Provide routines for sad functions for high bit depth, i.e. 10 bits.
Benchmarks run on AWS Gravtion 2 instances.

sad_4x4_c: 583
sad_4x4_neon: 273
sad_4x8_c: 1179
sad_4x8_neon: 366
sad_4x16_c: 2121
sad_4x16_neon: 550
sad_8x4_c: 924
sad_8x4_neon: 213
sad_8x8_c: 1711
sad_8x8_neon: 316
sad_8x16_c: 3505
sad_8x16_neon: 497
sad_16x8_c: 3070
sad_16x8_neon: 635
sad_16x16_c: 6113
sad_16x16_neon: 1118

Signed-off-by: Hubert Mazur <hum@semihalf.com>
Signed-off-by: Grzegorz Bernacki <gjb@semihalf.com>

On most systems any whitespace is fine, but MSYS2 wants ASCII 0x20.

Without the change parallel build occasionally fails as:

    $ make --shuffle
    ...
    gcc ... -c common/opencl.c -o common/opencl-8.o ...
    common/opencl.c:116:10: fatal error: common/oclobj.h: No such file or directory
      116 | #include "common/oclobj.h"
          |          ^~~~~~~~~~~~~~~~~

Best reproducible with `make --shuffle` mode:
   https://savannah.gnu.org/bugs/index.php?62100

This happens because `common/oclobj.h` is an autogenerated file.
Normally `.depend` would contain this autogenerated dependency.
But nothing forces `common/oclobj.h` to be generated.

The change moves dependency of $(GENERATED) from final binaries
to `.depend` itself:

    .depend: $(GENERATED)

Use pkg-config from the custom PATH.

Use perl for in-place editing because sed doesn't work with symlinks.

When operating on large blocks of data it's common to repeatedly use
an instruction on multiple registers. Using the REPX macro makes it
easy to quickly write dense code to achieve this without having to
explicitly duplicate the same instruction over and over.

For example,

    REPX {paddw x, m4}, m0, m1, m2, m3
    REPX {mova [r0+16*x], m5}, 0, 1, 2, 3

will expand to

    paddw       m0, m4
    paddw       m1, m4
    paddw       m2, m4
    paddw       m3, m4
    mova [r0+16*0], m5
    mova [r0+16*1], m5
    mova [r0+16*2], m5
    mova [r0+16*3], m5

Correctly handle emulation of 4-operand instructions (e.g. 'shufps')
where src1 is a memory operand.

With legacy encoding the last operand (the index) must be xmm0,
but aside from that emulating non-destructive forms works
the same as any other instruction.

Jessica Clarke authored Dec 19, 2021 and Anton Mitrofanov committed Dec 30, 2021

Building a shared library without -fPIC does not make sense. On most
architectures, especially recent ones, doing so will give link-time
errors due to relocations in read-only sections like .text. On some
legacy architectures, including i386, it is allowed by default, but will
warn, and is highly discouraged due to the overheads it adds at library
load time. Most architectures were already listed here as having shared
imply PIC, but not all, such as i386 which ends up with unwanted text
relocations, as well as architectures not known to the build system
currently like RISC-V, which does not permit text relocations by
default. There is no good reason to want shared without PIC on any
architecture, so just remove the architecture list.

Back in 2009 when this was added it improved scheduling of lookahead
threads on prevalent operating systems at the time.

According to more recent testing by Intel however, lowering thread
priorities does not improve performance on modern operating systems.
And more importantly, doing so on systems with heterogeneous CPU
topologies may actually result in a severe performance reduction.

Removing this code altogether eliminates the issue with performance
degradation on such systems, while having no noticeable impact on
regular systems with homogeneous CPU topologies.

/usr/ucb/bin/install on Solaris does not support creating multiple
directories in one go, issue multiple install commands instead.

Jean-Baptiste Kempf authored Sep 28, 2021 and Anton Mitrofanov committed Sep 29, 2021

The lookahead_thread main loop checks b_exit_thread and exits if it is set.
That flag is set by x264_lookahead_delete, which uses ifbuf.mutex to guard accessing it.
However, the read in the while-loop condition of lookahead_thread is not guarded,
and so TSAN sometimes reports a data race.

Martin Storsjö authored Sep 26, 2021 and Anton Mitrofanov committed Sep 29, 2021

This fixes rerunning checkasm with an earlier printed seed, when
it's outside of the signed range.