From 80fb7e2294768e47af1a84ccd47d93dfbed839bb Mon Sep 17 00:00:00 2001
From: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Date: Wed, 5 Feb 2025 19:43:08 +0100
Subject: [PATCH] [d3d9,d3d11] Adjust flush heuristic on tiling GPUs

Probably needs work, just disable most of the heuristic for the time being.
---
 src/d3d11/d3d11_context_imm.cpp | 14 +++++++++++++-
 src/d3d11/d3d11_context_imm.h   |  4 ++++
 src/d3d9/d3d9_device.cpp        | 12 +++++++++++-
 src/d3d9/d3d9_device.h          |  2 ++
 src/util/util_flush.cpp         |  7 +++----
 src/util/util_flush.h           |  5 ++---
 6 files changed, 35 insertions(+), 9 deletions(-)

diff --git a/src/d3d11/d3d11_context_imm.cpp b/src/d3d11/d3d11_context_imm.cpp
index 31426c850..90497b1e3 100644
--- a/src/d3d11/d3d11_context_imm.cpp
+++ b/src/d3d11/d3d11_context_imm.cpp
@@ -18,7 +18,7 @@ namespace dxvk {
   : D3D11CommonContext<D3D11ImmediateContext>(pParent, Device, 0, DxvkCsChunkFlag::SingleUse),
     m_csThread(Device, Device->createContext()),
     m_submissionFence(new sync::CallbackFence()),
-    m_flushTracker(pParent->GetOptions()->reproducibleCommandStream),
+    m_flushTracker(GetMaxFlushType(pParent, Device)),
     m_stagingBufferFence(new sync::Fence(0)),
     m_multithread(this, false, pParent->GetOptions()->enableContextLock),
     m_videoContext(this, Device) {
@@ -1074,4 +1074,16 @@ namespace dxvk {
     return stats;
   }
 
+
+  GpuFlushType D3D11ImmediateContext::GetMaxFlushType(
+          D3D11Device*    pParent,
+    const Rc<DxvkDevice>& Device) {
+    if (pParent->GetOptions()->reproducibleCommandStream)
+      return GpuFlushType::ExplicitFlush;
+    else if (Device->perfHints().preferRenderPassOps)
+      return GpuFlushType::ImplicitStrongHint;
+    else
+      return GpuFlushType::ImplicitWeakHint;
+  }
+
 }
diff --git a/src/d3d11/d3d11_context_imm.h b/src/d3d11/d3d11_context_imm.h
index 4bf35f83d..ca4fc5670 100644
--- a/src/d3d11/d3d11_context_imm.h
+++ b/src/d3d11/d3d11_context_imm.h
@@ -208,6 +208,10 @@ namespace dxvk {
 
     DxvkStagingBufferStats GetStagingMemoryStatistics();
 
+    static GpuFlushType GetMaxFlushType(
+            D3D11Device*    pParent,
+      const Rc<DxvkDevice>& Device);
+
   };
   
 }
diff --git a/src/d3d9/d3d9_device.cpp b/src/d3d9/d3d9_device.cpp
index bea0e9201..99d5bd3bf 100644
--- a/src/d3d9/d3d9_device.cpp
+++ b/src/d3d9/d3d9_device.cpp
@@ -58,7 +58,7 @@ namespace dxvk {
     , m_csThread           ( dxvkDevice, dxvkDevice->createContext() )
     , m_csChunk            ( AllocCsChunk() )
     , m_submissionFence    ( new sync::Fence() )
-    , m_flushTracker       ( m_d3d9Options.reproducibleCommandStream )
+    , m_flushTracker       ( GetMaxFlushType() )
     , m_d3d9Interop        ( this )
     , m_d3d9On12           ( this )
     , m_d3d8Bridge         ( this ) {
@@ -8731,4 +8731,14 @@ namespace dxvk {
     m_flags.clr(D3D9DeviceFlag::DirtySpecializationEntries);
   }
 
+
+  GpuFlushType D3D9DeviceEx::GetMaxFlushType() const {
+    if (m_d3d9Options.reproducibleCommandStream)
+      return GpuFlushType::ExplicitFlush;
+    else if (m_dxvkDevice->perfHints().preferRenderPassOps)
+      return GpuFlushType::ImplicitStrongHint;
+    else
+      return GpuFlushType::ImplicitWeakHint;
+  }
+
 }
diff --git a/src/d3d9/d3d9_device.h b/src/d3d9/d3d9_device.h
index ed86bb85e..cec35f3d4 100644
--- a/src/d3d9/d3d9_device.h
+++ b/src/d3d9/d3d9_device.h
@@ -1397,6 +1397,8 @@ namespace dxvk {
         && !m_state.renderTargets[Index]->IsNull();
     }
 
+    GpuFlushType GetMaxFlushType() const;
+
     Com<D3D9InterfaceEx>            m_parent;
     D3DDEVTYPE                      m_deviceType;
     HWND                            m_window;
diff --git a/src/util/util_flush.cpp b/src/util/util_flush.cpp
index c2b1ce6a6..62315b826 100644
--- a/src/util/util_flush.cpp
+++ b/src/util/util_flush.cpp
@@ -2,9 +2,8 @@
 
 namespace dxvk {
 
-  GpuFlushTracker::GpuFlushTracker(
-          bool ensureReproducibleHeuristic)
-  : m_ensureReproducibleHeuristic(ensureReproducibleHeuristic) {
+  GpuFlushTracker::GpuFlushTracker(GpuFlushType maxType)
+  : m_maxType(maxType) {
 
   }
 
@@ -23,7 +22,7 @@ namespace dxvk {
     if (!chunkCount)
       return false;
 
-    if (m_ensureReproducibleHeuristic && flushType != GpuFlushType::ExplicitFlush)
+    if (flushType > m_maxType)
       return false;
 
     // Take any earlier missed flush with a stronger hint into account, so
diff --git a/src/util/util_flush.h b/src/util/util_flush.h
index c3e71ebcb..d5c4957cf 100644
--- a/src/util/util_flush.h
+++ b/src/util/util_flush.h
@@ -34,7 +34,7 @@ namespace dxvk {
 
   public:
 
-    GpuFlushTracker(bool ensureReproducibleHeuristic);
+    GpuFlushTracker(GpuFlushType maxAllowed);
 
     /**
      * \brief Checks whether a context flush should be performed
@@ -63,8 +63,7 @@ namespace dxvk {
 
   private:
 
-    bool          m_ensureReproducibleHeuristic;
-
+    GpuFlushType  m_maxType               = GpuFlushType::ImplicitWeakHint;
     GpuFlushType  m_lastMissedType        = GpuFlushType::ImplicitWeakHint;
 
     uint64_t      m_lastFlushChunkId      = 0ull;