[Vulkan] Smarter transient cache

Detect cases where requested vertex data has already been uploaded in a larger buffer

src/xenia/gpu/vulkan/buffer_cache.cc

@@ -415,10 +415,22 @@ VkDeviceSize BufferCache::TryAllocateTransientData(VkDeviceSize length,
 
 VkDeviceSize BufferCache::FindCachedTransientData(uint32_t guest_address,
                                                   uint32_t guest_length) {
-  uint64_t key = uint64_t(guest_length) << 32 | uint64_t(guest_address);
+  if (transient_cache_.empty()) {
-  auto it = transient_cache_.find(key);
+    // Short-circuit exit.
-  if (it != transient_cache_.end()) {
+    return VK_WHOLE_SIZE;
-    return it->second;
+  }
+
+  // Find the first element > guest_address
+  auto it = transient_cache_.upper_bound(guest_address);
+  if (it != transient_cache_.begin()) {
+    // it = first element < guest_address
+    --it;
+
+    if (it->first <= guest_address &&
+        (it->first + it->second.first) >= (guest_address + guest_length)) {
+      // This element is contained within some existing transient data.
+      return it->second.second + (guest_address - it->first);
+    }
   }
 
   return VK_WHOLE_SIZE;
@@ -427,8 +439,17 @@ VkDeviceSize BufferCache::FindCachedTransientData(uint32_t guest_address,
 void BufferCache::CacheTransientData(uint32_t guest_address,
                                      uint32_t guest_length,
                                      VkDeviceSize offset) {
-  uint64_t key = uint64_t(guest_length) << 32 | uint64_t(guest_address);
+  transient_cache_[guest_address] = {guest_length, offset};
-  transient_cache_[key] = offset;
+
+  // Erase any entries contained within
+  auto it = transient_cache_.upper_bound(guest_address);
+  while (it != transient_cache_.end()) {
+    if ((guest_address + guest_length) >= (it->first + it->second.first)) {
+      it = transient_cache_.erase(it);
+    } else {
+      break;
+    }
+  }
 }
 
 void BufferCache::Flush(VkCommandBuffer command_buffer) {

src/xenia/gpu/vulkan/buffer_cache.h

@@ -122,7 +122,7 @@ class BufferCache {
   // Staging ringbuffer we cycle through fast. Used for data we don't
   // plan on keeping past the current frame.
   std::unique_ptr<ui::vulkan::CircularBuffer> transient_buffer_ = nullptr;
-  std::map<uint64_t, VkDeviceSize> transient_cache_;
+  std::map<uint32_t, std::pair<uint32_t, VkDeviceSize>> transient_cache_;
 
   VkDescriptorPool descriptor_pool_ = nullptr;
   VkDescriptorSetLayout descriptor_set_layout_ = nullptr;