Setting the PLL involves the calculation of a fixed point ratio
with 4 decimal digits fraction, referred to as "J.D". The
fraction "D" is stored separately from the integer part "J"
and is limited to 0..9999.

The current algorithm uses integer registers to calculate the
fraction part, but failed to compensate for rounding errors,
resulting in values larger than 9999 for the fraction part
occasionally, e.g. for 44.1kHz audio rate and pll_clkin =
3763400 it would set J to 11 and D to 10002, which will at
best result in wrong pitch.

The critical part is the "pll_clkin / 10000", which would be
ok with real numbers, but using integer arithmetic the rounding
decreases the divisor, thus increasing the final quotient.

The issue is solved by linear interpolation over the reciprocal
function between the two adjacent points with integer divisor,
i.e. pll_clkin / 10000 and pll_clkin / 10000 + 1, and doing
all rounding to the lower result.

As a side effect to the bug fix, the approximation to the
desired frequency is much better, for the above mentioned
example we get 11.9993, while the true ratio is 11.9993623.
Signed-off-by: Oskar Schirmer <oskar@scara.com>
Signed-off-by: Mark Brown <broonie@kernel.org>