Merge pull request #1292 from vlj/gl

gl: Use factored scale offset function call

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -1063,6 +1063,9 @@ void GLGSRender::oninit_thread()
 	m_vao.create();
 	m_vbo.create();
 	m_ebo.create();
+	m_scale_offset_buffer.create(16 * sizeof(float));
+
+	glBindBufferBase(GL_UNIFORM_BUFFER, 0, m_scale_offset_buffer.id());
 
 	m_vao.array_buffer = m_vbo;
 	m_vao.element_array_buffer = m_ebo;
@@ -1259,37 +1262,12 @@ bool GLGSRender::load_program()
 
 	(m_program.recreate() += { fp.compile(), vp.compile() }).make();
 #endif
+	glBindBuffer(GL_UNIFORM_BUFFER, m_scale_offset_buffer.id());
 
-	int viewport_x = int(rsx::method_registers[NV4097_SET_VIEWPORT_HORIZONTAL] & 0xffff);
+	void *buffer = glMapBuffer(GL_UNIFORM_BUFFER, GL_WRITE_ONLY);
-	int viewport_y = int(rsx::method_registers[NV4097_SET_VIEWPORT_VERTICAL] & 0xffff);
+	fill_scale_offset_data(buffer, false);
-	int viewport_w = int(rsx::method_registers[NV4097_SET_VIEWPORT_HORIZONTAL] >> 16);
+	glUnmapBuffer(GL_UNIFORM_BUFFER);
-	int viewport_h = int(rsx::method_registers[NV4097_SET_VIEWPORT_VERTICAL] >> 16);
 
-	f32 viewport_offset_x = (f32&)rsx::method_registers[NV4097_SET_VIEWPORT_OFFSET + 0];
-	f32 viewport_offset_y = (f32&)rsx::method_registers[NV4097_SET_VIEWPORT_OFFSET + 1];
-	f32 viewport_offset_z = (f32&)rsx::method_registers[NV4097_SET_VIEWPORT_OFFSET + 2];
-	f32 viewport_offset_w = (f32&)rsx::method_registers[NV4097_SET_VIEWPORT_OFFSET + 3];
-
-	f32 viewport_scale_x = (f32&)rsx::method_registers[NV4097_SET_VIEWPORT_SCALE + 0];
-	f32 viewport_scale_y = (f32&)rsx::method_registers[NV4097_SET_VIEWPORT_SCALE + 1];
-	f32 viewport_scale_z = (f32&)rsx::method_registers[NV4097_SET_VIEWPORT_SCALE + 2];
-	f32 viewport_scale_w = (f32&)rsx::method_registers[NV4097_SET_VIEWPORT_SCALE + 3];
-
-	f32 width = f32(rsx::method_registers[NV4097_SET_SURFACE_CLIP_HORIZONTAL] >> 16);
-	f32 height = f32(rsx::method_registers[NV4097_SET_SURFACE_CLIP_VERTICAL] >> 16);
-	glm::mat4 scaleOffsetMat(1.f);
-
-	//Scale
-	scaleOffsetMat[0][0] = viewport_scale_x * 2.f / width;
-	scaleOffsetMat[1][1] = viewport_scale_y * 2.f / height;
-	scaleOffsetMat[2][2] = viewport_scale_z;
-
-	// Offset
-	scaleOffsetMat[0][3] = viewport_offset_x * 2.f / width - 1.f;
-	scaleOffsetMat[1][3] = viewport_offset_y * 2.f / height - 1.f;
-	scaleOffsetMat[2][3] = viewport_offset_z - .5f;
-
-	__glcheck m_program->uniforms["scaleOffsetMat"] = scaleOffsetMat;
 
 	for (auto &constant : transform_constants)
 	{

rpcs3/Emu/RSX/GL/GLGSRender.h

@@ -75,6 +75,8 @@ private:
 	gl::fbo m_flip_fbo;
 	gl::texture m_flip_tex_color;
 
+	gl::buffer m_scale_offset_buffer;
+
 	gl::buffer m_vbo;
 	gl::buffer m_ebo;
 	gl::vao m_vao;

rpcs3/Emu/RSX/GL/GLProcTable.h

@@ -163,6 +163,9 @@ OPENGL_PROC(PFNGLGETINTEGER64VPROC, GetInteger64v);
 
 OPENGL_PROC(PFNGLCHECKFRAMEBUFFERSTATUSPROC, CheckFramebufferStatus);
 
+
+OPENGL_PROC(PFNGLBINDBUFFERBASEPROC, BindBufferBase);
+
 //KHR_debug
 OPENGL_PROC(PFNGLDEBUGMESSAGECONTROLARBPROC, DebugMessageControlARB);
 OPENGL_PROC(PFNGLDEBUGMESSAGEINSERTARBPROC, DebugMessageInsertARB);

rpcs3/Emu/RSX/GL/GLVertexProgram.cpp

@@ -28,8 +28,11 @@ std::string GLVertexDecompilerThread::compareFunction(COMPARE f, const std::stri
 
 void GLVertexDecompilerThread::insertHeader(std::stringstream &OS)
 {
-	OS << "#version 140" << std::endl << std::endl;
+	OS << "#version 420" << std::endl << std::endl;
-	OS << "uniform mat4 scaleOffsetMat = mat4(1.0);" << std::endl;
+	OS << "layout(std140, binding = 0) uniform ScaleOffsetBuffer" << std::endl;
+	OS << "{" << std::endl;
+	OS << "	mat4 scaleOffsetMat;" << std::endl;
+	OS << "};" << std::endl;
 }
 
 void GLVertexDecompilerThread::insertInputs(std::stringstream & OS, const std::vector<ParamType>& inputs)

rpcs3/Emu/RSX/RSXThread.cpp

@@ -1100,7 +1100,7 @@ namespace rsx
 		onexit_thread();
 	}
 
-	void thread::fill_scale_offset_data(void *buffer) const noexcept
+	void thread::fill_scale_offset_data(void *buffer, bool is_d3d) const noexcept
 	{
 		int clip_w = rsx::method_registers[NV4097_SET_SURFACE_CLIP_HORIZONTAL] >> 16;
 		int clip_h = rsx::method_registers[NV4097_SET_SURFACE_CLIP_VERTICAL] >> 16;
@@ -1109,12 +1109,15 @@ namespace rsx
 		float offset_x = (float&)rsx::method_registers[NV4097_SET_VIEWPORT_OFFSET] - (clip_w / 2.f);
 		offset_x /= clip_w / 2.f;
 
-		float scale_y = -(float&)rsx::method_registers[NV4097_SET_VIEWPORT_SCALE + 1] / (clip_h / 2.f);
+		float scale_y = (float&)rsx::method_registers[NV4097_SET_VIEWPORT_SCALE + 1] / (clip_h / 2.f);
-		float offset_y = -((float&)rsx::method_registers[NV4097_SET_VIEWPORT_OFFSET + 1] - (clip_h / 2.f));
+		float offset_y = ((float&)rsx::method_registers[NV4097_SET_VIEWPORT_OFFSET + 1] - (clip_h / 2.f));
 		offset_y /= clip_h / 2.f;
+		if (is_d3d) scale_y *= -1;
+		if (is_d3d) offset_y *= -1;
 
 		float scale_z = (float&)rsx::method_registers[NV4097_SET_VIEWPORT_SCALE + 2];
 		float offset_z = (float&)rsx::method_registers[NV4097_SET_VIEWPORT_OFFSET + 2];
+		if (!is_d3d) offset_z -= .5;
 
 		float one = 1.f;
 

rpcs3/Emu/RSX/RSXThread.h

@@ -256,8 +256,9 @@ namespace rsx
 		/**
 		 * Fill buffer with 4x4 scale offset matrix.
 		 * Vertex shader's position is to be multiplied by this matrix.
+		 * if is_d3d is set, the matrix is modified to use d3d convention.
 		 */
-		void fill_scale_offset_data(void *buffer) const noexcept;
+		void fill_scale_offset_data(void *buffer, bool is_d3d = true) const noexcept;
 
 		/**
 		* Fill buffer with vertex program constants.