Merge commit godotengine/godot@80f0b33313dae52d072ba2771a88ebcc4f0b4d6d

6 files changed, 75 insertions, 34 deletions

diff --git a/servers/rendering/renderer_canvas_cull.cpp b/servers/rendering/renderer_canvas_cull.cpp
index 9ec5d087ba..5fd4d12198 100644
--- a/servers/rendering/renderer_canvas_cull.cpp
+++ b/servers/rendering/renderer_canvas_cull.cpp
@@ -97,7 +97,7 @@ void RendererCanvasCull::_collect_ysort_children(RendererCanvasCull::Item *p_can
 			}
 
 			if (snapping_2d_transforms_to_pixel) {
-				child_xform.columns[2] = child_xform.columns[2].round();
+				child_xform.columns[2] = (child_xform.columns[2] + Point2(0.5, 0.5)).floor();
 			}
 
 			r_items[r_index] = child_items[i];
@@ -305,8 +305,8 @@ void RendererCanvasCull::_cull_canvas_item(Item *p_canvas_item, const Transform2
 		Transform2D parent_xform = p_parent_xform;
 
 		if (snapping_2d_transforms_to_pixel) {
-			self_xform.columns[2] = self_xform.columns[2].round();
-			parent_xform.columns[2] = parent_xform.columns[2].round();
+			self_xform.columns[2] = (self_xform.columns[2] + Point2(0.5, 0.5)).floor();
+			parent_xform.columns[2] = (parent_xform.columns[2] + Point2(0.5, 0.5)).floor();
 		}
 
 		final_xform = parent_xform * self_xform;
diff --git a/servers/rendering/renderer_rd/shaders/forward_mobile/scene_forward_mobile.glsl b/servers/rendering/renderer_rd/shaders/forward_mobile/scene_forward_mobile.glsl
index 9d47711599..2f0e4e0bea 100644
--- a/servers/rendering/renderer_rd/shaders/forward_mobile/scene_forward_mobile.glsl
+++ b/servers/rendering/renderer_rd/shaders/forward_mobile/scene_forward_mobile.glsl
@@ -105,10 +105,6 @@ layout(location = 6) mediump out vec3 binormal_interp;
 layout(location = 7) highp out vec4 diffuse_light_interp;
 layout(location = 8) highp out vec4 specular_light_interp;
 
-layout(constant_id = 9) const bool sc_disable_omni_lights = false;
-layout(constant_id = 10) const bool sc_disable_spot_lights = false;
-layout(constant_id = 12) const bool sc_disable_directional_lights = false;
-
 #include "../scene_forward_vertex_lights_inc.glsl"
 #endif // !defined(MODE_RENDER_DEPTH) && !defined(MODE_UNSHADED) && defined(USE_VERTEX_LIGHTING)
 #ifdef MATERIAL_UNIFORMS_USED
@@ -1606,7 +1602,6 @@ void main() {
 #endif
 #undef BIAS_FUNC
 			}
-#endif
 
 			if (i < 4) {
 				shadow0 |= uint(clamp(shadow * 255.0, 0.0, 255.0)) << (i * 8);
@@ -1614,6 +1609,7 @@ void main() {
 				shadow1 |= uint(clamp(shadow * 255.0, 0.0, 255.0)) << ((i - 4) * 8);
 			}
 		}
+#endif // SHADOWS_DISABLED
 
 #ifndef USE_VERTEX_LIGHTING
 		for (uint i = 0; i < scene_data.directional_light_count; i++) {
diff --git a/servers/rendering/rendering_device.cpp b/servers/rendering/rendering_device.cpp
index 74fe98f40a..1d01bb8b30 100644
--- a/servers/rendering/rendering_device.cpp
+++ b/servers/rendering/rendering_device.cpp
@@ -1245,6 +1245,7 @@ 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) {
@@ -1269,7 +1270,7 @@ 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;
@@ -1315,7 +1316,7 @@ 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;
@@ -1334,14 +1335,13 @@ 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);
@@ -5054,7 +5054,6 @@ RenderingDevice::TransferWorker *RenderingDevice::_acquire_transfer_worker(uint3
 			// No existing worker was picked, we create a new one.
 			transfer_worker = memnew(TransferWorker);
 			transfer_worker->command_fence = driver->fence_create();
-			transfer_worker->command_semaphore = driver->semaphore_create();
 			transfer_worker->command_pool = driver->command_pool_create(transfer_queue_family, RDD::COMMAND_BUFFER_TYPE_PRIMARY);
 			transfer_worker->command_buffer = driver->command_buffer_create(transfer_worker->command_pool);
 			transfer_worker->index = transfer_worker_pool.size();
@@ -5077,7 +5076,7 @@ RenderingDevice::TransferWorker *RenderingDevice::_acquire_transfer_worker(uint3
 		// If there's not enough bytes to use on the staging buffer, we submit everything pending from the worker and wait for the work to be finished.
 		if (transfer_worker->recording) {
 			_end_transfer_worker(transfer_worker);
-			_submit_transfer_worker(transfer_worker, false);
+			_submit_transfer_worker(transfer_worker);
 		}
 
 		if (transfer_worker->submitted) {
@@ -5130,12 +5129,12 @@ void RenderingDevice::_end_transfer_worker(TransferWorker *p_transfer_worker) {
 	p_transfer_worker->recording = false;
 }
 
-void RenderingDevice::_submit_transfer_worker(TransferWorker *p_transfer_worker, bool p_signal_semaphore) {
-	const VectorView<RDD::SemaphoreID> execute_semaphore = p_signal_semaphore ? p_transfer_worker->command_semaphore : VectorView<RDD::SemaphoreID>();
-	driver->command_queue_execute_and_present(transfer_queue, {}, p_transfer_worker->command_buffer, execute_semaphore, p_transfer_worker->command_fence, {});
-	if (p_signal_semaphore) {
+void RenderingDevice::_submit_transfer_worker(TransferWorker *p_transfer_worker, VectorView<RDD::SemaphoreID> p_signal_semaphores) {
+	driver->command_queue_execute_and_present(transfer_queue, {}, p_transfer_worker->command_buffer, p_signal_semaphores, p_transfer_worker->command_fence, {});
+
+	for (uint32_t i = 0; i < p_signal_semaphores.size(); i++) {
 		// Indicate the frame should wait on these semaphores before executing the main command buffer.
-		frames[frame].semaphores_to_wait_on.push_back(p_transfer_worker->command_semaphore);
+		frames[frame].semaphores_to_wait_on.push_back(p_signal_semaphores[i]);
 	}
 
 	p_transfer_worker->submitted = true;
@@ -5155,6 +5154,21 @@ 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) {
@@ -5196,10 +5210,11 @@ 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 (TransferWorker *worker : transfer_worker_pool) {
-		if (p_operations_used_by_draw) {
+	for (uint32_t i = 0; i < transfer_worker_pool.size(); i++) {
+		TransferWorker *worker = transfer_worker_pool[i];
+		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.
@@ -5210,11 +5225,21 @@ void RenderingDevice::_submit_transfer_workers(bool p_operations_used_by_draw) {
 		{
 			MutexLock lock(worker->thread_mutex);
 			if (worker->recording) {
+				VectorView<RDD::SemaphoreID> semaphores = p_draw_command_buffer ? frames[frame].transfer_worker_semaphores[i] : VectorView<RDD::SemaphoreID>();
 				_end_transfer_worker(worker);
-				_submit_transfer_worker(worker, true);
+				_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() {
@@ -5230,7 +5255,6 @@ void RenderingDevice::_wait_for_transfer_workers() {
 void RenderingDevice::_free_transfer_workers() {
 	MutexLock transfer_worker_lock(transfer_worker_pool_mutex);
 	for (TransferWorker *worker : transfer_worker_pool) {
-		driver->semaphore_free(worker->command_semaphore);
 		driver->fence_free(worker->command_fence);
 		driver->buffer_free(worker->staging_buffer);
 		driver->command_pool_free(worker->command_pool);
@@ -5809,10 +5833,10 @@ 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();
@@ -6016,6 +6040,9 @@ Error RenderingDevice::initialize(RenderingContextDriver *p_context, DisplayServ
 		present_queue_family = main_queue_family;
 	}
 
+	// Use the processor count as the max amount of transfer workers that can be created.
+	transfer_worker_pool_max_size = OS::get_singleton()->get_processor_count();
+
 	// Create data for all the frames.
 	for (uint32_t i = 0; i < frames.size(); i++) {
 		frames[i].index = 0;
@@ -6043,6 +6070,13 @@ Error RenderingDevice::initialize(RenderingContextDriver *p_context, DisplayServ
 
 		// Assign the main queue family and command pool to the command buffer pool.
 		frames[i].command_buffer_pool.pool = frames[i].command_pool;
+
+		// Create the semaphores for the transfer workers.
+		frames[i].transfer_worker_semaphores.resize(transfer_worker_pool_max_size);
+		for (uint32_t j = 0; j < transfer_worker_pool_max_size; j++) {
+			frames[i].transfer_worker_semaphores[j] = driver->semaphore_create();
+			ERR_FAIL_COND_V(!frames[i].transfer_worker_semaphores[j], FAILED);
+		}
 	}
 
 	// Start from frame count, so everything else is immediately old.
@@ -6089,9 +6123,6 @@ Error RenderingDevice::initialize(RenderingContextDriver *p_context, DisplayServ
 		ERR_FAIL_COND_V(err, FAILED);
 	}
 
-	// TODO: How should this max size be determined?
-	transfer_worker_pool_max_size = OS::get_singleton()->get_processor_count();
-
 	draw_list = nullptr;
 	compute_list = nullptr;
 
@@ -6382,7 +6413,7 @@ 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.
@@ -6454,6 +6485,10 @@ void RenderingDevice::finalize() {
 		for (uint32_t j = 0; j < buffer_pool.buffers.size(); j++) {
 			driver->semaphore_free(buffer_pool.semaphores[j]);
 		}
+
+		for (uint32_t j = 0; j < frames[i].transfer_worker_semaphores.size(); j++) {
+			driver->semaphore_free(frames[i].transfer_worker_semaphores[j]);
+		}
 	}
 
 	if (pipeline_cache_enabled) {
diff --git a/servers/rendering/rendering_device.h b/servers/rendering/rendering_device.h
index 60a8ef1a9c..3d6a8fc4a8 100644
--- a/servers/rendering/rendering_device.h
+++ b/servers/rendering/rendering_device.h
@@ -1269,7 +1269,7 @@ private:
 		RDD::CommandBufferID command_buffer;
 		RDD::CommandPoolID command_pool;
 		RDD::FenceID command_fence;
-		RDD::SemaphoreID command_semaphore;
+		LocalVector<RDD::TextureBarrier> texture_barriers;
 		bool recording = false;
 		bool submitted = false;
 		BinaryMutex thread_mutex;
@@ -1283,20 +1283,22 @@ private:
 	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;
 
 	TransferWorker *_acquire_transfer_worker(uint32_t p_transfer_size, uint32_t p_required_align, uint32_t &r_staging_offset);
 	void _release_transfer_worker(TransferWorker *p_transfer_worker);
 	void _end_transfer_worker(TransferWorker *p_transfer_worker);
-	void _submit_transfer_worker(TransferWorker *p_transfer_worker, bool p_signal_semaphore);
+	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();
 
@@ -1374,6 +1376,10 @@ private:
 		// Swap chains prepared for drawing during the frame that must be presented.
 		LocalVector<RDD::SwapChainID> swap_chains_to_present;
 
+		// Semaphores the transfer workers can use to wait before rendering the frame.
+		// This must have the same size of the transfer worker pool.
+		TightLocalVector<RDD::SemaphoreID> transfer_worker_semaphores;
+
 		// Extra command buffer pool used for driver workarounds.
 		RDG::CommandBufferPool command_buffer_pool;
 
diff --git a/servers/rendering/rendering_device_driver.cpp b/servers/rendering/rendering_device_driver.cpp
index 96876b4945..2ba75adef1 100644
--- a/servers/rendering/rendering_device_driver.cpp
+++ b/servers/rendering/rendering_device_driver.cpp
@@ -376,6 +376,8 @@ 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
index 0e83c2c42b..8730540509 100644
--- a/servers/rendering/rendering_device_driver.h
+++ b/servers/rendering/rendering_device_driver.h
@@ -222,6 +222,7 @@ public:
 
 	enum TextureLayout {
 		TEXTURE_LAYOUT_UNDEFINED,
+		TEXTURE_LAYOUT_GENERAL,
 		TEXTURE_LAYOUT_STORAGE_OPTIMAL,
 		TEXTURE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
 		TEXTURE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
@@ -752,6 +753,7 @@ public:
 		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 {