Busybox-w32 [1] works for building ffmpeg on Windows (as an
alternative to msys2, cygwin or WSL).
On busybox-w32, "uname" returns "Windows_NT"; recognize this
in exesuf() as having an .exe suffix.
If building in this environment with a mingw toolchain, one has
to explicitly set --target-os=mingw32. (We probably don't
want to imply that this uname, set as target_os_default, would
default to mingw?) But despite what is set with --target-os,
one can't override the configure variable "host_os", which
exesuf() has to recognize.
[1] https://github.com/rmyorston/busybox-w32
If Zbb is enabled at compilation (e.g. Ubuntu), the compiler should
compile the new C mid_pred() function correctly. But if Zbb is *not*
enabled (e.g. Debian), then we can at least fallback at run-time.
On SiFive-U74, before:
sub_median_pred_c: 1331.9 ( 1.00x)
sub_median_pred_rvb_b: 881.8 ( 1.51x)
After:
sub_median_pred_c: 1133.1 ( 1.00x)
sub_median_pred_rvb_b: 875.7 ( 1.29x)
This reduces the minimum instruction emission for mid_pred()
(i.e. median of 3) down to:
- 3 comparisons and 4 conditional moves, or
- 4 min/max.
With that the compiler can eliminate any branch. This optimal
situation is attainable with Clang 21 on Arm64, RVA22 and x86,
with GCC 15 on Arm64 and x86 (RVA22 goes from 2 to 1 branch).
These optimisations also work on Arm32 and LoongArch.
The same algorithm is already implemented via inline assembler for some
architectures such as x86 and Arm32, but notably not Arm64 and RVA22.
Besides, using C code allows the compiler to schedule instruction
properly.
Even on architectures with neither conditional moves nor min/max, this
leads to a visible performance improvement for C code, as seen here for
RVA20 code running on SiFive-U74:
Before:
sub_median_pred_c: 1657.5 ( 1.00x)
sub_median_pred_rvb_b: 875.9 ( 1.89x)
After:
sub_median_pred_c: 1331.9 ( 1.00x)
sub_median_pred_rvb_b: 881.8 ( 1.51x)
Note that this commit leaves the x86 and Arm32 code intact so it has
no effects on those ISA's.
When building DLLs on ARM64EC, the default use of `dumpbin
-linkermember:1` fails because ARM64EC static libraries use a
different linker member format. Use `-linkermember:32` for ARM64EC
to correctly extract symbols.
Additionally, MSVC inserts $exit_thunk and $entry_thunk symbols
for ARM64EC to handle x64 ↔ ARM64 transitions. These are internal
thunks and must not be exported. Filter them out when generating
the .def file to avoid unresolved symbols or invalid exports.
Trim out the leading '#' on ARM64EC function symbols. This is only
relevant on ARM64EC, but it is benign to do that filtering on
all architectures (such symbols aren't expected on other
architectures).
Simplify the sed command by removing the symbol address with a
sed expression instead of a later "cut" command.
This ensures correct symbol extraction and stable DLL generation
on ARM64EC targets, while keeping behavior unchanged for other
Windows architectures.
Explicitly spell it out that we are not going to modify the
individual libraries for the purposes of improving conformance
to ARM64EC.
We may (or may not) accept build system patches for making such
a build succeed, provided that it does not require changes to
the actual library code.
Since af97c9865f,
the return value of avio_read() has been compared against
an uint32_t, so that the int is promoted to uint32_t for
the comparison (on common systems with 32bit ints). The upshot was
that errors returned from avio_read() were ignored, so that
the buffer could be uninitialized on success.
Fix this by using ffio_read_size() instead.
Fixes: MemorySanitizer: use-of-uninitialized-value
Fixes: 443923343/clusterfuzz-testcase-minimized-ffmpeg_dem_FLAC_fuzzer-5458132865449984
Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Reviewed-by: James Almer <jamrial@gmail.com>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
read_frame_internal() may result in a stream being modified without also
returning a packet from it. Given said function only bothered to update the
internal stream context for the returned packet, the result would be a desync
between the stream's AVCodecParameters and the internal AVCodecContext.
This change makes sure all streams are updated within the
avformat_find_stream_info() loop.
Fixes #YWH-PGM40646-20
Signed-off-by: James Almer <jamrial@gmail.com>
vp8 encoder can be configured to drop frames, when e.g. bitrate
overshoot is detected. At present the code responsible for
managing an internal fifo assumes that we will get an output frame per
each frame fed into encoder. That is not the case if the encoder can
decide to drop frames.
Running:
ffmpeg -stream_loop 100 -i dash_video3.webm -c:v libvpx -b:v 50k
-drop-threshold 20 -screen-content-mode 2 output.webm
results in lots of warnings like:
[libvpx @ 0x563fd8aba100] Mismatching timestamps: libvpx 2187 queued
2185; this is a bug, please report it
[libvpx @ 0x563fd8aba100] Mismatching timestamps: libvpx 2189 queued
2186; this is a bug, please report it
followed by:
[vost#0:0/libvpx @ 0x563fd8ab9b40] [enc:libvpx @ 0x563fd8aba080] Error
submitting video frame to the encoder
[vost#0:0/libvpx @ 0x563fd8ab9b40] [enc:libvpx @ 0x563fd8aba080] Error
encoding a frame: No space left on device
[vost#0:0/libvpx @ 0x563fd8ab9b40] Task finished with error code: -28
(No space left on device)
[vost#0:0/libvpx @ 0x563fd8ab9b40] Terminating thread with return code
-28 (No space left on device)
The reason for the above error is that each dropped frame leaves an
extra item in the fifo, which eventually overflows.
The proposed fix is to keep popping elements from the fifo until the
one with the matching pts is found. A side effect of this change is that
the code no longer considers pts mismatch to be a bug.
This has likely regressed around 5bda4ec6c3
when fifo started to be universally used.
Signed-off-by: Dariusz Marcinkiewicz <darekm@google.com>
Use the concat protocol, to test the parser's capabilities to differentiate between
EOC maker before SOC marker, on top of false EOC marker positives and EOC maker on EOF.
Signed-off-by: James Almer <jamrial@gmail.com>
JPEG-XS streams can have the bytes corresponding to certain markers as part of
slice data, and no considerations were made for it, so we need to add checks
for false positives.
This fixes assembling several samples.
Signed-off-by: James Almer <jamrial@gmail.com>
The code assumed that the destination buffer was zeroed, a misbehaviour
with which checkasm is bug-compatible as it zeroes the destination
buffer. The fixed code is even faster:
SpacemiT X60:
sub_left_predict_c: 51792.5 ( 1.00x)
sub_left_predict_rvv_i32: 3504.4 (14.78x)
This requires updating the order of the dither matrix offsets as well. In
particular, this can only be done cleanly if the dither matrix offsets are
compatible between duplicated channels; but this should be guaranteed after
the previous commits.
As a side note, this also fixes a bug where we pushed SWS_OP_SWIZZLE past
SWS_OP_DITHER even for very low bit depth output (e.g. rgb4), which led to
a big loss in accuracy due to loss of per-channel dither noise.
Improves loss of e.g. gray -> rgb444 from 0.00358659 to 0.00261414,
now beating legacy swscale in these cases as well.
Otherwise, this is invalid; as the result of applying the next operation
after channel duplication may not commute with the result of duplicating
the result of applying the next operation on only one channel.
In practice, this was using the last seen channel's coefficients.
Note that this did not actually affect anything in practice, because the only
relevant ops (MIN/MAX) were always generated with identical coefficients for
identical channel ranges.
However, it will matter moving forwards, as e.g. dither ops may not be
commuted freely if their matrix offsets differ per channel.
The current usage is ambiguous between "affects each component equally" and
"affects each component independently" - and arguably, the current behavior
was a bug (since SWS_OP_DITHER should not commute with a SWIZZLE, at least
from a bit-exactness PoV).
However, when trying to define cleaner replacements for these concepts, I
realized there are too many special cases; and given that we only have two
use sites, I decided to just split them directly into "commute" functions
for those particular usage cases.
As an added benefit, this moves the commutation logic out of the already-long
ff_sws_ops_list_optimize().
On the surface, this trades a tiny bit of PSNR for not introducing chroma
noise into grayscale images. However, the main reason for this change is
actually motivated by a desire to avoid regressing the status quo of
duplicating swizzles being able to be commuted past dither ops.
Instead of computing y + N with a hard-coded index offset, calculate the
relative offset as a 16-bit integer in C and add that to the pointer directly.
Since we no longer mask the resulting combined address, this may result in
overread, but that's fine since we over-provisioned the array in the previous
commit.
I want to change the way offsets are calculated inside the dither asm;
and the cleanest way to solve that problem is to just over-allocate the
entire dither matrix based on the maximum offset range we expected.
This is an exceptionally unlikely (in fact, currently impossible) case to
actually hit, and not worth micro-optimizing for. More specifically, having
this special case prevents me from easily adding per-row offsets.
This check was wrong; 1 << 0 = 1. The intent was to skip allocating a 1x1
matrix by assuming it is a constant 0.5. But as written, the check never
actually executed.
This did not affect the runtime performance, nor did it leak memory; but it
did mean we didn't hit the intended `assert`.
Since we only need 8 bytes to store the dither matrix pointer, we actually
still have 8 bytes left-over. That means we could either store the 8-bit
row offset directly, or alternatively compute a 16-bit pointer offsets.
I have chosen to do the former for the C backend, in the interest of
simplicity.
The one downside of this approach is that it would fail on hypothetical
128-bit platforms; although I seriously hope that this code does not live
long enough to see the need for 128-bit addressable memory.
To improve decorrelation between components, we offset the dither matrix
slightly for each component. This is currently done by adding a hard-coded
offset of {0, 3, 2, 5} to each of the four components, respectively.
However, this represents a serious challenge when re-ordering SwsDitherOp
past a swizzle, or when splitting an SwsOpList into multiple sub-operations
(e.g. for decoupling luma from subsampled chroma when they are independent).
To fix this on a fundamental level, we have to keep track of the offset per
channel as part of the SwsDitherOp metadata, and respect those values at
runtime.
This commit merely adds the metadata; the update to the underlying backends
will come in a follow-up commit. The FATE change is merely due to the
added offsets in the op list print-out.
The shader needs ~3 loads per DCT coeff.
This data was not observed to get efficiently stored
in the upper cached levels, loading it explicitely in
shared memory fixes that.
Also reduce code size by moving the bitstream
initialization outside of the switch/case.
In commit 6e0034ab7e, the image crop adjustment
was moved after the fitting logic. However, this moved the adjustment inside the
`if (src == ref)` branch, thus missing applying the same un-rotation to input
frames that are *not* the reference frame.
Fix this by pulling the logic back out of the branch again. While we could just
move it after the fitting logic, I think it's more clear to the intent of the
code to just preserve the (rotated) crop rect as a separate variable
`crop_orig`.
Fixes: 6e0034ab7e
Martin Storsjö