On 1/5/2018 9:13 PM, Pierre-Louis Bossart wrote:
On 1/5/18 4:06 AM, Sagar Arun Kamble wrote:
On 1/3/2018 1:23 AM, Pierre-Louis Bossart wrote:
On 1/2/18 12:21 PM, Richard Cochran wrote:
On Tue, Jan 02, 2018 at 11:15:45AM -0600, Pierre-Louis Bossart wrote:
I wrote the code for HDaudio and I remember wasting time trying to figure out the gory details of the cycle counter stuff when all I wanted was a conversion from a 24MHz counter to ns values using a 125/3 operation in the right order - as explained in the comments
Would using clocks_calc_mult_shift() work for you?
In theory yes, but I'd need to re-check what the results would be. I remember applying the 1/3 factor separately to avoid wrap-around after 4 hours [1], but I can't remember the details on the analysis. I can't figure out what the 'maxsec' argument should be either.
I am not sure if I understood the wrap-around correctly. Is AZX_REG_WALL_CLK 64bit or 32bit and in the comments which 20 bits are being referred.
it's a 32-bit counter. off the top of my head, the idea was that the integer arithmetic should not degrade the precision (42ns) and that means you need to be careful with the fractional part, especially if the errors with the fractional part accumulate over time (I think this was the case when I looked several years ago). That's the main reason why I did the division by 3 last, after the read, so that the precision is not impacted by the interval between two reads.
timecounter interface ensures to add the fractional ns time so it has good precision. In the table below 3600 indicates interval of 1min between two counter reads and 14400 indicates that interval of 4 minutes. Shift factor controls the ns precision we want. Considering duration of 195 years of interval between counter reads at 24mhz we get mult=83 and shift=1 which will give cycle duration = 41.5. What is the maximum duration of interval over which this read can be done in audio driver.
You also need to be careful with the multiplication factor otherwise you will exceed the 64-bit resolution. For example with the 14400 factor, you cannot handle a counter larger than 2^64/14400, which gives you 14826 hours or 1.69 years. It's one of those 'nobody will ever need more than 640KB' value. The 125 factor gives you 195 years without saturating.
Keeping maxsec at lower value will ensure good precision but the mult factor derived then might lead to overflow if the interval of counter read is big. Keeping maxsec high will reduce the mult factor and will marginally impact the precision (of the order of 6 decimal places of fraction nano second).
For 24mhz clock I am getting following scale factors at different maxsec values. I think these are as good as 125/3 125/3 gives scale factor of 41.666666666666666666666666666667
maxsec, mult, shift, scale factor (mult/(2^shift)) 0, 2796202667, 26, 41.66666667163372039794921875 3600, 87381333, 21, 41.666666507720947265625 14400, 21845333, 19, 41.6666660308837890625
I see sound driver uses only timecounter_read so conversions should be fine. If there are usages of timecounter_cyc2time then we will have to take care of updating the timecounter often as timecounter API internally counts time backwards if counter is spaced more than 1/2 the range.
Thanks Sagar
[1] http://elixir.free-electrons.com/linux/latest/source/sound/hda/hdac_stream.c...