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.
One drawback of the current implementation is that the GPU device can be
unavailable at the time of the first pattern fill, which causes the
GPU-accelerated canvas to be move on the main thread because of putImageData.
Most of the shading patterns stuff will be moved to the GPU and in order
to avoid creating some useless data we've to know if the GPU is available or not.
So in this patch we create the GPU device during the worker initialization
and pass a flag to the evaluator to know if the GPU is available or not.
These classes, and various related code, became unused after PR 21023 with only unit-tests actually running that code now.
Also removes the `isEvalSupported` API option, since the `PostScriptCompiler` was the only remaining code where `eval` was used.
Currently we have no less than three different, but very similar, factories for reading built-in CMap files, standard font files, and wasm files on the main-thread.[1]
These factories were added at different points in time, since I cannot imagine that we'd add essentially three copies of the same code otherwise.
Nowadays these factories are often not even used[2], since worker-thread fetching is used whenever possible to improve performance. In particular, they will *only* be used when either:
- The PDF.js library runs in Node.js environments.
- The user manually sets `useWorkerFetch = false` when calling `getDocument`.
- The user provides custom `CMapReaderFactory`, `StandardFontDataFactory`, and/or `WasmFactory` instances when calling `getDocument`.
By replacing these factories with *a single* new `BinaryDataFactory` factory/option the number of `getDocument` options are thus reduced, which cannot hurt.
This also reduces the total bundle-size of the Firefox PDF Viewer a little bit, and it slightly reduces the number of import maps that need to be maintained.
*Please note:* For users that provide custom `CMapReaderFactory`, `StandardFontDataFactory`, and `WasmFactory` instances when calling `getDocument` this will be a breaking change, however it's unlikely that (many) such users exist.
(The *internal* format data-format of `CMapReaderFactory` was changed in PR 18951, and there hasn't been a single question/complaint about it in well over a year.)
---
[1] Any new functionality could easily lead to more such factories being added in the future, which wouldn't be great.
[2] Note that the Firefox PDF Viewer no longer use these factories, since it "forcibly" sets `useWorkerFetch = true` during building.
The `BaseCMapReaderFactory`, `BaseStandardFontDataFactory`, and `BaseWasmFactory` classes are all very similar, and the only difference is really in their respective `fetch` methods.
By have the worker-thread "compute" the complete `filename` it's possible to simplify the `BaseCMapReaderFactory.prototype.fetch` method, which will allow future improvements to all of these classes.
A couple of things to note:
- This code is unused, and it's not even bundled, in the Firefox PDF Viewer.
- In browsers it's unused by default, and worker-thread fetching will always be used when possible since that's more efficient.
*Please note:* For users that provide a custom `CMapReaderFactory` instance when calling `getDocument` this could be a breaking change, however it's unlikely that any such users exist.
(The *internal* format of this data was changed previously in PR 18951, and there hasn't been a single question/complaint about it in well over a year.)
The WebGPU feature hasn't been released yet but it's interesting to see how
we can use it in order to speed up the rendering of some objects.
This patch allows to render mesh patterns using WebGPU.
I didn't see any significant performance improvement on my machine (mac M2)
but it may be different on other platforms.
Given that we "forcibly" set `useWorkerFetch = true` for the MOZCENTRAL build-target there's a small amount of dead code as a result, which we can thus remove during building.
After the changes in PR 20197 the code in the `TranslatedFont.prototype.send` method is not all that readable[1] given how it handles e.g. the `charProcOperatorList` data used with Type3 fonts.
Since this is the only spot where `Font.prototype.exportData` is used, it seems much simpler to move the `compileFontInfo` call there and *directly* return the intended data rather than messing with it after the fact.
Finally, while it doesn't really matter, the patch flips the order of the `charProcOperatorList` and `extra` properties throughout the code-base since the former is used with Type3 fonts while the latter (effectively) requires that debugging is enabled.
---
[1] I had to re-read it twice, also looking at all the involved methods, in order to convince myself that it's actually correct.
On the worker-thread only the static `write` methods are actually used, and on the main-thread only class instances are being created.
Hence this, after PR 20197, leads to a bunch of dead code in both of the *built* `pdf.mjs` and `pdf.worker.js` files.
This patch reduces the size of the `gulp mozcentral` output by `21 419` bytes, i.e. `21` kilo-bytes, which I believe is way too large of a saving to not do this.
(I can't even remember the last time we managed to reduce build-size this much with a single patch.)
There's a number of classes where the constructors can be removed completely by instead using class fields, which help to slightly shorten the code.
It seems that `unicorn/prefer-class-fields` ESLint plugin, see PR 20657, unfortunately isn't able to detect all of these cases.
We tried to lookup the font metrics using the font name as-is, which didn't work since the PDF file in question has non-embedded fonts with names that include commas.
Hence the font names need to be normalized here as well, similar to elsewhere in the font code.
Usually, content stream or fonts are compressed using FlateDecode.
So use the DecompressionStream API to decompress those streams
in the async code path.
Follow up on https://github.com/mozilla/pdf.js/pull/20197,
This serializes pattern data into an ArrayBuffer which is
then transferred from the worker to the main thread.
It sets up the stage for us to eventually switch to a
SharedArrayBuffer in the future.
This PR serializes font data into an ArrayBuffer
that is then transfered from the worker to the
main thread. It's more efficient than the current
solution which clones the "export data" object
which includes the font data as a Uint8Array.
It prepares us to switch to a SharedArrayBuffer
in the future, which would allow us to share
the font data with multiple agents, which would be
crucial for the upcoming "renderer" worker.
We have a fallback for the common case of Type3 fonts without a /FontDescriptor dictionary, however we also need to handle the case where it's present but lacking the required /FontName entry.
For simplicity we will abort /Form XObject parsing *immediately* when encountering a circular reference, rather than letting it continue up until some limit (as e.g. PDFium appears to do), which should be fine since there are never any guarantees if/how *corrupt* PDF documents will render.
After the introduction of `OffscreenCanvas` support we now have *two separate* mask-methods in the `PDFImage` class, and the reason that they were not combined is likely that we need the "raw" bytes when parsing Type3-glyph image masks.
However, that case is easy to support simply by disabling `OffscreenCanvas` usage when parsing Type3-glyphs and that way we're able to reduce some code duplication.
Another slightly strange property of the `PDFImage.createMask` method is that it needs various image-dictionary parameters *manually* provided, which is probably because this is very old code.
That feels slightly unwieldy, and we instead change the method to pass in the image-stream directly and do the necessary data-lookup internally.
A side-effect of this re-factoring is that we now support using the custom `isSingleOpaquePixel` operator in Type3-glyphs, which shouldn't hurt even though it seems extremely unlikely for that to ever happen in Type3-glyphs.
