On nv50, this will be needed by applications using CUDA to know
how much stack/local memory to allocate.
Signed-off-by: Marcin Kościelnicki <koriakin@0x04.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>