On 64 bit, the stack seems to be aligned enough for our needs.
Signed-off-by: Martin Storsjö <martin@martin.st>

Generate quantization tables when they are not present in the
first chunk.
Signed-off-by: Martin Storsjö <martin@martin.st>

They belong in the init functions specific to each CPU capability.

They have been wrong since 11d957fb

New streams may be created at any time, e.g. on codec change.

Previously the file descriptor was stored in the priv_data pointer.
Signed-off-by: Martin Storsjö <martin@martin.st>

The 64 bit cl.exe version 16.00.30319.01 crashes with an internal
compiler error on the current check (and thus deduces it isn't
supported, even if the actual usage in libavuil/x86/cpu.c works
fine), but by assigning the value from the intrinsic to a variable,
or returning it, it works fine.

This error is fixed in cl.exe version 16.00.40219.01.
Signed-off-by: Martin Storsjö <martin@martin.st>

This can come in handy for testing and possibly other purposes.

Some msvcrt versions (the static 64 bit libc in MSVC 10) have
a log2 function, but there is no declaration for it in the headers.
Therefore, the normal configure check might find it, but it can fail
during build or at runtime, depending on whether implicit function
declarations are an error or not.

Therefore simply ignore this function on this platform.
Signed-off-by: Martin Storsjö <martin@martin.st>

Signed-off-by: Mans Rullgard <mans@mansr.com>

Signed-off-by: Mans Rullgard <mans@mansr.com>

Signed-off-by: Martin Storsjö <martin@martin.st>

This allows using it from libavformat as well. This will be used
by the RTP/JPEG depacketizer.
Signed-off-by: Martin Storsjö <martin@martin.st>

These symbols will be used from the RTP/JPEG depacketizer.
Signed-off-by: Martin Storsjö <martin@martin.st>

This reverts parts of e0c6cce4. There is external mmx asm that
requires this alignment.

This fixes crashes when using swscale in builds with external mmx,
without inline assembly.
Signed-off-by: Martin Storsjö <martin@martin.st>

This separates code relying on inline from that relying on external
assembly and fixes instances where the coalesced check was incorrect.

Signed-off-by: Mans Rullgard <mans@mansr.com>

Signed-off-by: Martin Storsjö <martin@martin.st>

If regularly parsing new chan atoms (as in rtpdec_qt), but the
chan atoms don't actually contain any channel layout, don't reset
the value that the caller has filled in (by guessing).
Signed-off-by: Martin Storsjö <martin@martin.st>

Signed-off-by: Martin Storsjö <martin@martin.st>

Signed-off-by: Alex Converse <alex.converse@gmail.com>

The SWAP macro does not work for explicit xmm/ymm usage, so instead just move
the scalar value from xmm2 to xmm0.

CAVS uses its own idct so using dsputil to set the permutation
is fragile.
Signed-off-by: Mans Rullgard <mans@mansr.com>

For some reason add_hfyu_median_prediction_cmov is only selected
on 3Dnow-capable CPUs, even though it uses no 3Dnow instructions.
This patch allows it to be selected on any cpu with cmov with the
possibility of being overridden by the mmxext version.
Signed-off-by: Mans Rullgard <mans@mansr.com>

This fixes crashes when called from rtpdec_qt, where
AVFormatContext->pb is null, a crash present since 3bab7cd1.
Signed-off-by: Martin Storsjö <martin@martin.st>

Fixes an infinite loop on flush when avresample_get_delay() still
reports some samples but avresample_convert() doesn't return any data.

It calculates the sum of power of two series, which can be done in one step.

Suggested by Michael Niedermayer <michaelni@gmx.at>
Signed-off-by: Kostya Shishkov <kostya.shishkov@gmail.com>

That division can be replaced with a comparison:
((c->value - c->low) << 1) + 1 >= range

By expanding 'range' definition and simplifying this inequation we obtain
the final expression.

Suggested by Michael Niedermayer <michaelni@gmx.at>
Signed-off-by: Kostya Shishkov <kostya.shishkov@gmail.com>

Reported by Michael Niedermayer <michaelni@gmx.at>
Signed-off-by: Kostya Shishkov <kostya.shishkov@gmail.com>

Prevents dangling pointers and makes access after free more obvious.
Setting AVFrame.qscale_table to NULL is required for successfully
allocating a previously freed Picture with ff_alloc_picture().

The default ones work, but outputs the preprocessed file on stdout
(into config.log).
Signed-off-by: Martin Storsjö <martin@martin.st>

Signed-off-by: Martin Storsjö <martin@martin.st>

The init functions check for CPU capabilities on their own already.