From b110c3622b35c752a94a7de1d5bb066e80e5db78 Mon Sep 17 00:00:00 2001 From: Thomas Schlichter Date: Wed, 21 Apr 2021 13:37:35 +0200 Subject: [PATCH] NR_UE: fix possible overflow in PRB based PDSCH channel estimation --- .../nr_dl_channel_estimation.c | 55 ++++++++++--------- 1 file changed, 30 insertions(+), 25 deletions(-) diff --git a/openair1/PHY/NR_UE_ESTIMATION/nr_dl_channel_estimation.c b/openair1/PHY/NR_UE_ESTIMATION/nr_dl_channel_estimation.c index 7fedecc426..ab9a42c60d 100644 --- a/openair1/PHY/NR_UE_ESTIMATION/nr_dl_channel_estimation.c +++ b/openair1/PHY/NR_UE_ESTIMATION/nr_dl_channel_estimation.c @@ -1200,49 +1200,52 @@ int nr_pdsch_channel_estimation(PHY_VARS_NR_UE *ue, else if (config_type == pdsch_dmrs_type1) { // this is case without frequency-domain linear interpolation, just take average of LS channel estimates of 6 DMRS REs and use a common value for the whole PRB for (pilot_cnt=0; pilot_cnt<6*nb_rb_pdsch; pilot_cnt += 6) { - ch[0] = (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] = (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); + int32_t ch_0, ch_1; + + ch_0 = ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 = ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0] += (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] += (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); + ch_0 += ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 += ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0] += (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] += (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); + ch_0 += ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 += ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0] += (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] += (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); + ch_0 += ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 += ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0] += (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] += (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); + ch_0 += ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 += ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0] += (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] += (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); + ch_0 += ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 += ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0]/=6; - ch[1]/=6; + + ch[0] = ch_0 / 6; + ch[1] = ch_1 / 6; ((__m128i*)dl_ch)[0]=_mm_set1_epi32(*(int32_t*)ch); ((__m128i*)dl_ch)[1]=((__m128i*)dl_ch)[0]; ((__m128i*)dl_ch)[2]=((__m128i*)dl_ch)[0]; @@ -1251,36 +1254,38 @@ int nr_pdsch_channel_estimation(PHY_VARS_NR_UE *ue, } else { // this is case without frequency-domain linear interpolation, just take average of LS channel estimates of 4 DMRS REs and use a common value for the whole PRB for (pilot_cnt=0; pilot_cnt<4*nb_rb_pdsch; pilot_cnt += 4) { - ch[0] = (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] = (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); + int32_t ch_0, ch_1; + + ch_0 = ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 = ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0] += (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] += (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); + ch_0 += ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 += ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0] += (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] += (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); + ch_0 += ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 += ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0] += (int16_t)(((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15); - ch[1] += (int16_t)(((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15); - + ch_0 += ((int32_t)pil[0]*rxF[0] - (int32_t)pil[1]*rxF[1])>>15; + ch_1 += ((int32_t)pil[0]*rxF[1] + (int32_t)pil[1]*rxF[0])>>15; pil += 2; re_offset = (re_offset+2) % ue->frame_parms.ofdm_symbol_size; rxF = (int16_t *)&rxdataF[aarx][(symbol_offset+nushift+re_offset)]; - ch[0] /=4 ; - ch[1] /=4 ; + + ch[0] = ch_0 / 4; + ch[1] = ch_1 / 4; ((__m128i*)dl_ch)[0]=_mm_set1_epi32(*(int32_t*)ch); ((__m128i*)dl_ch)[1]=((__m128i*)dl_ch)[0]; ((__m128i*)dl_ch)[2]=((__m128i*)dl_ch)[0];