The previous constant opj_c13318 was mysteriously equal to 2/K , and in
the DWT, we had to divide K and opj_c13318 by 2... The issue was that the
band->stepsize computation in tcd.c didn't take into account the log2gain of
the band.
The effect of this change is expected to be mostly equivalent to the previous
situation, except some difference in rounding. But it leads to a dramatic
reduction of the mean square error and peak error in the irreversible encoding
of issue141.tif !
Update the bench_dwt utility to have a -decode/-encode switch
Measured performance gains for DWT encoder on a
Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz (4 cores, hyper threaded)
Encoding time:
$ ./bin/bench_dwt -encode -num_threads 1
time for dwt_encode: total = 8.348 s, wallclock = 8.352 s
$ ./bin/bench_dwt -encode -num_threads 2
time for dwt_encode: total = 9.776 s, wallclock = 4.904 s
$ ./bin/bench_dwt -encode -num_threads 4
time for dwt_encode: total = 13.188 s, wallclock = 3.310 s
$ ./bin/bench_dwt -encode -num_threads 8
time for dwt_encode: total = 30.024 s, wallclock = 4.064 s
Scaling is probably limited by memory access patterns causing
memory access to be the bottleneck.
The slightly worse results with threads==8 than with thread==4
is due to hyperthreading being not appropriate here.
Instead of being the full tile size.
* Use a sparse array mechanism to store code-blocks and intermediate stages of
IDWT.
* IDWT, DC level shift and MCT stages are done just on that smaller array.
* Improve copy of tile component array to final image, by saving an intermediate
buffer.
* For full-tile decoding at reduced resolution, only allocate the tile buffer to
the reduced size, instead of the full-resolution size.
Thanks to our macros that abstract SSE use, the functions can use
AVX2 when available (at compile time)
This brings an extra 23% speed improvement on bench_dwt in 64bit builds
with AVX2 compared to SSE2.
* Use single-pass lifting inverse wavelet transform.
* For vertical pass, use SSE2 when available so as to process 8 columns
in parallel. This is the most beneficial improvement, since the
vertical pass involves a lot of cache trashing.
With the bench_dwt utility with default arguments (16383x16383 image),
time goes from 4.064 s to 1.212 s.
Visual Studio 2015 does not pass regression tests with `__restrict` so kept disabled for MSVC.
Need to check proper usage of OPJ_RESTRICT (if correct then there’s
probably a bug in vc14)
Closes#661