mirror of
https://github.com/par274/sharpemu.git
synced 2026-07-22 19:06:15 +08:00
João Victor Amorim
472fc96a37
The VOP3P emitter rejected any packed op with the clamp bit set. Clamp saturates each f16 output half to [0, 1] (and flushes NaN to 0, matching RDNA), so games that emit clamped packed arithmetic fell back to a loud emit failure. Apply the saturation to the f32 result of each lane, before it is narrowed back to f16. Because 0.0 and 1.0 are exact in both f32 and f16 and the clamp is monotonic, clamping in f32 and then rounding to f16 yields the same value as clamping the f16 result directly; for the fused multiply-add the pre-narrowing value is the round-to-odd f32, which preserves that equivalence through the final round-to-nearest-even. The saturation uses ordered compares so a NaN result collapses to 0 without a separate IsNan test. Verification: - The local exact-reference harness now also clamps: add, mul, and fma each compared against an f16-domain clamp reference (NaN -> 0, else [0, 1]) over directed boundary inputs and 24M random cases. 0 mismatches, alongside the existing 34M unclamped fma cases. - ShaderDump pk-f16 gains a clamped add and a clamped fma; all decode and emit. - The exec program computes the pinned fma with clamp (both lanes exceed 1.0, so each saturates to 0x3C00) and stores it at offset 28; GpuConformance checks it on device. All values match on an AMD Radeon RX 7700 XT.