D906 - A one-layer depth array is the held depth target

Status: decided Date: 2026-10-19

A dispatch may read a 32_FLOAT image in a depth swizzle as a 2D array of one layer: SQ_IMG_RSRC_WORD3.TYPE 2D_ARRAY, with DEPTH and BASE_ARRAY both zero, read by a program that addresses it as an array. That image is the depth target at its base, which the host holds, as a 2D image is (D894). Its depths come from the host and are bound as an array of one layer.

An array of more than one layer is not one target, and is read from guest memory as before.

Why: Grand Theft Auto V's loading-screen program reads its 1024 x 1024 depth this way. Read from guest memory, the image's metadata looked like DCC keys of which a quarter were compressed. They are HTILE, one word per 8 x 8 tile, every word 0xfffffff0: depth cleared to 1.0 (radv_cmd_buffer.c:15355-15360, ac_descriptors.h:148-154). Guest memory holds only the clear; the host holds the depths drawn since.

Rejected:

• Filling the image from its HTILE clear value: it would read 1.0 where the host has drawn depth.
• Classifying 64KB_Z_X metadata as DCC keys. addrlib sizes and addresses a 64KB_Z_X colour surface's keys as 64KB_R_X's (tools/addrlib/dcc under DCC_SWIZZLE), but no title is known to need it, and a depth image's metadata is not DCC.