Implement Program shaders cache. Seems to reduce a small amount of stuttering when F-Zero starts. Did it because I can :|

Source/Core/VideoCommon/Src/VideoConfig.h

@@ -167,6 +167,7 @@ struct VideoConfig
 		bool bSupportsGLSL;
 		bool bSupportsGLSLBinding;
 		bool bSupportsGLSLUBO;
+		bool bSupportsGLSLCache;
 	} backend_info;
 };
 

Source/Plugins/Plugin_VideoOGL/Src/ProgramShaderCache.cpp

@@ -26,6 +26,9 @@ namespace OGL
 	GLuint ProgramShaderCache::s_ps_vs_ubo;
 	GLintptr ProgramShaderCache::s_vs_data_offset;
 
+	LinearDiskCache<PROGRAMUID, u8> g_program_disk_cache;
+	GLenum ProgramFormat;
+
 	std::pair<u64, u64> ProgramShaderCache::CurrentShaderProgram;
 	const char *UniformNames[NUM_UNIFORMS] = {
 		// SAMPLERS
@@ -53,6 +56,48 @@ namespace OGL
 			I_DEPTHPARAMS,
 		};
 
+	void ProgramShaderCache::SetProgramVariables(PCacheEntry &entry, const PROGRAMUID &uid)
+	{
+		// Dunno why this is needed when I have the binding
+		// points statically set in the shader source
+		// We should only need these two functions when we don't support binding but do support UBO
+		// Driver Bug? Nvidia GTX 570, 290.xx Driver, Linux x64
+		if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
+		{
+			glUniformBlockBinding( entry.program.glprogid, 0, 1 );
+			if(uid.uid.vsid != 0) // Some things have no vertex shader
+				glUniformBlockBinding( entry.program.glprogid, 1, 2 );
+		}
+				
+		// We cache our uniform locations for now
+		// Once we move up to a newer version of GLSL, ~1.30
+		// We can remove this
+		
+		//For some reason this fails on my hardware     
+		//glGetUniformIndices(entry.program.glprogid, NUM_UNIFORMS, UniformNames, entry.program.UniformLocations);
+		//Got to do it this crappy way.
+		if (!g_ActiveConfig.backend_info.bSupportsGLSLUBO)
+			for(int a = 8; a < NUM_UNIFORMS; ++a)
+				entry.program.UniformLocations[a] = glGetUniformLocation(entry.program.glprogid, UniformNames[a]);
+		if (!g_ActiveConfig.backend_info.bSupportsGLSLBinding) 
+			for(int a = 0; a < 8; ++a)
+			{
+				// Still need to get sampler locations since we aren't binding them statically in the shaders
+				entry.program.UniformLocations[a] = glGetUniformLocation(entry.program.glprogid, UniformNames[a]);
+				if(entry.program.UniformLocations[a] != -1)
+					glUniform1i(entry.program.UniformLocations[a], a);
+			}
+
+		// Need to get some attribute locations
+		if(uid.uid.vsid != 0) // We have no vertex Shader
+		{
+			glBindAttribLocation(entry.program.glprogid, SHADER_NORM1_ATTRIB, "rawnorm1");
+			glBindAttribLocation(entry.program.glprogid, SHADER_NORM2_ATTRIB, "rawnorm2");
+			glBindAttribLocation(entry.program.glprogid, SHADER_POSMTX_ATTRIB, "fposmtx");
+		}
+	}
+
+	
 	void ProgramShaderCache::SetBothShaders(GLuint PS, GLuint VS)
 	{
 		PROGRAMUID uid;
@@ -92,49 +137,19 @@ namespace OGL
 			glAttachShader(entry.program.glprogid, entry.program.vsid);
 			
 		glAttachShader(entry.program.glprogid, entry.program.psid);
+
+		
+		glProgramParameteri(entry.program.glprogid, GL_PROGRAM_BINARY_RETRIEVABLE_HINT, GL_TRUE);
 		
 		glLinkProgram(entry.program.glprogid);
 		
 		glUseProgram(entry.program.glprogid);
 
-		// Dunno why this is needed when I have the binding
+		SetProgramVariables(entry, uid);
-		// points statically set in the shader source
-		// We should only need these two functions when we don't support binding but do support UBO
-		// Driver Bug? Nvidia GTX 570, 290.xx Driver, Linux x64
-		if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-		{
-			glUniformBlockBinding( entry.program.glprogid, 0, 1 );
-			if(uid.uid.vsid != 0) // Some things have no vertex shader
-				glUniformBlockBinding( entry.program.glprogid, 1, 2 );
-		}
-				
-		// We cache our uniform locations for now
-		// Once we move up to a newer version of GLSL, ~1.30
-		// We can remove this
-		
-		//For some reason this fails on my hardware     
-		//glGetUniformIndices(entry.program.glprogid, NUM_UNIFORMS, UniformNames, entry.program.UniformLocations);
-		//Got to do it this crappy way.
-		if (!g_ActiveConfig.backend_info.bSupportsGLSLUBO)
-			for(int a = 8; a < NUM_UNIFORMS; ++a)
-				entry.program.UniformLocations[a] = glGetUniformLocation(entry.program.glprogid, UniformNames[a]);
-		if (!g_ActiveConfig.backend_info.bSupportsGLSLBinding) 
-			for(int a = 0; a < 8; ++a)
-			{
-				// Still need to get sampler locations since we aren't binding them statically in the shaders
-				entry.program.UniformLocations[a] = glGetUniformLocation(entry.program.glprogid, UniformNames[a]);
-				if(entry.program.UniformLocations[a] != -1)
-					glUniform1i(entry.program.UniformLocations[a], a);
-			}
-
-		// Need to get some attribute locations
-		if(uid.uid.vsid != 0) // We have no vertex Shader
-		{
-			glBindAttribLocation(entry.program.glprogid, SHADER_NORM1_ATTRIB, "rawnorm1");
-			glBindAttribLocation(entry.program.glprogid, SHADER_NORM2_ATTRIB, "rawnorm2");
-			glBindAttribLocation(entry.program.glprogid, SHADER_POSMTX_ATTRIB, "fposmtx");
-		}
 
+		// Add it to our cache
+		if (g_ActiveConfig.backend_info.bSupportsGLSLCache)
+			g_program_disk_cache.Append(uid, entry.Data(), entry.Size());
 
 		pshaders[ShaderPair] = entry;
 		CurrentShaderProgram = ShaderPair;
@@ -186,10 +201,33 @@ namespace OGL
 			glBindBufferRange(GL_UNIFORM_BUFFER, 1, s_ps_vs_ubo, 0, ps_data_size);
 			glBindBufferRange(GL_UNIFORM_BUFFER, 2, s_ps_vs_ubo, s_vs_data_offset, vs_data_size);
 		}
+		
+		// Read our shader cache, only if supported
+		if (g_ActiveConfig.backend_info.bSupportsGLSLCache)
+		{
+			char cache_filename[MAX_PATH];
+			sprintf(cache_filename, "%sogl-%s-shaders.cache", File::GetUserPath(D_SHADERCACHE_IDX).c_str(),
+				SConfig::GetInstance().m_LocalCoreStartupParameter.m_strUniqueID.c_str());
+			ProgramShaderCacheInserter inserter;
+			g_program_disk_cache.OpenAndRead(cache_filename, inserter);
+			
+			GLint Supported;
+			glGetIntegerv(GL_NUM_PROGRAM_BINARY_FORMATS, &Supported);
+			
+			GLint Formats[Supported];
+			glGetIntegerv(GL_PROGRAM_BINARY_FORMATS, Formats);
+			ProgramFormat = (GLenum)Formats[0]; // We don't really care about format
+		}
 	}
 	
 	void ProgramShaderCache::Shutdown(void)
 	{
+		if (g_ActiveConfig.backend_info.bSupportsGLSLCache)
+		{
+			g_program_disk_cache.Sync();
+			g_program_disk_cache.Close();
+		}
+		
 		PCache::iterator iter = pshaders.begin();
 		for (; iter != pshaders.end(); ++iter)
 			iter->second.Destroy();

Source/Plugins/Plugin_VideoOGL/Src/ProgramShaderCache.h

@@ -25,6 +25,9 @@
 #include "PixelShaderGen.h"
 #include "VertexShaderGen.h"
 
+#include "LinearDiskCache.h"
+#include "ConfigManager.h"
+
 	union PID
 	{
 		struct {
@@ -65,14 +68,28 @@ namespace OGL
 {
 #define NUM_UNIFORMS 27
 extern const char *UniformNames[NUM_UNIFORMS];
-
+extern GLenum ProgramFormat;
 
 struct PROGRAMSHADER
 {
-	PROGRAMSHADER() : glprogid(0), vsid(0), psid(0){}
+	PROGRAMSHADER() : glprogid(0), vsid(0), psid(0), binaryLength(0){}
 	GLuint glprogid; // opengl program id
 	GLuint vsid, psid;
 	GLint UniformLocations[NUM_UNIFORMS];
+	GLint binaryLength;
+	u8 *Data()
+	{
+		glGetProgramiv(glprogid, GL_PROGRAM_BINARY_LENGTH, &binaryLength);
+        u8* binary = (u8*)malloc(binaryLength);
+        glGetProgramBinary(glprogid, binaryLength, NULL, &ProgramFormat, binary);
+        return binary;
+	}
+	GLint Size()
+	{
+		if(!binaryLength)
+			glGetProgramiv(glprogid, GL_PROGRAM_BINARY_LENGTH, &binaryLength);
+		return binaryLength;
+	}
 };
 
 
@@ -87,7 +104,44 @@ class ProgramShaderCache
 			glDeleteProgram(program.glprogid);
 			program.glprogid = 0;
 		}
+		u8* Data()
+		{
+			return program.Data();
+		}
+		GLint Size()
+		{
+			return program.Size();
+		}
 	};
+	
+	class ProgramShaderCacheInserter : public LinearDiskCacheReader<PROGRAMUID, u8>
+	{
+		public:
+			void Read(const PROGRAMUID &key, const u8 *value, u32 value_size)
+			{
+				PCacheEntry entry;
+
+				// The two shaders might not even exist anymore
+				// But it is fine, no need to worry about that
+				entry.program.vsid = key.uid.vsid;
+				entry.program.psid = key.uid.psid;
+				
+				entry.program.glprogid = glCreateProgram();
+
+				glProgramBinary(entry.program.glprogid, ProgramFormat, value, value_size);
+				
+				GLint success;
+				glGetProgramiv(entry.program.glprogid, GL_LINK_STATUS, &success);
+
+				if (success)
+				{
+					pshaders[std::make_pair(key.uid.psid, key.uid.vsid)] = entry;
+					glUseProgram(entry.program.glprogid);
+					SetProgramVariables(entry, key);
+				}
+			}
+	};
+	
 	typedef std::map<std::pair<u64, u64>, PCacheEntry> PCache;
 
 	static PCache pshaders;
@@ -96,6 +150,7 @@ class ProgramShaderCache
 
 	static GLuint s_ps_vs_ubo;
 	static GLintptr s_vs_data_offset;
+	static void SetProgramVariables(PCacheEntry &entry, const PROGRAMUID &uid);
 	
 public:
 	static PROGRAMSHADER GetShaderProgram(void);

Source/Plugins/Plugin_VideoOGL/Src/Render.cpp

@@ -315,16 +315,19 @@ Renderer::Renderer()
 	s_bHaveFramebufferBlit = strstr(ptoken, "GL_EXT_framebuffer_blit") != NULL;
 	s_bHaveCoverageMSAA = strstr(ptoken, "GL_NV_framebuffer_multisample_coverage") != NULL;
 
-
+	// TODO: Switch over to using glew once 1.6/1.7 becomes more mainstream, seems most people are stuck in 1.5
 	if (strstr((const char*)glGetString(GL_EXTENSIONS), "GL_ARB_shading_language_420pack") != NULL)
 		g_Config.backend_info.bSupportsGLSLBinding = true;
 	if (strstr((const char*)glGetString(GL_EXTENSIONS), "GL_ARB_uniform_buffer_object") != NULL)
 		g_Config.backend_info.bSupportsGLSLUBO = true;
+	if (strstr((const char*)glGetString(GL_EXTENSIONS), "GL_ARB_get_program_binary") != NULL)
+		g_Config.backend_info.bSupportsGLSLCache = true;
 	
 	UpdateActiveConfig();
-	OSD::AddMessage(StringFromFormat("Using GLSL. Supports Binding: %s UBOs: %s",
+	OSD::AddMessage(StringFromFormat("Using GLSL. Supports Binding: %s UBOs: %s Cache: %s",
 			g_ActiveConfig.backend_info.bSupportsGLSLBinding ? "True" : "False",
-			g_ActiveConfig.backend_info.bSupportsGLSLUBO ? "True" : "False").c_str(), 5000);
+			g_ActiveConfig.backend_info.bSupportsGLSLUBO ? "True" : "False",
+			g_ActiveConfig.backend_info.bSupportsGLSLCache ? "True" : "False").c_str(), 5000);
 			
 	s_LastMultisampleMode = g_ActiveConfig.iMultisampleMode;
 	s_MSAASamples = GetNumMSAASamples(s_LastMultisampleMode);