"unselecting" left-hand rows is a wasted i2c transaction.
On the left-hand side, the ergodox uses a GPIO expander. It
does *not* change "direction" (input/output) of pins, it just
sets pins high or low.
But all the pins are written at once. There's no way to
change just one pin's value; you send a full byte of all eight
row pins. (Not all of them are in use, but that doesn't matter.)
So every pin is either +V or ground. This is in contrast
with the right-hand side, which is using input mode to make pins
be neutral.
So there's no need to "deselect" the rows on the left side
at all. To select row 0, you set the GPIO register for the
rows to 0xFE. The previous code would then set it back to
0xFF, then set it to 0xFD on the next cycle. But we can just
omit the intervening step, and set it to 0xFD next cycle,
and get the same results.
And yes, I tested that the keyboard still works.
On my system, scan rate as reported by DEBUG_SCAN_RATE goes
from 445 or so to 579 or so, thus, from ~2.24ms to ~1.73ms.
Signed-off-by: seebs <seebs@seebs.net>
With these changes, the ergodox ez goes from 315 scans per second
when no keys are pressed (~3.17ms/scan) to 447 (~2.24ms/scan).
The changes to the pin read are just condensing the logic, and
replacing a lot of conditional operations with a single bitwise
inversion.
The change to row scanning is more significant, and merits
explanation. In general, you can only scan one row of a keyboard
at a time, because if you scan two rows, you no longer know
which row is pulling a given column down. But in the Ergodox
design, this isn't the case; the left hand is controlled by an
I2C-based GPIO expander, and the columns and rows are *completely
separate* electrically from the columns and rows on the right-hand
side.
So simply reading rows in parallel offers two significant
improvements. One is that we no longer need the 30us delay after
each right-hand row, because we're spending more than 30us
communicating with the left hand over i2c. Another is that we're
no longer wastefully sending i2c messages to the left hand
to unselect rows when no rows had actually been selected in the
first place. These delays were, between them, coming out to
nearly 30% of the time spent in each scan.
Signed-off-by: seebs <seebs@seebs.net>
This is particularly relevant for, e.g., the ergodox EZ and
other keyboards with slow scan rates. Without changing the API or
behavior of individual process_record() calls, we allow a
configuration flag to make multiple calls in a single scan.
This will probably have miniscule effects on non-steno users,
and it's not enabled by default for any keyboards. Added note
about it to ergodox README.
Signed-off-by: seebs <seebs@seebs.net>
Plover's steno support likes quasi-qwerty, and gaming likes qwerty,
and I like dvorak, so... what if I could have it all?
Signed-off-by: seebs <seebs@seebs.net>
Seebs