Skip to content

gfx-replay-gaps (2026-11-01, batch r3): D24 / D24S8 readback, separate texture / sampler handles, integer 3D / cube / multisample textures

Branch claude/gfx-replay-gaps-2026-11-01-r3 (Mac, Metal, base 84cb5a6). Contract: gfx.md section 13 ("Replay gaps").

What changed

  1. D24 / D24S8 texture -> buffer copies (mtl_replay.m copy_d24_from_tex). DEPTH24 (0x33) and DEPTH24_STENCIL8 (0x35) are 32-bit float depth (+ stencil) in Metal. A readback blits the depth (and stencil) aspect into staging memory, and a kernel writes NVN's packed words: unorm24 in the low 24 bits for 0x33, depth << 8 | stencil for 0x35. It is the exact inverse of the upload (copy_d24_to_tex). This replaces "reading back a 24-bit depth texture ... is not implemented".
  2. Separate texture / sampler handles:
  3. The replay stores BindSeparateTexture(s) / BindSeparateSampler(s) and their Deferred forms per stage (S.septex / S.sepsamp), replacing "separate texture/sampler binding is not supported yet".
  4. The translator recognises a bindless TEX whose handle register is the OR (LOP.OR, IADD, or LOP3 with an OR table) of two constant buffer words. It records the pair (ShaderTexture.separate, sep_cbuf / sep_offset).
  5. bind_stage ORs the two words, each from c2's handle tables (handle_word) or from a user uniform buffer.
  6. The c2 offsets of the separate tables (SEP_TEX_C2_BASE 0x120, SEP_SAMP_C2_BASE 0x520, 8 bytes per binding, in shader/maxwell.h) are a hypothesis: no title binds separate handles. The replay logs, once, the offsets a shader uses.
  7. Integer views of 3D, cube and 2D multisample targets are declared texture3d / texturecube / texture2d_ms<uint|int> and read through the new wrappers IntTex3D, IntTexCube (face and texel from the direction) and IntTex2DMS (read by sample). Before this, they kept the float declaration, which is a Metal validation error.
  8. TRANSLATOR_VERSION 16 -> 17.

Tests

  • New: t_d24_readback (64 / 64 words for both formats) and t_sep_sampler (clamp vs repeat sampler through c2 words, a uniform buffer pair with IADD).
  • Extended: t_int_tex cases 4-6 (a 3D RGBA8UI texture, a cube view of a 2D RGBA8UI texture, a 4x RGBA8UI target filled by a draw and read with TLD.MS).
  • t_xlate pins the integer wrappers (with Metal compiles) and the separate pair.
  • make -C gfx test: ALL TESTS PASSED. The LAS corpus still gives 5767 / 5767 / 5767.

Not verified on a title

None of the four titles reads back a 24-bit depth texture in the scenes reached, binds separate handles, or samples an integer 3D, cube or multisample view. All three items are tested (unit tests under the Metal debug layer), and the four-title boots show no regression (integrator file).