(In reply to comment #15)
> 
> The good news is that I can reproduce the problem, and it indeed looks like
> the firmware image gets corrupted. But that seems to always happen not only
> when the image is loaded from initrd.

That is a fact.  I run my locally-built kernels without initrd files; the
kernel has an initramfs facility which allows files to be manually selected for
inclusion in the vmlinuz image (I used to use this for uvesafb), and a separate
set of config options (CONFIG_EXTRA_FIRMWARE*, see comment 9 above)
specifically for including firmware files into a particular location in the
initial filesystem.

I have been doing this for several years.  It currently works for UVD on 2 out
of 3 systems.  (A fourth machine has an integrated Radeon GPU too old to be
supported by this UVD code, though it also boots Linux without an initrd.)