Vulkan: Properly invalidate resolve textures

Trace Writer memory caching

src/xenia/gpu/trace_writer.cc

@@ -47,6 +47,7 @@ bool TraceWriter::Open(const std::wstring& path, uint32_t title_id) {
   header.title_id = title_id;
   fwrite(&header, sizeof(header), 1, file_);
 
+  cached_memory_reads_.clear();
   return true;
 }
 
@@ -58,6 +59,8 @@ void TraceWriter::Flush() {
 
 void TraceWriter::Close() {
   if (file_) {
+    cached_memory_reads_.clear();
+
     fflush(file_);
     fclose(file_);
     file_ = nullptr;
@@ -132,6 +135,31 @@ void TraceWriter::WriteMemoryRead(uint32_t base_ptr, size_t length) {
   WriteMemoryCommand(TraceCommandType::kMemoryRead, base_ptr, length);
 }
 
+void TraceWriter::WriteMemoryReadCached(uint32_t base_ptr, size_t length) {
+  if (!file_) {
+    return;
+  }
+
+  // HACK: length is guaranteed to be within 32-bits (guest memory)
+  uint64_t key = uint64_t(base_ptr) << 32 | uint64_t(length);
+  if (cached_memory_reads_.find(key) == cached_memory_reads_.end()) {
+    WriteMemoryCommand(TraceCommandType::kMemoryRead, base_ptr, length);
+    cached_memory_reads_.insert(key);
+  }
+}
+
+void TraceWriter::WriteMemoryReadCachedNop(uint32_t base_ptr, size_t length) {
+  if (!file_) {
+    return;
+  }
+
+  // HACK: length is guaranteed to be within 32-bits (guest memory)
+  uint64_t key = uint64_t(base_ptr) << 32 | uint64_t(length);
+  if (cached_memory_reads_.find(key) == cached_memory_reads_.end()) {
+    cached_memory_reads_.insert(key);
+  }
+}
+
 void TraceWriter::WriteMemoryWrite(uint32_t base_ptr, size_t length) {
   if (!file_) {
     return;

src/xenia/gpu/trace_writer.h

@@ -10,6 +10,7 @@
 #ifndef XENIA_GPU_TRACE_WRITER_H_
 #define XENIA_GPU_TRACE_WRITER_H_
 
+#include <set>
 #include <string>
 
 #include "xenia/base/filesystem.h"
@@ -36,6 +37,8 @@ class TraceWriter {
   void WritePacketStart(uint32_t base_ptr, uint32_t count);
   void WritePacketEnd();
   void WriteMemoryRead(uint32_t base_ptr, size_t length);
+  void WriteMemoryReadCached(uint32_t base_ptr, size_t length);
+  void WriteMemoryReadCachedNop(uint32_t base_ptr, size_t length);
   void WriteMemoryWrite(uint32_t base_ptr, size_t length);
   void WriteEvent(EventCommand::Type event_type);
 
@@ -43,6 +46,7 @@ class TraceWriter {
   void WriteMemoryCommand(TraceCommandType type, uint32_t base_ptr,
                           size_t length);
 
+  std::set<uint64_t> cached_memory_reads_;
   uint8_t* membase_;
   FILE* file_;
 

src/xenia/gpu/vulkan/texture_cache.cc

@@ -348,9 +348,8 @@ TextureCache::Texture* TextureCache::AllocateTexture(
   // Check the device limits for the format before we create it.
   VkFormatProperties props;
   vkGetPhysicalDeviceFormatProperties(*device_, format, &props);
-  uint32_t required_flags = VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT |
-                            VK_FORMAT_FEATURE_BLIT_DST_BIT |
-                            VK_FORMAT_FEATURE_BLIT_SRC_BIT;
+  uint32_t required_flags =
+      VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT | VK_FORMAT_FEATURE_BLIT_SRC_BIT;
   if ((props.optimalTilingFeatures & required_flags) != required_flags) {
     // Texture needs conversion on upload to a native format.
     // assert_always();
@@ -484,10 +483,13 @@ TextureCache::Texture* TextureCache::Demand(const TextureInfo& texture_info,
     if (it->second->texture_info == texture_info) {
       if (it->second->pending_invalidation) {
         // This texture has been invalidated!
-        Scavenge();
+        RemoveInvalidatedTextures();
         break;
       }
 
+      trace_writer_->WriteMemoryReadCached(texture_info.guest_address,
+                                           texture_info.input_length);
+
       return it->second;
     }
   }
@@ -501,6 +503,12 @@ TextureCache::Texture* TextureCache::Demand(const TextureInfo& texture_info,
             texture->texture_info.size_2d.logical_width &&
         texture_info.size_2d.logical_height ==
             texture->texture_info.size_2d.logical_height) {
+      if (texture->pending_invalidation) {
+        // Texture invalidated! Remove.
+        RemoveInvalidatedTextures();
+        break;
+      }
+
       // Exact match.
       // TODO: Lazy match (at an offset)
       // Upgrade this texture to a full texture.
@@ -511,6 +519,10 @@ TextureCache::Texture* TextureCache::Demand(const TextureInfo& texture_info,
         memory_->CancelAccessWatch(texture->access_watch_handle);
       }
 
+      // Tell the trace writer to cache this memory but don't read it
+      trace_writer_->WriteMemoryReadCachedNop(texture_info.guest_address,
+                                              texture_info.input_length);
+
       texture->access_watch_handle = memory_->AddPhysicalAccessWatch(
           texture_info.guest_address, texture_info.input_length,
           cpu::MMIOHandler::kWatchWrite,
@@ -548,6 +560,9 @@ TextureCache::Texture* TextureCache::Demand(const TextureInfo& texture_info,
     return nullptr;
   }
 
+  trace_writer_->WriteMemoryRead(texture_info.guest_address,
+                                 texture_info.input_length);
+
   bool uploaded = false;
   switch (texture_info.dimension) {
     case Dimension::k1D: {
@@ -1479,9 +1494,6 @@ bool TextureCache::SetupTextureBinding(VkCommandBuffer command_buffer,
   uint16_t swizzle = static_cast<uint16_t>(fetch.swizzle);
   auto view = DemandView(texture, swizzle);
 
-  trace_writer_->WriteMemoryRead(texture_info.guest_address,
-                                 texture_info.input_length);
-
   auto image_info =
       &update_set_info->image_infos[update_set_info->image_write_count];
   auto image_write =
@@ -1507,35 +1519,7 @@ bool TextureCache::SetupTextureBinding(VkCommandBuffer command_buffer,
   return true;
 }
 
-void TextureCache::ClearCache() {
-  // TODO(DrChat): Nuke everything.
-}
-
-void TextureCache::Scavenge() {
-  // Close any open descriptor pool batches
-  if (descriptor_pool_->has_open_batch()) {
-    descriptor_pool_->EndBatch();
-  }
-
-  // Free unused descriptor sets
-  // TODO(DrChat): These sets could persist across frames, we just need a smart
-  // way to detect if they're unused and free them.
-  texture_bindings_.clear();
-  descriptor_pool_->Scavenge();
-  staging_buffer_.Scavenge();
-
-  // Kill all pending delete textures.
-  if (!pending_delete_textures_.empty()) {
-    for (auto it = pending_delete_textures_.begin();
-         it != pending_delete_textures_.end();) {
-      if (!FreeTexture(*it)) {
-        break;
-      }
-
-      it = pending_delete_textures_.erase(it);
-    }
-  }
-
+void TextureCache::RemoveInvalidatedTextures() {
   // Clean up any invalidated textures.
   invalidated_textures_mutex_.lock();
   std::vector<Texture*>& invalidated_textures = *invalidated_textures_;
@@ -1574,6 +1558,37 @@ void TextureCache::Scavenge() {
   invalidated_resolve_textures_mutex_.unlock();
 }
 
+void TextureCache::ClearCache() {
+  // TODO(DrChat): Nuke everything.
+}
+
+void TextureCache::Scavenge() {
+  // Close any open descriptor pool batches
+  if (descriptor_pool_->has_open_batch()) {
+    descriptor_pool_->EndBatch();
+  }
+
+  // Free unused descriptor sets
+  // TODO(DrChat): These sets could persist across frames, we just need a smart
+  // way to detect if they're unused and free them.
+  texture_bindings_.clear();
+  descriptor_pool_->Scavenge();
+  staging_buffer_.Scavenge();
+
+  // Kill all pending delete textures.
+  RemoveInvalidatedTextures();
+  if (!pending_delete_textures_.empty()) {
+    for (auto it = pending_delete_textures_.begin();
+         it != pending_delete_textures_.end();) {
+      if (!FreeTexture(*it)) {
+        break;
+      }
+
+      it = pending_delete_textures_.erase(it);
+    }
+  }
+}
+
 }  // namespace vulkan
 }  // namespace gpu
 }  // namespace xe

src/xenia/gpu/vulkan/texture_cache.h

@@ -180,6 +180,9 @@ class TextureCache {
                            UpdateSetInfo* update_set_info,
                            const Shader::TextureBinding& binding);
 
+  // Removes invalidated textures from the cache, queues them for delete.
+  void RemoveInvalidatedTextures();
+
   Memory* memory_ = nullptr;
 
   RegisterFile* register_file_ = nullptr;