[milescoviello]

What

sceAgcDcbSetNumInstances (libSceAgc/DcbDraw/src/Instancing.cpp) writes the PM4 NUM_INSTANCES packet that sets the instance count for the draws after it. No test covers it, and nine of the ten titles here call it: Borderlands 4 (PPSA01494, 7 call sites), Ghost of Tsushima (PPSA02225, 18), Forspoken (PPSA03984, 15), and Metal Gear Solid 4 (PPSA18194), Crimson Desert (PPSA25286), MXGP 26 (PPSA30839), Super Meat Boy 3D (PPSA31162), Call of Duty: Black Ops II (PPSA34502) and Steins;Gate Re:Boot (PPSA36591) with one each. Crimson Desert, Black Ops II and Steins;Gate also call sceAgcDcbSetNumInstancesGetSize. In the three titles with the most call sites, 23 of their 40 calls pass the constant 1, read from the instructions before each call. I didn't trace the other 17.

How

New test agc_num_instances (tests/NumInstances.cpp), registered in libSceAgc/CMakeLists.txt after agc_draw_index_multi_instanced, the same way. No code change. It checks:

• sceAgcDcbSetNumInstancesGetSize returns 8.
• Counts 1, 2 and 1000, written one after another into the same buffer, each come out as a 2-dword packet at the cursor: header 0xc0002f00 (type 3, one payload dword, opcode 0x2F) and the count. Each call returns the packet's start and moves the cursor by two dwords, and nothing is written after the third packet.

These are the words the driver's PM4 parser takes: it expects 0x2F to be 2 dwords long (Pm4.cpp:314) and reads the instance count from the second (Pm4.cpp:1065).

Tested

Linux (Ubuntu 26.04.1, GCC 15.2), on main 2227f13661:

• agc_num_instances passes 3 of 3 at this head (b7220cf006).
• With Instancing.cpp writing opcode 0x2E instead (local mutant a89f04e956), it fails with instance count 1 wrote 0xc0002e00 0x00000001, not NUM_INSTANCES with the count.
• Windows 11 VM (WinLibs GCC 15.2, Release): agc_num_instances passes 3 of 3 at this head.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Fixes: none
• Unblocks: none
• GPUs tested: n/a
• AI-assisted: yes

🤖 Generated with Claude Code

[MarcZX10]

I reviewed the packet layout against the implementation and parser references, then built and ran the new test with the exact-head and ; it passed. The local harness stubbed only (the driver notification), while retaining the real packet allocation and encoding path. Diff check and repository conventions also pass. No correctness issue found in this focused test addition.

[MarcZX10]

Correction to my approval summary: I compiled and ran the exact-head Instancing.cpp and Packet.cpp with the PR test; it passed. The local harness stubbed only AgcDriverPrepareCommandWrite_nid_no_patch (the driver notification), while keeping the real packet allocation and encoding path. The omitted identifiers in the review body were a quoting issue; the test result and review are unchanged.

[milescoviello]

Closing this as a duplicate of my own #11678, which already checks sceAgcDcbSetNumInstances and sceAgcDcbSetNumInstancesGetSize in tests/Commands.cpp. My duplicate search went by PR text, and #11678's title calls it the instance count packet, so I missed it. @MarcZX10, thanks for the review, and sorry for the extra work.

AI-assisted: yes (Claude Code)

[github-actions]

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 5021	core/libs/prx/libSceAgc/CMakeLists.txt