The baudrate in rkspi was calculated by using an integer division
(which implicitly discarded any fractional result), then rounding to
an even number and finally clamping to 0xfffe using a bitwise AND
operator. This introduced two issues:
1) for very small baudrates (overflowing the 0xfffe range), the
bitwise-AND generates rather random-looking (wildly varying)
actual output bitrates
2) for higher baudrates, the calculation tends to 'err towards a
higher baudrate' with the actual error increasing as the dividers
become very small. E.g., with a 99MHz input clock, a request
for a 20MBit baudrate (99/20 = 4.95), a 24.75 MBit would be use
(which amounts to a 23.75% error)... for a 34 MBit request this
would be an actual outbout of 49.5 Mbit (i.e. a 45% error).
This change rewrites the divider selection (i.e. baudrate calculation)
by making sure that
a) for the normal case: the largest representable baudrate below the
requested rate will be chosen;
b) for the denormal case (i.e. when the divider can no longer be
represented), the lowest representable baudrate is chosen.
Even though the denormal case (b) may be of little concern in real
world applications (even with a 198MHz input clock, this will only
happen at below approx. 3kHz/3kBit), our board-verification team kept
complaining.
Signed-off-by: Philipp Tomsich <philipp.tomsich@theobroma-systems.com>
Tested-by: Klaus Goger <klaus.goger@theobroma-systems.com>
9fc354e246