Entry 3

The original boots in a harness on rrt-emu, to the legal screen

2026-10-1417mtoolingenginedecompvideotest

rugrats-hle runs the unmodified SLUS_006.50 in retro_rt's R3000A (rrt::emu) for 600 frames in 4.8 seconds and reaches the opening cinematic. On the way it draws the legal screen, decoded from LEGAL512.BS through a model of the MDEC, and loads 32 files.

How it is put together

There is no emulator to compare against on this machine, so the reference is built from parts that already exist: rrt-emu's CPU, GTE and BIOS-table HLE, and hwtr's GPU rasterizer and I/O model, carried onto a new PsxBus (2 MB RAM mirrored to 8 MB, scratchpad, I/O, BIOS area reading zero). rrt-emu has no interrupts and no exceptions; the harness supplies both at the level the game sees them. libetc's ResetCallback (0x8009b2fc) hands the BIOS a setjmp buffer whose resume point calls the dispatcher at 0x8009b3d4, which reads I_STAT and I_MASK and calls one handler per bit. So an interrupt is: set the I_STAT bit, and when I_MASK lets it through, interrupts are enabled (the critical-section flag rrt's BIOS keeps) and the dispatcher is installed, call the dispatcher on top of the running program with every register saved and restored. The vertical blank is raised once per frame; a DMA's completion raises IRQ 3 through DICR 2000 steps after it started, which is what libgpu's drawing queue waits for.

Only four things are replaced at their entry points: VSync (0x8009af48) ends the frame and raises the vertical blank instead of spinning on the root counter; the game's own read_file (0x80019794) copies the file from the disc, since the original reads through libcd's ready callback; printf (0x800b1478) is logged; and libcd's drive calls (CdReset, CdSync, CdReady, CdControl, CdControlF, CdControlB) answer "done" because no drive is modelled yet. BIOS calls the game makes beyond rrt's table are added in the harness: A0:71 _96_init, A0:AB/AC/AD and B0:5C/5D for the memory card.

Three faults on the way

The nested dispatcher call first ran 5 million steps in libgpu's GPU time-out loop (0x8009e8e4): a VSync callback calls DrawSync(0), which waits for the DMA to finish, and the device clock only advanced between run slices. Nested calls now run in slices too and advance the devices, but never nest an interrupt. Before that, no interrupt was delivered at all: the flag rrt's critical_sections returns is "interrupts enabled", not "inside a critical section". And the hwtr jump-table fix of 2 was a precondition for reading any of this.

The MDEC

The logos and the legal screen are .BS stills: libpress decodes the variable-length codes on the CPU and feeds run-length blocks to the MDEC by DMA 0, reading pixels back by DMA 1. rugrats_data::mdec is the decoder proper (dequantise, zigzag, the two-pass IDCT through the uploaded scale table, YUV to RGB at the MDEC's coefficients) as a pure function, tested on hand-made blocks (a_dc_only_block_reads_back_at_its_zigzag_start, a_run_skips_coefficients_and_lands_in_zigzag_order, negative_levels_divide_toward_zero, a_flat_block_through_a_flat_matrix_stays_flat). The port will use the same code for its stills and movies. The harness wraps it in the command protocol (decode, quant tables, scale table) and lets a DMA 1 started before the data wait for it.

The libraries are mapped in docs/engine/libraries.md, the boot in docs/engine/boot.md.