Grid refactor

[christianparpart]

goals

• fix existing grid reflow related bugs and SEGV
• greatly improve text throughput performance
• reduce complexity

This PR also fixes: #371

Checklist

• ESC c (RIS, hard reset) not working when LF'ing is moving lines into scrollback.
• Change VT Parser to use char instead char32_t #471 VT Parser: use char instead of char32_t, later in Ground state and OSC (/DCS) payloads when UTF-8 is to be expected.
• Change VT Parser to use char instead char32_t #471 Parser: see if we can optimize the FSM to lower the L1/L2 cache pressure.
• Refactor Screen API to only deal with one grid and Terminal API has 2 Screen instances then. #472 - Screen: refactor Screen API to only know about one Grid, such that we can have a specialized Screen for primary screen (very thin grid cells) and another specialized Screen for the alternate screen whereas ThinCell is tuned for maximum throughput and FatCell is tuned for fast access of all attributes. - This requires factoring out all that is shared between the two screens, such as at least the modes.
• Refactor Screen API to only deal with one grid and Terminal API has 2 Screen instances then. #472 Terminal API class acts like a wall, i.e. it does NOT give direct access to Screen nor Grid nor Cell but only via proxy functions and RenderBuffer and RenderCell, with the goal to have everything behind Cell templated.
• Optimize LRU cache #470 LRU cache: do not use std::list but try a runtime fixed size list impl that has zero malloc's during use and it's elements stored in an array. This could even use an std::vector with a wrapper element type to point to prev and next
• optimize bulk ASCII writes #473
• Parser: SSE2 for plaintext ASCII throughput (no controls) and a fast-path for ASCII+LF
• Grid: rewrite grid system to be a dense array of Cell objects. (WIP)
• Grid: provide API for batched single-line text writes
• Grid: 0-based numbering for line and column adressing.
• Grid Cell: Use LRU cache for hyperlinks instead of shared_ptr<> and a unique uint as id key
• Grid Cell: Use LRU cache for image fragments instead of shared_ptr<> (is this GPU-effeciently possible?)` , use unique uint as key here too
• Grid Line: trivially copable, contain only start index and count into Grid's Cell array, line flags and initial SGR.
• make Grid and Line a template over T = Cell and possibly an allocator = std::allocator

All unchecked parts above have now its own dedicated ticket.

consequences checklist

• hyperlinks are still working and we'll tested (also LRU implications)
• images are still working (also LRU implications)
• notcurses demo and info works flawlessly :-D

[data-man]

Hurray! 🚀

[christianparpart]

Notcurses-demo did find some quirks I yet have to fix. But I had so many things pending. This is why 0.2.3 is now out to flush the pipeline. So master branch can head to 0.3.0 - mind, a 0.2.4 can still happen, just not from master branch. :)

…

On Mon, Dec 13, 2021, 9:05 PM Dmitry Atamanov ***@***.***> wrote: Hurray! 🚀 — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#342 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAN3O7FUCQVSQ6EQU7G3E3UQZGZZANCNFSM5A43K42Q> .