Eliminated a memcpy in DX11's vertex shader disk cache loading. Maybe games will boot faster now, who knows?

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@7422 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/Common/Src/LinearDiskCache.h

@@ -20,6 +20,7 @@
 
 #include "Common.h"
 #include <fstream>
+#include <functional>
 
 // Update this to the current SVN revision every time you change shader generation code.
 // We don't automatically get this from SVN_REV because that would mean regenerating the
@@ -43,12 +44,15 @@ enum
 // value_type[value_size]   value;
 //}
 
-template <typename K, typename V>
+// Example reader:
-class LinearDiskCacheReader
+//
-{
+//class LinearDiskCacheReader
-public:
+//{
-	virtual void Read(const K &key, const V *value, u32 value_size) = 0;
+//public:
-};
+//	template <typename F>
+//	void operator()(const K& key, u32 value_size, F get_data) const
+//	{...}
+//};
 
 // Dead simple unsorted key-value store with append functionality.
 // No random read functionality, all reading is done in OpenAndRead.
@@ -67,48 +71,45 @@ class LinearDiskCache
 {
 public:
 	// return number of read entries
-	u32 OpenAndRead(const char *filename, LinearDiskCacheReader<K, V> &reader)
+	template <typename LinearDiskCacheReader>
+	u32 OpenAndRead(const char* filename, LinearDiskCacheReader& reader)
 	{
-		using std::ios_base;
+		using std::ios;
 
 		// close any currently opened file
 		Close();
 
 		// try opening for reading/writing
-		m_file.open(filename, ios_base::in | ios_base::out | ios_base::binary | ios_base::app);
+		m_file.open(filename, ios::in | ios::out | ios::binary | ios::app);
 		
 		if (m_file.is_open() && ValidateHeader())
 		{
 			// good header, read some key/value pairs
 			u32 num_entries = 0;
-			K key;
 
-			V *value = NULL;
 			u32 value_size;
-
+			K key;
-			while (Read(&value_size))
+			while (Read(&value_size) && Read(&key))
 			{
-				delete[] value;
+				std::streamoff const pos = m_file.tellg();
-				value = new V[value_size];
 
-				// read key/value and pass to reader
+				// pass key and value_size to reader with callback function to read the data
-				if (Read(&key) && Read(value, value_size))
+				reader(key, value_size, [this, &value_size](V* data){ Read(data, value_size); });
-					reader.Read(key, value, value_size);
+
-				else
+				// seek past data (in case reader didn't read it for whatever reason)
-					break;
+				m_file.seekg(pos + (value_size * sizeof(V)), ios::beg);
 
 				++num_entries;
 			}
 			m_file.clear();
 
-			delete[] value;
 			return num_entries;
 		}
 
 		// failed to open file for reading or bad header
 		// close and recreate file
 		Close();
-		m_file.open(filename, ios_base::out | ios_base::trunc | ios_base::binary);
+		m_file.open(filename, ios::out | ios::trunc | ios::binary);
 		WriteHeader();
 		return 0;
 	}

Source/Plugins/Plugin_VideoDX11/Src/PixelShaderCache.cpp

@@ -306,12 +306,16 @@ ID3D11Buffer*const& PixelShaderCache::GetConstantBuffer()
 }
 
 // this class will load the precompiled shaders into our cache
-class PixelShaderCacheInserter : public LinearDiskCacheReader<PIXELSHADERUID, u8>
+class PixelShaderCacheInserter
 {
 public:
-	void Read(const PIXELSHADERUID &key, const u8 *value, u32 value_size)
+	template <typename F>
+	void operator()(const PIXELSHADERUID& key, u32 value_size, F get_data) const
 	{
-		PixelShaderCache::InsertByteCode(key, value, value_size);
+		std::unique_ptr<u8[]> value(new u8[value_size]);
+		get_data(value.get());
+
+		PixelShaderCache::InsertByteCode(key, value.get(), value_size);
 	}
 };
 
@@ -354,8 +358,8 @@ void PixelShaderCache::Init()
 	char cache_filename[MAX_PATH];
 	sprintf(cache_filename, "%sdx11-%s-ps.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
 		SConfig::GetInstance().m_LocalCoreStartupParameter.m_strUniqueID.c_str());
-	PixelShaderCacheInserter inserter;
+
-	g_ps_disk_cache.OpenAndRead(cache_filename, inserter);
+	g_ps_disk_cache.OpenAndRead(cache_filename, PixelShaderCacheInserter());
 }
 
 // ONLY to be used during shutdown.

Source/Plugins/Plugin_VideoDX11/Src/VertexShaderCache.cpp

@@ -68,14 +68,15 @@ ID3D11Buffer*const& VertexShaderCache::GetConstantBuffer()
 }
 
 // this class will load the precompiled shaders into our cache
-class VertexShaderCacheInserter : public LinearDiskCacheReader<VERTEXSHADERUID, u8>
+class VertexShaderCacheInserter
 {
 public:
-	void Read(const VERTEXSHADERUID &key, const u8 *value, u32 value_size)
+	template <typename F>
+	void operator()(const VERTEXSHADERUID& key, u32 value_size, F get_data) const
 	{
 		ID3D10Blob* blob = nullptr;
 		PD3D10CreateBlob(value_size, &blob);
-		memcpy(blob->GetBufferPointer(), value, value_size);	// TODO: i would like to eliminate this memcpy
+		get_data((u8*)blob->GetBufferPointer());	// read data from file into blob
 
 		VertexShaderCache::InsertByteCode(key, SharedPtr<ID3D10Blob>::FromPtr(blob));
 	}
@@ -177,8 +178,8 @@ void VertexShaderCache::Init()
 	char cache_filename[MAX_PATH];
 	sprintf(cache_filename, "%sdx11-%s-vs.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
 		SConfig::GetInstance().m_LocalCoreStartupParameter.m_strUniqueID.c_str());
-	VertexShaderCacheInserter inserter;
+
-	g_vs_disk_cache.OpenAndRead(cache_filename, inserter);
+	g_vs_disk_cache.OpenAndRead(cache_filename, VertexShaderCacheInserter());
 }
 
 void VertexShaderCache::Clear()

Source/Plugins/Plugin_VideoDX9/Src/PixelShaderCache.cpp

@@ -66,12 +66,16 @@ static LPDIRECT3DPIXELSHADER9 s_ClearProgram = NULL;
 static LPDIRECT3DPIXELSHADER9 s_rgba6_to_rgb8 = NULL;
 static LPDIRECT3DPIXELSHADER9 s_rgb8_to_rgba6 = NULL;
 
-class PixelShaderCacheInserter : public LinearDiskCacheReader<PIXELSHADERUID, u8>
+class PixelShaderCacheInserter
 {
 public:
-	void Read(const PIXELSHADERUID &key, const u8 *value, u32 value_size)
+	template <typename F>
+	void operator()(const PIXELSHADERUID& key, u32 value_size, F get_data) const
 	{
-		PixelShaderCache::InsertByteCode(key, value, value_size, false);
+		std::unique_ptr<u8[]> value(new u8[value_size]);
+		get_data(value.get());
+
+		PixelShaderCache::InsertByteCode(key, value.get(), value_size, false);
 	}
 };
 
@@ -281,8 +285,8 @@ void PixelShaderCache::Init()
 	char cache_filename[MAX_PATH];
 	sprintf(cache_filename, "%sdx9-%s-ps.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
 		SConfig::GetInstance().m_LocalCoreStartupParameter.m_strUniqueID.c_str());
-	PixelShaderCacheInserter inserter;
+
-	g_ps_disk_cache.OpenAndRead(cache_filename, inserter);
+	g_ps_disk_cache.OpenAndRead(cache_filename, PixelShaderCacheInserter());
 }
 
 // ONLY to be used during shutdown.

Source/Plugins/Plugin_VideoDX9/Src/VertexShaderCache.cpp

@@ -57,12 +57,16 @@ LPDIRECT3DVERTEXSHADER9 VertexShaderCache::GetClearVertexShader()
 }
 
 // this class will load the precompiled shaders into our cache
-class VertexShaderCacheInserter : public LinearDiskCacheReader<VERTEXSHADERUID, u8>
+class VertexShaderCacheInserter
 {
 public:
-	void Read(const VERTEXSHADERUID &key, const u8 *value, u32 value_size)
+	template <typename F>
+	void operator()(const VERTEXSHADERUID& key, u32 value_size, F get_data) const
 	{
-		VertexShaderCache::InsertByteCode(key, value, value_size, false);
+		std::unique_ptr<u8[]> value(new u8[value_size]);
+		get_data(value.get());
+
+		VertexShaderCache::InsertByteCode(key, value.get(), value_size, false);
 	}
 };
 
@@ -149,8 +153,8 @@ void VertexShaderCache::Init()
 	char cache_filename[MAX_PATH];
 	sprintf(cache_filename, "%sdx9-%s-vs.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
 		SConfig::GetInstance().m_LocalCoreStartupParameter.m_strUniqueID.c_str());
-	VertexShaderCacheInserter inserter;
+
-	g_vs_disk_cache.OpenAndRead(cache_filename, inserter);
+	g_vs_disk_cache.OpenAndRead(cache_filename, VertexShaderCacheInserter());
 }
 
 void VertexShaderCache::Clear()