PR #21101 narrowed `compose()`'s `clearRect` from full canvas to the
caller-supplied dirty box. That leaves pixels outside the current
dirty box on the SMask scratch canvas between `compose()` calls;
subsequent draws into scratch are then source-over-blended on top of
those leftovers, so the output depends on the cumulative draw history
rather than just the current draw.
It fixes#18032.
Only use the special inner-backdrop compositing path for nested non-isolated groups that actually need isolation.
This preserves the parent/page backdrop for simple nested groups inside knockout groups, preventing later group
compositing from erasing existing backdrop content.
In a knockout (KO) group each painting operator ("element") composites against
the group's initial backdrop instead of accumulating onto prior elements
of the same group. The backend renders each element to a per-group pooled
temp canvas (keyed off `#groupStackMeta`), builds a binary alpha mask via
a new `feFuncA` filter (`addKnockoutFilter`), `destination-out`s the
group canvas through that mask, restores the initial backdrop into the
cleared footprint for non-isolated groups (cropped to the same mask so
sparse groups don't bleed the whole rectangle), and finally paints the
element on top with the parent's blend mode. Path / clip / transform ops
are mirrored back to the group canvas via `mirrorContextOperations` so
graphics state stays in sync between elements; only the raster pixels
land on the temp canvas.
The temp canvas is forced to source-over for the element raster (`multiply`
on a transparent backdrop would zero the color) and the original GCO is
restored before `copyCtxState` writes back, so the parent's blend mode
survives for the final composite.
Also handled:
- Nested KO groups (the level is incremented for KO, reset to 0 for
non-KO subgroups so an ancestor KO doesn't leak in).
- Non-isolated non-KO subgroups inside a KO parent (`hasInnerBackdrop`
path: blend the elements against the subgroup's running backdrop for
color, mask with the elements-only canvas).
- Soft masks installed inside a KO element (`applySMaskInPlace` in
`compose`, which runs the SMask destination-in directly on the temp
canvas; the existing blit-to-suspended step is gated by `if (!ctx)`).
- Type-3 text, shading fills, image-mask groups, inline images and the
solid-color mask path: each is wrapped in `#begin/#endKnockoutElement`.
- `endDrawing` cleanup so cancelled rendering doesn't leak pooled
canvases or stale knockout state.
For now OffscreenCanvas in worker threads doesn't support ctx.filter,
so we need to fall back to a more expensive pixel-buffer SMask filtering in that case.
As a side effect, this also allows to support correctly smask in Safari.
Prepare reusable soft-mask canvases for filtered and backdrop-dependent masks,
and use a faster destination-in composition path where possible.
Handle Alpha SMask /BC correctly, preserve OOB alpha behavior, and mirror canvas path
operations needed while rendering inside soft-mask mode (mirrored clip was buggy).
Add reftest PDFs covering Alpha masks, transfer functions, backdrop/OOB
alpha, and the optimized composition paths.
remove as much as possible some intermediate canvases and avoid to use SVG filter
at each composition.
Rendering all the pages of issue17784.pdf takes 2x less time now.
1. Record `fill` dependencies even if we early return due to `isPatternFill``
2. Isolate the `drawPattern` inner `executeOperationList` in a
`CanvasNestedDependencyTracker` so that it does not consume pending
dependencies from the outer list.
Nowadays there's a lot of places in the code-base where we need to initialize or reset bounding boxes. Rather than spelling this out repeatedly, this patch adds new `Array`/`Float32Array` constants that can be copied or used as-is where appropriate.
We try to detect in the worker if some patterns or groups need to be drawn or not in isolation.
When they don't, we just draw them on the main canvas instead of drawing on a new canvas.
A pattern or a group is considered as being in isolation if it has some compositing rules or some transparency.
It improves the rendering performance of the pdf in bug 1731514.
The cache has been added in #3312 in 2013 and a lot of things changed since.
Having too many cached accelerated canvases can lead to have to move their data from the GPU to the RAM
which is costly.
So this patch:
- removes all the cached canvases;
- destroys the useless canvases in order to free their associated memory asap;
- slightly rewrite canvas.js::_scaleImage to avoid too much canvas creation.
This patch adds right-click support for images in the PDF, allowing
users to download them. To minimize memory consumption, we:
- Do not store the images separately, and instead crop them out of the
PDF page canvas
- Only extract the images when needed (i.e. when the user right-clicks
on them), rather than eagery having all of them available.
To do so, we layer one empty 0x0 canvas per image, stretched to cover
the whole image, and only populate its contents on right click.
These images need to be inside the text layer: they cannot be _behind_
it, otherwise they would be covered by the text layer's container and
not be clickable, and they cannot be in front of it, otherwise they
would make the text spans unselectable.
This feature is managed by a new preference, `imagesRightClickMinSize`:
- when it's set to `-1`, right-click support is disabled
- when set to `0`, all images are available for right click
- when set to a positive integer, only images whose width and height are
greater than or equal to that value (in the PDF page frame of
reference) are available for right click.
This features is disabled by default outside of MOZCENTRAL, as it
significantly degrades the text selection experience in non-Firefox
browsers.
With the exception of the first invocation the callback function is unused, which means that a lot of pointless functions may be created.
To avoid this we introduce helper functions for simple cases, such as creating `Map`s and `Objects`s.
We used a SVG string which can be pass to the Path2D ctor but it's a bit slower than
building the path step by step.
Having numerical data instead of a string will help the font data serialization.
The position of the text rendered by `showText` is affected
incrementally by the preceding `showText` operations "on the same line".
For this reason, we keep track of all of them (with their dependencies)
in `sameLineText`.
`sameLineText` can be reset whenever we explicitly position the text
somewhere else. We were previously only doing it for `moveText`, and
this patch updates the code to also do it on `setTextMatrix` (which
resets `this.current.x/y` in `CanvasGraphics` to 0).
The complexity of subsequent `sameLineText` operations dependency
tracking grows quadratically with the number of operations on the same
line, so this patch fixes the performance problem when there are whole
pages of text that only use `setTextMatrix` and not `moveText`.
This PR changes the way we store bounding boxes so that they use less
memory and can be more easily shared across threads in the future.
Instead of storing the bounding box and list of dependencies for each
operation that renders _something_, we now only store the bounding box
of _every_ operation and no dependencies list. The bounding box of
each operation covers the bounding box of all the operations affected
by it that render something. For example, the bounding box of a
`setFont` operation will be the bounding box of all the `showText`
operations that use that font.
This affects the debugging experience in pdfBug, since now the bounding
box of an operation may be larger than what it renders itself. To help
with this, now when hovering on an operation we also highlight (in red)
all its dependents. We highlight with white stripes operations that do
not affect any part of the page (i.e. with an empty bbox).
To save memory, we now save bounding box x/y coordinates as uint8
rather than float64. This effectively gives us a 256x256 uniform grid
that covers the page, which is high enough resolution for the usecase.
This commit is a first step towards #6419, and it can also help with
first compute which ops can affect what is visible in that part of
the page.
This commit adds logic to track operations with their respective
bounding boxes. Only operations that actually cause something to
be rendered have a bounding box and dependencies.
Consider the following example:
```
0. setFillRGBColor
1. beginText
2. showText "Hello"
3. endText
4. constructPath [...] -> eoFill
```
here we have three rendering operations: the showText op (2) and the
path (4). (2) depends on (0), (1) and (3), while (4) only depends on
(0). Both (2) and (4) have a bounding box.
This tracking happens when first rendering a PDF: we then use the
recorded information to optimize future partial renderings of a PDF, so
that we can skip operations that do not affected the PDF area on the
canvas.
All this logic only runs when the new `enableOptimizedPartialRendering`
preference, disabled by default, is enabled.
The bounding boxes and dependencies are also shown in the pdfBug
stepper. When hovering over a step now:
- it highlights the steps that they depend on
- it highlights on the PDF itself the bounding box
It fixes#20065.
The only to get a path (from the path generator) is when the font is embedded.
So when we need a path (disableFontFace: true or when we want to use a pattern for stroking/filling), it's impossible
to fulfil.
It seems that the `@napi-rs/canvas` dependency has *basic* canvas-filter support, whereas the "old" `canvas` dependency didn't, hence we no longer need the Node.js-specific checks in the `src/display/canvas.js` file.
Note that I've successfully tested the [`pdf2png` example](https://github.com/mozilla/pdf.js/tree/master/examples/node/pdf2png) with this patch applied and things appear to work as before.
