Merge pull request #3179 from phire/mesa-early-z

Support Conservative Depth as a fallback for Early Z (Mesa support for ZCompLoc)

Source/Core/VideoBackends/OGL/ProgramShaderCache.cpp

@@ -554,6 +554,22 @@ void ProgramShaderCache::CreateHeader()
 	break;
 	}
 
+	const char* earlyz_string = "";
+	if (!is_glsles && g_ActiveConfig.backend_info.bSupportsEarlyZ)
+	{
+		if (g_ogl_config.bSupportsEarlyFragmentTests)
+		{
+			earlyz_string = "#extension GL_ARB_shader_image_load_store : enable\n"
+			                "#define FORCE_EARLY_Z layout(early_fragment_tests) in\n";
+		}
+		else if(g_ogl_config.bSupportsConservativeDepth)
+		{
+			// See PixelShaderGen for details about this fallback.
+			earlyz_string = "#extension GL_ARB_conservative_depth : enable\n"
+			                "#define FORCE_EARLY_Z layout(depth_unchanged) out float gl_FragDepth\n";
+		}
+	}
+
 	snprintf(s_glsl_header, sizeof(s_glsl_header),
 		"%s\n"
 		"%s\n" // ubo
@@ -594,7 +610,7 @@ void ProgramShaderCache::CreateHeader()
 
 		, GetGLSLVersionString().c_str()
 		, v < GLSL_140 ? "#extension GL_ARB_uniform_buffer_object : enable" : ""
-		, !is_glsles && g_ActiveConfig.backend_info.bSupportsEarlyZ ? "#extension GL_ARB_shader_image_load_store : enable" : ""
+		, earlyz_string
 		, (g_ActiveConfig.backend_info.bSupportsBindingLayout && v < GLSLES_310) ? "#extension GL_ARB_shading_language_420pack : enable" : ""
 		, (g_ogl_config.bSupportsMSAA && v < GLSL_150) ? "#extension GL_ARB_texture_multisample : enable" : ""
 		, g_ActiveConfig.backend_info.bSupportsBindingLayout ? "#define SAMPLER_BINDING(x) layout(binding = x)" : "#define SAMPLER_BINDING(x)"

Source/Core/VideoBackends/OGL/Render.cpp

@@ -434,7 +434,6 @@ Renderer::Renderer()
 	                                                 GLExtensions::Supports("GL_EXT_blend_func_extended");
 	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");
 	g_Config.backend_info.bSupportsGSInstancing = GLExtensions::Supports("GL_ARB_gpu_shader5");
 	g_Config.backend_info.bSupportsSSAA = GLExtensions::Supports("GL_ARB_gpu_shader5") && GLExtensions::Supports("GL_ARB_sample_shading");
@@ -469,6 +468,8 @@ Renderer::Renderer()
 	g_ogl_config.bSupports3DTextureStorage = GLExtensions::Supports("GL_ARB_texture_storage_multisample") ||
 	                                         GLExtensions::Supports("GL_OES_texture_storage_multisample_2d_array");
 	g_ogl_config.bSupports2DTextureStorage = GLExtensions::Supports("GL_ARB_texture_storage_multisample");
+	g_ogl_config.bSupportsEarlyFragmentTests = GLExtensions::Supports("GL_ARB_shader_image_load_store");
+	g_ogl_config.bSupportsConservativeDepth = GLExtensions::Supports("GL_ARB_conservative_depth");
 
 	if (GLInterface->GetMode() == GLInterfaceMode::MODE_OPENGLES3)
 	{
@@ -488,7 +489,7 @@ Renderer::Renderer()
 			g_ogl_config.eSupportedGLSLVersion = GLSLES_310;
 			g_ogl_config.bSupportsAEP = GLExtensions::Supports("GL_ANDROID_extension_pack_es31a");
 			g_Config.backend_info.bSupportsBindingLayout = true;
-			g_Config.backend_info.bSupportsEarlyZ = true;
+			g_ogl_config.bSupportsEarlyFragmentTests = true;
 			g_Config.backend_info.bSupportsGeometryShaders = g_ogl_config.bSupportsAEP;
 			g_Config.backend_info.bSupportsGSInstancing = g_Config.backend_info.bSupportsGeometryShaders && g_ogl_config.SupportedESPointSize > 0;
 			g_Config.backend_info.bSupportsSSAA = g_ogl_config.bSupportsAEP;
@@ -507,7 +508,7 @@ Renderer::Renderer()
 			g_ogl_config.eSupportedGLSLVersion = GLSLES_320;
 			g_ogl_config.bSupportsAEP = GLExtensions::Supports("GL_ANDROID_extension_pack_es31a");
 			g_Config.backend_info.bSupportsBindingLayout = true;
-			g_Config.backend_info.bSupportsEarlyZ = true;
+			g_ogl_config.bSupportsEarlyFragmentTests = true;
 			g_Config.backend_info.bSupportsGeometryShaders = true;
 			g_Config.backend_info.bSupportsGSInstancing = g_ogl_config.SupportedESPointSize > 0;
 			g_Config.backend_info.bSupportsPaletteConversion = true;
@@ -533,13 +534,15 @@ Renderer::Renderer()
 		else if (strstr(g_ogl_config.glsl_version, "1.30"))
 		{
 			g_ogl_config.eSupportedGLSLVersion = GLSL_130;
-			g_Config.backend_info.bSupportsEarlyZ = false; // layout keyword is only supported on glsl150+
+			g_ogl_config.bSupportsEarlyFragmentTests = false; // layout keyword is only supported on glsl150+
+			g_ogl_config.bSupportsConservativeDepth = false; // layout keyword is only supported on glsl150+
 			g_Config.backend_info.bSupportsGeometryShaders = false; // geometry shaders are only supported on glsl150+
 		}
 		else if (strstr(g_ogl_config.glsl_version, "1.40"))
 		{
 			g_ogl_config.eSupportedGLSLVersion = GLSL_140;
-			g_Config.backend_info.bSupportsEarlyZ = false; // layout keyword is only supported on glsl150+
+			g_ogl_config.bSupportsEarlyFragmentTests = false; // layout keyword is only supported on glsl150+
+			g_ogl_config.bSupportsConservativeDepth = false; // layout keyword is only supported on glsl150+
 			g_Config.backend_info.bSupportsGeometryShaders = false; // geometry shaders are only supported on glsl150+
 		}
 		else if (strstr(g_ogl_config.glsl_version, "1.50"))
@@ -560,6 +563,9 @@ Renderer::Renderer()
 		g_ogl_config.bSupportsAEP = false;
 	}
 
+	// Either method can do early-z tests. See PixelShaderGen for details.
+	g_Config.backend_info.bSupportsEarlyZ = g_ogl_config.bSupportsEarlyFragmentTests || g_ogl_config.bSupportsConservativeDepth;
+
 	if (g_ogl_config.bSupportsDebug)
 	{
 		if (GLExtensions::Supports("GL_KHR_debug"))

Source/Core/VideoBackends/OGL/Render.h

@@ -50,6 +50,8 @@ struct VideoConfig
 	ES_TEXBUF_TYPE SupportedESTextureBuffer;
 	bool bSupports2DTextureStorage;
 	bool bSupports3DTextureStorage;
+	bool bSupportsEarlyFragmentTests;
+	bool bSupportsConservativeDepth;
 
 	const char* gl_vendor;
 	const char* gl_renderer;

Source/Core/VideoCommon/PixelShaderGen.cpp

@@ -306,14 +306,30 @@ static inline void GeneratePixelShader(T& out, DSTALPHA_MODE dstAlphaMode, API_T
 		// Most importantly, it was not possible to write to the depth buffer without also writing
 		// a color value (unless color writing was disabled altogether).
 
-		// OpenGL has a flag which allows the driver to still update the depth buffer if alpha
+		// OpenGL 4.2 actually provides two extensions which can force an early z test:
-		// test fails. The driver isn't required to do this, but I (degasus) assume all of them do
+		//  * ARB_image_load_store has 'layout(early_fragment_tests)' which forces the driver to do z and stencil tests early.
-		// because it's the much faster code path for the GPU.
+		//  * ARB_conservative_depth has 'layout(depth_unchanged) which signals to the driver that it can make optimisations
+		//    which assume the pixel shader won't update the depth buffer.
+
+		// early_fragment_tests is the best option, as it requires the driver to do early-z and defines early-z excatly as
+		// we expect, with discard causing the shader to exit with only the depth buffer updated.
+
+		// Conservative depth's 'depth_unchanged' only hints to the driver that an early-z optimistaion can be made and
+		// doesn't define what will happen if we discard the fragment. But the way modern graphics hardware is implemented
+		// means it is not unreasonable to expect the the same behaviour as early_fragment_tests.
+		// We can also assume that if a driver has gone out of it's way to support conservative depth and not image_load_store
+		// as required by OpenGL 4.2 that it will be doing the optimisation.
+		// If the driver doesn't actually do an early z optimisation, ZCompLoc will be broken and depth will only be written
+		// if the alpha test passes.
+
+		// We support Conservative as a fallback, because many drivers based on Mesa haven't implemented all of the
+		// ARB_image_load_store extension yet.
 
 		// D3D11 also has a way to force the driver to enable early-z, so we're fine here.
 		if(ApiType == API_OPENGL)
 		{
-			out.Write("layout(early_fragment_tests) in;\n");
+			 // This is a #define which signals whatever early-z method the driver supports.
+			out.Write("FORCE_EARLY_Z; \n");
 		}
 		else
 		{