GL3: ANGLE/D3D11 render-to-texture performance fixes

nel/include/nel/3d/driver.h

@@ -318,6 +318,9 @@ class IDriver : public NLMISC::CRefCount
 	/// Clear the current target surface stencil buffer. The function ignores the viewport settings but uses the scissor.
 	virtual bool			clearStencilBuffer(sint stencilval=0) = 0;
 
+	/// Clear both the depth and stencil buffers in one call. Drivers may override for efficiency.
+	virtual bool			clearDepthStencil(float zval=1, sint stencilval=0) { clearZBuffer(zval); clearStencilBuffer(stencilval); return true; }
+
 	/// Set the color mask filter through where the operation done will pass
 	virtual void			setColorMask(bool bRed, bool bGreen, bool bBlue, bool bAlpha) = 0;
 	// @}

nel/src/3d/driver/opengl3/driver_opengl3.cpp

@@ -710,6 +710,18 @@ bool CDriverGL3::activeFrameBufferObject(ITexture * tex)
 		}
 		else
 		{
+			// Invalidate depth/stencil attachments before unbinding the FBO.
+			// On ANGLE (Windows/Android WebGL), switching away from an FBO without
+			// invalidating forces an expensive depth/stencil resolve/copy. This hint
+			// tells the driver the data is no longer needed, avoiding the stall.
+			// (STORE_OP_DONT_CARE on Vulkan, DiscardView on D3D11)
+			const GLenum attachments[2] = { GL_DEPTH_ATTACHMENT, GL_STENCIL_ATTACHMENT };
+#ifdef USE_OPENGLES3
+			nglInvalidateFramebuffer(GL_FRAMEBUFFER, 2, attachments);
+#else
+			if (_Extensions.ARBInvalidateSubdata)
+				nglInvalidateFramebuffer(GL_FRAMEBUFFER, 2, attachments);
+#endif
 			_DriverGLStates.forceBindFramebuffer(0);
 			return true;
 		}
@@ -753,7 +765,8 @@ bool CDriverGL3::clearZBuffer(float zval)
 {
 	H_AUTO_OGL(CDriverGL3_clearZBuffer);
 
-	_DriverGLStates.enableZWrite(true);
+	// glClearBufferfv is not affected by the depth write mask (GL spec),
+	// so enableZWrite(true) is not needed here.
 	nglClearBufferfv(GL_DEPTH, 0, &zval);
 
 	return true;
@@ -769,6 +782,19 @@ bool CDriverGL3::clearStencilBuffer(sint stencilval)
 	return true;
 }
 
+// --------------------------------------------------
+bool CDriverGL3::clearDepthStencil(float zval, sint stencilval)
+{
+	H_AUTO_OGL(CDriverGL3_clearDepthStencil)
+
+	// Combined depth/stencil clear: one GL call instead of two.
+	// On ANGLE/D3D11 this maps to a single ClearDepthStencilView with
+	// D3D11_CLEAR_DEPTH | D3D11_CLEAR_STENCIL, avoiding two separate calls.
+	nglClearBufferfi(GL_DEPTH_STENCIL, 0, zval, stencilval);
+
+	return true;
+}
+
 // --------------------------------------------------
 void CDriverGL3::setColorMask (bool bRed, bool bGreen, bool bBlue, bool bAlpha)
 {
@@ -838,6 +864,21 @@ bool CDriverGL3::swapBuffers()
 	_SwapBufferSync[syncI] = nglFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
 	nlassert(_SwapBufferSync[syncI]);
 
+	// Invalidate the default framebuffer's depth/stencil before the swap.
+	// The depth/stencil data is never needed across frames. Without this hint,
+	// ANGLE stores the depth/stencil during present (an expensive resolve/copy
+	// on D3D11, and STORE_OP_STORE on Vulkan). With invalidation, ANGLE can
+	// skip the store (DiscardView on D3D11, STORE_OP_DONT_CARE on Vulkan).
+	{
+		const GLenum attachments[2] = { GL_DEPTH, GL_STENCIL };
+#ifdef USE_OPENGLES3
+		nglInvalidateFramebuffer(GL_FRAMEBUFFER, 2, attachments);
+#else
+		if (_Extensions.ARBInvalidateSubdata)
+			nglInvalidateFramebuffer(GL_FRAMEBUFFER, 2, attachments);
+#endif
+	}
+
 	++_SwapBufferCounter;
 
 	if (!_WndActive)

nel/src/3d/driver/opengl3/driver_opengl3.h

@@ -542,6 +542,7 @@ class CDriverGL3 : public IDriver
 
 	virtual bool			clearZBuffer(float zval=1);
 	virtual bool			clearStencilBuffer(sint stencilval=0);
+	virtual bool			clearDepthStencil(float zval=1, sint stencilval=0);
 	virtual void			setColorMask (bool bRed, bool bGreen, bool bBlue, bool bAlpha);
 	virtual void			setDepthRange(float znear, float zfar);
 	virtual	void			getDepthRange(float &znear, float &zfar) const;

nel/src/3d/driver/opengl3/driver_opengl3_extension.cpp

@@ -183,6 +183,7 @@ PFNGLGENERATEMIPMAPPROC							nglGenerateMipmap;
 PFNGLBLITFRAMEBUFFERPROC						nglBlitFramebuffer;
 PFNGLRENDERBUFFERSTORAGEMULTISAMPLEPROC			nglRenderbufferStorageMultisample;
 PFNGLFRAMEBUFFERTEXTURELAYERPROC				nglFramebufferTextureLayer;
+PFNGLINVALIDATEFRAMEBUFFERPROC					nglInvalidateFramebuffer;
 
 PFNGLACTIVETEXTUREPROC							nglActiveTexture;
 
@@ -503,6 +504,9 @@ static bool setupGLCore(std::vector<const char *> &glext)
 	CHECK_ADDRESS(PFNGLRENDERBUFFERSTORAGEMULTISAMPLEPROC, glRenderbufferStorageMultisample);
 	CHECK_ADDRESS(PFNGLFRAMEBUFFERTEXTURELAYERPROC, glFramebufferTextureLayer);
 
+	// GL 4.3 / ARB_invalidate_subdata: optional on GL 3.3, always available on GLES 3.0
+	nglInvalidateFramebuffer = (PFNGLINVALIDATEFRAMEBUFFERPROC)nglGetProcAddress("glInvalidateFramebuffer");
+
 	CHECK_ADDRESS(PFNGLACTIVETEXTUREPROC, glActiveTexture);
 
 	CHECK_ADDRESS(PFNGLCOMPRESSEDTEXIMAGE3DPROC, glCompressedTexImage3D);
@@ -603,6 +607,14 @@ static bool setupAMDPinnedMemory(std::vector<const char *> &glext)
 	return true;
 }
 
+// *********************************
+static bool setupARBInvalidateSubdata(std::vector<const char *> &glext)
+{
+	CHECK_EXT_2("GL_ARB_invalidate_subdata");
+
+	return true;
+}
+
 // *********************************
 static bool	setupNVXGPUMemoryInfo(std::vector<const char *> &glext)
 {
@@ -701,6 +713,9 @@ bool	registerGlExtensions(CGlExtensions &ext)
 	// Check GL_AMD_pinned_memory
 	ext.AMDPinnedMemory = false; // setupAMDPinnedMemory(glext); // TODO: Proper frame sync check
 
+	// Check GL_ARB_invalidate_subdata (GL 4.3 core; optional on GL 3.3)
+	ext.ARBInvalidateSubdata = setupARBInvalidateSubdata(glext);
+
 	// Memory info extensions
 	ext.NVXGPUMemoryInfo = setupNVXGPUMemoryInfo(glext);
 	ext.ATIMeminfo = setupATIMeminfo(glext);

nel/src/3d/driver/opengl3/driver_opengl3_extension.h

@@ -52,6 +52,7 @@ struct	CGlExtensions
 	bool	EXTTextureCompressionS3TC;
 	bool	EXTTextureFilterAnisotropic;
 	float	EXTTextureFilterAnisotropicMaximum;
+	bool	ARBInvalidateSubdata;
 	bool	AMDPinnedMemory;
 
 	// Extensions to get memory info
@@ -95,6 +96,7 @@ struct	CGlExtensions
 		EXTTextureCompressionS3TC = false;
 		EXTTextureFilterAnisotropic = false;
 		EXTTextureFilterAnisotropicMaximum = 0.f;
+		ARBInvalidateSubdata = false;
 		AMDPinnedMemory = false;
 
 		NVXGPUMemoryInfo = false;
@@ -133,6 +135,7 @@ struct	CGlExtensions
 
 		result += "\n  Buffers: ";
 		result += AMDPinnedMemory ? "AMDPinnedMemory " : "";
+		result += ARBInvalidateSubdata ? "ARBInvalidateSubdata " : "";
 
 		result += "\n  Memory info: ";
 		result += NVXGPUMemoryInfo ? "NVXGPUMemoryInfo " : "";
@@ -284,6 +287,7 @@ namespace NLDRIVERGL3 {
 #define nglBlitFramebuffer glBlitFramebuffer
 #define nglRenderbufferStorageMultisample glRenderbufferStorageMultisample
 #define nglFramebufferTextureLayer glFramebufferTextureLayer
+#define nglInvalidateFramebuffer glInvalidateFramebuffer
 
 #define nglActiveTexture glActiveTexture
 
@@ -491,6 +495,7 @@ extern PFNGLGENERATEMIPMAPPROC							nglGenerateMipmap;
 extern PFNGLBLITFRAMEBUFFERPROC							nglBlitFramebuffer;
 extern PFNGLRENDERBUFFERSTORAGEMULTISAMPLEPROC			nglRenderbufferStorageMultisample;
 extern PFNGLFRAMEBUFFERTEXTURELAYERPROC					nglFramebufferTextureLayer;
+extern PFNGLINVALIDATEFRAMEBUFFERPROC					nglInvalidateFramebuffer;
 
 extern PFNGLACTIVETEXTUREPROC							nglActiveTexture;
 

nel/src/3d/driver/opengl3/driver_opengl3_texture.cpp

@@ -248,12 +248,32 @@ bool CTextureDrvInfosGL3::activeFrameBufferObject(ITexture * tex)
 		{
 			_Driver->_DriverGLStates.forceBindTexture(TextureMode, 0);
 			_Driver->_DriverGLStates.forceBindFramebuffer(FBOId);
+			// Do NOT invalidate depth/stencil here at bind time.
+			// The application clears depth/stencil immediately after binding,
+			// which allows ANGLE to use LOAD_OP_CLEAR for the render pass.
+			// An invalidation here would downgrade that to LOAD_OP_DONT_CARE
+			// followed by a mid-pass vkCmdClearAttachments, which is slower.
 		}
 		else
 			return false;
 	}
 	else
 	{
+		// Invalidate depth/stencil attachments before unbinding the FBO.
+		// On ANGLE (Windows/Android WebGL), switching away from an FBO without
+		// invalidating forces an expensive depth/stencil resolve/copy. This hint
+		// tells the driver the data is no longer needed, avoiding the stall.
+		// (STORE_OP_DONT_CARE on Vulkan, DiscardView on D3D11)
+		if (AttachDepthStencil)
+		{
+			const GLenum attachments[2] = { GL_DEPTH_ATTACHMENT, GL_STENCIL_ATTACHMENT };
+#ifdef USE_OPENGLES3
+			nglInvalidateFramebuffer(GL_FRAMEBUFFER, 2, attachments);
+#else
+			if (_Driver->_Extensions.ARBInvalidateSubdata)
+				nglInvalidateFramebuffer(GL_FRAMEBUFFER, 2, attachments);
+#endif
+		}
 		_Driver->_DriverGLStates.forceBindFramebuffer(0);
 	}
 
@@ -1765,7 +1785,10 @@ bool CDriverGL3::setRenderTarget (ITexture *tex, uint32 x, uint32 y, uint32 widt
 	}
 	else if (_RenderTargetFBO)
 	{
-		activeFrameBufferObject(NULL);
+		// Deactivate through the texture-level method so it can use per-FBO
+		// state (e.g. AttachDepthStencil) for conditional invalidation.
+		CTextureDrvInfosGL3* gltext = (CTextureDrvInfosGL3*)(ITextureDrvInfos*)(_RenderTargetFBO->TextureDrvShare->DrvTexture);
+		gltext->activeFrameBufferObject(NULL);
 		setupViewport(_OldViewport);
 		_OldViewport = _CurrViewport;
 

nel/src/3d/driver_user.cpp

@@ -1357,8 +1357,7 @@ void			CDriverUser::clearBuffers(CRGBA col)
 	NL3D_HAUTO_CLEAR_DRIVER;
 
 	_Driver->clear2D(col);
-	_Driver->clearZBuffer();
-	_Driver->clearStencilBuffer();
+	_Driver->clearDepthStencil();
 }
 // ***************************************************************************
 void			CDriverUser::swapBuffers()