README

Short README, perhaps not too useful.

qscale (the "q" is for "quick" -- the confusion with "quantization scale" is
unfortunate, but there are only so many short names in the world) is a fast
JPEG-to-JPEG-only up- and downscaler. On my 1.2GHz Core Duo laptop (using
only one core), qscale is 3-4 times as fast as ImageMagick for downscaling
large JPEGs (~10Mpix from my digital camera) to more moderate-sized JPEGs
for web use etc. (like 640x480) without sacrificing quality. (Benchmarking must
be done with a bit of care, though, in particular due to different subsampling
options possible etc.) Most of the time in qscale's case is used on reading in
the original image using libjpeg, which is shared among the two. However, it
would probably not be realistic to exclude the libjpeg time, as most (although
not all) real-world scaling tasks would indeed need to read and decode the
source JPEG.

Note: This is not to denounce ImageMagick in any way. It is a fine library,
capable of doing much more than qscale can ever hope to do. Comparison between
the two are mainly to provide a well-known reference, and to demonstrate that
more specific tools than usually be made faster than generic tools.

qscale is not novel in any way, nor is it perfect (far from it; it's more like
a proof of concept) -- it is mainly a piece of engineering.
However, the following techniques deserve some kind of mention:

 - qscale recognizes that JPEGs are usually stored in the YCbCr colorspace and
   not RGB. (ImageMagick can, too, if you give it the right flags, but not all
   its operations are well-defined in YCbCr.) Although conversion between the
   two is cheap, it is not free, and it is not needed for scaling. Thus, qscale
   does not do it.
 - qscale recognizes that JPEGs are stored with the color channels mostly
   separate (planar) and not chunked. Scaling does not need to be done on
   chunked data -- in fact, mostly, scaling is easier to do on planar data.
   Thus, no conversion to chunked before scaling (and no need to convert back
   to planar afterwards). (Note: Some SIMD properties might be easier to
   exploit on a chunked representation. It's usually not worth it in total,
   though.)
 - qscale can utilize the SSE instruction set found in almost all modern
   x86-compatible processors to do more work in the same amount of instructions
   (It can also use the SSE3 instruction set, although the extra boost on top
   of SSE is smaller.) It takes care to align the image data and memory
   accesses on the right boundaries wherever it makes sense.
 - qscale recognizes (like almost any decent scaling program) that most
   practical filter kernels are separable, so scaling can be done in two
   sequential simpler passes (horizontal and vertical) instead of one. The
   order does matter, though -- I've found doing the vertical pass (in
   cache-friendly order, doing multiple neighboring pixels at a time to
   exploit that the processor reads in entire cache lines and not individual
   bytes at a time) before the horizontal to be superior, in particular
   because this case is easier to SIMD-wise.
 - qscale understands that JPEGs are typically subsampled; ie., that the
   different color components are not stored at the same resolution. On
   the web, this is typically because the eye is less sensitive to color
   (chroma) information and as such much of it can safely be stored in
   a lower resolution to reduce file size without much visible quality
   degradation; in the JPEGs stored by a digital camera, it is simply
   because much of the color information is interpolated anyway (since
   the individual CCD dots are sensitive to either red, green or blue,
   not all at the same time), so it would not make much sense to pretend
   there is full color information. qscale does not ask libjpeg to
   interpolate the "missing" color information nor to downscale the
   already-downscaled color channels as ImageMagick does, but instead
   does a single scaling pass from the original resolution to the final
   subsampled resolution. (This is impossible for any program working
   in RGB mode, or chunked YCbCr.) This increases both speed and quality,
   although the effect on the latter is not particularly large.

The following optimizations are possible but not done (yet?):

 - qscale does not do the IDCT itself, even though there is improvement
   potential over libjpeg's IDCT. (There is an unmaintained and little-used fork
   of libjpeg called libjpeg-mmx that demonstrates this quite well.) In fact,
   since the DCT can be viewed as just another (separable, but not
   time-invariant) FIR filter, the quantization scaling and IDCT could probably
   be folded into the scaling in many cases, in particular those where the
   filter kernel is large (ie. large amounts of scaling).
 - qscale does not use multiple processors or cores (although different cores 
   can of course work on different images at the same time).

Finally, a word of advice: Even the fastest scaler can never be as fast as not
scaling. If you can do less work, or move it to a time where it is less user
visible, do that no matter how you scale your pictures. In particular, cache
your downscaled pictures if you can, preferrably ahead of time. If you cannot
predict in advance exactly what size (or sizes) you'll need, consider making
mipmaps (ie. one image of half the resolution each way, another one of a quarter
the resolution each way etc., down to a reasonable size) and scaling from them
instead. It will be much faster (primarily due to less loading time) and the
difference in quality is nearly impossible to detect, at least in my tests.

qscale is Copyright 2008 Steinar H. Gunderson <sgunderson@bigfoot.com>, and
licensed under the GNU General Public License, version 2. The full text of
the GPLv2 can be found in the included LICENSE file.