This is implemented for a vector size of 128-bit. Since the scalar
product in the inner loop covers 5 samples or 160 bits, we need a group
multipler of 2.
To avoid reconfiguring the vector type, the outer loop, which loads
multiple input samples sticks to the same multipler. Consequently, the
outer loop loads 8 samples per iteration. This is safe since the minimum
period of the CELT codec is 15 samples.
The same code would also work, albeit needlessly inefficiently with a
vector length of 256 bits. A proper implementation will follow instead.
This is more spec-compliant because it does not rely
on dead-code elimination by the compiler. Especially
MSVC has problems with this, as can be seen in
https://ffmpeg.org/pipermail/ffmpeg-devel/2022-May/296373.html
or
https://ffmpeg.org/pipermail/ffmpeg-devel/2022-May/297022.html
This commit does not eliminate every instance where we rely
on dead code elimination: It only tackles branching to
the initialization of arch-specific dsp code, not e.g. all
uses of CONFIG_ and HAVE_ checks. But maybe it is already
enough to compile FFmpeg with MSVC with whole-programm-optimizations
enabled (if one does not disable too many components).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Some of these were made possible by moving several common macros to
libavutil/macros.h.
While just at it, also improve the other headers a bit.
Reviewed-by: Martin Storsjö <martin@martin.st>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The C and asm versions behaved differently _outside_ of the codec.
The C version returned pre-multiplied 'state' for the next execution
to use right away, while the assembly version outputted non-multiplied
'state' for the next execution to multiply to save instructions.
Since the initial state when initialized or seeking is always 0,
and since C and asm versions were never mixed, there was no issue.
However, comparing outputs directly in checkasm doesn't work without
dividing the initial state by CELT_EMPH_COEFF and multiplying the
returned state by CELT_EMPH_COEFF for the assembly function.
Since its actually faster to do this in C as well, copy the behavior the
asm versions use. As a reminder, the initial state 0 is divided by
CELT_EMPH_COEFF on seek and init (just in case in the future this is
changed, its technically more correct to init with CELT_EMPH_COEFF than 0,
however when seeking this will result in more audiable pops, unlike with 0
where the output gets in sync over a few samples).
153372 UNITS in postfilter_c, 65536 runs, 0 skips
73164 UNITS in postfilter_neon, 65536 runs, 0 skips -> 2.1x speedup
80591 UNITS in deemphasis_c, 131072 runs, 0 skips
43969 UNITS in deemphasis_neon, 131072 runs, 0 skips -> 1.83x speedup
Total decoder speedup: ~15% on a Raspberry Pi 3 (from 28.1x to 33.5x realtime)
Deemphasis SIMD based on the following unrolling:
const float c1 = CELT_EMPH_COEFF, c2 = c1*c1, c3 = c2*c1, c4 = c3*c1;
float state = coeff;
for (int i = 0; i < len; i += 4) {
y[0] = x[0] + c1*state;
y[1] = x[1] + c2*state + c1*x[0];
y[2] = x[2] + c3*state + c1*x[1] + c2*x[0];
y[3] = x[3] + c4*state + c1*x[2] + c2*x[1] + c3*x[0];
state = y[3];
y += 4;
x += 4;
}
Unlike the x86 version, duplication is used instead of pslldq so
the structure and tables are different.