aaudio: account for hardware jitter in clock model
This will allow AAudio to track the DSP better when
the DSP has variable timing, such as when resampling.
Keep track of the maximum lateness.
Do not assume it is only one burst.
Account for the lateness on the input streams
so we don't try to read the data before the DSP
has written it.
Bug: 123643363
Test: record through iRig UA using Oboe Tester
Test: see bug for details
Change-Id: I49eb852c6d0324e8a26ee912da5108021ec113fe
diff --git a/media/libaaudio/src/client/IsochronousClockModel.cpp b/media/libaaudio/src/client/IsochronousClockModel.cpp
index d26b352..9abdf53 100644
--- a/media/libaaudio/src/client/IsochronousClockModel.cpp
+++ b/media/libaaudio/src/client/IsochronousClockModel.cpp
@@ -19,12 +19,11 @@
#include <log/log.h>
#include <stdint.h>
+#include <algorithm>
#include "utility/AudioClock.h"
#include "IsochronousClockModel.h"
-#define MIN_LATENESS_NANOS (10 * AAUDIO_NANOS_PER_MICROSECOND)
-
using namespace aaudio;
IsochronousClockModel::IsochronousClockModel()
@@ -32,7 +31,7 @@
, mMarkerNanoTime(0)
, mSampleRate(48000)
, mFramesPerBurst(64)
- , mMaxLatenessInNanos(0)
+ , mMaxMeasuredLatenessNanos(0)
, mState(STATE_STOPPED)
{
}
@@ -41,8 +40,7 @@
}
void IsochronousClockModel::setPositionAndTime(int64_t framePosition, int64_t nanoTime) {
- ALOGV("setPositionAndTime(%lld, %lld)",
- (long long) framePosition, (long long) nanoTime);
+ ALOGV("setPositionAndTime, %lld, %lld", (long long) framePosition, (long long) nanoTime);
mMarkerFramePosition = framePosition;
mMarkerNanoTime = nanoTime;
}
@@ -54,7 +52,9 @@
}
void IsochronousClockModel::stop(int64_t nanoTime) {
- ALOGV("stop(nanos = %lld)\n", (long long) nanoTime);
+ ALOGD("stop(nanos = %lld) max lateness = %d micros\n",
+ (long long) nanoTime,
+ (int) (mMaxMeasuredLatenessNanos / 1000));
setPositionAndTime(convertTimeToPosition(nanoTime), nanoTime);
// TODO should we set position?
mState = STATE_STOPPED;
@@ -69,9 +69,10 @@
}
void IsochronousClockModel::processTimestamp(int64_t framePosition, int64_t nanoTime) {
-// ALOGD("processTimestamp() - framePosition = %lld at nanoTime %llu",
-// (long long)framePosition,
-// (long long)nanoTime);
+ mTimestampCount++;
+// Log position and time in CSV format so we can import it easily into spreadsheets.
+ //ALOGD("%s() CSV, %d, %lld, %lld", __func__,
+ //mTimestampCount, (long long)framePosition, (long long)nanoTime);
int64_t framesDelta = framePosition - mMarkerFramePosition;
int64_t nanosDelta = nanoTime - mMarkerNanoTime;
if (nanosDelta < 1000) {
@@ -110,22 +111,54 @@
// Earlier than expected timestamp.
// This data is probably more accurate, so use it.
// Or we may be drifting due to a fast HW clock.
-// int microsDelta = (int) (nanosDelta / 1000);
-// int expectedMicrosDelta = (int) (expectedNanosDelta / 1000);
-// ALOGD("processTimestamp() - STATE_RUNNING - %7d < %7d so %4d micros EARLY",
-// microsDelta, expectedMicrosDelta, (expectedMicrosDelta - microsDelta));
+ //int microsDelta = (int) (nanosDelta / 1000);
+ //int expectedMicrosDelta = (int) (expectedNanosDelta / 1000);
+ //ALOGD("%s() - STATE_RUNNING - #%d, %4d micros EARLY",
+ //__func__, mTimestampCount, expectedMicrosDelta - microsDelta);
setPositionAndTime(framePosition, nanoTime);
- } else if (nanosDelta > (expectedNanosDelta + mMaxLatenessInNanos)) {
- // Later than expected timestamp.
-// int microsDelta = (int) (nanosDelta / 1000);
-// int expectedMicrosDeadline = (int) ((expectedNanosDelta + mMaxLatenessInNanos) / 1000);
-// ALOGD("processTimestamp() - STATE_RUNNING - %7d > %7d so %4d micros LATE",
-// microsDelta, expectedMicrosDeadline, (microsDelta - expectedMicrosDeadline));
+ } else if (nanosDelta > (expectedNanosDelta + (2 * mBurstPeriodNanos))) {
+ // In this case we do not update mMaxMeasuredLatenessNanos because it
+ // would force it too high.
+ // mMaxMeasuredLatenessNanos should range from 1 to 2 * mBurstPeriodNanos
+ //int32_t measuredLatenessNanos = (int32_t)(nanosDelta - expectedNanosDelta);
+ //ALOGD("%s() - STATE_RUNNING - #%d, lateness %d - max %d = %4d micros VERY LATE",
+ //__func__,
+ //mTimestampCount,
+ //measuredLatenessNanos / 1000,
+ //mMaxMeasuredLatenessNanos / 1000,
+ //(measuredLatenessNanos - mMaxMeasuredLatenessNanos) / 1000
+ //);
- // When we are late it may be because of preemption in the kernel or
- // we may be drifting due to a slow HW clock.
- setPositionAndTime(framePosition, nanoTime - mMaxLatenessInNanos);
+ // This typically happens when we are modelling a service instead of a DSP.
+ setPositionAndTime(framePosition, nanoTime - (2 * mBurstPeriodNanos));
+ } else if (nanosDelta > (expectedNanosDelta + mMaxMeasuredLatenessNanos)) {
+ //int32_t previousLatenessNanos = mMaxMeasuredLatenessNanos;
+ mMaxMeasuredLatenessNanos = (int32_t)(nanosDelta - expectedNanosDelta);
+
+ //ALOGD("%s() - STATE_RUNNING - #%d, newmax %d - oldmax %d = %4d micros LATE",
+ //__func__,
+ //mTimestampCount,
+ //mMaxMeasuredLatenessNanos / 1000,
+ //previousLatenessNanos / 1000,
+ //(mMaxMeasuredLatenessNanos - previousLatenessNanos) / 1000
+ //);
+
+ // When we are late, it may be because of preemption in the kernel,
+ // or timing jitter caused by resampling in the DSP,
+ // or we may be drifting due to a slow HW clock.
+ // We add slight drift value just in case there is actual long term drift
+ // forward caused by a slower clock.
+ // If the clock is faster than the model will get pushed earlier
+ // by the code in the preceding branch.
+ // The two opposing forces should allow the model to track the real clock
+ // over a long time.
+ int64_t driftingTime = mMarkerNanoTime + expectedNanosDelta + kDriftNanos;
+ setPositionAndTime(framePosition, driftingTime);
+ //ALOGD("%s() - #%d, max lateness = %d micros",
+ //__func__,
+ //mTimestampCount,
+ //(int) (mMaxMeasuredLatenessNanos / 1000));
}
break;
default:
@@ -145,9 +178,12 @@
update();
}
+// Update expected lateness based on sampleRate and framesPerBurst
void IsochronousClockModel::update() {
- int64_t nanosLate = convertDeltaPositionToTime(mFramesPerBurst); // uses mSampleRate
- mMaxLatenessInNanos = (nanosLate > MIN_LATENESS_NANOS) ? nanosLate : MIN_LATENESS_NANOS;
+ mBurstPeriodNanos = convertDeltaPositionToTime(mFramesPerBurst); // uses mSampleRate
+ // Timestamps may be late by up to a burst because we are randomly sampling the time period
+ // after the DSP position is actually updated.
+ mMaxMeasuredLatenessNanos = mBurstPeriodNanos;
}
int64_t IsochronousClockModel::convertDeltaPositionToTime(int64_t framesDelta) const {
@@ -190,11 +226,25 @@
return position;
}
+int32_t IsochronousClockModel::getLateTimeOffsetNanos() const {
+ // This will never be < 0 because mMaxLatenessNanos starts at
+ // mBurstPeriodNanos and only gets bigger.
+ return (mMaxMeasuredLatenessNanos - mBurstPeriodNanos) + kExtraLatenessNanos;
+}
+
+int64_t IsochronousClockModel::convertPositionToLatestTime(int64_t framePosition) const {
+ return convertPositionToTime(framePosition) + getLateTimeOffsetNanos();
+}
+
+int64_t IsochronousClockModel::convertLatestTimeToPosition(int64_t nanoTime) const {
+ return convertTimeToPosition(nanoTime - getLateTimeOffsetNanos());
+}
+
void IsochronousClockModel::dump() const {
ALOGD("mMarkerFramePosition = %lld", (long long) mMarkerFramePosition);
ALOGD("mMarkerNanoTime = %lld", (long long) mMarkerNanoTime);
ALOGD("mSampleRate = %6d", mSampleRate);
ALOGD("mFramesPerBurst = %6d", mFramesPerBurst);
- ALOGD("mMaxLatenessInNanos = %6d", mMaxLatenessInNanos);
+ ALOGD("mMaxMeasuredLatenessNanos = %6d", mMaxMeasuredLatenessNanos);
ALOGD("mState = %6d", mState);
}