[dxvk,d3d11] Improve explicit UAV overlap behaviour

If the app explicitly enables UAV overlap, don't synchronize
back-to-back read-modify-write operations to the same UAV either.

src/d3d11/d3d11_context_ext.cpp

@@ -146,8 +146,10 @@ namespace dxvk {
     D3D11Device* parent = static_cast<D3D11Device*>(m_ctx->GetParentInterface());
     DxvkBarrierControlFlags flags = parent->GetOptionsBarrierControlFlags();
 
-    if (ControlFlags & D3D11_VK_BARRIER_CONTROL_IGNORE_WRITE_AFTER_WRITE)
-      flags.set(DxvkBarrierControl::IgnoreComputeWriteAfterWrite, DxvkBarrierControl::IgnoreGraphicsWriteAfterWrite);
+    if (ControlFlags & D3D11_VK_BARRIER_CONTROL_IGNORE_WRITE_AFTER_WRITE) {
+      flags.set(DxvkBarrierControl::ComputeAllowReadWriteOverlap,
+                DxvkBarrierControl::GraphicsAllowReadWriteOverlap);
+    }
 
     m_ctx->EmitCs([cFlags = flags] (DxvkContext* ctx) {
       ctx->setBarrierControl(cFlags);

src/d3d11/d3d11_device.h

@@ -472,13 +472,13 @@ namespace dxvk {
       const Rc<DxvkAdapter>&  Adapter);
 
     DxvkBarrierControlFlags GetOptionsBarrierControlFlags() {
-      DxvkBarrierControlFlags barrierControl;
+      DxvkBarrierControlFlags barrierControl = 0u;
 
       if (m_d3d11Options.relaxedBarriers)
-        barrierControl.set(DxvkBarrierControl::IgnoreComputeWriteAfterWrite);
+        barrierControl.set(DxvkBarrierControl::ComputeAllowWriteOnlyOverlap);
 
       if (m_d3d11Options.relaxedBarriers || m_d3d11Options.relaxedGraphicsBarriers)
-        barrierControl.set(DxvkBarrierControl::IgnoreGraphicsWriteAfterWrite);
+        barrierControl.set(DxvkBarrierControl::GraphicsAllowReadWriteOverlap);
 
       return barrierControl;
     }

src/dxvk/dxvk_context.cpp

@@ -2623,6 +2623,20 @@ namespace dxvk {
 
 
   void DxvkContext::setBarrierControl(DxvkBarrierControlFlags control) {
+    // If any currently relevant control flags change, play it safe and force
+    // a barrier the next time we encounter a write-after-write hazard, even
+    // if the same set of flags is restored by that time. Only check graphics
+    // flags inside a render pass to avoid performance regressions when an
+    // application uses this feature but we already have an app profile.
+    // Barriers get flushed when beginning or ending a render pass anyway.
+    DxvkBarrierControlFlags mask = m_flags.test(DxvkContextFlag::GpRenderPassBound)
+      ? DxvkBarrierControlFlags(DxvkBarrierControl::GraphicsAllowReadWriteOverlap)
+      : DxvkBarrierControlFlags(DxvkBarrierControl::ComputeAllowReadWriteOverlap,
+                                DxvkBarrierControl::ComputeAllowWriteOnlyOverlap);
+
+    if (!((m_barrierControl ^ control) & mask).isClear())
+      m_flags.set(DxvkContextFlag::ForceWriteAfterWriteSync);
+
     m_barrierControl = control;
   }
   
@@ -3740,12 +3754,14 @@ namespace dxvk {
       vk::makeSubresourceRange(imageSubresource), imageLayout,
       VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT, VK_ACCESS_2_SHADER_READ_BIT);
 
+    m_flags.set(DxvkContextFlag::ForceWriteAfterWriteSync);
+
     if (unlikely(m_features.test(DxvkContextFeature::DebugUtils)))
       m_cmd->cmdEndDebugUtilsLabel(DxvkCmdBuffer::ExecBuffer);
 
     m_cmd->track(buffer, DxvkAccess::Write);
     m_cmd->track(image, DxvkAccess::Read);
-  }
+}
 
 
   void DxvkContext::clearImageViewFb(
@@ -3951,6 +3967,9 @@ namespace dxvk {
       VK_IMAGE_LAYOUT_GENERAL, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
       VK_ACCESS_2_SHADER_WRITE_BIT);
 
+    if (cmdBuffer == DxvkCmdBuffer::ExecBuffer)
+      m_flags.set(DxvkContextFlag::ForceWriteAfterWriteSync);
+
     if (unlikely(m_features.test(DxvkContextFeature::DebugUtils)))
       m_cmd->cmdEndDebugUtilsLabel(cmdBuffer);
 
@@ -5528,6 +5547,10 @@ namespace dxvk {
         ctrOffsets[i] = physSlice.offset;
 
         if (physSlice.handle) {
+          // Just in case someone is mad enough to write to a
+          // transform feedback buffer from a shader as well
+          m_flags.set(DxvkContextFlag::ForceWriteAfterWriteSync);
+
           accessBuffer(DxvkCmdBuffer::ExecBuffer, m_state.xfb.activeCounters[i],
             VK_PIPELINE_STAGE_2_TRANSFORM_FEEDBACK_BIT_EXT,
             VK_ACCESS_2_TRANSFORM_FEEDBACK_COUNTER_READ_BIT_EXT |
@@ -7923,6 +7946,8 @@ namespace dxvk {
   void DxvkContext::flushBarriers() {
     m_execBarriers.flush(m_cmd);
     m_barrierTracker.clear();
+
+    m_flags.clr(DxvkContextFlag::ForceWriteAfterWriteSync);
   }
 
 

src/dxvk/dxvk_context.h

@@ -1785,23 +1785,33 @@ namespace dxvk {
             VkAccessFlags             access);
 
     template<VkPipelineBindPoint BindPoint>
-    bool canIgnoreWawHazards() {
-      constexpr auto controlFlag = BindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS
-        ? DxvkBarrierControl::IgnoreGraphicsWriteAfterWrite
-        : DxvkBarrierControl::IgnoreComputeWriteAfterWrite;
+    DxvkAccessFlags getAllowedStorageHazards() {
+      if (m_barrierControl.isClear() || m_flags.test(DxvkContextFlag::ForceWriteAfterWriteSync))
+        return DxvkAccessFlags();
 
-      if (!m_barrierControl.test(controlFlag))
-        return false;
-
-      if (BindPoint == VK_PIPELINE_BIND_POINT_COMPUTE) {
+      if constexpr (BindPoint == VK_PIPELINE_BIND_POINT_COMPUTE) {
+        // If there are any pending accesses that are not directly related
+        // to shader dispatches, always insert a barrier if there is a hazard.
         VkPipelineStageFlags2 stageMask = VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT
                                         | VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT;
-        return !m_execBarriers.hasPendingStages(~stageMask);
+
+        if (!m_execBarriers.hasPendingStages(~stageMask)) {
+          if (m_barrierControl.test(DxvkBarrierControl::ComputeAllowReadWriteOverlap))
+            return DxvkAccessFlags(DxvkAccess::Write, DxvkAccess::Read);
+          else if (m_barrierControl.test(DxvkBarrierControl::ComputeAllowWriteOnlyOverlap))
+            return DxvkAccessFlags(DxvkAccess::Write);
+        }
+      } else {
+        // For graphics, the only type of unrelated access we have to worry about
+        // is transform feedback writes, in which case inserting a barrier is fine.
+        if (m_barrierControl.test(DxvkBarrierControl::GraphicsAllowReadWriteOverlap))
+          return DxvkAccessFlags(DxvkAccess::Write, DxvkAccess::Read);
       }
 
-      return true;
+      return DxvkAccessFlags();
     }
 
+
     void emitMemoryBarrier(
             VkPipelineStageFlags      srcStages,
             VkAccessFlags             srcAccess,
@@ -2039,18 +2049,17 @@ namespace dxvk {
       if (hasPendingWrite) {
         // If there is a write-after-write hazard and synchronization
         // for those is not explicitly disabled, insert a barrier.
-        if (!canIgnoreWawHazards<BindPoint>())
+        DxvkAccessFlags allowedHazards = getAllowedStorageHazards<BindPoint>();
+
+        if (!allowedHazards.test(DxvkAccess::Write))
           return true;
 
-        // If write-after-write checking is disabled and we're on graphics,
-        // be aggressive about avoiding barriers and ignore any reads if we
-        // do find a write-after-write hazard. This essentially assumes that
-        // back-to-back read-modify-write operations are safe, but will still
-        // consider read-only or transform feedback operations as unsafe.
-        if (BindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS)
-          return !(access & VK_ACCESS_SHADER_WRITE_BIT);
+        // Skip barrier if overlapping read-modify-write ops are allowed.
+        // This includes shader atomics, but also non-atomic load-stores.
+        if (allowedHazards.test(DxvkAccess::Read))
+          return false;
 
-        // On compute, if we are reading the resource, add a barrier.
+        // Otherwise, check if there is a read-after-write hazard.
         if (access & vk::AccessReadMask)
           return true;
       }
@@ -2059,6 +2068,8 @@ namespace dxvk {
       return pred(DxvkAccess::Read);
     }
 
+    void invalidateWriteAfterWriteTracking();
+
     void beginRenderPassDebugRegion();
 
     void beginInternalDebugRegion(

src/dxvk/dxvk_context_state.h

@@ -57,6 +57,8 @@ namespace dxvk {
     DirtyDrawBuffer,            ///< Indirect argument buffer is dirty
     DirtyPushConstants,         ///< Push constant data has changed
 
+    ForceWriteAfterWriteSync,   ///< Ignores barrier control flags for write-after-write hazards
+
     Count
   };
 
@@ -86,8 +88,11 @@ namespace dxvk {
    * synchronize implicitly.
    */
   enum class DxvkBarrierControl : uint32_t {
-    IgnoreComputeWriteAfterWrite  = 0,
-    IgnoreGraphicsWriteAfterWrite = 1,
+    // Ignores write-after-write hazard
+    ComputeAllowWriteOnlyOverlap  = 0,
+    ComputeAllowReadWriteOverlap  = 1,
+
+    GraphicsAllowReadWriteOverlap = 2,
   };
 
   using DxvkBarrierControlFlags  = Flags<DxvkBarrierControl>;