prometeu-runtime/docs/specs/runtime/04-gfx-peripheral.md
2026-07-03 23:06:35 +01:00

801 lines
24 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# GFX Peripheral (Graphics System)
Domain: virtual hardware: graphics
Function: normative
Didactic companion: [`../learn/mental-model-gfx.md`](../runtime/learn/mental-model-gfx.md)
## 1. Overview
The **GFX** peripheral is responsible for generating images in PROMETEU.
`DEC-0030` defines the current logical render boundary. PROMETEU render
production is split from render implementation:
```text
domain buffers during logical frame
-> RenderManager closes buffers
-> RenderSubmission snapshot
-> render surface / implementation consumes submission
-> RGBA8888 surface publication
```
The GFX peripheral remains the classic local raster implementation for
PROMETEU's 2D output. Logical render APIs do not target `Gfx` or the
framebuffer directly.
`RenderSubmission` is the closed snapshot passed across the render boundary. It
MUST contain a frame id, the active app mode, and one typed packet:
`Game2DFramePacket` for `AppMode::Game` or `ShellUiFramePacket` for
`AppMode::Shell`. Once closed, producers MUST treat the submission as
immutable. The runtime backpressure policy is latest-complete-submission-wins;
it MUST NOT accumulate an unbounded frame queue.
The platform layer is the runtime-facing implementation boundary for render
handoff. Runtime and firmware code publish completed frames through typed
platform services such as `RenderSubmissionSink`, and they access Game 2D
composition through `Game2DFrameComposer`. They MUST NOT depend on a monolithic
hardware bridge, mutable `Hardware`, mutable `Gfx`, or live `FrameComposer`
reference as the render handoff contract.
The current 2D graphics model is based on:
- framebuffer
- tilemaps
- tile banks
- priority-based sprites
- composition by drawing order
---
## 2. Resolution and Framebuffer
### Base resolution
- **480 × 270 pixels**
- aspect ratio close to 16:9
- scalable by the host (nearest-neighbor)
### Pixel format
- **RGBA8888**
- 8 bits Red
- 8 bits Green
- 8 bits Blue
- 8 bits Alpha
- canonical raw channel order: **RGBA**
The framebuffer alpha channel may carry meaningful runtime output. The runtime
MUST NOT force the published front buffer alpha channel to `255` as a blanket
contract.
Color values in the runtime, HAL, host-facing framebuffer, and GFX ABI are
logical RGBA8888 values. RGB565 is not a supported runtime framebuffer,
palette, host presentation, or compatibility contract.
### Published owned frame
The canonical worker-published frame value is `OwnedRgba8888Frame`.
`OwnedRgba8888Frame` contains:
- the logical `FrameId`;
- render ownership metadata;
- width and height in pixels;
- stride in pixels;
- an owned `Vec<u32>` containing packed RGBA8888 pixels in canonical RGBA
channel order.
The pixel vector is owned by the published frame. Consumers may copy it into a
host upload buffer, retain it as the latest published frame, or repeat it for a
host redraw. Consumers MUST NOT interpret the frame as a native texture,
swapchain image, window handle, or host GPU resource.
---
## 3. Double Buffering
The render surface implementation maintains the concrete buffers needed to
publish a frame. The classic software implementation uses:
- **Back Buffer** — where the frame is built
- **Front Buffer** — where the frame is displayed
Per-frame flow:
1. Game or Shell code mutates mode-specific domain buffers during the logical frame.
2. `RenderManager` closes the active buffers into an immutable `RenderSubmission`.
3. The render surface implementation consumes the submission and rasterizes to its back buffer.
4. The render surface publishes the completed RGBA8888 surface.
5. The host displays the published surface.
This guarantees:
- no tearing
- clear per-frame synchronization
- deterministic behavior
---
## 4. Asynchronous Render Boundary
`DEC-0031` defines the asynchronous render boundary. The contract applies even
when the current implementation consumes submissions synchronously.
### 4.1 Complete packet contract
`RenderSubmissionPacket::Game2D` MUST be a complete frame description for the
Game 2D render consumer. The consumer MUST NOT bypass the packet by consulting
live VM state, live `FrameComposer` state, or mutable runtime state.
For Game 2D, the required composition order is:
```text
composer scene/layers/sprites/HUD
-> buffered gfx2d primitives
-> publication/present
```
`gfx2d.*` primitives are therefore a final Game 2D overlay. This is true for
both active-scene frames and no-scene frames.
### 4.2 Resource boundary
Submissions MUST remain small and owned. Heavy resident resources such as glyph
banks, scene banks, asset payloads, and viewport cache materializations MUST
NOT be copied into each submission.
Submissions SHALL carry stable resource IDs or handles. The render consumer may
resolve those IDs through read-only resource APIs. Resource installation,
resolver updates, viewport cache refreshes, and bank residency changes belong
to the logical/runtime side before handoff, or to an owning service. They MUST
NOT require the render consumer to hold mutable VM, `Hardware`, `Gfx`, or
`FrameComposer` references.
Read-only resource APIs expose compact lookup by ID. They do not expose mutable
bank state, `Arc` ownership as part of the contract, copies of whole banks, or
snapshot payloads. The implementation may choose its internal sharing mechanism,
but the render boundary contract is ID-based read-only access.
Local host implementations may keep a concrete hardware object internally as a
platform implementation detail. That object is not part of the runtime-facing
render boundary; the boundary is the typed platform service set.
The runtime does not guarantee visual integrity if a developer or framework
replaces resources behind an in-flight submission in a way that violates asset
discipline.
### 4.3 Handoff
The first asynchronous Game render model uses a single pending slot with
latest-wins semantics:
- publishing a new submission replaces the previous pending submission;
- replacement before consumption is counted as a render drop;
- the render consumer takes ownership of the pending submission;
- the VM/producer MUST NOT block on render consumption, raster completion, or
present completion.
The consumer status is telemetry. It is not a semantic ACK to the VM.
The real render worker consumes owned `RenderSubmission` values from this
single-slot handoff and publishes owned `OwnedRgba8888Frame` values. The
producer path MUST remain non-blocking with respect to worker rasterization and
host present. A slow worker can cause replacement/drop telemetry, but it MUST
NOT create an unbounded queue or stall VM logical frame production.
### 4.4 Frame pacing
Game logical frames are paced by the runtime frame scheduler, not by render
worker ACK. The intended cadence is one Game logical frame per frame tick and
at most one pending submission ahead of the consumer.
`FRAME_SYNC` remains the canonical end of a VM logical frame. Cycle/time budget
is evidence for certification, watchdog, diagnostics, and overrun reporting; it
MUST NOT be used as the normal mechanism for cutting a logical frame short.
If a Game logical frame overruns the display cadence, logical frames remain
sequential. The render consumer may repeat the last valid frame, and telemetry
records the overrun/repeat. The VM MUST NOT produce catch-up frames to skip
from frame `N` to frame `N+k`.
Repeating a frame means reusing the latest published `OwnedRgba8888Frame`
without recomposing, rerunning VM code, or consuming a new submission. Repeat
behavior is presentation cadence behavior, not guest-visible execution.
### 4.5 AppMode policy
Render execution policy is explicit by pipeline/AppMode:
- `AppMode::Game` is frame-paced and may use a render worker when the
host/runtime supports it, with local synchronous fallback.
- `AppMode::Shell` is lifecycle-driven and local/synchronous by default.
- Shell VM-backed apps follow Shell lifecycle; they do not declare an
independent frame-paced game workload under this contract.
- Splash, crash, and hub/system screens follow Shell/local policy unless a
later decision defines a more specific policy.
### 4.6 Ownership and epoch
Every submission that may cross an asynchronous boundary MUST carry render
ownership metadata: at minimum frame identity, app mode, app identity where
available, and render epoch/generation.
Foreground visual-owner transitions MUST advance the active render epoch or
equivalent generation through a central runtime render manager. The render
consumer MUST check ownership before present and MUST discard stale submissions
whose ownership no longer matches the active owner.
Transitions that invalidate stale render work include:
- Game to Shell/Hub;
- Shell/Hub to Game;
- crash screen takeover;
- splash or system screen takeover;
- cartridge or app swap, even if `AppMode` remains unchanged;
- shutdown/stop.
The same physical surface may be reused by multiple visual owners, but logical
ownership MUST remain explicit.
Game pause/resume, foreground stack behavior, and coexistence of a paused Game
with VM-backed Shell apps are defined by the foreground/lifecycle contract. The
render boundary's responsibility is narrower: foreground owner changes MUST
advance ownership, stale submissions MUST be discarded, and visual return to a
resumed Game MUST wait for a valid submission from the current ownership epoch.
### 4.7 Render telemetry
Asynchronous render is best-effort observable, not a VM-visible handshake.
Render drops, stale epoch discards, repeated presents, render errors, and
present errors MUST be recorded for host diagnostics, debugging, profiling, and
certification evidence. VM program semantics MUST NOT depend on whether a
submission was consumed or presented.
Minimum render telemetry includes:
- produced submissions;
- replaced-before-consume submissions;
- consumed submissions;
- presented frames;
- repeated presents;
- render errors;
- present errors;
- stale epoch discards;
- shutdown discards;
- last produced, consumed, presented, dropped, and error frame IDs;
- active render epoch.
### 4.8 Shutdown and typed failures
Render worker shutdown is bounded and observable. A shutdown request MUST
wake a waiting worker, discard pending submissions that will not be consumed,
and either join within the configured timeout or report a typed shutdown
failure.
Worker backend failures, sink publication failures, stale ownership discards,
panic capture, and shutdown timeout are typed render worker outcomes. They are
recorded through telemetry and diagnostics; they do not become VM-visible ACKs.
---
## 5. PROMETEU Graphical Structure
The graphical world is composed of:
- Up to **16 Tile Banks**
- **4 Tile Layers** (scrollable)
- **1 HUD Layer** (fixed, always on top)
- Sprites with priority between layers
### 5.1 Tile Banks
- There are up to **16 banks**
- Each bank has a fixed tile size:
- 8×8, 16×16, or 32×32
- A bank is a graphics library:
- environment
- characters
- UI
- effects
- `assets.pa` tile-bank payloads use a serialized representation distinct from runtime memory:
- serialized pixels are `4bpp` packed in payload order
- runtime memory may expand pixels to one `u8` palette index per pixel after decode
### 5.2 Layers
- There are:
- 4 Tile Layers
- 1 HUD Layer
- Each layer points to **a single bank**
- Sprites can use **any bank**
- HUD:
- does not scroll
- maximum priority
- generally uses 8×8 tiles
---
## 6. Internal Model of a Tile Layer
A Tile Layer **is not a bitmap of pixels**.
It is composed of:
- A **logical Tilemap** (tile indices)
- A **Border Cache** (window of visible tiles)
- A **Scroll Offset**
### Structure:
- `bank_id`
- `tile_size`
- `tilemap` (large matrix)
- `scroll_x`, `scroll_y`
- `cache_origin_x`, `cache_origin_y`
- `cache_tiles[w][h]`
---
## 7. Logical Tilemap
The tilemap represents the world:
Each cell contains:
- `tile_id`
- `flip_x`
- `flip_y`
- `priority` (optional)
- `palette_id` (optional)
The tilemap can be much larger than the screen.
---
## 8. Border Cache (Tile Cache)
The cache is a window of tiles around the camera.
Example:
- Screen: 480×270
- 16×16 tiles → approximately 30×17 visible
- Current runtime cache sizing with 16×16 layers: 35×21 tiles
It stores tiles already resolved from the tilemap.
---
## 9. Cache Update
Every frame:
1. Calculate:
- `tile_x = scroll_x / tile_size`
- `tile_y = scroll_y / tile_size`
- `offset_x = scroll_x % tile_size`
- `offset_y = scroll_y % tile_size`
2. If `tile_x` changed:
- Advance `cache_origin_x`
- Reload only the new column
3. If `tile_y` changed:
- Advance `cache_origin_y`
- Reload only the new line
Only **one row and/or column** is updated per frame.
---
## 10. Cache as Ring Buffer
The cache is circular:
- Does not physically move data
- Only moves logical indices
Access:
- `real_x = (cache_origin_x + logical_x) % cache_width`
- `real_y = (cache_origin_y + logical_y) % cache_height`
---
## 11. Canonical Game Projection
Game mode uses a typed Game 2D submission. `composer.*` owns high-level Game 2D
frame composition: scene binding, camera, sprites, HUD, and frame orchestration.
`gfx2d.*` owns Game 2D primitives only. Both domains are mutable while the
logical frame is being produced and are closed by `RenderManager` into a
`Game2DFramePacket`.
For each Game 2D packet:
1. For each Tile Layer, in order:
- Rasterize visible tiles from the cache
- Apply scroll, flip, and transparency
- Write to the render surface's working buffer
2. Draw sprites:
- With priority between layers
- Drawing order defines depth
3. Draw HUD layer last
4. Draw buffered `gfx2d.*` primitives according to the Game 2D packet contract.
This section describes only the Game 2D packet rendering path. Shell/system UI
uses `gfxui.*` and `ShellUiFramePacket`; it is never part of Game HUD or
`composer.*`.
---
## 12. Drawing Order and Priority
- There is no Z-buffer
- There is no automatic sorting
- Whoever draws later is in front
Base order:
1. Tile Layer 0
2. Tile Layer 1
3. Tile Layer 2
4. Tile Layer 3
5. Sprites (by priority between layers)
6. HUD Layer
7. Buffered `gfx2d.*` Game primitives
Normative boundary:
- Items 1 through 6 belong to `composer.*` Game-frame composition.
- Item 7 belongs to `gfx2d.*`.
- `gfx2d.*` primitives MUST NOT be interpreted as scene, sprite, camera, HUD,
or frame orchestration.
- Shell/system UI belongs to `gfxui.*` and `ShellUiFramePacket`, not Game HUD.
---
## 13. Transparency
Transparency is represented by the alpha channel of the resolved RGBA8888
color.
Palette indices are ordinary indices. The runtime MUST NOT reserve palette
index `0` as a special transparent index. A palette entry may still be authored
as transparent by setting its alpha channel to `0`, but that is ordinary palette
data rather than a special index rule.
```
if resolved_color.alpha == 0:
skip
else:
draw_or_blend(resolved_color)
```
---
## 14. Color Math (Discrete Blending)
Inspired by the SNES.
Official modes:
- `BLEND_NONE`
- `BLEND_HALF`
- `BLEND_HALF_PLUS`
- `BLEND_HALF_MINUS`
- `BLEND_FULL`
Alpha is available through RGBA8888 colors. Discrete blend modes remain part of
the classic GFX contract, and later optimization work may specialize opaque,
masked, and alpha paths.
Everything is:
- integer
- cheap
- deterministic
---
## 15. Where Blend is Applied
- Blending occurs during drawing
- For canonical game composition, the result goes to the back buffer during composition
- For `gfx2d.*` and `gfxui.*` primitives, the result is applied when the render
surface consumes the closed packet for the active app mode
- There is no automatic GPU-style post-processing pipeline
---
## 16. What the GFX DOES NOT support
By design:
- Shaders
- Modern GPU pipeline
- HDR
- Gamma correction
- RGB565 compatibility framebuffers
- fade fields, fade syscalls, or fade packet members
- multi-format backend selection
- render-thread ownership as part of this contract
---
## 17. Performance Rule
- Layers:
- only update the border when crossing a tile
- never redraw the entire world
- Rasterization:
- always per frame, only the visible area
- Sprites:
- always redrawn per frame
---
## 18. Transitions and Removed Fade Contract
The previous special fade model is not part of the canonical render contract.
Current render packets, syscalls, and ABI domains MUST NOT expose scene fade,
HUD fade, fade levels, fade colors, or equivalent inherited fade state.
Future visual transitions between Shell and Game belong to `RenderManager`.
They are not owned by `composer.*`, `gfx2d.*`, `gfxui.*`, `Game2DFramePacket`,
or `ShellUiFramePacket`.
Rules:
- render domain packets MUST NOT contain fade fields;
- public render syscalls MUST NOT expose fade controls;
- render implementations MUST NOT treat fade as an inherited post-processing
habit;
- transition work requires an explicit `RenderManager` contract or a later
decision.
---
## 19. Palette System
### 19.1. Overview
PROMETEU uses **exclusively** palette-indexed graphics.
There is no direct RGB-per-pixel mode.
Every graphical pixel is an **index** pointing to a real color in a palette.
---
### 19.2. Pixel Format
Each pixel of a tile or sprite is:
- **4 bits per pixel (4bpp)**
- values: `0..15`
Fixed rule:
- Indices `0..15` are ordinary valid palette indices
- Transparency comes from the alpha channel of the resolved palette entry
---
### 19.3. Palette Structure
Each **Tile Bank** contains:
- **64 palettes** in runtime-facing v1
- Each palette has:
- **16 colors**
- each color in **RGBA8888** with canonical RGBA channel order
Size:
- 1 palette = 16 × 4 bytes = **64 bytes**
- 64 palettes = **4 KB per bank**
- 16 banks = **64 KB maximum palettes**
---
### 19.4. Palette Association
#### Fundamental Rule
- Each **tile** uses **a single palette**
- Each **sprite** uses **a single palette**
- The palette must be provided **explicitly** in every draw
There is no palette swap within the same tile or sprite.
---
### 19.5. Where the Palette is Defined
#### Tilemap
Each tilemap cell contains:
- `tile_id`
- `palette_id (u8)`
- `flip_x`
- `flip_y`
Runtime-facing validity rule for v1:
- `palette_id` values are valid only in the range `0..63`
#### Sprite
Each sprite draw contains:
- `bank_id`
- `tile_id`
- `palette_id (u8)`
- `x`, `y`
- `flip_x`, `flip_y`
- `priority`
Runtime-facing validity rule for v1:
- `palette_id` values are valid only in the range `0..63`
---
### 19.6. Color Resolution
The pipeline works like this:
1. Read indexed pixel from tile (value 0..15)
2. Resolve:
- real_color = palette[palette_id][index]
3. Apply:
- flip
- discrete blend
- alpha/skip behavior from the resolved RGBA8888 color
- writing to back buffer
In other words:
```
pixel_index = tile_pixel(x,y)
color = bank.palettes[palette_id][pixel_index]
if color.alpha == 0:
skip
else:
draw_or_blend(color)
```
---
### 19.7. Organization of Tile Banks
Tile Banks are "strong assets":
- Tiles and palettes live together
- Export/import always carries:
- tiles + palettes
- In `assets.pa` v1, the serialized payload is:
- packed indexed pixels for the whole sheet
- followed by the palette table for the same bank
- The hardware does not impose semantic organization:
- grouping is the creator's decision
- Tooling and scripts can create conventions:
- e.g.: palettes 0..15 = enemies
- 16..31 = scenery
- etc.
Runtime-facing v1 baseline:
- sheet pixels are authored and resolved as indexed values `0..15`
- serialized tile-bank payload uses packed `u4` pixel indices
- runtime may materialize the decoded bank as expanded `u8` pixel indices plus palette table
---
### 19.8. Metrics for Certification (CAP)
The system can measure:
- `palettes_loaded_total`
- `palettes_referenced_this_frame`
- `tiles_drawn_by_palette_id`
- `sprites_drawn_by_palette_id`
---
## 20. Syscall Return and Fault Policy
Graphics-related public ABI in v1 is split between:
- `composer.*` for Game 2D high-level frame composition;
- `gfx2d.*` for Game 2D primitives only;
- `gfxui.*` for Shell UI primitives only.
`composer.*` and `gfx2d.*` are available to `AppMode::Game`.
`gfxui.*` is available to `AppMode::Shell`. Renderers and capabilities may
differ by app mode.
Only operations with real operational rejection paths return explicit status values.
Fault boundary:
- `Trap`: structural ABI misuse (type/arity/capability/shape mismatch);
- `status`: operational failure;
- `Panic`: internal runtime invariant break only.
### 20.1 Return-shape matrix in v1
| Syscall | Return | Policy basis |
| ----------------------- | ------------- | --------------------------------------------------- |
| `gfx2d.clear` | `void` | no real operational failure path in v1 |
| `gfx2d.fill_rect` | `void` | no real operational failure path in v1 |
| `gfx2d.draw_line` | `void` | no real operational failure path in v1 |
| `gfx2d.draw_circle` | `void` | no real operational failure path in v1 |
| `gfx2d.draw_disc` | `void` | no real operational failure path in v1 |
| `gfx2d.draw_square` | `void` | no real operational failure path in v1 |
| `gfx2d.draw_text` | `void` | no real operational failure path in v1 |
| `gfxui.*` primitives | `void` | no real operational failure path in v1 |
| `composer.bind_scene` | `status:int` | status-returning API, but missing scene glyph dependencies are fatal runtime errors |
| `composer.unbind_scene` | `status:int` | explicit orchestration-domain operational result |
| `composer.set_camera` | `void` | no real operational failure path in v1 |
| `composer.emit_sprite` | `status:int` | explicit orchestration-domain operational rejection |
### 20.1.a Primitive domain semantics
The primitive domains have stable operational meaning:
- `gfx2d.*` is Game 2D primitive command buffering;
- `gfx2d.*` is screen-space and primitive-only relative to `composer.*`;
- `gfx2d.*` is outside scene, camera, sprites, HUD, and Game 2D frame orchestration;
- `gfxui.*` is Shell UI primitive command buffering;
- `gfxui.*` does not contain widget or layout policy.
Callers MUST NOT rely on stable immediate writes to the working back buffer as
the public contract for primitive drawing. Primitive calls mutate domain command
buffers that close into the active typed submission.
### 20.1.b Scene dependency fatal boundary
`composer.bind_scene` remains a status-returning syscall in the public ABI, but scene glyph dependency absence is outside the accepted passive operational-error model.
Rules:
- the target scene slot may still be rejected through ordinary status-returning behavior when the scene itself is unavailable;
- missing glyph dependencies referenced by a resident scene are not a passive `status:int` case;
- if scene activation discovers that a layer dependency cannot be resolved to a committed glyph asset, the machine MUST fail fatally and emit a clear log;
- if scene composition later discovers that a layer dependency can no longer be resolved, the machine MUST fail fatally and emit a clear log;
- runtime MUST NOT continue canonical scene composition after such a dependency failure.
### 20.2 `composer.emit_sprite`
`composer.emit_sprite` returns `status:int`.
ABI:
1. `glyph_id: int` — glyph index within the bank
2. `palette_id: int` — palette index
3. `x: int` — x coordinate
4. `y: int` — y coordinate
5. `layer: int` — composition layer reference
6. `bank_id: int` — glyph bank index
7. `flip_x: bool` — horizontal flip
8. `flip_y: bool` — vertical flip
9. `priority: int` — within-layer ordering priority
Minimum status table:
- `0` = `OK`
- `1` = `SCENE_UNAVAILABLE`
- `2` = `INVALID_ARG_RANGE`
- `3` = `BANK_INVALID`
- `4` = `LAYER_INVALID`
- `5` = `SPRITE_OVERFLOW`
Operational notes:
- the canonical public sprite contract is frame-emission based;
- no caller-provided sprite index exists in the v1 canonical ABI;
- no `active` flag exists in the v1 canonical ABI;
- overflow remains non-fatal and must not escalate to trap in v1.