notcurses/doc/HACKING.md

Hacking!

notcurses vs notcurses-core

I wanted to achieve three things:

• Administrators decide whether they want multimedia support installed.
• Clients decide whether they want to use multimedia, and write one program.
• No dlopen(3) or weak symbols -- they're unportable, and break static linking.

If the administrator doesn't want multimedia support installed, they can refrain from installing the notcurses library built with it. Building with USE_MULTIMEDIA=none results in a shim notcurses. This notcurses allows programs that want multimedia to still link; attempting to actually use notcurses_from_file() will result in an error, and the client application can test ahead of time with e.g. notcurses_canopen_images().

Packaging

The ideal packaging IMHO involves two builds, one with USE_MULTIMEDIA set to either ffmpeg or oiio (ffmpeg is preferred to oiio), and one with USE_MULTIMEDIA=none. These ought result in equivalent notcurses-core objects, but two different notcurses objects. Package notcurses-core into its own package, which recommends or even depends on either of the notcurses packages. Name the notcurses packages, say, libnotcurses-ffmpeg and libnotcurses-nomedia, have them conflict with one another, and have both depend on notcurses-core. Defining a virtual package libnotcurses, provided by either of the libnotcurses-* packages, is desirable if supported.

Rows

There are four kinds of ys: physical, rational, logical, and virtual. Physical and rational ys are independent of any particular plane. A physical y refers to a particular row of the terminal. A rational y refers to a particular row of the rendering area. They are related by:

• physical y - margin top == rational y
• rational y + margin top == physical y

In the absence of a top margin, physical y == rational y.

Logical and virtual ys are relative to a plane (possibly the standard plane). A logical y refers to a row of a plane, independent of scrolling. A virtual y refers to a row-sized chunk of the plane's framebuffer, which might be mapped to any row within the plane. They are related by:

• (logical y + plane logrow) % plane leny == virtual y
• (virtual y + plane leny - plane logrow) % plane leny == logical y

All API points expressing a y, whether writing it (e.g. ncplane_cursor_yx()) or reading it (e.g. ncplane_cursor_move_yx()), are working with a logical y. The y member of an ncplane is also a logical y.

Whenever we initiate a write past the end of the line, and the virtual y is equal to ncplane->lenx - 1, we must scroll. Scrolling:

• plane logrow = (plane logrow + 1) % plane leny

As a result, logical y is unchanged, but virtual y has advanced.

Virtual y is useful for only two things:

• Determining whether to scroll, and
• Indexing into the plane's framebuffer

Thus we usually keep y logical.

Right-to-left text

We want to fully support Unicode and international text. But what does it mean to use right-to-left text with a fullscreen, random-access application? In particular, what happens in the case where we've written the right-to-left string SHRDLU (which ought appear as ULDRHS) to a plane, starting at (0, 0), and then we place say a U+1F982 SCORPION (🦂) at (0, 2)? Ought this yield UL🦂HS, or ought it instead yield HS🦂UL? If the original string had been SH🦂LU, it would have been displayed by most terminals as HS🦂UL, due to treating it as a right-to-left segment, a left-to-right segment, and finally a right-to-left segment. Alternatively, it might have been displayed as UL🦂HS, especially if aligned on the right. It's difficult to know. So, we instead force text direction by appending U+200E LEFT-TO-RIGHT MARK to any EGCs we believe to provoke right-to-left. The user is thus solely responsible for managing right-to-left presentation.

I hate everything about this terrible, fragile, wasteful "solution".

Rendering/rasterizing/writeout, and resizing

The scope of rendering is a pile. The scope of rasterization is a pile, the last frame, and the screen. These latter two are shared, and thus concurrent rasterizations are illegal and an error. Concurrent rendering of different piles is explicitly supported.

In Notcurses prior to 2.1.0, there was only one pile. Rendering and rasterizing were a single function, notcurses_render(). Since this proceeded end-to-end, and didn't need worry about concurrency, it could perform an optimal strategy:

• Check for a resize, resizing the last frame and standard plane if appropriate
• Render the (single) pile, taking full advantage of an enlarged terminal
• Rasterize the (single) render, carrying through plenty of state from render
• Write out the (single) rasterization

It is an ineluctable fact that we cannot guarantee proper writeout, since the terminal can be resized in the middle of a writeout, and the signal is both unreliable and asynchronous. Receipt of the SIGWINCH signal is async with regards to the actual geometry change; processing of the signal is async with regards to its delivery. Even if this was all synchronous, signals are fundamentally unreliable, and can be missed. Internalize and accept this.

If we write more data than the terminal has geometry (either with regards to rows or columns), we will produce some garbage. If we write less, we'll simply fail to fill up the screen (so long as we explicitly move to new rows, which we do). Both are undesirable, but neither is catastrophic.

Writeout is a blocking process. We do not support non-blocking writeout at this time. An error at any point while writing out the frame will abort the writeout and be considered a failure. Writeout takes a buffer, a buffer length, and an output descriptor; it attempts to write until the buffer has been written in its entirety. The buffer might only partially update the screen, due to damage detection (undamaged cells are never placed into the buffer); the buffer is thus relative to our concept of the current state of the terminal (the "last frame"). The "last frame" is updated in rasterization, as the buffer is generated. It is thus critical that rasterized frames be written out in order. Writeout is thus bound to rasterization, except special cases that always rasterize total frames:

• notcurses_refresh() (writes last frame to terminal following clear screen)
• notcurses_rasterize_to_buffer() (copies last frame to buffer)
• notcurses_rasterize_to_file() (appends last frame to file)

Rasterization always results in at least one writeout. Henceforth, we will consider only rendering and rasterizing, the latter with an implicit writeout.

The output of rendering is fed into rasterization. Especially given multiple piles, it is possible that another render will take place between rendering and rasterizing of a given pile (this can happen with even a single pile, though, now that rendering and rasterizing are decoupled). It is thus necessary that rendering never refer to the "last written frame", since that last written frame might change by the time the render is written out. Similarly, the rasterizer may not assume that the size of the render it is given is equal to the current conception of the screen size.

The last frame and standard frame are resized in notcurses_resize() to match the recovered terminal geometry. notcurses_resize() acquires the geometry via an ioctl(), and resizes these framebuffers, zero-initializing any new area. Since it's possible that the terminal was resized without our receipt of a signal, we want to call it in somewhere in the render/resize cycle.

It is undesirable to call notcurses_resize() in the multiple render path, since this would need internal locking to deal with concurrent renders. It is desirable to call notcurses_resize() prior to rendering, since otherwise we might not render portions of the pile only just made visible (in the case of the terminal being enlarged). It is not desirable to call notcurses_resize() prior to rendering, since if the terminal shrinks following the render but before the raster, we'd like to know that and thus avoid overwriting. Remember from above that an underwrite is less damaging than an overwrite. We thus perform notcurses_resize() in the rasterization path. The upshot is that a rendered frame can be larger or smaller than the screen at the time we rasterize--but since this could happen anyway, it's no great loss.

EXCEPT for one case: imagine that we have a single plane, 1000x1000, that is all green. Our program starts at 80x24, renders, rasterizes, and enters an input loop. It performs another render+raster for each input (remember, a SIGWINCH manifests as NCKEY_RESIZE). The terminal is then resized to 100x100. The following happens:

• initial render renders an 80x24 frame
• initial raster writes out this 80x24 frame, screen is green
• block on input
• terminal is resized to 100x100
• NCKEY_RESIZE is read
• second render renders an 80x24 frame
• second raster learns of 100x100 size, writes out 80x24 in upper left
• block on input, screen is partially green and partially background

At the end of our second writeout, we have an incomplete screen, despite the geometry change happening well before (and indeed triggering) our second cycle. We do not simply move rendering into the top of rasterization, since resizes are presumably rare, and we want to facilitate maximum parallelism, which we can't do if rendering is part of a serial section).

Actually, this suggests (and I then confirmed) that this means the top half itself is using the screen geometry, and thus already accessing shared data. So a mutex is happening there no matter what.

By the time we rasterize, we thus have three different geometries in play:

• the most recently-acquired actual screen geometry (as reported by ioctl()),
• the geometry of the supplied render (as determined at render time), and
• the geometry of the last-rendered frame.

Rasterization, remember, is a function of the supplied render, the last frame, and the output geometry--all three of these distinct geometries. So long as there is no resize step between rasterizing and writing, writing deals with the same geometries as rasterization, so we ignore it.

Rasterization can be split into two virtual phases: postpaint and rastering. postpaint corrects for CELL_ALPHA_HIGHCONTRAST, performs damage detection, and copies any necessary EGCs from their source pools to the common pool (copying these EGCs is why a pile cannot be modified between rendering and rasterizing--such modifications might invalidate the EGC references). The rasterizing phase takes this final rendered plane, pool, damage map, and the current rendering state (e.g. cursor position, last style+color), and generates a buffer. At this point, the last frame is updated, and a new rasterization could technically begin. It is probably possible to unite the two phases, though this has not been done, and might never be.

So, rasterization must:

• use the rendered frame's geometry to create a damage map
• iterate over each cell of the rendered frame (postpaint)
  • if the cell was present in the last frame, check for damage
  • if the cell was not present in the last frame, assume damage
• iterate over each cell of the visual area (rasterization)
  • if the cell was present in the damage map, check for damage
    • if there was damage, emit the data (plus a move if applicable)
    • if there was not damage, skip the cell
  • if the cell was not present in the damage map, skip the cell

We skip the cell if it was not present in the damage map because an enlarged terminal is filled with default cells, which is all we could generate in any case, having not rendered the cell. This implies that the damage map must be two-dimensional, as must the render. Only the rasterized buffer is flattened to a single dimension.

Given our requirement that a pile not be mutated between render and raster, we know that at render time the pile is suitable for rendering. We could thus check to see if the screen has grown relative to the render, and call for a fresh render. This would be a great solution for our 1000x1000 case above, but it doesn't help when the user has only been generating enough output for the visible area. In this case, new data will not be available should raster call for a new render; it is instead necessary that the "userspace" resize actions be taken.

This raises a new issue: given cascading resize callbacks, notcurses_resize() can result in arbitrary changes to the pile. This suggests that the resize operation cannot occur between render and raster...

Alternatives to the Painter's Algorithm

The rendering area is RY * RX, where RY and RX are positive integers.

A plane is either active or inactive for a given cell in the rendering area. The plane is active if it is defined at that cell. It is inactive otherwise.

There is an initial (possibly empty) inactive region before the plane is first reached. There then follow A' (A' >= 0) active regions, separated by (I' = A'-1) inactive regions (I' is 0 if A' is 0). These active regions A_0, A_1, ... all have the same size, and these inactive regions I_0, I_1, ... likewise all have the same size. I_0 + A_0 == RX. There is then a final (possibly empty) inactive region following the plane's lowermost, rightmost intersection with the visual area.

I_init + A' * A_0 + I' * I_0 + I_final == RX * RY.

Given RX and RY, we can describe a plane's activity pattern completely with three numbers: I_init, A', and A_0.

Keep two ordered structures, an active set and an inactive set. The active set is counting down until they become inactive. The inactive set is counting down until the become active.

Initialization:

For each plane, calculate I_init and A_0. Planes with I_init values of 0 go into the active set, sorted first by A_0 and secondarily by plane depth. Planes with I_init values >= 0 go into the inactive set, sorted first by I_init and secondarily by plane depth.

For rendering area RY * RX and plane py * px at offset y, x, I_init is:

 infinite for x >= RX
 infinite for x + px <= 0
 infinite for y >= RY
 infinite for y + py <= 0
 0 for y <= 0, y + py >= 0, x <= 0
 x for y <= 0, y + py >= 0, x > 0
 y * RX + x for y > 0, x >= 0
 y * RX for y > 0

max finite initial gap is RY * RX - 1. min initial gap is 0.

Each node is a pointer to a plane, and the scalar coordinate xy (0 <= xy < PX * PY) at which the current state changes (A_0 and I_init).

assuming finite initial gap (i.e. that the plane overlaps the rendering area), the active length (can exceed practical length) is:

 x <= 0:
  x + px >= RX: (spans horizontal range)
   y <= 0:
     RX * py + y, from origin
   y > 0:
     RX * py, from column 0
  x + px < RX:
    x + px,
 x > 0:
  x + px >= RX:
    RX - x
  x + px < RX:
    px

max active length is RY * RX (for a plane covering the entirety of the horizontal viewing area), otherwise RX - 1. min active length is 1.

inactive gap is undefined if plane spans visual region or is invisible. otherwise, inactive gap is calculated at right edge of plane (column C), and is equal to PX - (C + 1) + x if x >= 0, or PX - (C + 1) otherwise.

at each step we check to see if the foremost planes of either set need flip to the other set. this suggests an extra sort per flip. unless we've eclipsed a plane's I_init, or entered a plane's I_final, an element moving from one set to another must have the same previous element as it did before. each node thus keeps an additional element, a double pointer to the previous element's next link. upon flip, check this pointer to ensure it's NULL. if it is NULL, link ourselves. otherwise, chase to the end, and link ourselves.

ANALYSIS

There's a sort at the beginning of O(PlgP) on P planes. We then check P * PX * PY cells. In the worst case, where all cells actually need be used, our new algorithm is worse by the cost of a sort.