[YuutaTsubasa]

Follow-up to #215: the per-draw shader constant cost on the main thread.

What changed

• Only the constants a shader reads. Every draw used to byte-swap both stages' whole 4 KiB constant files into the upload ring (8 KiB a draw). XenosRecomp reads a float4 constant only through the name the shader's constant table gives it, so each native shader now records those registers (src/shader_constant_use.cpp, from the guest container's table) and a draw swaps only them. The rest of each 4 KiB block keeps what an earlier draw left; the shader never reads it.
• Relative addressing gets everything. A shader whose ALU instructions use a0/aL constant addressing may read past its array, so it still uploads all 256 registers: 33 of the game's 287 vertex shaders, no pixel shaders. The microcode walk vertex_palette.cpp already had moved to src/xenos_microcode.h so both use it.
• Skinning palette copied once. It is swapped from guest memory straight into the ring instead of copy, swap, copy.
• Switches. SFR_CONSTANT_RUNS=0 (also on the Test options page, and kept in the Android diagnostic ZIP) restores the full upload and the old palette copy. SFR_TEST_CONSTANT_POISON=1 fills the registers a draw does not write with NaNs. SFR_CONSTANT_TIMING=1 prints the constant writes' ns and bytes per draw, plus what each shader uploads.

Dirty tracking at write time was not needed after this: the writes left are ~620 bytes a draw.

A/B on the AYN Thor

One APK (versionCode 61), SFR_CONSTANT_RUNS=0 for the baseline. Moving Dolphin Resort race, HOST_THREADS over frames 13000-21000, frame metrics off, runs interleaved. drawn_ms stays near 16 ms because of the 60 fps pacing.

Constants	draw_ms	record_ms	cpu_ms	undrawn_ms	frames skipped of 8040
all (baseline)	3.75 / 3.73 / 3.73	1.04 / 1.01 / 1.01	14.36 / 14.27 / 14.25	9.82 / 9.81 / 9.57	214 / 127 / 124
only the ones read	3.45 / 3.39 / 3.38	0.75 / 0.73 / 0.72	13.77 / 13.84 / 13.63	9.21 / 9.62 / 8.42	35 / 34 / 15

Mean change: draw_ms -0.33, record_ms -0.29, cpu_ms -0.54 ms a frame. Race speed was 0.9994 in every run.

Per-draw timing of the constant writes (SFR_CONSTANT_TIMING=1, one run each, two clock reads of ~0.13 us included):

	ns / draw	bytes / draw
Thor, before	561	8192
Thor, after	237	623
Windows i9 Vulkan, before / after	776 / 199	8192 / 914
Windows i9 D3D12, before / after	778 / 140	8192 / 868

A frame has about 820 draws. The Windows machine had other load on it, so its frame-level numbers are not reported.

Pixel check

Screenshots from two runs can't be compared pixel for pixel in this harness. The menus animate in real time and the fixed-step race diverges: two runs of the same build differ in about 520,000 pixels at present 5000. Behaviour was checked three other ways instead:

• Statically, all 476 shaders of the 0.10.13 pack.
  • In SPIR-V, every OpAccessChain into the stage's constant block uses a constant row inside the uploaded registers.
  • In DXIL, every cbufferLoadLegacy on the stage's cbuffer does the same.
  • Dynamic rows appear in exactly the 33 relative shaders, which get the full upload.
  • The game's own split (SHADER_CONSTANT_RUNS) matches.
• NaN poison. SFR_TEST_CONSTANT_POISON=1 races on Windows Vulkan and D3D12 render normally (characters, skinning, effects, HUD). Any read of an unwritten register would show up as NaN garbage.
• Visual comparison. Vulkan and D3D12 races with the change on and off look the same.

Tests

• Windows: ctest 166/166.
• Linux (WSL, clang 15): ctest 159/159.
• Python: 353 tests OK, 11 skipped without the game files.
• Android arm64 build.
• Windows Vulkan race with the Khronos validation layer: 0 errors, 455 shipped pipelines loaded.
• Windows D3D12 race.
• Thor races at speed 0.9994.

New unit test: tests/shader_constant_use_test.cpp.

Risks

• A future translator change. The ranges rely on how XenosRecomp reads constants. If a new translator version read a register outside the constant table, that read would see stale data. The poison switch and the pack checks above would catch it. SFR_CONSTANT_RUNS=0 turns the change off.
• Shaders outside the pack (mods, runtime translation) are classified the same way from their own containers. Anything this code doesn't understand falls back to the full upload.
• SFR_CONSTANT_UPLOAD_REUSE and SFR_CONSTANT_REUSE_TRACE still compare and upload whole arrays, as before.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.