TEST-71, commit c2967a5e, Windows x64, AMD Radeon AI PRO R9700, driver 26.11.2 (LLPC), LLVM 23.1.6 from conda-forge. Both native and emulated runs had exactly one mismatch out of 100, wave64 index 23. All 64 lanes of v31 retained the buffer fill pattern. No skips, untranslated programs or harness errors. The cause is unconfirmed. #1051 records a similar symptom with a different seed and register.
Reproduce with the synthetic generator; no game files required:
python -I tools/valu_fuzz.py --family f16 --seed 221122 --programs 50
python -I tools/valu_fuzz.py --family f16 --seed 221122 --programs 50 --emulateThe comparison is in tools/valu_fuzz.py:603. GPU report: https://github.com/ItsNotRexed/Leonida/issues/166#issuecomment-7078580740. AI-assisted.
Activity
2Same fill-pattern (0xeeeeeeee) retention reproduces on a consumer AMD Radeon RX 9070 XT (RDNA 4 / GFX12) with an earlier LLPC driver (26.10.44), and this time in wave32 (seed 1, program 237, cs_f16fuzz_s1_237_w32) in both native and --emulate modes, at commit f85582e.
- Failing GPU: AMD Radeon RX 9070 XT (RDNA 4 / GFX12), AMD
26.10.44 (LLPC),driverVersion 2.0.395, Vulkan1.4.349, Windows 11 (VK_LOADER_LAYERS_DISABLE=~implicit~) - Passing GPU (same machine): AMD Radeon(TM) Graphics (Ryzen 7 9800X3D RDNA 2 iGPU), AMD
24.30.18 (AMD proprietary shader compiler),driverVersion 2.0.331, Vulkan1.3.302— passes 500/500 natively on the exact same run (--family f16 --seed 1 --programs 250).
Result on RX 9070 XT (499/500 in both native and --emulate):
MISMATCH cs_f16fuzz_s1_237_w32 (128 words); print it with --family f16 --show 1:237:32
thread 0 v20: got 0xeeeeeeee, want 0x00007e00 (compared bits 0xffff0000; last written by op 1: v_fmac_f16_e32 v20, v21, v25)
thread 0 v30: got 0xeeeeeeee, want 0x49117e00 (compared bits 0xffff0000; last written by op 5: v_pk_add_f16 v30, v21, v20 op_sel:[1,0] op_sel_hi:[1,1])Same compared bits 0xffff0000 pattern as your reduced v31 case: the lower 16 bits are NaN (0x7e00 from v_mul_f16_e64 v25, v32, v28 consuming sNaN v28.lo = 0xfd9d), the upper 16 bits are valid and judged, yet all 64 lanes retain 0xeeeeeeee in v20 (after v_fmac_f16_e32 v20, v21, v25) and v30 (after v_pk_add_f16 v30, v21, v20 op_sel:[1,0] op_sel_hi:[1,1]), while v25 and all other registers are stored normally.
So this is not wave64-only or specific to 26.11.2 / R9700 — it reproduces across RDNA 4 LLPC drivers (26.10.44 and 26.11.2) in both wave32 and wave64, while passing on the non-LLPC AMD iGPU driver.
Synthetic fuzzer program only; no game files. AI-assisted.
Reduced on main bb2e1842, using the matching compute harness, Windows x64, LLVM 23.1.6, AMD Vulkan 26.11.2. Keeping only generated ops 5, 6 and 10 (zero-based, out of 15) reproduces the discrepancy. The generator initialization and stores are unchanged. Removing any one of those three makes the native R9700 check pass. Three repeats per cell:
Only the upper 16 bits of v31 are judged here (lane 0 expects 0x3e4e). The different emulated readback on this commit is worth keeping distinct from the original report on c2967a5e. These measurements leave the cause open.
From the repo root, save the following as repro.py and run python -I repro.py --family f16 --build , then repeat with --emulate. Set LLVM_MC and LEONIDA_VK_DEVICE for your machine. This uses the upstream generator, translator, model and GPU runner with one synthetic input.
AI-assisted.