bugfix for 4096 and 8192-point dfts.

This commit is contained in:
Raymond Knopp
2016-09-26 12:08:33 -07:00
parent 3124cb2c00
commit ff90a54ad9
6 changed files with 112 additions and 37 deletions
+39 -13
View File
@@ -46,7 +46,7 @@
#include "SCHED/extern.h"
#include "UTIL/LOG/vcd_signal_dumper.h"
//#define PRACH_DEBUG 1
#define PRACH_DEBUG 1
uint16_t NCS_unrestricted[16] = {0,13,15,18,22,26,32,38,46,59,76,93,119,167,279,419};
uint16_t NCS_restricted[15] = {15,18,22,26,32,38,46,55,68,82,100,128,158,202,237}; // high-speed case
@@ -818,11 +818,14 @@ int32_t generate_prach( PHY_VARS_UE *ue, uint8_t eNB_id, uint8_t subframe, uint1
break;
case 75:
memset((void*)prachF,0,4*19432);
memset((void*)prachF,0,4*18432);
break;
case 100:
memset((void*)prachF,0,4*24576);
if (ue->frame_parms.threequarter_fs == 0)
memset((void*)prachF,0,4*24576);
else
memset((void*)prachF,0,4*18432);
break;
}
@@ -886,6 +889,9 @@ int32_t generate_prach( PHY_VARS_UE *ue, uint8_t eNB_id, uint8_t subframe, uint1
break;
}
if (ue->frame_parms.threequarter_fs == 1)
Ncp=(Ncp*3)>>2;
prach2 = prach+(Ncp<<1);
// do IDFT
@@ -986,21 +992,41 @@ int32_t generate_prach( PHY_VARS_UE *ue, uint8_t eNB_id, uint8_t subframe, uint1
break;
case 100:
if (ue->frame_parms.threequarter_fs == 0) {
if (prach_fmt == 4) {
idft4096(prachF,prach2,1);
memmove( prach, prach+8192, Ncp<<2 );
prach_len = 4096+Ncp;
} else {
idft24576(prachF,prach2);
memmove( prach, prach+49152, Ncp<<2 );
prach_len = 24576+Ncp;
if (prach_fmt>1) {
memmove( prach2+49152, prach2, 98304 );
prach_len = 2* 24576+Ncp;
}
}
}
else {
if (prach_fmt == 4) {
idft4096(prachF,prach2,1);
memmove( prach, prach+8192, Ncp<<2 );
prach_len = 4096+Ncp;
idft3072(prachF,prach2);
//TODO: account for repeated format in dft output
memmove( prach, prach+6144, Ncp<<2 );
prach_len = 3072+Ncp;
} else {
idft24576(prachF,prach2);
memmove( prach, prach+49152, Ncp<<2 );
prach_len = 24576+Ncp;
idft18432(prachF,prach2);
memmove( prach, prach+36864, Ncp<<2 );
prach_len = 18432+Ncp;
if (prach_fmt>1) {
memmove( prach2+49152, prach2, 98304 );
prach_len = 2* 24576+Ncp;
memmove( prach2+36834, prach2, 73728 );
prach_len = 2*18432+Ncp;
}
}
}
break;
}
@@ -1496,7 +1522,7 @@ void rx_prach(PHY_VARS_eNB *eNB,
#ifdef PRACH_DEBUG
if (en>40) {
// if (en>40) {
k = (12*n_ra_prb) - 6*eNB->frame_parms.N_RB_UL;
if (k<0)
@@ -1510,7 +1536,7 @@ void rx_prach(PHY_VARS_eNB *eNB,
write_output("prach_rxF_comp0.m","prach_rxF_comp0",prachF,1024,1,1);
write_output("prach_ifft0.m","prach_t0",prach_ifft[0],1024,1,1);
exit(-1);
}
// }
#endif
} // new dft
+55 -13
View File
@@ -5229,7 +5229,7 @@ void dft8192(int16_t *x,int16_t *y,int scale)
xtmpp = xtmp;
for (i=0; i<32; i++) {
for (i=0; i<16; i++) {
transpose4_ooff_simd256(x256 ,xtmpp,512);
transpose4_ooff_simd256(x256+2,xtmpp+1,512);
transpose4_ooff_simd256(x256+4,xtmpp+2,512);
@@ -5267,7 +5267,7 @@ void dft8192(int16_t *x,int16_t *y,int scale)
}
dft4096((int16_t*)(xtmp),(int16_t*)ytmp,1);
dft4096((int16_t*)(xtmp+1024),(int16_t*)(ytmp+512),1);
dft4096((int16_t*)(xtmp+512),(int16_t*)(ytmp+512),1);
for (i=0; i<512; i++) {
@@ -5319,7 +5319,7 @@ void idft8192(int16_t *x,int16_t *y,int scale)
xtmpp = xtmp;
for (i=0; i<32; i++) {
for (i=0; i<16; i++) {
transpose4_ooff_simd256(x256 ,xtmpp,512);
transpose4_ooff_simd256(x256+2,xtmpp+1,512);
transpose4_ooff_simd256(x256+4,xtmpp+2,512);
@@ -5357,7 +5357,7 @@ void idft8192(int16_t *x,int16_t *y,int scale)
}
idft4096((int16_t*)(xtmp),(int16_t*)ytmp,1);
idft4096((int16_t*)(xtmp+1024),(int16_t*)(ytmp+512),1);
idft4096((int16_t*)(xtmp+512),(int16_t*)(ytmp+512),1);
for (i=0; i<512; i++) {
@@ -5733,8 +5733,8 @@ void dft24576(int16_t *input, int16_t *output)
void idft24576(int16_t *input, int16_t *output)
{
int i,i2,j;
uint32_t tmp[3][16384] __attribute__((aligned(32)));
uint32_t tmpo[3][16384] __attribute__((aligned(32)));
uint32_t tmp[3][8192] __attribute__((aligned(32)));
uint32_t tmpo[3][8192] __attribute__((aligned(32)));
for (i=0,j=0; i<8192; i++) {
tmp[0][i] = ((uint32_t *)input)[j++];
@@ -5746,13 +5746,6 @@ void idft24576(int16_t *input, int16_t *output)
idft8192((int16_t*)(tmp[1]),(int16_t*)(tmpo[1]),1);
idft8192((int16_t*)(tmp[2]),(int16_t*)(tmpo[2]),1);
/*
for (i=1; i<8192; i++) {
tmpo[0][i] = tmpo[0][i<<1];
tmpo[1][i] = tmpo[1][i<<1];
tmpo[2][i] = tmpo[2][i<<1];
}*/
/*
write_output("in.m","in",input,24576,1,1);
write_output("out0.m","o0",tmpo[0],8192,1,1);
@@ -19129,6 +19122,55 @@ int main(int argc, char**argv)
write_output("y2048.m","y2048",y,2048,1,1);
write_output("x2048.m","x2048",x,2048,1,1);
memset((void*)x,0,2048*sizeof(int32_t));
for (i=2;i<2402;i++) {
if ((taus() & 1)==0)
((int16_t*)x)[i] = 364;
else
((int16_t*)x)[i] = -364;
}
for (i=2*(4096-1200);i<8192;i++) {
if ((taus() & 1)==0)
((int16_t*)x)[i] = 364;
else
((int16_t*)x)[i] = -364;
}
reset_meas(&ts);
for (i=0; i<10000; i++) {
start_meas(&ts);
idft4096((int16_t *)x,(int16_t *)y,1);
stop_meas(&ts);
}
printf("\n\n4096-point(%f cycles)\n",(double)ts.diff/(double)ts.trials);
write_output("y4096.m","y4096",y,4096,1,1);
write_output("x4096.m","x4096",x,4096,1,1);
memset((void*)x,0,8192*sizeof(int32_t));
for (i=2;i<4802;i++) {
if ((taus() & 1)==0)
((int16_t*)x)[i] = 364;
else
((int16_t*)x)[i] = -364;
}
for (i=2*(4096-1200);i<8192;i++) {
if ((taus() & 1)==0)
((int16_t*)x)[i] = 364;
else
((int16_t*)x)[i] = -364;
}
reset_meas(&ts);
for (i=0; i<10000; i++) {
start_meas(&ts);
idft8192((int16_t *)x,(int16_t *)y,1);
stop_meas(&ts);
}
printf("\n\n8192-point(%f cycles)\n",(double)ts.diff/(double)ts.trials);
write_output("y8192.m","y8192",y,8192,1,1);
write_output("x8192.m","x8192",x,8192,1,1);
return(0);
}
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
File diff suppressed because one or more lines are too long
+12 -5
View File
@@ -104,13 +104,15 @@ int main(int argc, char **argv)
double delay_avg=0;
double ue_speed = 0;
int NCS_config = 1,rootSequenceIndex=0;
int threequarter_fs = 0;
logInit();
number_of_cards = 1;
while ((c = getopt (argc, argv, "hHaA:Cr:p:g:n:s:S:t:x:y:v:V:z:N:F:d:Z:L:R:")) != -1) {
while ((c = getopt (argc, argv, "hHaA:Cr:p:g:n:s:S:t:x:y:v:V:z:N:F:d:Z:L:R:E")) != -1) {
switch (c) {
case 'a':
printf("Running AWGN simulation\n");
@@ -182,6 +184,10 @@ int main(int argc, char **argv)
break;
case 'E':
threequarter_fs=1;
break;
case 'n':
n_frames = atoi(optarg);
break;
@@ -316,7 +322,7 @@ int main(int argc, char **argv)
Nid_cell,
3,
N_RB_DL,
0,
threequarter_fs,
osf,
0);
@@ -395,8 +401,9 @@ int main(int argc, char **argv)
eNB->frame_parms.prach_config_common.prach_ConfigInfo.highSpeedFlag=hs_flag;
eNB->frame_parms.prach_config_common.prach_ConfigInfo.prach_FreqOffset=0;
eNB->node_function = eNodeB_3GPP;
eNB->proc.subframe_rx = subframe;
eNB->node_function = eNodeB_3GPP;
eNB->proc.subframe_rx = subframe;
eNB->proc.subframe_prach = subframe;
/* N_ZC not used later, so prach_fmt is also useless, don't set */
//prach_fmt = get_prach_fmt(eNB->frame_parms.prach_config_common.prach_ConfigInfo.prach_ConfigIndex,
@@ -530,7 +537,7 @@ int main(int argc, char **argv)
write_output("rxsig0.m","rxs0",
&eNB->common_vars.rxdata[0][0][subframe*frame_parms->samples_per_tti],
frame_parms->samples_per_tti,1,1);
write_output("rxsigF0.m","rxsF0", &eNB->common_vars.rxdataF[0][0][0],512*nsymb*2,2,1);
write_output("rxsigF0.m","rxsF0", eNB->prach_vars.rxsigF[0],6144,1,1);
write_output("prach_preamble.m","prachp",&eNB->X_u[0],839,1,1);
}
}