Screenshot

Port of krom (Peter Lemon)'s HiColor64PerTileRow demo (PeterLemon/SNES, PPU/HDMA/HiColor64PerTileRow). An H-timer IRQ fires on every scanline and general-DMAs 16 bytes (8 colors) into CGRAM while the PPU draws; over each 8-line tile row that streams a fresh 64-color set. Even tile rows render from CGRAM 0-63 while the stream refills 64-127 for the odd rows, and vice versa — 28 rows × 64 = 1792 palette slots from a mode that nominally allows 128. HDMA cannot do this: its widest mode moves 4 bytes per scanline; the technique needs 16.

The sunset art is original (procedural, res/sunset.png — 15,885 source colors); devtools/hicolor64.py reimplements krom's converter contract (64×8-pixel segments quantized to 15 colors + black). 357 distinct colors land on screen — any static 4bpp screen caps at 128.

SNES Concepts

H-timer IRQ (irqSet / irqSetHTimer / irqEnable) — per-scanline interrupts
Raw ASM IRQ handlers (irq_stream.asm) — see the contract in <snes/interrupt.h>
General DMA to CGDATA during active display; DMA source auto-advance across transfers
CGRAM double-banking via tilemap palette bits (rows alternate palettes 0-3 / 4-7)
OPHCT/OPVCT 2-read latch discipline ($213F read-pointer reset)

Register fidelity vs the original

Register	krom	this port	note
BGMODE	%00001011	same (setMode(BG_MODE3, 0x08))	Mode 3, BG2 8×8
BG2SC	$3C (map $3C00w, 32×32)	same (bgSetMapPtr)	map loaded at row 4 ($3C80w)
BG12NBA	$00	same (bgSetGfxPtr)	tiles at $0000
BG2VOFS	31	same (bgSetScroll, NMI-synced)	aligns rows with reset cadence
TM	%00000010	same (setMainScreen(LAYER_BG2))	BG2 only
DMAP0/BBAD0	$00/$22	same	1 byte → CGDATA, inc source
HTIME	190	same (irqSetHTimer(190))	writes land in H-blank
NMITIMEN	%10010000	%10010001	we keep auto-joypad (SDK default)
IRQ handler	HTIMERIRQ verbatim	+ save/restore, $213F reset	ours interrupts arbitrary C
VBlank rewind	VBLANKIRQ verbatim	C callback via nmiSetBank	CGADD=0, 128 B, source reset

Measured parity (luna v1.9.0 + Mesen2 cross-check)

Stream cadence is cycle-exact in both emulators: DMA source at scanline 100 = pal + 128 + 100×16 = $86C0 in luna AND Mesen2 (in-ROM probe reading $4302 back).
All 28 bands render from their own palette row (28/28 best-match).
Residual vs the static per-row expectation concentrates on the first scanline of each band — the palette-race window inherent to the technique. Same structure as the reference: krom's ROM in luna diverges from its own shipped PNG on exactly those lines (first-line residual 149.0 on his test pattern vs 16.7 on our smooth gradients; other lines 3.0 vs 6.5).
Render is frame-stable (0 differing pixels across consecutive frames).

Two SDK bugs this port surfaced

consoleInit() seeded the RNG with single reads of OPHCT/OPVCT, leaving the latch read pointers mid-sequence for the whole session — every later latched V read returned hi/lo-swapped garbage. Fixed in lib/source/console.c (STAT78 read after seeding resets the pointers).
nmiSetBank() restored NMITIMEN from a literal, which would have clobbered IRQ enable bits — replaced by the nmitimen_shadow scheme (all $4200 writes compose through the shadow).

How to Build

cd examples/color/hicolor_1792 && make

To regenerate the assets from different art (requires Pillow):

python3 ../../../../devtools/hicolor64.py res/sunset.png res/sunset

Modules Used

console, dma, background