Merge pull request #98271 from DarioSamo/d3d12-enhanced-barrier-fix

clayjohn · web-flow · commit 58a7f9b4d838 · 2024-10-18T12:03:50.000-07:00
Move transitions of textures from transfer workers to the graphics queue.
diff --git a/drivers/d3d12/rendering_device_driver_d3d12.cpp b/drivers/d3d12/rendering_device_driver_d3d12.cpp
@@ -2003,6 +2003,8 @@ static D3D12_BARRIER_LAYOUT _rd_texture_layout_to_d3d12_barrier_layout(RDD::Text
 	switch (p_texture_layout) {
 		case RDD::TEXTURE_LAYOUT_UNDEFINED:
 			return D3D12_BARRIER_LAYOUT_UNDEFINED;
+		case RDD::TEXTURE_LAYOUT_GENERAL:
+			return D3D12_BARRIER_LAYOUT_COMMON;
 		case RDD::TEXTURE_LAYOUT_STORAGE_OPTIMAL:
 			return D3D12_BARRIER_LAYOUT_UNORDERED_ACCESS;
 		case RDD::TEXTURE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL:
@@ -6175,6 +6177,8 @@ uint64_t RenderingDeviceDriverD3D12::api_trait_get(ApiTrait p_trait) {
 			return false;
 		case API_TRAIT_CLEARS_WITH_COPY_ENGINE:
 			return false;
+		case API_TRAIT_USE_GENERAL_IN_COPY_QUEUES:
+			return true;
 		default:
 			return RenderingDeviceDriver::api_trait_get(p_trait);
 	}
diff --git a/drivers/vulkan/rendering_device_driver_vulkan.cpp b/drivers/vulkan/rendering_device_driver_vulkan.cpp
@@ -266,6 +266,7 @@ static const VkFormat RD_TO_VK_FORMAT[RDD::DATA_FORMAT_MAX] = {
 
 static VkImageLayout RD_TO_VK_LAYOUT[RDD::TEXTURE_LAYOUT_MAX] = {
 	VK_IMAGE_LAYOUT_UNDEFINED, // TEXTURE_LAYOUT_UNDEFINED
+	VK_IMAGE_LAYOUT_GENERAL, // TEXTURE_LAYOUT_GENERAL
 	VK_IMAGE_LAYOUT_GENERAL, // TEXTURE_LAYOUT_STORAGE_OPTIMAL
 	VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, // TEXTURE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
 	VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL, // TEXTURE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
diff --git a/servers/rendering/rendering_device.cpp b/servers/rendering/rendering_device.cpp
@@ -1243,6 +1243,7 @@ Error RenderingDevice::_texture_initialize(RID p_texture, uint32_t p_layer, cons
 	TransferWorker *transfer_worker = nullptr;
 	const uint8_t *read_ptr = p_data.ptr();
 	uint8_t *write_ptr = nullptr;
+	const RDD::TextureLayout copy_dst_layout = driver->api_trait_get(RDD::API_TRAIT_USE_GENERAL_IN_COPY_QUEUES) ? RDD::TEXTURE_LAYOUT_GENERAL : RDD::TEXTURE_LAYOUT_COPY_DST_OPTIMAL;
 	for (uint32_t pass = 0; pass < 2; pass++) {
 		const bool copy_pass = (pass == 1);
 		if (copy_pass) {
@@ -1267,7 +1268,7 @@ Error RenderingDevice::_texture_initialize(RID p_texture, uint32_t p_layer, cons
 				tb.texture = texture->driver_id;
 				tb.dst_access = RDD::BARRIER_ACCESS_COPY_WRITE_BIT;
 				tb.prev_layout = RDD::TEXTURE_LAYOUT_UNDEFINED;
-				tb.next_layout = RDD::TEXTURE_LAYOUT_COPY_DST_OPTIMAL;
+				tb.next_layout = copy_dst_layout;
 				tb.subresources.aspect = texture->barrier_aspect_flags;
 				tb.subresources.mipmap_count = texture->mipmaps;
 				tb.subresources.base_layer = p_layer;
@@ -1313,7 +1314,7 @@ Error RenderingDevice::_texture_initialize(RID p_texture, uint32_t p_layer, cons
 					copy_region.texture_subresources.layer_count = 1;
 					copy_region.texture_offset = Vector3i(0, 0, z);
 					copy_region.texture_region_size = Vector3i(logic_width, logic_height, 1);
-					driver->command_copy_buffer_to_texture(transfer_worker->command_buffer, transfer_worker->staging_buffer, texture->driver_id, RDD::TEXTURE_LAYOUT_COPY_DST_OPTIMAL, copy_region);
+					driver->command_copy_buffer_to_texture(transfer_worker->command_buffer, transfer_worker->staging_buffer, texture->driver_id, copy_dst_layout, copy_region);
 				}
 
 				staging_local_offset += to_allocate;
@@ -1332,14 +1333,13 @@ Error RenderingDevice::_texture_initialize(RID p_texture, uint32_t p_layer, cons
 				RDD::TextureBarrier tb;
 				tb.texture = texture->driver_id;
 				tb.src_access = RDD::BARRIER_ACCESS_COPY_WRITE_BIT;
-				tb.prev_layout = RDD::TEXTURE_LAYOUT_COPY_DST_OPTIMAL;
+				tb.prev_layout = copy_dst_layout;
 				tb.next_layout = RDD::TEXTURE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 				tb.subresources.aspect = texture->barrier_aspect_flags;
 				tb.subresources.mipmap_count = texture->mipmaps;
 				tb.subresources.base_layer = p_layer;
 				tb.subresources.layer_count = 1;
-
-				driver->command_pipeline_barrier(transfer_worker->command_buffer, RDD::PIPELINE_STAGE_COPY_BIT, RDD::PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, {}, {}, tb);
+				transfer_worker->texture_barriers.push_back(tb);
 			}
 
 			_release_transfer_worker(transfer_worker);
@@ -5152,6 +5152,21 @@ void RenderingDevice::_wait_for_transfer_worker(TransferWorker *p_transfer_worke
 		MutexLock lock(p_transfer_worker->operations_mutex);
 		p_transfer_worker->operations_processed = p_transfer_worker->operations_submitted;
 	}
+
+	if (!p_transfer_worker->texture_barriers.is_empty()) {
+		MutexLock transfer_worker_lock(transfer_worker_pool_mutex);
+		_flush_barriers_for_transfer_worker(p_transfer_worker);
+	}
+}
+
+void RenderingDevice::_flush_barriers_for_transfer_worker(TransferWorker *p_transfer_worker) {
+	if (!p_transfer_worker->texture_barriers.is_empty()) {
+		for (uint32_t i = 0; i < p_transfer_worker->texture_barriers.size(); i++) {
+			transfer_worker_pool_texture_barriers.push_back(p_transfer_worker->texture_barriers[i]);
+		}
+
+		p_transfer_worker->texture_barriers.clear();
+	}
 }
 
 void RenderingDevice::_check_transfer_worker_operation(uint32_t p_transfer_worker_index, uint64_t p_transfer_worker_operation) {
@@ -5193,11 +5208,11 @@ void RenderingDevice::_check_transfer_worker_index_array(IndexArray *p_index_arr
 	}
 }
 
-void RenderingDevice::_submit_transfer_workers(bool p_operations_used_by_draw) {
+void RenderingDevice::_submit_transfer_workers(RDD::CommandBufferID p_draw_command_buffer) {
 	MutexLock transfer_worker_lock(transfer_worker_pool_mutex);
 	for (uint32_t i = 0; i < transfer_worker_pool.size(); i++) {
 		TransferWorker *worker = transfer_worker_pool[i];
-		if (p_operations_used_by_draw) {
+		if (p_draw_command_buffer) {
 			MutexLock lock(worker->operations_mutex);
 			if (worker->operations_processed >= transfer_worker_operation_used_by_draw[worker->index]) {
 				// The operation used by the draw has already been processed, we don't need to wait on the worker.
@@ -5208,12 +5223,21 @@ void RenderingDevice::_submit_transfer_workers(bool p_operations_used_by_draw) {
 		{
 			MutexLock lock(worker->thread_mutex);
 			if (worker->recording) {
-				VectorView<RDD::SemaphoreID> semaphores = p_operations_used_by_draw ? frames[frame].transfer_worker_semaphores[i] : VectorView<RDD::SemaphoreID>();
+				VectorView<RDD::SemaphoreID> semaphores = p_draw_command_buffer ? frames[frame].transfer_worker_semaphores[i] : VectorView<RDD::SemaphoreID>();
 				_end_transfer_worker(worker);
 				_submit_transfer_worker(worker, semaphores);
 			}
+
+			if (p_draw_command_buffer) {
+				_flush_barriers_for_transfer_worker(worker);
+			}
 		}
 	}
+
+	if (p_draw_command_buffer && !transfer_worker_pool_texture_barriers.is_empty()) {
+		driver->command_pipeline_barrier(p_draw_command_buffer, RDD::PIPELINE_STAGE_COPY_BIT, RDD::PIPELINE_STAGE_ALL_COMMANDS_BIT, {}, {}, transfer_worker_pool_texture_barriers);
+		transfer_worker_pool_texture_barriers.clear();
+	}
 }
 
 void RenderingDevice::_wait_for_transfer_workers() {
@@ -5807,10 +5831,10 @@ 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).");
 	}
 
-	_submit_transfer_workers(true);
-
 	// 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].command_buffer;
+	_submit_transfer_workers(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();
@@ -6387,7 +6411,7 @@ void RenderingDevice::finalize() {
 	}
 
 	// Wait for transfer workers to finish.
-	_submit_transfer_workers(false);
+	_submit_transfer_workers();
 	_wait_for_transfer_workers();
 
 	// Delete everything the graph has created.
diff --git a/servers/rendering/rendering_device.h b/servers/rendering/rendering_device.h
@@ -1267,6 +1267,7 @@ class RenderingDevice : public RenderingDeviceCommons {
 		RDD::CommandBufferID command_buffer;
 		RDD::CommandPoolID command_pool;
 		RDD::FenceID command_fence;
+		LocalVector<RDD::TextureBarrier> texture_barriers;
 		bool recording = false;
 		bool submitted = false;
 		BinaryMutex thread_mutex;
@@ -1280,6 +1281,7 @@ class RenderingDevice : public RenderingDeviceCommons {
 	uint32_t transfer_worker_pool_max_size = 1;
 	LocalVector<uint64_t> transfer_worker_operation_used_by_draw;
 	LocalVector<uint32_t> transfer_worker_pool_available_list;
+	LocalVector<RDD::TextureBarrier> transfer_worker_pool_texture_barriers;
 	BinaryMutex transfer_worker_pool_mutex;
 	ConditionVariable transfer_worker_pool_condition;
 
@@ -1288,12 +1290,13 @@ class RenderingDevice : public RenderingDeviceCommons {
 	void _end_transfer_worker(TransferWorker *p_transfer_worker);
 	void _submit_transfer_worker(TransferWorker *p_transfer_worker, VectorView<RDD::SemaphoreID> p_signal_semaphores = VectorView<RDD::SemaphoreID>());
 	void _wait_for_transfer_worker(TransferWorker *p_transfer_worker);
+	void _flush_barriers_for_transfer_worker(TransferWorker *p_transfer_worker);
 	void _check_transfer_worker_operation(uint32_t p_transfer_worker_index, uint64_t p_transfer_worker_operation);
 	void _check_transfer_worker_buffer(Buffer *p_buffer);
 	void _check_transfer_worker_texture(Texture *p_texture);
 	void _check_transfer_worker_vertex_array(VertexArray *p_vertex_array);
 	void _check_transfer_worker_index_array(IndexArray *p_index_array);
-	void _submit_transfer_workers(bool p_operations_used_by_draw);
+	void _submit_transfer_workers(RDD::CommandBufferID p_draw_command_buffer = RDD::CommandBufferID());
 	void _wait_for_transfer_workers();
 	void _free_transfer_workers();
 
diff --git a/servers/rendering/rendering_device_driver.cpp b/servers/rendering/rendering_device_driver.cpp
@@ -374,6 +374,8 @@ uint64_t RenderingDeviceDriver::api_trait_get(ApiTrait p_trait) {
 			return 1;
 		case API_TRAIT_CLEARS_WITH_COPY_ENGINE:
 			return true;
+		case API_TRAIT_USE_GENERAL_IN_COPY_QUEUES:
+			return false;
 		default:
 			ERR_FAIL_V(0);
 	}
diff --git a/servers/rendering/rendering_device_driver.h b/servers/rendering/rendering_device_driver.h
@@ -220,6 +220,7 @@ class RenderingDeviceDriver : public RenderingDeviceCommons {
 
 	enum TextureLayout {
 		TEXTURE_LAYOUT_UNDEFINED,
+		TEXTURE_LAYOUT_GENERAL,
 		TEXTURE_LAYOUT_STORAGE_OPTIMAL,
 		TEXTURE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
 		TEXTURE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
@@ -750,6 +751,7 @@ class RenderingDeviceDriver : public RenderingDeviceCommons {
 		API_TRAIT_TEXTURE_DATA_ROW_PITCH_STEP,
 		API_TRAIT_SECONDARY_VIEWPORT_SCISSOR,
 		API_TRAIT_CLEARS_WITH_COPY_ENGINE,
+		API_TRAIT_USE_GENERAL_IN_COPY_QUEUES,
 	};
 
 	enum ShaderChangeInvalidation {

Original file line number	Diff line number	Diff line change
`@@ -374,6 +374,8 @@ uint64_t RenderingDeviceDriver::api_trait_get(ApiTrait p_trait) {`
`374`	`374`	`return 1;`
`375`	`375`	`case API_TRAIT_CLEARS_WITH_COPY_ENGINE:`
`376`	`376`	`return true;`
	`377`	`+ case API_TRAIT_USE_GENERAL_IN_COPY_QUEUES:`
	`378`	`+ return false;`
`377`	`379`	`default:`
`378`	`380`	`ERR_FAIL_V(0);`
`379`	`381`	`}`