[dxvk] Correctly use VK_SHARING_MODE_CONCURRENT for buffers

Fixes some out-of-spec behaviour introduced with the global buffer rework.

src/dxvk/dxvk_barrier.cpp

@@ -117,50 +117,6 @@ namespace dxvk {
   }
 
 
-  void DxvkBarrierSet::releaseBuffer(
-          DxvkBarrierSet&           acquire,
-    const DxvkBufferSliceHandle&    bufSlice,
-          uint32_t                  srcQueue,
-          VkPipelineStageFlags      srcStages,
-          VkAccessFlags             srcAccess,
-          uint32_t                  dstQueue,
-          VkPipelineStageFlags      dstStages,
-          VkAccessFlags             dstAccess) {
-    auto& release = *this;
-
-    m_allBarrierSrcStages |= srcStages;
-
-    VkBufferMemoryBarrier2 barrier = { VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2 };
-    barrier.srcStageMask                = srcStages & vk::StageDeviceMask;
-    barrier.srcAccessMask               = srcAccess & vk::AccessWriteMask;
-    barrier.dstStageMask                = VK_PIPELINE_STAGE_2_BOTTOM_OF_PIPE_BIT;
-    barrier.dstAccessMask               = 0;
-    barrier.srcQueueFamilyIndex         = srcQueue;
-    barrier.dstQueueFamilyIndex         = dstQueue;
-    barrier.buffer                      = bufSlice.handle;
-    barrier.offset                      = bufSlice.offset;
-    barrier.size                        = bufSlice.length;
-    release.m_bufBarriers.push_back(barrier);
-
-    barrier.srcStageMask                = VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT;
-    barrier.srcAccessMask               = 0;
-    barrier.dstStageMask                = dstStages;
-    barrier.dstAccessMask               = dstAccess;
-    acquire.m_bufBarriers.push_back(barrier);
-
-    if (dstAccess & vk::AccessHostMask) {
-      acquire.m_hostBarrierSrcStages |= srcStages & vk::StageDeviceMask;
-      acquire.m_hostBarrierDstAccess |= dstAccess & vk::AccessHostMask;
-    }
-
-    DxvkAccessFlags access(DxvkAccess::Read, DxvkAccess::Write);
-    release.m_bufSlices.insert(bufSlice.handle,
-      DxvkBarrierBufferSlice(bufSlice.offset, bufSlice.length, access));
-    acquire.m_bufSlices.insert(bufSlice.handle,
-      DxvkBarrierBufferSlice(bufSlice.offset, bufSlice.length, access));
-  }
-
-
   void DxvkBarrierSet::releaseImage(
           DxvkBarrierSet&           acquire,
     const Rc<DxvkImage>&            image,

src/dxvk/dxvk_barrier.h

@@ -533,16 +533,6 @@ namespace dxvk {
             VkPipelineStageFlags      dstStages,
             VkAccessFlags             dstAccess);
 
-    void releaseBuffer(
-            DxvkBarrierSet&           acquire,
-      const DxvkBufferSliceHandle&    bufSlice,
-            uint32_t                  srcQueue,
-            VkPipelineStageFlags      srcStages,
-            VkAccessFlags             srcAccess,
-            uint32_t                  dstQueue,
-            VkPipelineStageFlags      dstStages,
-            VkAccessFlags             dstAccess);
-
     void releaseImage(
             DxvkBarrierSet&           acquire,
       const Rc<DxvkImage>&            image,

src/dxvk/dxvk_buffer.cpp

@@ -15,6 +15,7 @@ namespace dxvk {
     m_allocator     (&allocator),
     m_properties    (memFlags),
     m_shaderStages  (util::shaderStages(createInfo.stages)),
+    m_sharingMode   (device->getSharingMode()),
     m_info          (createInfo) {
     // Create and assign actual buffer resource
     assignSlice(allocateSlice());
@@ -31,12 +32,13 @@ namespace dxvk {
     m_allocator     (&allocator),
     m_properties    (memFlags),
     m_shaderStages  (util::shaderStages(createInfo.stages)),
+    m_sharingMode   (device->getSharingMode()),
     m_info          (createInfo) {
     VkBufferCreateInfo info = { VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO };
     info.flags = m_info.flags;
     info.usage = m_info.usage;
     info.size = m_info.size;
-    info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+    m_sharingMode.fill(info);
 
     assignSlice(allocator.importBufferResource(info, importInfo));
   }

src/dxvk/dxvk_buffer.h

@@ -317,7 +317,7 @@ namespace dxvk {
       info.flags = m_info.flags;
       info.usage = m_info.usage;
       info.size = m_info.size;
-      info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+      m_sharingMode.fill(info);
 
       return m_allocator->createBufferResource(info, m_properties, cache);
     }
@@ -391,6 +391,7 @@ namespace dxvk {
     DxvkMemoryAllocator*        m_allocator     = nullptr;
     VkMemoryPropertyFlags       m_properties    = 0u;
     VkShaderStageFlags          m_shaderStages  = 0u;
+    DxvkSharingModeInfo         m_sharingMode   = { };
 
     DxvkBufferCreateInfo        m_info          = { };
 

src/dxvk/dxvk_context.cpp

@@ -1435,11 +1435,9 @@ namespace dxvk {
       slice.handle, slice.offset,
       dxvk::align(slice.length, 4), 0);
 
-    m_initBarriers.accessBuffer(slice,
+    m_initBarriers.accessMemory(
-      VK_PIPELINE_STAGE_TRANSFER_BIT,
+      VK_PIPELINE_STAGE_TRANSFER_BIT, VK_ACCESS_TRANSFER_WRITE_BIT,
-      VK_ACCESS_TRANSFER_WRITE_BIT,
+      buffer->info().stages, buffer->info().access);
-      buffer->info().stages,
-      buffer->info().access);
 
     m_cmd->trackResource<DxvkAccess::Write>(buffer);
   }
@@ -2556,15 +2554,23 @@ namespace dxvk {
 
     m_cmd->cmdCopyBuffer(DxvkCmdBuffer::SdmaBuffer, &copyInfo);
 
-    m_sdmaBarriers.releaseBuffer(
+    if (m_device->hasDedicatedTransferQueue()) {
-      m_initBarriers, bufferSlice,
+      // Buffers use SHARING_MODE_CONCURRENT, so no explicit queue
-      m_device->queues().transfer.queueFamily,
+      // family ownership transfer is required. Access is serialized
-      VK_PIPELINE_STAGE_TRANSFER_BIT,
+      // via a semaphore.
-      VK_ACCESS_TRANSFER_WRITE_BIT,
+      m_sdmaBarriers.accessMemory(
-      m_device->queues().graphics.queueFamily,
+        VK_PIPELINE_STAGE_TRANSFER_BIT, VK_ACCESS_TRANSFER_WRITE_BIT,
-      buffer->info().stages,
+        VK_PIPELINE_STAGE_NONE, VK_ACCESS_NONE);
-      buffer->info().access);
+
-    
+      m_initBarriers.accessMemory(
+        VK_PIPELINE_STAGE_NONE, VK_ACCESS_NONE,
+        buffer->info().stages, buffer->info().access);
+    } else {
+      m_sdmaBarriers.accessMemory(
+        VK_PIPELINE_STAGE_TRANSFER_BIT, VK_ACCESS_TRANSFER_WRITE_BIT,
+        buffer->info().stages, buffer->info().access);
+    }
+
     m_cmd->trackResource<DxvkAccess::Read>(stagingSlice.buffer());
     m_cmd->trackResource<DxvkAccess::Write>(buffer);
   }

src/dxvk/dxvk_device.cpp

@@ -14,11 +14,11 @@ namespace dxvk {
     m_instance          (instance),
     m_adapter           (adapter),
     m_vkd               (vkd),
+    m_queues            (queues),
     m_features          (features),
     m_properties        (adapter->devicePropertiesExt()),
     m_perfHints         (getPerfHints()),
     m_objects           (this),
-    m_queues            (queues),
     m_submissionQueue   (this, queueCallback) {
 
   }

src/dxvk/dxvk_device.h

@@ -134,7 +134,18 @@ namespace dxvk {
       return m_queues.transfer.queueHandle
           != m_queues.graphics.queueHandle;
     }
-    
+
+    /**
+     * \brief Queries sharing mode info
+     * \returns Sharing mode info
+     */
+    DxvkSharingModeInfo getSharingMode() const {
+      DxvkSharingModeInfo result = { };
+      result.queueFamilies[0] = m_queues.graphics.queueFamily;
+      result.queueFamilies[1] = m_queues.transfer.queueFamily;
+      return result;
+    }
+
     /**
      * \brief The instance
      * 
@@ -539,6 +550,8 @@ namespace dxvk {
     Rc<DxvkAdapter>             m_adapter;
     Rc<vk::DeviceFn>            m_vkd;
 
+    DxvkDeviceQueueSet          m_queues;
+
     DxvkDeviceFeatures          m_features;
     DxvkDeviceInfo              m_properties;
     
@@ -548,8 +561,6 @@ namespace dxvk {
     sync::Spinlock              m_statLock;
     DxvkStatCounters            m_statCounters;
     
-    DxvkDeviceQueueSet          m_queues;
-    
     DxvkRecycler<DxvkCommandList, 16> m_recycledCommandLists;
     
     DxvkSubmissionQueue         m_submissionQueue;

src/dxvk/dxvk_memory.cpp

@@ -434,7 +434,7 @@ namespace dxvk {
 
 
   DxvkMemoryAllocator::DxvkMemoryAllocator(DxvkDevice* device)
-  : m_device(device) {
+  : m_device(device), m_sharingModeInfo(m_device->getSharingMode()) {
     VkPhysicalDeviceMemoryProperties memInfo = m_device->adapter()->memoryProperties();
 
     m_memTypeCount = memInfo.memoryTypeCount;
@@ -616,7 +616,7 @@ namespace dxvk {
           DxvkLocalAllocationCache*   allocationCache) {
     Rc<DxvkResourceAllocation> allocation;
 
-    if (likely(!createInfo.flags && createInfo.sharingMode == VK_SHARING_MODE_EXCLUSIVE)) {
+    if (likely(!createInfo.flags)) {
       VkMemoryRequirements memoryRequirements = { };
       memoryRequirements.size = createInfo.size;
       memoryRequirements.alignment = GlobalBufferAlignment;
@@ -998,7 +998,7 @@ namespace dxvk {
       VkBufferCreateInfo bufferInfo = { VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO };
       bufferInfo.size = size;
       bufferInfo.usage = type.bufferUsage;
-      bufferInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+      m_sharingModeInfo.fill(bufferInfo);
 
       VkResult status = vk->vkCreateBuffer(vk->device(), &bufferInfo, nullptr, &buffer);
 
@@ -1499,7 +1499,7 @@ namespace dxvk {
                             | VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT
                             | VK_BUFFER_USAGE_TRANSFER_DST_BIT
                             | VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
-    bufferInfo.sharingMode  = VK_SHARING_MODE_EXCLUSIVE;
+    m_sharingModeInfo.fill(bufferInfo);
 
     if (getBufferMemoryRequirements(bufferInfo, requirements))
       typeMask &= requirements.memoryRequirements.memoryTypeBits;
@@ -1561,7 +1561,7 @@ namespace dxvk {
     // to be supported, but we need to be robust around buffer creation anyway.
     VkBufferCreateInfo bufferInfo = { VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO };
     bufferInfo.size = 65536;
-    bufferInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+    m_sharingModeInfo.fill(bufferInfo);
 
     VkMemoryRequirements2 requirements = { VK_STRUCTURE_TYPE_MEMORY_REQUIREMENTS_2 };
 

src/dxvk/dxvk_memory.h

@@ -213,6 +213,33 @@ namespace dxvk {
   };
 
 
+  /**
+   * \brief Sharing mode info
+   *
+   * Stores available queue families and provides methods
+   * to fill in sharing mode infos for resource creation.
+   */
+  struct DxvkSharingModeInfo {
+    std::array<uint32_t, 2u> queueFamilies = { };
+
+    VkSharingMode sharingMode() const {
+      return queueFamilies[0] != queueFamilies[1]
+        ? VK_SHARING_MODE_CONCURRENT
+        : VK_SHARING_MODE_EXCLUSIVE;
+    }
+
+    template<typename CreateInfo>
+    void fill(CreateInfo& info) const {
+      info.sharingMode = sharingMode();
+
+      if (info.sharingMode == VK_SHARING_MODE_CONCURRENT) {
+        info.queueFamilyIndexCount = queueFamilies.size();
+        info.pQueueFamilyIndices = queueFamilies.data();
+      }
+    }
+  };
+
+
   /**
    * \brief Buffer view key
    *
@@ -1121,6 +1148,8 @@ namespace dxvk {
 
     DxvkDevice* m_device;
 
+    DxvkSharingModeInfo       m_sharingModeInfo;
+
     dxvk::mutex               m_mutex;
     dxvk::condition_variable  m_cond;