Commit 34cf585a authored by Robert Schmidt's avatar Robert Schmidt

Merge remote-tracking branch 'origin/fix-4layers-nr_dlsim' into integration_2024_w45 (!3082)

fix blocking bugs in command nr_dlsim -n100 -e27 -s35 -b 273 -R273 -x4 -y4 -z4

errors in main OAI code:

- one stack overflow
- one variable array of size 0
- one direct usage of unaligned SIMD

and small fixes inside the tests tool nr_dlsim to handle this large configuration (4layers, 100MHz)

use a better complex vector multiplication function
parents f6184c9f 8f688c00
......@@ -93,6 +93,11 @@ void dft_lte(int32_t *z,struct complex16 *input, int32_t Msc_PUSCH, uint8_t Nsym
// printf("\n");
dft_size_idx_t dftsize = get_dft(Msc_PUSCH);
if (dftsize == DFT_SIZE_IDXTABLESIZE) {
LOG_E(PHY, "Internal error, not modulating the slot, Msc_PUSCH = %d\n", Msc_PUSCH);
return;
}
switch (Msc_PUSCH) {
case 12:
dft(dftsize, (int16_t *)dft_in0, (int16_t *)dft_out0, 0);
......
......@@ -589,20 +589,21 @@ int nr_rx_pdsch(PHY_VARS_NR_UE *ue,
symbol,
nb_re_pdsch);
if (nl >= 2) // Apply MMSE for 2, 3, and 4 Tx layers
nr_dlsch_mmse(rx_size_symbol,
n_rx,
nl,
rxdataF_comp,
dl_ch_mag,
dl_ch_magb,
dl_ch_magr,
dl_ch_estimates_ext,
nb_rb_pdsch,
dlsch_config->qamModOrder,
*log2_maxh,
symbol,
nb_re_pdsch,
nvar);
if (nb_re_pdsch)
nr_dlsch_mmse(rx_size_symbol,
n_rx,
nl,
rxdataF_comp,
dl_ch_mag,
dl_ch_magb,
dl_ch_magr,
dl_ch_estimates_ext,
nb_rb_pdsch,
dlsch_config->qamModOrder,
*log2_maxh,
symbol,
nb_re_pdsch,
nvar);
}
stop_meas_nr_ue_phy(ue, DLSCH_MRC_MMSE_STATS);
......@@ -1333,50 +1334,6 @@ void nr_a_sum_b(c16_t *input_x, c16_t *input_y, unsigned short nb_rb)
}
}
/* Zero Forcing Rx function: nr_a_mult_b()
* Compute the complex Multiplication c=a*b
*
* */
void nr_a_mult_b(c16_t *a, c16_t *b, c16_t *c, unsigned short nb_rb, unsigned char output_shift0)
{
//This function is used to compute complex multiplications
short nr_conjugate[8]__attribute__((aligned(16))) = {1,-1,1,-1,1,-1,1,-1};
unsigned short rb;
simde__m128i *a_128,*b_128, *c_128, mmtmpD0,mmtmpD1,mmtmpD2,mmtmpD3;
a_128 = (simde__m128i *)a;
b_128 = (simde__m128i *)b;
c_128 = (simde__m128i *)c;
for (rb=0; rb<3*nb_rb; rb++) {
// the real part
mmtmpD0 = simde_mm_sign_epi16(a_128[0],*(simde__m128i*)&nr_conjugate[0]);
mmtmpD0 = simde_mm_madd_epi16(mmtmpD0,b_128[0]); //Re: (a_re*b_re - a_im*b_im)
// the imag part
mmtmpD1 = simde_mm_shufflelo_epi16(a_128[0],SIMDE_MM_SHUFFLE(2,3,0,1));
mmtmpD1 = simde_mm_shufflehi_epi16(mmtmpD1,SIMDE_MM_SHUFFLE(2,3,0,1));
mmtmpD1 = simde_mm_madd_epi16(mmtmpD1,b_128[0]);//Im: (x_im*y_re + x_re*y_im)
mmtmpD0 = simde_mm_srai_epi32(mmtmpD0,output_shift0);
mmtmpD1 = simde_mm_srai_epi32(mmtmpD1,output_shift0);
mmtmpD2 = simde_mm_unpacklo_epi32(mmtmpD0,mmtmpD1);
mmtmpD3 = simde_mm_unpackhi_epi32(mmtmpD0,mmtmpD1);
c_128[0] = simde_mm_packs_epi32(mmtmpD2,mmtmpD3);
/*printf("\n Computing mult \n");
print_shorts("a:",(int16_t*)&a_128[0]);
print_shorts("b:",(int16_t*)&b_128[0]);
print_shorts("pack:",(int16_t*)&c_128[0]);*/
a_128+=1;
b_128+=1;
c_128+=1;
}
}
/* Zero Forcing Rx function: nr_element_sign()
* Compute b=sign*a
*
......@@ -1448,7 +1405,7 @@ static void nr_determin(int size,
nb_rb,
((rtx & 1) == 1 ? -1 : 1) * ((ctx & 1) == 1 ? -1 : 1) * sign,
shift0);
nr_a_mult_b(a44[ctx][rtx], outtemp, rtx == 0 ? ad_bc : outtemp1, nb_rb, shift0);
mult_complex_vectors(a44[ctx][rtx], outtemp, rtx == 0 ? ad_bc : outtemp1, sizeofArray(outtemp1), shift0);
if (rtx != 0)
nr_a_sum_b(ad_bc, outtemp1, nb_rb);
......@@ -1759,11 +1716,11 @@ static void nr_dlsch_mmse(uint32_t rx_size_symbol,
// printf("Computing r_%d c_%d\n",rtx,ctx);
// print_shorts(" H_h_H=",(int16_t*)&conjH_H_elements[ctx*nl+rtx][0][0]);
// print_shorts(" Inv_H_h_H=",(int16_t*)&inv_H_h_H[ctx*nl+rtx][0]);
nr_a_mult_b(inv_H_h_H[ctx][rtx],
(c16_t *)(rxdataF_comp[ctx][0] + symbol * rx_size_symbol),
outtemp,
nb_rb_0,
shift - (fp_flag == 1 ? 2 : 0));
mult_complex_vectors(inv_H_h_H[ctx][rtx],
(c16_t *)(rxdataF_comp[ctx][0] + symbol * rx_size_symbol),
outtemp,
sizeofArray(outtemp),
shift - (fp_flag == 1 ? 2 : 0));
nr_a_sum_b(rxdataF_zforcing[rtx], outtemp, nb_rb_0); // a = a + b
}
#ifdef DEBUG_DLSCH_DEMOD
......
......@@ -16,8 +16,15 @@ add_dependencies(tests test_log2_approx)
add_test(NAME test_log2_approx
COMMAND ./test_log2_approx --gtest_filter=-log2_approx.complete)
add_executable(dft_test test_dft.c ../dfts_load.c)
target_link_libraries(dft_test minimal_lib shlib_loader SIMU m)
add_dependencies(tests dft_test)
add_dependencies(dft_test dfts) # trigger build of dfts (shared lib for DFT)
add_test(NAME dft_test COMMAND ./dft_test)
add_executable(test_vector_op test_vector_op.cpp)
target_link_libraries(test_vector_op PRIVATE LOG minimal_lib)
add_dependencies(tests test_vector_op)
add_test(NAME test_vector_op
COMMAND test_vector_op)
......@@ -45,7 +45,7 @@ static void BM_rotate_cpx_vector(benchmark::State &state)
{
int vector_size = state.range(0);
auto input_complex_16 = generate_random_c16(vector_size);
auto input_alpha = generate_random_c16(vector_size);
auto input_alpha = generate_random_c16(1);
AlignedVector512<c16_t> output;
output.resize(vector_size);
int shift = 2;
......
......@@ -19,3 +19,52 @@
* contact@openairinterface.org
*/
#include <stdint.h>
#include <vector>
#include <algorithm>
#include <numeric>
extern "C" {
#include "openair1/PHY/TOOLS/tools_defs.h"
struct configmodule_interface_s;
struct configmodule_interface_s *uniqCfg = NULL;
void exit_function(const char *file, const char *function, const int line, const char *s, const int assert)
{
if (assert) {
abort();
} else {
exit(EXIT_SUCCESS);
}
}
}
#include <cstdio>
#include "common/utils/LOG/log.h"
#include "openair1/PHY/TOOLS/phy_test_tools.hpp"
int main()
{
const int shift = 15; // it should always be 15 to keep int16 in same range
for (int vector_size = 1237; vector_size < 1237 + 8; vector_size++) {
auto input1 = generate_random_c16(vector_size);
auto input2 = generate_random_c16(vector_size);
AlignedVector512<c16_t> output;
output.resize(vector_size);
mult_complex_vectors(input1.data(), input2.data(), output.data(), vector_size, shift);
for (int i = 0; i < vector_size; i++) {
c16_t res = c16mulShift(input1[i], input2[i], shift);
if (output[i].r != res.r || output[i].i != res.i) {
printf("Error at %d: (%d,%d) * (%d,%d) = (%d,%d) (should be (%d,%d))\n",
i,
input1[i].r,
input1[i].i,
input2[i].r,
input2[i].i,
output[i].r,
output[i].i,
res.r,
res.i);
return 1;
}
}
}
return 0;
}
......@@ -38,6 +38,7 @@
#include <simde/simde-common.h>
#include <simde/x86/sse.h>
#include <simde/x86/avx2.h>
#include "common/utils/LOG/log.h"
#define simd_q15_t simde__m128i
#define simdshort_q15_t simde__m64
......@@ -90,7 +91,7 @@ extern "C" {
int dim2;
int dim3;
int dim4;
uint8_t data[];
uint8_t data[] __attribute__((aligned(32)));
} fourDimArray_t;
static inline fourDimArray_t *allocateFourDimArray(int elmtSz, int dim1, int dim2, int dim3, int dim4)
......@@ -357,6 +358,74 @@ static __attribute__((always_inline)) inline void multadd_real_four_symbols_vect
simde_mm_storeu_si128((simd_q15_t *)y, y_128);
}
// Multiply two vectors of complex int16 and take the most significant bits (shift by 15 in normal case)
// works only with little endian storage (for big endian, modify the srai/ssli at the end)
static __attribute__((always_inline)) inline void mult_complex_vectors(const c16_t *in1,
const c16_t *in2,
c16_t *out,
const int size,
const int shift)
{
const simde__m256i complex_shuffle256 = simde_mm256_set_epi8(29,
28,
31,
30,
25,
24,
27,
26,
21,
20,
23,
22,
17,
16,
19,
18,
13,
12,
15,
14,
9,
8,
11,
10,
5,
4,
7,
6,
1,
0,
3,
2);
const simde__m256i conj256 = simde_mm256_set_epi16(-1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1);
int i;
// do 8 multiplications at a time
for (i = 0; i < size - 7; i += 8) {
const simde__m256i i1 = simde_mm256_loadu_epi32((simde__m256i *)(in1 + i));
const simde__m256i i2 = simde_mm256_loadu_epi32((simde__m256i *)(in2 + i));
const simde__m256i i2swap = simde_mm256_shuffle_epi8(i2, complex_shuffle256);
const simde__m256i i2conj = simde_mm256_sign_epi16(i2, conj256);
const simde__m256i re = simde_mm256_madd_epi16(i1, i2conj);
const simde__m256i im = simde_mm256_madd_epi16(i1, i2swap);
simde_mm256_storeu_si256(
(simde__m256i *)(out + i),
simde_mm256_blend_epi16(simde_mm256_srai_epi32(re, shift), simde_mm256_slli_epi32(im, 16 - shift), 0xAA));
}
if (size - i > 4) {
const simde__m128i i1 = simde_mm_loadu_epi32((simde__m128i *)(in1 + i));
const simde__m128i i2 = simde_mm_loadu_epi32((simde__m128i *)(in2 + i));
const simde__m128i i2swap = simde_mm_shuffle_epi8(i2, *(simde__m128i *)&complex_shuffle256);
const simde__m128i i2conj = simde_mm_sign_epi16(i2, *(simde__m128i *)&conj256);
const simde__m128i re = simde_mm_madd_epi16(i1, i2conj);
const simde__m128i im = simde_mm_madd_epi16(i1, i2swap);
simde_mm_storeu_si128((simde__m128i *)(out + i),
simde_mm_blend_epi16(simde_mm_srai_epi32(re, shift), simde_mm_slli_epi32(im, 16 - shift), 0xAA));
i += 4;
}
for (; i < size; i++)
out[i] = c16mulShift(in1[i], in2[i], shift);
}
/*!\fn void multadd_complex_vector_real_scalar(int16_t *x,int16_t alpha,int16_t *y,uint8_t zero_flag,uint32_t N)
This function performs componentwise multiplication and accumulation of a real scalar and a complex vector.
@param x Vector input (Q1.15) in the format |Re0 Im0|Re1 Im 1| ...
......@@ -550,17 +619,15 @@ typedef enum dft_size_idx {
case Sz: \
return DFT_##Sz; \
break;
static inline
dft_size_idx_t get_dft(int ofdm_symbol_size)
static inline dft_size_idx_t get_dft(int size)
{
switch (ofdm_symbol_size) {
switch (size) {
FOREACH_DFTSZ(FIND_ENUM)
default:
printf("function get_dft : unsupported ofdm symbol size \n");
assert(0);
LOG_E(UTIL, "function get_dft : unsupported DFT size %d\n", size);
break;
}
return DFT_SIZE_IDXTABLESIZE; // never reached and will trigger assertion in idft function;
return DFT_SIZE_IDXTABLESIZE;
}
#define SZ_iENUM(Sz) IDFT_##Sz,
......@@ -610,14 +677,12 @@ struct {
return IDFT_##iSz; \
break;
static inline
idft_size_idx_t get_idft(int ofdm_symbol_size)
static inline idft_size_idx_t get_idft(int size)
{
switch (ofdm_symbol_size) {
switch (size) {
FOREACH_IDFTSZ(FIND_iENUM)
default:
printf("function get_idft : unsupported ofdm symbol size \n");
assert(0);
AssertFatal(false, "function get_idft : unsupported iDFT size %d\n", size);
break;
}
return IDFT_SIZE_IDXTABLESIZE; // never reached and will trigger assertion in idft function
......
......@@ -141,14 +141,13 @@ void nr_64qam_llr(int32_t *rxdataF_comp, int32_t *ch_mag, int32_t *ch_mag2, int1
int32_t *llr_32 = (int32_t *)llr;
#ifndef USE_128BIT
simde__m256i xmm0, xmm1, xmm2;
for (int i = 0; i < (nb_re >> 3); i++) {
xmm0 = *rxF;
simde__m256i xmm0 = simde_mm256_loadu_si256(rxF);
// registers of even index in xmm0-> |y_R|, registers of odd index in xmm0-> |y_I|
xmm1 = simde_mm256_abs_epi16(xmm0);
simde__m256i xmm1 = simde_mm256_abs_epi16(xmm0);
// registers of even index in xmm0-> |y_R|-|h|^2, registers of odd index in xmm0-> |y_I|-|h|^2
xmm1 = simde_mm256_subs_epi16(*ch_maga, xmm1);
xmm2 = simde_mm256_abs_epi16(xmm1);
simde__m256i xmm2 = simde_mm256_abs_epi16(xmm1);
xmm2 = simde_mm256_subs_epi16(*ch_magb, xmm2);
// xmm0 |1st 4th 7th 10th 13th 16th 19th 22ed|
// xmm1 |2ed 5th 8th 11th 14th 17th 20th 23rd|
......
......@@ -505,8 +505,8 @@ static int nr_ue_pdsch_procedures(PHY_VARS_NR_UE *ue,
dlsch0->Nl);
const uint32_t pdsch_est_size = ((ue->frame_parms.symbols_per_slot * ue->frame_parms.ofdm_symbol_size + 15) / 16) * 16;
__attribute__((aligned(32))) int32_t pdsch_dl_ch_estimates[ue->frame_parms.nb_antennas_rx * dlsch0->Nl][pdsch_est_size];
memset(pdsch_dl_ch_estimates, 0, sizeof(int32_t) * ue->frame_parms.nb_antennas_rx * dlsch0->Nl * pdsch_est_size);
fourDimArray_t *toFree = NULL;
allocCast2D(pdsch_dl_ch_estimates, int32_t, toFree, ue->frame_parms.nb_antennas_rx * dlsch0->Nl, pdsch_est_size, false);
c16_t ptrs_phase_per_slot[ue->frame_parms.nb_antennas_rx][NR_SYMBOLS_PER_SLOT];
memset(ptrs_phase_per_slot, 0, sizeof(ptrs_phase_per_slot));
......@@ -515,8 +515,14 @@ static int nr_ue_pdsch_procedures(PHY_VARS_NR_UE *ue,
memset(ptrs_re_per_slot, 0, sizeof(ptrs_re_per_slot));
const uint32_t rx_size_symbol = (dlsch[0].dlsch_config.number_rbs * NR_NB_SC_PER_RB + 15) & ~15;
__attribute__((aligned(32))) int32_t rxdataF_comp[dlsch[0].Nl][ue->frame_parms.nb_antennas_rx][rx_size_symbol * NR_SYMBOLS_PER_SLOT];
memset(rxdataF_comp, 0, sizeof(rxdataF_comp));
fourDimArray_t *toFree2 = NULL;
allocCast3D(rxdataF_comp,
int32_t,
toFree2,
dlsch[0].Nl,
ue->frame_parms.nb_antennas_rx,
rx_size_symbol * NR_SYMBOLS_PER_SLOT,
false);
uint32_t nvar = 0;
......@@ -620,6 +626,8 @@ static int nr_ue_pdsch_procedures(PHY_VARS_NR_UE *ue,
return -1;
} // CRNTI active
stop_meas_nr_ue_phy(ue, RX_PDSCH_STATS);
free(toFree);
free(toFree2);
}
return 0;
}
......
......@@ -858,15 +858,12 @@ int main(int argc, char **argv)
UE_mac->if_module = nr_ue_if_module_init(0);
unsigned int available_bits=0;
unsigned char *estimated_output_bit;
unsigned char *test_input_bit;
unsigned char *estimated_output_bit=NULL;
unsigned char *test_input_bit=NULL;
unsigned int errors_bit = 0;
initFloatingCoresTpool(dlsch_threads, &nrUE_params.Tpool, false, "UE-tpool");
test_input_bit = (unsigned char *) malloc16(sizeof(unsigned char) * 16 * 68 * 384);
estimated_output_bit = (unsigned char *) malloc16(sizeof(unsigned char) * 16 * 68 * 384);
// generate signal
AssertFatal(input_fd==NULL,"Not ready for input signal file\n");
......@@ -905,13 +902,14 @@ int main(int argc, char **argv)
gNB->msgDataTx = msgDataTx;
// Buffers to store internal memory of slot process
int rx_size = (((14 * frame_parms->N_RB_DL * 12 * sizeof(int32_t)) + 15) >> 4) << 4;
UE->phy_sim_rxdataF = calloc(sizeof(int32_t *) * frame_parms->nb_antennas_rx * g_nrOfLayers, frame_parms->samples_per_slot_wCP * sizeof(int32_t));
int rx_size = (((14 * UE->frame_parms.N_RB_DL * 12 * sizeof(int32_t)) + 15) >> 4) << 4;
UE->phy_sim_rxdataF = calloc(sizeof(int32_t *) * UE->frame_parms.nb_antennas_rx * g_nrOfLayers,
UE->frame_parms.samples_per_slot_wCP * sizeof(int32_t));
UE->phy_sim_pdsch_llr = calloc(1, (8 * (3 * 8 * 8448)) * sizeof(int16_t)); // Max length
UE->phy_sim_pdsch_rxdataF_ext = calloc(sizeof(int32_t *) * frame_parms->nb_antennas_rx * g_nrOfLayers, rx_size);
UE->phy_sim_pdsch_rxdataF_comp = calloc(sizeof(int32_t *) * frame_parms->nb_antennas_rx * g_nrOfLayers, rx_size);
UE->phy_sim_pdsch_dl_ch_estimates = calloc(sizeof(int32_t *) * frame_parms->nb_antennas_rx * g_nrOfLayers, rx_size);
UE->phy_sim_pdsch_dl_ch_estimates_ext = calloc(sizeof(int32_t *) * frame_parms->nb_antennas_rx * g_nrOfLayers, rx_size);
UE->phy_sim_pdsch_rxdataF_ext = calloc(sizeof(int32_t *) * UE->frame_parms.nb_antennas_rx * g_nrOfLayers, rx_size);
UE->phy_sim_pdsch_rxdataF_comp = calloc(sizeof(int32_t *) * UE->frame_parms.nb_antennas_rx * g_nrOfLayers, rx_size);
UE->phy_sim_pdsch_dl_ch_estimates = calloc(sizeof(int32_t *) * UE->frame_parms.nb_antennas_rx * g_nrOfLayers, rx_size);
UE->phy_sim_pdsch_dl_ch_estimates_ext = calloc(sizeof(int32_t *) * UE->frame_parms.nb_antennas_rx * g_nrOfLayers, rx_size);
int a_segments = MAX_NUM_NR_DLSCH_SEGMENTS_PER_LAYER*NR_MAX_NB_LAYERS; //number of segments to be allocated
if (g_rbSize != 273) {
a_segments = a_segments*g_rbSize;
......@@ -1118,7 +1116,17 @@ int main(int argc, char **argv)
// Apply MIMO Channel
multipath_channel(gNB2UE, s_re, s_im, r_re, r_im, slot_length, 0, (n_trials == 1) ? 1 : 0);
add_noise(UE->common_vars.rxdata, (const double **) r_re, (const double **) r_im, sigma2, slot_length, slot_offset, ts, delay, pdu_bit_map, 0x1, frame_parms->nb_antennas_rx);
add_noise(UE->common_vars.rxdata,
(const double **)r_re,
(const double **)r_im,
sigma2,
slot_length,
slot_offset,
ts,
delay,
pdu_bit_map,
0x1,
UE->frame_parms.nb_antennas_rx);
dl_config.sfn = frame;
dl_config.slot = slot;
ue_dci_configuration(UE_mac, &dl_config, frame, slot);
......@@ -1168,6 +1176,10 @@ int main(int argc, char **argv)
round++;
} // round
if (test_input_bit == NULL) {
test_input_bit = (unsigned char *)malloc16(sizeof(unsigned char) * TBS);
estimated_output_bit = (unsigned char *)malloc16(sizeof(unsigned char) * TBS);
}
for (i = 0; i < TBS; i++) {
estimated_output_bit[i] = (UE->phy_sim_dlsch_b[i/8] & (1 << (i & 7))) >> (i & 7);
......@@ -1226,10 +1238,11 @@ int main(int argc, char **argv)
fprintf(csv_file,"%.2f,%.4f,%.2f,%u\n", roundStats, effRate, effRate / TBS * 100, TBS);
}
if (print_perf==1) {
printf("\ngNB TX function statistics (per %d us slot, NPRB %d, mcs %d, block %d)\n",
printf("\ngNB TX function statistics (per %d us slot, NPRB %d, mcs %d, C %d, block %d)\n",
1000 >> *scc->ssbSubcarrierSpacing,
g_rbSize,
g_mcsIndex,
UE->dl_harq_processes[0][slot].C,
msgDataTx->dlsch[0][0].harq_process.pdsch_pdu.pdsch_pdu_rel15.TBSize[0] << 3);
printDistribution(&gNB->phy_proc_tx,table_tx,"PHY proc tx");
printStatIndent2(&gNB->dlsch_encoding_stats,"DLSCH encoding time");
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment