generates the following errors on some platforms (android-x86;
android-x86_64; macos; ios-simulator-x86_64):

../../modules/packetizer/startcode_helper.h:91:41: error: invalid input size for constraint 'x'
                : [v]"r"(p), [czero]"x"(zeros)

'x' is supposedly the right constraint for both xmm and ymm registers, and
this compiles fine on other platforms. googling got me nowhere.

helps performance should SSE code also be in use at the same time

helps performance should SSE code also be in use, say from a filter

In a previous patch, bits of deinterlace that were MMX/MMXEXT accelerated
only were upgraded to SSE2 in preparation for a complete purge of
MMX/MMXEXT, however this upgrade was minimal, not going as far as actually
making proper use of the wider SSE2 registers; todo notes were left
with regard to this.

This commit addresses those todo notes.

With algo-x, the todos were simply removed; in that case using more of the
SSE2 registers would involve moving from the 8x8 processing model to a
16x16/8x16 one, which would then change the results of processing, per how
it determines which change to apply to each 8x8 block on the content of
each 8x8 block.

there is no MMX, it was incorrectly used with pure SSE2+

there's no MMX code, it was being incorrectly used with pure SSE2

notes:
 - the whole of mmx.h was removed, despite actually also containing 3dNow
   and SSE items, because none of those are in use
 - the `EndMMX` function of merge.[h|c] has been left, because it is used
   by the SSE2 code (incorrectly?) in deinterlace.c

(prep work for purging MMX/MMXEXT)

the code changed here has no greater acceleration than MMX/MMXEXT, so to
prepare for the MMX/MMXEXT purge, convert this to SSE2

note, this is a basic conversion only, *using* SSE2 vector registers (Xmm)
but not actually taking advantage of their greater width; that may
require significant revision to the code and thus is left for a separate
commit (and/or person) to possibly tackle later. note the following
benefits, vs. leaving as it was:
 - switching to SSE registers avoids the MMX<->FP register clash
 - thus it avoids the need for issuing `emms` instructions
 - with a little more work, could take full advantage of these wider registers
 - allows us to proceed with the purge of all old MMX/MMXEXT code

also, I inlined the asm, removing dependency on the abstraction macro set

[revised edition]
 - corrected the misunderstanding that NASM syntax could be used
   as normal inline asm by switching back to MASM syntax (apparently
   GCC does support intel syntax, but it's not up to scratch); a
   proper switch to NASM which requires moving to separate files
   for asm can be done later perhaps.
 - fixed a few small bugs

 - prep startcode_FindAnnexB() to more simply support multiple SIMD
   implementations.
 - adjust the test unit per that change, and itself to properly
   support testing each simd implementation.

 - use `__asm__` rather than `asm`.
 - better align a few asm statements.

i'd guessed at what the right AVX2 instruction might be when i added
the new check, then forgot to later go back and do it properly
before it got merged.

Zhao Zhili authored 4 years ago and Steve Lhomme committed 4 years ago

1. FF_THREAD_FRAME is in the FFmpeg code base since 2011.

2. codec/avcodec/video.c and codec/avcodec/directx_va.c use
FF_THREAD_FRAME directly without check.

3. FF_THREAD_FRAME is a thread type, it's not a multithreads decoding
support flag.

Signed-off-by: Steve Lhomme <robux4@ycbcr.xyz>

Zhao Zhili authored 4 years ago and Steve Lhomme committed 4 years ago

Signed-off-by: Steve Lhomme <robux4@ycbcr.xyz>

Zhao Zhili authored 4 years ago and Steve Lhomme committed 4 years ago

It's useless to set struct fields to NULL just before free the struct.

Signed-off-by: Steve Lhomme <robux4@ycbcr.xyz>

We should not rely on the old PTS values after a flush.

The macOS fix has been merged upstream.

  it is no longer dependant of the medialibrary

  isFocusOnContextButton no longer exists

  bookmark are stored in the medialibrary

Fatih Uzunoğlu authored 4 years ago and Pierre Lamot committed 4 years ago

lower duration, set easing type sine