audio: use monotonic counter-based PTS for AAC RTSP streams

Replace gettimeofday() with a wallclock-anchored monotonic counter for
AAC audio fPresentationTime in IMPDeviceSource. On the first AAC frame
we snapshot gettimeofday() as the base, then advance by exactly
frameCount * 1024 / sampleRate microseconds per frame.

This produces strictly monotonic +1024-sample RTP PTS increments
(verified via ffprobe) while staying in the gettimeofday() clock domain
required by live555 RTCP SR (RFC 3550 §6.4.1) for correct NTP-RTP
timestamp pairing and inter-stream synchronization.

The previous gettimeofday()-per-frame approach introduced OS scheduling
jitter causing non-monotonic DTS that broke FFmpeg/Frigate MP4 muxing.

Also fix AudioReframer timestamp increment from microseconds (*1000000)
to milliseconds (*1000) to match the actual unit used by callers.

Fixes: themactep/thingino-firmware#1082

Author: themactep

src/AudioReframer.cpp

@@ -42,7 +42,7 @@ void AudioReframer::getReframedFrame(uint8_t *frameData, int64_t &timestamp) {
 
   timestamp = currentTimestamp;
   if (inputSampleRate > 0) {
-    currentTimestamp += (static_cast<int64_t>(outputSamplesPerFrame) * 1000000LL) / inputSampleRate;
+    currentTimestamp += (outputSamplesPerFrame * 1000) / inputSampleRate;
   } // else: keep currentTimestamp stable if misconfigured
 }
 

src/IMPDeviceSource.cpp

@@ -1,19 +1,7 @@
 #include "IMPDeviceSource.hpp"
 #include "GroupsockHelper.hh"
 #include <iostream>
-
-// Use live555's own clock (gettimeofday on the live555 thread) for the audio
-// presentation time so that the RTCP SR NTP and RTP timestamps are computed
-// from the same clock, keeping pts values near-zero and non-negative.
-// Hardware timestamps are still used by AudioWorker for the AudioReframer
-// (uniform 1024-sample frame boundaries) and for MP4 recording.
-// Video falls back to gettimeofday because H264NALUnit.time is not yet populated.
-static void assignPresentationTime(const AudioFrame &, struct timeval &tv) {
-  gettimeofday(&tv, nullptr);
-}
-static void assignPresentationTime(const H264NALUnit &, struct timeval &tv) {
-  gettimeofday(&tv, nullptr);
-}
+#include <type_traits>
 
 // explicit instantiation
 template class IMPDeviceSource<H264NALUnit, video_stream>;
@@ -76,7 +64,32 @@ template <typename FrameType, typename Stream> void IMPDeviceSource<FrameType, S
       fFrameSize = nal.data.size();
     }
 
-    assignPresentationTime(nal, fPresentationTime);
+    /* Use monotonic counter-based timestamps for AAC to produce perfectly
+       steady +1024 RTP PTS increments and avoid gettimeofday jitter. */
+    if constexpr (std::is_same_v<FrameType, AudioFrame>) {
+      auto *imp_audio = global_audio[encChn]->imp_audio;
+      if (imp_audio && imp_audio->format == IMPAudioFormat::AAC) {
+        if (audioFirstFrame) {
+          gettimeofday(&audioStartTime, NULL);
+          audioFrameCount = 0;
+          audioFirstFrame = false;
+        }
+        int sampleRate = (imp_audio->sample_rate > 0) ? imp_audio->sample_rate : 16000;
+        constexpr uint64_t kSamplesPerFrame = 1024;
+        uint64_t usec_offset = (audioFrameCount * kSamplesPerFrame * 1000000ULL) / static_cast<uint64_t>(sampleRate);
+        fPresentationTime.tv_sec  = audioStartTime.tv_sec  + static_cast<time_t>(usec_offset / 1000000ULL);
+        fPresentationTime.tv_usec = audioStartTime.tv_usec + static_cast<suseconds_t>(usec_offset % 1000000ULL);
+        if (fPresentationTime.tv_usec >= 1000000) {
+          fPresentationTime.tv_sec++;
+          fPresentationTime.tv_usec -= 1000000;
+        }
+        audioFrameCount++;
+      } else {
+        gettimeofday(&fPresentationTime, NULL);
+      }
+    } else {
+      gettimeofday(&fPresentationTime, NULL);
+    }
 
     memcpy(fTo, &nal.data[0], fFrameSize);