[dxvk] Limit amount of memory to relocate per submission

If the resources we're moving around are in slow memory, moving a full
chunk at once can already take several milliseconds. Try to avoid that.

src/dxvk/dxvk_context.cpp

@@ -6497,8 +6497,13 @@ namespace dxvk {
 
 
   void DxvkContext::relocateQueuedResources() {
+    // Limit the number and size of resources to process per submission to
+    // something reasonable. We don't know if we are transferring over PCIe.
     constexpr static uint32_t MaxRelocationsPerSubmission = 128u;
-    auto resourceList = m_common->memoryManager().pollRelocationList(MaxRelocationsPerSubmission);
+    constexpr static uint32_t MaxRelocatedMemoryPerSubmission = 16u << 20;
+
+    auto resourceList = m_common->memoryManager().pollRelocationList(
+      MaxRelocationsPerSubmission, MaxRelocatedMemoryPerSubmission);
 
     if (resourceList.empty())
       return;

src/dxvk/dxvk_memory.cpp

@@ -469,7 +469,9 @@ namespace dxvk {
   }
 
 
-  std::vector<DxvkRelocationEntry> DxvkRelocationList::poll(uint32_t count) {
+  std::vector<DxvkRelocationEntry> DxvkRelocationList::poll(
+          uint32_t                    count,
+          VkDeviceSize                size) {
     std::lock_guard lock(m_mutex);
 
     std::vector<DxvkRelocationEntry> result;
@@ -480,10 +482,17 @@ namespace dxvk {
 
     result.reserve(count);
 
+    VkDeviceSize totalSize = 0u;
+
     for (uint32_t i = 0; i < count; i++) {
       auto iter = m_entries.begin();
 
-      result.push_back({ iter->first, iter->second });
+      if (totalSize && totalSize + iter->second.size > size)
+        break;
+
+      totalSize += iter->second.size;
+
+      result.push_back({ iter->first, iter->second.mode });
       m_entries.erase(iter);
     }
 
@@ -493,9 +502,10 @@ namespace dxvk {
 
   void DxvkRelocationList::addResource(
           Rc<DxvkPagedResource>&&     resource,
-          DxvkAllocationModes         mode) {
+          DxvkAllocationModes         mode,
+          VkDeviceSize                size) {
     std::lock_guard lock(m_mutex);
-    m_entries.emplace(std::move(resource), mode);
+    m_entries.emplace(std::move(resource), Entry { mode, size });
   }
 
 
@@ -2110,7 +2120,7 @@ namespace dxvk {
         continue;
 
       // Acquired the resource, add it to the relocation list.
-      m_relocations.addResource(std::move(resource), mode);
+      m_relocations.addResource(std::move(resource), mode, a->m_size);
     }
   }
 

src/dxvk/dxvk_memory.h

@@ -998,20 +998,24 @@ namespace dxvk {
      * Removes items from the internally stored list.
      * Any duplicate entries will be removed.
      * \param [in] count Number of entries to return
+     * \param [in] size Maximum total resource size
      * \returns List of resources to move
      */
     std::vector<DxvkRelocationEntry> poll(
-            uint32_t                    count);
+            uint32_t                    count,
+            VkDeviceSize                size);
 
     /**
      * \brief Adds relocation entry to the list
      *
      * \param [in] resource Resource to add
      * \param [in] mode Allocation mode
+     * \param [in] size Allocation size
      */
     void addResource(
             Rc<DxvkPagedResource>&&     resource,
-            DxvkAllocationModes         mode);
+            DxvkAllocationModes         mode,
+            VkDeviceSize                size);
 
     /**
      * \brief Clears list
@@ -1028,9 +1032,14 @@ namespace dxvk {
 
   private:
 
+    struct Entry {
+      DxvkAllocationModes mode = 0u;
+      VkDeviceSize        size = 0u;
+    };
+
     dxvk::mutex                               m_mutex;
     std::unordered_map<Rc<DxvkPagedResource>,
-      DxvkAllocationModes, RcHash>            m_entries;
+      Entry, RcHash>                          m_entries;
 
   };
 
@@ -1243,11 +1252,12 @@ namespace dxvk {
     /**
      * \brief Polls relocation list
      *
-     * \param [in] count Desired entry count
+     * \param [in] count Maximum resource count
+     * \param [in] size Maximum total size
      * \returns Relocation entries
      */
-    auto pollRelocationList(uint32_t count) {
-      return m_relocations.poll(count);
+    auto pollRelocationList(uint32_t count, VkDeviceSize size) {
+      return m_relocations.poll(count, size);
     }
 
   private: