Fix: audio RTP timestamp calculation (#1213)

Potentially fixes #1185

Remove (set and unused) pacing_time_stamp struct member
Use TAI instead of epoch to calculate RTP time stamp
For code consistency use pkt_time instead of pacing->trs

---------

Co-authored-by: desaulov <[email protected]>

Author: tszumski

lib/src/st2110/st_header.h

@@ -747,8 +747,6 @@ struct st_tx_audio_session_pacing {
   uint64_t cur_epochs;      /* epoch of current pkt */
   /* timestamp for rtp header */
   uint32_t rtp_time_stamp;
-  /* timestamp for pacing */
-  uint32_t pacing_time_stamp;
   uint64_t cur_epoch_time;
   /* in ns, tsc time cursor for packet pacing */
   uint64_t tsc_time_cursor;

lib/src/st2110/st_tx_audio_session.c

@@ -264,7 +264,7 @@ static int tx_audio_session_sync_pacing(struct mtl_main_impl* impl,
                                         struct st_tx_audio_session_impl* s, bool sync,
                                         uint64_t required_tai) {
   struct st_tx_audio_session_pacing* pacing = &s->pacing;
-  double pkt_time = pacing->trs;
+  long double pkt_time = pacing->trs;
   /* always use MTL_PORT_P for ptp now */
   uint64_t ptp_time = mt_get_ptp_time(impl, MTL_PORT_P);
   uint64_t next_epochs = pacing->cur_epochs + 1;
@@ -302,7 +302,18 @@ static int tx_audio_session_sync_pacing(struct mtl_main_impl* impl,
     }
   }
 
-  to_epoch = tx_audio_pacing_time(pacing, epochs) - ptp_time;
+  if (required_tai) {
+    to_epoch = (double)required_tai - ptp_time;
+    if (to_epoch > NS_PER_S) {
+      dbg("%s(%d), required tai %" PRIu64 " ptp_epochs %" PRIu64 " epochs %" PRIu64 "\n",
+          __func__, s->idx, required_tai, ptp_epochs, epochs);
+      s->stat_error_user_timestamp++;
+      to_epoch = NS_PER_S;  // do our best to slow down
+    }
+  } else {
+    to_epoch = tx_audio_pacing_time(pacing, epochs) - ptp_time;
+  }
+
   if (to_epoch < 0) {
     /* time bigger than the assigned epoch time */
     s->stat_epoch_mismatch++;
@@ -313,13 +324,27 @@ static int tx_audio_session_sync_pacing(struct mtl_main_impl* impl,
   if (epochs < next_epochs) s->stat_epoch_onward += (next_epochs - epochs);
 
   pacing->cur_epochs = epochs;
-  pacing->cur_epoch_time = tx_audio_pacing_time(pacing, epochs);
-  pacing->pacing_time_stamp = tx_audio_pacing_time_stamp(pacing, epochs);
-  pacing->rtp_time_stamp = pacing->pacing_time_stamp;
+
+  if (required_tai) {
+    pacing->cur_epoch_time = required_tai + pkt_time;  // prepare next packet
+    /*
+     * Cast [double] to intermediate [uint64_t] to extract 32 least significant bits.
+     * If calculated time stored in [double] is larger than max uint32_t,
+     * then result of direct cast to [uint32_t] results in max uint32_t which is not
+     * what we want. "& 0xffffffff" is used to extract 32 least significant bits
+     * without compiler trying to optimize-out intermediate cast.
+     */
+    pacing->rtp_time_stamp =
+        ((uint64_t)((required_tai / pkt_time) * pacing->pkt_time_sampling) & 0xffffffff);
+  } else {
+    pacing->cur_epoch_time = tx_audio_pacing_time(pacing, epochs);
+    pacing->rtp_time_stamp = tx_audio_pacing_time_stamp(pacing, epochs);
+  }
+
   if (s->ops.rtp_timestamp_delta_us) {
     double rtp_timestamp_delta_us = s->ops.rtp_timestamp_delta_us;
     int32_t rtp_timestamp_delta =
-        (rtp_timestamp_delta_us * NS_PER_US) * pacing->pkt_time_sampling / pacing->trs;
+        (rtp_timestamp_delta_us * NS_PER_US) * pacing->pkt_time_sampling / pkt_time;
     pacing->rtp_time_stamp += rtp_timestamp_delta;
   }
   pacing->tsc_time_cursor = (double)mt_get_tsc(impl) + to_epoch;