[FFmpeg-devel] [PATCH] lavu/hwcontext_vulkan: check both vendor and PCI IDs

[Lynne]

Mar 18, 2024, 07:33 by haihao.xiang-at-intel.com at ffmpeg.org:

> On Ma, 2024-03-18 at 07:18 +0100, Lynne wrote:
>
>> Mar 18, 2024, 06:57 by haihao.xiang-at-intel.com at ffmpeg.org:
>>
>> > On Vr, 2024-03-08 at 15:13 +0800, Xiang, Haihao wrote:
>> > 
>> > > From: Haihao Xiang <haihao.xiang at intel.com>
>> > > 
>> > > Otherwise the derived device and the source device might have different
>> > > PCI ID or vendor ID in a multiple-device system.
>> > > 
>> > > Signed-off-by: Haihao Xiang <haihao.xiang at intel.com>
>> > > ---
>> > >  libavutil/hwcontext_vulkan.c | 31 +++++++++++++++++++++++++++++--
>> > >  1 file changed, 29 insertions(+), 2 deletions(-)
>> > > 
>> > > diff --git a/libavutil/hwcontext_vulkan.c b/libavutil/hwcontext_vulkan.c
>> > > index 855f099e26..9d94f74d78 100644
>> > > --- a/libavutil/hwcontext_vulkan.c
>> > > +++ b/libavutil/hwcontext_vulkan.c
>> > > @@ -975,6 +975,20 @@ static int find_device(AVHWDeviceContext *ctx,
>> > > VulkanDeviceSelection *select)
>> > >                 select->name);
>> > >          err = AVERROR(ENODEV);
>> > >          goto end;
>> > > +    } else if (select->vendor_id && select->pci_device) {
>> > > +        av_log(ctx, AV_LOG_VERBOSE, "Requested vendor:device
>> > > %04x:%04x\n",
>> > > +               select->vendor_id, select->pci_device);
>> > > +        for (int i = 0; i < num; i++) {
>> > > +            if (select->vendor_id == prop[i].properties.vendorID &&
>> > > +                select->pci_device == prop[i].properties.deviceID) {
>> > > +                choice = i;
>> > > +                goto end;
>> > > +            }
>> > > +        }
>> > > +        av_log(ctx, AV_LOG_ERROR, "Unable to find device with vendor ID
>> > > 0x%x
>> > > "
>> > > +               "and PCI ID 0x%x!\n", select->vendor_id, select-
>> > > >pci_device);
>> > > +        err = AVERROR(EINVAL);
>> > > +        goto end;
>> > >      } else if (select->pci_device) {
>> > >          av_log(ctx, AV_LOG_VERBOSE, "Requested device: 0x%x\n", select-
>> > > > pci_device);
>> > >          for (int i = 0; i < num; i++) {
>> > > @@ -1597,8 +1611,14 @@ static int vulkan_device_derive(AVHWDeviceContext
>> > > *ctx,
>> > >  #if CONFIG_VAAPI
>> > >      case AV_HWDEVICE_TYPE_VAAPI: {
>> > >          AVVAAPIDeviceContext *src_hwctx = src_ctx->hwctx;
>> > > -
>> > > -        const char *vendor = vaQueryVendorString(src_hwctx->display);
>> > > +        VADisplay dpy = src_hwctx->display;
>> > > +#if VA_CHECK_VERSION(1, 15, 0)
>> > > +        VAStatus vas;
>> > > +        VADisplayAttribute attr = {
>> > > +            .type = VADisplayPCIID,
>> > > +        };
>> > > +#endif
>> > > +        const char *vendor = vaQueryVendorString(dpy);
>> > >          if (!vendor) {
>> > >              av_log(ctx, AV_LOG_ERROR, "Unable to get device info from
>> > > VAAPI!\n");
>> > >              return AVERROR_EXTERNAL;
>> > > @@ -1607,6 +1627,13 @@ static int vulkan_device_derive(AVHWDeviceContext
>> > > *ctx,
>> > >          if (strstr(vendor, "AMD"))
>> > >              dev_select.vendor_id = 0x1002;
>> > >  
>> > > +#if VA_CHECK_VERSION(1, 15, 0)
>> > > +        vas = vaGetDisplayAttributes(dpy, &attr, 1);
>> > > +        if (vas == VA_STATUS_SUCCESS && attr.flags !=
>> > > VA_DISPLAY_ATTRIB_NOT_SUPPORTED) {
>> > > +            dev_select.vendor_id = ((attr.value >> 16) & 0xFFFF);
>> > > +            dev_select.pci_device = (attr.value & 0xFFFF);
>> > > +        }
>> > > +#endif
>> > >          return vulkan_device_create_internal(ctx, &dev_select, 0, opts,
>> > > flags);
>> > >      }
>> > >  #endif
>> > > 
>> > 
>> > Hi,
>> > 
>> > Any comment for this patch ? 
>> > 
>>
>> Is this possible? For two devices from different vendors to have the same PCI
>> ID?
>>
>
> I'm not sure for different vendors. But when two devices come from the same
> vendor (e.g. 8086:56a0 and 8086:4682), vulkan always picks up 8086:56a0 no
> matter the source device is 8086:4682 if we don't set the device id here.
>

Do you need to check for both vendor and PCI ID?
If not, could you resend without the custom vendor+pci check
and just using the regular pci ID check?