The fastest fast Fourier transform in not just the west, but the world,
now for the most popular toy ISA.
On a high level, it follows the design of the AVX2 version closely,
with the exception that the input is slightly less permuted as we don't have
to do lane switching with the input on double 4pt and 8pt.
On a low level, the lack of subadd/addsub instructions REALLY penalizes
any attempt at writing an FFT. That single register matters a lot,
and reloading it simply takes unacceptably long.
In x86 land, vendors would've noticed developers need this.
In ARM land, you get a badly designed complex multiplication instruction
we cannot use, that's not present on 95% of devices. Because only
compilers matter, right?
Future optimization options are very few, perhaps better register
management to use more ld1/st1s.
All timings below are in cycles:
A53:
Length | C | New (lavu) | Old (lavc) | FFTW
------ |-------------|-------------|-------------|-----
4 | 842 | 420 | 1210 | 1460
8 | 1538 | 1020 | 1850 | 2520
16 | 3717 | 1900 | 3700 | 3990
32 | 9156 | 4070 | 8289 | 8860
64 | 21160 | 9931 | 18600 | 19625
128 | 49180 | 23278 | 41922 | 41922
256 | 112073 | 53876 | 93202 | 101092
512 | 252864 | 122884 | 205897 | 207868
1024 | 560512 | 278322 | 458071 | 453053
2048 | 1295402 | 775835 | 1038205 | 1020265
4096 | 3281263 | 2021221 | 2409718 | 2577554
8192 | 8577845 | 4780526 | 5673041 | 6802722
Apple M1
New - Total for len 512 reps 2097152 = 1.459141 s
Old - Total for len 512 reps 2097152 = 2.251344 s
FFTW - Total for len 512 reps 2097152 = 1.868429 s
New - Total for len 1024 reps 4194304 = 6.490080 s
Old - Total for len 1024 reps 4194304 = 9.604949 s
FFTW - Total for len 1024 reps 4194304 = 7.889281 s
New - Total for len 16384 reps 262144 = 10.374001 s
Old - Total for len 16384 reps 262144 = 15.266713 s
FFTW - Total for len 16384 reps 262144 = 12.341745 s
New - Total for len 65536 reps 8192 = 1.769812 s
Old - Total for len 65536 reps 8192 = 4.209413 s
FFTW - Total for len 65536 reps 8192 = 3.012365 s
New - Total for len 131072 reps 4096 = 1.942836 s
Old - Segfaults
FFTW - Total for len 131072 reps 4096 = 3.713713 s
Thanks to wbs for some simplifications, assembler fixes and a review
and to jannau for giving it a look.
It is unnecessary since ee551a21ddcbf81afe183d9489c534ee80f263a0;
but it was redundant even before that, because decode_simple_internal()
calls emms_c().
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
There's no warranty that vpx_codec_encode() will generate a list with the same
amount of packets for both the yuv planes encoder and the alpha plane encoder,
so queueing packets based on what the main encoder returns will fail when the
amount of packets in both lists differ.
Queue all data packets for every vpx_codec_encode() call from both encoders
before attempting to assemble output AVPackets out of them.
Fixes ticket #9884
Reviewed-by: Vignesh Venkatasubramanian <vigneshv@google.com>
Signed-off-by: James Almer <jamrial@gmail.com>
This is somewhat redundant with the is_decoded check. Maybe
there is a nicer solution
Fixes: Null pointer dereference
Fixes: 49584/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_HEVC_fuzzer-5297367351427072
Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Fixes: left shift of 192 by 24 places cannot be represented in type 'int'
Fixes: 49577/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_MP1FLOAT_fuzzer-5205996678545408
Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Fixes: signed integer overflow: -1948269928 * 10 cannot be represented in type 'int'
Fixes: 49451/clusterfuzz-testcase-minimized-ffmpeg_dem_SUBVIEWER_fuzzer-6344614822412288
Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Fixes: out of array read
Fixes: 49434/clusterfuzz-testcase-minimized-ffmpeg_AV_CODEC_ID_TIFF_fuzzer-5208501080686592
Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Don't silently replace it with the default layout for the amount of channels
from the requested layout.
Should fix ticket #9869
Signed-off-by: James Almer <jamrial@gmail.com>
Fixes FATE-failures with the the filter-2xbr filter-3xbr filter-4xbr
filter-ep2x filter-ep3x filter-hq2x filter-hq3x filter-hq4x
filter-paletteuse-bayer filter-paletteuse-bayer0
filter-paletteuse-nodither and filter-paletteuse-sierra2_4a tests
when using 32bit x86 with CPUFLAGS ranging from "mmx+mmxext" to
"mmx+mmxext+sse+sse2+sse3" (the relevant function is only overwritten
when using SSSE3).
Reviewed-by: Lynne <dev@lynne.ee>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
By checking immediately whether the first allocation was successfull
one can simplify the cleanup code in case of errors.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
APNG works with a single reference frame and an output frame.
According to the spec, decoding APNG works by decoding
the current IDAT/fdAT chunks (which decodes to a rectangular
subregion of the whole image region), followed by either
overwriting the region of the output frame with the newly
decoded data or by blending the newly decoded data with
the data from the reference frame onto the current subregion
of the output frame. The remainder of the output frame
is just copied from the reference frame.
Then the reference frame might be left untouched
(APNG_DISPOSE_OP_PREVIOUS), it might be replaced by the output
frame (APNG_DISPOSE_OP_NONE) or the rectangular subregion
corresponding to the just decoded frame has to be reset
to black (APNG_DISPOSE_OP_BACKGROUND).
The latter case is not handled correctly by our decoder:
It only performs resetting the rectangle in the reference frame
when decoding the next frame; and since commit
b593abda6c it does not reset
the reference frame permanently, but only temporarily (i.e.
it only affects decoding the frame after the frame with
APNG_DISPOSE_OP_BACKGROUND). This is a problem if the
frame after the APNG_DISPOSE_OP_BACKGROUND frame uses
APNG_DISPOSE_OP_PREVIOUS, because then the frame after
the APNG_DISPOSE_OP_PREVIOUS frame has an incorrect reference
frame. (If it is not followed by an APNG_DISPOSE_OP_PREVIOUS
frame, the decoder only keeps a reference to the output frame,
which is ok.)
This commit fixes this by being much closer to the spec
than the earlier code: Resetting the background is no longer
postponed until the next frame; instead it is applied to
the reference frame.
Fixes ticket #9602.
(For multithreaded decoding it was actually already broken
since commit 5663301560d77486c7f7c03c1aa5f542fab23c24.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Some code in alimiter assumes that there are 2 channels, resulting in
clipping if the loudest channel is 3 or above and an out-of-bounds read if
the input is monophonic. Fix that in 2 places.
Signed-off-by: David Flater <dave@flaterco.com>
Add adaptive_i/b feature to hevc_qsv. Adaptive_i allows changing of
frame type from P and B to I. Adaptive_b allows changing of frame type
frome B to P.
Signed-off-by: Wenbin Chen <wenbin.chen@intel.com>
Signed-off-by: Haihao Xiang <haihao.xiang@intel.com>
According to its documentation it returns "pts of the last muxed packet
+ its duration", but the value it actually returns right now is
(possibly guessed) dts after muxer-internal bitstream filtering (if
any).
This function was added for ffmpeg.c, but it is not used there anymore.
Since the value it returns is ill-defined and so inappropriate for any
serious use, deprecate it.
Just return AVERROR_EXTERNAL immediately when encode error.
The other logic should keep the old behavior before commit 7c05b7951.
Suggested-By: Zhao Zhili <zhilizhao@tencent.com>
Signed-off-by: Steven Liu <lq@chinaffmpeg.org>
This commit moves the encoder-only allocations of the tables
owned solely by the main encoder context to mpegvideo_enc.c.
This avoids checks in mpegvideo.c for whether we are dealing
with an encoder; it also improves modularity (if encoders are
disabled, this code will no longer be included in the binary).
And it also avoids having to reset these pointers at the beginning
of ff_mpv_common_init() (in case the dst context is uninitialized,
ff_mpeg_update_thread_context() simply copies the src context
into dst which therefore may contain pointers not owned by it,
but this does not happen for encoders at all).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Fixes frame-threaded decoding of samples created by concatting
a video with data partitioning and a video not using it.
(Only the MPEG-4 decoder sets this, so it is synced in
mpeg4_update_thread_context() despite being a MpegEncContext-field.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
