what does this mean?

mfxVariant Data is an c-like union. each configuration parameter specifically fills its defined type in union and mfxImplDescription.AccelerationMode has U32 type value

so why do we handle only that case?

because accel_mode.Data is c-union, and only one member U32 is correct for the value determined by "mfxImplDescription.AccelerationMode" description

let's look at this as std::variant<...> and as VPL mentioned in their Docs, value of type of mfxImplDescription.AccelerationMode should be stored as U32.

this case is handled in previous constuctor, too - should that logic be unified?

the first ctor uses device creation from default adapter - no external pointers from user
but this ctor just wrap user-created device passed by void* pointer AND uses actual device type through CfgParams to restore type and work out some checks

Should the full test bodies be enclosed under HAVE_ macros?

I mean, once and for all

..or the idea is to extend this with other flags as they arrive?

I'd prefer different tests cover different cases otherwise you can't say for sure what actually passed/failed in the test looking at its name when it has that lots of compile conditionals inside

Do you mean
TEST(OneVPL_Source_Device_Selector_CfgParam, PtrDeviceFromCfgParam)
then
TEST(OneVPL_Source_Device_Selector_CfgParam, DX11_PtrDeviceFromCfgParam)
then
TEST(OneVPL_Source_Device_Selector_CfgParam, YYY_PtrDeviceFromCfgParam)

agree, it makes sense

yes