nb/intel/x4x: Implement both read and write training
This training find the optimal write DQ delay and read DQS delay
settings. It does so on all lanes at the same time, like
vendor (training each lane individually has poor results).
The results are stored in the sysinfo struct and restored on next
boots and S3 resume.
This potentially increases stability as optimal settings are chosen
and is more necessary for DDR3 raminit where the write DQS delays are
leveled/variable due to the flyby topology.
TESTED on Intel DG43GT with (2G + 1G) on each channel, see that the
results are quite close to the safe original ones (that previous
worked fine) and tested with memtest86+.
Signed-off-by: Arthur Heymans <firstname.lastname@example.org>
Tested-by: build bot (Jenkins) <email@example.com>
Reviewed-by: Felix Held <firstname.lastname@example.org>
4 files changed