Family 2 (Backgrounds) · the offset-per-tile modes (2/4/6), after mode2 and mode4.

Screenshot

The screenshot is luna's native 512×448 output: the stripes are one half-pixel wide.

Mode 6 is the hi-res member of the offset-per-tile family. It gives you one 4bpp (16-colour) BG1 drawn 512 half-pixels wide, like Mode 5, and BG3 as the offset table, read like Mode 2's: one row of horizontal offset words, one row of vertical ones. Its tiles are always 16 half-pixels wide, and so are the table's columns, so 32 words still cover the screen.

BG1 here is diagonal colour bands drawn in one-half-pixel stripes, which no 256-pixel mode can show. A sine wave written into the vertical row makes the columns ripple; press A and the wave moves to the horizontal row, and the columns slide sideways. B shows a still test card for a question the references leave open (below).

Build & Run

cd $OPENSNES_HOME

make -C examples/backgrounds/mode6

Open mode6.sfc in luna (or any SNES emulator) and press A, then B.

SNES Concepts

Mode 6: setMode(BG_MODE6, 0) — BG1 4bpp in hi-res, BG3 the offset table (not drawn), no BG2
Hi-res needs both screens: the sub screen draws the even half-pixel columns and the main screen the odd ones, so BG1 goes on both (setMainScreen(LAYER_BG1) and setSubScreen(LAYER_BG1)), as in Mode 5
16-half-pixel tiles: a map entry for character N draws N on the left and N + 1 on the right; this applies to BG3 too, so each offset word moves a 16-half-pixel column (anomie-regs, Backgrounds / Mode 6)
The table, as in Mode 2: BG3 row 0 = horizontal words, row 1 = vertical words; bit 13 = apply to BG1; value in bits 0-9; column 0 is never offset
BG3VOFS picks the rows: bgSetScroll(2, 0, 0); the lib writes BG3's VOFS raw in Modes 2/4/6
4bpp tiles at run time: 16 characters from pixels with tileEncode4bpp(), no asset
Upload timing: the row is built during the picture and DMA'd at the start of VBlank, as in mode2

What You'll See

Diagonal rainbow bands with fine vertical stripes, whose columns ripple up and down as a travelling wave; the leftmost column stays put.
A: the columns slide sideways instead. A again: back to vertical.
B: the wave stops and the test card appears. B again: back.

The test card: bit 3 of a hi-res offset

Test card

luna's native capture after B: every odd column is shifted half a tile.

In a 256-pixel mode the low 3 bits of a horizontal offset are not applied (fullsnes e7b0c5d598767874: "Lower 3bit of HORIZONTAL offsets are ignored"; snesdev-wiki d594aeedde1b87c2: "the fine scroll of 0-7 pixels is retained"), so offsets move whole 8-pixel tiles. In Mode 6 the scroll counts half-pixels and a tile is 16 of them, so bit 3 — the value 8 — falls inside a tile, and no reference says whether it is applied. The card writes 8 into every odd column and 0 into the even ones, nothing else, so the answer is in the picture:

luna and ares apply it (and bsnes, by luna's team's reading of its code): odd columns move by half a tile, and the steps of the diagonal bands are cut into half steps (the image above). ares: sfc/ppu/background.cpp, hoffset = hpixel + (hlookup & ~7) + … (corpus chunk a81897dc4e20fe4b).
Mesen2 drops it: it puts the offset into a doubled scroll and halves the result (Core/SNES/SnesPpu.cpp, corpus chunk 008321f6eaec1f44), so the card looks like the plain bands. luna's team measured the two disagreeing exactly on the columns whose offset has bit 3 set.
A console: not measured yet. It is row 23 of the hardware protocol (docs/HARDWARE_VERIFICATION.md). ares and bsnes share an author, so their agreement is one implementation, not two witnesses.

For a game: horizontal offsets that are multiples of 16 look the same everywhere. If you check this example in Mesen2, its wave columns differ from luna's wherever an offset has bit 3 set; neither the example nor the lib is at fault. The corpus note is cartouche-fiches chunk 61be8d9e9e954b38.

Testing

testing/manifests/backgrounds_mode6.toml checks the mode, both screens and BG3VOFS, the wave in the vertical row at frame 100, then in the horizontal row after A with the vertical row back to zero (VRAM $6000 and $6040), and that every upload lands in blank. Without the A press it fails. backgrounds_mode6_card.toml checks the card: B at frame 110, then at 140 the wave is frozen and row 0 holds $2000/$2008 alternately with row 1 zero. It pins what the ROM writes, not what the PPU draws.

Modules Used

Module	Purpose
console	System initialization, setMode, WaitForVBlank
dma	Palette, tiles, map and the offset rows
background	BG1 and BG3 bases, BG3 scroll
input	The A button
math	fixSin for the wave
tile	The 4bpp characters, built from pixels