Add optional driver workaround to RenderingDevice for Adreno 6XX.

Co-authored-by: Clay John <[email protected]>

drivers/d3d12/rendering_context_driver_d3d12.cpp

@@ -173,6 +173,7 @@ Error RenderingContextDriverD3D12::_initialize_devices() {
 		Device &device = driver_devices[i];
 		device.name = desc.Description;
 		device.vendor = Vendor(desc.VendorId);
+		device.workarounds = Workarounds();
 
 		if (desc.Flags & DXGI_ADAPTER_FLAG_SOFTWARE) {
 			device.type = DEVICE_TYPE_CPU;

drivers/vulkan/rendering_context_driver_vulkan.cpp

@@ -502,6 +502,9 @@ Error RenderingContextDriverVulkan::_initialize_devices() {
 		driver_device.name = String::utf8(props.deviceName);
 		driver_device.vendor = Vendor(props.vendorID);
 		driver_device.type = DeviceType(props.deviceType);
+		driver_device.workarounds = Workarounds();
+
+		_check_driver_workarounds(props, driver_device);
 
 		uint32_t queue_family_properties_count = 0;
 		vkGetPhysicalDeviceQueueFamilyProperties(physical_devices[i], &queue_family_properties_count, nullptr);
@@ -515,6 +518,31 @@ Error RenderingContextDriverVulkan::_initialize_devices() {
 	return OK;
 }
 
+void RenderingContextDriverVulkan::_check_driver_workarounds(const VkPhysicalDeviceProperties &p_device_properties, Device &r_device) {
+	// Workaround for the Adreno 6XX family of devices.
+	//
+	// There's a known issue with the Vulkan driver in this family of devices where it'll crash if a dynamic state for drawing is
+	// used in a command buffer before a dispatch call is issued. As both dynamic scissor and viewport are basic requirements for
+	// the engine to not bake this state into the PSO, the only known way to fix this issue is to reset the command buffer entirely.
+	//
+	// As the render graph has no built in limitations of whether it'll issue compute work before anything needs to draw on the
+	// frame, and there's no guarantee that compute work will never be dependent on rasterization in the future, this workaround
+	// will end recording on the current command buffer any time a compute list is encountered after a draw list was executed.
+	// A new command buffer will be created afterwards and the appropriate synchronization primitives will be inserted.
+	//
+	// Executing this workaround has the added cost of synchronization between all the command buffers that are created as well as
+	// all the individual submissions. This performance hit is accepted for the sake of being able to support these devices without
+	// limiting the design of the renderer.
+	//
+	// This bug was fixed in driver version 512.503.0, so we only enabled it on devices older than this.
+	//
+	r_device.workarounds.avoid_compute_after_draw =
+			r_device.vendor == VENDOR_QUALCOMM &&
+			p_device_properties.deviceID >= 0x6000000 && // Adreno 6xx
+			p_device_properties.driverVersion < VK_MAKE_VERSION(512, 503, 0) &&
+			r_device.name.find("Turnip") < 0;
+}
+
 bool RenderingContextDriverVulkan::_use_validation_layers() const {
 	return Engine::get_singleton()->is_validation_layers_enabled();
 }

drivers/vulkan/rendering_context_driver_vulkan.h

@@ -105,6 +105,7 @@ private:
 	Error _initialize_instance_extensions();
 	Error _initialize_instance();
 	Error _initialize_devices();
+	void _check_driver_workarounds(const VkPhysicalDeviceProperties &p_device_properties, Device &r_device);
 
 	// Static callbacks.
 	static VKAPI_ATTR VkBool32 VKAPI_CALL _debug_messenger_callback(VkDebugUtilsMessageSeverityFlagBitsEXT p_message_severity, VkDebugUtilsMessageTypeFlagsEXT p_message_type, const VkDebugUtilsMessengerCallbackDataEXT *p_callback_data, void *p_user_data);

servers/rendering/rendering_context_driver.h

@@ -73,10 +73,15 @@ public:
 		DEVICE_TYPE_MAX = 0x5
 	};
 
+	struct Workarounds {
+		bool avoid_compute_after_draw = false;
+	};
+
 	struct Device {
 		String name = "Unknown";
 		Vendor vendor = VENDOR_UNKNOWN;
 		DeviceType type = DEVICE_TYPE_OTHER;
+		Workarounds workarounds;
 	};
 
 	virtual ~RenderingContextDriver();

servers/rendering/rendering_device.cpp

@@ -4877,25 +4877,78 @@ void RenderingDevice::_end_frame() {
 		ERR_PRINT("Found open compute list at the end of the frame, this should never happen (further compute will likely not work).");
 	}
 
-	draw_graph.end(frames[frame].draw_command_buffer, RENDER_GRAPH_REORDER, RENDER_GRAPH_FULL_BARRIERS);
 	driver->command_buffer_end(frames[frame].setup_command_buffer);
-	driver->command_buffer_end(frames[frame].draw_command_buffer);
+
+	// The command buffer must be copied into a stack variable as the driver workarounds can change the command buffer in use.
+	RDD::CommandBufferID command_buffer = frames[frame].draw_command_buffer;
+	draw_graph.end(RENDER_GRAPH_REORDER, RENDER_GRAPH_FULL_BARRIERS, command_buffer, frames[frame].command_buffer_pool);
+	driver->command_buffer_end(command_buffer);
 	driver->end_segment();
 }
 
 void RenderingDevice::_execute_frame(bool p_present) {
+	// Check whether this frame should present the swap chains and in which queue.
 	const bool frame_can_present = p_present && !frames[frame].swap_chains_to_present.is_empty();
 	const bool separate_present_queue = main_queue != present_queue;
-	const VectorView<RDD::SemaphoreID> execute_draw_semaphore = frame_can_present && separate_present_queue ? frames[frame].draw_semaphore : VectorView<RDD::SemaphoreID>();
-	const VectorView<RDD::SwapChainID> execute_draw_swap_chains = frame_can_present && !separate_present_queue ? frames[frame].swap_chains_to_present : VectorView<RDD::SwapChainID>();
+	thread_local LocalVector<RDD::SwapChainID> swap_chains;
+	swap_chains.clear();
+
+	// Execute the setup command buffer.
 	driver->command_queue_execute_and_present(main_queue, {}, frames[frame].setup_command_buffer, frames[frame].setup_semaphore, {}, {});
-	driver->command_queue_execute_and_present(main_queue, frames[frame].setup_semaphore, frames[frame].draw_command_buffer, execute_draw_semaphore, frames[frame].draw_fence, execute_draw_swap_chains);
+
+	// Execute command buffers and use semaphores to wait on the execution of the previous one. Normally there's only one command buffer,
+	// but driver workarounds can force situations where there'll be more.
+	uint32_t command_buffer_count = 1;
+	RDG::CommandBufferPool &buffer_pool = frames[frame].command_buffer_pool;
+	if (buffer_pool.buffers_used > 0) {
+		command_buffer_count += buffer_pool.buffers_used;
+		buffer_pool.buffers_used = 0;
+	}
+
+	RDD::SemaphoreID wait_semaphore = frames[frame].setup_semaphore;
+	for (uint32_t i = 0; i < command_buffer_count; i++) {
+		RDD::CommandBufferID command_buffer;
+		RDD::SemaphoreID signal_semaphore;
+		RDD::FenceID signal_fence;
+		if (i > 0) {
+			command_buffer = buffer_pool.buffers[i - 1];
+			signal_semaphore = buffer_pool.semaphores[i - 1];
+		} else {
+			command_buffer = frames[frame].draw_command_buffer;
+			signal_semaphore = frames[frame].draw_semaphore;
+		}
+
+		bool signal_semaphore_valid;
+		if (i == (command_buffer_count - 1)) {
+			// This is the last command buffer, it should signal the fence.
+			signal_fence = frames[frame].draw_fence;
+			signal_semaphore_valid = false;
+
+			if (frame_can_present && separate_present_queue) {
+				// The semaphore is required if the frame can be presented and a separate present queue is used.
+				signal_semaphore_valid = true;
+			} else if (frame_can_present) {
+				// Just present the swap chains as part of the last command execution.
+				swap_chains = frames[frame].swap_chains_to_present;
+			}
+		} else {
+			// Semaphores always need to be signaled if it's not the last command buffer.
+			signal_semaphore_valid = true;
+		}
+
+		driver->command_queue_execute_and_present(main_queue, wait_semaphore, command_buffer, signal_semaphore_valid ? signal_semaphore : VectorView<RDD::SemaphoreID>(), signal_fence, swap_chains);
+
+		// Make the next command buffer wait on the semaphore signaled by this one.
+		wait_semaphore = signal_semaphore;
+	}
+
+	// Indicate the fence has been signaled so the next time the frame's contents need to be used, the CPU needs to wait on the work to be completed.
 	frames[frame].draw_fence_signaled = true;
 
 	if (frame_can_present) {
 		if (separate_present_queue) {
 			// Issue the presentation separately if the presentation queue is different from the main queue.
-			driver->command_queue_execute_and_present(present_queue, frames[frame].draw_semaphore, {}, {}, {}, frames[frame].swap_chains_to_present);
+			driver->command_queue_execute_and_present(present_queue, wait_semaphore, {}, {}, {}, frames[frame].swap_chains_to_present);
 		}
 
 		frames[frame].swap_chains_to_present.clear();
@@ -5044,6 +5097,9 @@ Error RenderingDevice::initialize(RenderingContextDriver *p_context, DisplayServ
 		frames[i].timestamp_cpu_result_values.resize(max_timestamp_query_elements);
 		frames[i].timestamp_result_values.resize(max_timestamp_query_elements);
 		frames[i].timestamp_result_count = 0;
+
+		// Assign the main queue family and command pool to the command buffer pool.
+		frames[i].command_buffer_pool.pool = frames[i].command_pool;
 	}
 
 	// Start from frame count, so everything else is immediately old.
@@ -5055,7 +5111,7 @@ Error RenderingDevice::initialize(RenderingContextDriver *p_context, DisplayServ
 	driver->command_buffer_begin(frames[0].draw_command_buffer);
 
 	// Create draw graph and start it initialized as well.
-	draw_graph.initialize(driver, frames.size(), main_queue_family, SECONDARY_COMMAND_BUFFERS_PER_FRAME);
+	draw_graph.initialize(driver, device, frames.size(), main_queue_family, SECONDARY_COMMAND_BUFFERS_PER_FRAME);
 	draw_graph.begin();
 
 	for (uint32_t i = 0; i < frames.size(); i++) {
@@ -5388,6 +5444,11 @@ void RenderingDevice::finalize() {
 		driver->semaphore_free(frames[i].setup_semaphore);
 		driver->semaphore_free(frames[i].draw_semaphore);
 		driver->fence_free(frames[i].draw_fence);
+
+		RDG::CommandBufferPool &buffer_pool = frames[i].command_buffer_pool;
+		for (uint32_t j = 0; j < buffer_pool.buffers.size(); j++) {
+			driver->semaphore_free(buffer_pool.semaphores[j]);
+		}
 	}
 
 	if (pipeline_cache_enabled) {

servers/rendering/rendering_device.h

@@ -1261,6 +1261,9 @@ private:
 		// Swap chains prepared for drawing during the frame that must be presented.
 		LocalVector<RDD::SwapChainID> swap_chains_to_present;
 
+		// Extra command buffer pool used for driver workarounds.
+		RDG::CommandBufferPool command_buffer_pool;
+
 		struct Timestamp {
 			String description;
 			uint64_t value = 0;

servers/rendering/rendering_device_graph.cpp

@@ -753,71 +753,96 @@ void RenderingDeviceGraph::_wait_for_secondary_command_buffer_tasks() {
 	}
 }
 
-void RenderingDeviceGraph::_run_render_commands(RDD::CommandBufferID p_command_buffer, int32_t p_level, const RecordedCommandSort *p_sorted_commands, uint32_t p_sorted_commands_count, int32_t &r_current_label_index, int32_t &r_current_label_level) {
+void RenderingDeviceGraph::_run_render_commands(int32_t p_level, const RecordedCommandSort *p_sorted_commands, uint32_t p_sorted_commands_count, RDD::CommandBufferID &r_command_buffer, CommandBufferPool &r_command_buffer_pool, int32_t &r_current_label_index, int32_t &r_current_label_level) {
 	for (uint32_t i = 0; i < p_sorted_commands_count; i++) {
 		const uint32_t command_index = p_sorted_commands[i].index;
 		const uint32_t command_data_offset = command_data_offsets[command_index];
 		const RecordedCommand *command = reinterpret_cast<RecordedCommand *>(&command_data[command_data_offset]);
-		_run_label_command_change(p_command_buffer, command->label_index, p_level, false, true, &p_sorted_commands[i], p_sorted_commands_count - i, r_current_label_index, r_current_label_level);
+		_run_label_command_change(r_command_buffer, command->label_index, p_level, false, true, &p_sorted_commands[i], p_sorted_commands_count - i, r_current_label_index, r_current_label_level);
 
 		switch (command->type) {
 			case RecordedCommand::TYPE_BUFFER_CLEAR: {
 				const RecordedBufferClearCommand *buffer_clear_command = reinterpret_cast<const RecordedBufferClearCommand *>(command);
-				driver->command_clear_buffer(p_command_buffer, buffer_clear_command->buffer, buffer_clear_command->offset, buffer_clear_command->size);
+				driver->command_clear_buffer(r_command_buffer, buffer_clear_command->buffer, buffer_clear_command->offset, buffer_clear_command->size);
 			} break;
 			case RecordedCommand::TYPE_BUFFER_COPY: {
 				const RecordedBufferCopyCommand *buffer_copy_command = reinterpret_cast<const RecordedBufferCopyCommand *>(command);
-				driver->command_copy_buffer(p_command_buffer, buffer_copy_command->source, buffer_copy_command->destination, buffer_copy_command->region);
+				driver->command_copy_buffer(r_command_buffer, buffer_copy_command->source, buffer_copy_command->destination, buffer_copy_command->region);
 			} break;
 			case RecordedCommand::TYPE_BUFFER_GET_DATA: {
 				const RecordedBufferGetDataCommand *buffer_get_data_command = reinterpret_cast<const RecordedBufferGetDataCommand *>(command);
-				driver->command_copy_buffer(p_command_buffer, buffer_get_data_command->source, buffer_get_data_command->destination, buffer_get_data_command->region);
+				driver->command_copy_buffer(r_command_buffer, buffer_get_data_command->source, buffer_get_data_command->destination, buffer_get_data_command->region);
 			} break;
 			case RecordedCommand::TYPE_BUFFER_UPDATE: {
 				const RecordedBufferUpdateCommand *buffer_update_command = reinterpret_cast<const RecordedBufferUpdateCommand *>(command);
 				const RecordedBufferCopy *command_buffer_copies = buffer_update_command->buffer_copies();
 				for (uint32_t j = 0; j < buffer_update_command->buffer_copies_count; j++) {
-					driver->command_copy_buffer(p_command_buffer, command_buffer_copies[j].source, buffer_update_command->destination, command_buffer_copies[j].region);
+					driver->command_copy_buffer(r_command_buffer, command_buffer_copies[j].source, buffer_update_command->destination, command_buffer_copies[j].region);
 				}
 			} break;
 			case RecordedCommand::TYPE_COMPUTE_LIST: {
+				if (device.workarounds.avoid_compute_after_draw && workarounds_state.draw_list_found) {
+					// Avoid compute after draw workaround. Refer to the comment that enables this in the Vulkan driver for more information.
+					workarounds_state.draw_list_found = false;
+
+					// Create or reuse a command buffer and finish recording the current one.
+					driver->command_buffer_end(r_command_buffer);
+
+					while (r_command_buffer_pool.buffers_used >= r_command_buffer_pool.buffers.size()) {
+						RDD::CommandBufferID command_buffer = driver->command_buffer_create(r_command_buffer_pool.pool);
+						RDD::SemaphoreID command_semaphore = driver->semaphore_create();
+						r_command_buffer_pool.buffers.push_back(command_buffer);
+						r_command_buffer_pool.semaphores.push_back(command_semaphore);
+					}
+
+					// Start recording on the next usable command buffer from the pool.
+					uint32_t command_buffer_index = r_command_buffer_pool.buffers_used++;
+					r_command_buffer = r_command_buffer_pool.buffers[command_buffer_index];
+					driver->command_buffer_begin(r_command_buffer);
+				}
+
 				const RecordedComputeListCommand *compute_list_command = reinterpret_cast<const RecordedComputeListCommand *>(command);
-				_run_compute_list_command(p_command_buffer, compute_list_command->instruction_data(), compute_list_command->instruction_data_size);
+				_run_compute_list_command(r_command_buffer, compute_list_command->instruction_data(), compute_list_command->instruction_data_size);
 			} break;
 			case RecordedCommand::TYPE_DRAW_LIST: {
+				if (device.workarounds.avoid_compute_after_draw) {
+					// Indicate that a draw list was encountered for the workaround.
+					workarounds_state.draw_list_found = true;
+				}
+
 				const RecordedDrawListCommand *draw_list_command = reinterpret_cast<const RecordedDrawListCommand *>(command);
 				const VectorView clear_values(draw_list_command->clear_values(), draw_list_command->clear_values_count);
-				driver->command_begin_render_pass(p_command_buffer, draw_list_command->render_pass, draw_list_command->framebuffer, draw_list_command->command_buffer_type, draw_list_command->region, clear_values);
-				_run_draw_list_command(p_command_buffer, draw_list_command->instruction_data(), draw_list_command->instruction_data_size);
-				driver->command_end_render_pass(p_command_buffer);
+				driver->command_begin_render_pass(r_command_buffer, draw_list_command->render_pass, draw_list_command->framebuffer, draw_list_command->command_buffer_type, draw_list_command->region, clear_values);
+				_run_draw_list_command(r_command_buffer, draw_list_command->instruction_data(), draw_list_command->instruction_data_size);
+				driver->command_end_render_pass(r_command_buffer);
 			} break;
 			case RecordedCommand::TYPE_TEXTURE_CLEAR: {
 				const RecordedTextureClearCommand *texture_clear_command = reinterpret_cast<const RecordedTextureClearCommand *>(command);
-				driver->command_clear_color_texture(p_command_buffer, texture_clear_command->texture, RDD::TEXTURE_LAYOUT_TRANSFER_DST_OPTIMAL, texture_clear_command->color, texture_clear_command->range);
+				driver->command_clear_color_texture(r_command_buffer, texture_clear_command->texture, RDD::TEXTURE_LAYOUT_TRANSFER_DST_OPTIMAL, texture_clear_command->color, texture_clear_command->range);
 			} break;
 			case RecordedCommand::TYPE_TEXTURE_COPY: {
 				const RecordedTextureCopyCommand *texture_copy_command = reinterpret_cast<const RecordedTextureCopyCommand *>(command);
-				driver->command_copy_texture(p_command_buffer, texture_copy_command->from_texture, RDD::TEXTURE_LAYOUT_TRANSFER_SRC_OPTIMAL, texture_copy_command->to_texture, RDD::TEXTURE_LAYOUT_TRANSFER_DST_OPTIMAL, texture_copy_command->region);
+				driver->command_copy_texture(r_command_buffer, texture_copy_command->from_texture, RDD::TEXTURE_LAYOUT_TRANSFER_SRC_OPTIMAL, texture_copy_command->to_texture, RDD::TEXTURE_LAYOUT_TRANSFER_DST_OPTIMAL, texture_copy_command->region);
 			} break;
 			case RecordedCommand::TYPE_TEXTURE_GET_DATA: {
 				const RecordedTextureGetDataCommand *texture_get_data_command = reinterpret_cast<const RecordedTextureGetDataCommand *>(command);
 				const VectorView<RDD::BufferTextureCopyRegion> command_buffer_texture_copy_regions_view(texture_get_data_command->buffer_texture_copy_regions(), texture_get_data_command->buffer_texture_copy_regions_count);
-				driver->command_copy_texture_to_buffer(p_command_buffer, texture_get_data_command->from_texture, RDD::TEXTURE_LAYOUT_TRANSFER_SRC_OPTIMAL, texture_get_data_command->to_buffer, command_buffer_texture_copy_regions_view);
+				driver->command_copy_texture_to_buffer(r_command_buffer, texture_get_data_command->from_texture, RDD::TEXTURE_LAYOUT_TRANSFER_SRC_OPTIMAL, texture_get_data_command->to_buffer, command_buffer_texture_copy_regions_view);
 			} break;
 			case RecordedCommand::TYPE_TEXTURE_RESOLVE: {
 				const RecordedTextureResolveCommand *texture_resolve_command = reinterpret_cast<const RecordedTextureResolveCommand *>(command);
-				driver->command_resolve_texture(p_command_buffer, texture_resolve_command->from_texture, RDD::TEXTURE_LAYOUT_TRANSFER_SRC_OPTIMAL, texture_resolve_command->src_layer, texture_resolve_command->src_mipmap, texture_resolve_command->to_texture, RDD::TEXTURE_LAYOUT_TRANSFER_DST_OPTIMAL, texture_resolve_command->dst_layer, texture_resolve_command->dst_mipmap);
+				driver->command_resolve_texture(r_command_buffer, texture_resolve_command->from_texture, RDD::TEXTURE_LAYOUT_TRANSFER_SRC_OPTIMAL, texture_resolve_command->src_layer, texture_resolve_command->src_mipmap, texture_resolve_command->to_texture, RDD::TEXTURE_LAYOUT_TRANSFER_DST_OPTIMAL, texture_resolve_command->dst_layer, texture_resolve_command->dst_mipmap);
 			} break;
 			case RecordedCommand::TYPE_TEXTURE_UPDATE: {
 				const RecordedTextureUpdateCommand *texture_update_command = reinterpret_cast<const RecordedTextureUpdateCommand *>(command);
 				const RecordedBufferToTextureCopy *command_buffer_to_texture_copies = texture_update_command->buffer_to_texture_copies();
 				for (uint32_t j = 0; j < texture_update_command->buffer_to_texture_copies_count; j++) {
-					driver->command_copy_buffer_to_texture(p_command_buffer, command_buffer_to_texture_copies[j].from_buffer, texture_update_command->to_texture, RDD::TEXTURE_LAYOUT_TRANSFER_DST_OPTIMAL, command_buffer_to_texture_copies[j].region);
+					driver->command_copy_buffer_to_texture(r_command_buffer, command_buffer_to_texture_copies[j].from_buffer, texture_update_command->to_texture, RDD::TEXTURE_LAYOUT_TRANSFER_DST_OPTIMAL, command_buffer_to_texture_copies[j].region);
 				}
 			} break;
 			case RecordedCommand::TYPE_CAPTURE_TIMESTAMP: {
 				const RecordedCaptureTimestampCommand *texture_capture_timestamp_command = reinterpret_cast<const RecordedCaptureTimestampCommand *>(command);
-				driver->command_timestamp_write(p_command_buffer, texture_capture_timestamp_command->pool, texture_capture_timestamp_command->index);
+				driver->command_timestamp_write(r_command_buffer, texture_capture_timestamp_command->pool, texture_capture_timestamp_command->index);
 			} break;
 			default: {
 				DEV_ASSERT(false && "Unknown recorded command type.");
@@ -1229,8 +1254,9 @@ void RenderingDeviceGraph::_print_compute_list(const uint8_t *p_instruction_data
 	}
 }
 
-void RenderingDeviceGraph::initialize(RDD *p_driver, uint32_t p_frame_count, RDD::CommandQueueFamilyID p_secondary_command_queue_family, uint32_t p_secondary_command_buffers_per_frame) {
+void RenderingDeviceGraph::initialize(RDD *p_driver, RenderingContextDriver::Device p_device, uint32_t p_frame_count, RDD::CommandQueueFamilyID p_secondary_command_queue_family, uint32_t p_secondary_command_buffers_per_frame) {
 	driver = p_driver;
+	device = p_device;
 	frames.resize(p_frame_count);
 
 	for (uint32_t i = 0; i < p_frame_count; i++) {
@@ -1805,7 +1831,7 @@ void RenderingDeviceGraph::end_label() {
 	command_label_index = -1;
 }
 
-void RenderingDeviceGraph::end(RDD::CommandBufferID p_command_buffer, bool p_reorder_commands, bool p_full_barriers) {
+void RenderingDeviceGraph::end(bool p_reorder_commands, bool p_full_barriers, RDD::CommandBufferID &r_command_buffer, CommandBufferPool &r_command_buffer_pool) {
 	if (command_count == 0) {
 		// No commands have been logged, do nothing.
 		return;
@@ -1919,7 +1945,12 @@ void RenderingDeviceGraph::end(RDD::CommandBufferID p_command_buffer, bool p_reo
 	if (command_count > 0) {
 		int32_t current_label_index = -1;
 		int32_t current_label_level = -1;
-		_run_label_command_change(p_command_buffer, -1, -1, true, true, nullptr, 0, current_label_index, current_label_level);
+		_run_label_command_change(r_command_buffer, -1, -1, true, true, nullptr, 0, current_label_index, current_label_level);
+
+		if (device.workarounds.avoid_compute_after_draw) {
+			// Reset the state of the workaround.
+			workarounds_state.draw_list_found = false;
+		}
 
 		if (p_reorder_commands) {
 #if PRINT_RENDER_GRAPH
@@ -1946,8 +1977,8 @@ void RenderingDeviceGraph::end(RDD::CommandBufferID p_command_buffer, bool p_reo
 					RecordedCommandSort *level_command_ptr = &commands_sorted[current_level_start];
 					uint32_t level_command_count = i - current_level_start;
 					_boost_priority_for_render_commands(level_command_ptr, level_command_count, boosted_priority);
-					_group_barriers_for_render_commands(p_command_buffer, level_command_ptr, level_command_count, p_full_barriers);
-					_run_render_commands(p_command_buffer, current_level, level_command_ptr, level_command_count, current_label_index, current_label_level);
+					_group_barriers_for_render_commands(r_command_buffer, level_command_ptr, level_command_count, p_full_barriers);
+					_run_render_commands(current_level, level_command_ptr, level_command_count, r_command_buffer, r_command_buffer_pool, current_label_index, current_label_level);
 					current_level = commands_sorted[i].level;
 					current_level_start = i;
 				}
@@ -1956,20 +1987,20 @@ void RenderingDeviceGraph::end(RDD::CommandBufferID p_command_buffer, bool p_reo
 			RecordedCommandSort *level_command_ptr = &commands_sorted[current_level_start];
 			uint32_t level_command_count = command_count - current_level_start;
 			_boost_priority_for_render_commands(level_command_ptr, level_command_count, boosted_priority);
-			_group_barriers_for_render_commands(p_command_buffer, level_command_ptr, level_command_count, p_full_barriers);
-			_run_render_commands(p_command_buffer, current_level, level_command_ptr, level_command_count, current_label_index, current_label_level);
+			_group_barriers_for_render_commands(r_command_buffer, level_command_ptr, level_command_count, p_full_barriers);
+			_run_render_commands(current_level, level_command_ptr, level_command_count, r_command_buffer, r_command_buffer_pool, current_label_index, current_label_level);
 
 #if PRINT_RENDER_GRAPH
 			print_line("COMMANDS", command_count, "LEVELS", current_level + 1);
 #endif
 		} else {
 			for (uint32_t i = 0; i < command_count; i++) {
-				_group_barriers_for_render_commands(p_command_buffer, &commands_sorted[i], 1, p_full_barriers);
-				_run_render_commands(p_command_buffer, i, &commands_sorted[i], 1, current_label_index, current_label_level);
+				_group_barriers_for_render_commands(r_command_buffer, &commands_sorted[i], 1, p_full_barriers);
+				_run_render_commands(i, &commands_sorted[i], 1, r_command_buffer, r_command_buffer_pool, current_label_index, current_label_level);
 			}
 		}
 
-		_run_label_command_change(p_command_buffer, -1, -1, true, false, nullptr, 0, current_label_index, current_label_level);
+		_run_label_command_change(r_command_buffer, -1, -1, true, false, nullptr, 0, current_label_index, current_label_level);
 
 #if PRINT_COMMAND_RECORDING
 		print_line(vformat("Recorded %d commands", command_count));

servers/rendering/rendering_device_graph.h

@@ -184,6 +184,20 @@ public:
 		}
 	};
 
+	struct CommandBufferPool {
+		// Provided by RenderingDevice.
+		RDD::CommandPoolID pool;
+
+		// Created internally by RenderingDeviceGraph.
+		LocalVector<RDD::CommandBufferID> buffers;
+		LocalVector<RDD::SemaphoreID> semaphores;
+		uint32_t buffers_used = 0;
+	};
+
+	struct WorkaroundsState {
+		bool draw_list_found = false;
+	};
+
 private:
 	struct InstructionList {
 		LocalVector<uint8_t> data;
@@ -560,6 +574,7 @@ private:
 	};
 
 	RDD *driver = nullptr;
+	RenderingContextDriver::Device device;
 	int64_t tracking_frame = 0;
 	LocalVector<uint8_t> command_data;
 	LocalVector<uint32_t> command_data_offsets;
@@ -582,6 +597,7 @@ private:
 	bool command_synchronization_pending = false;
 	BarrierGroup barrier_group;
 	bool driver_honors_barriers = false;
+	WorkaroundsState workarounds_state;
 	TightLocalVector<Frame> frames;
 	uint32_t frame = 0;
 
@@ -608,7 +624,7 @@ private:
 	void _run_draw_list_command(RDD::CommandBufferID p_command_buffer, const uint8_t *p_instruction_data, uint32_t p_instruction_data_size);
 	void _run_secondary_command_buffer_task(const SecondaryCommandBuffer *p_secondary);
 	void _wait_for_secondary_command_buffer_tasks();
-	void _run_render_commands(RDD::CommandBufferID p_command_buffer, int32_t p_level, const RecordedCommandSort *p_sorted_commands, uint32_t p_sorted_commands_count, int32_t &r_current_label_index, int32_t &r_current_label_level);
+	void _run_render_commands(int32_t p_level, const RecordedCommandSort *p_sorted_commands, uint32_t p_sorted_commands_count, RDD::CommandBufferID &r_command_buffer, CommandBufferPool &r_command_buffer_pool, int32_t &r_current_label_index, int32_t &r_current_label_level);
 	void _run_label_command_change(RDD::CommandBufferID p_command_buffer, int32_t p_new_label_index, int32_t p_new_level, bool p_ignore_previous_value, bool p_use_label_for_empty, const RecordedCommandSort *p_sorted_commands, uint32_t p_sorted_commands_count, int32_t &r_current_label_index, int32_t &r_current_label_level);
 	void _boost_priority_for_render_commands(RecordedCommandSort *p_sorted_commands, uint32_t p_sorted_commands_count, uint32_t &r_boosted_priority);
 	void _group_barriers_for_render_commands(RDD::CommandBufferID p_command_buffer, const RecordedCommandSort *p_sorted_commands, uint32_t p_sorted_commands_count, bool p_full_memory_barrier);
@@ -619,7 +635,7 @@ private:
 public:
 	RenderingDeviceGraph();
 	~RenderingDeviceGraph();
-	void initialize(RDD *p_driver, uint32_t p_frame_count, RDD::CommandQueueFamilyID p_secondary_command_queue_family, uint32_t p_secondary_command_buffers_per_frame);
+	void initialize(RDD *p_driver, RenderingContextDriver::Device p_device, uint32_t p_frame_count, RDD::CommandQueueFamilyID p_secondary_command_queue_family, uint32_t p_secondary_command_buffers_per_frame);
 	void finalize();
 	void begin();
 	void add_buffer_clear(RDD::BufferID p_dst, ResourceTracker *p_dst_tracker, uint32_t p_offset, uint32_t p_size);
@@ -664,7 +680,7 @@ public:
 	void add_synchronization();
 	void begin_label(const String &p_label_name, const Color &p_color);
 	void end_label();
-	void end(RDD::CommandBufferID p_command_buffer, bool p_reorder_commands, bool p_full_barriers);
+	void end(bool p_reorder_commands, bool p_full_barriers, RDD::CommandBufferID &r_command_buffer, CommandBufferPool &r_command_buffer_pool);
 	static ResourceTracker *resource_tracker_create();
 	static void resource_tracker_free(ResourceTracker *tracker);
 };