OGL: implement bounding box support with ssbo

This implemention tries to be as accurate as the old SW implemention, but it will remove the dependcy of our vertexloader on videosw.

Source/Core/VideoBackends/OGL/BoundingBox.cpp

@@ -0,0 +1,54 @@
+// Copyright 2014 Dolphin Emulator Project
+// Licensed under GPLv2+
+// Refer to the license.txt file included.
+
+#include "VideoBackends/OGL/BoundingBox.h"
+#include "VideoBackends/OGL/GLUtil.h"
+
+#include "VideoCommon/VideoConfig.h"
+
+static GLuint s_bbox_buffer_id;
+
+namespace OGL
+{
+
+void BoundingBox::Init()
+{
+	if (g_ActiveConfig.backend_info.bSupportsBBox)
+	{
+		int initial_values[4] = {0,0,0,0};
+		glGenBuffers(1, &s_bbox_buffer_id);
+		glBindBuffer(GL_SHADER_STORAGE_BUFFER, s_bbox_buffer_id);
+		glBufferData(GL_SHADER_STORAGE_BUFFER, 4 * sizeof(s32), initial_values, GL_DYNAMIC_DRAW);
+		glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, s_bbox_buffer_id);
+	}
+}
+
+void BoundingBox::Shutdown()
+{
+	if (g_ActiveConfig.backend_info.bSupportsBBox)
+		glDeleteBuffers(1, &s_bbox_buffer_id);
+}
+
+void BoundingBox::Set(int index, int value)
+{
+	glBindBuffer(GL_SHADER_STORAGE_BUFFER, s_bbox_buffer_id);
+	glBufferSubData(GL_SHADER_STORAGE_BUFFER, index * sizeof(int), sizeof(int), &value);
+	glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+}
+
+int BoundingBox::Get(int index)
+{
+	int data = 0;
+	glBindBuffer(GL_SHADER_STORAGE_BUFFER, s_bbox_buffer_id);
+	void* ptr = glMapBufferRange(GL_SHADER_STORAGE_BUFFER, index * sizeof(int), sizeof(int), GL_MAP_READ_BIT);
+	if (ptr)
+	{
+		data = *(int*)ptr;
+		glUnmapBuffer(GL_SHADER_STORAGE_BUFFER);
+	}
+	glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+	return data;
+}
+
+};

Source/Core/VideoBackends/OGL/BoundingBox.h

@@ -0,0 +1,20 @@
+// Copyright 2014 Dolphin Emulator Project
+// Licensed under GPLv2+
+// Refer to the license.txt file included.
+
+#pragma once
+
+namespace OGL
+{
+
+class BoundingBox
+{
+public:
+	static void Init();
+	static void Shutdown();
+
+	static void Set(int index, int value);
+	static int Get(int index);
+};
+
+};

Source/Core/VideoBackends/OGL/CMakeLists.txt

@@ -1,4 +1,5 @@
 set(SRCS GLExtensions/GLExtensions.cpp
+	   BoundingBox.cpp
 	   FramebufferManager.cpp
 	   GLUtil.cpp
 	   main.cpp

Source/Core/VideoBackends/OGL/OGL.vcxproj

@@ -35,6 +35,7 @@
   </ImportGroup>
   <PropertyGroup Label="UserMacros" />
   <ItemGroup>
+    <ClCompile Include="BoundingBox.cpp" />
     <ClCompile Include="FramebufferManager.cpp" />
     <ClCompile Include="GLExtensions\GLExtensions.cpp" />
     <ClCompile Include="GLInterface\GLInterface.cpp" />
@@ -54,6 +55,7 @@
     <ClCompile Include="VertexManager.cpp" />
   </ItemGroup>
   <ItemGroup>
+    <ClInclude Include="BoundingBox.h" />
     <ClInclude Include="FramebufferManager.h" />
     <ClInclude Include="GLExtensions\ARB_blend_func_extended.h" />
     <ClInclude Include="GLExtensions\ARB_buffer_storage.h" />
@@ -111,4 +113,4 @@
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">
   </ImportGroup>
-</Project>
+</Project>

Source/Core/VideoBackends/OGL/OGL.vcxproj.filters

@@ -36,6 +36,9 @@
     <ClCompile Include="RasterFont.cpp">
       <Filter>Logging</Filter>
     </ClCompile>
+    <ClCompile Include="BoundingBox.cpp">
+      <Filter>Render</Filter>
+    </ClCompile>
     <ClCompile Include="FramebufferManager.cpp">
       <Filter>Render</Filter>
     </ClCompile>
@@ -82,6 +85,9 @@
     <ClInclude Include="RasterFont.h">
       <Filter>Logging</Filter>
     </ClInclude>
+    <ClInclude Include="BoundingBox.h">
+      <Filter>Render</Filter>
+    </ClInclude>
     <ClInclude Include="FramebufferManager.h">
       <Filter>Render</Filter>
     </ClInclude>

Source/Core/VideoBackends/OGL/ProgramShaderCache.cpp

@@ -485,6 +485,7 @@ void ProgramShaderCache::CreateHeader()
 		"%s\n" // msaa
 		"%s\n" // sample shading
 		"%s\n" // Sampler binding
+		"%s\n" // storage buffer
 
 		// Precision defines for GLSL ES
 		"%s\n"
@@ -516,6 +517,7 @@ void ProgramShaderCache::CreateHeader()
 		, (g_ogl_config.bSupportsMSAA && v < GLSL_150) ? "#extension GL_ARB_texture_multisample : enable" : ""
 		, (g_ogl_config.bSupportSampleShading) ? "#extension GL_ARB_sample_shading : enable" : ""
 		, g_ActiveConfig.backend_info.bSupportsBindingLayout ? "#define SAMPLER_BINDING(x) layout(binding = x)" : "#define SAMPLER_BINDING(x)"
+		, g_ActiveConfig.backend_info.bSupportsBBox ? "#extension GL_ARB_shader_storage_buffer_object : enable" : ""
 
 		, v>=GLSLES_300 ? "precision highp float;" : ""
 		, v>=GLSLES_300 ? "precision highp int;" : ""

Source/Core/VideoBackends/OGL/Render.cpp

@@ -20,6 +20,7 @@
 #include "Core/Core.h"
 #include "Core/Movie.h"
 
+#include "VideoBackends/OGL/BoundingBox.h"
 #include "VideoBackends/OGL/FramebufferManager.h"
 #include "VideoBackends/OGL/GLInterfaceBase.h"
 #include "VideoBackends/OGL/GLUtil.h"
@@ -465,6 +466,7 @@ Renderer::Renderer()
 	g_Config.backend_info.bSupportsPrimitiveRestart = !DriverDetails::HasBug(DriverDetails::BUG_PRIMITIVERESTART) &&
 				((GLExtensions::Version() >= 310) || GLExtensions::Supports("GL_NV_primitive_restart"));
 	g_Config.backend_info.bSupportsEarlyZ = GLExtensions::Supports("GL_ARB_shader_image_load_store");
+	g_Config.backend_info.bSupportsBBox = GLExtensions::Supports("GL_ARB_shader_storage_buffer_object");
 
 	// Desktop OpenGL supports the binding layout if it supports 420pack
 	// OpenGL ES 3.1 supports it implicitly without an extension
@@ -1161,6 +1163,52 @@ u32 Renderer::AccessEFB(EFBAccessType type, u32 x, u32 y, u32 poke_data)
 	return 0;
 }
 
+u16 Renderer::BBoxRead(int index)
+{
+	int swapped_index = index;
+	if (index >= 2)
+		swapped_index ^= 1; // swap 2 and 3 for top/bottom
+
+	// Here we get the min/max value of the truncated position of the upscaled and swapped framebuffer.
+	// So we have to correct them to the unscaled EFB sizes.
+	int value = BoundingBox::Get(swapped_index);
+
+	if (index < 2)
+	{
+		// left/right
+		value = value * EFB_WIDTH / s_target_width;
+	}
+	else
+	{
+		// up/down -- we have to swap up and down
+		value = value * EFB_HEIGHT / s_target_height;
+		value = EFB_HEIGHT - value - 1;
+	}
+	if (index & 1)
+		value++; // fix max values to describe the outer border
+
+	return value;
+}
+
+void Renderer::BBoxWrite(int index, u16 _value)
+{
+	int value = _value; // u16 isn't enough to multiply by the efb width
+	if (index & 1)
+		value--;
+	if (index < 2)
+	{
+		value = value * s_target_width / EFB_WIDTH;
+	}
+	else
+	{
+		index ^= 1; // swap 2 and 3 for top/bottom
+		value = EFB_HEIGHT - value - 1;
+		value = value * s_target_height / EFB_HEIGHT;
+	}
+
+	BoundingBox::Set(index, value);
+}
+
 void Renderer::SetViewport()
 {
 	// reversed gxsetviewport(xorig, yorig, width, height, nearz, farz)

Source/Core/VideoBackends/OGL/Render.h

@@ -71,6 +71,9 @@ public:
 
 	u32 AccessEFB(EFBAccessType type, u32 x, u32 y, u32 poke_data) override;
 
+	u16 BBoxRead(int index) override;
+	void BBoxWrite(int index, u16 value) override;
+
 	void ResetAPIState() override;
 	void RestoreAPIState() override;
 

Source/Core/VideoBackends/OGL/main.cpp

@@ -48,6 +48,7 @@ Make AA apply instantly during gameplay if possible
 #include "Core/Core.h"
 #include "Core/Host.h"
 
+#include "VideoBackends/OGL/BoundingBox.h"
 #include "VideoBackends/OGL/FramebufferManager.h"
 #include "VideoBackends/OGL/GLInterfaceBase.h"
 #include "VideoBackends/OGL/GLUtil.h"
@@ -205,6 +206,7 @@ void VideoBackend::Video_Prepare()
 	Renderer::Init();
 	VertexLoaderManager::Init();
 	TextureConverter::Init();
+	BoundingBox::Init();
 
 	// Notify the core that the video backend is ready
 	Host_Message(WM_USER_CREATE);
@@ -229,6 +231,7 @@ void VideoBackend::Video_Cleanup()
 		// The following calls are NOT Thread Safe
 		// And need to be called from the video thread
 		Renderer::Shutdown();
+		BoundingBox::Shutdown();
 		TextureConverter::Shutdown();
 		VertexLoaderManager::Shutdown();
 		delete g_sampler_cache;