Split direct v210 read and copy timing

This commit is contained in:
Johanness
2026-05-19 15:02:22 +03:00
parent 973c547c53
commit b096f32d75
6 changed files with 102 additions and 7 deletions
+35 -1
View File
@@ -47,6 +47,8 @@ struct PerfStats
uint64_t feedReadTicks = 0;
uint64_t uploadDecodeTicks = 0;
uint64_t stageCopyTicks = 0;
uint64_t directV210SourceReadNs = 0;
uint64_t directV210PayloadCopyNs = 0;
uint64_t uploadRecordTicks = 0;
uint64_t submitPresentTicks = 0;
uint64_t idleDelayTicks = 0;
@@ -65,6 +67,8 @@ struct FramePerfStats
uint64_t fenceWaitTicks = 0;
uint64_t feedReadTicks = 0;
uint64_t stageCopyTicks = 0;
uint64_t directV210SourceReadNs = 0;
uint64_t directV210PayloadCopyNs = 0;
uint64_t maxStageCopyTicks = 0;
uint64_t maxStageCopyBytes = 0;
uint32_t maxStageCopyFeed = 0;
@@ -1610,6 +1614,7 @@ int main(int argc, char* argv[])
if (directV210Read)
{
V210ReadTiming directReadTiming{};
const size_t v210Bytes =
static_cast<size_t>(
feeds[i]->v210Stride()
@@ -1622,7 +1627,8 @@ int main(int argc, char* argv[])
feeds[i]->readV210FrameInto(
frameCounter,
v210Feeds[i].v210MappedData,
v210Bytes
v210Bytes,
&directReadTiming
);
const uint64_t version =
feeds[i]->frameVersion();
@@ -1632,6 +1638,18 @@ int main(int argc, char* argv[])
recordFeedVersionStats(i, uploadNeeded);
if (config.logPerf)
{
perfStats.directV210SourceReadNs +=
directReadTiming.sourceReadNs;
perfStats.directV210PayloadCopyNs +=
directReadTiming.payloadCopyNs;
framePerf.directV210SourceReadNs +=
directReadTiming.sourceReadNs;
framePerf.directV210PayloadCopyNs +=
directReadTiming.payloadCopyNs;
}
if (!uploadNeeded)
{
continue;
@@ -2117,6 +2135,14 @@ int main(int argc, char* argv[])
<< (framePerf.feedReadTicks * invMs)
<< "ms stageCopy="
<< (framePerf.stageCopyTicks * invMs)
<< "ms v210Read="
<< (static_cast<double>(
framePerf.directV210SourceReadNs) /
1000000.0)
<< "ms v210Copy="
<< (static_cast<double>(
framePerf.directV210PayloadCopyNs) /
1000000.0)
<< "ms stageCopies="
<< framePerf.stageCopies
<< " v210Copies="
@@ -2168,6 +2194,14 @@ int main(int argc, char* argv[])
<< (perfStats.uploadDecodeTicks * invMs / frames)
<< "ms stageCopy="
<< (perfStats.stageCopyTicks * invMs / frames)
<< "ms v210Read="
<< (static_cast<double>(
perfStats.directV210SourceReadNs) /
1000000.0 / frames)
<< "ms v210Copy="
<< (static_cast<double>(
perfStats.directV210PayloadCopyNs) /
1000000.0 / frames)
<< "ms uploadRecord="
<< (perfStats.uploadRecordTicks * invMs / frames)
<< "ms submitPresent="