[d3d11] Add d3d11.maxImplicitDiscardSize option

dxvk.conf

@@ -203,6 +203,14 @@
 # d3d11.zeroWorkgroupMemory = False
 
 
+# Resource size limit for implicit discards, in kilobytes. For small staging
+# resources mapped with MAP_WRITE, DXVK will sometimes allocate new backing
+# storage in order to avoid GPU synchronization, so setting this too high
+# may cause memory issues, setting it to 0 disables the feature.
+
+# d3d11.maxImplicitDiscardSize = 256
+
+
 # Sets number of pipeline compiler threads.
 # 
 # Supported values:

src/d3d11/d3d11_context_imm.cpp

@@ -7,8 +7,6 @@ constexpr static uint32_t MinFlushIntervalUs = 750;
 constexpr static uint32_t IncFlushIntervalUs = 250;
 constexpr static uint32_t MaxPendingSubmits  = 6;
 
-constexpr static VkDeviceSize MaxImplicitDiscardSize = 256ull << 10;
-
 namespace dxvk {
   
   D3D11ImmediateContext::D3D11ImmediateContext(
@@ -16,6 +14,7 @@ namespace dxvk {
     const Rc<DxvkDevice>& Device)
   : D3D11DeviceContext(pParent, Device, DxvkCsChunkFlag::SingleUse),
     m_csThread(Device, Device->createContext()),
+    m_maxImplicitDiscardSize(pParent->GetOptions()->maxImplicitDiscardSize),
     m_videoContext(this, Device) {
     EmitCs([
       cDevice                 = m_device,
@@ -393,7 +392,7 @@ namespace dxvk {
       auto buffer = pResource->GetBuffer();
       auto sequenceNumber = pResource->GetSequenceNumber();
 
-      if (MapType != D3D11_MAP_READ && !MapFlags && bufferSize <= MaxImplicitDiscardSize) {
+      if (MapType != D3D11_MAP_READ && !MapFlags && bufferSize <= m_maxImplicitDiscardSize) {
         SynchronizeCsThread(sequenceNumber);
 
         bool hasWoAccess = buffer->isInUse(DxvkAccess::Write);
@@ -514,11 +513,9 @@ namespace dxvk {
 
         // Don't implicitly discard large buffers or buffers of images with
         // multiple subresources, as that is likely to cause memory issues.
-        VkDeviceSize bufferSize = pResource->CountSubresources() == 1
-          ? pResource->GetMappedSlice(Subresource).length
-          : MaxImplicitDiscardSize;
+        VkDeviceSize bufferSize = pResource->GetMappedSlice(Subresource).length;
 
-        if (bufferSize >= MaxImplicitDiscardSize) {
+        if (bufferSize >= m_maxImplicitDiscardSize || pResource->CountSubresources() > 1) {
           // Don't check access flags, WaitForResource will return
           // early anyway if the resource is currently in use
           doFlags = DoWait;

src/d3d11/d3d11_context_imm.h

@@ -125,6 +125,7 @@ namespace dxvk {
     uint64_t                m_eventCount = 0ull;
     uint32_t                m_mappedImageCount = 0u;
 
+    VkDeviceSize            m_maxImplicitDiscardSize = 0ull;
 
     dxvk::high_resolution_clock::time_point m_lastFlush
       = dxvk::high_resolution_clock::now();

src/d3d11/d3d11_options.cpp

@@ -25,6 +25,11 @@ namespace dxvk {
     this->syncInterval          = config.getOption<int32_t>("dxgi.syncInterval", -1);
     this->tearFree              = config.getOption<Tristate>("dxgi.tearFree", Tristate::Auto);
 
+    int32_t maxImplicitDiscardSize = config.getOption<int32_t>("d3d11.maxImplicitDiscardSize", 256);
+    this->maxImplicitDiscardSize = maxImplicitDiscardSize >= 0
+      ? VkDeviceSize(maxImplicitDiscardSize) << 10
+      : VkDeviceSize(~0ull);
+
     this->constantBufferRangeCheck = config.getOption<bool>("d3d11.constantBufferRangeCheck", false)
       && DxvkGpuVendor(devInfo.core.properties.vendorID) != DxvkGpuVendor::Amd;
 

src/d3d11/d3d11_options.h

@@ -92,6 +92,9 @@ namespace dxvk {
     /// Limit frame rate
     int32_t maxFrameRate;
 
+    /// Limit discardable resource size
+    VkDeviceSize maxImplicitDiscardSize;
+
     /// Defer surface creation until first present call. This
     /// fixes issues with games that create multiple swap chains
     /// for a single window that may interfere with each other.