Entry 57

Stream 2's effect task: feedback, raster noise and fades, and frame-buffer textures

2026-09-23renderenginetest

Filescrates/piney-stream/src/effect.rscrates/piney-draw/src/lib.rscrates/piney-gs/src/lib.rscrates/piney-desktop/src/soft.rscrates/piney-game/src/main.rstools/test_stream_rs.pydocs/engine/stream.md

After 54 the first cutscene's shapes were right, but its look was not. Each stream may have its own effect task (StreamDemoFuncTbl). The one for str0001, Func_str0001 (0x001885d0), draws straight into and from the frame buffer. A helper agent ported it and checked it against the game's code; I put it in the runtime. The reference is "Stream 2's effect task" in the stream page.

What the task does

Timing.

• It runs at priority 96, after the scene's task (24), every frame. Its first pass is one frame before the scene's first drawn frame.
• It takes the stream's 0x8010 cues last-queued first.
• After a natural end it runs in the first gap frame only.

At start:

• The floor. OBJ_se1_6flo1 stops writing Z (NEVER with FB_ONLY).
• Fog. The scene's fog is set to black between depths 7,500 and 20,000, with the sky, two cloud layers and the moon exempt.
• Raster noise. 8 bands are set up, which takes 216 rand calls.
• Feedback. It is switched on for layer 8 (LYR_jm).
• A fade. 30 frames in from black, on the font layer.

The cues:

• 5, 6 and 15 set the feedback's alpha to 0x58, 0x50 and 0x60;
• 11 starts the noise and 12 stops it;
• 999 fades to white;
• 800 and 16 do nothing.

The feedback. The previous frame's picture, drawn 1.015 times larger and filtered, blended over layers 0-7. Its Z is 0x7fffffff, because the EE's division by zero saturates, so it passes every model. This makes the cutscene's dreamy trails.

The raster noise (layer 100). Each band copies 512 x 32 rows of the current frame to spare VRAM, then draws one sprite per row, shifted by rand. rand is newlib's generator; seed 1, since only the staff roll calls srand.

Frame-buffer textures

Two new TexRef variants in piney-draw:

• FrameBuffer { x, y, width, height }: that rectangle of the frame as drawn so far.
• PreviousFrame: the last finished picture.

On the GPU, a FrameBuffer draw ends the render pass, copies the rectangle and goes on in a new pass that keeps colour and Z. PreviousFrame is a copy taken before the clear. The CPU GS model has both.

Because the feedback samples the previous picture, the runtime now renders every game frame. Before, the window loop drew only the newest frame when it caught up, and --shot only the last one.

Checked

• Against the game's code. test_stream_rs.py effects runs Func_str0001 natively in eemu over all 2,043 steps of stream 2: the noise, feedback, fade and frame-copy code, the layer lists and rand are all the game's own. tests/str0001.rs compares every step bit for bit:

  • the step each cue is raised on;
  • all 4,164 primitives in GS order, with every register the draw sets and every vertex;
  • the rand state after each pass.

  I regenerated the fixture from the game's code in a clean worktree, and it is byte-identical to the one committed.

• The new textures. frame_buffer_textures (piney-gs, on the GPU) and a CPU GS model test check that the copy sees earlier draws and not later ones, and that rows past the frame read 0.

• The rest. The workspace's tests pass, and clippy and fmt are clean.

• A shot. Frame 822 shows the trails around the figure and the floor.