[dxvk] Optimize flush heuristic for low-latency

Possibly can be optimized more, but just changing these numbers already had a huge effect, especially for games having a small number of submissions to begin with.

src/dxvk/framepacer/dxvk_framepacer.cpp

@@ -2,6 +2,7 @@
 #include "dxvk_framepacer_mode_low_latency.h"
 #include "dxvk_framepacer_mode_min_latency.h"
 #include "dxvk_options.h"
+#include "../../util/util_flush.h"
 #include "../../util/util_env.h"
 #include "../../util/log/log.h"
 
@@ -40,11 +41,15 @@ namespace dxvk {
 
       case FramePacerMode::LOW_LATENCY:
         Logger::info( "Frame pace: low-latency" );
+        GpuFlushTracker::m_minPendingSubmissions = 1;
+        GpuFlushTracker::m_minChunkCount = 1;
         m_mode = std::make_unique<LowLatencyMode>(mode, &m_latencyMarkersStorage, options);
         break;
 
       case FramePacerMode::MIN_LATENCY:
         Logger::info( "Frame pace: min-latency" );
+        GpuFlushTracker::m_minPendingSubmissions = 1;
+        GpuFlushTracker::m_minChunkCount = 1;
         m_mode = std::make_unique<MinLatencyMode>(mode, &m_latencyMarkersStorage);
         break;
     }

src/dxvk/framepacer/dxvk_framepacer.h

@@ -136,7 +136,7 @@ namespace dxvk {
     // not implemented methods
 
 
-    void notifyCpuPresentBegin( uint64_t frameId) override { }
+    void notifyCpuPresentBegin( uint64_t frameId ) override { }
     void notifyCpuPresentEnd( uint64_t frameId ) override { }
     void notifyQueuePresentEnd( uint64_t frameId, VkResult status) override { }
     void notifyGpuExecutionBegin( uint64_t frameId ) override { }

src/util/util_flush.cpp

@@ -2,6 +2,10 @@
 
 namespace dxvk {
 
+  std::atomic<uint32_t> GpuFlushTracker::m_minPendingSubmissions = { 2 };
+  std::atomic<uint32_t> GpuFlushTracker::m_minChunkCount         = { 3 };
+  std::atomic<uint32_t> GpuFlushTracker::m_maxChunkCount         = { 20 };
+
   GpuFlushTracker::GpuFlushTracker(GpuFlushType maxType)
   : m_maxType(maxType) {
 
@@ -11,10 +15,6 @@ namespace dxvk {
           GpuFlushType          flushType,
           uint64_t              chunkId,
           uint32_t              lastCompleteSubmissionId) {
-    constexpr uint32_t minPendingSubmissions = 2;
-
-    constexpr uint32_t minChunkCount =  3u;
-    constexpr uint32_t maxChunkCount = 20u;
 
     // Do not flush if there is nothing to flush
     uint32_t chunkCount = uint32_t(chunkId - m_lastFlushChunkId);
@@ -42,14 +42,14 @@ namespace dxvk {
 
       case GpuFlushType::ImplicitStrongHint: {
         // Flush aggressively with a strong hint to reduce readback latency.
-        return chunkCount >= minChunkCount;
+        return chunkCount >= m_minChunkCount;
       }
 
       case GpuFlushType::ImplicitMediumHint:
       case GpuFlushType::ImplicitWeakHint: {
         // Aim for a higher number of chunks per submission with
         // a weak hint in order to avoid submitting too often.
-        if (chunkCount < 2 * minChunkCount)
+        if (chunkCount < 2 * m_minChunkCount)
           return false;
 
         // Actual heuristic is shared with synchronization commands
@@ -60,13 +60,13 @@ namespace dxvk {
         // required if the application is spinning on a query or resource.
         uint32_t pendingSubmissions = uint32_t(m_lastFlushSubmissionId - lastCompleteSubmissionId);
 
-        if (pendingSubmissions < minPendingSubmissions)
+        if (pendingSubmissions < m_minPendingSubmissions)
           return true;
 
         // Use the number of pending submissions to decide whether to flush. Other
         // than ignoring the minimum chunk count condition, we should treat this
         // the same as weak hints to avoid unnecessary synchronization.
-        uint32_t threshold = std::min(maxChunkCount, pendingSubmissions * minChunkCount);
+        uint32_t threshold = std::min(m_maxChunkCount.load(), pendingSubmissions * m_minChunkCount.load());
         return chunkCount >= threshold;
       }
     }

src/util/util_flush.h

@@ -3,6 +3,7 @@
 #include <cstddef>
 #include <cstdint>
 #include <vector>
+#include <atomic>
 
 namespace dxvk {
 
@@ -64,6 +65,10 @@ namespace dxvk {
             uint64_t              chunkId,
             uint64_t              submissionId);
 
+    static std::atomic<uint32_t> m_minPendingSubmissions;
+    static std::atomic<uint32_t> m_minChunkCount;
+    static std::atomic<uint32_t> m_maxChunkCount;
+
   private:
 
     GpuFlushType  m_maxType               = GpuFlushType::ImplicitWeakHint;