99 lines
4.6 KiB
Text
99 lines
4.6 KiB
Text
|
The official guide to swscale for confused developers.
|
||
|
|
========================================================
|
||
|
|
|
||
|
|
Current (simplified) Architecture:
|
||
|
|
---------------------------------
|
||
|
|
Input
|
||
|
|
v
|
||
|
|
_______OR_________
|
||
|
|
/ \
|
||
|
|
/ \
|
||
|
|
special converter [Input to YUV converter]
|
||
|
|
| |
|
||
|
|
| (8-bit YUV 4:4:4 / 4:2:2 / 4:2:0 / 4:0:0 )
|
||
|
|
| |
|
||
|
|
| v
|
||
|
|
| Horizontal scaler
|
||
|
|
| |
|
||
|
|
| (15-bit YUV 4:4:4 / 4:2:2 / 4:2:0 / 4:1:1 / 4:0:0 )
|
||
|
|
| |
|
||
|
|
| v
|
||
|
|
| Vertical scaler and output converter
|
||
|
|
| |
|
||
|
|
v v
|
||
|
|
output
|
||
|
|
|
||
|
|
|
||
|
|
Swscale has 2 scaler paths. Each side must be capable of handling
|
||
|
|
slices, that is, consecutive non-overlapping rectangles of dimension
|
||
|
|
(0,slice_top) - (picture_width, slice_bottom).
|
||
|
|
|
||
|
|
special converter
|
||
|
|
These generally are unscaled converters of common
|
||
|
|
formats, like YUV 4:2:0/4:2:2 -> RGB12/15/16/24/32. Though it could also
|
||
|
|
in principle contain scalers optimized for specific common cases.
|
||
|
|
|
||
|
|
Main path
|
||
|
|
The main path is used when no special converter can be used. The code
|
||
|
|
is designed as a destination line pull architecture. That is, for each
|
||
|
|
output line the vertical scaler pulls lines from a ring buffer. When
|
||
|
|
the ring buffer does not contain the wanted line, then it is pulled from
|
||
|
|
the input slice through the input converter and horizontal scaler.
|
||
|
|
The result is also stored in the ring buffer to serve future vertical
|
||
|
|
scaler requests.
|
||
|
|
When no more output can be generated because lines from a future slice
|
||
|
|
would be needed, then all remaining lines in the current slice are
|
||
|
|
converted, horizontally scaled and put in the ring buffer.
|
||
|
|
[This is done for luma and chroma, each with possibly different numbers
|
||
|
|
of lines per picture.]
|
||
|
|
|
||
|
|
Input to YUV Converter
|
||
|
|
When the input to the main path is not planar 8 bits per component YUV or
|
||
|
|
8-bit gray, it is converted to planar 8-bit YUV. Two sets of converters
|
||
|
|
exist for this currently: One performs horizontal downscaling by 2
|
||
|
|
before the conversion, the other leaves the full chroma resolution,
|
||
|
|
but is slightly slower. The scaler will try to preserve full chroma
|
||
|
|
when the output uses it. It is possible to force full chroma with
|
||
|
|
SWS_FULL_CHR_H_INP even for cases where the scaler thinks it is useless.
|
||
|
|
|
||
|
|
Horizontal scaler
|
||
|
|
There are several horizontal scalers. A special case worth mentioning is
|
||
|
|
the fast bilinear scaler that is made of runtime-generated MMXEXT code
|
||
|
|
using specially tuned pshufw instructions.
|
||
|
|
The remaining scalers are specially-tuned for various filter lengths.
|
||
|
|
They scale 8-bit unsigned planar data to 16-bit signed planar data.
|
||
|
|
Future >8 bits per component inputs will need to add a new horizontal
|
||
|
|
scaler that preserves the input precision.
|
||
|
|
|
||
|
|
Vertical scaler and output converter
|
||
|
|
There is a large number of combined vertical scalers + output converters.
|
||
|
|
Some are:
|
||
|
|
* unscaled output converters
|
||
|
|
* unscaled output converters that average 2 chroma lines
|
||
|
|
* bilinear converters (C, MMX and accurate MMX)
|
||
|
|
* arbitrary filter length converters (C, MMX and accurate MMX)
|
||
|
|
And
|
||
|
|
* Plain C 8-bit 4:2:2 YUV -> RGB converters using LUTs
|
||
|
|
* Plain C 17-bit 4:4:4 YUV -> RGB converters using multiplies
|
||
|
|
* MMX 11-bit 4:2:2 YUV -> RGB converters
|
||
|
|
* Plain C 16-bit Y -> 16-bit gray
|
||
|
|
...
|
||
|
|
|
||
|
|
RGB with less than 8 bits per component uses dither to improve the
|
||
|
|
subjective quality and low-frequency accuracy.
|
||
|
|
|
||
|
|
|
||
|
|
Filter coefficients:
|
||
|
|
--------------------
|
||
|
|
There are several different scalers (bilinear, bicubic, lanczos, area,
|
||
|
|
sinc, ...). Their coefficients are calculated in initFilter().
|
||
|
|
Horizontal filter coefficients have a 1.0 point at 1 << 14, vertical ones at
|
||
|
|
1 << 12. The 1.0 points have been chosen to maximize precision while leaving
|
||
|
|
a little headroom for convolutional filters like sharpening filters and
|
||
|
|
minimizing SIMD instructions needed to apply them.
|
||
|
|
It would be trivial to use a different 1.0 point if some specific scaler
|
||
|
|
would benefit from it.
|
||
|
|
Also, as already hinted at, initFilter() accepts an optional convolutional
|
||
|
|
filter as input that can be used for contrast, saturation, blur, sharpening
|
||
|
|
shift, chroma vs. luma shift, ...
|