maki/openairinterface5g
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

remove a race condition in viberti_lte and a useless global table

Browse Source
This commit is contained in:
Laurent THOMAS
2025-01-14 10:07:30 +01:00
parent a0d6057294
commit d5c72cf843
5 changed files with 68 additions and 100 deletions
Download Patch File Download Diff File Expand all files Collapse all files
openair1/PHY/CODING/TESTBENCH/pdcch_test.c
+1 -2
View File
@@ -246,9 +246,8 @@ int main(int argc, char *argv[])
set_taus_seed(0);
ccodelte_init();
ccodelte_init_inv();
phy_generate_viterbi_tables_lte();
phy_generate_viterbi_tables_lte(false);
lte_frame_parms = &(PHY_config->lte_frame_parms);
lte_frame_parms->N_RB_DL = 25; //50 for 10MHz and 25 for 5 MHz
lte_frame_parms->Ncp = 1;
openair1/PHY/CODING/ccoding_byte_lte.c
+5 -8
View File
@@ -32,9 +32,7 @@ static const unsigned short glte[] = {0133, 0171, 0165}; // {A,B}
static const unsigned short glte_rev[] = {0155, 0117, 0127}; // {A,B}
static const unsigned short gdab[] = {0133, 0171, 0145}; // {A,B}
static const unsigned short gdab_rev[] = {0155, 0117, 0123}; // {A,B}
unsigned char ccodelte_table[128]; // for transmitter
unsigned char ccodelte_table_rev[128]; // for receiver
static unsigned char ccodelte_table[128]; // for transmitter
/*************************************************************************
@@ -45,8 +43,6 @@ unsigned char ccodelte_table_rev[128]; // for receiver
Trellis tail-biting is included here
*************************************************************************/
void
ccodelte_encode (int32_t numbits,
uint8_t add_crc,
@@ -238,7 +234,8 @@ void ccodelte_init(void) {
}
/* Input in LSB, followed by state in 6 MSBs */
void ccodelte_init_inv(void) {
void ccodelte_init_inv(unsigned char ccodelte_table_rev[128])
{
unsigned int i, j, k, sum;
for (i = 0; i < 128; i++) {
@@ -279,7 +276,8 @@ void ccodedab_init(void) {
}
/* Input in LSB, followed by state in 6 MSBs */
void ccodedab_init_inv(void) {
void ccodedab_init_inv(unsigned char ccodelte_table_rev[128])
{
unsigned int i, j, k, sum;
for (i = 0; i < 128; i++) {
@@ -299,7 +297,6 @@ void ccodedab_init_inv(void) {
}
}
/*****************************************************************/
/**
Test program
openair1/PHY/CODING/coding_defs.h
+5 -5
View File
@@ -304,17 +304,17 @@ void ccodelte_encode(int32_t numbits, uint8_t add_crc, uint8_t *inPtr, uint8_t *
\brief This function initializes the generator polynomials for an LTE convolutional code.*/
void ccodelte_init(void);
/*!\fn void ccodelte_init_inv(void)
/*!\fn void ccodelte_init_inv(unsigned char ccodelte_table_rev[128])
\brief This function initializes the trellis structure for decoding an LTE convolutional code.*/
void ccodelte_init_inv(void);
void ccodelte_init_inv(unsigned char ccodelte_table_rev[128]);
/*!\fn void ccodelte_init(void)
\brief This function initializes the generator polynomials for an DAB convolutional code (first 3 bits).*/
void ccodedab_init(void);
/*!\fn void ccodelte_init_inv(void)
/*!\fn void ccodelte_init_inv(unsigned char ccodelte_table_rev[128])
\brief This function initializes the trellis structure for decoding an DAB convolutional code (first 3 bits).*/
void ccodedab_init_inv(void);
void ccodedab_init_inv(unsigned char ccodelte_table_rev[128]);
/*!\fn void crcTableInit(void)
\brief This function initializes the different crc tables.*/
@@ -399,7 +399,7 @@ void phy_viterbi_lte_sse2(int8_t *y, uint8_t *decoded_bytes, uint16_t n);
*/
void phy_generate_viterbi_tables(void);
/*!\fn void phy_generate_viterbi_tables_lte(void)
/*!\fn void phy_generate_viterbi_tables_lte(bool dabflag)
\brief This routine initializes metric tables for the optimized LTE Viterbi decoder.
*/
void phy_generate_viterbi_tables_lte(void);
openair1/PHY/CODING/viterbi_lte.c
+56 -83
View File
@@ -40,16 +40,20 @@
#include "PHY/sse_intrin.h"
extern uint8_t ccodelte_table[128],ccodelte_table_rev[128];
#include "coding_defs.h"
static int8_t m0_table[64*16*16*16] __attribute__ ((aligned(16)));
static int8_t m1_table[64*16*16*16] __attribute__ ((aligned(16)));
// Set up Viterbi tables for SSE2 implementation
void phy_generate_viterbi_tables_lte( void )
void phy_generate_viterbi_tables_lte(bool dabflag)
{
uint8_t ccodelte_table_rev[128];
if (dabflag)
ccodedab_init_inv(ccodelte_table_rev);
else
ccodelte_init_inv(ccodelte_table_rev);
int8_t w[8],in0,in1,in2;
uint8_t state,index0,index1;
@@ -126,52 +130,29 @@ void print_shorts(simde__m128i x,char *s) {
void phy_viterbi_lte_sse2(int8_t *y,uint8_t *decoded_bytes,uint16_t n)
{
simde__m128i TB[4 * 8192];
simde__m128i *m0_ptr, *m1_ptr, *TB_ptr = &TB[0];
simde__m128i metrics0_15, metrics16_31, metrics32_47, metrics48_63, even0_30a, even0_30b, even32_62a, even32_62b, odd1_31a,
odd1_31b, odd33_63a, odd33_63b, TBeven0_30, TBeven32_62, TBodd1_31, TBodd33_63;
simde__m128i min_state, min_state2;
int8_t *in = y;
uint8_t prev_state0,maxm,s;
static uint8_t *TB_ptr2;
uint32_t table_offset;
uint8_t iter;
int16_t position;
// set initial metrics
//debug_msg("Doing viterbi\n");
metrics0_15 = simde_mm_setzero_si128();
metrics16_31 = simde_mm_setzero_si128();
metrics32_47 = simde_mm_setzero_si128();
metrics48_63 = simde_mm_setzero_si128();
for (iter=0; iter<2; iter++) {
in = y;
TB_ptr=&TB[0];
for (position=0; position<n; position++) {
simde__m128i metrics0_15 = {0}, metrics16_31 = {0}, metrics32_47 = {0}, metrics48_63 = {0};
for (int iter = 0; iter < 2; iter++) {
int8_t *in = y;
simde__m128i *TB_ptr = TB;
for (int position = 0; position < n; position++) {
// get branch metric offsets for the 64 states
table_offset = (in[0]+8 + ((in[1]+8)<<4) + ((in[2]+8)<<8))<<6;
uint table_offset = (in[0] + 8 + ((in[1] + 8) << 4) + ((in[2] + 8) << 8)) << 6;
m0_ptr = (simde__m128i *)&m0_table[table_offset];
m1_ptr = (simde__m128i *)&m1_table[table_offset];
simde__m128i *m0_ptr = (simde__m128i *)&m0_table[table_offset];
simde__m128i *m1_ptr = (simde__m128i *)&m1_table[table_offset];
// even states
even0_30a = simde_mm_adds_epu8(metrics0_15, m0_ptr[0]);
even32_62a = simde_mm_adds_epu8(metrics16_31, m0_ptr[1]);
even0_30b = simde_mm_adds_epu8(metrics32_47, m0_ptr[2]);
even32_62b = simde_mm_adds_epu8(metrics48_63, m0_ptr[3]);
simde__m128i even0_30a = simde_mm_adds_epu8(metrics0_15, m0_ptr[0]);
simde__m128i even32_62a = simde_mm_adds_epu8(metrics16_31, m0_ptr[1]);
simde__m128i even0_30b = simde_mm_adds_epu8(metrics32_47, m0_ptr[2]);
simde__m128i even32_62b = simde_mm_adds_epu8(metrics48_63, m0_ptr[3]);
// odd states
odd1_31a = simde_mm_adds_epu8(metrics0_15, m1_ptr[0]);
odd33_63a = simde_mm_adds_epu8(metrics16_31, m1_ptr[1]);
odd1_31b = simde_mm_adds_epu8(metrics32_47, m1_ptr[2]);
odd33_63b = simde_mm_adds_epu8(metrics48_63, m1_ptr[3]);
simde__m128i odd1_31a = simde_mm_adds_epu8(metrics0_15, m1_ptr[0]);
simde__m128i odd33_63a = simde_mm_adds_epu8(metrics16_31, m1_ptr[1]);
simde__m128i odd1_31b = simde_mm_adds_epu8(metrics32_47, m1_ptr[2]);
simde__m128i odd33_63b = simde_mm_adds_epu8(metrics48_63, m1_ptr[3]);
// select maxima
@@ -182,33 +163,31 @@ void phy_viterbi_lte_sse2(int8_t *y,uint8_t *decoded_bytes,uint16_t n)
// Traceback information
TBeven0_30 = simde_mm_cmpeq_epi8(even0_30a, even0_30b);
TBeven32_62 = simde_mm_cmpeq_epi8(even32_62a, even32_62b);
TBodd1_31 = simde_mm_cmpeq_epi8(odd1_31a, odd1_31b);
TBodd33_63 = simde_mm_cmpeq_epi8(odd33_63a, odd33_63b);
simde__m128i TBeven0_30 = simde_mm_cmpeq_epi8(even0_30a, even0_30b);
simde__m128i TBeven32_62 = simde_mm_cmpeq_epi8(even32_62a, even32_62b);
simde__m128i TBodd1_31 = simde_mm_cmpeq_epi8(odd1_31a, odd1_31b);
simde__m128i TBodd33_63 = simde_mm_cmpeq_epi8(odd33_63a, odd33_63b);
metrics0_15 = simde_mm_unpacklo_epi8(even0_30a, odd1_31a);
metrics16_31 = simde_mm_unpackhi_epi8(even0_30a, odd1_31a);
metrics32_47 = simde_mm_unpacklo_epi8(even32_62a, odd33_63a);
metrics48_63 = simde_mm_unpackhi_epi8(even32_62a, odd33_63a);
TB_ptr[0] = simde_mm_unpacklo_epi8(TBeven0_30, TBodd1_31);
TB_ptr[1] = simde_mm_unpackhi_epi8(TBeven0_30, TBodd1_31);
TB_ptr[2] = simde_mm_unpacklo_epi8(TBeven32_62, TBodd33_63);
TB_ptr[3] = simde_mm_unpackhi_epi8(TBeven32_62, TBodd33_63);
in+=3;
TB_ptr += 4;
*TB_ptr++ = simde_mm_unpacklo_epi8(TBeven0_30, TBodd1_31);
*TB_ptr++ = simde_mm_unpackhi_epi8(TBeven0_30, TBodd1_31);
*TB_ptr++ = simde_mm_unpacklo_epi8(TBeven32_62, TBodd33_63);
*TB_ptr++ = simde_mm_unpackhi_epi8(TBeven32_62, TBodd33_63);
in += 3;
// rescale by subtracting minimum
/****************************************************
USE SSSE instruction phminpos!!!!!!!
****************************************************/
min_state = simde_mm_min_epu8(metrics0_15, metrics16_31);
simde__m128i min_state = simde_mm_min_epu8(metrics0_15, metrics16_31);
min_state = simde_mm_min_epu8(min_state, metrics32_47);
min_state = simde_mm_min_epu8(min_state, metrics48_63);
min_state2 = min_state;
simde__m128i min_state2 = min_state;
min_state = simde_mm_unpacklo_epi8(min_state, min_state);
min_state2 = simde_mm_unpackhi_epi8(min_state2, min_state2);
min_state = simde_mm_min_epu8(min_state, min_state2);
@@ -237,37 +216,34 @@ void phy_viterbi_lte_sse2(int8_t *y,uint8_t *decoded_bytes,uint16_t n)
} // iteration
// Traceback
prev_state0 = 0;
maxm = 0;
for (s=0; s<16; s++)
if (((uint8_t *)&metrics0_15)[s] > maxm) {
maxm = ((uint8_t *)&metrics0_15)[s];
uint prev_state0 = 0;
uint maxm = 0;
uint s = 0;
for (uint8_t *ptr = (uint8_t *)&metrics0_15; s < 16; s++, ptr++)
if (*ptr > maxm) {
maxm = *ptr;
prev_state0 = s;
}
for (s=0; s<16; s++)
if (((uint8_t *)&metrics16_31)[s] > maxm) {
maxm = ((uint8_t *)&metrics16_31)[s];
prev_state0 = s+16;
for (uint8_t *ptr = (uint8_t *)&metrics16_31; s < 32; s++, ptr++)
if (*ptr > maxm) {
maxm = *ptr;
prev_state0 = s;
}
for (uint8_t *ptr = (uint8_t *)&metrics32_47; s < 48; s++, ptr++)
if (*ptr > maxm) {
maxm = *ptr;
prev_state0 = s;
}
for (uint8_t *ptr = (uint8_t *)&metrics48_63; s < 64; s++, ptr++)
if (*ptr > maxm) {
maxm = *ptr;
prev_state0 = s;
}
for (s=0; s<16; s++)
if (((uint8_t *)&metrics32_47)[s] > maxm) {
maxm = ((uint8_t *)&metrics32_47)[s];
prev_state0 = s+32;
}
for (s=0; s<16; s++)
if (((uint8_t *)&metrics48_63)[s] > maxm) {
maxm = ((uint8_t *)&metrics48_63)[s];
prev_state0 = s+48;
}
TB_ptr2 = (uint8_t *)&TB[(n-1)*4];
for (position = n-1 ; position>-1; position--) {
uint8_t *TB_ptr2 = (uint8_t *)&TB[(n - 1) * 4];
for (int position = n - 1; position > -1; position--) {
decoded_bytes[(position)>>3] += (prev_state0 & 0x1)<<(7-(position & 0x7));
@@ -278,7 +254,6 @@ void phy_viterbi_lte_sse2(int8_t *y,uint8_t *decoded_bytes,uint16_t n)
TB_ptr2-=64;
}
}
#ifdef TEST_DEBUG
@@ -304,10 +279,8 @@ int test_viterbi(uint8_t dabflag)
if (dabflag==0) {
ccodelte_init();
ccodelte_init_inv();
} else {
ccodedab_init();
ccodedab_init_inv();
printf("Running with DAB polynomials\n");
}
openair1/PHY/INIT/init_top.c
+1 -2
View File
@@ -63,8 +63,7 @@ void generate_qpsk_table(void) {
void init_lte_top(LTE_DL_FRAME_PARMS *frame_parms) {
ccodelte_init();
ccodelte_init_inv();
phy_generate_viterbi_tables_lte();
phy_generate_viterbi_tables_lte(false);
load_codinglib();
generate_ul_ref_sigs();
generate_ul_ref_sigs_rx();