- Aug 02, 2021
-
-
Lyndon Brown authoredthe feature-specific macros each make a call to `vlc_CPU()` which involves making an atomic read of a global variable, then ANDing the return with a specific feature flag and comparing with zero. rather than do that over and over pointlessly, just get the flag set once with a direct call to `vlc_CPU()` and then directly compare with the individual feature flags. since we've removed the build-time conditional definition of the feature test macros which defined them as just returning `1` if matching a build target, this produces no difference in results, whereas if we had not removed that, the results would not have corresponded with build-time force-enabled features.
-
Lyndon Brown authored
the `vlc_CPU_SSE2()` type macros are provided for run-time feature detection. built-time target settings should have no influence on these, except perhaps in rare cases where we happen to know that a feature to absolutely definitely be available and wish to force enable it in the build for that platform. as an example, windows 8+ requires SSE2 (apparently), and so if we only want to support windows 8+ for our windows builds, then we could force enable SSE2 via use of the `-msse2` compile flag. in general though, i'm not aware that we ever do such a thing; i'm not sure we ever will want to, and we can easily hack in a necessary change if we did at such time; it risks mistakes being made; and as things were, this was just adding a lot of unnecessary mess. i find it questionable whether we even need to conditionally define the `VLC_SSE` type defines, but at leats there's only a couple of those. with the arm stuff: - note that while linux specific feature detection works, we are missing detection for other platforms. - `HAVE_FPU` is built-time dependant, which possibly needs replacing with runtime detection?
-
Lyndon Brown authored
- added detection for: FMA3, BMI1, BMI2, POPCNT, LZCNT, AVX512. - added missing detection on non-linux for: AVX, AVX2, FMA4, XOP, SSE4a. (with thanks to libx264 in combination with intel docs). - fixed broken detection of fma3 on linux - the string checked for needs to be "fma" not "fma3" at least on my system. in the latter case, these were already detected on linux via the alternative `/proc/cpuid` based cpuid detection, but were not handled by the generic solution used on other platforms. a xor operation is performed upon ecx to zero it before calling cpuid since when calling cpuid with eax=7 to get extended features, we need to also have ecx=0 to get the right data set for our needs, as detailed on wikipedia ([1]). the added AVX2 detection will enable my AVX2 chroma plugin to work on non-linux systems as well as linux. the rest i have no plans for, but we already have ones that are unused so why not expand coverage. i removed the obsolete `i_` prefix from the integer variables to avoid having to use them in all of the new code added. the AVX512 detection looks for a specific subset of available AVX512 flags (there are several). i've simply copied that used in libx264. i don't expect `__AVX512__` will actually work, but it'll do for now. [1]: https://en.wikipedia.org/wiki/CPUID
-
Lyndon Brown authored
to use SSE you both need a CPU with the SSE feature, and an OS that has support for saving the SSE registers during context switches. whilst detecting SSE CPU support is easy via the `cpuid` instruction and checking flags, it is not easy to check for OS support which involves trying to use an SSE instruction in a process fork, and seeing whether or not this results in SIGILL. (thankfully cpu designers made things easier for AVX & AVX512). since SSE support has been around since 1999, and all operating systems from that time are long since unsupported, we should be able to safely assume now that SSE OS support is available and thus remove this check. note, the `vlc_CPU_check()` function is now only used for ppc and only with `CAN_COMPILE_ALTIVEC` so i've adjusted the preprocessor conditions accordingly.
-
Lyndon Brown authored
the `cpuid` instruction was introduced back in 1993 (according to wikipedia). we can thus safely assume that this exists on all platforms we care about. this also removes the obsolete ability to bypass the level=0 check for certain very old cpus.
-
Lyndon Brown authored
notes: - the `b_amd` detection was only used in connection to MMXEXT so was also removed. - the `goto out`s were replaced with `return 0` to avoid the warning that can now occur as i experienced.
-
Lyndon Brown authored
the cpu SIMD selection code removed here dates from a time when vlc had options for disabling use of select SIMD variants, from before postproc added cpu auto-detection ([1]), and from before postproc seems to have added SSE2 ([2] and [3]). we are purging MMX/MMXEXT from vlc v4.0-dev, and thus have an interest in removing the corresponding MMX/MMXEXT bits here. rather than just removing those lines, alongside adding an entry for SSE2 though, i have instead chosen to convert the code to use auto-detection, which avoids having to keep the block of code explicitly enabling implementations in sync with the set of implementations available. note, the version of postproc in contribs is very old, pre-dating the SSE2 and CPU feature auto-detection enhancements. accordingly i have had to ensure that we define `PP_CPU_CAPS_AUTO` ourselves when not found, as had been done for `PP_CPU_CAPS_ALTIVEC`. effectively, for users like myself on linux with a new enough version, the auto-detection will work correctly and now make use of SSE2, which we were ignoring before; while where the contrib package is used, its use will fall back to the C implementation until such time that the contrib gets updated. [1]: https://github.com/FFmpeg/FFmpeg/commit/59d686f100863d00b8f171dd891e893c2bfd951e [2]: https://github.com/FFmpeg/FFmpeg/commit/4e264d1c79cfae8c3e05aacf77e350ed1b6d7e4b [3]: https://github.com/FFmpeg/FFmpeg/commit/f48cddfe4cf04e2d6e802d12e973301ff5a1a9a8
-
-
- Aug 01, 2021
-
-
the line removed from avcodec appears to just be a stray and obsolete code remenant.
-
- Jul 31, 2021
-
-
00d1f50e introduced a regression for PPC since the `signal()` based detection for altivec in libmpeg2 is not thread safe unlike our own. here we restore use of our own detection for PPC, bypassing libmpeg2's unsafe auto-detection otherwise triggered in `mpeg2_init()`. fixes #25948.
-
Hugo Beauzée-Luyssen authoredThis version has protobuf-compiler prebuilt in order to speed up the android jobs
-
-
-
-
-
-
-
-
-
this will be reused in different parts as well
-
-
defines background color for local toolbar section of BannerSources
-
-
update banner and bgAlt contants
-
use 'bgAlt' constant for background color
-
-
the old text here has been obsolete since we stopped "blindly" copying between video_palette_t and ffmpeg/libavcodec's AVPaletteControl struct in 61146f7e.
-
fixes #25748
-
-
-
recent introduction of Loaders broke the navigation. This patch intends to fix the regression. fixes #25762.
-
drop obsolete entries.
-
- Jul 30, 2021
-
-
fixes cut off edges in maximized window when using CSD
-
-
fixes one of the accessibility issues reported in #25896.
-
in the qt codebase properties are often used but they are not overridden by a derived class. according to the Qt docs, the presence of final attribute can be used for performance optimizations.
-
-
-
-
-
