New version: 3.7.0, 4th November, 2013New drivers, lots of code updates, new features, lots of bugfixes, major upgrade. The new drivers are NOT built into the binaries since KnC hardware only runs on beaglebone boards and Hashfast hardware isn't in the wild yet. Binaries now include klondike support on windows as well.
Human readable changelog:- Driver for KnC miner ASIC hardware.
- Driver for Hashfast ASIC hardware.
- Fix for blue fury to work as well as red fury.
- Busloads of little tweaks to make AMU much more reliable on windows - they may go zombie if you have unreliable communications but should hotplug themselves again soon after. Thanks greatly to the generous testing time dedicated so far by aigeezer and jmc1517. No doubt they will still be more reliable on linux, but hopefully we should be close to getting them reliably hashing in a sustained fashion on windows.
- Major rewrites of all the "diff" code in preparation for endless changes in diff and hashrates. It should now seamlessly support any diff levels (even microdiffs for development) for network, pool and share diff, while being lower overhead than before. Diff will now be rounded to nearest integer instead of down as well (63.99 will be 64, not 63).
- Fix for scrypt showing a block solve with every share.
- Faster flushing of work across all devices on block change should lead to lower rejects than ever on devices that can abort work quickly.
- Numerous usb communication improvements should lead to less software induced artificial hardware errors.
- Lower CPU usage from multiple areas.
- Current block now only shows the first 8 characters after the paired zeroes end.
- Accepted share will trim off all paired zeroes.
- Display the amount of work items worked on per pool now in the API and summary on exit for when load balance with quotas is in use to allow comparison.
- Account for discarded work when calculating quotas to try and keep them as close to set values as possible.
- Know which pools are on the new block during a block change and decide accordingly whether to update work depending on the pool failover strategy.
- Changed the verbose message to share is above target instead of below since it was incorrect.
- Fix for the corrupted console output after shutting down cgminer if we've done one or more restarts.
- Fix for the rare scenario where it would stall completely and no work would be grabbed or done.
- Klondike driver updates.
- Numerous low level features for upcoming drivers.
- Lots of low level code updates and bugfixes.
Full changelog:- Use WRITEIOERR macro check for all usb writes.
- Always use a usb read buffer instead of having to explicitly enable it.
- Force unlocking of the console lock on restart to avoid corrupting the console
state when we finally quit.
- Never wait indefinitely for a pthread conditional in the hash_pop loop in case
the work scheduler misses the last wakeup.
- Make hash_pop signal the work scheduler each time it waits on the conditional
that it should look for more work.
- Discriminate between libusb transfer errors and regular libusb errors and make
sure to capture them all.
- Always read a full sized transfer for bulk reads.
- Deprecate preferred packet size functions in usbutils since they're unhelpful.
- Copy known transferred amount back to buffer for usb reads instead of
requested length.
- Treat timeout errors on usb writes as IO errors.
- Ignore iManufacturer from bitfury devices to support bluefury as well as
redfury.
- Add more debugging info for when usb details don't match.
- Look for timeout overruns in usb read/write.
- Use an int for usb_read/write to identify overruns.
- Use the callback timeout as a safety mechanism only on windows.
- Instead of using complicated sleeps to emulate characters per second on usb
writes, submit only as many characters as can be transferred per usb poll of
1ms, and use timeouts in bulk transfers, cancelling transfers only as a
failsafe.
- Remove discarded work from quota used.
- Display works completed in summary and API data.
- Store how many work items are worked on per pool.
- Make each pool store its on reference for what the most current block is and
fine tune management of block change in shared pool failover strategies using
the information.
- Rationalise use of current_hash to a single hex string the length of the
previous block and display only the first non zero hex chars of the block in the
status window.
- Update uthash to latest.
- show_hash doesn't know the size of the string so hard code the max size.
- Remove as many initial zeroes as exist on share display, abstracting out a
hash show function to use across different submission mechanisms.
- Add missing endian swap functions for 64bits.
- Sanity check for absurd target setting and divide by zero.
- Abstract out conversion of a 256 bit endian number to a double, correcting
errors and use it for determining any magnitude share diff.
- Avoid the extra generation of a byte flipped hash2 in struct work and directly
use the LE work hash.
- Add a sanity check to avoid divide by zero crashes in set_target
- Calculate diff from target accurately for all 256 bits.
- Set a true 256bit binary target based on any diff value in set_target()
- Provide a copy_work_noffset function for copying a work struct but changing
its ntime.
- Make calls to flush queue and flush work asynchronous wrt to the main work
loops.
- Share is also above target for submit noffset nonce.
- Use round for displaying current pool diff.
- Use round for stratum share diff display instead of floor.
- Use round instead of floor for displayed pool difficulty.
- Allow arbitrary diffs to be tested against nonces via a test_nonce_diff
function.
- Abstract out the rebuilding of hash2 in work.
- Share is above, not below target, when it doesn't meet it.
- Add the ability to add uint8 and uint16 entities to api data.
- Use a non blocking connect with a 1 second select timeout when initiating
stratum to allow us to iterate over all IPs returned by getaddrinfo in round
robin DNS pools.
- Minor style changes to output.
- Revert two different hash_sequence(_head)'s to one variable, use
HF_SEQUENCE_DISTANCE in both places
- Remove duplicate HF_SEQUENCE_DISTANCE() macro, and duplicate hash_sequence
from info structure
- Change SEQUENCE_DISTANCE() macro to HF_SEQUENCE_DISTANCE()
- Structure changes for OP_NONCE, add big endian header
- klondike - initialise stat_lock
- klondike - better to unlock locks than to lock them twice
- Add copyright notice to knc driver.
- Trivial style changes to knc driver.
- Improve performance of work generation by optimizing hex2bin and bin2hex
- klondike - change options to clock and temptarget only
- klondike - fix another uninit dev warning
- klondike - downgrade 'late update' but add an idle detect - and correct error
levels
- klondike - fix isc uninit warning
- Use a mutex to protect data in the knc structure, to prevent loading more work
during a flush, and unlock and return to main between calls to get_queued_work.
- Use the existing device_data for knc state data.
- Only count successful nonces as hashrate in the knc driver.
- Fix trivial warnings in knc driver.
- Add KNC to api
- klondike - drop the device for hotplug if it's unresponsive
- usbutils - usb_nodev() allow a driver to drop a device
- klondike - single 'shutdown' and ensure it happens
- klondike remove SCNu8 - unsupported on windows
- Correctly calculate sleep_estimate in usbutils that may have been preventing
usecps from working.
- Use a sanity check on timeout on windows.
- Better HW error count; disable permanently those cores which fail often
- KnC driver: knc-spi-fpga ASIC driver
- Fixup jansson & libusb include paths when using separate build directory
- 'llround' is more suitable here than 'roundl'
- Silence warning if MAX/MIN is already defined
- Remove prebuild ccan/opt dependencies
- Reinstate block solve testing.
- Dramatically simplify the calculation of blockdiff.
- Simplify the set_target function, allowing it to work properly for fractional
diffs.
- Merge hashfast driver
- Merge KnC driver
Show Posts
