Implement OGL sampler cache. Allows binding a texture multiple times with different parameters. Also possibly gives a very small speed improvement.

8 files changed, 244 insertions, 71 deletions

diff --git a/Source/Plugins/Plugin_VideoOGL/CMakeLists.txt b/Source/Plugins/Plugin_VideoOGL/CMakeLists.txt
index 05b9eac08f..7f34f9e355 100644
--- a/Source/Plugins/Plugin_VideoOGL/CMakeLists.txt
+++ b/Source/Plugins/Plugin_VideoOGL/CMakeLists.txt
@@ -7,6 +7,7 @@ set(SRCS	Src/FramebufferManager.cpp
 			Src/ProgramShaderCache.cpp
 			Src/RasterFont.cpp
 			Src/Render.cpp
+			Src/SamplerCache.cpp
 			Src/StreamBuffer.cpp
 			Src/TextureCache.cpp
 			Src/TextureConverter.cpp
diff --git a/Source/Plugins/Plugin_VideoOGL/Plugin_VideoOGL.vcxproj b/Source/Plugins/Plugin_VideoOGL/Plugin_VideoOGL.vcxproj
index e0458d9a01..e1ede098cd 100644
--- a/Source/Plugins/Plugin_VideoOGL/Plugin_VideoOGL.vcxproj
+++ b/Source/Plugins/Plugin_VideoOGL/Plugin_VideoOGL.vcxproj
@@ -205,6 +205,7 @@
     <ClCompile Include="Src\ProgramShaderCache.cpp" />
     <ClCompile Include="Src\RasterFont.cpp" />
     <ClCompile Include="Src\Render.cpp" />
+    <ClCompile Include="Src\SamplerCache.cpp" />
     <ClCompile Include="Src\StreamBuffer.cpp" />
     <ClCompile Include="Src\stdafx.cpp">
       <PrecompiledHeader Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">Create</PrecompiledHeader>
@@ -228,6 +229,7 @@
     <ClInclude Include="Src\ProgramShaderCache.h" />
     <ClInclude Include="Src\RasterFont.h" />
     <ClInclude Include="Src\Render.h" />
+	<ClInclude Include="Src\SamplerCache.h" />
     <ClInclude Include="Src\StreamBuffer.h" />
     <ClInclude Include="Src\stdafx.h" />
     <ClInclude Include="Src\TextureCache.h" />
diff --git a/Source/Plugins/Plugin_VideoOGL/Src/Render.cpp b/Source/Plugins/Plugin_VideoOGL/Src/Render.cpp
index b172ce4b71..9fc6773c5b 100644
--- a/Source/Plugins/Plugin_VideoOGL/Src/Render.cpp
+++ b/Source/Plugins/Plugin_VideoOGL/Src/Render.cpp
@@ -63,6 +63,7 @@
 #include "FPSCounter.h"
 #include "ConfigManager.h"
 #include "VertexManager.h"
+#include "SamplerCache.h"
 
 #include "main.h" // Local
 #ifdef _WIN32
@@ -128,6 +129,7 @@ static const u32 EFB_CACHE_HEIGHT = (EFB_HEIGHT + EFB_CACHE_RECT_SIZE - 1) / EFB
 static bool s_efbCacheValid[2][EFB_CACHE_WIDTH * EFB_CACHE_HEIGHT];
 static std::vector<u32> s_efbCache[2][EFB_CACHE_WIDTH * EFB_CACHE_HEIGHT]; // 2 for PEEK_Z and PEEK_COLOR
 
+static SamplerCache s_sampler_cache;
 
 int GetNumMSAASamples(int MSAAMode)
 {
@@ -366,6 +368,8 @@ void Renderer::Shutdown()
 	delete s_pfont;
 	s_pfont = 0;
 	s_ShowEFBCopyRegions.Destroy();
+	
+	s_sampler_cache.Clear();
 }
 
 void Renderer::Init()
@@ -1464,7 +1468,11 @@ void Renderer::SetLineWidth()
 
 void Renderer::SetSamplerState(int stage, int texindex)
 {
-	// TODO
+	auto const& tex = bpmem.tex[texindex];
+	auto const& tm0 = tex.texMode0[stage];
+	auto const& tm1 = tex.texMode1[stage];
+	
+	s_sampler_cache.SetSamplerState((texindex * 4) + stage, tm0, tm1);
 }
 
 void Renderer::SetInterlacingMode()
diff --git a/Source/Plugins/Plugin_VideoOGL/Src/SamplerCache.cpp b/Source/Plugins/Plugin_VideoOGL/Src/SamplerCache.cpp
new file mode 100644
index 0000000000..b0e6e1f8aa
--- /dev/null
+++ b/Source/Plugins/Plugin_VideoOGL/Src/SamplerCache.cpp
@@ -0,0 +1,133 @@
+// Copyright (C) 2003 Dolphin Project.
+
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, version 2.0.
+
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU General Public License 2.0 for more details.
+
+// A copy of the GPL 2.0 should have been included with the program.
+// If not, see http://www.gnu.org/licenses/
+
+// Official SVN repository and contact information can be found at
+// http://code.google.com/p/dolphin-emu/
+
+#include "SamplerCache.h"
+
+namespace OGL
+{
+
+SamplerCache::SamplerCache()
+	: m_last_max_anisotropy()
+{}
+
+SamplerCache::~SamplerCache()
+{
+	Clear();
+}
+
+void SamplerCache::SetSamplerState(int stage, const TexMode0& tm0, const TexMode1& tm1)
+{
+	// TODO: can this go somewhere else?
+	if (m_last_max_anisotropy != g_ActiveConfig.iMaxAnisotropy)
+	{
+		m_last_max_anisotropy = g_ActiveConfig.iMaxAnisotropy;
+		Clear();
+	}
+	
+	Params params(tm0, tm1);
+	
+	// take equivalent forced linear when bForceFiltering
+	if (g_ActiveConfig.bForceFiltering)
+	{
+		params.tm0.min_filter |= 0x4;
+		params.tm0.mag_filter |= 0x1;
+	}
+	
+	// TODO: Should keep a circular buffer for each stage of recently used samplers.
+	
+	auto const& active_sampler = m_active_samplers[stage];
+	if (active_sampler.first != params || !active_sampler.second.sampler_id)
+	{
+		// Active sampler does not match parameters (or is invalid), bind the proper one.
+		auto const new_sampler = GetEntry(params);
+		
+		glBindSampler(stage, new_sampler.second.sampler_id);
+		m_active_samplers[stage] = new_sampler;
+	}
+	
+	//active_it->second.last_frame_used = frameCount;
+}
+
+auto SamplerCache::GetEntry(const Params& params) -> std::pair<Params, Value>
+{
+	auto it = m_cache.find(params);
+	if (m_cache.end() == it)
+	{
+		// Sampler not found in cache, create it.
+		Value val;
+		glGenSamplers(1, &val.sampler_id);
+		SetParameters(val.sampler_id, params);
+		
+		it = m_cache.insert(std::make_pair(params, val)).first;
+		
+		// TODO: Maybe kill old samplers if the cache gets huge. It doesn't seem to get huge though.
+		//ERROR_LOG(VIDEO, "Sampler cache size is now %ld.", m_cache.size());
+	}
+	
+	return *it;
+}
+
+void SamplerCache::SetParameters(GLuint sampler_id, const Params& params)
+{
+	static const GLint min_filters[8] =
+	{
+		GL_NEAREST,
+		GL_NEAREST_MIPMAP_NEAREST,
+		GL_NEAREST_MIPMAP_LINEAR,
+		GL_NEAREST,
+		GL_LINEAR,
+		GL_LINEAR_MIPMAP_NEAREST,
+		GL_LINEAR_MIPMAP_LINEAR,
+		GL_LINEAR,
+	};
+	
+	static const GLint wrap_settings[4] =
+	{
+		GL_CLAMP_TO_EDGE,
+		GL_REPEAT,
+		GL_MIRRORED_REPEAT,
+		GL_REPEAT,
+	};
+	
+	auto& tm0 = params.tm0;
+	auto& tm1 = params.tm1;
+
+	glSamplerParameteri(sampler_id, GL_TEXTURE_MIN_FILTER, min_filters[tm0.min_filter % ARRAYSIZE(min_filters)]);
+	glSamplerParameteri(sampler_id, GL_TEXTURE_MAG_FILTER, tm0.mag_filter ? GL_LINEAR : GL_NEAREST);
+
+	glSamplerParameteri(sampler_id, GL_TEXTURE_WRAP_S, wrap_settings[tm0.wrap_s]);
+	glSamplerParameteri(sampler_id, GL_TEXTURE_WRAP_T, wrap_settings[tm0.wrap_t]);
+	
+	glSamplerParameterf(sampler_id, GL_TEXTURE_LOD_BIAS, tm0.lod_bias / 32.f);
+	
+	glSamplerParameterf(sampler_id, GL_TEXTURE_MIN_LOD, tm1.min_lod / 16.f);
+	glSamplerParameterf(sampler_id, GL_TEXTURE_MAX_LOD, tm1.max_lod / 16.f);
+
+	if (g_ActiveConfig.iMaxAnisotropy > 0)
+		glSamplerParameterf(sampler_id, GL_TEXTURE_MAX_ANISOTROPY_EXT, (float)(1 << g_ActiveConfig.iMaxAnisotropy));
+}
+
+void SamplerCache::Clear()
+{
+	for (auto it = m_cache.begin(); it != m_cache.end(); ++it)
+	{
+		glDeleteSamplers(1, &it->second.sampler_id);
+	}
+	m_cache.clear();
+}
+
+}
diff --git a/Source/Plugins/Plugin_VideoOGL/Src/SamplerCache.h b/Source/Plugins/Plugin_VideoOGL/Src/SamplerCache.h
new file mode 100644
index 0000000000..5c1d951e1a
--- /dev/null
+++ b/Source/Plugins/Plugin_VideoOGL/Src/SamplerCache.h
@@ -0,0 +1,80 @@
+
+#ifndef INCLUDE_SAMPLER_CACHE_H_
+#define INCLUDE_SAMPLER_CACHE_H_
+
+#include <map>
+
+#include "Render.h"
+#include "GLUtil.h"
+
+namespace OGL
+{
+
+class SamplerCache : NonCopyable
+{
+public:
+	SamplerCache();
+	~SamplerCache();
+	
+	void SetSamplerState(int stage, const TexMode0& tm0, const TexMode1& tm1);
+	void Clear();
+	
+private:
+	struct Params
+	{
+		union
+		{
+			struct
+			{
+				TexMode0 tm0;
+				TexMode1 tm1;
+			};
+			
+			u64 hex;
+		};
+		
+		Params()
+			: hex()
+		{}
+		
+		Params(const TexMode0& _tm0, const TexMode1& _tm1)
+			: tm0(_tm0)
+			, tm1(_tm1)
+		{
+			static_assert(sizeof(Params) == 8, "Assuming I can treat this as a 64bit int.");
+		}
+		
+		bool operator<(const Params& other) const
+		{
+			return hex < other.hex;
+		}
+		
+		bool operator!=(const Params& other) const
+		{
+			return hex != other.hex;
+		}
+	};
+	
+	struct Value
+	{
+		Value()
+			: sampler_id()
+		//	, last_frame_used()
+		{}
+		
+		GLuint sampler_id;
+		//int last_frame_used;
+	};
+	
+	void SetParameters(GLuint sampler_id, const Params& params);
+	std::pair<Params, Value> GetEntry(const Params& params);
+
+	std::map<Params, Value> m_cache;
+	std::pair<Params, Value> m_active_samplers[8];
+	
+	int m_last_max_anisotropy;
+};
+
+}
+
+#endif
diff --git a/Source/Plugins/Plugin_VideoOGL/Src/TextureCache.cpp b/Source/Plugins/Plugin_VideoOGL/Src/TextureCache.cpp
index 815ca4433d..7bb877f8e3 100644
--- a/Source/Plugins/Plugin_VideoOGL/Src/TextureCache.cpp
+++ b/Source/Plugins/Plugin_VideoOGL/Src/TextureCache.cpp
@@ -111,34 +111,20 @@ TextureCache::TCacheEntry::~TCacheEntry()
 TextureCache::TCacheEntry::TCacheEntry()
 {
 	glGenTextures(1, &texture);
-	currmode.hex = 0;
-	currmode1.hex = 0;
 	GL_REPORT_ERRORD();
 }
 
 void TextureCache::TCacheEntry::Bind(unsigned int stage)
 {
-	// TODO: is this already done somewhere else?
-	TexMode0 &tm0 = bpmem.tex[stage >> 2].texMode0[stage & 3];
-	TexMode1 &tm1 = bpmem.tex[stage >> 2].texMode1[stage & 3];
-	
-	if(currmode.hex != tm0.hex || currmode1.hex != tm1.hex)
+	if (s_Textures[stage] != texture)
 	{
-		if(s_ActiveTexture != stage)
+		if (s_ActiveTexture != stage)
+		{
 			glActiveTexture(GL_TEXTURE0 + stage);
-		if(s_Textures[stage] != texture)
-			glBindTexture(GL_TEXTURE_2D, texture);
+			s_ActiveTexture = stage;
+		}
 		
-		SetTextureParameters(tm0, tm1);
-		s_ActiveTexture = stage;
-		s_Textures[stage] = texture;
-	} 
-	else if (s_Textures[stage] != texture) 
-	{
-		if(s_ActiveTexture != stage)
-			glActiveTexture(GL_TEXTURE0 + stage);
 		glBindTexture(GL_TEXTURE_2D, texture);
-		s_ActiveTexture = stage;
 		s_Textures[stage] = texture;
 	}
 }
@@ -225,15 +211,23 @@ TextureCache::TCacheEntryBase* TextureCache::CreateTexture(unsigned int width,
 void TextureCache::TCacheEntry::Load(unsigned int stage, unsigned int width, unsigned int height,
 	unsigned int expanded_width, unsigned int level)
 {
-	if(s_ActiveTexture != stage)
+	if (s_ActiveTexture != stage)
+	{
 		glActiveTexture(GL_TEXTURE0 + stage);
-	if(s_Textures[stage] != texture)
+		s_ActiveTexture = stage;
+	}
+	
+	if (s_Textures[stage] != texture)
+	{
 		glBindTexture(GL_TEXTURE_2D, texture);
-	s_ActiveTexture = stage;
-	s_Textures[stage] = texture;
+		s_Textures[stage] = texture;
+	}
 	
-	if(level == 0 && m_tex_levels != 0)
+	// TODO: sloppy, just do this on creation?
+	if (level == 0)
+	{
 		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, m_tex_levels - 1);
+	}
 
 	if (pcfmt != PC_TEX_FMT_DXT1)
 	{
@@ -417,47 +411,6 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
 	g_renderer->RestoreAPIState();
 }
 
-void TextureCache::TCacheEntry::SetTextureParameters(const TexMode0 &newmode, const TexMode1 &newmode1)
-{
-	const GLint c_MinLinearFilter[8] =
-	{
-		GL_NEAREST,
-		GL_NEAREST_MIPMAP_NEAREST,
-		GL_NEAREST_MIPMAP_LINEAR,
-		GL_NEAREST,
-		GL_LINEAR,
-		GL_LINEAR_MIPMAP_NEAREST,
-		GL_LINEAR_MIPMAP_LINEAR,
-		GL_LINEAR,
-	};
-	const GLint c_WrapSettings[4] =
-	{
-		GL_CLAMP_TO_EDGE,
-		GL_REPEAT,
-		GL_MIRRORED_REPEAT,
-		GL_REPEAT,
-	};
-
-	int filt = newmode.min_filter;
-	if (g_ActiveConfig.bForceFiltering && newmode.min_filter < 4)
-		filt += 4; // take equivalent forced linear
-
-	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, c_MinLinearFilter[filt & 7]);
-	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, (newmode.mag_filter || g_Config.bForceFiltering) ? GL_LINEAR : GL_NEAREST);
-
-	glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MIN_LOD, newmode1.min_lod / 16.0f);
-	glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MAX_LOD, newmode1.max_lod / 16.0f);
-	glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_LOD_BIAS, newmode.lod_bias / 32.0f);
-
-	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, c_WrapSettings[newmode.wrap_s]);
-	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, c_WrapSettings[newmode.wrap_t]);
-
-	// TODO: Reset anisotrop when changed to 1
-	if (g_Config.iMaxAnisotropy >= 1)
-		glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MAX_ANISOTROPY_EXT,
-						(float)(1 << g_ActiveConfig.iMaxAnisotropy));
-}
-
 TextureCache::TextureCache()
 {
 	const char *pColorMatrixProg = 
diff --git a/Source/Plugins/Plugin_VideoOGL/Src/TextureCache.h b/Source/Plugins/Plugin_VideoOGL/Src/TextureCache.h
index 136a349c59..f91a6ed0e4 100644
--- a/Source/Plugins/Plugin_VideoOGL/Src/TextureCache.h
+++ b/Source/Plugins/Plugin_VideoOGL/Src/TextureCache.h
@@ -65,11 +65,6 @@ private:
 
 		void Bind(unsigned int stage);
 		bool Save(const char filename[], unsigned int level);
-
-	private:
-		void SetTextureParameters(const TexMode0 &newmode, const TexMode1 &newmode1);
-		TexMode0 currmode;
-		TexMode1 currmode1;
 	};
 
 	~TextureCache();
diff --git a/Source/Plugins/Plugin_VideoOGL/Src/VertexManager.cpp b/Source/Plugins/Plugin_VideoOGL/Src/VertexManager.cpp
index c7497ed6ed..f00ec47479 100644
--- a/Source/Plugins/Plugin_VideoOGL/Src/VertexManager.cpp
+++ b/Source/Plugins/Plugin_VideoOGL/Src/VertexManager.cpp
@@ -217,6 +217,7 @@ void VertexManager::vFlush()
 	{
 		if (usedtextures & (1 << i))
 		{
+			g_renderer->SetSamplerState(i % 4, i / 4);
 			FourTexUnits &tex = bpmem.tex[i >> 2];
 			TextureCache::TCacheEntryBase* tentry = TextureCache::Load(i, 
 				(tex.texImage3[i&3].image_base/* & 0x1FFFFF*/) << 5,