rsx/vk: Implement rendering using no attachments

rpcs3/Emu/RSX/RSXThread.cpp

@@ -1477,7 +1477,8 @@ namespace rsx
 		layout.width = rsx::method_registers.surface_clip_width();
 		layout.height = rsx::method_registers.surface_clip_height();
 
-		m_graphics_state.clear(rsx::rtt_config_contested | rsx::rtt_config_valid | rsx::rtt_config_no_attachments);
+		// NOTE: rtt_config_no_attachments is intentionally not reset.
+		m_graphics_state.clear(rsx::rtt_config_contested | rsx::rtt_config_valid);
 		m_current_framebuffer_context = context;
 
 		if (layout.width == 0 || layout.height == 0)
@@ -1801,7 +1802,22 @@ namespace rsx
 			}
 
 			// Context must be draw now...
-			m_graphics_state.set(rsx::rtt_config_no_attachments);
+		}
+
+		if (framebufferless != m_graphics_state.test(rsx::rtt_config_no_attachments))
+		{
+			// Programmable blending reads back from the color attachments and is disabled when there are none.
+			// Force the blend configuration to be re-evaluated when moving in or out of this state.
+			if (framebufferless)
+			{
+				m_graphics_state.set(rsx::rtt_config_no_attachments);
+			}
+			else
+			{
+				m_graphics_state.clear(rsx::rtt_config_no_attachments);
+			}
+
+			m_graphics_state.set(rsx::pipeline_config_dirty);
 		}
 
 		// At least one attachment exists
@@ -2145,8 +2161,9 @@ namespace rsx
 	{
 		u32 expected_ctrl = 0;
 
-		// Programmable blending
-		if (backend_config.supports_programmable_blending)
+		// Programmable blending. Requires color attachments to read from.
+		if (backend_config.supports_programmable_blending &&
+			!m_graphics_state.test(rsx::rtt_config_no_attachments))
 		{
 			expected_ctrl = current_fragment_program.ctrl & RSX_SHADER_CONTROL_PROGRAMMABLE_BLENDING;
 

rpcs3/Emu/RSX/VK/VKDraw.cpp

@@ -219,7 +219,9 @@ void VKGSRender::update_draw_state()
 		const auto polygon_offset_scale = rsx::method_registers.poly_offset_scale();
 		auto polygon_offset_bias = rsx::method_registers.poly_offset_bias();
 
-		if (m_draw_fbo->depth_format() == VK_FORMAT_D24_UNORM_S8_UINT && is_NVIDIA(vk::get_chip_family()))
+		if (!m_draw_fbo->attachments.empty() &&
+			m_draw_fbo->depth_format() == VK_FORMAT_D24_UNORM_S8_UINT &&
+			is_NVIDIA(vk::get_chip_family()))
 		{
 			// Empirically derived to be 0.5 * (2^24 - 1) for fixed type on Pascal. The same seems to apply for other NVIDIA GPUs.
 			// RSX seems to be using 2^24 - 1 instead making the biases twice as large when using fixed type Z-buffer on NVIDIA.
@@ -1261,7 +1263,6 @@ void VKGSRender::begin()
 	rsx::thread::begin();
 
 	if (skip_current_frame ||
-		swapchain_unavailable ||
 		cond_render_ctrl.disable_rendering())
 	{
 		return;
@@ -1279,7 +1280,7 @@ void VKGSRender::begin()
 
 void VKGSRender::end()
 {
-	if (skip_current_frame || !m_graphics_state.test(rsx::rtt_config_valid) || swapchain_unavailable || cond_render_ctrl.disable_rendering())
+	if (should_skip_draw())
 	{
 		execute_nop_draw();
 		rsx::thread::end();

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -677,6 +677,21 @@ VKGSRender::VKGSRender(utils::serial* ar) noexcept : GSRender(ar)
 		backend_config.supports_hw_a2one = m_device->get_alpha_to_one_support();
 	}
 
+	// Framebufferless rendering. Zero-attachment subpasses can only rasterize at the sample counts reported for them.
+	const VkSampleCountFlags required_sample_counts = (g_cfg.video.antialiasing_level == msaa_level::_auto)
+		? VK_SAMPLE_COUNT_1_BIT | VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT
+		: VK_SAMPLE_COUNT_1_BIT;
+
+	if (const auto supported_sample_counts = m_device->gpu().get_limits().framebufferNoAttachmentsSampleCounts & required_sample_counts;
+		supported_sample_counts == required_sample_counts)
+	{
+		backend_config.supports_framebufferless_rendering = true;
+	}
+	else
+	{
+		rsx_log.warning("Framebufferless rendering is not supported with the current MSAA configuration. Some occlusion queries may return incorrect results.");
+	}
+
 	// NOTE: On NVIDIA cards going back decades (including the PS3) there is a slight normalization inaccuracy in compressed formats.
 	// Confirmed in BLES01916 (The Evil Within) which uses RGB565 for some virtual texturing data.
 	backend_config.supports_hw_renormalization = vk::is_NVIDIA(vk::get_driver_vendor());
@@ -1299,7 +1314,7 @@ void VKGSRender::on_exit()
 
 void VKGSRender::clear_surface(u32 mask)
 {
-	if (skip_current_frame || swapchain_unavailable) return;
+	if (skip_current_frame) return;
 
 	// If stencil write mask is disabled, remove clear_stencil bit
 	if (!rsx::method_registers.stencil_mask()) mask &= ~RSX_GCM_CLEAR_STENCIL_BIT;
@@ -2709,7 +2724,19 @@ void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
 		std::iota(input_attachments.begin(), input_attachments.end(), 0);
 	}
 
-	m_current_renderpass_key = vk::get_renderpass_key(m_fbo_images, input_attachments);
+	if (m_graphics_state.test(rsx::rtt_config_no_attachments))
+	{
+		// Framebufferless rendering. The raster sample count must match what real surfaces would have used.
+		ensure(m_fbo_images.empty() && input_attachments.empty());
+
+		const u8 raster_samples = (g_cfg.video.antialiasing_level == msaa_level::_auto) ? samples : 1;
+		m_current_renderpass_key = vk::get_renderpass_key_no_attachments(raster_samples);
+	}
+	else
+	{
+		m_current_renderpass_key = vk::get_renderpass_key(m_fbo_images, input_attachments);
+	}
+
 	m_cached_renderpass = vk::get_renderpass(*m_device, m_current_renderpass_key);
 
 	// Search old framebuffers for this same configuration
@@ -2749,9 +2776,6 @@ void VKGSRender::renderctl(u32 request_code, void* args)
 
 bool VKGSRender::scaled_image_from_memory(const rsx::blit_src_info& src, const rsx::blit_dst_info& dst, bool interpolate)
 {
-	if (swapchain_unavailable)
-		return false;
-
 	if (m_texture_cache.blit(src, dst, interpolate, m_rtts, *m_current_command_buffer))
 	{
 		m_samplers_dirty.store(true);