Running any OpenCL program on a big endian machine using a radeonsi card fails.

- union si_vgt_param_key uses a bitfield shared with a 32 bit value. This
causes the index to be wrong on a big endian machine, and crashes any OpenCL
program with a segfault.
- si_setup_user_sgprs_co_v2 in si_compute.c fills a dispatch_packet struct with
values which are uploaded to the GPU. To ensure this works correctly, these
values need to be converted to little endian.

I'll send patches to the mesa-dev list for both these issues.

Bas Vermeulen