To support non-aligned buffers during the post-transform step, just iterate
backwards over the array.
This allows using the 15xN-point FFT, with which the speed is 2.1 times
faster than our old libavcodec implementation.
~4x faster than the C version.
The shuffles in the 15pt dim1 are seriously expensive. Not happy with it,
but I'm contempt.
Can be easily converted to pure AVX by removing all vpermpd/vpermps
instructions.
This commit implements an iMDCT in pure assembly.
This is capable of processing any mod-8 transforms, rather than just
power of two, but since power of two is all we have assembly for
currently, that's what's supported.
It would really benefit if we could somehow use the C code to decide
which function to jump into, but exposing function labels from assebly
into C is anything but easy.
The post-transform loop could probably be improved.
This was somewhat annoying to write, as we must support arbitrary
strides during runtime. There's a fast branch for stride == 4 bytes
and a slower one which uses vgatherdps.
Zen 3 benchmarks for stride == 4 for old (av_imdct_half) vs new (av_tx):
128pt:
2811 decicycles in av_tx (imdct),16775916 runs, 1300 skips
3082 decicycles in av_imdct_half,16776751 runs, 465 skips
256pt:
4920 decicycles in av_tx (imdct),16775820 runs, 1396 skips
5378 decicycles in av_imdct_half,16776411 runs, 805 skips
512pt:
9668 decicycles in av_tx (imdct),16775774 runs, 1442 skips
10626 decicycles in av_imdct_half,16775647 runs, 1569 skips
1024pt:
19812 decicycles in av_tx (imdct),16777144 runs, 72 skips
23036 decicycles in av_imdct_half,16777167 runs, 49 skips
When the SLOW_GATHER flag was added to the AVX2 version, this
made FMA3-features not enabled on Zen CPUs.
As FMA3 adds 6-7% across all platforms that support it, in
the interest of saving space, this commit removes the AVX
version and replaces it with an FMA3 version.
The only CPUs affected are Sandy Bridge and Bulldozer, which
have AVX support, but no FMA3 support.
In the future, if there's a demand for it, a version of the
function duplicated for AVX can be added.
This reverts commit 82a68a8771.
Smarter slow ISA penalties makes gathers still useful.
The intention is to use gathers with the final stage of non-ptwo iMDCTs,
where they give benefit.
Its performance loss ranges from either being just as fast as individual loads
(Skylake), a few percent slower (Alderlake), 8% slower (Zen 3), to completely
disasterous (older/other CPUs).
Sadly, gathers never panned out fast on x86, even with the benefit of time and
implementation experience.
This also saves a register, as there's no need to fill out an additional
register mask.
Zen 3 (16384-point transform):
Before: 1561050 decicycles in av_tx (fft), 131072 runs, 0 skips
After: 1449621 decicycles in av_tx (fft), 131072 runs, 0 skips
Alderlake:
2% slower on big transforms (65536), to 1% (131072), to a few percent for smaller
sizes.
This commit does some refactoring to make defining assembly codelets
smaller, and fixes compiler redefinition warnings. It also allows
for other assembly versions to reuse the same boilerplate code as
x86.
Finally, it also adds the out_of_place flag to all assembly codelets.
This changes nothing, as out-of-place operation was assumed to be
available anyway, but this makes it more explicit.
Makes Bulldozer prefer AVX functions rather than AVX2,
which are 64% slower:
AVX: 117653 decicycles in av_tx (fft), 1048535 runs, 41 skips
AVX2: 193385 decicycles in av_tx (fft), 1048561 runs, 15 skips
The only difference between both is that vgatherdpd is used in
the former. We don't want to mark them with the new SLOW_GATHER
flag however, since gathers are still faster on Haswell/Zen 2/3
than plain loads.
This commit rewrites the internal transform code into a constructor
that stitches transforms (codelets).
This allows for transforms to reuse arbitrary parts of other
transforms, and allows transforms to be stacked onto one
another (such as a full iMDCT using a half-iMDCT which in turn
uses an FFT). It also permits for each step to be individually
replaced by assembly or a custom implementation (such as an ASIC).
This commit adds a pure x86 assembly SIMD version of the FFT in libavutil/tx.
The design of this pure assembly FFT is pretty unconventional.
On the lowest level, instead of splitting the complex numbers into
real and imaginary parts, we keep complex numbers together but split
them in terms of parity. This saves a number of shuffles in each transform,
but more importantly, it splits each transform into two independent
paths, which we process using separate registers in parallel.
This allows us to keep all units saturated and lets us use all available
registers to avoid dependencies.
Moreover, it allows us to double the granularity of our per-load permutation,
skipping many expensive lookups and allowing us to use just 4 loads per register,
rather than 8, or in case FMA3 (and by extension, AVX2), use the vgatherdpd
instruction, which is at least as fast as 4 separate loads on old hardware,
and quite a bit faster on modern CPUs).
Higher up, we go for a bottom-up construction of large transforms, foregoing
the traditional per-transform call-return recursion chains. Instead, we always
start at the bottom-most basis transform (in this case, a 32-point transform),
and continue constructing larger and larger transforms until we return to the
top-most transform.
This way, we only touch the stack 3 times per a complete target transform:
once for the 1/2 length transform and two times for the 1/4 length transform.
The combination algorithm we use is a standard Split-Radix algorithm,
as used in our C code. Although a version with less operations exists
(Steven G. Johnson and Matteo Frigo's "A modified split-radix FFT with fewer
arithmetic operations", IEEE Trans. Signal Process. 55 (1), 111–119 (2007),
which is the one FFTW uses), it only has 2% less operations and requires at least 4x
the binary code (due to it needing 4 different paths to do a single transform).
That version also has other issues which prevent it from being implemented
with SIMD code as efficiently, which makes it lose the marginal gains it offered,
and cannot be performed bottom-up, requiring many recursive call-return chains,
whose overhead adds up.
We go through a lot of effort to minimize load/stores by keeping as much in
registers in between construcring transforms. This saves us around 32 cycles,
on paper, but in reality a lot more due to load/store aliasing (a load from a
memory location cannot be issued while there's a store pending, and there are
only so many (2 for Zen 3) load/store units in a CPU).
Also, we interleave coefficients during the last stage to save on a store+load
per register.
Each of the smallest, basis transforms (4, 8 and 16-point in our case)
has been extremely optimized. Our 8-point transform is barely 20 instructions
in total, beating our old implementation 8-point transform by 1 instruction.
Our 2x8-point transform is 23 instructions, beating our old implementation by
6 instruction and needing 50% less cycles. Our 16-point transform's combination
code takes slightly more instructions than our old implementation, but makes up
for it by requiring a lot less arithmetic operations.
Overall, the transform was optimized for the timings of Zen 3, which at the
time of writing has the most IPC from all documented CPUs. Shuffles were
preferred over arithmetic operations due to their 1/0.5 latency/throughput.
On average, this code is 30% faster than our old libavcodec implementation.
It's able to trade blows with the previously-untouchable FFTW on small transforms,
and due to its tiny size and better prediction, outdoes FFTW on larger transforms
by 11% on the largest currently supported size.
