Average vertically before horizontally; horizontal averaging is more
worksome. Doing the vertical averaging first reduces the number of
horizontal averages by half.
Use pmaddubsw and pavgw to do the horizontal averaging for a slight
performance improvement.
Minor tweaks.
Improve the SSSE3 dyadic downsample routines and drop the SSE4 routines.
The non-temporal loads used in the SSE4 routines do nothing for cache-
backed memory AFAIK.
Adjust tests because averaging vertically first gives slightly different
output.
~2.39x speedup for the widthx32 routine on Haswell when not memory-bound.
~2.20x speedup for the widthx16 routine on Haswell when not memory-bound.
Note that the widthx16 routine can be unrolled for further speedup.
Keep track of relative pixel offsets and utilize pshufb to efficiently
extract relevant pixels for horizontal scaling ratios <= 4.
Fall back to a generic approach for ratios > 4. Note that the generic
approach can be backported to SSE2.
The implementation assumes that data beyond the end of each line,
before the next line begins, can be dirtied; which AFAICT is safe with
the current usage of these routines.
Speedup is ~6.67x/~3.26x (32-bit/64-bit) for horizontal ratios <= 2,
~6.24x/~3.00x for ratios within (2, 4], and ~4.89x/~2.17x for ratios
> 4 when not memory-bound on Haswell as compared with the current SSE2
implementation.
WelsQuantFour4x4Max_avx2 (~2.06x speedup over SSE2)
WelsQuantFour4x4_avx2 (~2.32x speedup over SSE2)
WelsQuant4x4Dc_avx2 (~1.49x speedup over SSE2)
WelsQuant4x4_avx2 (~1.42x speedup over SSE2)
At lower bitrates, it is overall faster to conditionally do one block
at a time with SSE2 on Haswell and likely other common architectures.
At higher bitrates, it is faster to use the wider routine that IDCTs
four blocks at a time. To avoid potential performance regressions
as compared to MMX, stick with single-block IDCTs with SSE2. There
is still a performance advantage as compared to MMX because the
single-block SSE2 routine is faster than the corresponding MMX
routine.
Stick with four blocks at a time with AVX2 for which that appears
to be consistently faster on Haswell.
Move asm routines to common. Delete obsolete decoder routines.
Use wider routines where applicable.
~1.07x overall faster decode on a quick 720p30 4Mbps test on Haswell.
WelsSampleSatd16x16_avx2 (~2.31x speedup over SSE4.1 on Haswell).
WelsSampleSatd16x8_avx2 (~2.19x speedup over SSE4.1 on Haswell).
WelsSampleSatd8x16_avx2 (~1.68x speedup over SSE4.1 on Haswell).
WelsSampleSatd8x8_avx2 (~1.53x speedup over SSE4.1 on Haswell).
The pFeatureC pointer is an uint32_t pointer, therefore only
load 32 bits into ecx.
This avoids loading potentially uninitialized data into the upper
half of the rcx register, fixing valgrind warnings in some build
setups (depending on how the compiler chooses to layout the stack
in the calling function).
Previously the assembly sources had mixed indentation consisting
of both spaces and tabs, making it quite hard to read unless
the right tab size was used in the editor.
Tabs have been interpreted as 4 spaces in most cases, matching
the surrounding code.
