[FFmpeg-devel] [PATCH 01/18] hwcontext_vulkan: disable descriptor buffer extension on Intel

[Jerome Martinez]

I tested the patches as a whole, for the moment with a focus on the 
performance.

Tested with a RTX 4070 Ti on Linux & a RTX 3050 on Windows, globally no 
regression in speed on both test platforms, and I find a +25% (up to 
+35% in some cases, especially high resolutions) in average with 10-bit 
or 16-bit content for most cases (it is weird sometimes, no speed 
improvement, I'll check).
Easy to decode complex 4K 10-bit in real time with the RTX 4070 Ti now, 
which is good performance for lossless compressed content.

Additionally, the cards can handle higher resolutions e.g. 12k9k16bit 
succeeds now even on the low end RTX 3050 (at 0.5 fps :) ).


Le 12/04/2025 à 09:22, Lynne a écrit :
> Temporary workaround. Will be replaced with a version check once a fix is
> in the works and a known next version for Mesa with a fix is known.
> ---
>   libavutil/hwcontext_vulkan.c | 5 +++++
>   1 file changed, 5 insertions(+)
>
> diff --git a/libavutil/hwcontext_vulkan.c b/libavutil/hwcontext_vulkan.c
> index 319b71ed04..d11c0274d2 100644
> --- a/libavutil/hwcontext_vulkan.c
> +++ b/libavutil/hwcontext_vulkan.c
> @@ -773,6 +773,11 @@ static int check_extensions(AVHWDeviceContext *ctx, int dev, AVDictionary *opts,
>           tstr = optional_exts[i].name;
>           found = 0;
>   
> +        /* Intel has had a bad descriptor buffer implementation for a while */
> +        if (p->vkctx.driver_props.driverID == VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA &&
> +            !strcmp(tstr, VK_EXT_DESCRIPTOR_BUFFER_EXTENSION_NAME))
> +            continue;
> +
>           if (dev &&
>               ((debug_mode == FF_VULKAN_DEBUG_VALIDATE) ||
>                (debug_mode == FF_VULKAN_DEBUG_PRINTF) ||