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 1/4] 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
 
 

From f46240ca2ba34291d2d9e12232541ce34fb9e680 Mon Sep 17 00:00:00 2001
From: Laurent THOMAS <laurent.thomas@open-cells.com>
Date: Fri, 8 Apr 2022 18:22:38 +0200
Subject: [PATCH 2/4] fix code review remarks

---
 openair1/PHY/INIT/nr_init_ue.c         |  6 ++-
 openair1/PHY/TOOLS/tools_defs.h        | 62 ++++++++++++++------------
 openair2/LAYER2/NR_MAC_gNB/mac_proto.h |  3 ++
 3 files changed, 41 insertions(+), 30 deletions(-)

diff --git a/openair1/PHY/INIT/nr_init_ue.c b/openair1/PHY/INIT/nr_init_ue.c
index 1417a16819..fe4eabb34e 100644
--- a/openair1/PHY/INIT/nr_init_ue.c
+++ b/openair1/PHY/INIT/nr_init_ue.c
@@ -468,16 +468,18 @@ void term_nr_ue_signal(PHY_VARS_NR_UE *ue, int nb_connected_gNB)
   }
   free_and_zero(ue->nr_gold_pdsch[0]);
 
-  for (int gNB_id = 0; gNB_id < ue->n_connected_gNB; gNB_id++) {
+  for (int gNB_id = 0; gNB_id < ue->n_connected_gNB+1; gNB_id++) {
 
     // PDSCH
     for (int th_id = 0; th_id < RX_NB_TH_MAX; th_id++) {
-
       free_and_zero(ue->pdsch_vars[th_id][gNB_id]->llr_shifts);
       free_and_zero(ue->pdsch_vars[th_id][gNB_id]->llr128_2ndstream);
       phy_term_nr_ue__PDSCH(ue->pdsch_vars[th_id][gNB_id], fp);
       free_and_zero(ue->pdsch_vars[th_id][gNB_id]);
     }
+  }
+  
+  for (int gNB_id = 0; gNB_id < ue->n_connected_gNB; gNB_id++) {
 
     for (int th_id = 0; th_id < RX_NB_TH_MAX; th_id++) {
       for (int i = 0; i < fp->nb_antennas_rx; i++) {
diff --git a/openair1/PHY/TOOLS/tools_defs.h b/openair1/PHY/TOOLS/tools_defs.h
index 30327cf689..7cee04fc59 100644
--- a/openair1/PHY/TOOLS/tools_defs.h
+++ b/openair1/PHY/TOOLS/tools_defs.h
@@ -272,58 +272,64 @@ This function performs optimized fixed-point radix-2 FFT/IFFT.
   SZ_DEF(36864) \
   SZ_DEF(49152) \
   SZ_DEF(73728) \
-  SZ_DEF(98304) 
+  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);     
+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);
 
 #define SZ_FUNC(Sz) void dft ## Sz(int16_t *x,int16_t *y,uint8_t scale_flag);
 
-  FOREACH_DFTSZ(SZ_FUNC)
+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)
+FOREACH_IDFTSZ(SZ_iFUNC)
 
 #else
-  typedef  void(*dftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);  
-  typedef  void(*idftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);  
+typedef  void(*dftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);
+typedef  void(*idftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);
 #  ifdef OAIDFTS_LOADER
-  dftfunc_t dft;
-  idftfunc_t idft;
+dftfunc_t dft;
+idftfunc_t idft;
 #  else
-  extern dftfunc_t dft;
-  extern idftfunc_t idft;
-  extern int load_dftslib(void);
+extern dftfunc_t dft;
+extern idftfunc_t idft;
+extern int load_dftslib(void);
 #  endif
 #endif
-  
+
 #define SZ_ENUM(Sz) DFT_ ## Sz,
 
-  typedef enum dft_size_idx {
-    FOREACH_DFTSZ(SZ_ENUM)
-    DFT_SIZE_IDXTABLESIZE
+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 {
+typedef enum idft_size_idx {
   FOREACH_IDFTSZ(SZ_iENUM)
   IDFT_SIZE_IDXTABLESIZE
 }  idft_size_idx_t;
-  
+
 #ifdef OAIDFTS_MAIN
-  
+
 #define SZ_PTR(Sz) {dft ## Sz,Sz},
-  struct {adftfunc_t func; int size;} dft_ftab[]={
-    FOREACH_DFTSZ(SZ_PTR)
-    };
+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)
-    };
+struct {
+  adftfunc_t func;
+  int size;
+} idft_ftab[]= {
+  FOREACH_IDFTSZ(SZ_iPTR)
+};
 
 #endif
 
@@ -378,7 +384,7 @@ int32_t add_real_vector64(int16_t *x,
                           uint32_t N);
 
 int32_t sub_real_vector64(int16_t *x,
-                          int16_t* y,
+                          int16_t *y,
                           int16_t *z,
                           uint32_t N);
 
@@ -435,7 +441,7 @@ int32_t signal_energy_amp_shift(int32_t *input, uint32_t length);
 /*!\fn int32_t signal_energy(int *,uint32_t);
 \brief Computes the signal energy per subcarrier
 */
-int32_t subcarrier_energy(int32_t *,uint32_t, int32_t* subcarrier_energy, uint16_t rx_power_correction);
+int32_t subcarrier_energy(int32_t *,uint32_t, int32_t *subcarrier_energy, uint16_t rx_power_correction);
 #endif
 
 /*!\fn int32_t signal_energy_nodc(int32_t *,uint32_t);
diff --git a/openair2/LAYER2/NR_MAC_gNB/mac_proto.h b/openair2/LAYER2/NR_MAC_gNB/mac_proto.h
index 784a71f23f..399033abd0 100644
--- a/openair2/LAYER2/NR_MAC_gNB/mac_proto.h
+++ b/openair2/LAYER2/NR_MAC_gNB/mac_proto.h
@@ -519,4 +519,7 @@ bool nr_find_nb_rb(uint16_t Qm,
 void nr_sr_reporting(int Mod_idP, frame_t frameP, sub_frame_t slotP);
 
 void dump_mac_stats(gNB_MAC_INST *gNB, char *output, int strlen, bool reset_rsrp);
+
+void process_CellGroup(NR_CellGroupConfig_t *CellGroup, NR_UE_sched_ctrl_t *sched_ctrl);
+
 #endif /*__LAYER2_NR_MAC_PROTO_H__*/

From 1a2be8fc892c3b2f2826ff3fc5c8835f6eec24e8 Mon Sep 17 00:00:00 2001
From: Laurent THOMAS <laurent.thomas@open-cells.com>
Date: Wed, 6 Apr 2022 21:48:01 +0200
Subject: [PATCH 3/4] fix dft modifications related to alignment

---
 openair1/PHY/LTE_UE_TRANSPORT/transport_ue.h  |  2 -
 .../PHY/LTE_UE_TRANSPORT/ulsch_modulation.c   | 45 +++++------
 openair1/PHY/MODULATION/slot_fep.c            |  2 +-
 openair1/PHY/TOOLS/oai_dfts.c                 | 78 +++++++++++--------
 openair1/PHY/TOOLS/tools_defs.h               |  2 +-
 5 files changed, 69 insertions(+), 60 deletions(-)

diff --git a/openair1/PHY/LTE_UE_TRANSPORT/transport_ue.h b/openair1/PHY/LTE_UE_TRANSPORT/transport_ue.h
index 6b84a5f969..b4ae44f6ea 100644
--- a/openair1/PHY/LTE_UE_TRANSPORT/transport_ue.h
+++ b/openair1/PHY/LTE_UE_TRANSPORT/transport_ue.h
@@ -155,8 +155,6 @@ typedef struct {
   uint8_t h[MAX_NUM_CHANNEL_BITS];
   /// Scrambled "b"-sequences (for definition see 36-211 V8.6 2009-03, p.14)
   uint8_t b_tilde[MAX_NUM_CHANNEL_BITS];
-  /// Modulated "d"-sequences (for definition see 36-211 V8.6 2009-03, p.14)
-  int32_t d[MAX_NUM_RE];
   /// Transform-coded "z"-sequences (for definition see 36-211 V8.6 2009-03, p.14-15)
   int32_t z[MAX_NUM_RE];
   /// "q" sequences for CQI/PMI (for definition see 36-212 V8.6 2009-03, p.27)
diff --git a/openair1/PHY/LTE_UE_TRANSPORT/ulsch_modulation.c b/openair1/PHY/LTE_UE_TRANSPORT/ulsch_modulation.c
index 7ec3e7bfb6..6a663e8940 100644
--- a/openair1/PHY/LTE_UE_TRANSPORT/ulsch_modulation.c
+++ b/openair1/PHY/LTE_UE_TRANSPORT/ulsch_modulation.c
@@ -42,7 +42,7 @@
 
 //#define DEBUG_ULSCH_MODULATION
 
-void dft_lte(int32_t *z,int32_t *d, int32_t Msc_PUSCH, uint8_t Nsymb)
+void dft_lte(int32_t *z,struct complex16 *input, int32_t Msc_PUSCH, uint8_t Nsymb)
 {
 
 #if defined(__x86_64__) || defined(__i386__)
@@ -66,7 +66,7 @@ void dft_lte(int32_t *z,int32_t *d, int32_t Msc_PUSCH, uint8_t Nsymb)
 #endif
   //  printf("Doing lte_dft for Msc_PUSCH %d\n",Msc_PUSCH);
 
-  d0 = (uint32_t *)d;
+  d0 = (uint32_t *)input;
   d1 = d0+Msc_PUSCH;
   d2 = d1+Msc_PUSCH;
   d3 = d2+Msc_PUSCH;
@@ -476,7 +476,8 @@ void ulsch_modulation(int32_t **txdataF,
   // Modulation
 
   ulsch_Msymb = G/Q_m;
-
+  /// Modulated "d"-sequences (for definition see 36-211 V8.6 2009-03, p.14)
+  struct complex16 d[MAX_NUM_RE] __attribute__((aligned(32)));
   if(ulsch->cooperation_flag == 2)
     // For Distributed Alamouti Scheme in Collabrative Communication
   {
@@ -488,14 +489,14 @@ void ulsch_modulation(int32_t **txdataF,
 
 
         //UE1, -x1*
-        ((int16_t*)&ulsch->d[i])[0] = (ulsch->b_tilde[j] == 1)  ? (gain_lin_QPSK) : -gain_lin_QPSK;
-        ((int16_t*)&ulsch->d[i])[1] = (ulsch->b_tilde[j+1] == 1)? (-gain_lin_QPSK) : gain_lin_QPSK;
+        d[i].r = (ulsch->b_tilde[j] == 1)  ? (gain_lin_QPSK) : -gain_lin_QPSK;
+        d[i].i = (ulsch->b_tilde[j+1] == 1)? (-gain_lin_QPSK) : gain_lin_QPSK;
         //      if (i<Msc_PUSCH)
         //  printf("input %d (%p): %d,%d\n", i,&ulsch->d[i],((int16_t*)&ulsch->d[i])[0],((int16_t*)&ulsch->d[i])[1]);
 
         // UE1, x0*
-        ((int16_t*)&ulsch->d[i+1])[0] = (ulsch->b_tilde[j-2] == 1)  ? (-gain_lin_QPSK) : gain_lin_QPSK;
-        ((int16_t*)&ulsch->d[i+1])[1] = (ulsch->b_tilde[j-1] == 1)? (gain_lin_QPSK) : -gain_lin_QPSK;
+        d[i+1].r = (ulsch->b_tilde[j-2] == 1)  ? (-gain_lin_QPSK) : gain_lin_QPSK;
+        d[i+1].i = (ulsch->b_tilde[j-1] == 1)? (gain_lin_QPSK) : -gain_lin_QPSK;
 
         break;
 
@@ -521,8 +522,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam16_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i])[0]=-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i])[1]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
+        d[i].r =-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
+        d[i].i =(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
 
         //UE1,x0*
         qam16_table_offset_re = 0;
@@ -544,8 +545,8 @@ void ulsch_modulation(int32_t **txdataF,
 
         //    ((int16_t*)&ulsch->d[i+1])[0]=-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
         //    ((int16_t*)&ulsch->d[i+1])[1]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
-        ((int16_t*)&ulsch->d[i+1])[0]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i+1])[1]=-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
+        d[i+1].r=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
+        d[i+1].i=-(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
 
 
         break;
@@ -578,8 +579,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam64_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i])[0]=-(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i])[1]=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
+        d[i].r=-(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
+        d[i].i=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
 
         //UE1,x0*
         qam64_table_offset_re = 0;
@@ -605,8 +606,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam64_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i+1])[0]=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i+1])[1]=-(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
+        d[i+1].r=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
+        d[i+1].i=-(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
 
         break;
 
@@ -621,8 +622,8 @@ void ulsch_modulation(int32_t **txdataF,
       case 2:
         // TODO: this has to be updated!!!
 
-        ((int16_t*)&ulsch->d[i])[0] = (ulsch->b_tilde[j] == 1)  ? (-gain_lin_QPSK) : gain_lin_QPSK;
-        ((int16_t*)&ulsch->d[i])[1] = (ulsch->b_tilde[j+1] == 1)? (-gain_lin_QPSK) : gain_lin_QPSK;
+        d[i].r = (ulsch->b_tilde[j] == 1)  ? (-gain_lin_QPSK) : gain_lin_QPSK;
+        d[i].i = (ulsch->b_tilde[j+1] == 1)? (-gain_lin_QPSK) : gain_lin_QPSK;
         //        if (i<Msc_PUSCH)
         //    printf("input %d/%d Msc_PUSCH %d (%p): %d,%d\n", i,Msymb,Msc_PUSCH,&ulsch->d[i],((int16_t*)&ulsch->d[i])[0],((int16_t*)&ulsch->d[i])[1]);
 
@@ -646,8 +647,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam16_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i])[0]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i])[1]=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
+        d[i].r=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_re])>>15);
+        d[i].i=(int16_t)(((int32_t)amp*qam16_table[qam16_table_offset_im])>>15);
         //      printf("input(16qam) %d (%p): %d,%d\n", i,&ulsch->d[i],((int16_t*)&ulsch->d[i])[0],((int16_t*)&ulsch->d[i])[1]);
         break;
 
@@ -676,8 +677,8 @@ void ulsch_modulation(int32_t **txdataF,
           qam64_table_offset_im+=1;
 
 
-        ((int16_t*)&ulsch->d[i])[0]=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
-        ((int16_t*)&ulsch->d[i])[1]=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
+        d[i].r=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_re])>>15);
+        d[i].i=(int16_t)(((int32_t)amp*qam64_table[qam64_table_offset_im])>>15);
 
         break;
 
@@ -688,7 +689,7 @@ void ulsch_modulation(int32_t **txdataF,
 
   // Transform Precoding
 
-  dft_lte(ulsch->z,ulsch->d,Msc_PUSCH,ulsch->Nsymb_pusch);
+  dft_lte(ulsch->z,d,Msc_PUSCH,ulsch->Nsymb_pusch);
 
   DevAssert(txdataF);
 
diff --git a/openair1/PHY/MODULATION/slot_fep.c b/openair1/PHY/MODULATION/slot_fep.c
index 8f41735f10..b2c7d934c6 100644
--- a/openair1/PHY/MODULATION/slot_fep.c
+++ b/openair1/PHY/MODULATION/slot_fep.c
@@ -91,7 +91,7 @@ int slot_fep(PHY_VARS_UE *ue,
   }
 
   //  subframe_offset_F = frame_parms->ofdm_symbol_size * frame_parms->symbols_per_tti * (Ns>>1);
-
+ 
   if (l<0 || l>=7-frame_parms->Ncp) {
     printf("slot_fep: l must be between 0 and %d\n",7-frame_parms->Ncp);
     return(-1);
diff --git a/openair1/PHY/TOOLS/oai_dfts.c b/openair1/PHY/TOOLS/oai_dfts.c
index c266a7f82b..c2b6a55210 100644
--- a/openair1/PHY/TOOLS/oai_dfts.c
+++ b/openair1/PHY/TOOLS/oai_dfts.c
@@ -7082,43 +7082,40 @@ static inline void dft12f(simd_q15_t *x0,
 
   simd_q15_t tmp_dft12[12];
 
-  simd_q15_t *tmp_dft12_ptr = &tmp_dft12[0];
-
   // msg("dft12\n");
 
   bfly4_tw1(x0,
             x3,
             x6,
             x9,
-            tmp_dft12_ptr,
-            tmp_dft12_ptr+3,
-            tmp_dft12_ptr+6,
-            tmp_dft12_ptr+9);
-
+            tmp_dft12,
+            tmp_dft12+3,
+            tmp_dft12+6,
+            tmp_dft12+9);
 
   bfly4_tw1(x1,
             x4,
             x7,
             x10,
-            tmp_dft12_ptr+1,
-            tmp_dft12_ptr+4,
-            tmp_dft12_ptr+7,
-            tmp_dft12_ptr+10);
+            tmp_dft12+1,
+            tmp_dft12+4,
+            tmp_dft12+7,
+            tmp_dft12+10);
 
 
   bfly4_tw1(x2,
             x5,
             x8,
             x11,
-            tmp_dft12_ptr+2,
-            tmp_dft12_ptr+5,
-            tmp_dft12_ptr+8,
-            tmp_dft12_ptr+11);
+            tmp_dft12+2,
+            tmp_dft12+5,
+            tmp_dft12+8,
+            tmp_dft12+11);
 
   //  k2=0;
-  bfly3_tw1(tmp_dft12_ptr,
-            tmp_dft12_ptr+1,
-            tmp_dft12_ptr+2,
+  bfly3_tw1(tmp_dft12,
+            tmp_dft12+1,
+            tmp_dft12+2,
             y0,
             y4,
             y8);
@@ -7126,9 +7123,9 @@ static inline void dft12f(simd_q15_t *x0,
 
 
   //  k2=1;
-  bfly3(tmp_dft12_ptr+3,
-        tmp_dft12_ptr+4,
-        tmp_dft12_ptr+5,
+  bfly3(tmp_dft12+3,
+        tmp_dft12+4,
+        tmp_dft12+5,
         y1,
         y5,
         y9,
@@ -7138,9 +7135,9 @@ static inline void dft12f(simd_q15_t *x0,
 
 
   //  k2=2;
-  bfly3(tmp_dft12_ptr+6,
-        tmp_dft12_ptr+7,
-        tmp_dft12_ptr+8,
+  bfly3(tmp_dft12+6,
+        tmp_dft12+7,
+        tmp_dft12+8,
         y2,
         y6,
         y10,
@@ -7148,9 +7145,9 @@ static inline void dft12f(simd_q15_t *x0,
         W4_12);
 
   //  k2=3;
-  bfly3(tmp_dft12_ptr+9,
-        tmp_dft12_ptr+10,
-        tmp_dft12_ptr+11,
+  bfly3(tmp_dft12+9,
+        tmp_dft12+10,
+        tmp_dft12+11,
         y3,
         y7,
         y11,
@@ -10607,11 +10604,19 @@ int dfts_autoinit(void)
 #ifndef MR_MAIN
 
 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) {
+	AssertFatal((sizeidx >= 0 && sizeidx<DFT_SIZE_IDXTABLESIZE),"Invalid dft size index %i\n",sizeidx);
+        int algn=0xF;
+        #ifdef __AVX2__
+        if ( (dft_ftab[sizeidx].size%3) != 0 ) // there is no AVX2 implementation for multiples of 3 DFTs
+          algn=0x1F;
+        #endif 
+        AssertFatal(((intptr_t)output&algn)==0,"Buffers should be aligned %p",output);
+        if (((intptr_t)input)&algn) {
           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)
+          int sz=dft_ftab[sizeidx].size;
+          if (sizeidx==DFT_12) // This case does 8 DFTs in //
+            sz*=8;
+          int16_t tmp[sz*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
@@ -10620,10 +10625,15 @@ void dft(uint8_t sizeidx, int16_t *input,int16_t *output,unsigned char scale_fla
 
 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) {
+        int algn=0xF;
+        #ifdef __AVX2__
+          algn=0x1F;
+        #endif
+        AssertFatal( ((intptr_t)output&algn)==0,"Buffers should be 16 bytes aligned %p",output);
+        if (((intptr_t)input)&algn ) {  
           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)
+          int sz=idft_ftab[sizeidx].size;
+          int16_t tmp[sz*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
diff --git a/openair1/PHY/TOOLS/tools_defs.h b/openair1/PHY/TOOLS/tools_defs.h
index 7cee04fc59..86bce9614f 100644
--- a/openair1/PHY/TOOLS/tools_defs.h
+++ b/openair1/PHY/TOOLS/tools_defs.h
@@ -253,7 +253,7 @@ This function performs optimized fixed-point radix-2 FFT/IFFT.
   SZ_DEF(73728) \
   SZ_DEF(98304)
 
-#define FOREACH_IDFTSZ(SZ_DEF)                  \
+#define FOREACH_IDFTSZ(SZ_DEF)\
   SZ_DEF(64) \
   SZ_DEF(128) \
   SZ_DEF(256) \

From 8623bc6690b6f7e9e1aed9580bc66822f382bccc Mon Sep 17 00:00:00 2001
From: Laurent Thomas <laurent.thomas@open-cells.com>
Date: Thu, 7 Apr 2022 12:05:50 +0200
Subject: [PATCH 4/4] a small out of buffer access, likely only santize addr
 detection

---
 openair1/PHY/NR_TRANSPORT/nr_dlsch.c | 81 ++++++++++++++--------------
 1 file changed, 39 insertions(+), 42 deletions(-)

diff --git a/openair1/PHY/NR_TRANSPORT/nr_dlsch.c b/openair1/PHY/NR_TRANSPORT/nr_dlsch.c
index 47d6c48373..e242f989ea 100644
--- a/openair1/PHY/NR_TRANSPORT/nr_dlsch.c
+++ b/openair1/PHY/NR_TRANSPORT/nr_dlsch.c
@@ -98,7 +98,6 @@ void nr_generate_pdsch(processingData_L1tx_t *msgTx,
     uint16_t nb_re = ((12*rel15->NrOfSymbols)-nb_re_dmrs*dmrs_len-xOverhead)*rel15->rbSize*rel15->nrOfLayers;
     uint8_t Qm = rel15->qamModOrder[0];
     uint32_t encoded_length = nb_re*Qm;
-    uint32_t scrambled_output[rel15->NrOfCodewords][(encoded_length>>5)+1];
     int16_t mod_dmrs[n_dmrs<<1] __attribute__ ((aligned(16)));
 
     /* PTRS */
@@ -147,63 +146,61 @@ void nr_generate_pdsch(processingData_L1tx_t *msgTx,
     printf("\n");
 #endif
 
-    /// scrambling
-    start_meas(dlsch_scrambling_stats);
     for (int q=0; q<rel15->NrOfCodewords; q++) {
-      memset((void*)scrambled_output[q], 0, ((encoded_length>>5)+1)*sizeof(uint32_t));
+      /// scrambling
+      start_meas(dlsch_scrambling_stats);
+      uint32_t scrambled_output[(encoded_length>>5)+4]; // modulator acces by 4 bytes in some cases
+      memset(scrambled_output, 0, sizeof(scrambled_output));
+      if ( encoded_length > rel15->rbSize * NR_SYMBOLS_PER_SLOT * NR_NB_SC_PER_RB * Qm * rel15->nrOfLayers) abort();
       nr_pdsch_codeword_scrambling(output,
                                    encoded_length,
                                    q,
                                    rel15->dataScramblingId,
                                    rel15->rnti,
-                                   scrambled_output[q]);
-    }
-    
-    stop_meas(dlsch_scrambling_stats);
+                                   scrambled_output);
 #ifdef DEBUG_DLSCH
-    printf("PDSCH scrambling:\n");
-    for (int i=0; i<encoded_length>>8; i++) {
-      for (int j=0; j<8; j++)
-	printf("0x%08x\t", scrambled_output[0][(i<<3)+j]);
-      printf("\n");
-    }
-#endif
-    
-    /// Modulation
-    start_meas(dlsch_modulation_stats);
-    VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_gNB_PDSCH_MODULATION, 1);
-    for (int q=0; q<rel15->NrOfCodewords; q++)
-      nr_modulation(scrambled_output[q],
-		    encoded_length,
-		    Qm,
-		    mod_symbs[q]);
-    VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_gNB_PDSCH_MODULATION, 0);
-    stop_meas(dlsch_modulation_stats);
-#ifdef DEBUG_DLSCH
-    printf("PDSCH Modulation: Qm %d(%d)\n", Qm, nb_re);
-    for (int i=0; i<nb_re>>3; i++) {
-      for (int j=0; j<8; j++) {
-	printf("%d %d\t", mod_symbs[0][((i<<3)+j)<<1], mod_symbs[0][(((i<<3)+j)<<1)+1]);
+      printf("PDSCH scrambling:\n");
+      for (int i=0; i<encoded_length>>8; i++) {
+        for (int j=0; j<8; j++)
+          printf("0x%08x\t", scrambled_output[(i<<3)+j]);
+        printf("\n");
       }
-      printf("\n");
-    }
 #endif
+      stop_meas(dlsch_scrambling_stats);
+      /// Modulation
+      start_meas(dlsch_modulation_stats);
+      nr_modulation(scrambled_output,
+                    encoded_length,
+                    Qm,
+                    mod_symbs[q]);
+      VCD_SIGNAL_DUMPER_DUMP_FUNCTION_BY_NAME(VCD_SIGNAL_DUMPER_FUNCTIONS_gNB_PDSCH_MODULATION, 0);
+      stop_meas(dlsch_modulation_stats);
+#ifdef DEBUG_DLSCH
+      printf("PDSCH Modulation: Qm %d(%d)\n", Qm, nb_re);
+      for (int i=0; i<nb_re>>3; i++) {
+        for (int j=0; j<8; j++) {
+          printf("%d %d\t", mod_symbs[0][((i<<3)+j)<<1], mod_symbs[0][(((i<<3)+j)<<1)+1]);
+        }
+        printf("\n");
+      }
+#endif
+    }
     
     start_meas(&gNB->dlsch_layer_mapping_stats); 
     /// Layer mapping
     nr_layer_mapping(mod_symbs,
-		     rel15->nrOfLayers,
-		     nb_re,
-		     tx_layers);
+                     rel15->nrOfLayers,
+                     nb_re,
+                     tx_layers);
 #ifdef DEBUG_DLSCH
     printf("Layer mapping (%d layers):\n", rel15->nrOfLayers);
     for (int l=0; l<rel15->nrOfLayers; l++)
       for (int i=0; i<(nb_re/rel15->nrOfLayers)>>3; i++) {
-	printf("layer %d, Re %d..%d : ",l,i<<3,(i<<3)+7);
-	for (int j=0; j<8; j++) {
-	  printf("l%d %d\t", tx_layers[l][((i<<3)+j)<<1], tx_layers[l][(((i<<3)+j)<<1)+1]);
-	}
-	printf("\n");
+        printf("layer %d, Re %d..%d : ",l,i<<3,(i<<3)+7);
+        for (int j=0; j<8; j++) {
+          printf("l%d %d\t", tx_layers[l][((i<<3)+j)<<1], tx_layers[l][(((i<<3)+j)<<1)+1]);
+        }
+        printf("\n");
       }
 #endif
 
@@ -222,7 +219,7 @@ void nr_generate_pdsch(processingData_L1tx_t *msgTx,
 
 #ifdef DEBUG_DLSCH_MAPPING
     printf("PDSCH resource mapping started (start SC %d\tstart symbol %d\tN_PRB %d\tnb_re %d,nb_layers %d)\n",
-	   start_sc, rel15->StartSymbolIndex, rel15->rbSize, nb_re,rel15->nrOfLayers);
+           start_sc, rel15->StartSymbolIndex, rel15->rbSize, nb_re,rel15->nrOfLayers);
 #endif
 
     start_meas(&gNB->dlsch_resource_mapping_stats);