Add the remaining code for the global UBO in PushBuffer

Post-process shaders to actually add the `globalUniformBlock`, add `SetConstUniforms()` and `SetFrameUniforms()` functions to the core and material system renderers, and add the supporting glsl code.

Author: VReaperV

src/engine/renderer/GLMemory.cpp

@@ -117,6 +117,10 @@ void GLStagingBuffer::FlushAll() {
 	FlushStagingCopyQueue();
 }
 
+bool GLStagingBuffer::Active() const {
+	return buffer.id;
+}
+
 void GLStagingBuffer::InitGLBuffer() {
 	buffer.GenBuffer();
 

src/engine/renderer/GLMemory.h

@@ -175,6 +175,7 @@ class GLBuffer {
 
 	void DelBuffer() {
 		glDeleteBuffers( 1, &id );
+		id = 0;
 		mapped = false;
 	}
 
@@ -303,6 +304,8 @@ class GLStagingBuffer {
 	void FlushStagingCopyQueue();
 	void FlushAll();
 
+	bool Active() const;
+
 	void InitGLBuffer();
 	void FreeGLBuffer();
 

src/engine/renderer/Material.cpp

@@ -1593,9 +1593,24 @@ void MaterialSystem::UpdateDynamicSurfaces() {
 	GL_CheckErrors();
 }
 
+void MaterialSystem::SetConstUniforms() {
+	globalUBOProxy->SetUniform_SurfaceDescriptorsCount( surfaceDescriptorsCount );
+	uint32_t globalWorkGroupX = surfaceDescriptorsCount % MAX_COMMAND_COUNTERS == 0 ?
+		surfaceDescriptorsCount / MAX_COMMAND_COUNTERS : surfaceDescriptorsCount / MAX_COMMAND_COUNTERS + 1;
+
+	globalUBOProxy->SetUniform_FirstPortalGroup( globalWorkGroupX );
+	globalUBOProxy->SetUniform_TotalPortals( totalPortals );
+}
+
+void MaterialSystem::SetFrameUniforms() {
+	globalUBOProxy->SetUniform_Frame( nextFrame );
+
+	globalUBOProxy->SetUniform_UseFrustumCulling( r_gpuFrustumCulling.Get() );
+	globalUBOProxy->SetUniform_UseOcclusionCulling( r_gpuOcclusionCulling.Get() );
+}
+
 void MaterialSystem::UpdateFrameData() {
 	gl_clearSurfacesShader->BindProgram( 0 );
-	gl_clearSurfacesShader->SetUniform_Frame( nextFrame );
 	gl_clearSurfacesShader->DispatchCompute( MAX_VIEWS, 1, 1 );
 
 	GL_CheckErrors();
@@ -1681,15 +1696,9 @@ void MaterialSystem::CullSurfaces() {
 		uint32_t globalWorkGroupX = surfaceDescriptorsCount % MAX_COMMAND_COUNTERS == 0 ?
 			surfaceDescriptorsCount / MAX_COMMAND_COUNTERS : surfaceDescriptorsCount / MAX_COMMAND_COUNTERS + 1;
 		GL_Bind( depthImage );
-		gl_cullShader->SetUniform_Frame( nextFrame );
 		gl_cullShader->SetUniform_ViewID( view );
-		gl_cullShader->SetUniform_SurfaceDescriptorsCount( surfaceDescriptorsCount );
-		gl_cullShader->SetUniform_UseFrustumCulling( r_gpuFrustumCulling.Get() );
-		gl_cullShader->SetUniform_UseOcclusionCulling( r_gpuOcclusionCulling.Get() );
 		gl_cullShader->SetUniform_CameraPosition( origin );
 		gl_cullShader->SetUniform_ModelViewMatrix( viewMatrix );
-		gl_cullShader->SetUniform_FirstPortalGroup( globalWorkGroupX );
-		gl_cullShader->SetUniform_TotalPortals( totalPortals );
 		gl_cullShader->SetUniform_ViewWidth( depthImage->width );
 		gl_cullShader->SetUniform_ViewHeight( depthImage->height );
 		gl_cullShader->SetUniform_SurfaceCommandsOffset( surfaceCommandsCount * ( MAX_VIEWS * nextFrame + view ) );
@@ -1721,7 +1730,6 @@ void MaterialSystem::CullSurfaces() {
 		gl_cullShader->DispatchCompute( globalWorkGroupX, 1, 1 );
 
 		gl_processSurfacesShader->BindProgram( 0 );
-		gl_processSurfacesShader->SetUniform_Frame( nextFrame );
 		gl_processSurfacesShader->SetUniform_ViewID( view );
 		gl_processSurfacesShader->SetUniform_SurfaceCommandsOffset( surfaceCommandsCount * ( MAX_VIEWS * nextFrame + view ) );
 

src/engine/renderer/Material.h

@@ -362,6 +362,8 @@ class MaterialSystem {
 
 	void StartFrame();
 	void EndFrame();
+	void SetConstUniforms();
+	void SetFrameUniforms();
 
 	void GenerateDepthImages( const int width, const int height, imageParams_t imageParms );
 

src/engine/renderer/gl_shader.cpp

@@ -1332,6 +1332,12 @@ void GLShaderManager::InitShader( GLShader* shader ) {
 			ShaderDescriptor* desc = FindShader( shader->_name, shaderType.mainText, shaderType.GLType, shaderType.headers,
 				uniqueMacros, compileMacros, true );
 
+			if ( desc && glConfig.pushBufferAvailable ) {
+				desc->shaderSource = RemoveUniformsFromShaderText( desc->shaderSource, shader->_pushUniforms );
+
+				desc->shaderSource.insert( shaderType.offset, globalUniformBlock );
+			}
+
 			if ( desc && glConfig.usingMaterialSystem && shader->_useMaterialSystem ) {
 				desc->shaderSource = ShaderPostProcess( shader, desc->shaderSource, shaderType.offset );
 			}
@@ -1585,7 +1591,7 @@ void GLShaderManager::PostProcessGlobalUniforms() {
 	GLuint padding;
 	std::vector<GLUniform*>* uniforms = &( ( GLShader* ) globalUBOProxy )->_uniforms;
 	std::vector<GLUniform*> constUniforms =
-		ProcessUniforms( GLUniform::CONST, GLUniform::CONST, false, *uniforms, size, padding );
+		ProcessUniforms( GLUniform::CONST, GLUniform::CONST, !glConfig.usingBindlessTextures, *uniforms, size, padding );
 
 	GenerateUniformStructDefinesText( constUniforms, padding, 0, "globalUniforms", uniformStruct, uniformDefines );
 
@@ -1594,7 +1600,7 @@ void GLShaderManager::PostProcessGlobalUniforms() {
 	pushBuffer.constUniformsSize = size + padding;
 
 	std::vector<GLUniform*> frameUniforms =
-		ProcessUniforms( GLUniform::FRAME, GLUniform::FRAME, false, *uniforms, size, padding );
+		ProcessUniforms( GLUniform::FRAME, GLUniform::FRAME, !glConfig.usingBindlessTextures, *uniforms, size, padding );
 
 	GenerateUniformStructDefinesText( frameUniforms, padding, paddingCount, "globalUniforms", uniformStruct, uniformDefines );
 
@@ -2191,6 +2197,29 @@ GLuint GLShaderManager::SortUniforms( std::vector<GLUniform*>& uniforms ) {
 	return structSize;
 }
 
+std::vector<GLUniform*> GLShaderManager::ProcessUniforms( const GLUniform::UpdateType minType, const GLUniform::UpdateType maxType,
+	const bool skipTextures,
+	std::vector<GLUniform*>& uniforms, GLuint& structSize, GLuint& padding ) {
+	std::vector<GLUniform*> tmp;
+
+	tmp.reserve( uniforms.size() );
+	for ( GLUniform* uniform : uniforms ) {
+		if ( uniform->_updateType >= minType && uniform->_updateType <= maxType
+			&& ( !uniform->_isTexture || !skipTextures ) ) {
+			tmp.emplace_back( uniform );
+		}
+	}
+
+	structSize = SortUniforms( tmp );
+
+	const GLuint structAlignment = 4; // Material buffer is now a UBO, so it uses std140 layout, which is aligned to vec4
+	if ( structSize > 0 ) {
+		padding = ( structAlignment - ( structSize % structAlignment ) ) % structAlignment;
+	}
+
+	return tmp;
+}
+
 void GLShader::PostProcessUniforms() {
 	if ( _useMaterialSystem ) {
 		_materialSystemUniforms = gl_shaderManager.ProcessUniforms( GLUniform::MATERIAL_OR_PUSH, GLUniform::MATERIAL_OR_PUSH,
@@ -2200,7 +2229,7 @@ void GLShader::PostProcessUniforms() {
 	if ( glConfig.pushBufferAvailable && !pushSkip ) {
 		GLuint unused;
 		_pushUniforms = gl_shaderManager.ProcessUniforms( GLUniform::CONST, GLUniform::FRAME,
-			false, _uniforms, unused, unused );
+			!glConfig.usingBindlessTextures, _uniforms, unused, unused );
 	}
 }
 

src/engine/renderer/glsl_source/cameraEffects_fp.glsl

@@ -22,6 +22,8 @@ Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
 
 /* cameraEffects_fp.glsl */
 
+#define CAMERAEFFECTS_GLSL
+
 uniform sampler2D u_CurrentMap;
 
 #if defined(r_colorGrading)
@@ -73,6 +75,8 @@ DECLARE_OUTPUT(vec4)
 
 void main()
 {
+	#insert material_fp
+
 	// calculate the screen texcoord in the 0.0 to 1.0 range
 	vec2 st = gl_FragCoord.st / r_FBufSize;
 

src/engine/renderer/glsl_source/depthReduction_cp.glsl

@@ -34,6 +34,8 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 /* depthReduction_cp.glsl */
 
+#define DEPTHMAP_GLSL
+
 #insert common_cp
 
 // Keep this to 8x8 because we don't want extra shared mem etc. to be allocated, and to minimize wasted lanes
@@ -48,6 +50,8 @@ uniform uint u_ViewHeight;
 uniform bool u_InitialDepthLevel;
 
 void main() {
+    #insert material_fp
+
     const uint globalInvocationID = GLOBAL_INVOCATION_ID;
 
     const ivec2 position = ivec2( gl_GlobalInvocationID.xy );

src/engine/renderer/glsl_source/depthtile1_fp.glsl

@@ -34,6 +34,8 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 /* depthtile1_fp.glsl */
 
+#define DEPTHMAP_GLSL
+
 uniform sampler2D u_DepthMap;
 IN(flat) vec3 unprojectionParams;
 
@@ -63,6 +65,8 @@ float min16(in vec4 data0, in vec4 data1, in vec4 data2, in vec4 data3)
 }
 void	main()
 {
+  #insert material_fp
+
   vec2 st = gl_FragCoord.st * 4.0 * pixelScale;
   vec4 depth[4], mask[4];
 

src/engine/renderer/glsl_source/depthtile2_fp.glsl

@@ -34,6 +34,8 @@ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 /* depthtile2_fp.glsl */
 
+#define DEPTHTILE1_GLSL
+
 uniform sampler2D u_DepthTile1;
 
 #if __VERSION__ > 120
@@ -44,6 +46,8 @@ out vec4 outputColor;
 
 void	main()
 {
+  #insert material_fp
+
   vec2 st = gl_FragCoord.st * r_tileStep;
   float x, y;
   vec4 accum = vec4( 0.0, 99999.0, 0.0, 0.0 );

src/engine/renderer/glsl_source/fogGlobal_fp.glsl

@@ -37,6 +37,8 @@ DECLARE_OUTPUT(vec4)
 
 void	main()
 {
+	#insert material_fp
+
 	// calculate the screen texcoord in the 0.0 to 1.0 range
 	vec2 st = gl_FragCoord.st / r_FBufSize;