PclFix.txt

PCL Event Handling Performance Fix Summary
==========================================

Problem:
--------
When using a gamepad, high-frequency PCLStatsInput/ReflexStatsInput events (100-1000Hz)
caused system-wide frame time issues. The bottleneck was TdhGetEventInformation calls
for every TraceLogging event.

Root Cause:
-----------
1. TraceLogging events require TdhGetEventInformation to decode event names and properties
2. This is an expensive Windows API call
3. Gamepad input generates thousands of StatsInput events per second
4. Original caching attempt using (EventId, Opcode, Level) failed because TraceLogging
   events share identical EVENT_DESCRIPTOR values - they distinguish events by name
   embedded in metadata, not by descriptor fields

Solution:
---------
Implemented multi-level optimization to eliminate TDH calls and minimize processing
for high-frequency events.

Files Modified:
---------------
1. PresentData/PresentMonTraceConsumer.hpp
2. PresentData/PresentMonTraceConsumer.cpp

Changes in Detail:
------------------

1. EVENT TYPE CACHING BY UserDataLength
   - Different PCL event types have different payload sizes
   - Cache the UserDataLength for StatsInput and StatsEvent after first TDH decode
   - Use direct integer comparison instead of hash map lookup

   New members in PMTraceConsumer:
   - USHORT mPclStatsInputLength = 0;   // UserDataLength for StatsInput events
   - USHORT mPclStatsEventLength = 0;   // UserDataLength for StatsEvent events

2. DIRECT PAYLOAD READING FOR StatsEvent
   - TraceLogging payload layout: Marker (uint32, offset 0) + FrameID (uint64, offset 4)
   - Read values directly from pEventRecord->UserData without TDH decode
   - Extracted HandlePclStatsEvent() helper function to process markers

3. CACHED POINTER FOR StatsInput TIMESTAMP UPDATES
   - Most gamepad input comes from single process
   - Cache pointer to hash map entry to avoid repeated lookups

   New members in PMTraceConsumer:
   - uint32_t mLastPingProcessId = 0;
   - uint64_t* mLastPingTimestampPtr = nullptr;

   - If same ProcessId: direct pointer write (no hash map lookup)
   - If different ProcessId: hash map lookup + cache update
   - Cache invalidated on process shutdown/exit

4. REMOVED TRY-CATCH FROM FAST PATH
   - Exception handling only for slow path (first occurrence, rare events)

Performance Characteristics:
----------------------------

StatsInput Fast Path (gamepad input - most frequent):
- 2 integer comparisons (event type check)
- 2 integer comparisons (ProcessId + nullptr check)
- 1 direct memory write
- NO hash map lookup
- NO TDH decode
- NO exception handling overhead

StatsEvent Fast Path (frame markers):
- 2 integer comparisons (event type check)
- 2 memory reads from UserData (Marker, FrameID)
- HandlePclStatsEvent() call with hash map operations
- NO TDH decode

Slow Path (first occurrence, Init, Shutdown):
- Full TDH decode via TraceLoggingDecoder
- Event type cached for future fast path
- Only happens once per event type per session

Cache Invalidation:
-------------------
The cached timestamp pointer is invalidated when:
- PCLStatsShutdown event received
- Process exit detected (two locations in HandleProcessEvent)

This prevents use-after-free if hash map entry is erased.

Testing Notes:
--------------
- Setting mTrackPcLatency = false completely disables PCL provider at ETW level
- If performance issues persist with optimizations, the overhead is in ETW event
  delivery infrastructure, not in HandlePclEvent processing
- Consider throttling/sampling if ETW overhead is still problematic

Header Include Added:
---------------------
#include "ETW/Nvidia_PCL.h" added to PresentMonTraceConsumer.hpp for Nvidia_PCL::PCLMarker

Function Declaration Added:
---------------------------
void HandlePclStatsEvent(uint32_t processId, uint64_t timestamp,
                         Nvidia_PCL::PCLMarker marker, uint32_t frameId);