[FFmpeg-devel] [PATCH 1/2] libavutil/hwcontext_vulkan: Fix VK_FORMAT_R8G8_UNORM and VK_FORMAT_R16G16_UNORM map problem on Vulkan

[Lynne]

13 Apr 2022, 08:05 by haihao.xiang-at-intel.com at ffmpeg.org:

> On Wed, 2022-03-30 at 11:41 +0800, Wenbin Chen wrote:
>
>> Vulkan map both DRM_FORMAT_RG88 and DRM_FORMAT_GR88 to
>> VK_FORMAT_R8G8_UNORM. This cannot distinguish nv12/nv24 and nv21/nv42.
>> Vulkan also map both DRM_FORMAT_RG1616 and DRM_FORMAT_GR1616 to
>> VK_FORMAT_R16G16_UNORM. This causes issue when map back.
>> VK_FORMAT_R16G16 will be mapped to DRM_FORMAT_GR1616, while p010 need
>> DRM_FORMAT_RG1616.
>> Add sw_format check to vulkan_fmt_to_drm() to fix this problem.
>>
>> Signed-off-by: Wenbin Chen <wenbin.chen at intel.com>
>> ---
>>  libavutil/hwcontext_vulkan.c | 13 +++++++++++--
>>  1 file changed, 11 insertions(+), 2 deletions(-)
>>
>> diff --git a/libavutil/hwcontext_vulkan.c b/libavutil/hwcontext_vulkan.c
>> index 237caa4bc0..791649001f 100644
>> --- a/libavutil/hwcontext_vulkan.c
>> +++ b/libavutil/hwcontext_vulkan.c
>> @@ -3290,8 +3290,17 @@ static void vulkan_unmap_to_drm(AVHWFramesContext
>> *hwfc, HWMapDescriptor *hwmap)
>>  av_free(drm_desc);
>>  }
>>  
>> -static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt)
>> +static inline uint32_t vulkan_fmt_to_drm(VkFormat vkfmt,
>> +                                         enum AVPixelFormat sw_format)
>>  {
>> +    if (vkfmt == VK_FORMAT_R8G8_UNORM &&
>> +       (sw_format == AV_PIX_FMT_NV12 || sw_format == AV_PIX_FMT_NV24))
>> +            return DRM_FORMAT_RG88;
>> +
>> +    if (vkfmt == VK_FORMAT_R16G16_UNORM &&
>> +       (sw_format == AV_PIX_FMT_P010 || sw_format == AV_PIX_FMT_P016))
>> +            return DRM_FORMAT_RG1616;
>> +
>>  for (int i = 0; i < FF_ARRAY_ELEMS(vulkan_drm_format_map); i++)
>>  if (vulkan_drm_format_map[i].vk_format == vkfmt)
>>  return vulkan_drm_format_map[i].drm_fourcc;
>> @@ -3373,7 +3382,7 @@ static int vulkan_map_to_drm(AVHWFramesContext *hwfc,
>> AVFrame *dst,
>>  };
>>  VkFormat plane_vkfmt = av_vkfmt_from_pixfmt(hwfc->sw_format)[i];
>>  
>> -        drm_desc->layers[i].format    = vulkan_fmt_to_drm(plane_vkfmt);
>> +        drm_desc->layers[i].format    = vulkan_fmt_to_drm(plane_vkfmt, hwfc-
>> >sw_format);
>>  drm_desc->layers[i].nb_planes = 1;
>>  
>>  if (drm_desc->layers[i].format == DRM_FORMAT_INVALID) {
>>
>
> Patchset LGTM
>

I'll hold off on pushing all of your and Wenbin's Vulkan patches for now.

Your hacks for cheap hardware have seriously broken image
and memory allocation for anyone who doesn't care about
your niche use cases, and of course it's up to me to salvage it.
What's worse, I don't know if it's your drivers that are at fault,
or if it's your patches that are at fault, or both.
I already had to rip all of your patches off just to get Vulkan
Video decoding working.

Pushing any more code will just create even more mess for me
to deal with, and I'm already knee-deep in it trying to work around
the Khronos user synchronized object madness.

You'll just have to wait for me to fix them, unless you think you
can fix them yourselves, which I doubt. I may have to partially
revert your patches too, because right now, everything's
pretty broken.