizeidx);
-	dft_ftab[sizeidx](sigF,sig,scale_flag);
+void dft(uint8_t sizeidx, int16_t *input,int16_t *output,unsigned char scale_flag){
+	AssertFatal((sizeidx>=0 && sizeidx<DFT_SIZE_IDXTABLESIZE),"Invalid dft size index %i\n",sizeidx);
+        AssertFatal( ((intptr_t)output&0x1F)==0,"Buffers should be 32 bytes aligned %p",output);
+        if ((intptr_t)input&0x1F) {
+          LOG_D(PHY, "DFT called with input not aligned, add a memcpy, size %d\n", sizeidx);
+          int16_t tmp[dft_ftab[sizeidx].size*2] __attribute__ ((aligned(32))); // input and output are not in right type (int16_t instead of c16_t)
+          memcpy(tmp, input, sizeof tmp);
+          dft_ftab[sizeidx].func(tmp,output,scale_flag);
+        } else
+          dft_ftab[sizeidx].func(input,output,scale_flag);
 };
 
-void idft(uint8_t sizeidx, int16_t *sigF,int16_t *sig,unsigned char scale_flag){
-	AssertFatal((sizeidx>=0 && sizeidx<(int)IDFT_SIZE_IDXTABLESIZE),"Invalid idft size index %i\n",sizeidx);
-	idft_ftab[sizeidx](sigF,sig,scale_flag);
+void idft(uint8_t sizeidx, int16_t *input,int16_t *output,unsigned char scale_flag){
+	AssertFatal((sizeidx>=0 && sizeidx<DFT_SIZE_IDXTABLESIZE),"Invalid idft size index %i\n",sizeidx);
+        AssertFatal( ((intptr_t)output&0x1F)==0,"Buffers should be 32 bytes aligned %p",output);
+        if ((intptr_t)input&0x1F) {
+          LOG_D(PHY, "DFT called with input not aligned, add a memcpy\n");
+          int16_t tmp[idft_ftab[sizeidx].size*2] __attribute__ ((aligned(32))); // input and output are not in right type (int16_t instead of c16_t)
+          memcpy(tmp, input, sizeof tmp);
+          dft_ftab[sizeidx].func(tmp,output,scale_flag);
+        } else
+          idft_ftab[sizeidx].func(input,output,scale_flag);
 };
+
 #endif
 
 /*---------------------------------------------------------------------------------------*/
diff --git a/openair1/PHY/TOOLS/tools_defs.h b/openair1/PHY/TOOLS/tools_defs.h
index 966e93440e..30327cf689 100644
--- a/openair1/PHY/TOOLS/tools_defs.h
+++ b/openair1/PHY/TOOLS/tools_defs.h
@@ -181,106 +181,110 @@ This function performs optimized fixed-point radix-2 FFT/IFFT.
         );
 */
 
+#define FOREACH_DFTSZ(SZ_DEF) \
+  SZ_DEF(12) \
+  SZ_DEF(24) \
+  SZ_DEF(36) \
+  SZ_DEF(48) \
+  SZ_DEF(60) \
+  SZ_DEF(64) \
+  SZ_DEF(72) \
+  SZ_DEF(96) \
+  SZ_DEF(108) \
+  SZ_DEF(120) \
+  SZ_DEF(128) \
+  SZ_DEF(144) \
+  SZ_DEF(180) \
+  SZ_DEF(192) \
+  SZ_DEF(216) \
+  SZ_DEF(240) \
+  SZ_DEF(256) \
+  SZ_DEF(288) \
+  SZ_DEF(300) \
+  SZ_DEF(324) \
+  SZ_DEF(360) \
+  SZ_DEF(384) \
+  SZ_DEF(432) \
+  SZ_DEF(480) \
+  SZ_DEF(512) \
+  SZ_DEF(540) \
+  SZ_DEF(576) \
+  SZ_DEF(600) \
+  SZ_DEF(648) \
+  SZ_DEF(720) \
+  SZ_DEF(768) \
+  SZ_DEF(864) \
+  SZ_DEF(900) \
+  SZ_DEF(960) \
+  SZ_DEF(972) \
+  SZ_DEF(1024) \
+  SZ_DEF(1080) \
+  SZ_DEF(1152) \
+  SZ_DEF(1200) \
+  SZ_DEF(1296) \
+  SZ_DEF(1440) \
+  SZ_DEF(1500) \
+  SZ_DEF(1536) \
+  SZ_DEF(1620) \
+  SZ_DEF(1728) \
+  SZ_DEF(1800) \
+  SZ_DEF(1920) \
+  SZ_DEF(1944) \
+  SZ_DEF(2048) \
+  SZ_DEF(2160) \
+  SZ_DEF(2304) \
+  SZ_DEF(2400) \
+  SZ_DEF(2592) \
+  SZ_DEF(2700) \
+  SZ_DEF(2880) \
+  SZ_DEF(2916) \
+  SZ_DEF(3000) \
+  SZ_DEF(3072) \
+  SZ_DEF(3240) \
+  SZ_DEF(4096) \
+  SZ_DEF(6144) \
+  SZ_DEF(8192) \
+  SZ_DEF(9216) \
+  SZ_DEF(12288) \
+  SZ_DEF(18432) \
+  SZ_DEF(24576) \
+  SZ_DEF(36864) \
+  SZ_DEF(49152) \
+  SZ_DEF(73728) \
+  SZ_DEF(98304)
 
+#define FOREACH_IDFTSZ(SZ_DEF)                  \
+  SZ_DEF(64) \
+  SZ_DEF(128) \
+  SZ_DEF(256) \
+  SZ_DEF(512) \
+  SZ_DEF(1024) \
+  SZ_DEF(1536) \
+  SZ_DEF(2048) \
+  SZ_DEF(3072) \
+  SZ_DEF(4096) \
+  SZ_DEF(6144) \
+  SZ_DEF(8192) \
+  SZ_DEF(9216) \
+  SZ_DEF(12288) \
+  SZ_DEF(18432) \
+  SZ_DEF(24576) \
+  SZ_DEF(36864) \
+  SZ_DEF(49152) \
+  SZ_DEF(73728) \
+  SZ_DEF(98304) 
 
 #ifdef OAIDFTS_MAIN
 typedef  void(*adftfunc_t)(int16_t *sigF,int16_t *sig,unsigned char scale_flag);  
 typedef  void(*aidftfunc_t)(int16_t *sigF,int16_t *sig,unsigned char scale_flag);     
 
-void dft12(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft24(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft36(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft48(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft60(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft64(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft72(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft96(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft108(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft120(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft128(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft144(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft180(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft192(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft216(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft240(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft256(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft288(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft300(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft324(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft360(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft384(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft432(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft480(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft512(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft540(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft576(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft600(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft648(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft720(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft768(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft864(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft900(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft960(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft972(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1024(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1080(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1152(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1200(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1296(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1440(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1500(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1536(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void dft1620(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1728(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1800(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1920(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft1944(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2048(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2160(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2304(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2400(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2592(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2700(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2880(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft2916(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft3000(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft3072(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void dft3240(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft4096(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft6144(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void dft8192(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft9216(int16_t *x,int16_t *y,uint8_t scale_flag);
-void dft12288(int16_t *x,int16_t *y,uint8_t scale_flag);  
-void dft18432(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft24576(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft36864(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft49152(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft73728(int16_t *x,int16_t *y,uint8_t scale_flag); 
-void dft98304(int16_t *x,int16_t *y,uint8_t scale_flag);
-
-
-void idft64(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft128(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft256(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft512(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft1024(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft1536(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft2048(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft3072(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft4096(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft6144(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft8192(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft9216(int16_t *x,int16_t *y,uint8_t scale_flag);
-void idft12288(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft18432(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft24576(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft36864(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft49152(int16_t *sigF,int16_t *sig,uint8_t scale_flag); 
-void idft73728(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
-void idft98304(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
+#define SZ_FUNC(Sz) void dft ## Sz(int16_t *x,int16_t *y,uint8_t scale_flag);
 
+  FOREACH_DFTSZ(SZ_FUNC)
 
+#define SZ_iFUNC(Sz) void idft ## Sz(int16_t *x,int16_t *y,uint8_t scale_flag);
 
+  FOREACH_IDFTSZ(SZ_iFUNC)
 
 #else
   typedef  void(*dftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);  
@@ -294,46 +298,33 @@ void idft98304(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
   extern int load_dftslib(void);
 #  endif
 #endif
+  
+#define SZ_ENUM(Sz) DFT_ ## Sz,
 
-typedef enum DFT_size_idx {
-	DFT_12,    DFT_24,    DFT_36,   DFT_48,     DFT_60,   DFT_72,   DFT_96,
-	DFT_108,   DFT_120,   DFT_128,  DFT_144,    DFT_180,  DFT_192,  DFT_216,   DFT_240,
-	DFT_256,   DFT_288,   DFT_300,  DFT_324,    DFT_360,  DFT_384,  DFT_432,   DFT_480,
-	DFT_512,   DFT_540,   DFT_576,  DFT_600,    DFT_648,  DFT_720,  DFT_768,   DFT_864,
-	DFT_900,   DFT_960,   DFT_972,  DFT_1024,   DFT_1080, DFT_1152, DFT_1200,  DFT_1296,
-	DFT_1440,  DFT_1500,  DFT_1536, DFT_1620,   DFT_1728, DFT_1800, DFT_1920,  DFT_1944,
-	DFT_2048,  DFT_2160,  DFT_2304, DFT_2400,   DFT_2592, DFT_2700, DFT_2880,  DFT_2916,
-	DFT_3000,  DFT_3072,  DFT_3240, DFT_4096,   DFT_6144, DFT_8192, DFT_9216,  DFT_12288,
-	DFT_18432, DFT_24576, DFT_36864, DFT_49152, DFT_73728, DFT_98304,
-	DFT_SIZE_IDXTABLESIZE
-} dft_size_idx_t;
+  typedef enum dft_size_idx {
+    FOREACH_DFTSZ(SZ_ENUM)
+    DFT_SIZE_IDXTABLESIZE
+}  dft_size_idx_t;
+  
+#define SZ_iENUM(Sz) IDFT_ ## Sz,
 
+  typedef enum idft_size_idx {
+  FOREACH_IDFTSZ(SZ_iENUM)
+  IDFT_SIZE_IDXTABLESIZE
+}  idft_size_idx_t;
+  
 #ifdef OAIDFTS_MAIN
-adftfunc_t dft_ftab[]={
-	dft12,    dft24,    dft36,    dft48,    dft60,   dft72,   dft96,
-	dft108,   dft120,   dft128,   dft144,   dft180,  dft192,  dft216,   dft240,
-	dft256,   dft288,   dft300,   dft324,   dft360,  dft384,  dft432,   dft480,
-	dft512,   dft540,   dft576,   dft600,   dft648,  dft720,  dft768,   dft864,
-	dft900,   dft960,   dft972,   dft1024,  dft1080, dft1152, dft1200,  dft1296,
-	dft1440,  dft1500,  dft1536,  dft1620,  dft1728, dft1800, dft1920,  dft1944,
-	dft2048,  dft2160,  dft2304,  dft2400,  dft2592, dft2700, dft2880,  dft2916,
-	dft3000,  dft3072,  dft3240,  dft4096,  dft6144, dft8192, dft9216,  dft12288,
-	dft18432, dft24576, dft36864, dft49152, dft73728, dft98304
-};
-#endif
+  
+#define SZ_PTR(Sz) {dft ## Sz,Sz},
+  struct {adftfunc_t func; int size;} dft_ftab[]={
+    FOREACH_DFTSZ(SZ_PTR)
+    };
+
+#define SZ_iPTR(Sz)  {idft ## Sz,Sz},
+ struct {adftfunc_t func; int size;} idft_ftab[]={
+    FOREACH_IDFTSZ(SZ_iPTR)
+    };
 
-typedef enum idft_size_idx {
-	IDFT_128,   IDFT_256,  IDFT_512,   IDFT_1024,  IDFT_1536,  IDFT_2048,  IDFT_3072,  IDFT_4096,
-	IDFT_6144,  IDFT_8192, IDFT_9216,  IDFT_12288, IDFT_18432, IDFT_24576, IDFT_36864, IDFT_49152, 
-	IDFT_73728, IDFT_98304, 
-	IDFT_SIZE_IDXTABLESIZE
-} idft_size_idx_t;
-#ifdef OAIDFTS_MAIN
-aidftfunc_t idft_ftab[]={
-	idft128,   idft256,  idft512,   idft1024,  idft1536,  idft2048,  idft3072,  idft4096,
-	idft6144,  idft8192, idft9216,  idft12288, idft18432, idft24576, idft36864, idft49152,
-	idft73728, idft98304
-};
 #endif