use one call to multiply the full size vector

This commit is contained in:
Laurent THOMAS
2025-04-25 14:47:59 +02:00
parent f470ea5946
commit bb93b06663
+17 -28
View File
@@ -27,42 +27,40 @@
void remove_7_5_kHz(RU_t *ru,uint8_t slot)
{
int32_t **rxdata=ru->common.rxdata;
int32_t **rxdata_7_5kHz=ru->common.rxdata_7_5kHz;
uint32_t *kHz7_5ptr;
simde__m128i *rxptr128, *rxptr128_7_5kHz, *kHz7_5ptr128;
c16_t **rxdata=(c16_t **)ru->common.rxdata;
c16_t **rxdata_7_5kHz=(c16_t **)ru->common.rxdata_7_5kHz;
int16_t *kHz7_5;
LTE_DL_FRAME_PARMS *frame_parms=ru->frame_parms;
switch (frame_parms->N_RB_UL) {
case 6:
kHz7_5ptr = (frame_parms->Ncp==0) ? (uint32_t*)s6n_kHz_7_5 : (uint32_t*)s6e_kHz_7_5;
kHz7_5 = (frame_parms->Ncp==0) ? s6n_kHz_7_5 : s6e_kHz_7_5;
break;
case 15:
kHz7_5ptr = (frame_parms->Ncp==0) ? (uint32_t*)s15n_kHz_7_5 : (uint32_t*)s15e_kHz_7_5;
kHz7_5 = (frame_parms->Ncp==0) ? s15n_kHz_7_5 : s15e_kHz_7_5;
break;
case 25:
kHz7_5ptr = (frame_parms->Ncp==0) ? (uint32_t*)s25n_kHz_7_5 : (uint32_t*)s25e_kHz_7_5;
kHz7_5 = (frame_parms->Ncp==0) ? s25n_kHz_7_5 : s25e_kHz_7_5;
break;
case 50:
kHz7_5ptr = (frame_parms->Ncp==0) ? (uint32_t*)s50n_kHz_7_5 : (uint32_t*)s50e_kHz_7_5;
kHz7_5 = (frame_parms->Ncp==0) ? s50n_kHz_7_5 : s50e_kHz_7_5;
break;
case 75:
kHz7_5ptr = (frame_parms->Ncp==0) ? (uint32_t*)s75n_kHz_7_5 : (uint32_t*)s75e_kHz_7_5;
kHz7_5 = (frame_parms->Ncp==0) ? s75n_kHz_7_5 : s75e_kHz_7_5;
break;
case 100:
kHz7_5ptr = (frame_parms->Ncp==0) ? (uint32_t*)s100n_kHz_7_5 : (uint32_t*)s100e_kHz_7_5;
kHz7_5 = (frame_parms->Ncp==0) ? s100n_kHz_7_5 : s100e_kHz_7_5;
break;
default:
kHz7_5ptr = (frame_parms->Ncp==0) ? (uint32_t*)s25n_kHz_7_5 : (uint32_t*)s25e_kHz_7_5;
kHz7_5 = (frame_parms->Ncp==0) ? s25n_kHz_7_5 : s25e_kHz_7_5;
break;
}
@@ -72,25 +70,16 @@ void remove_7_5_kHz(RU_t *ru,uint8_t slot)
uint16_t len = frame_parms->samples_per_tti/2;
for (uint8_t aa=0; aa<ru->nb_rx; aa++) {
rxptr128 = (simde__m128i *)&rxdata[aa][slot_offset];
rxptr128_7_5kHz = (simde__m128i *)&rxdata_7_5kHz[aa][slot_offset2];
kHz7_5ptr128 = (simde__m128i *)kHz7_5ptr;
c16_t *rxptr = rxdata[aa]+slot_offset;
c16_t *rxptr_7_5kHz = rxdata_7_5kHz[aa]+slot_offset2;
// apply 7.5 kHz
// if (((slot>>1)&1) == 0) { // apply the sinusoid from the table directly
for (uint32_t i=0; i<(len>>2); i++) {
rxptr128_7_5kHz[i] = oai_mm_cpx_mult(kHz7_5ptr128[i], rxptr128[i], 15);
}
mult_complex_vectors((c16_t*)kHz7_5,rxptr,rxptr_7_5kHz,len, 15);
// undo 7.5 kHz offset for symbol 3 in case RU is slave (for OTA synchronization)
if (ru->is_slave == 1 && slot == 2){
memcpy((void*)&rxdata_7_5kHz[aa][(3*frame_parms->ofdm_symbol_size)+
(2*frame_parms->nb_prefix_samples)+
frame_parms->nb_prefix_samples0],
(void*)&rxdata[aa][slot_offset+ru->N_TA_offset+
(3*frame_parms->ofdm_symbol_size)+
(2*frame_parms->nb_prefix_samples)+
frame_parms->nb_prefix_samples0],
(frame_parms->ofdm_symbol_size + frame_parms->nb_prefix_samples)*sizeof(int32_t));
int offset=3*frame_parms->ofdm_symbol_size+2*frame_parms->nb_prefix_samples+frame_parms->nb_prefix_samples0;
memcpy(rxdata_7_5kHz[aa]+offset,
rxdata[aa]+offset,
(frame_parms->ofdm_symbol_size + frame_parms->nb_prefix_samples)*sizeof(c16_t));
}
}
}