Apply some small fixes/enhancements to the Vulkan RD

- Initialize queue indices to values meaning 'unset'
- Remove unused parameters & members
- Make texture update access flags consistent with texture copy
- Fix style and pass type of some parameters
- Synchronize setup-draw in flush with a semaphore
- Add no current list validation to draw_list_begin_splits()
- Update texture usage flags on destination of copy
- Fix misuse of Vulkan flag

4 files changed, 40 insertions, 30 deletions

diff --git a/drivers/vulkan/rendering_device_vulkan.cpp b/drivers/vulkan/rendering_device_vulkan.cpp
index f9189aacf7..4c97ef87fc 100644
--- a/drivers/vulkan/rendering_device_vulkan.cpp
+++ b/drivers/vulkan/rendering_device_vulkan.cpp
@@ -1400,7 +1400,7 @@ Error RenderingDeviceVulkan::_insert_staging_block() {
 	return OK;
 }
 
-Error RenderingDeviceVulkan::_staging_buffer_allocate(uint32_t p_amount, uint32_t p_required_align, uint32_t &r_alloc_offset, uint32_t &r_alloc_size, bool p_can_segment, bool p_on_draw_command_buffer) {
+Error RenderingDeviceVulkan::_staging_buffer_allocate(uint32_t p_amount, uint32_t p_required_align, uint32_t &r_alloc_offset, uint32_t &r_alloc_size, bool p_can_segment) {
 	//determine a block to use
 
 	r_alloc_size = p_amount;
@@ -1542,7 +1542,7 @@ Error RenderingDeviceVulkan::_buffer_update(Buffer *p_buffer, size_t p_offset, c
 		uint32_t block_write_offset;
 		uint32_t block_write_amount;
 
-		Error err = _staging_buffer_allocate(MIN(to_submit, staging_buffer_block_size), p_required_align, block_write_offset, block_write_amount, p_use_draw_command_buffer);
+		Error err = _staging_buffer_allocate(MIN(to_submit, staging_buffer_block_size), p_required_align, block_write_offset, block_write_amount);
 		if (err) {
 			return err;
 		}
@@ -2474,7 +2474,7 @@ Error RenderingDeviceVulkan::_texture_update(RID p_texture, uint32_t p_layer, co
 					to_allocate >>= get_compressed_image_format_pixel_rshift(texture->format);
 
 					uint32_t alloc_offset, alloc_size;
-					Error err = _staging_buffer_allocate(to_allocate, required_align, alloc_offset, alloc_size, false, !p_use_setup_queue);
+					Error err = _staging_buffer_allocate(to_allocate, required_align, alloc_offset, alloc_size, false);
 					ERR_FAIL_COND_V(err, ERR_CANT_CREATE);
 
 					uint8_t *write_ptr;
@@ -2572,11 +2572,11 @@ Error RenderingDeviceVulkan::_texture_update(RID p_texture, uint32_t p_layer, co
 		uint32_t access_flags = 0;
 		if (p_post_barrier & BARRIER_MASK_COMPUTE) {
 			barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
-			access_flags |= VK_ACCESS_SHADER_READ_BIT;
+			access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
 		}
 		if (p_post_barrier & BARRIER_MASK_RASTER) {
 			barrier_flags |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
-			access_flags |= VK_ACCESS_SHADER_READ_BIT;
+			access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
 		}
 		if (p_post_barrier & BARRIER_MASK_TRANSFER) {
 			barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT;
@@ -2990,7 +2990,7 @@ Error RenderingDeviceVulkan::texture_copy(RID p_from_texture, RID p_to_texture,
 			image_memory_barrier.subresourceRange.baseArrayLayer = p_src_layer;
 			image_memory_barrier.subresourceRange.layerCount = 1;
 
-			vkCmdPipelineBarrier(command_buffer, VK_ACCESS_TRANSFER_WRITE_BIT, barrier_flags, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+			vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_TRANSFER_BIT, barrier_flags, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
 		}
 
 		{ //make dst readable
@@ -3016,6 +3016,13 @@ Error RenderingDeviceVulkan::texture_copy(RID p_from_texture, RID p_to_texture,
 		}
 	}
 
+	if (dst_tex->used_in_frame != frames_drawn) {
+		dst_tex->used_in_raster = false;
+		dst_tex->used_in_compute = false;
+		dst_tex->used_in_frame = frames_drawn;
+	}
+	dst_tex->used_in_transfer = true;
+
 	return OK;
 }
 
@@ -4771,6 +4778,7 @@ Vector<uint8_t> RenderingDeviceVulkan::shader_compile_binary_from_spirv(const Ve
 								//just append stage mask and return
 								uniform_info.write[set].write[k].stages |= 1 << stage;
 								exists = true;
+								break;
 							}
 						}
 
@@ -7202,6 +7210,9 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin(RID p_framebu
 Error RenderingDeviceVulkan::draw_list_begin_split(RID p_framebuffer, uint32_t p_splits, DrawListID *r_split_ids, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, const Vector<Color> &p_clear_color_values, float p_clear_depth, uint32_t p_clear_stencil, const Rect2 &p_region, const Vector<RID> &p_storage_textures) {
 	_THREAD_SAFE_METHOD_
 
+	ERR_FAIL_COND_V_MSG(draw_list != nullptr, ERR_BUSY, "Only one draw list can be active at the same time.");
+	ERR_FAIL_COND_V_MSG(compute_list != nullptr && !compute_list->state.allow_draw_overlap, ERR_BUSY, "Only one draw/compute list can be active at the same time.");
+
 	ERR_FAIL_COND_V(p_splits < 1, ERR_INVALID_DECLARATION);
 
 	Framebuffer *framebuffer = framebuffer_owner.get_or_null(p_framebuffer);
diff --git a/drivers/vulkan/rendering_device_vulkan.h b/drivers/vulkan/rendering_device_vulkan.h
index 2b92c74db9..10bdd7f4de 100644
--- a/drivers/vulkan/rendering_device_vulkan.h
+++ b/drivers/vulkan/rendering_device_vulkan.h
@@ -206,7 +206,7 @@ class RenderingDeviceVulkan : public RenderingDevice {
 	uint64_t staging_buffer_max_size = 0;
 	bool staging_buffer_used = false;
 
-	Error _staging_buffer_allocate(uint32_t p_amount, uint32_t p_required_align, uint32_t &r_alloc_offset, uint32_t &r_alloc_size, bool p_can_segment = true, bool p_on_draw_command_buffer = false);
+	Error _staging_buffer_allocate(uint32_t p_amount, uint32_t p_required_align, uint32_t &r_alloc_offset, uint32_t &r_alloc_size, bool p_can_segment = true);
 	Error _insert_staging_block();
 
 	struct Buffer {
@@ -637,7 +637,6 @@ class RenderingDeviceVulkan : public RenderingDevice {
 		};
 
 		bool is_compute = false;
-		int max_output = 0;
 		Vector<Set> sets;
 		Vector<uint32_t> set_formats;
 		Vector<VkPipelineShaderStageCreateInfo> pipeline_stages;
@@ -870,11 +869,9 @@ class RenderingDeviceVulkan : public RenderingDevice {
 			uint32_t pipeline_dynamic_state = 0;
 			VertexFormatID pipeline_vertex_format = INVALID_ID;
 			RID pipeline_shader;
-			uint32_t invalid_set_from = 0;
 			bool pipeline_uses_restart_indices = false;
 			uint32_t pipeline_primitive_divisor = 0;
 			uint32_t pipeline_primitive_minimum = 0;
-			Vector<uint32_t> pipeline_set_formats;
 			uint32_t pipeline_push_constant_size = 0;
 			bool pipeline_push_constant_supplied = false;
 		} validation;
@@ -948,7 +945,6 @@ class RenderingDeviceVulkan : public RenderingDevice {
 			bool pipeline_active = false;
 			RID pipeline_shader;
 			uint32_t invalid_set_from = 0;
-			Vector<uint32_t> pipeline_set_formats;
 			uint32_t pipeline_push_constant_size = 0;
 			bool pipeline_push_constant_supplied = false;
 		} validation;
diff --git a/drivers/vulkan/vulkan_context.cpp b/drivers/vulkan/vulkan_context.cpp
index 7944057041..0301f5b7fa 100644
--- a/drivers/vulkan/vulkan_context.cpp
+++ b/drivers/vulkan/vulkan_context.cpp
@@ -1860,16 +1860,16 @@ Error VulkanContext::initialize() {
 	return OK;
 }
 
-void VulkanContext::set_setup_buffer(const VkCommandBuffer &pCommandBuffer) {
-	command_buffer_queue.write[0] = pCommandBuffer;
+void VulkanContext::set_setup_buffer(VkCommandBuffer p_command_buffer) {
+	command_buffer_queue.write[0] = p_command_buffer;
 }
 
-void VulkanContext::append_command_buffer(const VkCommandBuffer &pCommandBuffer) {
+void VulkanContext::append_command_buffer(VkCommandBuffer p_command_buffer) {
 	if (command_buffer_queue.size() <= command_buffer_count) {
 		command_buffer_queue.resize(command_buffer_count + 1);
 	}
 
-	command_buffer_queue.write[command_buffer_count] = pCommandBuffer;
+	command_buffer_queue.write[command_buffer_count] = p_command_buffer;
 	command_buffer_count++;
 }
 
@@ -1879,7 +1879,10 @@ void VulkanContext::flush(bool p_flush_setup, bool p_flush_pending) {
 
 	//flush the pending setup buffer
 
-	if (p_flush_setup && command_buffer_queue[0]) {
+	bool setup_flushable = p_flush_setup && command_buffer_queue[0];
+	bool pending_flushable = p_flush_pending && command_buffer_count > 1;
+
+	if (setup_flushable) {
 		//use a fence to wait for everything done
 		VkSubmitInfo submit_info;
 		submit_info.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
@@ -1889,33 +1892,33 @@ void VulkanContext::flush(bool p_flush_setup, bool p_flush_pending) {
 		submit_info.pWaitSemaphores = nullptr;
 		submit_info.commandBufferCount = 1;
 		submit_info.pCommandBuffers = command_buffer_queue.ptr();
-		submit_info.signalSemaphoreCount = 0;
-		submit_info.pSignalSemaphores = nullptr;
+		submit_info.signalSemaphoreCount = pending_flushable ? 1 : 0;
+		submit_info.pSignalSemaphores = pending_flushable ? &draw_complete_semaphores[frame_index] : nullptr;
 		VkResult err = vkQueueSubmit(graphics_queue, 1, &submit_info, VK_NULL_HANDLE);
 		command_buffer_queue.write[0] = nullptr;
 		ERR_FAIL_COND(err);
-		vkDeviceWaitIdle(device);
 	}
 
-	if (p_flush_pending && command_buffer_count > 1) {
+	if (pending_flushable) {
 		//use a fence to wait for everything done
 
 		VkSubmitInfo submit_info;
 		submit_info.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
 		submit_info.pNext = nullptr;
-		submit_info.pWaitDstStageMask = nullptr;
-		submit_info.waitSemaphoreCount = 0;
-		submit_info.pWaitSemaphores = nullptr;
+		VkPipelineStageFlags wait_stage_mask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+		submit_info.pWaitDstStageMask = setup_flushable ? &wait_stage_mask : nullptr;
+		submit_info.waitSemaphoreCount = setup_flushable ? 1 : 0;
+		submit_info.pWaitSemaphores = setup_flushable ? &draw_complete_semaphores[frame_index] : nullptr;
 		submit_info.commandBufferCount = command_buffer_count - 1;
 		submit_info.pCommandBuffers = command_buffer_queue.ptr() + 1;
 		submit_info.signalSemaphoreCount = 0;
 		submit_info.pSignalSemaphores = nullptr;
 		VkResult err = vkQueueSubmit(graphics_queue, 1, &submit_info, VK_NULL_HANDLE);
-		ERR_FAIL_COND(err);
-		vkDeviceWaitIdle(device);
-
 		command_buffer_count = 1;
+		ERR_FAIL_COND(err);
 	}
+
+	vkDeviceWaitIdle(device);
 }
 
 Error VulkanContext::prepare_buffers() {
diff --git a/drivers/vulkan/vulkan_context.h b/drivers/vulkan/vulkan_context.h
index 236e3bf35f..e96facfacb 100644
--- a/drivers/vulkan/vulkan_context.h
+++ b/drivers/vulkan/vulkan_context.h
@@ -117,8 +117,8 @@ private:
 
 	// Present queue.
 	bool queues_initialized = false;
-	uint32_t graphics_queue_family_index = 0;
-	uint32_t present_queue_family_index = 0;
+	uint32_t graphics_queue_family_index = UINT32_MAX;
+	uint32_t present_queue_family_index = UINT32_MAX;
 	bool separate_present_queue = false;
 	VkQueue graphics_queue = VK_NULL_HANDLE;
 	VkQueue present_queue = VK_NULL_HANDLE;
@@ -289,8 +289,8 @@ public:
 	VkFormat get_screen_format() const;
 	VkPhysicalDeviceLimits get_device_limits() const;
 
-	void set_setup_buffer(const VkCommandBuffer &pCommandBuffer);
-	void append_command_buffer(const VkCommandBuffer &pCommandBuffer);
+	void set_setup_buffer(VkCommandBuffer p_command_buffer);
+	void append_command_buffer(VkCommandBuffer p_command_buffer);
 	void resize_notify();
 	void flush(bool p_flush_setup = false, bool p_flush_pending = false);
 	Error prepare_buffers();