Skip to content

tabby-v1.2

Choose a tag to compare

@SeattleUser0 SeattleUser0 released this 20 Nov 10:59
· 6 commits to classic-tabby/linux-build since this release

Messages moved to the top of the screen, so they never cover the pads' cells. The firmware can now give back the pixels it sent to its screen; the release test compares them with the simulator's on every page, and the manual's pictures are the device's own. Reinstall to get it (Full backup first; storage does not move).

Classic Tabby tabby-v1.2: messages at the top; the screen proven on the device

  • Firmware package build ID: fef558ab16e275b4 (first 16 hex digits). App image 488548 bytes (measured), margin 30909 bytes
    below the 519457-byte limit (docs/FLASH-BUDGET.md 51). From tabby-v1.1 (f0bd5047ca742907): +1824 bytes. At least
    15360 bytes stay free for fixes. Free RAM 143244 bytes (192 fewer).
  • Installing over tabby-v04.2 to tabby-v1.1 keeps storage as it is. Take a Full backup first anyway.
  • Going back to older firmware: songs saved on tabby-v04.10 or later cannot be read by older firmware. Go back to
    an older build only with a Full backup made before tabby-v04.10: install the older build, then restore that backup.
    A clip is refused (its track is silent) by firmware that does not know one of its settings: TRANSPOSE before
    tabby-v04.15, SHUFFLE and PLAY QUANTIZE before tabby-v04.17, STEP SIZE before tabby-v04.19, DIRECTION before
    tabby-v04.20. A song with EVENT LEAD set plays on older firmware, without the lead.
  • Released through the candidate flow, as candidate 7. Candidate 1 was held back: its screen copy asked the device
    for one block of 57600 bytes and the device's heap refused. Candidate 2 (a smaller copy) was held back: one line
    measured 2.01 and 2.03 ms on two of eight pages, and its screen copy showed why the device never finished a
    picture while a song played (below). Candidate 3 (no more repaint on every step) was held back: the pages were still
    repainted on every eighth note, now exactly on the notes, and one line measured 2.3 to 2.5 ms on CLIPS and
    SCENES. Candidate 4 has the fix for that and for the standing CPU GUARD box, was installed and tested and made
    the device's screen pictures. Candidate 5 (candidate 4's firmware, those pictures in the manual) was held back:
    its own pixel comparison found leftover pieces of the page before on HOME, and eight lines on SCENES measured
    3.9 ms once. Candidate 6 has both fixed and passed everything but one soak segment, in which the
    song stayed in a scene: a fault of the song player at SEQ PLAY that is as old as song mode (below). Candidate 7
    has it fixed, was tested in full, and is the one promoted. Its manual pictures are the ones taken on candidate
    4's build; its own release run compared its own pixels again. docs/BUILD.md 65.1 to 65.7.

What changed

  • A message is an amber box across the top, over the top line (in tabby-v1.1 it was across the bottom and
    covered the lower pad row and part of the queue line of tracks 5 to 8). One line covers the top line; two or
    three lines also cover the title row and the row under it. It stays about a second and a half, as before. The
    KNOB BANK list stays at the bottom and is readable under a shortcut's message.
  • CLIPS and SCENES are no longer painted again on every 16th note while a song plays. Found by the screen
    copy: the page's own "did anything change" test included the step counter, so the whole page (15 bands) was
    repainted eight times a second, more than the painter manages beside the MIDI sends (84 bands a second): a
    picture was never finished, and every note had a band in its way. It has been so since these pages exist. Now a
    page is repainted when something it shows changes (the scene, a launch, a queue); the bar count repaints only its
    own digits; on STEP only the rows of the step cells follow the running step. The pads' blink phase, which
    repainted the page on every eighth note without anything on the screen blinking, is out of that test too.
  • A song could start wrong at SEQ PLAY: fixed. Rarely (once in the roughly 300 starts of this release's test
    runs; never seen before, though the code was the same), the song played scene 1, then scene 2, then scene 2 for
    ever, and sent its notes on the device's MIDI channels instead of the song's. The control task looked at "is it
    playing?" and then at "is a start pending?"; when the audio interrupt took the start between those two looks,
    it saw neither, took the song for ended and let go of it while it played on. The two looks are in the safe order
    now, and a test puts the interrupt at that very place. If this ever happened to you on stage it looked like a
    song stuck in its second scene with some parts on the wrong synth; stopping and starting cured it.
  • A band left over from the page before is painted within a few passes. Found by the pixel comparison on
    candidate 5: 2.4 seconds after PATCH from CLIPS to HOME, while a song played, the project name's row still showed
    pieces of CLIPS. On HOME the level meters ask for their bands on every frame and were served first every time.
    Every second band now goes in plain rotation.
  • A page is not composed when a note is due before the composing would end (as for a band): SCENES is composed
    again at each new scene, the moment its first notes are due.
  • CPU GUARD - VOICES ADAPTED is no longer a message box. Seen in the device's own pictures: with eight parts on
    the internal sounds the guard takes voices back about once a second, and its message stood over the page for the
    whole song (every Wrong Room soak since tabby-v1.0 counted 125 of them in 125 seconds; nobody was looking at the
    screen). It is CPU! in amber in the top line while it happens. AUDIO OVERLOAD and MIDI FULL stay message boxes.
  • The wait before a due MIDI note counts the band's own last painting time, not only that of the band before it
    (large text makes bands differ more: 0.6 to 2.5 ms).
  • Kept as they were, by your answers: MASTER MUTED small in the top line (K8=0); no pattern cells and no knob strip
    on HOME.
  • The screen copy (remote command 29 {19}; it only reads). The LCD cannot be read back: one data line, driven by
    the SoC only. So the firmware can keep a copy of every rectangle of pixels it hands to the LCD, freeze it when the
    picture is whole, and send it over USB. It takes 29 KB of memory, in fifteen small blocks, only while a test has
    switched it on.
  • The comparison (a step of every release run from now on): one tour of every page of docs/CONTROL-MAP.md
    section 20, 22 pictures at rest and 22 while the song plays, run on the device with remote panel keys and in the
    simulator with the same song at the same place in the song; every pixel compared. Left out, and named in the
    result: the level meters and MIDI-in marks on HOME (they follow the audio), FREE / RECL on ACTIONS and the stored
    project name on PRESETS and SAVE PROJECT (the device holds your data, the simulator starts empty).
  • The timing figures: this computer measures when a note arrives; the device also counts, by its own clock, how
    long after its due time each note left. From this release the release run reads both, and one line must be
    under 2 ms by both. The test computer is kept awake while it measures: idle, its own timestamps wander by 0.1 to
    0.3 ms (docs/BUILD.md 65.4).
  • The manual's pictures are the device's, taken by that test, in place of simulator renders. Two are still the
    simulator's and say so: SAVE PROJECT and STOP TO SAVE need the > button, which the device does not let a remote
    command press.
  • What this does not prove: the last step from the LCD's input to the glass (colours, contrast, viewing angle, a
    dead pixel), and the start screen (gone before a command can ask). Test T97 of the guide is that look, by you.

Validation

Candidate 7 (builds/tabby-v1.2-rc7/) on the device, unattended, before the promotion. Times at the test computer
were taken with its cores kept awake; "own" is the device's own figure: the latest a timed event left after its
due time.

  • The device's pixels: 44 of 44 pictures identical to the simulator's, 22 at rest and 22 while the song played.
    Left out, by name: the level meters and MIDI-in marks on HOME, FREE / RECL on ACTIONS, the stored project name on
    PRESETS, the last played note's name on STEP, and the unsaved-project mark and CPU! in the top line where the two
    differ. Not taken on the device: SAVE PROJECT and STOP TO SAVE (the > button).
One line, 960 notes DIN ms USB ms own ms
Overview (three runs) 1.66, 1.75, 1.85 1.62, 1.60, 1.62 1.49, 1.51, 1.48
CLIPS 1.75 1.60 1.49
SCENES 1.68 1.68 1.44
SYNC / MIDI 1.79 1.59 1.44
ACTIONS 1.74 1.64 2.96
Mixer 1.79 1.62 1.45
Eight lines, worst line DIN ms USB ms own ms
Overview 3.04 2.35 2.21
CLIPS 2.27 2.17 2.19
SCENES 2.36 2.24 2.12
  • One line stays under 2 ms on every page, at the test computer and by the device's own figure (tabby-v1.1, own:
    CLIPS 1.66, SCENES 1.87, mixer 1.49). On ACTIONS the device's own figure is about 3 ms as in every version (the
    page reads the song storage; the line's notes arrive within 1.74 ms); it is reported, not judged.
  • Soak with EVENT LEAD OFF, the Wrong Room song looping, 40 minutes, 19 segments on six pages, none bad: every
    scene entry sent its events; USB band 1.73 to 2.19 ms; on the one DIN cable 14.88 to 15.79 ms, as before.
  • Soak with EVENT LEAD 1/16, 40 minutes, 19 segments, none bad: DIN band 1.91 to 2.33 ms, USB 1.75 to 1.98 ms.
  • CLIP SETTING 40 of 40, take edit 26 of 26, note repeat and step loop 20 of 20, program and knobs 10 of 10, clip
    chains 11 of 11, scene chain 13 of 13, scene edit 16 of 16, recording 25 of 25, STEP page 18 of 18, CLIPS page 40
    of 40 twice, the timing song 960 of 960 notes (worst 1.16 ms).
  • Full backup, restore of that backup, verify: PASS (41 items). The panel was left as it was found.
  • Still not looked at by a person: the glass itself. Test T97 of the guide.