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

create dft shared lib

Browse Source
This commit is contained in:
frtabu
2020-03-30 18:23:32 +02:00
committed by oai
parent 82e97fce59
commit 296494ea4c
11 changed files with 494 additions and 316 deletions
Download Patch File Download Diff File Expand all files Collapse all files
cmake_targets/CMakeLists.txt
+5 -1
View File
@@ -1327,6 +1327,9 @@ add_library(ldpc MODULE ${PHY_LDPC_OPTIM8SEGMULTI_SRC} )
add_library(coding MODULE ${PHY_TURBOSRC} ) add_library(coding MODULE ${PHY_TURBOSRC} )
add_library(dfts MODULE ${OPENAIR1_DIR}/PHY/TOOLS/oai_dfts.c )
set(PHY_SRC_COMMON set(PHY_SRC_COMMON
${OPENAIR1_DIR}/PHY/LTE_TRANSPORT/dci_tools_common.c ${OPENAIR1_DIR}/PHY/LTE_TRANSPORT/dci_tools_common.c
${OPENAIR1_DIR}/PHY/LTE_TRANSPORT/lte_mcs.c ${OPENAIR1_DIR}/PHY/LTE_TRANSPORT/lte_mcs.c
@@ -1368,7 +1371,7 @@ set(PHY_SRC_COMMON
${OPENAIR1_DIR}/PHY/INIT/lte_parms.c ${OPENAIR1_DIR}/PHY/INIT/lte_parms.c
${OPENAIR1_DIR}/PHY/INIT/lte_param_init.c ${OPENAIR1_DIR}/PHY/INIT/lte_param_init.c
${OPENAIR1_DIR}/PHY/TOOLS/cadd_vv.c ${OPENAIR1_DIR}/PHY/TOOLS/cadd_vv.c
${OPENAIR1_DIR}/PHY/TOOLS/lte_dfts.c # ${OPENAIR1_DIR}/PHY/TOOLS/lte_dfts.c
${OPENAIR1_DIR}/PHY/TOOLS/log2_approx.c ${OPENAIR1_DIR}/PHY/TOOLS/log2_approx.c
${OPENAIR1_DIR}/PHY/TOOLS/cmult_sv.c ${OPENAIR1_DIR}/PHY/TOOLS/cmult_sv.c
${OPENAIR1_DIR}/PHY/TOOLS/cmult_vv.c ${OPENAIR1_DIR}/PHY/TOOLS/cmult_vv.c
@@ -1576,6 +1579,7 @@ endif ()
add_library(PHY_COMMON ${PHY_SRC_COMMON}) add_library(PHY_COMMON ${PHY_SRC_COMMON})
add_dependencies(PHY_COMMON rrc_flag) add_dependencies(PHY_COMMON rrc_flag)
add_dependencies(PHY_COMMON dfts)
add_library(PHY ${PHY_SRC}) add_library(PHY ${PHY_SRC})
add_dependencies(PHY rrc_flag) add_dependencies(PHY rrc_flag)
add_library(PHY_UE ${PHY_SRC_UE}) add_library(PHY_UE ${PHY_SRC_UE})
openair1/PHY/LTE_ESTIMATION/lte_adjust_sync_eNB.c
+4 -4
View File
@@ -56,22 +56,22 @@ int lte_est_timing_advance(LTE_DL_FRAME_PARMS *frame_parms,
// do ifft of channel estimate // do ifft of channel estimate
switch(frame_parms->N_RB_DL) { switch(frame_parms->N_RB_DL) {
case 6: case 6:
dft128((int16_t*) &lte_eNB_srs->srs_ch_estimates[aa][0], dft(DFT_128,(int16_t*) &lte_eNB_srs->srs_ch_estimates[aa][0],
(int16_t*) lte_eNB_srs->srs_ch_estimates_time[aa], (int16_t*) lte_eNB_srs->srs_ch_estimates_time[aa],
1); 1);
break; break;
case 25: case 25:
dft512((int16_t*) &lte_eNB_srs->srs_ch_estimates[aa][0], dft(DFT_512,(int16_t*) &lte_eNB_srs->srs_ch_estimates[aa][0],
(int16_t*) lte_eNB_srs->srs_ch_estimates_time[aa], (int16_t*) lte_eNB_srs->srs_ch_estimates_time[aa],
1); 1);
break; break;
case 50: case 50:
dft1024((int16_t*) &lte_eNB_srs->srs_ch_estimates[aa][0], dft(DFT_1024,(int16_t*) &lte_eNB_srs->srs_ch_estimates[aa][0],
(int16_t*) lte_eNB_srs->srs_ch_estimates_time[aa], (int16_t*) lte_eNB_srs->srs_ch_estimates_time[aa],
1); 1);
break; break;
case 100: case 100:
dft2048((int16_t*) &lte_eNB_srs->srs_ch_estimates[aa][0], dft(DFT_2048,(int16_t*) &lte_eNB_srs->srs_ch_estimates[aa][0],
(int16_t*) lte_eNB_srs->srs_ch_estimates_time[aa], (int16_t*) lte_eNB_srs->srs_ch_estimates_time[aa],
1); 1);
break; break;
openair1/PHY/LTE_ESTIMATION/lte_sync_time.c
+20 -20
View File
@@ -141,28 +141,28 @@ int lte_sync_time_init(LTE_DL_FRAME_PARMS *frame_parms ) // LTE_UE_COMMON *com
switch (frame_parms->N_RB_DL) { switch (frame_parms->N_RB_DL) {
case 6: case 6:
idft128((short*)syncF_tmp, /// complex input idft(IDFT_128,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 25: case 25:
idft512((short*)syncF_tmp, /// complex input idft(IDFT_512,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 50: case 50:
idft1024((short*)syncF_tmp, /// complex input idft(IDFT_1024,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 75: case 75:
idft1536((short*)syncF_tmp, /// complex input idft(IDFT_1536,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, (short*)sync_tmp,
1); /// complex output 1); /// complex output
break; break;
case 100: case 100:
idft2048((short*)syncF_tmp, /// complex input idft(IDFT_2048,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
@@ -189,28 +189,28 @@ int lte_sync_time_init(LTE_DL_FRAME_PARMS *frame_parms ) // LTE_UE_COMMON *com
switch (frame_parms->N_RB_DL) { switch (frame_parms->N_RB_DL) {
case 6: case 6:
idft128((short*)syncF_tmp, /// complex input idft(IDFT_128,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 25: case 25:
idft512((short*)syncF_tmp, /// complex input idft(IDFT_512,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 50: case 50:
idft1024((short*)syncF_tmp, /// complex input idft(IDFT_1024,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 75: case 75:
idft1536((short*)syncF_tmp, /// complex input idft(IDFT_1536,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 100: case 100:
idft2048((short*)syncF_tmp, /// complex input idft(IDFT_2048,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
@@ -237,28 +237,28 @@ int lte_sync_time_init(LTE_DL_FRAME_PARMS *frame_parms ) // LTE_UE_COMMON *com
switch (frame_parms->N_RB_DL) { switch (frame_parms->N_RB_DL) {
case 6: case 6:
idft128((short*)syncF_tmp, /// complex input idft(IDFT_128,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 25: case 25:
idft512((short*)syncF_tmp, /// complex input idft(IDFT_512,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 50: case 50:
idft1024((short*)syncF_tmp, /// complex input idft(IDFT_1024,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 75: case 75:
idft1536((short*)syncF_tmp, /// complex input idft(IDFT_1536,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
case 100: case 100:
idft2048((short*)syncF_tmp, /// complex input idft(IDFT_2048,(short*)syncF_tmp, /// complex input
(short*)sync_tmp, /// complex output (short*)sync_tmp, /// complex output
1); 1);
break; break;
@@ -492,28 +492,28 @@ int ru_sync_time_init(RU_t *ru) // LTE_UE_COMMON *common_vars
switch (ru->frame_parms->N_RB_DL) { switch (ru->frame_parms->N_RB_DL) {
case 6: case 6:
idft128((int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]), idft(IDFT_128,(int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]),
ru->dmrssync, /// complex output ru->dmrssync, /// complex output
1); 1);
break; break;
case 25: case 25:
idft512((int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]), idft(IDFT_512,(int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]),
ru->dmrssync, /// complex output ru->dmrssync, /// complex output
1); 1);
break; break;
case 50: case 50:
idft1024((int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]), idft(IDFT_1024,(int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]),
ru->dmrssync, /// complex output ru->dmrssync, /// complex output
1); 1);
break; break;
case 75: case 75:
idft1536((int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]), idft(IDFT_1536,(int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]),
ru->dmrssync, ru->dmrssync,
1); /// complex output 1); /// complex output
break; break;
case 100: case 100:
idft2048((int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]), idft(IDFT_2048,(int16_t*)(&dmrsp[0][3*ru->frame_parms->ofdm_symbol_size]),
ru->dmrssync, /// complex output ru->dmrssync, /// complex output
1); 1);
break; break;
openair1/PHY/LTE_ESTIMATION/lte_ul_channel_estimation.c
+8 -8
View File
@@ -246,25 +246,25 @@ int32_t lte_ul_channel_estimation(PHY_VARS_eNB *eNB,
switch(frame_parms->N_RB_DL) { switch(frame_parms->N_RB_DL) {
case 6: case 6:
idft128((int16_t *) temp_in_ifft_0, idft(IDFT_128,(int16_t *) temp_in_ifft_0,
(int16_t *) ul_ch_estimates_time[aa], (int16_t *) ul_ch_estimates_time[aa],
1); 1);
break; break;
case 25: case 25:
idft512((int16_t *) temp_in_ifft_0, idft(IDFT_512,(int16_t *) temp_in_ifft_0,
(int16_t *) ul_ch_estimates_time[aa], (int16_t *) ul_ch_estimates_time[aa],
1); 1);
break; break;
case 50: case 50:
idft1024((int16_t *) temp_in_ifft_0, idft(IDFT_1024,(int16_t *) temp_in_ifft_0,
(int16_t *) ul_ch_estimates_time[aa], (int16_t *) ul_ch_estimates_time[aa],
1); 1);
break; break;
case 100: case 100:
idft2048((int16_t *) temp_in_ifft_0, idft(IDFT_2048,(int16_t *) temp_in_ifft_0,
(int16_t *) ul_ch_estimates_time[aa], (int16_t *) ul_ch_estimates_time[aa],
1); 1);
break; break;
@@ -575,25 +575,25 @@ int32_t lte_ul_channel_estimation_RRU(LTE_DL_FRAME_PARMS *frame_parms,
switch(frame_parms->N_RB_DL) { switch(frame_parms->N_RB_DL) {
case 6: case 6:
idft128((int16_t *) temp_in_ifft_0, idft(IDFT_128,(int16_t *) temp_in_ifft_0,
(int16_t *) ul_ch_estimates_time[aa], (int16_t *) ul_ch_estimates_time[aa],
1); 1);
break; break;
case 25: case 25:
idft512((int16_t *) temp_in_ifft_0, idft(IDFT_512,(int16_t *) temp_in_ifft_0,
(int16_t *) ul_ch_estimates_time[aa], (int16_t *) ul_ch_estimates_time[aa],
1); 1);
break; break;
case 50: case 50:
idft1024((int16_t *) temp_in_ifft_0, idft(IDFT_1024,(int16_t *) temp_in_ifft_0,
(int16_t *) ul_ch_estimates_time[aa], (int16_t *) ul_ch_estimates_time[aa],
1); 1);
break; break;
case 100: case 100:
idft2048((int16_t *) temp_in_ifft_0, idft(IDFT_2048,(int16_t *) temp_in_ifft_0,
(int16_t *) ul_ch_estimates_time[aa], (int16_t *) ul_ch_estimates_time[aa],
1); 1);
break; break;
openair1/PHY/LTE_TRANSPORT/prach.c
+23 -23
View File
@@ -298,24 +298,24 @@ void rx_prach0(PHY_VARS_eNB *eNB,
switch (fp->N_RB_UL) { switch (fp->N_RB_UL) {
case 6: case 6:
if (prach_fmt == 4) { if (prach_fmt == 4) {
dft256(prach2,rxsigF[aa],1); dft(DFT_256,prach2,rxsigF[aa],1);
} else { } else {
dft1536(prach2,rxsigF[aa],1); dft(DFT_1536,prach2,rxsigF[aa],1);
if (prach_fmt>1) if (prach_fmt>1)
dft1536(prach2+3072,rxsigF[aa]+3072,1); dft(DFT_1536,prach2+3072,rxsigF[aa]+3072,1);
} }
break; break;
case 15: case 15:
if (prach_fmt == 4) { if (prach_fmt == 4) {
dft256(prach2,rxsigF[aa],1); dft(DFT_256,prach2,rxsigF[aa],1);
} else { } else {
dft3072(prach2,rxsigF[aa],1); dft(DFT_3072,prach2,rxsigF[aa],1);
if (prach_fmt>1) if (prach_fmt>1)
dft3072(prach2+6144,rxsigF[aa]+6144,1); dft(DFT_3072,prach2+6144,rxsigF[aa]+6144,1);
} }
break; break;
@@ -323,13 +323,13 @@ void rx_prach0(PHY_VARS_eNB *eNB,
case 25: case 25:
default: default:
if (prach_fmt == 4) { if (prach_fmt == 4) {
dft1024(prach2,rxsigF[aa],1); dft(DFT_1024,prach2,rxsigF[aa],1);
fft_size = 1024; fft_size = 1024;
} else { } else {
dft6144(prach2,rxsigF[aa],1); dft(DFT_6144,prach2,rxsigF[aa],1);
if (prach_fmt>1) if (prach_fmt>1)
dft6144(prach2+12288,rxsigF[aa]+12288,1); dft(DFT_6144,prach2+12288,rxsigF[aa]+12288,1);
fft_size = 6144; fft_size = 6144;
} }
@@ -338,24 +338,24 @@ void rx_prach0(PHY_VARS_eNB *eNB,
case 50: case 50:
if (prach_fmt == 4) { if (prach_fmt == 4) {
dft2048(prach2,rxsigF[aa],1); dft(DFT_2048,prach2,rxsigF[aa],1);
} else { } else {
dft12288(prach2,rxsigF[aa],1); dft(DFT_12288,prach2,rxsigF[aa],1);
if (prach_fmt>1) if (prach_fmt>1)
dft12288(prach2+24576,rxsigF[aa]+24576,1); dft(DFT_12288,prach2+24576,rxsigF[aa]+24576,1);
} }
break; break;
case 75: case 75:
if (prach_fmt == 4) { if (prach_fmt == 4) {
dft3072(prach2,rxsigF[aa],1); dft(DFT_3072,prach2,rxsigF[aa],1);
} else { } else {
dft18432(prach2,rxsigF[aa],1); dft(DFT_18432,prach2,rxsigF[aa],1);
if (prach_fmt>1) if (prach_fmt>1)
dft18432(prach2+36864,rxsigF[aa]+36864,1); dft(DFT_18432,prach2+36864,rxsigF[aa]+36864,1);
} }
break; break;
@@ -363,21 +363,21 @@ void rx_prach0(PHY_VARS_eNB *eNB,
case 100: case 100:
if (fp->threequarter_fs==0) { if (fp->threequarter_fs==0) {
if (prach_fmt == 4) { if (prach_fmt == 4) {
dft4096(prach2,rxsigF[aa],1); dft(DFT_4096,prach2,rxsigF[aa],1);
} else { } else {
dft24576(prach2,rxsigF[aa],1); dft(DFT_24576,prach2,rxsigF[aa],1);
if (prach_fmt>1) if (prach_fmt>1)
dft24576(prach2+49152,rxsigF[aa]+49152,1); dft(DFT_24576,prach2+49152,rxsigF[aa]+49152,1);
} }
} else { } else {
if (prach_fmt == 4) { if (prach_fmt == 4) {
dft3072(prach2,rxsigF[aa],1); dft(DFT_3072,prach2,rxsigF[aa],1);
} else { } else {
dft18432(prach2,rxsigF[aa],1); dft(DFT_18432,prach2,rxsigF[aa],1);
if (prach_fmt>1) if (prach_fmt>1)
dft18432(prach2+36864,rxsigF[aa]+36864,1); dft(DFT_18432,prach2+36864,rxsigF[aa]+36864,1);
} }
} }
@@ -589,13 +589,13 @@ void rx_prach0(PHY_VARS_eNB *eNB,
// Now do IFFT of size 1024 (N_ZC=839) or 256 (N_ZC=139) // Now do IFFT of size 1024 (N_ZC=839) or 256 (N_ZC=139)
if (N_ZC == 839) { if (N_ZC == 839) {
log2_ifft_size = 10; log2_ifft_size = 10;
idft1024(prachF,prach_ifft_tmp,1); idft(IDFT_1024,prachF,prach_ifft_tmp,1);
// compute energy and accumulate over receive antennas and repetitions for BR // compute energy and accumulate over receive antennas and repetitions for BR
for (i=0; i<2048; i++) for (i=0; i<2048; i++)
prach_ifft[i] += (prach_ifft_tmp[i<<1]*prach_ifft_tmp[i<<1] + prach_ifft_tmp[1+(i<<1)]*prach_ifft_tmp[1+(i<<1)])>>9; prach_ifft[i] += (prach_ifft_tmp[i<<1]*prach_ifft_tmp[i<<1] + prach_ifft_tmp[1+(i<<1)]*prach_ifft_tmp[1+(i<<1)])>>9;
} else { } else {
idft256(prachF,prach_ifft_tmp,1); idft(IDFT_256,prachF,prach_ifft_tmp,1);
log2_ifft_size = 8; log2_ifft_size = 8;
// compute energy and accumulate over receive antennas and repetitions for BR // compute energy and accumulate over receive antennas and repetitions for BR
openair1/PHY/LTE_TRANSPORT/ulsch_demodulation.c
+102 -102
View File
@@ -154,9 +154,9 @@ void lte_idft(LTE_DL_FRAME_PARMS *frame_parms,uint32_t *z, uint16_t Msc_PUSCH) {
switch (Msc_PUSCH) { switch (Msc_PUSCH) {
case 12: case 12:
dft12((int16_t *)idft_in0,(int16_t *)idft_out0); dft(DFT_12,(int16_t *)idft_in0,(int16_t *)idft_out0,0);
dft12((int16_t *)idft_in1,(int16_t *)idft_out1); dft(DFT_12,(int16_t *)idft_in1,(int16_t *)idft_out1,0);
dft12((int16_t *)idft_in2,(int16_t *)idft_out2); dft(DFT_12,(int16_t *)idft_in2,(int16_t *)idft_out2,0);
#if defined(__x86_64__)||defined(__i386__) #if defined(__x86_64__)||defined(__i386__)
norm128 = _mm_set1_epi16(9459); norm128 = _mm_set1_epi16(9459);
#elif defined(__arm__) #elif defined(__arm__)
@@ -178,201 +178,201 @@ void lte_idft(LTE_DL_FRAME_PARMS *frame_parms,uint32_t *z, uint16_t Msc_PUSCH) {
break; break;
case 24: case 24:
dft24(idft_in0,idft_out0,1); dft(DFT_24,idft_in0,idft_out0,1);
dft24(idft_in1,idft_out1,1); dft(DFT_24,idft_in1,idft_out1,1);
dft24(idft_in2,idft_out2,1); dft(DFT_24,idft_in2,idft_out2,1);
break; break;
case 36: case 36:
dft36(idft_in0,idft_out0,1); dft(DFT_36,idft_in0,idft_out0,1);
dft36(idft_in1,idft_out1,1); dft(DFT_36,idft_in1,idft_out1,1);
dft36(idft_in2,idft_out2,1); dft(DFT_36,idft_in2,idft_out2,1);
break; break;
case 48: case 48:
dft48(idft_in0,idft_out0,1); dft(DFT_48,idft_in0,idft_out0,1);
dft48(idft_in1,idft_out1,1); dft(DFT_48,idft_in1,idft_out1,1);
dft48(idft_in2,idft_out2,1); dft(DFT_48,idft_in2,idft_out2,1);
break; break;
case 60: case 60:
dft60(idft_in0,idft_out0,1); dft(DFT_60,idft_in0,idft_out0,1);
dft60(idft_in1,idft_out1,1); dft(DFT_60,idft_in1,idft_out1,1);
dft60(idft_in2,idft_out2,1); dft(DFT_60,idft_in2,idft_out2,1);
break; break;
case 72: case 72:
dft72(idft_in0,idft_out0,1); dft(DFT_72,idft_in0,idft_out0,1);
dft72(idft_in1,idft_out1,1); dft(DFT_72,idft_in1,idft_out1,1);
dft72(idft_in2,idft_out2,1); dft(DFT_72,idft_in2,idft_out2,1);
break; break;
case 96: case 96:
dft96(idft_in0,idft_out0,1); dft(DFT_96,idft_in0,idft_out0,1);
dft96(idft_in1,idft_out1,1); dft(DFT_96,idft_in1,idft_out1,1);
dft96(idft_in2,idft_out2,1); dft(DFT_96,idft_in2,idft_out2,1);
break; break;
case 108: case 108:
dft108(idft_in0,idft_out0,1); dft(DFT_108,idft_in0,idft_out0,1);
dft108(idft_in1,idft_out1,1); dft(DFT_108,idft_in1,idft_out1,1);
dft108(idft_in2,idft_out2,1); dft(DFT_108,idft_in2,idft_out2,1);
break; break;
case 120: case 120:
dft120(idft_in0,idft_out0,1); dft(DFT_120,idft_in0,idft_out0,1);
dft120(idft_in1,idft_out1,1); dft(DFT_120,idft_in1,idft_out1,1);
dft120(idft_in2,idft_out2,1); dft(DFT_120,idft_in2,idft_out2,1);
break; break;
case 144: case 144:
dft144(idft_in0,idft_out0,1); dft(DFT_144,idft_in0,idft_out0,1);
dft144(idft_in1,idft_out1,1); dft(DFT_144,idft_in1,idft_out1,1);
dft144(idft_in2,idft_out2,1); dft(DFT_144,idft_in2,idft_out2,1);
break; break;
case 180: case 180:
dft180(idft_in0,idft_out0,1); dft(DFT_180,idft_in0,idft_out0,1);
dft180(idft_in1,idft_out1,1); dft(DFT_180,idft_in1,idft_out1,1);
dft180(idft_in2,idft_out2,1); dft(DFT_180,idft_in2,idft_out2,1);
break; break;
case 192: case 192:
dft192(idft_in0,idft_out0,1); dft(DFT_192,idft_in0,idft_out0,1);
dft192(idft_in1,idft_out1,1); dft(DFT_192,idft_in1,idft_out1,1);
dft192(idft_in2,idft_out2,1); dft(DFT_192,idft_in2,idft_out2,1);
break; break;
case 216: case 216:
dft216(idft_in0,idft_out0,1); dft(DFT_216,idft_in0,idft_out0,1);
dft216(idft_in1,idft_out1,1); dft(DFT_216,idft_in1,idft_out1,1);
dft216(idft_in2,idft_out2,1); dft(DFT_216,idft_in2,idft_out2,1);
break; break;
case 240: case 240:
dft240(idft_in0,idft_out0,1); dft(DFT_240,idft_in0,idft_out0,1);
dft240(idft_in1,idft_out1,1); dft(DFT_240,idft_in1,idft_out1,1);
dft240(idft_in2,idft_out2,1); dft(DFT_240,idft_in2,idft_out2,1);
break; break;
case 288: case 288:
dft288(idft_in0,idft_out0,1); dft(DFT_288,idft_in0,idft_out0,1);
dft288(idft_in1,idft_out1,1); dft(DFT_288,idft_in1,idft_out1,1);
dft288(idft_in2,idft_out2,1); dft(DFT_288,idft_in2,idft_out2,1);
break; break;
case 300: case 300:
dft300(idft_in0,idft_out0,1); dft(DFT_300,idft_in0,idft_out0,1);
dft300(idft_in1,idft_out1,1); dft(DFT_300,idft_in1,idft_out1,1);
dft300(idft_in2,idft_out2,1); dft(DFT_300,idft_in2,idft_out2,1);
break; break;
case 324: case 324:
dft324((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_324,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft324((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_324,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft324((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_324,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 360: case 360:
dft360((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_360,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft360((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_360,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft360((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_360,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 384: case 384:
dft384((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_384,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft384((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_384,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft384((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_384,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 432: case 432:
dft432((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_432,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft432((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_432,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft432((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_432,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 480: case 480:
dft480((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_480,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft480((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_480,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft480((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_480,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 540: case 540:
dft540((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_540,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft540((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_540,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft540((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_540,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 576: case 576:
dft576((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_576,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft576((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_576,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft576((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_576,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 600: case 600:
dft600((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_600,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft600((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_600,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft600((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_600,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 648: case 648:
dft648((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_648,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft648((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_648,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft648((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_648,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 720: case 720:
dft720((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_720,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft720((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_720,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft720((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_720,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 768: case 768:
dft768((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_768,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft768((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_768,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft768((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_768,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 864: case 864:
dft864((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_864,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft864((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_864,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft864((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_864,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 900: case 900:
dft900((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_900,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft900((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_900,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft900((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_900,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 960: case 960:
dft960((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_960,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft960((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_960,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft960((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_960,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 972: case 972:
dft972((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_972,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft972((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_972,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft972((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_972,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 1080: case 1080:
dft1080((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_1080,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft1080((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_1080,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft1080((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_1080,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 1152: case 1152:
dft1152((int16_t *)idft_in0,(int16_t *)idft_out0,1); dft(DFT_1152,(int16_t *)idft_in0,(int16_t *)idft_out0,1);
dft1152((int16_t *)idft_in1,(int16_t *)idft_out1,1); dft(DFT_1152,(int16_t *)idft_in1,(int16_t *)idft_out1,1);
dft1152((int16_t *)idft_in2,(int16_t *)idft_out2,1); dft(DFT_1152,(int16_t *)idft_in2,(int16_t *)idft_out2,1);
break; break;
case 1200: case 1200:
dft1200(idft_in0,idft_out0,1); dft(DFT_1200,idft_in0,idft_out0,1);
dft1200(idft_in1,idft_out1,1); dft(DFT_1200,idft_in1,idft_out1,1);
dft1200(idft_in2,idft_out2,1); dft(DFT_1200,idft_in2,idft_out2,1);
break; break;
default: default:
openair1/PHY/MODULATION/ofdm_mod.c
+12 -12
View File
@@ -95,41 +95,41 @@ void PHY_ofdm_mod(int *input, /// pointer to complex input
volatile int *output_ptr=(int*)0; volatile int *output_ptr=(int*)0;
int *temp_ptr=(int*)0; int *temp_ptr=(int*)0;
void (*idft)(int16_t *,int16_t *, int); idft_size_idx_t idftsize;
switch (fftsize) { switch (fftsize) {
case 128: case 128:
idft = idft128; idftsize = IDFT_128;
break; break;
case 256: case 256:
idft = idft256; idftsize = IDFT_256;
break; break;
case 512: case 512:
idft = idft512; idftsize = IDFT_512;
break; break;
case 1024: case 1024:
idft = idft1024; idftsize = IDFT_1024;
break; break;
case 1536: case 1536:
idft = idft1536; idftsize = IDFT_1536;
break; break;
case 2048: case 2048:
idft = idft2048; idftsize = IDFT_2048;
break; break;
case 3072: case 3072:
idft = idft3072; idftsize = IDFT_3072;
break; break;
case 4096: case 4096:
idft = idft4096; idftsize = IDFT_4096;
break; break;
default: default:
idft = idft512; idftsize = IDFT_512;
break; break;
} }
@@ -148,12 +148,12 @@ void PHY_ofdm_mod(int *input, /// pointer to complex input
#ifndef __AVX2__ #ifndef __AVX2__
// handle 128-bit alignment for 128-bit SIMD (SSE4,NEON,AltiVEC) // handle 128-bit alignment for 128-bit SIMD (SSE4,NEON,AltiVEC)
idft((int16_t *)&input[i*fftsize], idft(idftsize,(int16_t *)&input[i*fftsize],
(fftsize==128) ? (int16_t *)temp : (int16_t *)&output[(i*fftsize) + ((1+i)*nb_prefix_samples)], (fftsize==128) ? (int16_t *)temp : (int16_t *)&output[(i*fftsize) + ((1+i)*nb_prefix_samples)],
1); 1);
#else #else
// on AVX2 need 256-bit alignment // on AVX2 need 256-bit alignment
idft((int16_t *)&input[i*fftsize], idft(idftsize,(int16_t *)&input[i*fftsize],
(int16_t *)temp, (int16_t *)temp,
1); 1);
openair1/PHY/MODULATION/slot_fep_nr.c
+25 -25
View File
@@ -61,44 +61,44 @@ int nr_slot_fep(PHY_VARS_NR_UE *ue,
unsigned int rx_offset; unsigned int rx_offset;
void (*dft)(int16_t *,int16_t *, int); dft_size_idx_t dftsize;
int tmp_dft_in[8192] __attribute__ ((aligned (32))); // This is for misalignment issues for 6 and 15 PRBs int tmp_dft_in[8192] __attribute__ ((aligned (32))); // This is for misalignment issues for 6 and 15 PRBs
switch (frame_parms->ofdm_symbol_size) { switch (frame_parms->ofdm_symbol_size) {
case 128: case 128:
dft = dft128; dftsize = DFT_128;
break; break;
case 256: case 256:
dft = dft256; dftsize = DFT_256;
break; break;
case 512: case 512:
dft = dft512; dftsize = DFT_512;
break; break;
case 1024: case 1024:
dft = dft1024; dftsize = DFT_1024;
break; break;
case 1536: case 1536:
dft = dft1536; dftsize = DFT_1536;
break; break;
case 2048: case 2048:
dft = dft2048; dftsize = DFT_2048;
break; break;
case 3072: case 3072:
dft = dft3072; dftsize = DFT_3072;
break; break;
case 4096: case 4096:
dft = dft4096; dftsize = DFT_4096;
break; break;
case 8192: case 8192:
dft = dft8192; dftsize = DFT_8192;
break; break;
default: default:
@@ -146,14 +146,14 @@ int nr_slot_fep(PHY_VARS_NR_UE *ue,
memcpy((void *)tmp_dft_in, memcpy((void *)tmp_dft_in,
(void *) &common_vars->rxdata[aa][rx_offset % frame_length_samples], (void *) &common_vars->rxdata[aa][rx_offset % frame_length_samples],
frame_parms->ofdm_symbol_size*sizeof(int)); frame_parms->ofdm_symbol_size*sizeof(int));
dft((int16_t *)tmp_dft_in, dft(dftsize,(int16_t *)tmp_dft_in,
(int16_t *)&common_vars->common_vars_rx_data_per_thread[ue->current_thread_id[Ns]].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],1); (int16_t *)&common_vars->common_vars_rx_data_per_thread[ue->current_thread_id[Ns]].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],1);
} else { // use dft input from RX buffer directly } else { // use dft input from RX buffer directly
#if UE_TIMING_TRACE #if UE_TIMING_TRACE
start_meas(&ue->rx_dft_stats); start_meas(&ue->rx_dft_stats);
#endif #endif
dft((int16_t *) &common_vars->rxdata[aa][(rx_offset) % frame_length_samples], dft(dftsize,(int16_t *) &common_vars->rxdata[aa][(rx_offset) % frame_length_samples],
(int16_t *)&common_vars->common_vars_rx_data_per_thread[ue->current_thread_id[Ns]].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],1); (int16_t *)&common_vars->common_vars_rx_data_per_thread[ue->current_thread_id[Ns]].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],1);
#if UE_TIMING_TRACE #if UE_TIMING_TRACE
stop_meas(&ue->rx_dft_stats); stop_meas(&ue->rx_dft_stats);
@@ -176,11 +176,11 @@ int nr_slot_fep(PHY_VARS_NR_UE *ue,
memcpy((void *)tmp_dft_in, memcpy((void *)tmp_dft_in,
(void *) &common_vars->rxdata[aa][(rx_offset) % frame_length_samples], (void *) &common_vars->rxdata[aa][(rx_offset) % frame_length_samples],
frame_parms->ofdm_symbol_size*sizeof(int)); frame_parms->ofdm_symbol_size*sizeof(int));
dft((int16_t *)tmp_dft_in, dft(dftsize,(int16_t *)tmp_dft_in,
(int16_t *)&common_vars->common_vars_rx_data_per_thread[ue->current_thread_id[Ns]].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],1); (int16_t *)&common_vars->common_vars_rx_data_per_thread[ue->current_thread_id[Ns]].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],1);
} else { // use dft input from RX buffer directly } else { // use dft input from RX buffer directly
dft((int16_t *) &common_vars->rxdata[aa][(rx_offset) % frame_length_samples], dft(dftsize,(int16_t *) &common_vars->rxdata[aa][(rx_offset) % frame_length_samples],
(int16_t *)&common_vars->common_vars_rx_data_per_thread[ue->current_thread_id[Ns]].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],1); (int16_t *)&common_vars->common_vars_rx_data_per_thread[ue->current_thread_id[Ns]].rxdataF[aa][frame_parms->ofdm_symbol_size*symbol],1);
} }
#if UE_TIMING_TRACE #if UE_TIMING_TRACE
@@ -218,43 +218,43 @@ int nr_slot_fep_ul(NR_DL_FRAME_PARMS *frame_parms,
unsigned int nb_prefix_samples = (no_prefix ? 0 : frame_parms->nb_prefix_samples); unsigned int nb_prefix_samples = (no_prefix ? 0 : frame_parms->nb_prefix_samples);
unsigned int nb_prefix_samples0 = (no_prefix ? 0 : frame_parms->nb_prefix_samples0); unsigned int nb_prefix_samples0 = (no_prefix ? 0 : frame_parms->nb_prefix_samples0);
void (*dft)(int16_t *,int16_t *, int); dft_size_idx_t dftsize;
switch (frame_parms->ofdm_symbol_size) { switch (frame_parms->ofdm_symbol_size) {
case 128: case 128:
dft = dft128; dftsize = DFT_128;
break; break;
case 256: case 256:
dft = dft256; dftsize = DFT_256;
break; break;
case 512: case 512:
dft = dft512; dftsize = DFT_512;
break; break;
case 1024: case 1024:
dft = dft1024; dftsize = DFT_1024;
break; break;
case 1536: case 1536:
dft = dft1536; dftsize = DFT_1536;
break; break;
case 2048: case 2048:
dft = dft2048; dftsize = DFT_2048;
break; break;
case 4096: case 4096:
dft = dft4096; dftsize = DFT_4096;
break; break;
case 8192: case 8192:
dft = dft8192; dftsize = DFT_8192;
break; break;
default: default:
dft = dft512; dftsize = DFT_512;
break; break;
} }
@@ -266,7 +266,7 @@ int nr_slot_fep_ul(NR_DL_FRAME_PARMS *frame_parms,
else else
rxdata_offset = slot_offset + nb_prefix_samples0 + (symbol * (frame_parms->ofdm_symbol_size + nb_prefix_samples)) - SOFFSET; rxdata_offset = slot_offset + nb_prefix_samples0 + (symbol * (frame_parms->ofdm_symbol_size + nb_prefix_samples)) - SOFFSET;
dft((int16_t *)&rxdata[rxdata_offset], dft(dftsize,(int16_t *)&rxdata[rxdata_offset],
(int16_t *)&rxdataF[symbol * frame_parms->ofdm_symbol_size], 1); (int16_t *)&rxdataF[symbol * frame_parms->ofdm_symbol_size], 1);
return(0); return(0);
openair1/PHY/MODULATION/slot_fep_ul.c
+11 -11
View File
@@ -44,7 +44,7 @@ int slot_fep_ul(RU_t *ru,
unsigned int slot_offset; unsigned int slot_offset;
void (*dft)(int16_t *,int16_t *, int); dft_size_idx_t dftsize;
int tmp_dft_in[2048] __attribute__ ((aligned (32))); // This is for misalignment issues for 6 and 15 PRBs int tmp_dft_in[2048] __attribute__ ((aligned (32))); // This is for misalignment issues for 6 and 15 PRBs
unsigned int frame_length_samples = fp->samples_per_tti * 10; unsigned int frame_length_samples = fp->samples_per_tti * 10;
@@ -52,31 +52,31 @@ int slot_fep_ul(RU_t *ru,
switch (fp->ofdm_symbol_size) { switch (fp->ofdm_symbol_size) {
case 128: case 128:
dft = dft128; dftsize = DFT_128;
break; break;
case 256: case 256:
dft = dft256; dftsize = DFT_256;
break; break;
case 512: case 512:
dft = dft512; dftsize = DFT_512;
break; break;
case 1024: case 1024:
dft = dft1024; dftsize = DFT_1024;
break; break;
case 1536: case 1536:
dft = dft1536; dftsize = DFT_1536;
break; break;
case 2048: case 2048:
dft = dft2048; dftsize = DFT_2048;
break; break;
default: default:
dft = dft512; dftsize = DFT_512;
break; break;
} }
@@ -109,7 +109,7 @@ int slot_fep_ul(RU_t *ru,
LOG_D(PHY,"slot_fep: symbol 0 %d dB\n", LOG_D(PHY,"slot_fep: symbol 0 %d dB\n",
dB_fixed(signal_energy(&common->rxdata_7_5kHz[aa][rx_offset],fp->ofdm_symbol_size))); dB_fixed(signal_energy(&common->rxdata_7_5kHz[aa][rx_offset],fp->ofdm_symbol_size)));
#endif #endif
dft( (int16_t *)&common->rxdata_7_5kHz[aa][rx_offset], dft( dftsize,(int16_t *)&common->rxdata_7_5kHz[aa][rx_offset],
(int16_t *)&common->rxdataF[aa][fp->ofdm_symbol_size*symbol], (int16_t *)&common->rxdataF[aa][fp->ofdm_symbol_size*symbol],
1 1
); );
@@ -121,13 +121,13 @@ int slot_fep_ul(RU_t *ru,
memcpy((void *)&tmp_dft_in, memcpy((void *)&tmp_dft_in,
(void *)&common->rxdata_7_5kHz[aa][(rx_offset % frame_length_samples)], (void *)&common->rxdata_7_5kHz[aa][(rx_offset % frame_length_samples)],
fp->ofdm_symbol_size*sizeof(int)); fp->ofdm_symbol_size*sizeof(int));
dft( (short *) tmp_dft_in, dft( dftsize,(short *) tmp_dft_in,
(short*) &common->rxdataF[aa][fp->ofdm_symbol_size*symbol], (short*) &common->rxdataF[aa][fp->ofdm_symbol_size*symbol],
1 1
); );
} }
else{ else{
dft( (short *)&common->rxdata_7_5kHz[aa][rx_offset], dft( dftsize,(short *)&common->rxdata_7_5kHz[aa][rx_offset],
(short*)&common->rxdataF[aa][fp->ofdm_symbol_size*symbol], (short*)&common->rxdataF[aa][fp->ofdm_symbol_size*symbol],
1 1
); );
openair1/PHY/TOOLS/lte_dfts.c → openair1/PHY/TOOLS/oai_dfts.c
+67 -48
View File
@@ -31,11 +31,10 @@
#ifndef M_PI #ifndef M_PI
#define M_PI 3.14159265358979323846 #define M_PI 3.14159265358979323846
#endif #endif
#define OAIDFTS
#ifndef MR_MAIN #ifndef MR_MAIN
#include "PHY/defs_common.h" #include "PHY/defs_common.h"
#include "PHY/impl_defs_top.h" #include "PHY/impl_defs_top.h"
#include "tools_defs.h"
#else #else
#include "time_meas.h" #include "time_meas.h"
#include "LOG/log.h" #include "LOG/log.h"
@@ -64,6 +63,9 @@ const static int16_t reflip[32] __attribute__((aligned(32))) = {1,-1,1,-1,1,-1,
#if defined(__x86_64__) || defined(__i386__) #if defined(__x86_64__) || defined(__i386__)
static inline void cmac(__m128i a,__m128i b, __m128i *re32, __m128i *im32) __attribute__((always_inline)); static inline void cmac(__m128i a,__m128i b, __m128i *re32, __m128i *im32) __attribute__((always_inline));
static inline void cmac(__m128i a,__m128i b, __m128i *re32, __m128i *im32) static inline void cmac(__m128i a,__m128i b, __m128i *re32, __m128i *im32)
@@ -2588,7 +2590,7 @@ const static int16_t tw64c[96] __attribute__((aligned(32))) = {
#endif #endif
#ifndef __AVX2__ #ifndef __AVX2__
void dft64(int16_t *x,int16_t *y,int scale) void dft64(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd_q15_t xtmp[16],ytmp[16],*tw64a_128=(simd_q15_t *)tw64a,*tw64b_128=(simd_q15_t *)tw64b,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y; simd_q15_t xtmp[16],ytmp[16],*tw64a_128=(simd_q15_t *)tw64a,*tw64b_128=(simd_q15_t *)tw64b,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y;
@@ -2702,7 +2704,7 @@ void dft64(int16_t *x,int16_t *y,int scale)
} }
#else // __AVX2__ #else // __AVX2__
void dft64(int16_t *x,int16_t *y,int scale) void dft64(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[16],ytmp[16],*tw64a_256=(simd256_q15_t *)tw64a,*tw64b_256=(simd256_q15_t *)tw64b,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y; simd256_q15_t xtmp[16],ytmp[16],*tw64a_256=(simd256_q15_t *)tw64a,*tw64b_256=(simd256_q15_t *)tw64b,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y;
@@ -2850,7 +2852,7 @@ void dft64(int16_t *x,int16_t *y,int scale)
#endif #endif
#ifndef __AVX2__ #ifndef __AVX2__
void idft64(int16_t *x,int16_t *y,int scale) void idft64(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd_q15_t xtmp[16],ytmp[16],*tw64a_128=(simd_q15_t *)tw64,*tw64b_128=(simd_q15_t *)tw64c,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y; simd_q15_t xtmp[16],ytmp[16],*tw64a_128=(simd_q15_t *)tw64,*tw64b_128=(simd_q15_t *)tw64c,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y;
@@ -2942,7 +2944,7 @@ void idft64(int16_t *x,int16_t *y,int scale)
} }
#else // __AVX2__ #else // __AVX2__
void idft64(int16_t *x,int16_t *y,int scale) void idft64(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[16],ytmp[16],*tw64a_256=(simd256_q15_t *)tw64,*tw64b_256=(simd256_q15_t *)tw64c,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y; simd256_q15_t xtmp[16],ytmp[16],*tw64a_256=(simd256_q15_t *)tw64,*tw64b_256=(simd256_q15_t *)tw64c,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y;
@@ -3052,7 +3054,7 @@ int16_t tw128b[128] __attribute__((aligned(32))) = {0,32767,-1608,32727,-3212,32
int16_t tw128c[128] __attribute__((aligned(32))) = {0,32767,1608,32727,3212,32609,4808,32412,6393,32137,7962,31785,9512,31356,11039,30851,12540,30272,14010,29621,15447,28897,16846,28105,18205,27244,19520,26318,20788,25329,22005,24278,23170,23169,24279,22004,25330,20787,26319,19519,27245,18204,28106,16845,28898,15446,29622,14009,30273,12539,30852,11038,31357,9511,31786,7961,32138,6392,32413,4807,32610,3211,32728,1607,32767,0,32728,-1608,32610,-3212,32413,-4808,32138,-6393,31786,-7962,31357,-9512,30852,-11039,30273,-12540,29622,-14010,28898,-15447,28106,-16846,27245,-18205,26319,-19520,25330,-20788,24279,-22005,23170,-23170,22005,-24279,20788,-25330,19520,-26319,18205,-27245,16846,-28106,15447,-28898,14010,-29622,12540,-30273,11039,-30852,9512,-31357,7962,-31786,6393,-32138,4808,-32413,3212,-32610,1608,-32728}; int16_t tw128c[128] __attribute__((aligned(32))) = {0,32767,1608,32727,3212,32609,4808,32412,6393,32137,7962,31785,9512,31356,11039,30851,12540,30272,14010,29621,15447,28897,16846,28105,18205,27244,19520,26318,20788,25329,22005,24278,23170,23169,24279,22004,25330,20787,26319,19519,27245,18204,28106,16845,28898,15446,29622,14009,30273,12539,30852,11038,31357,9511,31786,7961,32138,6392,32413,4807,32610,3211,32728,1607,32767,0,32728,-1608,32610,-3212,32413,-4808,32138,-6393,31786,-7962,31357,-9512,30852,-11039,30273,-12540,29622,-14010,28898,-15447,28106,-16846,27245,-18205,26319,-19520,25330,-20788,24279,-22005,23170,-23170,22005,-24279,20788,-25330,19520,-26319,18205,-27245,16846,-28106,15447,-28898,14010,-29622,12540,-30273,11039,-30852,9512,-31357,7962,-31786,6393,-32138,4808,-32413,3212,-32610,1608,-32728};
#ifndef __AVX2__ #ifndef __AVX2__
void dft128(int16_t *x,int16_t *y,int scale) void dft128(int16_t *x,int16_t *y,unsigned char scale)
{ {
simdshort_q15_t xtmp[64],*x64 = (simdshort_q15_t *)x; simdshort_q15_t xtmp[64],*x64 = (simdshort_q15_t *)x;
@@ -3159,7 +3161,7 @@ void dft128(int16_t *x,int16_t *y,int scale)
} }
#else // __AVX2__ #else // __AVX2__
void dft128(int16_t *x,int16_t *y,int scale) void dft128(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[16],*x256 = (simd256_q15_t *)x; simd256_q15_t xtmp[16],*x256 = (simd256_q15_t *)x;
@@ -3229,7 +3231,7 @@ void dft128(int16_t *x,int16_t *y,int scale)
#endif #endif
#ifndef __AVX2__ #ifndef __AVX2__
void idft128(int16_t *x,int16_t *y,int scale) void idft128(int16_t *x,int16_t *y,unsigned char scale)
{ {
simdshort_q15_t xtmp[64],*x64 = (simdshort_q15_t *)x; simdshort_q15_t xtmp[64],*x64 = (simdshort_q15_t *)x;
@@ -3328,7 +3330,7 @@ void idft128(int16_t *x,int16_t *y,int scale)
} }
#else // __AVX2__ #else // __AVX2__
void idft128(int16_t *x,int16_t *y,int scale) void idft128(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[16],*x256 = (simd256_q15_t *)x; simd256_q15_t xtmp[16],*x256 = (simd256_q15_t *)x;
@@ -3401,7 +3403,7 @@ int16_t tw256b[384] __attribute__((aligned(32))) = {0,32767,-805,32757,-1608,327
0,32767,-2411,32678,-4808,32412,-7180,31970,-9512,31356,-11793,30571,-14010,29621,-16151,28510,-18205,27244,-20160,25831,-22005,24278,-23732,22594,-25330,20787,-26790,18867,-28106,16845,-29269,14732,-30273,12539,-31114,10278,-31786,7961,-32285,5601,-32610,3211,-32758,804,-32728,-1608,-32521,-4012,-32138,-6393,-31581,-8740,-30852,-11039,-29956,-13279,-28898,-15447,-27684,-17531,-26319,-19520,-24812,-21403,-23170,-23170,-21403,-24812,-19520,-26319,-17531,-27684,-15447,-28898,-13279,-29956,-11039,-30852,-8740,-31581,-6393,-32138,-4012,-32521,-1608,-32728,804,-32758,3211,-32610,5601,-32285,7961,-31786,10278,-31114,12539,-30273,14732,-29269,16845,-28106,18867,-26790,20787,-25330,22594,-23732,24278,-22005,25831,-20160,27244,-18205,28510,-16151,29621,-14010,30571,-11793,31356,-9512,31970,-7180,32412,-4808,32678,-2411 0,32767,-2411,32678,-4808,32412,-7180,31970,-9512,31356,-11793,30571,-14010,29621,-16151,28510,-18205,27244,-20160,25831,-22005,24278,-23732,22594,-25330,20787,-26790,18867,-28106,16845,-29269,14732,-30273,12539,-31114,10278,-31786,7961,-32285,5601,-32610,3211,-32758,804,-32728,-1608,-32521,-4012,-32138,-6393,-31581,-8740,-30852,-11039,-29956,-13279,-28898,-15447,-27684,-17531,-26319,-19520,-24812,-21403,-23170,-23170,-21403,-24812,-19520,-26319,-17531,-27684,-15447,-28898,-13279,-29956,-11039,-30852,-8740,-31581,-6393,-32138,-4012,-32521,-1608,-32728,804,-32758,3211,-32610,5601,-32285,7961,-31786,10278,-31114,12539,-30273,14732,-29269,16845,-28106,18867,-26790,20787,-25330,22594,-23732,24278,-22005,25831,-20160,27244,-18205,28510,-16151,29621,-14010,30571,-11793,31356,-9512,31970,-7180,32412,-4808,32678,-2411
}; };
#ifndef __AVX2__ #ifndef __AVX2__
void dft256(int16_t *x,int16_t *y,int scale) void dft256(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd_q15_t xtmp[64],ytmp[64],*tw256a_128p=(simd_q15_t *)tw256a,*tw256b_128p=(simd_q15_t *)tw256b,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y; simd_q15_t xtmp[64],ytmp[64],*tw256a_128p=(simd_q15_t *)tw256a,*tw256b_128p=(simd_q15_t *)tw256b,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y;
@@ -3513,7 +3515,7 @@ void dft256(int16_t *x,int16_t *y,int scale)
void idft256(int16_t *x,int16_t *y,int scale) void idft256(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd_q15_t xtmp[64],ytmp[64],*tw256_128p=(simd_q15_t *)tw256,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y; simd_q15_t xtmp[64],ytmp[64],*tw256_128p=(simd_q15_t *)tw256,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y;
@@ -3571,7 +3573,7 @@ void idft256(int16_t *x,int16_t *y,int scale)
#else //__AVX2__ #else //__AVX2__
void dft256(int16_t *x,int16_t *y,int scale) void dft256(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[32],ytmp[32],*tw256a_256p=(simd256_q15_t *)tw256a,*tw256b_256p=(simd256_q15_t *)tw256b,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y; simd256_q15_t xtmp[32],ytmp[32],*tw256a_256p=(simd256_q15_t *)tw256a,*tw256b_256p=(simd256_q15_t *)tw256b,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y;
@@ -3663,7 +3665,7 @@ void dft256(int16_t *x,int16_t *y,int scale)
} }
void idft256(int16_t *x,int16_t *y,int scale) void idft256(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[32],ytmp[32],*tw256_256p=(simd256_q15_t *)tw256,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y; simd256_q15_t xtmp[32],ytmp[32],*tw256_256p=(simd256_q15_t *)tw256,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y;
@@ -3768,7 +3770,7 @@ int16_t tw512c[512] __attribute__((aligned(32))) = {
}; };
#ifndef __AVX2__ #ifndef __AVX2__
void dft512(int16_t *x,int16_t *y,int scale) void dft512(int16_t *x,int16_t *y,unsigned char scale)
{ {
simdshort_q15_t xtmp[256],*xtmpp,*x64 = (simdshort_q15_t *)x; simdshort_q15_t xtmp[256],*xtmpp,*x64 = (simdshort_q15_t *)x;
@@ -3888,7 +3890,7 @@ void dft512(int16_t *x,int16_t *y,int scale)
} }
void idft512(int16_t *x,int16_t *y,int scale) void idft512(int16_t *x,int16_t *y,unsigned char scale)
{ {
simdshort_q15_t xtmp[256],*xtmpp,*x64 = (simdshort_q15_t *)x; simdshort_q15_t xtmp[256],*xtmpp,*x64 = (simdshort_q15_t *)x;
@@ -3980,7 +3982,7 @@ void idft512(int16_t *x,int16_t *y,int scale)
#else //__AVX2__ #else //__AVX2__
void dft512(int16_t *x,int16_t *y,int scale) void dft512(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[64],*x256 = (simd256_q15_t *)x; simd256_q15_t xtmp[64],*x256 = (simd256_q15_t *)x;
@@ -4062,7 +4064,7 @@ void dft512(int16_t *x,int16_t *y,int scale)
} }
void idft512(int16_t *x,int16_t *y,int scale) void idft512(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[64],*x256 = (simd256_q15_t *)x; simd256_q15_t xtmp[64],*x256 = (simd256_q15_t *)x;
@@ -4149,7 +4151,7 @@ void idft512(int16_t *x,int16_t *y,int scale)
int16_t tw1024[1536] __attribute__((aligned(32))); int16_t tw1024[1536] __attribute__((aligned(32)));
#ifndef __AVX2__ #ifndef __AVX2__
void dft1024(int16_t *x,int16_t *y,int scale) void dft1024(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd_q15_t xtmp[256],ytmp[256],*tw1024_128p=(simd_q15_t *)tw1024,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y; simd_q15_t xtmp[256],ytmp[256],*tw1024_128p=(simd_q15_t *)tw1024,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y;
@@ -4205,7 +4207,7 @@ void dft1024(int16_t *x,int16_t *y,int scale)
} }
void idft1024(int16_t *x,int16_t *y,int scale) void idft1024(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd_q15_t xtmp[256],ytmp[256],*tw1024_128p=(simd_q15_t *)tw1024,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y; simd_q15_t xtmp[256],ytmp[256],*tw1024_128p=(simd_q15_t *)tw1024,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y;
@@ -4262,7 +4264,7 @@ void idft1024(int16_t *x,int16_t *y,int scale)
} }
#else //__AVX2__ #else //__AVX2__
void dft1024(int16_t *x,int16_t *y,int scale) void dft1024(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[128],ytmp[128],*tw1024_256p=(simd256_q15_t *)tw1024,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y; simd256_q15_t xtmp[128],ytmp[128],*tw1024_256p=(simd256_q15_t *)tw1024,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y;
@@ -4318,7 +4320,7 @@ void dft1024(int16_t *x,int16_t *y,int scale)
} }
void idft1024(int16_t *x,int16_t *y,int scale) void idft1024(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[128],ytmp[128],*tw1024_256p=(simd256_q15_t *)tw1024,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y; simd256_q15_t xtmp[128],ytmp[128],*tw1024_256p=(simd256_q15_t *)tw1024,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y;
@@ -4378,7 +4380,7 @@ void idft1024(int16_t *x,int16_t *y,int scale)
int16_t tw2048[2048] __attribute__((aligned(32))); int16_t tw2048[2048] __attribute__((aligned(32)));
#ifndef __AVX2__ #ifndef __AVX2__
void dft2048(int16_t *x,int16_t *y,int scale) void dft2048(int16_t *x,int16_t *y,unsigned char scale)
{ {
simdshort_q15_t xtmp[1024],*xtmpp,*x64 = (simdshort_q15_t *)x; simdshort_q15_t xtmp[1024],*xtmpp,*x64 = (simdshort_q15_t *)x;
@@ -4469,7 +4471,7 @@ void dft2048(int16_t *x,int16_t *y,int scale)
} }
void idft2048(int16_t *x,int16_t *y,int scale) void idft2048(int16_t *x,int16_t *y,unsigned char scale)
{ {
simdshort_q15_t xtmp[1024],*xtmpp,*x64 = (simdshort_q15_t *)x; simdshort_q15_t xtmp[1024],*xtmpp,*x64 = (simdshort_q15_t *)x;
@@ -4561,7 +4563,7 @@ void idft2048(int16_t *x,int16_t *y,int scale)
#else // __AVX2__ #else // __AVX2__
void dft2048(int16_t *x,int16_t *y,int scale) void dft2048(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[256],*xtmpp,*x256 = (simd256_q15_t *)x; simd256_q15_t xtmp[256],*xtmpp,*x256 = (simd256_q15_t *)x;
@@ -4652,7 +4654,7 @@ void dft2048(int16_t *x,int16_t *y,int scale)
} }
void idft2048(int16_t *x,int16_t *y,int scale) void idft2048(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[256],*xtmpp,*x256 = (simd256_q15_t *)x; simd256_q15_t xtmp[256],*xtmpp,*x256 = (simd256_q15_t *)x;
@@ -4749,7 +4751,7 @@ void idft2048(int16_t *x,int16_t *y,int scale)
int16_t tw4096[3*2*1024]; int16_t tw4096[3*2*1024];
#ifndef __AVX2__ #ifndef __AVX2__
void dft4096(int16_t *x,int16_t *y,int scale) void dft4096(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd_q15_t xtmp[1024],ytmp[1024],*tw4096_128p=(simd_q15_t *)tw4096,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y; simd_q15_t xtmp[1024],ytmp[1024],*tw4096_128p=(simd_q15_t *)tw4096,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y;
@@ -4807,7 +4809,7 @@ void dft4096(int16_t *x,int16_t *y,int scale)
void idft4096(int16_t *x,int16_t *y,int scale) void idft4096(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd_q15_t xtmp[1024],ytmp[1024],*tw4096_128p=(simd_q15_t *)tw4096,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y; simd_q15_t xtmp[1024],ytmp[1024],*tw4096_128p=(simd_q15_t *)tw4096,*x128=(simd_q15_t *)x,*y128=(simd_q15_t *)y,*y128p=(simd_q15_t *)y;
@@ -4864,7 +4866,7 @@ void idft4096(int16_t *x,int16_t *y,int scale)
} }
#else //__AVX2__ #else //__AVX2__
void dft4096(int16_t *x,int16_t *y,int scale) void dft4096(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[512],ytmp[512],*tw4096_256p=(simd256_q15_t *)tw4096,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y; simd256_q15_t xtmp[512],ytmp[512],*tw4096_256p=(simd256_q15_t *)tw4096,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y;
@@ -4920,7 +4922,7 @@ void dft4096(int16_t *x,int16_t *y,int scale)
} }
void idft4096(int16_t *x,int16_t *y,int scale) void idft4096(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[512],ytmp[512],*tw4096_256p=(simd256_q15_t *)tw4096,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y; simd256_q15_t xtmp[512],ytmp[512],*tw4096_256p=(simd256_q15_t *)tw4096,*x256=(simd256_q15_t *)x,*y256=(simd256_q15_t *)y,*y256p=(simd256_q15_t *)y;
@@ -4982,7 +4984,7 @@ void idft4096(int16_t *x,int16_t *y,int scale)
int16_t tw8192[2*4096] __attribute__((aligned(32))); int16_t tw8192[2*4096] __attribute__((aligned(32)));
#ifndef __AVX2__ #ifndef __AVX2__
void dft8192(int16_t *x,int16_t *y,int scale) void dft8192(int16_t *x,int16_t *y,unsigned char scale)
{ {
simdshort_q15_t xtmp[4096],*xtmpp,*x64 = (simdshort_q15_t *)x; simdshort_q15_t xtmp[4096],*xtmpp,*x64 = (simdshort_q15_t *)x;
@@ -5072,7 +5074,7 @@ void dft8192(int16_t *x,int16_t *y,int scale)
} }
void idft8192(int16_t *x,int16_t *y,int scale) void idft8192(int16_t *x,int16_t *y,unsigned char scale)
{ {
simdshort_q15_t xtmp[4096],*xtmpp,*x64 = (simdshort_q15_t *)x; simdshort_q15_t xtmp[4096],*xtmpp,*x64 = (simdshort_q15_t *)x;
@@ -5163,7 +5165,7 @@ void idft8192(int16_t *x,int16_t *y,int scale)
} }
#else // __AVX2__ #else // __AVX2__
void dft8192(int16_t *x,int16_t *y,int scale) void dft8192(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[1024],*xtmpp,*x256 = (simd256_q15_t *)x; simd256_q15_t xtmp[1024],*xtmpp,*x256 = (simd256_q15_t *)x;
@@ -5254,7 +5256,7 @@ void dft8192(int16_t *x,int16_t *y,int scale)
} }
void idft8192(int16_t *x,int16_t *y,int scale) void idft8192(int16_t *x,int16_t *y,unsigned char scale)
{ {
simd256_q15_t xtmp[1024],*xtmpp,*x256 = (simd256_q15_t *)x; simd256_q15_t xtmp[1024],*xtmpp,*x256 = (simd256_q15_t *)x;
@@ -5350,7 +5352,7 @@ void idft8192(int16_t *x,int16_t *y,int scale)
int16_t twa1536[1024],twb1536[1024]; int16_t twa1536[1024],twb1536[1024];
// 512 x 3 // 512 x 3
void idft1536(int16_t *input, int16_t *output, int scale) void idft1536(int16_t *input, int16_t *output, unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][512 ]__attribute__((aligned(32))); uint32_t tmp[3][512 ]__attribute__((aligned(32)));
@@ -5402,7 +5404,7 @@ void idft1536(int16_t *input, int16_t *output, int scale)
} }
void dft1536(int16_t *input, int16_t *output, int scale) void dft1536(int16_t *input, int16_t *output, unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][512] __attribute__((aligned(32))); uint32_t tmp[3][512] __attribute__((aligned(32)));
@@ -5467,7 +5469,7 @@ void dft1536(int16_t *input, int16_t *output, int scale)
int16_t twa3072[2048] __attribute__((aligned(32))); int16_t twa3072[2048] __attribute__((aligned(32)));
int16_t twb3072[2048] __attribute__((aligned(32))); int16_t twb3072[2048] __attribute__((aligned(32)));
// 1024 x 3 // 1024 x 3
void dft3072(int16_t *input, int16_t *output,int scale) void dft3072(int16_t *input, int16_t *output,unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][1024] __attribute__((aligned(32))); uint32_t tmp[3][1024] __attribute__((aligned(32)));
@@ -5517,7 +5519,7 @@ void dft3072(int16_t *input, int16_t *output,int scale)
_m_empty(); _m_empty();
} }
void idft3072(int16_t *input, int16_t *output,int scale) void idft3072(int16_t *input, int16_t *output,unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][1024]__attribute__((aligned(32))); uint32_t tmp[3][1024]__attribute__((aligned(32)));
@@ -5571,7 +5573,7 @@ void idft3072(int16_t *input, int16_t *output,int scale)
int16_t twa6144[4096] __attribute__((aligned(32))); int16_t twa6144[4096] __attribute__((aligned(32)));
int16_t twb6144[4096] __attribute__((aligned(32))); int16_t twb6144[4096] __attribute__((aligned(32)));
void idft6144(int16_t *input, int16_t *output,int scale) void idft6144(int16_t *input, int16_t *output,unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][2048] __attribute__((aligned(32))); uint32_t tmp[3][2048] __attribute__((aligned(32)));
@@ -5631,7 +5633,7 @@ void idft6144(int16_t *input, int16_t *output,int scale)
} }
void dft6144(int16_t *input, int16_t *output,int scale) void dft6144(int16_t *input, int16_t *output,unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][2048] __attribute__((aligned(32))); uint32_t tmp[3][2048] __attribute__((aligned(32)));
@@ -5696,7 +5698,7 @@ void dft6144(int16_t *input, int16_t *output,int scale)
int16_t twa12288[8192] __attribute__((aligned(32))); int16_t twa12288[8192] __attribute__((aligned(32)));
int16_t twb12288[8192] __attribute__((aligned(32))); int16_t twb12288[8192] __attribute__((aligned(32)));
// 4096 x 3 // 4096 x 3
void dft12288(int16_t *input, int16_t *output,int scale) void dft12288(int16_t *input, int16_t *output,unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][4096] __attribute__((aligned(32))); uint32_t tmp[3][4096] __attribute__((aligned(32)));
@@ -5756,7 +5758,7 @@ void dft12288(int16_t *input, int16_t *output,int scale)
} }
void idft12288(int16_t *input, int16_t *output,int scale) void idft12288(int16_t *input, int16_t *output,unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][4096] __attribute__((aligned(32))); uint32_t tmp[3][4096] __attribute__((aligned(32)));
@@ -5820,7 +5822,7 @@ void idft12288(int16_t *input, int16_t *output,int scale)
int16_t twa18432[12288] __attribute__((aligned(32))); int16_t twa18432[12288] __attribute__((aligned(32)));
int16_t twb18432[12288] __attribute__((aligned(32))); int16_t twb18432[12288] __attribute__((aligned(32)));
// 6144 x 3 // 6144 x 3
void dft18432(int16_t *input, int16_t *output,int scale) { void dft18432(int16_t *input, int16_t *output,unsigned char scale) {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][6144] __attribute__((aligned(32))); uint32_t tmp[3][6144] __attribute__((aligned(32)));
@@ -5868,7 +5870,7 @@ void dft18432(int16_t *input, int16_t *output,int scale) {
_m_empty(); _m_empty();
} }
void idft18432(int16_t *input, int16_t *output,int scale) { void idft18432(int16_t *input, int16_t *output,unsigned char scale) {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][6144] __attribute__((aligned(32))); uint32_t tmp[3][6144] __attribute__((aligned(32)));
@@ -5920,7 +5922,7 @@ void idft18432(int16_t *input, int16_t *output,int scale) {
int16_t twa24576[16384] __attribute__((aligned(32))); int16_t twa24576[16384] __attribute__((aligned(32)));
int16_t twb24576[16384] __attribute__((aligned(32))); int16_t twb24576[16384] __attribute__((aligned(32)));
// 8192 x 3 // 8192 x 3
void dft24576(int16_t *input, int16_t *output,int scale) void dft24576(int16_t *input, int16_t *output,unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][8192] __attribute__((aligned(32))); uint32_t tmp[3][8192] __attribute__((aligned(32)));
@@ -5983,7 +5985,7 @@ void dft24576(int16_t *input, int16_t *output,int scale)
} }
} }
void idft24576(int16_t *input, int16_t *output,int scale) void idft24576(int16_t *input, int16_t *output,unsigned char scale)
{ {
int i,i2,j; int i,i2,j;
uint32_t tmp[3][8192] __attribute__((aligned(32))); uint32_t tmp[3][8192] __attribute__((aligned(32)));
@@ -6194,7 +6196,7 @@ static inline void dft12f(simd_q15_t *x0,
void dft12(int16_t *x,int16_t *y) void dft12(int16_t *x,int16_t *y ,unsigned char scale_flag)
{ {
simd_q15_t *x128 = (simd_q15_t *)x,*y128 = (simd_q15_t *)y; simd_q15_t *x128 = (simd_q15_t *)x,*y128 = (simd_q15_t *)y;
@@ -8644,7 +8646,8 @@ void init_rad5_rep(int N,int16_t *twa,int16_t *twb,int16_t *twc,int16_t *twd) {
twd+=8; twd+=8;
} }
} }
/*----------------------------------------------------------------*/
/* dft library entry points: */
void init_dfts(void) void init_dfts(void)
{ {
@@ -8693,8 +8696,24 @@ void init_dfts(void)
init_rad3_rep(1080,twa1080,twb1080); init_rad3_rep(1080,twa1080,twb1080);
init_rad4_rep(1152,twa1152,twb1152,twc1152); init_rad4_rep(1152,twa1152,twb1152,twc1152);
init_rad4_rep(1200,twa1200,twb1200,twc1200); init_rad4_rep(1200,twa1200,twb1200,twc1200);
} }
void dft(uint8_t sizeidx, int16_t *sigF,int16_t *sig,unsigned char scale_flag){
AssertFatal((sizeidx>=0 && sizeidx<(int)DFT_SIZE_IDXTABLESIZE),"Invalid dft size index %i\n",sizeidx);
dft_ftab[sizeidx](sigF,sig,scale_flag);
};
void idft(uint8_t sizeidx, int16_t *sigF,int16_t *sig,unsigned char scale_flag){
AssertFatal((sizeidx>=0 && sizeidx<(int)IDFT_SIZE_IDXTABLESIZE),"Invalid idft size index %i\n",sizeidx);
idft_ftab[sizeidx](sigF,sig,scale_flag);
};
/*---------------------------------------------------------------------------------------*/
#ifdef MR_MAIN #ifdef MR_MAIN
#include <string.h> #include <string.h>
#include <stdio.h> #include <stdio.h>
openair1/PHY/TOOLS/tools_defs.h
+217 -62
View File
@@ -181,29 +181,234 @@ This function performs optimized fixed-point radix-2 FFT/IFFT.
); );
*/ */
void idft1536(int16_t *sigF,int16_t *sig,int scale);
void idft6144(int16_t *sigF,int16_t *sig,int scale);
void idft12288(int16_t *sigF,int16_t *sig,int scale);
void idft18432(int16_t *sigF,int16_t *sig,int scale);
void idft3072(int16_t *sigF,int16_t *sig,int scale);
void idft24576(int16_t *sigF,int16_t *sig,int scale);
void dft1536(int16_t *sigF,int16_t *sig,int scale);
void dft3072(int16_t *sigF,int16_t *sig,int scale); #ifdef OAIDFTS
typedef void(*adftfunc_t)(int16_t *sigF,int16_t *sig,unsigned char scale_flag);
typedef void(*aidftfunc_t)(int16_t *sigF,int16_t *sig,unsigned char scale_flag);
void dft6144(int16_t *sigF,int16_t *sig,int scale); void dft12(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft24(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft36(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft48(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft60(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft72(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft96(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft108(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft120(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft144(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft180(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft192(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft216(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft240(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft288(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft300(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft324(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft360(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft384(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft432(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft480(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft540(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft576(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft600(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft648(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft720(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft768(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft864(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft900(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft960(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft972(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft1080(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft1152(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft1200(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft12288(int16_t *sigF,int16_t *sig,int scale); void dft64(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft128(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft256(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft512(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft1024(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft2048(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft4096(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft8192(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft18432(int16_t *sigF,int16_t *sig,int scale); void idft64(int16_t *x,int16_t *y,uint8_t scale_flag);
void idft128(int16_t *x,int16_t *y,uint8_t scale_flag);
void idft256(int16_t *x,int16_t *y,uint8_t scale_flag);
void idft512(int16_t *x,int16_t *y,uint8_t scale_flag);
void idft1024(int16_t *x,int16_t *y,uint8_t scale_flag);
void idft2048(int16_t *x,int16_t *y,uint8_t scale_flag);
void idft4096(int16_t *x,int16_t *y,uint8_t scale_flag);
void idft8192(int16_t *x,int16_t *y,uint8_t scale_flag);
void idft1536(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void idft6144(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void idft12288(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void idft18432(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void idft3072(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void idft24576(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void dft1536(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void dft3072(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void dft6144(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void dft12288(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void dft18432(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
void dft24576(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
#else
typedef void(*dftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);
typedef void(*idftfunc_t)(uint8_t sizeidx,int16_t *sigF,int16_t *sig,unsigned char scale_flag);
extern dftfunc_t dft;
extern idftfunc_t idft;
#endif
typedef enum DFT_size_idx {
DFT_12,
DFT_24,
DFT_36,
DFT_48,
DFT_60,
DFT_72,
DFT_96,
DFT_108,
DFT_120,
DFT_128,
DFT_144,
DFT_180,
DFT_192,
DFT_216,
DFT_240,
DFT_256,
DFT_288,
DFT_300,
DFT_324,
DFT_360,
DFT_384,
DFT_432,
DFT_480,
DFT_512,
DFT_540,
DFT_576,
DFT_600,
DFT_648,
DFT_720,
DFT_768,
DFT_864,
DFT_900,
DFT_960,
DFT_972,
DFT_1024,
DFT_1080,
DFT_1152,
DFT_1200,
DFT_1536,
DFT_2048,
DFT_3072,
DFT_4096,
DFT_6144,
DFT_8192,
DFT_12288,
DFT_18432,
DFT_24576,
DFT_SIZE_IDXTABLESIZE,
} dft_size_idx_t;
#ifdef OAIDFTS
adftfunc_t dft_ftab[]={
dft12,
dft24,
dft36,
dft48,
dft60,
dft72,
dft96,
dft108,
dft120,
dft128,
dft144,
dft180,
dft192,
dft216,
dft240,
dft256,
dft288,
dft300,
dft324,
dft360,
dft384,
dft432,
dft480,
dft512,
dft540,
dft576,
dft600,
dft648,
dft720,
dft768,
dft864,
dft900,
dft960,
dft972,
dft1024,
dft1080,
dft1152,
dft1200,
dft1536,
dft2048,
dft3072,
dft4096,
dft6144,
dft8192,
dft12288,
dft18432,
dft24576,
};
#endif
typedef enum idft_size_idx {
IDFT_128,
IDFT_256,
IDFT_288,
IDFT_300,
IDFT_324,
IDFT_360,
IDFT_384,
IDFT_432,
IDFT_480,
IDFT_512,
IDFT_1024,
IDFT_1536,
IDFT_2048,
IDFT_3072,
IDFT_4096,
IDFT_6144,
IDFT_8192,
IDFT_12288,
IDFT_18432,
IDFT_24576,
IDFT_SIZE_IDXTABLESIZE,
} idft_size_idx_t;
#ifdef OAIDFTS
aidftfunc_t idft_ftab[]={
idft128,
idft256,
idft512,
idft1024,
idft1536,
idft2048,
idft3072,
idft4096,
idft6144,
idft8192,
idft12288,
idft18432,
idft24576,
};
#endif
void dft24576(int16_t *sigF,int16_t *sig,int scale);
/*!\fn int32_t rotate_cpx_vector(int16_t *x,int16_t *alpha,int16_t *y,uint32_t N,uint16_t output_shift) /*!\fn int32_t rotate_cpx_vector(int16_t *x,int16_t *alpha,int16_t *y,uint32_t N,uint16_t output_shift)
@@ -374,57 +579,7 @@ int64_t dot_product64(int16_t *x,
uint32_t N, //must be a multiple of 8 uint32_t N, //must be a multiple of 8
uint8_t output_shift); uint8_t output_shift);
void dft12(int16_t *x,int16_t *y);
void dft24(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft36(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft48(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft60(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft72(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft96(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft108(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft120(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft144(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft180(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft192(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft216(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft240(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft288(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft300(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft324(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft360(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft384(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft432(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft480(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft540(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft576(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft600(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft648(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft720(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft768(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft864(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft900(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft960(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft972(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft1080(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft1152(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft1200(int16_t *x,int16_t *y,uint8_t scale_flag);
void dft64(int16_t *x,int16_t *y,int scale);
void dft128(int16_t *x,int16_t *y,int scale);
void dft256(int16_t *x,int16_t *y,int scale);
void dft512(int16_t *x,int16_t *y,int scale);
void dft1024(int16_t *x,int16_t *y,int scale);
void dft2048(int16_t *x,int16_t *y,int scale);
void dft4096(int16_t *x,int16_t *y,int scale);
void dft8192(int16_t *x,int16_t *y,int scale);
void idft64(int16_t *x,int16_t *y,int scale);
void idft128(int16_t *x,int16_t *y,int scale);
void idft256(int16_t *x,int16_t *y,int scale);
void idft512(int16_t *x,int16_t *y,int scale);
void idft1024(int16_t *x,int16_t *y,int scale);
void idft2048(int16_t *x,int16_t *y,int scale);
void idft4096(int16_t *x,int16_t *y,int scale);
void idft8192(int16_t *x,int16_t *y,int scale);
/** @} */ /** @} */