From fd68a08572d2be961294c323e6487720eb191ab2 Mon Sep 17 00:00:00 2001
From: netborg <137700136+netborg-afps@users.noreply.github.com>
Date: Fri, 28 Feb 2025 08:48:36 +0100
Subject: [PATCH] [dxvk] Improve handling of cpu-frametime related stutters for
 low-latency frame pacing

Stutters less this way because we increase the sensitivity to mark frames as outliers, so that they don't get used for predicting the next frame. The actual "optimal" threshold is still to be fine-tuned, but this one worked really well.
---
 .../dxvk_framepacer_mode_low_latency.h        | 23 ++++++++++---------
 1 file changed, 12 insertions(+), 11 deletions(-)

diff --git a/src/dxvk/framepacer/dxvk_framepacer_mode_low_latency.h b/src/dxvk/framepacer/dxvk_framepacer_mode_low_latency.h
index 9193f438e..24d8ab60e 100644
--- a/src/dxvk/framepacer/dxvk_framepacer_mode_low_latency.h
+++ b/src/dxvk/framepacer/dxvk_framepacer_mode_low_latency.h
@@ -116,9 +116,7 @@ namespace dxvk {
       // where gpuSubmit[i] <= gpuRun[i] for all i
 
       std::vector<int32_t>& gpuRun = m_tempGpuRun;
-      std::vector<int32_t>& gpuRunDurations = m_tempGpuRunDurations;
       gpuRun.clear();
-      gpuRunDurations.clear();
       int32_t optimizedGpuTime = 0;
       gpuRun.push_back(optimizedGpuTime);
 
@@ -127,7 +125,6 @@ namespace dxvk {
         int32_t duration = duration_cast<microseconds>( m->gpuReady[i+1] - _gpuRun ).count();
         optimizedGpuTime += duration;
         gpuRun.push_back(optimizedGpuTime);
-        gpuRunDurations.push_back(duration);
       }
 
       int32_t alignment = duration_cast<microseconds>( m->gpuSubmit[numLoop-1] - m->gpuSubmit[0] ).count()
@@ -177,7 +174,7 @@ namespace dxvk {
 
     struct SyncProps {
       int32_t optimizedGpuTime;   // gpu executing packed submits in one go
-      int32_t gpuSync;            // us after gpuStart
+      int32_t gpuSync;            // gpuStart to this sync point, in microseconds
       int32_t cpuUntilGpuSync;
       int32_t cpuUntilGpuStart;
       int32_t csStart;
@@ -187,9 +184,14 @@ namespace dxvk {
 
 
     SyncProps getSyncPrediction() {
-      // in the future we might use more samples to get a prediction
-      // however, simple averaging gives a slightly artificial mouse input
-      // more advanced methods will be investigated
+      // In the future we might use more samples to get a prediction.
+      // Possibly this will be optional, as until now, basing it on
+      // just the previous frame gave us the best mouse input feel.
+      // Simple averaging or median filtering is surely not the way
+      // to go, but more advanced methods will be investigated.
+      // The best place to filter should be on the Present() timeline,
+      // so not sure if we really will do any filtering here other
+      // than outlier removal, which will dampen stuttering effects.
       SyncProps res = {};
       uint64_t id = m_propsFinished;
       if (id < DXGI_MAX_SWAP_CHAIN_BUFFERS+7)
@@ -210,14 +212,14 @@ namespace dxvk {
     bool isOutlier( uint64_t frameId ) {
       constexpr size_t numLoop = 7;
       int32_t totalCpuTime = 0;
-      for (size_t i=0; i<numLoop; ++i) {
+      for (size_t i=1; i<numLoop; ++i) {
         const LatencyMarkers* m = m_latencyMarkersStorage->getConstMarkers(frameId-i);
         totalCpuTime += m->cpuFinished;
       }
 
-      int32_t avgCpuTime = totalCpuTime / numLoop;
+      int32_t avgCpuTime = totalCpuTime / (numLoop-1);
       const LatencyMarkers* m = m_latencyMarkersStorage->getConstMarkers(frameId);
-      if (m->cpuFinished > 1.7*avgCpuTime || m->gpuSubmit.empty() || m->gpuReady.size() != (m->gpuSubmit.size()+1) )
+      if (m->cpuFinished > 1.3*avgCpuTime || m->gpuSubmit.empty() || m->gpuReady.size() != (m->gpuSubmit.size()+1) )
         return true;
 
       return false;
@@ -232,7 +234,6 @@ namespace dxvk {
     std::atomic<uint64_t> m_propsFinished = { 0 };
 
     std::vector<int32_t>  m_tempGpuRun;
-    std::vector<int32_t>  m_tempGpuRunDurations;
 
   };