From 06d488bb23ac6d17e72eda8fa0a6a5f75fdb2644 Mon Sep 17 00:00:00 2001 From: francescomani Date: Wed, 9 Mar 2022 17:35:38 +0100 Subject: [PATCH] ulsch memory optimization --- openair1/PHY/INIT/nr_init_ue.c | 4 +- openair1/PHY/NR_TRANSPORT/nr_ulsch_decoding.c | 26 ++++++---- .../NR_UE_TRANSPORT/nr_transport_proto_ue.h | 6 ++- .../PHY/NR_UE_TRANSPORT/nr_transport_ue.h | 4 +- .../PHY/NR_UE_TRANSPORT/nr_ulsch_coding.c | 47 ++++++++++++------- openair1/PHY/defs_gNB.h | 10 ++-- openair1/PHY/defs_nr_common.h | 3 +- openair1/SIMULATION/NR_PHY/ulschsim.c | 4 +- openair2/LAYER2/NR_MAC_UE/nr_ra_procedures.c | 2 +- 9 files changed, 61 insertions(+), 45 deletions(-) diff --git a/openair1/PHY/INIT/nr_init_ue.c b/openair1/PHY/INIT/nr_init_ue.c index 6626d4614c..5a26bfcfac 100644 --- a/openair1/PHY/INIT/nr_init_ue.c +++ b/openair1/PHY/INIT/nr_init_ue.c @@ -684,7 +684,7 @@ void term_nr_ue_transport(PHY_VARS_NR_UE *ue) for (int j = 0; j < 2; j++) { for (int k = 0; k < RX_NB_TH_MAX; k++) { free_nr_ue_dlsch(&ue->dlsch[k][i][j], N_RB_DL); - free_nr_ue_ulsch(&ue->ulsch[k][i][j], N_RB_DL); + free_nr_ue_ulsch(&ue->ulsch[k][i][j], N_RB_DL, &ue->frame_parms); } } @@ -702,7 +702,7 @@ void init_nr_ue_transport(PHY_VARS_NR_UE *ue) for (int k=0; kdlsch[k][i][j] = new_nr_ue_dlsch(1,NR_MAX_DLSCH_HARQ_PROCESSES,NSOFT,MAX_LDPC_ITERATIONS,ue->frame_parms.N_RB_DL))!=NULL,"Can't get ue dlsch structures\n"); LOG_D(PHY,"dlsch[%d][%d][%d] => %p\n",k,i,j,ue->dlsch[k][i][j]); - AssertFatal((ue->ulsch[k][i][j] = new_nr_ue_ulsch(ue->frame_parms.N_RB_UL, NR_MAX_ULSCH_HARQ_PROCESSES))!=NULL,"Can't get ue ulsch structures\n"); + AssertFatal((ue->ulsch[k][i][j] = new_nr_ue_ulsch(ue->frame_parms.N_RB_UL, NR_MAX_ULSCH_HARQ_PROCESSES,&ue->frame_parms))!=NULL,"Can't get ue ulsch structures\n"); LOG_D(PHY,"ulsch[%d][%d][%d] => %p\n",k,i,j,ue->ulsch[k][i][j]); } } diff --git a/openair1/PHY/NR_TRANSPORT/nr_ulsch_decoding.c b/openair1/PHY/NR_TRANSPORT/nr_ulsch_decoding.c index 304dad96f0..4a824dc678 100644 --- a/openair1/PHY/NR_TRANSPORT/nr_ulsch_decoding.c +++ b/openair1/PHY/NR_TRANSPORT/nr_ulsch_decoding.c @@ -57,8 +57,7 @@ void free_gNB_ulsch(NR_gNB_ULSCH_t **ulschptr, uint16_t N_RB_UL) { - int i,r; - uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS; //number of segments to be allocated + uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS_PER_LAYER*NR_MAX_NB_LAYERS; //number of segments to be allocated NR_gNB_ULSCH_t *ulsch = *ulschptr; if (N_RB_UL != 273) { @@ -66,19 +65,21 @@ void free_gNB_ulsch(NR_gNB_ULSCH_t **ulschptr, uint16_t N_RB_UL) a_segments = a_segments/273 +1; } - for (i=0; iharq_processes[i]) { if (ulsch->harq_processes[i]->b) { free_and_zero(ulsch->harq_processes[i]->b); ulsch->harq_processes[i]->b = NULL; } - for (r=0; rharq_processes[i]->c[r]); free_and_zero(ulsch->harq_processes[i]->d[r]); - free_and_zero(ulsch->harq_processes[i]->w[r]); nrLDPC_free_mem(ulsch->harq_processes[i]->p_nrLDPC_procBuf[r]); ulsch->harq_processes[i]->p_nrLDPC_procBuf[r] = NULL; } + free_and_zero(ulsch->harq_processes[i]->c); + free_and_zero(ulsch->harq_processes[i]->d); + free_and_zero(ulsch->harq_processes[i]->p_nrLDPC_procBuf); free_and_zero(ulsch->harq_processes[i]); ulsch->harq_processes[i] = NULL; } @@ -92,7 +93,7 @@ NR_gNB_ULSCH_t *new_gNB_ulsch(uint8_t max_ldpc_iterations, uint16_t N_RB_UL) NR_gNB_ULSCH_t *ulsch; uint8_t i,r; - uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS; //number of segments to be allocated + uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS_PER_LAYER*NR_MAX_NB_LAYERS; //number of segments to be allocated if (N_RB_UL != 273) { a_segments = a_segments*N_RB_UL; @@ -109,11 +110,13 @@ NR_gNB_ULSCH_t *new_gNB_ulsch(uint8_t max_ldpc_iterations, uint16_t N_RB_UL) ulsch->harq_processes[i] = (NR_UL_gNB_HARQ_t *)malloc16_clear(sizeof(NR_UL_gNB_HARQ_t)); ulsch->harq_processes[i]->b = (uint8_t*)malloc16_clear(ulsch_bytes); + ulsch->harq_processes[i]->c = (uint8_t**)malloc16_clear(a_segments); + ulsch->harq_processes[i]->d = (int16_t**)malloc16_clear(a_segments); + ulsch->harq_processes[i]->p_nrLDPC_procBuf = (t_nrLDPC_procBuf **)malloc16_clear(a_segments); for (r=0; rharq_processes[i]->p_nrLDPC_procBuf[r] = nrLDPC_init_mem(); ulsch->harq_processes[i]->c[r] = (uint8_t*)malloc16_clear(8448*sizeof(uint8_t)); ulsch->harq_processes[i]->d[r] = (int16_t*)malloc16_clear((68*384)*sizeof(int16_t)); - ulsch->harq_processes[i]->w[r] = (int16_t*)malloc16_clear((3*(6144+64))*sizeof(int16_t)); } } @@ -265,9 +268,12 @@ void nr_processULSegment(void* arg) { //////////////////////////// ulsch_llr =====> ulsch_harq->e ////////////////////////////// + /// code blocks after bit selection in rate matching for LDPC code (38.212 V15.4.0 section 5.4.2.1) + int16_t harq_e[3*8448]; + nr_deinterleaving_ldpc(E, Qm, - ulsch_harq->e[r], + harq_e, ulsch_llr+r_offset); //for (int i =0; i<16; i++) @@ -292,7 +298,7 @@ void nr_processULSegment(void* arg) { p_decoderParms->BG, p_decoderParms->Z, ulsch_harq->d[r], - ulsch_harq->e[r], + harq_e, ulsch_harq->C, rv_index, ulsch_harq->new_rx, @@ -539,7 +545,7 @@ uint32_t nr_ulsch_decoding(PHY_VARS_gNB *phy_vars_gNB, r_offset = 0; - uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS; //number of segments to be allocated + uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS_PER_LAYER*n_layers; //number of segments to be allocated if (nb_rb != 273) { a_segments = a_segments*nb_rb; diff --git a/openair1/PHY/NR_UE_TRANSPORT/nr_transport_proto_ue.h b/openair1/PHY/NR_UE_TRANSPORT/nr_transport_proto_ue.h index 9bc3643557..822edcc3d9 100644 --- a/openair1/PHY/NR_UE_TRANSPORT/nr_transport_proto_ue.h +++ b/openair1/PHY/NR_UE_TRANSPORT/nr_transport_proto_ue.h @@ -63,10 +63,12 @@ void free_nr_ue_dlsch(NR_UE_DLSCH_t **dlsch, uint16_t N_RB_DL); */ NR_UE_DLSCH_t *new_nr_ue_dlsch(uint8_t Kmimo,uint8_t Mdlharq,uint32_t Nsoft,uint8_t max_turbo_iterations,uint16_t N_RB_DL); +void free_nr_ue_ulsch(NR_UE_ULSCH_t **ulschptr, + uint16_t N_RB_UL, + NR_DL_FRAME_PARMS* frame_parms); -void free_nr_ue_ulsch(NR_UE_ULSCH_t **ulsch, uint16_t N_RB_UL); -NR_UE_ULSCH_t *new_nr_ue_ulsch(uint16_t N_RB_UL, int number_of_harq_pids); +NR_UE_ULSCH_t *new_nr_ue_ulsch(uint16_t N_RB_UL, int number_of_harq_pids, NR_DL_FRAME_PARMS* frame_parms); /** \brief This function computes the LLRs for ML (max-logsum approximation) dual-stream QPSK/QPSK reception. @param stream0_in Input from channel compensated (MR combined) stream 0 diff --git a/openair1/PHY/NR_UE_TRANSPORT/nr_transport_ue.h b/openair1/PHY/NR_UE_TRANSPORT/nr_transport_ue.h index caf3a19d85..9d5beb9d20 100644 --- a/openair1/PHY/NR_UE_TRANSPORT/nr_transport_ue.h +++ b/openair1/PHY/NR_UE_TRANSPORT/nr_transport_ue.h @@ -84,9 +84,9 @@ typedef struct { /// Pointer to the payload + CRC uint8_t *b; /// Pointers to transport block segments - uint8_t *c[MAX_NUM_NR_ULSCH_SEGMENTS]; + uint8_t **c; /// LDPC-code outputs - uint8_t *d[MAX_NUM_NR_ULSCH_SEGMENTS]; + uint8_t **d; /// LDPC-code outputs (TS 36.212 V15.4.0, Sec 5.3.2 p. 17) uint8_t *e; /// Rate matching (Interleaving) outputs (TS 36.212 V15.4.0, Sec 5.4.2.2 p. 30) diff --git a/openair1/PHY/NR_UE_TRANSPORT/nr_ulsch_coding.c b/openair1/PHY/NR_UE_TRANSPORT/nr_ulsch_coding.c index 5160a643fb..b1966aaa43 100644 --- a/openair1/PHY/NR_UE_TRANSPORT/nr_ulsch_coding.c +++ b/openair1/PHY/NR_UE_TRANSPORT/nr_ulsch_coding.c @@ -43,10 +43,10 @@ //#define DEBUG_ULSCH_CODING -void free_nr_ue_ulsch(NR_UE_ULSCH_t **ulschptr, uint16_t N_RB_UL) +void free_nr_ue_ulsch(NR_UE_ULSCH_t **ulschptr, + uint16_t N_RB_UL, + NR_DL_FRAME_PARMS* frame_parms) { -{ - int i, r; NR_UE_ULSCH_t *ulsch = *ulschptr; if (ulsch) { @@ -54,15 +54,15 @@ void free_nr_ue_ulsch(NR_UE_ULSCH_t **ulschptr, uint16_t N_RB_UL) printf("Freeing ulsch %p\n",ulsch); #endif - uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS; //number of segments to be allocated + int max_layers = (frame_parms->nb_antennas_txnb_antennas_tx : NR_MAX_NB_LAYERS; + uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS_PER_LAYER*max_layers; //number of segments to be allocated - if (N_RB_UL != 273) { - a_segments = a_segments*N_RB_UL; - a_segments = a_segments/273 +1; - } + if (N_RB_UL != 273) { + a_segments = a_segments*N_RB_UL; + a_segments = a_segments/273 +1; + } - - for (i=0; iharq_processes[i]) { if (ulsch->harq_processes[i]->a) { @@ -81,7 +81,7 @@ void free_nr_ue_ulsch(NR_UE_ULSCH_t **ulschptr, uint16_t N_RB_UL) free16(ulsch->harq_processes[i]->f,14*N_RB_UL*12*8); ulsch->harq_processes[i]->f = NULL; } - for (r=0; rharq_processes[i]->c[r]) { free16(ulsch->harq_processes[i]->c[r],((r==0)?8:0) + 3+768); ulsch->harq_processes[i]->c[r] = NULL; @@ -94,6 +94,15 @@ void free_nr_ue_ulsch(NR_UE_ULSCH_t **ulschptr, uint16_t N_RB_UL) } + if (ulsch->harq_processes[i]->c) { + free16(ulsch->harq_processes[i]->c,a_segments); + ulsch->harq_processes[i]->c = NULL; + } + if (ulsch->harq_processes[i]->d) { + free16(ulsch->harq_processes[i]->d,a_segments); + ulsch->harq_processes[i]->d = NULL; + } + free16(ulsch->harq_processes[i],sizeof(NR_UL_UE_HARQ_t)); ulsch->harq_processes[i] = NULL; } @@ -101,13 +110,13 @@ void free_nr_ue_ulsch(NR_UE_ULSCH_t **ulschptr, uint16_t N_RB_UL) free16(ulsch,sizeof(NR_UE_ULSCH_t)); *ulschptr = NULL; } - } -NR_UE_ULSCH_t *new_nr_ue_ulsch(uint16_t N_RB_UL, int number_of_harq_pids) -{ - uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS; //number of segments to be allocated +NR_UE_ULSCH_t *new_nr_ue_ulsch(uint16_t N_RB_UL, int number_of_harq_pids, NR_DL_FRAME_PARMS* frame_parms) { + + int max_layers = (frame_parms->nb_antennas_txnb_antennas_tx : NR_MAX_NB_LAYERS; + uint16_t a_segments = MAX_NUM_NR_ULSCH_SEGMENTS_PER_LAYER*max_layers; //number of segments to be allocated if (N_RB_UL != 273) { a_segments = a_segments*N_RB_UL; @@ -140,6 +149,8 @@ NR_UE_ULSCH_t *new_nr_ue_ulsch(uint16_t N_RB_UL, int number_of_harq_pids) DevAssert(ulsch->harq_processes[i]->b); bzero(ulsch->harq_processes[i]->b,ulsch_bytes); + ulsch->harq_processes[i]->c = malloc16(a_segments); + ulsch->harq_processes[i]->d = malloc16(a_segments); for (int r = 0; r < a_segments; r++) { // account for filler in first segment and CRCs for multiple segment case ulsch->harq_processes[i]->c[r] = malloc16(8448); @@ -240,6 +251,8 @@ int nr_ulsch_encoding(PHY_VARS_NR_UE *ue, printf("\n"); */ + int max_payload_bytes = MAX_NUM_NR_ULSCH_SEGMENTS_PER_LAYER*harq_process->pusch_pdu.nrOfLayers*1056; + if (A > 3824) { // Add 24-bit crc (polynomial A) to payload crc = crc24a(harq_process->a,A)>>8; @@ -251,7 +264,7 @@ int nr_ulsch_encoding(PHY_VARS_NR_UE *ue, harq_process->B = A+24; - AssertFatal((A/8)+4 <= MAX_NR_ULSCH_PAYLOAD_BYTES,"A %d is too big (A/8+4 = %d > %d)\n",A,(A/8)+4,MAX_NR_ULSCH_PAYLOAD_BYTES); + AssertFatal((A/8)+4 <= max_payload_bytes,"A %d is too big (A/8+4 = %d > %d)\n",A,(A/8)+4,max_payload_bytes); memcpy(harq_process->b,harq_process->a,(A/8)+4); } @@ -265,7 +278,7 @@ int nr_ulsch_encoding(PHY_VARS_NR_UE *ue, harq_process->B = A+16; - AssertFatal((A/8)+3 <= MAX_NR_ULSCH_PAYLOAD_BYTES,"A %d is too big (A/8+3 = %d > %d)\n",A,(A/8)+3,MAX_NR_ULSCH_PAYLOAD_BYTES); + AssertFatal((A/8)+3 <= max_payload_bytes,"A %d is too big (A/8+3 = %d > %d)\n",A,(A/8)+3,max_payload_bytes); memcpy(harq_process->b,harq_process->a,(A/8)+3); // using 3 bytes to mimic the case of 24 bit crc } diff --git a/openair1/PHY/defs_gNB.h b/openair1/PHY/defs_gNB.h index 297b58e88e..2bd0bd301a 100644 --- a/openair1/PHY/defs_gNB.h +++ b/openair1/PHY/defs_gNB.h @@ -273,7 +273,7 @@ typedef struct { /// The payload + CRC (24 bits) in bits (38.212 V15.4.0 section 5.1) uint32_t B; /// Pointers to code blocks after code block segmentation and CRC attachment (38.212 V15.4.0 section 5.2.2) - uint8_t *c[MAX_NUM_NR_ULSCH_SEGMENTS]; + uint8_t **c; /// Number of bits in each code block (38.212 V15.4.0 section 5.2.2) uint32_t K; /// Number of "Filler" bits added in the code block segmentation (38.212 V15.4.0 section 5.2.2) @@ -281,13 +281,11 @@ typedef struct { /// Number of code blocks after code block segmentation (38.212 V15.4.0 section 5.2.2) uint32_t C; /// Pointers to code blocks after LDPC coding (38.212 V15.4.0 section 5.3.2) - int16_t *d[MAX_NUM_NR_ULSCH_SEGMENTS]; + int16_t **d; /// LDPC processing buffer - t_nrLDPC_procBuf* p_nrLDPC_procBuf[MAX_NUM_NR_ULSCH_SEGMENTS]; + t_nrLDPC_procBuf** p_nrLDPC_procBuf; /// LDPC lifting size (38.212 V15.4.0 table 5.3.2-1) uint32_t Z; - /// code blocks after bit selection in rate matching for LDPC code (38.212 V15.4.0 section 5.4.2.1) - int16_t e[MAX_NUM_NR_DLSCH_SEGMENTS][3*8448]; /// Number of bits in each code block after rate matching for LDPC code (38.212 V15.4.0 section 5.4.2.1) uint32_t E; /// Number of segments processed so far @@ -342,8 +340,6 @@ typedef struct { int16_t q_RI[MAX_RI_PAYLOAD]; /// Temporary h sequence to flag PUSCH_x/PUSCH_y symbols which are not scrambled uint8_t h[MAX_NUM_CHANNEL_BITS]; - /// soft bits for each received segment ("w"-sequence)(for definition see 36-212 V8.6 2009-03, p.15) - int16_t *w[MAX_NUM_NR_ULSCH_SEGMENTS]; ////////////////////////////////////////////////////////////// } NR_UL_gNB_HARQ_t; diff --git a/openair1/PHY/defs_nr_common.h b/openair1/PHY/defs_nr_common.h index 5574be705f..a912a896af 100644 --- a/openair1/PHY/defs_nr_common.h +++ b/openair1/PHY/defs_nr_common.h @@ -104,8 +104,7 @@ #define MAX_NUM_NR_DLSCH_SEGMENTS (NR_MAX_NB_LAYERS*34) #define MAX_NR_DLSCH_PAYLOAD_BYTES (MAX_NUM_NR_DLSCH_SEGMENTS*1056) -#define MAX_NUM_NR_ULSCH_SEGMENTS (NR_MAX_NB_LAYERS*34) -#define MAX_NR_ULSCH_PAYLOAD_BYTES (MAX_NUM_NR_ULSCH_SEGMENTS*1056) +#define MAX_NUM_NR_ULSCH_SEGMENTS_PER_LAYER 34 #define MAX_NUM_NR_CHANNEL_BITS (4*14*273*12*8) // 14 symbols, 273 RB #define MAX_NUM_NR_RE (4*14*273*12) diff --git a/openair1/SIMULATION/NR_PHY/ulschsim.c b/openair1/SIMULATION/NR_PHY/ulschsim.c index 11b94a2221..1a6c788c8d 100644 --- a/openair1/SIMULATION/NR_PHY/ulschsim.c +++ b/openair1/SIMULATION/NR_PHY/ulschsim.c @@ -417,7 +417,7 @@ int main(int argc, char **argv) for (sf = 0; sf < 2; sf++) { for (i = 0; i < 2; i++) { - UE->ulsch[sf][0][i] = new_nr_ue_ulsch(N_RB_UL, 8); + UE->ulsch[sf][0][i] = new_nr_ue_ulsch(N_RB_UL, 8, frame_parms); if (!UE->ulsch[sf][0][i]) { printf("Can't get ue ulsch structures.\n"); @@ -628,7 +628,7 @@ int main(int argc, char **argv) for (sf = 0; sf < 2; sf++) for (i = 0; i < 2; i++) - free_nr_ue_ulsch(&UE->ulsch[sf][0][i], N_RB_UL); + free_nr_ue_ulsch(&UE->ulsch[sf][0][i], N_RB_UL, frame_parms); term_nr_ue_signal(UE, 1); free(UE); diff --git a/openair2/LAYER2/NR_MAC_UE/nr_ra_procedures.c b/openair2/LAYER2/NR_MAC_UE/nr_ra_procedures.c index 3851ee011d..1c9ebf3eba 100644 --- a/openair2/LAYER2/NR_MAC_UE/nr_ra_procedures.c +++ b/openair2/LAYER2/NR_MAC_UE/nr_ra_procedures.c @@ -743,7 +743,7 @@ uint8_t nr_ue_get_rach(NR_PRACH_RESOURCES_t *prach_resources, } else if (get_softmodem_params()->nsa) { - uint8_t mac_sdus[MAX_NR_ULSCH_PAYLOAD_BYTES]; + uint8_t mac_sdus[MAX_NUM_NR_ULSCH_SEGMENTS_PER_LAYER*1056]; uint16_t sdu_lengths[NB_RB_MAX] = {0}; int TBS_bytes = 848; int mac_ce_len = 0;