rsx: Drop format information from the vertex cache and use a simpler model that is much faster

rpcs3/Emu/RSX/GL/GLGSRender.h

@@ -20,8 +20,8 @@
 
 namespace gl
 {
-	using vertex_cache = rsx::vertex_cache::default_vertex_cache<rsx::vertex_cache::uploaded_range<GLenum>, GLenum>;
+	using vertex_cache = rsx::vertex_cache::default_vertex_cache<rsx::vertex_cache::uploaded_range>;
-	using weak_vertex_cache = rsx::vertex_cache::weak_vertex_cache<GLenum>;
+	using weak_vertex_cache = rsx::vertex_cache::weak_vertex_cache;
 	using null_vertex_cache = vertex_cache;
 
 	using shader_cache = rsx::shaders_cache<void*, GLProgramBuffer>;

rpcs3/Emu/RSX/GL/GLVertexBuffers.cpp

@@ -195,7 +195,7 @@ gl::vertex_upload_info GLGSRender::set_vertex_buffer()
 			const auto data_offset = (vertex_base * m_vertex_layout.interleaved_blocks[0]->attribute_stride);
 			storage_address = m_vertex_layout.interleaved_blocks[0]->real_offset_address + data_offset;
 
-			if (auto cached = m_vertex_cache->find_vertex_range(storage_address, GL_R8UI, required.first))
+			if (auto cached = m_vertex_cache->find_vertex_range(storage_address, required.first))
 			{
 				ensure(cached->local_address == storage_address);
 
@@ -216,7 +216,7 @@ gl::vertex_upload_info GLGSRender::set_vertex_buffer()
 			if (to_store)
 			{
 				//store ref in vertex cache
-				m_vertex_cache->store_range(storage_address, GL_R8UI, required.first, persistent_mapping.second);
+				m_vertex_cache->store_range(storage_address, required.first, persistent_mapping.second);
 			}
 		}
 

rpcs3/Emu/RSX/VK/VKGSRenderTypes.hpp

@@ -31,8 +31,8 @@ namespace vk
 	struct program_cache;
 	struct pipeline_props;
 
-	using vertex_cache = rsx::vertex_cache::default_vertex_cache<rsx::vertex_cache::uploaded_range<VkFormat>, VkFormat>;
+	using vertex_cache = rsx::vertex_cache::default_vertex_cache<rsx::vertex_cache::uploaded_range>;
-	using weak_vertex_cache = rsx::vertex_cache::weak_vertex_cache<VkFormat>;
+	using weak_vertex_cache = rsx::vertex_cache::weak_vertex_cache;
 	using null_vertex_cache = vertex_cache;
 
 	using shader_cache = rsx::shaders_cache<vk::pipeline_props, vk::program_cache>;

rpcs3/Emu/RSX/VK/VKVertexBuffers.cpp

@@ -251,7 +251,7 @@ vk::vertex_upload_info VKGSRender::upload_vertex_data()
 			const auto data_offset = (vertex_base * m_vertex_layout.interleaved_blocks[0]->attribute_stride);
 			storage_address = m_vertex_layout.interleaved_blocks[0]->real_offset_address + data_offset;
 
-			if (auto cached = m_vertex_cache->find_vertex_range(storage_address, VK_FORMAT_R8_UINT, required.first))
+			if (auto cached = m_vertex_cache->find_vertex_range(storage_address, required.first))
 			{
 				ensure(cached->local_address == storage_address);
 
@@ -274,7 +274,7 @@ vk::vertex_upload_info VKGSRender::upload_vertex_data()
 			if (to_store)
 			{
 				//store ref in vertex cache
-				m_vertex_cache->store_range(storage_address, VK_FORMAT_R8_UINT, required.first, static_cast<u32>(persistent_offset));
+				m_vertex_cache->store_range(storage_address, required.first, static_cast<u32>(persistent_offset));
 			}
 		}
 	}

rpcs3/Emu/RSX/rsx_cache.h

@@ -3,6 +3,7 @@
 #include "Utilities/lockless.h"
 #include "Utilities/Thread.h"
 #include "Common/bitfield.hpp"
+#include "Common/unordered_map.hpp"
 #include "Emu/System.h"
 #include "Emu/cache_utils.hpp"
 #include "Program/ProgramStateCache.h"
@@ -10,7 +11,6 @@
 #include "Overlays/Shaders/shader_loading_dialog.h"
 
 #include <chrono>
-#include <unordered_map>
 
 #include "util/sysinfo.hpp"
 #include "util/fnv_hash.hpp"
@@ -447,67 +447,62 @@ namespace rsx
 	namespace vertex_cache
 	{
 		// A null vertex cache
-		template <typename storage_type, typename upload_format>
+		template <typename storage_type>
 		class default_vertex_cache
 		{
 		public:
 			virtual ~default_vertex_cache() = default;
-			virtual storage_type* find_vertex_range(uptr /*local_addr*/, upload_format, u32 /*data_length*/) { return nullptr; }
+			virtual const storage_type* find_vertex_range(u32 /*local_addr*/, u32 /*data_length*/) { return nullptr; }
-			virtual void store_range(uptr /*local_addr*/, upload_format, u32 /*data_length*/, u32 /*offset_in_heap*/) {}
+			virtual void store_range(u32 /*local_addr*/, u32 /*data_length*/, u32 /*offset_in_heap*/) {}
 			virtual void purge() {}
 		};
 
-		// A weak vertex cache with no data checks or memory range locks
-		// Of limited use since contents are only guaranteed to be valid once per frame
-		// TODO: Strict vertex cache with range locks
-		template <typename upload_format>
 		struct uploaded_range
 		{
 			uptr local_address;
-			upload_format buffer_format;
 			u32 offset_in_heap;
 			u32 data_length;
 		};
 
-		template <typename upload_format>
+		// A weak vertex cache with no data checks or memory range locks
-		class weak_vertex_cache : public default_vertex_cache<uploaded_range<upload_format>, upload_format>
+		// Of limited use since contents are only guaranteed to be valid once per frame
+		// Supports upto 1GiB block lengths if typed and full 4GiB otherwise.
+		// Using a 1:1 hash-value with robin-hood is 2x faster than what we had before with std-map-of-arrays.
+		class weak_vertex_cache : public default_vertex_cache<uploaded_range>
 		{
-			using storage_type = uploaded_range<upload_format>;
+			using storage_type = uploaded_range;
 
 		private:
-			std::unordered_map<uptr, std::vector<storage_type>> vertex_ranges;
+			rsx::unordered_map<uptr, storage_type> vertex_ranges;
+
+			FORCE_INLINE u64 hash(u32 local_addr, u32 data_length) const
+			{
+				return u64(local_addr) | (u64(data_length) << 32);
+			}
 
 		public:
 
-			storage_type* find_vertex_range(uptr local_addr, upload_format fmt, u32 data_length) override
+			const storage_type* find_vertex_range(u32 local_addr, u32 data_length) override
 			{
-				auto found = vertex_ranges.find(local_addr);
+				const auto key = hash(local_addr, data_length);
+				const auto found = vertex_ranges.find(key);
 				if (found == vertex_ranges.end())
 				{
 					return nullptr;
 				}
 
-				for (auto &v : found->second)
+				return std::addressof(found->second);
-				{
-					// NOTE: This has to match exactly. Using sized shortcuts such as >= comparison causes artifacting in some applications (UC1)
-					if (v.data_length == data_length && v.buffer_format == fmt)
-					{
-						return &v;
-					}
 			}
 
-				return nullptr;
+			void store_range(u32 local_addr, u32 data_length, u32 offset_in_heap) override
-			}
-
-			void store_range(uptr local_addr, upload_format fmt, u32 data_length, u32 offset_in_heap) override
 			{
 				storage_type v = {};
-				v.buffer_format = fmt;
 				v.data_length = data_length;
 				v.local_address = local_addr;
 				v.offset_in_heap = offset_in_heap;
 
-				vertex_ranges[local_addr].push_back(v);
+				const auto key = hash(local_addr, data_length);
+				vertex_ranges[key] = v;
 			}
 
 			void purge() override

rpcs3/emucore.vcxproj

@@ -556,6 +556,7 @@
     <ClInclude Include="Emu\RSX\Common\simple_array.hpp" />
     <ClInclude Include="Emu\RSX\Common\surface_cache_dma.hpp" />
     <ClInclude Include="Emu\RSX\Common\time.hpp" />
+    <ClInclude Include="Emu\RSX\Common\unordered_map.hpp" />
     <ClInclude Include="Emu\RSX\Core\RSXEngLock.hpp" />
     <ClInclude Include="Emu\RSX\Core\RSXFrameBuffer.h" />
     <ClInclude Include="Emu\RSX\Core\RSXIOMap.hpp" />

rpcs3/emucore.vcxproj.filters

@@ -2365,6 +2365,9 @@
     <ClInclude Include="Emu\Io\emulated_pad_config.h">
       <Filter>Emu\Io</Filter>
     </ClInclude>
+    <ClInclude Include="Emu\RSX\Common\unordered_map.hpp">
+      <Filter>Emu\GPU\RSX\Common</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <None Include="Emu\RSX\Program\GLSLSnippets\GPUDeswizzle.glsl">