This prepares allowing to iterate over a TIFF file with invalid images
while `Image` maintains invariants. For instance we precalculate the
amount of buffer space and verify that the count of strips makes sense
in the constructor—so we do not have to redo that for every other entry
point.
Do we even want to error early here? This isn't exactly fatal to reading
the image it is a only nuisance. It becomes crucial only when we are at
the point of reading all the tiles themselves.
Two tests are disabled, both related to tiles, since they do not follow
the fully released spec themselves but incidentally decode with libtiff
anyways. The two images (`cramps-tile.tif`, `quad-tile.tif`) are tiled
layout which libtiff detects by the presence of TileWidth/TileHeight
tags. We would detect that by chunks encoded via TileOffsets and
TileByteLengths tags. However the images encoded their chunk data into
the StripOffsets tag; additionally they contain a RowsPerStrip value
anyways.
Tests disabled as this is to be resolved in a separate issue (#338) with
a decision on treatment of this situation (strictness setting or copying
libtiff, also error late for mismatched lengths).
This complements the entry API of the encoder with the byte slice based
method. Note we don't provide it for writing raw data, here a
`std::io::Write` would be more appropriate and we should not attempt to
make it `write_all`, nor compress. You'd get all the benefits of the
other API through `[u8]: TiffValue` since that interface is otherwise
type agnostic and `u8` specifically is endian agnostic.
We have not released this yet, let's align these names before it is too
late. The decoder adopted the `_buf` suffix and I find it more
convincing than the more generic `_value` (that also confusingly shares
it with `TiffValue`).
Implement the encoder's `TiffValue` trait for the tags commonly defined
and represented as separate types in our own code. We don't tie the
tag for an entry of these to the value type (yet). This lets us keep the
trait exactly as-is and we avoid unnecessary churn. We can always add a
separate trait on top.
Seems they were introduced with BigTiff but since the enumeration of the
decoder had no obligation of introducing discriminant values the list
was not cross-referenced to libtiff. Hence, some rational types exist
with no basis in decoding.
We have a fundamental choice here: We could add more proxy
implementations for `Vec` and `Box` to follow `&'_ T`, for some types
that indirections is clear. Vec and arrays are also both very natural
ways of expressing a repeated value. However, passing an allocated value
to the interface forces us to drop it inside. We might be able to
utilize the allocation for an in-place byteswap in a future increment
but not with the current design.
So, overall, I think that inherently `Drop`-y types should be avoided in
these implementations. Instead, we can use the amazing diagnostic tools
in the compiler to just help the users navigate the type system. We can
still provide the impl for arrays.
For context, see: https://users.rust-lang.org/t/what-type-is-u16/137395
Now they see:
|
38 | .write_tag(tiff::tags::Tag::GeoKeyDirectoryTag, &geokey);
| ^^^^^^ the trait `TiffValue` is not implemented for `&Vec<u16>`
|
= note: the trait is implemented for primitive types (`u8`, `i16`, `f32`, etc.)
= note: the trait is implemented for shared references to values
= note: the trait is implemented for slices, pass them by reference (e.g. `&[u8]`)
= note: the trait is implemented for arrays if it is implemented for a slice
= note: values in a `Vec` or `Box` should be dereferenced, e.g. `&vec[..]`
= help: the following other types implement trait `TiffValue`:
Previously the compiler would both strip the `&` in its first message
(due to our forwarding impl) and not list any way of getting a
reference-to-slice.
I didn't realize we support this in a slightly weird manner where the
tag is completely ignored and zune-jpeg performs the upsampling as
indicated in the JPEG stream. This is a recurring problem with the
compression scheme and tags in TIFF. In the end, decoder can only handle
the data as indicated by the JPEG steam and while the specification
requires the tags to be consistent with the underlying data, sometimes
they are not. The useful fallback is of course to ignore the tag.
Ignoring the tag of course buts us in a weird spot for zero-copy
decoding since the buffer indication to the caller will be incorrect.
For now, maybe the best way of treating subsampling is to upsample all
the cases when we add support.
In particular we avoid mistaking the absence of this tag for an absence
of subsampling. The specification instead requires it to mean half the
samples in both directions—a mode that we do not support yet. Instead of
returning erroneous samples or an error in the data stream (which would
appear too short) we return an error when asked to layout the read
buffer for these images.
We previously handed it a wrapper that was supposedly type-safe but
really internally cast the buffer to bytes. That was rather pointless as
the `_bytes` method could always be used instead. We replace the
semantics by a utility which allocates an appropriate buffer with the
internal limits, wrapping the exact sequence identified previously in
the example as a pattern.
By adding the type interpretation to `BufferLayoutPreference` (simply
computed from the same fields that are trivial to fetch in
`ReadoutLayout` with the rest) we can allocate an appropriate
`DecoderResult` without any `Decoder` instance. We still want a Limits
instance though. This lets a caller read _typed_ data from the `_bytes`
methods which perform less allocations. It also lets them read multiple
planes in a more convenient interface. See, `read_image` will keep its
planar treatment bug for a little bit so you don't accidentally break
during an upgrade.
This also implies less code is being monomorphized, just a bonus.
This is well-defined even if we add sub-sampling support later on. We
can adjust the layout type accordingly with factors per each plane. Note
that the layout, in theory, does not depend on the index of the unit if
we describe all planes for a multi-planar layout. The specification
requires that the subsampling factors divide the tile sizes. However we
have to handle non-planar images here and this leaves us open to other
potential parameters.
We currently have one type in decoder/image.rs and another in
decoder/mod.rs which each contributes part of the computation for laying
out the output buffer. Confusingly though they rely on shared
implementation details such as determining the number of channels that
are written and their bit depth. We should have one shared method and
resulting struct instead.
These samples are normal bits in the encoded (e.g. compressed) stream of
each tile or row of the image. However, when considering the photometric
interpretation of an image these samples should be discarded. They are
only relevant for multiband.
For the most part we previously rejected these images due to a mismatch
between sample count and the color as indicated by the interpretation.
Only, for alpha we never really verified if the extra channel is indeed
alpha or something unrelated. We don't have a perfect solution as
GeoTiff motivates us having a MultiBand pseudo-color output
configuration that has no related photometric interpretation and borrows
the usual gray tone instead. We can only really distinguish this from a
gray-alpha image through looking at the values of the ExtraSamples
vector. Then again we do not yet support GrayAlpha anyways.
As a simplification for the discarding process we only support full-byte
extra samples and require all samples to be the same bit-depth, not only
those which we return. This is blessed by Tiff's recommendations for
baseline decoding as well as libtiff.
The idea is to allow codegen to exploit specific known constants. It
seems that LLVM by itself will not generate split loop blocks by an
input argument that is a loop constant. Whereas we know, from PNG, that
some constants yield much better code. In this case
coef[i] = coef[i] + coef[i - n]
This has a destructive dependency chain for `n = 1`, should never get an
argument of 0 and is almost embarrassingly parallel in SIMD if n >= 8
where we can increase the amount of data loaded at once for each
independent loop iteration.
By splitting the loop we make the compiler apply independent
optimization passes to each of case, then have a fallback for things we
did not cover with vectorized possibilities.
A. Molzer