rsx/gl: Implement support for framebuffer-less rendering

rpcs3/Emu/RSX/GL/GLGSRender.cpp

@@ -192,6 +192,11 @@ void GLGSRender::on_init_thread()
 		backend_config.supports_hw_instanced_rendering = false;
 	}
 
+	if (gl_caps.ARB_framebuffer_no_attachments_supported)
+	{
+		backend_config.supports_framebufferless_rendering = true;
+	}
+
 	if (!gl_caps.ARB_bindless_texture_supported)
 	{
 		switch (shadermode)

rpcs3/Emu/RSX/GL/GLProcTable.h

@@ -158,6 +158,8 @@ OPENGL_PROC(PFNGLNAMEDFRAMEBUFFERDRAWBUFFERSPROC, NamedFramebufferDrawBuffers);
 OPENGL_PROC(PFNGLFRAMEBUFFERDRAWBUFFERSEXTPROC, FramebufferDrawBuffersEXT);
 OPENGL_PROC(PFNGLNAMEDFRAMEBUFFERREADBUFFERPROC, NamedFramebufferReadBuffer);
 OPENGL_PROC(PFNGLFRAMEBUFFERREADBUFFEREXTPROC, FramebufferReadBufferEXT);
+OPENGL_PROC(PFNGLNAMEDFRAMEBUFFERPARAMETERIPROC, NamedFramebufferParameteri);
+OPENGL_PROC(PFNGLNAMEDFRAMEBUFFERPARAMETERIEXTPROC, NamedFramebufferParameteriEXT);
 
 OPENGL_PROC(PFNGLENABLEIPROC, Enablei);
 OPENGL_PROC(PFNGLDISABLEIPROC, Disablei);

rpcs3/Emu/RSX/GL/GLRenderTargets.cpp

@@ -268,6 +268,18 @@ void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool /*
 
 	ensure(m_draw_fbo);
 
+	if (!depth_stencil_target &&
+		std::all_of(color_targets.begin(), color_targets.end(), FN(x == GL_NONE)))
+	{
+		// Framebufferless rendering. Raster dimensions must be supplied explicitly and match what real surfaces would have used.
+		const auto [raster_width, raster_height] = rsx::apply_resolution_scale<true>(
+			resolution_scaling_config, m_framebuffer_layout.width, m_framebuffer_layout.height);
+
+		const GLint raster_samples = (g_cfg.video.antialiasing_level == msaa_level::_auto) ? samples : 1;
+		m_draw_fbo->set_default_extents({ raster_width, raster_height });
+		m_draw_fbo->set_default_samples(raster_samples);
+	}
+
 	switch (rsx::method_registers.surface_color_target())
 	{
 	case rsx::surface_target::none: break;

rpcs3/Emu/RSX/GL/glutils/capabilities.cpp

@@ -96,6 +96,8 @@ namespace gl
 
 		CHECK_EXTENSION_SUPPORT(ARB_shader_storage_buffer_object);
 
+		CHECK_EXTENSION_SUPPORT(ARB_framebuffer_no_attachments);
+
 #undef CHECK_EXTENSION_SUPPORT
 
 		// Set GLSL version

rpcs3/Emu/RSX/GL/glutils/capabilities.h

@@ -44,6 +44,7 @@ namespace gl
 		bool ARB_shader_texture_image_samples_supported = false;
 		bool EXT_texture_compression_s3tc_supported = false;
 		bool ARB_shader_storage_buffer_object_supported = false;
+		bool ARB_framebuffer_no_attachments_supported = false;
 
 		bool vendor_INTEL = false;  // has broken GLSL compiler
 		bool vendor_AMD = false;    // has broken ARB_multidraw

rpcs3/Emu/RSX/GL/glutils/fbo.cpp

@@ -255,6 +255,18 @@ namespace gl
 		return m_size;
 	}
 
+	void fbo::set_default_extents(const size2i& extents) const
+	{
+		DSA_CALL2(NamedFramebufferParameteri, m_id, GL_FRAMEBUFFER_DEFAULT_WIDTH, extents.width);
+		DSA_CALL2(NamedFramebufferParameteri, m_id, GL_FRAMEBUFFER_DEFAULT_HEIGHT, extents.height);
+	}
+
+	void fbo::set_default_samples(GLint samples) const
+	{
+		DSA_CALL2(NamedFramebufferParameteri, m_id, GL_FRAMEBUFFER_DEFAULT_SAMPLES, samples > 1 ? samples : 0);
+		DSA_CALL2(NamedFramebufferParameteri, m_id, GL_FRAMEBUFFER_DEFAULT_FIXED_SAMPLE_LOCATIONS, GL_TRUE);
+	}
+
 	bool fbo::matches(const std::array<GLuint, 4>& color_targets, GLuint depth_stencil_target) const
 	{
 		for (u32 index = 0; index < 4; ++index)

rpcs3/Emu/RSX/GL/glutils/fbo.h

@@ -257,6 +257,9 @@ namespace gl
 		void set_extents(const size2i& extents);
 		size2i get_extents() const;
 
+		void set_default_extents(const size2i& extents) const;
+		void set_default_samples(GLint samples) const;
+
 		bool matches(const std::array<GLuint, 4>& color_targets, GLuint depth_stencil_target) const;
 		bool references_any(const std::vector<GLuint>& resources) const;
 

rpcs3/Emu/RSX/RSXThread.cpp

@@ -1773,10 +1773,19 @@ namespace rsx
 			m_graphics_state.set(rsx::rtt_config_valid);
 		}
 
-		if (!m_graphics_state.test(rsx::rtt_config_valid) && !layout.zeta_address)
+		const bool framebufferless = !m_graphics_state.test(rsx::rtt_config_valid) && !layout.zeta_address;
+		if (framebufferless)
 		{
-			rsx_log.warning("Framebuffer setup failed. Draw calls may have been lost");
-			return;
+			// No attachments. The hardware still rasterizes in this case and side effects such as ZPASS counters are still updated.
+			if (context != rsx::framebuffer_creation_context::context_draw ||
+				!backend_config.supports_framebufferless_rendering)
+			{
+				rsx_log.warning("Framebuffer setup failed. Draw calls may have been lost");
+				return;
+			}
+
+			// Context must be draw now...
+			// TODO: Also disable rendering if ZPASS stat counting is not enabled (side-effect)
 		}
 
 		// At least one attachment exists
@@ -1852,7 +1861,8 @@ namespace rsx
 			}
 		}
 
-		if (!really_changed)
+		// Framebufferless setups have no surface info to compare against. Always rebuild.
+		if (!really_changed && !framebufferless)
 		{
 			if (layout.zeta_address == m_depth_surface_info.address &&
 				layout.depth_format == m_depth_surface_info.depth_format &&

rpcs3/Emu/RSX/RSXThread.h

@@ -79,21 +79,22 @@ namespace rsx
 
 	struct backend_configuration
 	{
-		bool supports_multidraw;               // Draw call batching
-		bool supports_hw_a2c;                  // Alpha to coverage
-		bool supports_hw_a2c_1spp;             // Alpha to coverage at 1 sample per pixel
-		bool supports_hw_renormalization;      // Should be true on NV hardware which matches PS3 texture renormalization behaviour
-		bool supports_hw_msaa;                 // MSAA support
-		bool supports_hw_a2one;                // Alpha to one
-		bool supports_hw_conditional_render;   // Conditional render
-		bool supports_hw_instanced_rendering;  // Instanced draws
-		bool supports_passthrough_dma;         // DMA passthrough
-		bool supports_asynchronous_compute;    // Async compute
-		bool supports_host_gpu_labels;         // Advanced host synchronization
-		bool supports_normalized_barycentrics; // Basically all GPUs except NVIDIA have properly normalized barycentrics
-		bool supports_last_provoking_vertex;   // Flat shading using RSX's last-vertex convention
-		bool supports_extended_depth_range;    // Depth values above 1 can be stored and tested. Float depth targets are emulated otherwise.
-		bool supports_programmable_blending;   // Can handle programmable blending requests
+		bool supports_multidraw;                 // Draw call batching
+		bool supports_hw_a2c;                    // Alpha to coverage
+		bool supports_hw_a2c_1spp;               // Alpha to coverage at 1 sample per pixel
+		bool supports_hw_renormalization;        // Should be true on NV hardware which matches PS3 texture renormalization behaviour
+		bool supports_hw_msaa;                   // MSAA support
+		bool supports_hw_a2one;                  // Alpha to one
+		bool supports_hw_conditional_render;     // Conditional render
+		bool supports_hw_instanced_rendering;    // Instanced draws
+		bool supports_passthrough_dma;           // DMA passthrough
+		bool supports_asynchronous_compute;      // Async compute
+		bool supports_host_gpu_labels;           // Advanced host synchronization
+		bool supports_normalized_barycentrics;   // Basically all GPUs except NVIDIA have properly normalized barycentrics
+		bool supports_last_provoking_vertex;     // Flat shading using RSX's last-vertex convention
+		bool supports_programmable_blending;     // Can handle programmable blending requests
+		bool supports_extended_depth_range;      // Depth values above 1 can be stored and tested. Float depth targets are emulated otherwise.
+		bool supports_framebufferless_rendering; // Can rasterize without any attachments bound (e.g for occlusion queries)
 	};
 
 	struct desync_fifo_cmd_info