Quelle  render_task_graph.rs

// This Source Code Form is subject to the terms of the Mozilla Public
// License, v. 2.0. If a copy of the MPL was not distributed with this
// file, You can obtain one at http://mozilla.org/MPL/2.0/.

//! This module contains the render task graph.
//!
//! Code associated with creating specific render tasks is in the render_task
//! module.

use api::units::*;
use api::ImageFormat;
use crate::gpu_types::ImageSource;
use crate::internal_types::{TextureSource, CacheTextureId, FastHashMap, FastHashSet, FrameId};
use crate::internal_types::size_of_frame_vec;
use crate::render_task::{StaticRenderTaskSurface, RenderTaskLocation, RenderTask, SubTask};
use crate::render_target::RenderTargetKind;
use crate::render_task::{RenderTaskData, RenderTaskKind};
use crate::renderer::GpuBufferAddress;
use crate::renderer::GpuBufferBuilder;
use crate::resource_cache::ResourceCache;
use crate::texture_pack::GuillotineAllocator;
use crate::prim_store::DeferredResolve;
use crate::image_source::{resolve_image, resolve_cached_render_task};
use smallvec::SmallVec;
use topological_sort::TopologicalSort;

use crate::render_target::{RenderTargetList, PictureCacheTarget, RenderTarget};
use crate::util::{Allocation, VecHelper};
use std::{usize, f32};

use crate::internal_types::{FrameVec, FrameMemory};

#[cfg(test)]
use crate::frame_allocator::FrameAllocator;

/// If we ever need a larger texture than the ideal, we better round it up to a
/// reasonable number in order to have a bit of leeway in case the size of this
/// this target is changing each frame.
const TEXTURE_DIMENSION_MASK: i32 = 0xFF;

/// Allows initializing a render task directly into the render task buffer.
///
/// See utils::VecHelpers. RenderTask is fairly large so avoiding the move when
/// pushing into the vector can save a lot of expensive memcpys on pages with many
/// render tasks.
pub struct RenderTaskAllocation<'a> {
    pub alloc: Allocation<'a, RenderTask>,
}

impl<'l> RenderTaskAllocation<'l> {
    #[inline(always)]
    pub fn init(self, value: RenderTask) -> RenderTaskId {
        RenderTaskId::from_index(self.alloc.init(value))
    }
}

#[derive(Copy, Clone, PartialEq, Eq, Hash)]
#[derive(MallocSizeOf)]
#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
pub struct RenderTaskId {
    pub index: u32,
    pub sub_rect_index: u16,
}

impl RenderTaskId {
    pub const INVALID: RenderTaskId = RenderTaskId {
        index: u32::MAX,
        sub_rect_index: u16::MAX,
    };

    #[inline]
    fn from_index(index: usize) -> Self {
        RenderTaskId { index: index as u32, sub_rect_index: u16::MAX }
    }

    #[inline]
    pub fn index(&self) -> usize {
        self.index as usize
    }

    pub fn has_sub_rect(&self) -> bool {
        self.sub_rect_index != u16::MAX
    }
}

impl std::fmt::Debug for RenderTaskId {
    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
        if *self == Self::INVALID {
            write!(f, "<invalid>")
        } else {
            write!(f, "#{}", self.index)
        }
    }
}

#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
#[derive(Clone)]
pub struct SubTaskRange(std::ops::Range<u32>);

impl SubTaskRange {
    pub fn empty() -> Self { SubTaskRange(0..0) }
    pub fn is_empty(&self) -> bool { self.0.is_empty() }
}

impl std::fmt::Debug for SubTaskRange {
    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
        if self.is_empty() {
            write!(f, "<empty>")
        } else {
            self.0.fmt(f)
        }
    }
}

#[derive(Copy, Clone, PartialEq, Eq, Hash)]
#[derive(MallocSizeOf)]
#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
pub struct SubTaskId(u32);

impl std::fmt::Debug for SubTaskId {
    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
        write!(f, "#{}", self.0)
    }
}

impl Iterator for SubTaskRange {
    type Item = SubTaskId;
    fn next(&mut self) -> Option<Self::Item> {
        Some(SubTaskId(self.0.next()?))
    }
}

#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
#[derive(Debug, Copy, Clone, Hash, Eq, PartialEq, PartialOrd, Ord)]
pub struct PassId(usize);

impl PassId {
    pub const MIN: PassId = PassId(0);
    pub const MAX: PassId = PassId(!0 - 1);
    pub const INVALID: PassId = PassId(!0 - 2);
}

/// An internal representation of a dynamic surface that tasks can be
/// allocated into. Maintains some extra metadata about each surface
/// during the graph build.
#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
struct Surface {
    /// Whether this is a color or alpha render target
    kind: RenderTargetKind,
    /// Allocator for this surface texture
    allocator: GuillotineAllocator,
    /// We can only allocate into this for reuse if it's a shared surface
    is_shared: bool,
    /// The lifetime group of this surface: only tasks with a matching
    /// lifetime_group can share it, to avoid holding the surface longer
    /// than necessary.
    lifetime_group: PassId,
    /// Reference count: number of tasks whose individual free_after will trigger
    /// a decrement. Surface is returned to the pool when this reaches 0.
    pending_frees: usize,
}

impl Surface {
    /// Allocate a rect within a shared surfce. Returns None if the
    /// format doesn't match, or allocation fails.
    fn alloc_rect(
        &mut self,
        size: DeviceIntSize,
        kind: RenderTargetKind,
        is_shared: bool,
        lifetime_group: PassId,
    ) -> Option<DeviceIntPoint> {
        if self.kind == kind && self.is_shared == is_shared && self.lifetime_group == lifetime_group {
            self.allocator
                .allocate(&size)
                .map(|(_slice, origin)| origin)
        } else {
            None
        }
    }
}

/// A sub-pass can draw to either a dynamic (temporary render target) surface,
/// or a persistent surface (texture or picture cache).
#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
#[derive(Debug)]
pub enum SubPassSurface {
    /// A temporary (intermediate) surface.
    Dynamic {
        /// The renderer texture id
        texture_id: CacheTextureId,
        /// Color / alpha render target
        target_kind: RenderTargetKind,
        /// The rectangle occupied by tasks in this surface. Used as a clear
        /// optimization on some GPUs.
        used_rect: DeviceIntRect,
    },
    Persistent {
        /// Reference to the texture or picture cache surface being drawn to.
        surface: StaticRenderTaskSurface,
    },
}

/// A subpass is a specific render target, and a list of tasks to draw to it.
#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
pub struct SubPass {
    /// The surface this subpass draws to
    pub surface: SubPassSurface,
    /// The tasks assigned to this subpass.
    pub task_ids: FrameVec<RenderTaskId>,
}

/// A pass expresses dependencies between tasks. Each pass consists of a number
/// of subpasses.
#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
pub struct Pass {
    /// The tasks assigned to this render pass
    pub task_ids: FrameVec<RenderTaskId>,
    /// The subpasses that make up this dependency pass
    pub sub_passes: FrameVec<SubPass>,
    /// A list of intermediate surfaces that can be invalidated after
    /// this pass completes.
    pub textures_to_invalidate: FrameVec<CacheTextureId>,
}

#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
pub struct TaskSubRect {
    pub sub_rect: DeviceIntRect,
    source_task: RenderTaskId,
    uv_address: GpuBufferAddress,
}

/// The RenderTaskGraph is the immutable representation of the render task graph. It is
/// built by the RenderTaskGraphBuilder, and is constructed once per frame.
#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
pub struct RenderTaskGraph {
    /// List of tasks added to the graph
    pub tasks: FrameVec<RenderTask>,
    /// List of sub-tasks, executing after the regular tasks for a given
    /// pass.
    pub sub_tasks: FrameVec<SubTask>,

    pub sub_rects: FrameVec<TaskSubRect>,

    /// The passes that were created, based on dependencies between tasks
    pub passes: FrameVec<Pass>,

    /// Current frame id, used for debug validation
    frame_id: FrameId,

    /// GPU specific data for each task that is made available to shaders
    pub task_data: FrameVec<RenderTaskData>,

    /// Total number of intermediate surfaces that will be drawn to, used for test validation.
    #[cfg(test)]
    surface_count: usize,

    /// Total number of real allocated textures that will be drawn to, used for test validation.
    #[cfg(test)]
    unique_surfaces: FastHashSet<CacheTextureId>,
}

/// The persistent interface that is used during frame building to construct the
/// frame graph.
pub struct RenderTaskGraphBuilder {
    /// List of tasks added to the builder
    tasks: Vec<RenderTask>,
    /// List of sub-tasks added to the builder
    sub_tasks: Vec<SubTask>,
    /// render task ids can optionally refer to sub-rects of a render task
    /// by indexing into this vector.
    sub_rects: Vec<TaskSubRect>,

    /// List of task roots
    roots: FastHashSet<RenderTaskId>,

    /// Current frame id, used for debug validation
    frame_id: FrameId,

    /// A list of texture surfaces that can be freed at the end of a pass. Retained
    /// here to reduce heap allocations.
    textures_to_free: FastHashSet<CacheTextureId>,

    /// Set of task ids already processed for freeing in the current pass,
    /// used to avoid double-counting when multiple parents reference
    /// the same child task.
    freed_tasks: FastHashSet<RenderTaskId>,

    // Keep a map of `texture_id` to metadata about surfaces that are currently
    // borrowed from the render target pool.
    active_surfaces: FastHashMap<CacheTextureId, Surface>,
}

impl RenderTaskGraphBuilder {
    /// Construct a new graph builder. Typically constructed once and maintained
    /// over many frames, to avoid extra heap allocations where possible.
    pub fn new() -> Self {
        RenderTaskGraphBuilder {
            tasks: Vec::new(),
            sub_tasks: Vec::new(),
            sub_rects: Vec::new(),
            roots: FastHashSet::default(),
            frame_id: FrameId::INVALID,
            textures_to_free: FastHashSet::default(),
            freed_tasks: FastHashSet::default(),
            active_surfaces: FastHashMap::default(),
        }
    }

    pub fn frame_id(&self) -> FrameId {
        self.frame_id
    }

    /// Begin a new frame
    pub fn begin_frame(&mut self, frame_id: FrameId) {
        self.frame_id = frame_id;
        self.roots.clear();
    }

    /// Get immutable access to a task
    // TODO(gw): There's only a couple of places that existing code needs to access
    //           a task during the building step. Perhaps we can remove this?
    pub fn get_task(
        &self,
        task_id: RenderTaskId,
    ) -> &RenderTask {
        &self.tasks[task_id.index as usize]
    }

    /// Get mutable access to a task
    // TODO(gw): There's only a couple of places that existing code needs to access
    //           a task during the building step. Perhaps we can remove this?
    pub fn get_task_mut(
        &mut self,
        task_id: RenderTaskId,
    ) -> &mut RenderTask {
        &mut self.tasks[task_id.index as usize]
    }

    /// Add a new task to the graph.
    pub fn add(&mut self) -> RenderTaskAllocation {
        // Assume every task is a root to start with
        self.roots.insert(
            RenderTaskId::from_index(self.tasks.len()),
        );

        RenderTaskAllocation {
            alloc: self.tasks.alloc(),
        }
    }

    pub fn begin_sub_tasks(&self) -> SubTaskRange {
        let first = self.sub_tasks.len() as u32;
        SubTaskRange(first..first)
    }

    pub fn push_sub_task(&mut self, in_range: &mut SubTaskRange, task: SubTask) {
        // Check that we are pushing sub-tasks as contiguous sequences.
        assert_eq!(in_range.0.end as usize, self.sub_tasks.len());
        in_range.0.end += 1;

        self.sub_tasks.push(task);
    }

    /// Express a dependency, such that `task_id` depends on `input` as a texture source.
    pub fn add_dependency(
        &mut self,
        task_id: RenderTaskId,
        input: RenderTaskId,
    ) {
        self.tasks[task_id.index as usize].children.push(input);

        // Once a task is an input, it's no longer a root
        self.roots.remove(&input);
    }

    pub fn add_sub_rect(&mut self, source_task: RenderTaskId, sub_rect: &DeviceIntRect) -> RenderTaskId {
        assert!(self.sub_rects.len() < u16::MAX as usize);
        if source_task == RenderTaskId::INVALID {
            return RenderTaskId::INVALID;
        }

        let sub_rect_index = self.sub_rects.len() as u16;
        self.sub_rects.push(TaskSubRect {
            source_task,
            sub_rect: *sub_rect,
            uv_address: GpuBufferAddress::INVALID,
        });

        RenderTaskId {
            index: source_task.index,
            sub_rect_index,
        }
    }

    /// End the graph building phase and produce the immutable task graph for this frame
    pub fn end_frame(
        &mut self,
        resource_cache: &mut ResourceCache,
        gpu_buffers: &mut GpuBufferBuilder,
        deferred_resolves: &mut FrameVec<DeferredResolve>,
        max_shared_surface_size: i32,
        memory: &FrameMemory,
    ) -> RenderTaskGraph {
        // Copy the render tasks over to the immutable graph output
        let task_count = self.tasks.len();

        // Copy from the frame_builder's task vector to the frame's instead of stealing it
        // because they use different memory allocators. TODO: The builder should use the
        // frame allocator, however since the builder lives longer than the frame, it's a
        // bit more risky to do so.
        let mut tasks = memory.new_vec_with_capacity(task_count);
        for task in self.tasks.drain(..) {
            tasks.push(task)
        }

        let mut sub_tasks = memory.new_vec_with_capacity(self.sub_tasks.len());
        for task in self.sub_tasks.drain(..) {
            sub_tasks.push(task)
        }

        let mut graph = RenderTaskGraph {
            tasks,
            sub_tasks,
            sub_rects: memory.new_vec(),
            passes: memory.new_vec(),
            task_data: memory.new_vec_with_capacity(task_count),
            frame_id: self.frame_id,
            #[cfg(test)]
            surface_count: 0,
            #[cfg(test)]
            unique_surfaces: FastHashSet::default(),
        };

        // First, use a topological sort of the dependency graph to split the task set in to
        // a list of passes. This is necessary because when we have a complex graph (e.g. due
        // to a large number of sibling backdrop-filter primitives) traversing it via a simple
        // recursion can be too slow. The second pass determines when the last time a render task
        // is used as an input, and assigns what pass the surface backing that render task can
        // be freed (the surface is then returned to the render target pool and may be aliased
        // or reused during subsequent passes).

        let mut pass_count = 0;
        let mut passes = memory.new_vec();
        let mut task_sorter = TopologicalSort::<RenderTaskId>::new();

        // Iterate the task list, and add all the dependencies to the topo sort
        for (parent_id, task) in graph.tasks.iter().enumerate() {
            let parent_id = RenderTaskId::from_index(parent_id);

            for child_id in &task.children {
                task_sorter.add_dependency(
                    parent_id,
                    *child_id,
                );
            }
        }

        // Pop the sorted passes off the topological sort
        loop {
            // Get the next set of tasks that can be drawn
            let tasks = task_sorter.pop_all();

            // If there are no tasks left, we're done
            if tasks.is_empty() {
                // If the task sorter itself isn't empty but we couldn't pop off any
                // tasks, that implies a circular dependency in the task graph
                assert!(task_sorter.is_empty());
                break;
            } else {
                // Assign the `render_on` field to the task
                for task_id in &tasks {
                    graph.tasks[task_id.index as usize].render_on = PassId(pass_count);
                }

                // Store the task list for this pass, used later for `assign_free_pass`.
                passes.push(tasks);
                pass_count += 1;
            }
        }

        // Always create at least one pass for root tasks
        pass_count = pass_count.max(1);

        // Determine which pass each task can be freed on, which depends on which is
        // the last task that has this as an input. This must be done in top-down
        // pass order to ensure that RenderTaskLocation::Existing references are
        // visited in the correct order
        for pass in passes {
            for task_id in pass {
                assign_free_pass(
                    task_id,
                    &mut graph,
                );
            }
        }

        // Construct passes array for tasks to be assigned to below
        for _ in 0 .. pass_count {
            graph.passes.push(Pass {
                task_ids: memory.new_vec(),
                sub_passes: memory.new_vec(),
                textures_to_invalidate: memory.new_vec(),
            });
        }

        // Assign tasks to each pass based on their `render_on` attribute
        for (index, task) in graph.tasks.iter().enumerate() {
            if task.kind.is_a_rendering_operation() {
                let id = RenderTaskId::from_index(index);
                graph.passes[task.render_on.0].task_ids.push(id);
            }
        }

        // At this point, tasks are assigned to each dependency pass. Now we
        // can go through each pass and create sub-passes, assigning each task
        // to a target and destination rect.
        assert!(self.active_surfaces.is_empty());

        for (pass_id, pass) in graph.passes.iter_mut().enumerate().rev() {
            assert!(self.textures_to_free.is_empty());

            // Phase 1: Allocate all tasks in this pass to surfaces, incrementing
            // pending_frees on each surface for every task (including Existing).
            for task_id in &pass.task_ids {

                let task_location = graph.tasks[task_id.index as usize].location.clone();

                match task_location {
                    RenderTaskLocation::Unallocated { size } => {
                        let task = &mut graph.tasks[task_id.index as usize];

                        let mut location = None;
                        let kind = task.kind.target_kind();

                        let can_use_shared_surface =
                            task.kind.can_use_shared_surface();

                        if can_use_shared_surface {
                            // If we can use a shared surface, step through the existing shared
                            // surfaces for this subpass, and see if we can allocate the task
                            // to one of these targets.
                            for sub_pass in &mut pass.sub_passes {
                                if let SubPassSurface::Dynamic { texture_id, ref mut used_rect, .. } = sub_pass.surface {
                                    let surface = self.active_surfaces.get_mut(&texture_id).unwrap();
                                    if let Some(p) = surface.alloc_rect(size, kind, true, task.free_after) {
                                        surface.pending_frees += 1;
                                        location = Some((texture_id, p));
                                        *used_rect = used_rect.union(&DeviceIntRect::from_origin_and_size(p, size));
                                        sub_pass.task_ids.push(*task_id);
                                        break;
                                    }
                                }
                            }
                        }

                        if location.is_none() {
                            // If it wasn't possible to allocate the task to a shared surface, get a new
                            // render target from the resource cache pool/

                            // If this is a really large task, don't bother allocating it as a potential
                            // shared surface for other tasks.

                            let can_use_shared_surface = can_use_shared_surface &&
                                size.width <= max_shared_surface_size &&
                                size.height <= max_shared_surface_size;

                            let surface_size = if can_use_shared_surface {
                                DeviceIntSize::new(
                                    max_shared_surface_size,
                                    max_shared_surface_size,
                                )
                            } else {
                                // Round up size here to avoid constant re-allocs during resizing
                                DeviceIntSize::new(
                                    (size.width + TEXTURE_DIMENSION_MASK) & !TEXTURE_DIMENSION_MASK,
                                    (size.height + TEXTURE_DIMENSION_MASK) & !TEXTURE_DIMENSION_MASK,
                                )
                            };

                            if surface_size.is_empty() {
                                // We would panic in the guillotine allocator. Instead, panic here
                                // with some context.
                                let task_name = graph.tasks[task_id.index as usize].kind.as_str();
                                panic!("{} render task has invalid size {:?}", task_name, surface_size);
                            }

                            let format = match kind {
                                RenderTargetKind::Color => ImageFormat::RGBA8,
                                RenderTargetKind::Alpha => ImageFormat::R8,
                            };

                            // Get render target of appropriate size and format from resource cache
                            let texture_id = resource_cache.get_or_create_render_target_from_pool(
                                surface_size,
                                format,
                            );

                            // Allocate metadata we need about this surface while it's active
                            let mut surface = Surface {
                                kind,
                                allocator: GuillotineAllocator::new(Some(surface_size)),
                                is_shared: can_use_shared_surface,
                                lifetime_group: task.free_after,
                                pending_frees: 1,
                            };

                            // Allocation of the task must fit in this new surface!
                            let p = surface.alloc_rect(
                                size,
                                kind,
                                can_use_shared_surface,
                                task.free_after,
                            ).expect("bug: alloc must succeed!");

                            location = Some((texture_id, p));

                            // Store the metadata about this newly active surface. We should never
                            // get a target surface with the same texture_id as a currently active surface.
                            let _prev_surface = self.active_surfaces.insert(texture_id, surface);
                            assert!(_prev_surface.is_none());

                            // Store some information about surface allocations if in test mode
                            #[cfg(test)]
                            {
                                graph.surface_count += 1;
                                graph.unique_surfaces.insert(texture_id);
                            }

                            let mut task_ids = memory.new_vec();
                            task_ids.push(*task_id);

                            // Add the target as a new subpass for this render pass.
                            pass.sub_passes.push(SubPass {
                                surface: SubPassSurface::Dynamic {
                                    texture_id,
                                    target_kind: kind,
                                    used_rect: DeviceIntRect::from_origin_and_size(p, size),
                                },
                                task_ids,
                            });
                        }

                        // By now, we must have allocated a surface and rect for this task, so assign it!
                        assert!(location.is_some());
                        task.location = RenderTaskLocation::Dynamic {
                            texture_id: location.unwrap().0,
                            rect: DeviceIntRect::from_origin_and_size(location.unwrap().1, size),
                        };
                    }
                    RenderTaskLocation::Existing { parent_task_id, size: existing_size, .. } => {
                        let parent_task_location = graph.tasks[parent_task_id.index as usize].location.clone();

                        match parent_task_location {
                            RenderTaskLocation::Unallocated { .. } |
                            RenderTaskLocation::CacheRequest { .. } |
                            RenderTaskLocation::Existing { .. } => {
                                panic!("bug: reference to existing task must be allocated by now");
                            }
                            RenderTaskLocation::Dynamic { texture_id, rect, .. } => {
                                assert_eq!(existing_size, rect.size());

                                let surface = self.active_surfaces.get_mut(&texture_id).unwrap();
                                surface.pending_frees += 1;

                                let kind = graph.tasks[parent_task_id.index as usize].kind.target_kind();
                                let mut task_ids = memory.new_vec();
                                task_ids.push(*task_id);
                                // A sub-pass is always created in this case, as existing tasks by definition can't be shared.
                                pass.sub_passes.push(SubPass {
                                    surface: SubPassSurface::Dynamic {
                                        texture_id,
                                        target_kind: kind,
                                        used_rect: rect,        // clear will be skipped due to no-op check anyway
                                    },
                                    task_ids,
                                });

                                let task = &mut graph.tasks[task_id.index as usize];
                                task.location = parent_task_location;
                            }
                            RenderTaskLocation::Static { .. } => {
                                unreachable!("bug: not possible since we don't dup static locations");
                            }
                        }
                    }
                    RenderTaskLocation::Static { ref surface, .. } => {
                        // No need to allocate for this surface, since it's a persistent
                        // target. Instead, just create a new sub-pass for it.
                        let mut task_ids = memory.new_vec();
                        task_ids.push(*task_id);
                        pass.sub_passes.push(SubPass {
                            surface: SubPassSurface::Persistent {
                                surface: surface.clone(),
                            },
                            task_ids,
                        });
                    }
                    RenderTaskLocation::CacheRequest { .. } => {
                        // No need to allocate nor to create a sub-path for read-only locations.
                    }
                    RenderTaskLocation::Dynamic { .. } => {
                        // Dynamic tasks shouldn't be allocated by this point
                        panic!("bug: encountered an already allocated task");
                    }
                }
            }

            // Phase 2: For each child task whose free_after matches this pass,
            // decrement the pending_frees on its surface. When pending_frees
            // reaches 0, the surface can be returned to the pool.
            assert!(self.freed_tasks.is_empty());
            for task_id in &pass.task_ids {
                let task = &graph.tasks[task_id.index as usize];
                for child_id in &task.children {
                    let child_task = &graph.tasks[child_id.index as usize];
                    match child_task.location {
                        RenderTaskLocation::Unallocated { .. } |
                        RenderTaskLocation::Existing { .. } => panic!("bug: must be allocated"),
                        RenderTaskLocation::Dynamic { texture_id, .. } => {
                            if child_task.free_after == PassId(pass_id) &&
                               self.freed_tasks.insert(*child_id)
                            {
                                let surface = self.active_surfaces.get_mut(&texture_id).unwrap();
                                surface.pending_frees -= 1;
                                if surface.pending_frees == 0 {
                                    self.textures_to_free.insert(texture_id);
                                }
                            }
                        }
                        RenderTaskLocation::Static { .. } => {}
                        RenderTaskLocation::CacheRequest { .. } => {}
                    }
                }
            }
            self.freed_tasks.clear();

            // Return no longer used textures to the pool, so that they can be reused / aliased
            // by later passes.
            for texture_id in self.textures_to_free.drain() {
                resource_cache.return_render_target_to_pool(texture_id);
                self.active_surfaces.remove(&texture_id).unwrap();
                pass.textures_to_invalidate.push(texture_id);
            }
        }

        if !self.active_surfaces.is_empty() {
            graph.print();
            // By now, all surfaces that were borrowed from the render target pool must
            // be returned to the resource cache, or we are leaking intermediate surfaces!
            assert!(self.active_surfaces.is_empty());
        }

        // Each task is now allocated to a surface and target rect. Write that to the
        // GPU blocks and task_data. After this point, the graph is returned and is
        // considered to be immutable for the rest of the frame building process.

        for task in &mut graph.tasks {
            // Check whether the render task texture and uv rects are managed externally.
            // This is the case for image tasks and cached tasks. In both cases it
            // results in a finding the information in the texture cache.
            let cache_item = if let Some(ref cache_handle) = task.cache_handle {
                Some(resolve_cached_render_task(
                    cache_handle,
                    resource_cache,
                ))
            } else if let RenderTaskKind::Image(info) = &task.kind {
                Some(resolve_image(
                    info.request,
                    resource_cache,
                    &mut gpu_buffers.f32,
                    deferred_resolves,
                    info.is_composited,
                ))
            } else {
                // General case (non-cached non-image tasks).
                None
            };

            if let Some(cache_item) = &cache_item {
                task.uv_rect_handle = gpu_buffers.f32.resolve_handle(cache_item.uv_rect_handle);

                // Update the render task even if the item is invalid.
                // We'll handle it later and it's easier to not have to
                // deal with unexpected location variants like
                // RenderTaskLocation::CacheRequest when we do.
                if let RenderTaskLocation::CacheRequest { .. } = &task.location {
                    let source = cache_item.texture_id;
                    task.location = RenderTaskLocation::Static {
                        surface: StaticRenderTaskSurface::ReadOnly { source },
                        rect: cache_item.uv_rect,
                    };
                }
            }

            // This has to be done after we do the task location fixup above.
            let target_rect = task.get_target_rect();

            // If the uv rect is not managed externally, generate it now.
            if cache_item.is_none() {
                let image_source = ImageSource {
                    p0: target_rect.min.to_f32(),
                    p1: target_rect.max.to_f32(),
                    user_data: [0.0; 4],
                    uv_rect_kind: task.uv_rect_kind,
                };

                let uv_rect_handle = image_source.write_gpu_blocks(&mut gpu_buffers.f32);
                task.uv_rect_handle = gpu_buffers.f32.resolve_handle(uv_rect_handle);
            }

            // Give the render task an opportunity to add any
            // information to the GPU cache, if appropriate.
            task.kind.write_gpu_blocks(gpu_buffers);

            graph.task_data.push(
                task.kind.write_task_data(target_rect)
            );
        }

        graph.sub_rects.reserve(self.sub_rects.len());
        for item in self.sub_rects.drain(..) {
            let task = &graph.tasks[item.source_task.index()];
            let task_rect = task.get_target_rect();
            let rect = item.sub_rect
                .translate(task_rect.min.to_vector())
                .intersection_unchecked(&task_rect);

            let image_source = ImageSource {
                p0: rect.min.to_f32(),
                p1: rect.max.to_f32(),
                user_data: [0.0; 4],
                uv_rect_kind: task.uv_rect_kind,
            };

            let uv_rect_handle = image_source.write_gpu_blocks(&mut gpu_buffers.f32);

            graph.sub_rects.push(TaskSubRect {
                source_task: item.source_task,
                sub_rect: item.sub_rect,
                uv_address: gpu_buffers.f32.resolve_handle(uv_rect_handle),
            });
        }

        graph
    }
}

impl RenderTaskGraph {
    /// Print the render task graph to console
    #[allow(dead_code)]
    pub fn print(
        &self,
    ) {
        print!("-- RenderTaskGraph --\n");

        for (i, task) in self.tasks.iter().enumerate() {
            print!("Task {} [{}]: render_on={} free_after={} children={:?} target_size={:?}\n",
                i,
                task.kind.as_str(),
                task.render_on.0,
                task.free_after.0,
                task.children,
                task.get_target_size(),
            );
        }

        for (p, pass) in self.passes.iter().enumerate() {
            print!("Pass {}:\n", p);

            for (s, sub_pass) in pass.sub_passes.iter().enumerate() {
                print!("\tSubPass {}: {:?}\n",
                    s,
                    sub_pass.surface,
                );

                for task_id in &sub_pass.task_ids {
                    print!("\t\tTask {:?}\n", task_id.index);
                }
            }
        }
    }

    pub fn resolve_texture(
        &self,
        task_id: impl Into<Option<RenderTaskId>>,
    ) -> Option<TextureSource> {
        let task_id = task_id.into()?;
        let task = &self[task_id];

        match task.get_texture_source() {
            TextureSource::Invalid => None,
            source => Some(source),
        }
    }

    pub fn resolve_location(
        &self,
        task_id: impl Into<Option<RenderTaskId>>,
    ) -> Option<(GpuBufferAddress, TextureSource)> {
        self.resolve_impl(task_id.into()?)
    }

    fn resolve_impl(
        &self,
        task_id: RenderTaskId,
    ) -> Option<(GpuBufferAddress, TextureSource)> {
        let task = &self[task_id];
        let texture_source = task.get_texture_source();

        if let TextureSource::Invalid = texture_source {
            return None;
        }

        let uv_address = if task_id.has_sub_rect() {
            self.sub_rects[task_id.sub_rect_index as usize].uv_address
        } else {
            task.get_texture_address()
        };

        assert!(uv_address.is_valid());

        Some((uv_address, texture_source))
    }

    pub fn report_memory(&self) -> usize {
        // We can't use wr_malloc_sizeof here because the render task
        // graph's memory is mainly backed by frame's custom allocator.
        // So we calulate the memory footprint manually.

        let mut mem = size_of_frame_vec(&self.tasks)
            +  size_of_frame_vec(&self.task_data)
            +  size_of_frame_vec(&self.passes);

        for pass in &self.passes {
            mem += size_of_frame_vec(&pass.task_ids)
                + size_of_frame_vec(&pass.sub_passes)
                + size_of_frame_vec(&pass.textures_to_invalidate);
            for sub_pass in &pass.sub_passes {
                mem += size_of_frame_vec(&sub_pass.task_ids);
            }
        }

        mem
    }

    #[cfg(test)]
    pub fn new_for_testing() -> Self {
        let allocator = FrameAllocator::fallback();
        RenderTaskGraph {
            tasks: allocator.clone().new_vec(),
            sub_tasks: allocator.clone().new_vec(),
            sub_rects: allocator.clone().new_vec(),
            passes: allocator.clone().new_vec(),
            frame_id: FrameId::INVALID,
            task_data: allocator.clone().new_vec(),
            surface_count: 0,
            unique_surfaces: FastHashSet::default(),
        }
    }

    /// Return the surface and texture counts, used for testing
    #[cfg(test)]
    pub fn surface_counts(&self) -> (usize, usize) {
        (self.surface_count, self.unique_surfaces.len())
    }

    /// Return current frame id, used for validation
    #[cfg(debug_assertions)]
    pub fn frame_id(&self) -> FrameId {
        self.frame_id
    }
}

/// Batching uses index access to read information about tasks
impl std::ops::Index<RenderTaskId> for RenderTaskGraph {
    type Output = RenderTask;
    fn index(&self, id: RenderTaskId) -> &RenderTask {
        &self.tasks[id.index as usize]
    }
}

impl std::ops::Index<SubTaskId> for RenderTaskGraph {
    type Output = SubTask;
    fn index(&self, id: SubTaskId) -> &SubTask {
        &self.sub_tasks[id.0 as usize]
    }
}

fn assign_free_pass(
    id: RenderTaskId,
    graph: &mut RenderTaskGraph,
) {
    let task = &mut graph.tasks[id.index as usize];
    let render_on = task.render_on;

    let mut child_task_ids: SmallVec<[RenderTaskId; 8]> = SmallVec::new();
    child_task_ids.extend_from_slice(&task.children);

    for child_id in child_task_ids {
        let child_location = graph.tasks[child_id.index as usize].location.clone();

        // Each dynamic child task can free its backing surface after the last
        // task that references it as an input. Using min here ensures the
        // safe time to free this surface in the presence of multiple paths
        // to this task from the root(s).
        match child_location {
            RenderTaskLocation::CacheRequest { .. } => {}
            RenderTaskLocation::Static { .. } => {
                // never get freed anyway, so can leave untouched
                // (could validate that they remain at PassId::MIN)
            }
            RenderTaskLocation::Dynamic { .. } => {
                panic!("bug: should not be allocated yet");
            }
            RenderTaskLocation::Unallocated { .. } |
            RenderTaskLocation::Existing { .. } => {
                let child_task = &mut graph.tasks[child_id.index as usize];
                child_task.free_after = child_task.free_after.min(render_on);
            }
        }
    }
}

/// A render pass represents a set of rendering operations that don't depend on one
/// another.
///
/// A render pass can have several render targets if there wasn't enough space in one
/// target to do all of the rendering for that pass. See `RenderTargetList`.
#[cfg_attr(feature = "capture", derive(Serialize))]
#[cfg_attr(feature = "replay", derive(Deserialize))]
pub struct RenderPass {
    /// The subpasses that describe targets being rendered to in this pass
    pub alpha: RenderTargetList,
    pub color: RenderTargetList,
    pub texture_cache: FastHashMap<CacheTextureId, RenderTarget>,
    pub picture_cache: FrameVec<PictureCacheTarget>,
    pub textures_to_invalidate: FrameVec<CacheTextureId>,
}

impl RenderPass {
    /// Creates an intermediate off-screen pass.
    pub fn new(src: &Pass, memory: &mut FrameMemory) -> Self {
        RenderPass {
            color: RenderTargetList::new(memory.allocator()),
            alpha: RenderTargetList::new(memory.allocator()),
            texture_cache: FastHashMap::default(),
            picture_cache: memory.allocator().new_vec(),
            textures_to_invalidate: src.textures_to_invalidate.clone(),
        }
    }
}

// Dump an SVG visualization of the render graph for debugging purposes
#[cfg(feature = "capture")]
pub fn dump_render_tasks_as_svg(
    render_tasks: &RenderTaskGraph,
    output: &mut dyn std::io::Write,
) -> std::io::Result<()> {
    use svg_fmt::*;

    let node_width = 80.0;
    let node_height = 30.0;
    let vertical_spacing = 8.0;
    let horizontal_spacing = 20.0;
    let margin = 10.0;
    let text_size = 10.0;

    let mut pass_rects = Vec::new();
    let mut nodes = vec![None; render_tasks.tasks.len()];

    let mut x = margin;
    let mut max_y: f32 = 0.0;

    #[derive(Clone)]
    struct Node {
        rect: Rectangle,
        label: Text,
        size: Text,
    }

    for pass in render_tasks.passes.iter().rev() {
        let mut layout = VerticalLayout::new(x, margin, node_width);

        for task_id in &pass.task_ids {
            let task_index = task_id.index as usize;
            let task = &render_tasks.tasks[task_index];

            let rect = layout.push_rectangle(node_height);

            let tx = rect.x + rect.w / 2.0;
            let ty = rect.y + 10.0;

            let label = text(tx, ty, format!("{}", task.kind.as_str()));
            let size = text(tx, ty + 12.0, format!("{:?}", task.location.size()));

            nodes[task_index] = Some(Node { rect, label, size });

            layout.advance(vertical_spacing);
        }

        pass_rects.push(layout.total_rectangle());

        x += node_width + horizontal_spacing;
        max_y = max_y.max(layout.y + margin);
    }

    let mut links = Vec::new();
    for node_index in 0..nodes.len() {
        if nodes[node_index].is_none() {
            continue;
        }

        let task = &render_tasks.tasks[node_index];
        for dep in &task.children {
            let dep_index = dep.index as usize;

            if let (&Some(ref node), &Some(ref dep_node)) = (&nodes[node_index], ;&nodes[dep_index]) {
                links.push((
                    dep_node.rect.x + dep_node.rect.w,
                    dep_node.rect.y + dep_node.rect.h / 2.0,
                    node.rect.x,
                    node.rect.y + node.rect.h / 2.0,
                ));
            }
        }
    }

    let svg_w = x + margin;
    let svg_h = max_y + margin;
    writeln!(output, "{}", BeginSvg { w: svg_w, h: svg_h })?;

    // Background.
    writeln!(output,
        "    {}",
        rectangle(0.0, 0.0, svg_w, svg_h)
            .inflate(1.0, 1.0)
            .fill(rgb(50, 50, 50))
    )?;

    // Passes.
    for rect in pass_rects {
        writeln!(output,
            "    {}",
            rect.inflate(3.0, 3.0)
                .border_radius(4.0)
                .opacity(0.4)
                .fill(black())
        )?;
    }

    // Links.
    for (x1, y1, x2, y2) in links {
        dump_task_dependency_link(output, x1, y1, x2, y2);
    }

    // Tasks.
    for node in &nodes {
        if let Some(node) = node {
            writeln!(output,
                "    {}",
                node.rect
                    .clone()
                    .fill(black())
                    .border_radius(3.0)
                    .opacity(0.5)
                    .offset(0.0, 2.0)
            )?;
            writeln!(output,
                "    {}",
                node.rect
                    .clone()
                    .fill(rgb(200, 200, 200))
                    .border_radius(3.0)
                    .opacity(0.8)
            )?;

            writeln!(output,
                "    {}",
                node.label
                    .clone()
                    .size(text_size)
                    .align(Align::Center)
                    .color(rgb(50, 50, 50))
            )?;
            writeln!(output,
                "    {}",
                node.size
                    .clone()
                    .size(text_size * 0.7)
                    .align(Align::Center)
                    .color(rgb(50, 50, 50))
            )?;
        }
    }

    writeln!(output, "{}", EndSvg)
}

#[allow(dead_code)]
fn dump_task_dependency_link(
    output: &mut dyn std::io::Write,
    x1: f32, y1: f32,
    x2: f32, y2: f32,
) {
    use svg_fmt::*;

    // If the link is a straight horizontal line and spans over multiple passes, it
    // is likely to go straight though unrelated nodes in a way that makes it look like
    // they are connected, so we bend the line upward a bit to avoid that.
    let simple_path = (y1 - y2).abs() > 1.0 || (x2 - x1) < 45.0;

    let mid_x = (x1 + x2) / 2.0;
    if simple_path {
        write!(output, "    {}",
            path().move_to(x1, y1)
                .cubic_bezier_to(mid_x, y1, mid_x, y2, x2, y2)
                .fill(Fill::None)
                .stroke(Stroke::Color(rgb(100, 100, 100), 3.0))
        ).unwrap();
    } else {
        let ctrl1_x = (mid_x + x1) / 2.0;
        let ctrl2_x = (mid_x + x2) / 2.0;
        let ctrl_y = y1 - 25.0;
        write!(output, "    {}",
            path().move_to(x1, y1)
                .cubic_bezier_to(ctrl1_x, y1, ctrl1_x, ctrl_y, mid_x, ctrl_y)
                .cubic_bezier_to(ctrl2_x, ctrl_y, ctrl2_x, y2, x2, y2)
                .fill(Fill::None)
                .stroke(Stroke::Color(rgb(100, 100, 100), 3.0))
        ).unwrap();
    }
}

/// Construct a picture cache render task location for testing
#[cfg(test)]
fn pc_target(
    surface_id: u64,
    tile_x: i32,
    tile_y: i32,
) -> RenderTaskLocation {
    use crate::{
        composite::{NativeSurfaceId, NativeTileId},
        picture::ResolvedSurfaceTexture,
    };

    let width = 512;
    let height = 512;

    RenderTaskLocation::Static {
        surface: StaticRenderTaskSurface::PictureCache {
            surface: ResolvedSurfaceTexture::Native {
                id: NativeTileId {
                    surface_id: NativeSurfaceId(surface_id),
                    x: tile_x,
                    y: tile_y,
                },
                size: DeviceIntSize::new(width, height),
            },
        },
        rect: DeviceIntSize::new(width, height).into(),
    }
}

#[cfg(test)]
impl RenderTaskGraphBuilder {
    fn test_expect(
        mut self,
        pass_count: usize,
        total_surface_count: usize,
        unique_surfaces: &[(i32, i32, ImageFormat)],
    ) {
        use crate::{internal_types::FrameStamp, renderer::{GpuBufferBuilderF, GpuBufferBuilderI}};
        use api::{DocumentId, IdNamespace};

        let mut rc = ResourceCache::new_for_testing();

        let mut frame_stamp = FrameStamp::first(DocumentId::new(IdNamespace(1), 1));
        frame_stamp.advance();

        let frame_memory = FrameMemory::fallback();
        let mut gpu_buffers = GpuBufferBuilder {
            f32: GpuBufferBuilderF::new(&frame_memory, 0, FrameId::first()),
            i32: GpuBufferBuilderI::new(&frame_memory, 0, FrameId::first()),
        };
        let g = self.end_frame(&mut rc, &mut gpu_buffers, &mut frame_memory.new_vec(), 2048, &frame_memory);
        g.print();

        assert_eq!(g.passes.len(), pass_count);
        assert_eq!(g.surface_counts(), (total_surface_count, unique_surfaces.len()));

        rc.validate_surfaces(unique_surfaces);
    }
}

/// Construct a testing render task with given location
#[cfg(test)]
fn task_location(location: RenderTaskLocation) -> RenderTask {
    RenderTask::new_test(
        location,
        RenderTargetKind::Color,
    )
}

/// Construct a dynamic render task location for testing
#[cfg(test)]
fn task_dynamic(size: i32) -> RenderTask {
    RenderTask::new_test(
        RenderTaskLocation::Unallocated { size: DeviceIntSize::new(size, size) },
        RenderTargetKind::Color,
    )
}

#[test]
fn fg_test_1() {
    // Test that a root target can be used as an input for readbacks
    // This functionality isn't currently used, but will be in future.

    let mut gb = RenderTaskGraphBuilder::new();

    let root_target = pc_target(0, 0, 0);

    let root = gb.add().init(task_location(root_target.clone()));

    let readback = gb.add().init(task_dynamic(100));
    gb.add_dependency(readback, root);

    let mix_blend_content = gb.add().init(task_dynamic(50));

    let content = gb.add().init(task_location(root_target));
    gb.add_dependency(content, readback);
    gb.add_dependency(content, mix_blend_content);

    gb.test_expect(3, 1, &[
        (2048, 2048, ImageFormat::RGBA8),
    ]);
}

#[test]
fn fg_test_3() {
    // Test that small targets are allocated in a shared surface, and that large
    // tasks are allocated in a rounded up texture size.

    let mut gb = RenderTaskGraphBuilder::new();

    let pc_root = gb.add().init(task_location(pc_target(0, 0, 0)));

    let child_pic_0 = gb.add().init(task_dynamic(128));
    let child_pic_1 = gb.add().init(task_dynamic(3000));

    gb.add_dependency(pc_root, child_pic_0);
    gb.add_dependency(pc_root, child_pic_1);

    gb.test_expect(2, 2, &[
        (2048, 2048, ImageFormat::RGBA8),
        (3072, 3072, ImageFormat::RGBA8),
    ]);
}

#[test]
fn fg_test_4() {
    // Test that for a simple dependency chain of tasks, that render
    // target surfaces are aliased and reused between passes where possible.

    let mut gb = RenderTaskGraphBuilder::new();

    let pc_root = gb.add().init(task_location(pc_target(0, 0, 0)));

    let child_pic_0 = gb.add().init(task_dynamic(128));
    let child_pic_1 = gb.add().init(task_dynamic(128));
    let child_pic_2 = gb.add().init(task_dynamic(128));

    gb.add_dependency(pc_root, child_pic_0);
    gb.add_dependency(child_pic_0, child_pic_1);
    gb.add_dependency(child_pic_1, child_pic_2);

    gb.test_expect(4, 3, &[
        (2048, 2048, ImageFormat::RGBA8),
        (2048, 2048, ImageFormat::RGBA8),
    ]);
}

#[test]
fn fg_test_5() {
    // Test that a task that is used as an input by direct parent and also
    // distance ancestor are scheduled correctly, and allocates the correct
    // number of passes, taking advantage of surface reuse / aliasing where feasible.

    let mut gb = RenderTaskGraphBuilder::new();

    let pc_root = gb.add().init(task_location(pc_target(0, 0, 0)));

    let child_pic_0 = gb.add().init(task_dynamic(128));
    let child_pic_1 = gb.add().init(task_dynamic(64));
    let child_pic_2 = gb.add().init(task_dynamic(32));
    let child_pic_3 = gb.add().init(task_dynamic(16));

    gb.add_dependency(pc_root, child_pic_0);
    gb.add_dependency(child_pic_0, child_pic_1);
    gb.add_dependency(child_pic_1, child_pic_2);
    gb.add_dependency(child_pic_2, child_pic_3);
    gb.add_dependency(pc_root, child_pic_3);

    gb.test_expect(5, 4, &[
        (2048, 2048, ImageFormat::RGBA8),
        (2048, 2048, ImageFormat::RGBA8),
        (2048, 2048, ImageFormat::RGBA8),
    ]);
}

#[test]
fn fg_test_6() {
    // Test that a task that is used as an input dependency by two parent
    // tasks is correctly allocated and freed.

    let mut gb = RenderTaskGraphBuilder::new();

    let pc_root_1 = gb.add().init(task_location(pc_target(0, 0, 0)));
    let pc_root_2 = gb.add().init(task_location(pc_target(0, 1, 0)));

    let child_pic = gb.add().init(task_dynamic(128));

    gb.add_dependency(pc_root_1, child_pic);
    gb.add_dependency(pc_root_2, child_pic);

    gb.test_expect(2, 1, &[
        (2048, 2048, ImageFormat::RGBA8),
    ]);
}

#[test]
fn fg_test_7() {
    // Test that a standalone surface is not incorrectly used to
    // allocate subsequent shared task rects.

    let mut gb = RenderTaskGraphBuilder::new();

    let pc_root = gb.add().init(task_location(pc_target(0, 0, 0)));

    let child0 = gb.add().init(task_dynamic(16));
    let child1 = gb.add().init(task_dynamic(16));

    let child2 = gb.add().init(task_dynamic(16));
    let child3 = gb.add().init(task_dynamic(16));

    gb.add_dependency(pc_root, child0);
    gb.add_dependency(child0, child1);
    gb.add_dependency(pc_root, child1);

    gb.add_dependency(pc_root, child2);
    gb.add_dependency(child2, child3);

    gb.test_expect(3, 3, &[
        (2048, 2048, ImageFormat::RGBA8),
        (2048, 2048, ImageFormat::RGBA8),
        (2048, 2048, ImageFormat::RGBA8),
    ]);
}