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

Remove spinlock from threadpool task return

Browse Source 
The idea to use atomic aligned to cache line for counting finished jobs
is reused, but instead of polling a semaphore was added
This commit is contained in:
Bartosz Podrygajlo
2025-01-14 15:17:19 +01:00
parent f7d3b72851
commit 0657ab2275
13 changed files with 137 additions and 100 deletions
Download Patch File Download Diff File Expand all files Collapse all files
openair1/PHY/CODING/nrLDPC_coding/nrLDPC_coding_segment/nrLDPC_coding_segment_decoder.c
+7 -10
View File
@@ -250,7 +250,7 @@ int nrLDPC_prepare_TB_decoding(nrLDPC_slot_decoding_parameters_t *nrLDPC_slot_de
for (int r = 0; r < nrLDPC_TB_decoding_parameters->C; r++) {
nrLDPC_decoding_parameters_t *rdata = &((nrLDPC_decoding_parameters_t *)t_info->buf)[t_info->len];
DevAssert(t_info->len < t_info->cap);
rdata->ans = &t_info->ans[t_info->len];
rdata->ans = t_info->ans;
t_info->len += 1;
decParams.R = nrLDPC_TB_decoding_parameters->segments[r].R;
@@ -309,19 +309,16 @@ int32_t nrLDPC_coding_decoder(nrLDPC_slot_decoding_parameters_t *nrLDPC_slot_dec
nbSegments += nrLDPC_TB_decoding_parameters->C;
}
nrLDPC_decoding_parameters_t arr[nbSegments];
task_ans_t ans[nbSegments];
memset(ans, 0, nbSegments * sizeof(task_ans_t));
thread_info_tm_t t_info = {.buf = (uint8_t *)arr, .len = 0, .cap = nbSegments, .ans = ans};
task_ans_t ans;
init_task_ans(&ans, nbSegments);
thread_info_tm_t t_info = {.buf = (uint8_t *)arr, .len = 0, .cap = nbSegments, .ans = &ans};
int nbDecode = 0;
for (int pusch_id = 0; pusch_id < nrLDPC_slot_decoding_parameters->nb_TBs; pusch_id++) {
nbDecode += nrLDPC_prepare_TB_decoding(nrLDPC_slot_decoding_parameters, pusch_id, &t_info);
(void)nrLDPC_prepare_TB_decoding(nrLDPC_slot_decoding_parameters, pusch_id, &t_info);
}
DevAssert(nbDecode == t_info.len);
// Execute thread poool tasks
join_task_ans(t_info.ans, t_info.len);
// Execute thread pool tasks
join_task_ans(t_info.ans);
for (int pusch_id = 0; pusch_id < nrLDPC_slot_decoding_parameters->nb_TBs; pusch_id++) {
nrLDPC_TB_decoding_parameters_t *nrLDPC_TB_decoding_parameters = &nrLDPC_slot_decoding_parameters->TBs[pusch_id];
openair1/PHY/CODING/nrLDPC_coding/nrLDPC_coding_segment/nrLDPC_coding_segment_encoder.c
+9 -9
View File
@@ -189,7 +189,7 @@ static int nrLDPC_prepare_TB_encoding(nrLDPC_slot_encoding_parameters_t *nrLDPC_
for (int j = 0; j < n_seg; j++) {
ldpc8blocks_args_t *perJobImpp = &((ldpc8blocks_args_t *)t_info->buf)[t_info->len];
DevAssert(t_info->len < t_info->cap);
impp.ans = &t_info->ans[t_info->len];
impp.ans = t_info->ans;
t_info->len += 1;
impp.macro_num = j;
@@ -211,19 +211,19 @@ int nrLDPC_coding_encoder(nrLDPC_slot_encoding_parameters_t *nrLDPC_slot_encodin
nbTasks += n_seg;
}
ldpc8blocks_args_t arr[nbTasks];
task_ans_t ans[nbTasks];
memset(ans, 0, nbTasks * sizeof(task_ans_t));
thread_info_tm_t t_info = {.buf = (uint8_t *)arr, .len = 0, .cap = nbTasks, .ans = ans};
task_ans_t ans;
init_task_ans(&ans, nbTasks);
thread_info_tm_t t_info = {.buf = (uint8_t *)arr, .len = 0, .cap = nbTasks, .ans = &ans};
int nbEncode = 0;
for (int dlsch_id = 0; dlsch_id < nrLDPC_slot_encoding_parameters->nb_TBs; dlsch_id++) {
nbEncode += nrLDPC_prepare_TB_encoding(nrLDPC_slot_encoding_parameters, dlsch_id, &t_info);
}
DevAssert(nbEncode == t_info.len);
// Execute thread poool tasks
join_task_ans(ans, nbEncode);
if (nbEncode < nbTasks) {
completed_many_task_ans(&ans, nbTasks - nbEncode);
}
// Execute thread pool tasks
join_task_ans(&ans);
return 0;
}