D868 - A half-float target is drawn as floats and held as halves

Status: assumed Date: 2026-10-14

A colour target whose CB_COLOR0_INFO names COLOR_16, NUMBER_FLOAT and SWAP_STD holds one IEEE half float a texel, red. It lies by the two-byte layout, as a COLOR_16 UINT target does (D852), and a host carries each texel as a word with the half in its low sixteen bits. A drawer draws it through R32_SFLOAT: a half is widened to its float exactly where words enter the attachment - seeded, reloaded, or filled uniformly - and a float is rounded to the nearest half, a tie to the even, where they leave it - read back, by rows, or through a snapshot. Draws and blending run in 32-bit float.

Known by: assumed. That the hardware rounds a half-float target's result to the nearest, ties to even, is the IEEE default and what Vulkan's R16_SFLOAT attachments do on the hosts here; it is not measured on the console. Blending in 32 bits where the hardware may blend nearer 16 can differ in the last bit of a half.

Why: NUMBER_FLOAT is 7 and COLOR_16 is 2 (gfx103.json). Grand Theft Auto V's second submission draws into such a target once its 8_8_8_8 SNORM target is drawn (D866). Widening at the attachment's edges, where every word crosses, keeps the host's word-a-texel handling of two-byte targets as it is.

Rejected:

• An R16_SFLOAT attachment: its texels are two bytes, and every buffer, read and reload sized by a texel's bytes would carry half the words the host's two-byte handling gives it.
• Converting in the host's texel load and store: those take only the component order, which is carried through every frame and band structure that reaches them.