[2026春季][T1-1-1] HyosungSink

[HyosungSink]

Summary

• Implements the T1-1-1 ntops operators: rad2deg, copysign, lcm, lgamma, and nextafter.
• Adds NineToothed kernels, ntops.torch wrappers, registration, operator-scoped correctness tests, and operator-scoped performance tests.
• Covers 30 correctness cases and 20 performance cases per operator, for 150 correctness cases and 100 performance cases in total.
• Includes NVIDIA, Iluvatar, and MetaX adaptations behind vendor/device gates, keeping platform-specific fast paths isolated from the NVIDIA path.
• Connects the five operators to the InfiniCore use_ntops path while using PyTorch only as the test reference, not as the runtime implementation.
• Reports per-case performance for all 100 cases below; total minimum ratios are NVIDIA 0.913x, Iluvatar 0.932x, and MetaX 0.901x.

Validation

• Correctness passes on all reported platforms: NVIDIA L4/sm_89, Iluvatar MR-V100, and MetaX C500/MACA each pass the full 150-case T1-1-1 suite.
• Performance passes on all reported platforms: NVIDIA, Iluvatar, and MetaX each pass all 100 performance cases, with every case meeting torch_ms / ntops_ms >= 0.9.
• InfiniCore integration passes for all five operators through the use_ntops path on NVIDIA, Iluvatar, and MetaX.
• Platform isolation was reviewed: Iluvatar and MetaX specializations are guarded by vendor/device checks, and the NVIDIA path does not enter those architecture-specific branches.
• Submission hygiene was reviewed: no PyTorch runtime wrapping, no skipped or xfailed validation cases, and no lowered performance thresholds are used.

Performance Summary

Operator	Cases	NVIDIA min	Iluvatar min	MetaX min
rad2deg	20	0.976x	0.996x	0.926x
copysign	20	0.966x	0.975x	0.917x
lcm	20	0.913x	1.242x	0.901x
lgamma	20	0.953x	0.932x	0.969x
nextafter	20	0.961x	0.946x	0.946x
Total	100	0.913x	0.932x	0.901x

rad2deg Performance Cases

Case	NVIDIA ntops	NVIDIA torch	NVIDIA ratio	Iluvatar ntops	Iluvatar torch	Iluvatar ratio	MetaX ntops	MetaX torch	MetaX ratio
f16 large 1d	0.2693	0.2692	1.000x	0.1186	0.1257	1.060x	0.0525	0.0489	0.933x
f32 large 1d	0.5848	0.5710	0.976x	0.2335	0.2327	0.996x	0.0959	0.0941	0.981x
f64 large 1d	1.1509	1.1407	0.991x	0.1998	0.2032	1.017x	0.1887	0.1843	0.977x
f32 large 2d	0.5830	0.5769	0.989x	0.2333	0.2329	0.998x	0.0964	0.0942	0.978x
f16 large 3d	0.2691	0.2698	1.003x	0.1184	0.1254	1.058x	0.0521	0.0487	0.935x
f32 large 3d	0.5827	0.5738	0.985x	0.2333	0.2330	0.999x	0.0961	0.0944	0.983x
f64 large 3d	1.1546	1.1424	0.990x	0.2002	0.2037	1.017x	0.1857	0.1843	0.992x
f32 large out 1d	0.5828	0.5814	0.998x	0.2327	0.2335	1.003x	0.0956	0.0940	0.983x
f64 large out 2d	1.1540	1.1435	0.991x	0.2004	0.2029	1.013x	0.1855	0.1839	0.992x
f16 large out 3d	0.2661	0.2680	1.007x	0.1185	0.1238	1.045x	0.0514	0.0484	0.941x
f32 mid 1d	0.5809	0.5726	0.986x	0.2331	0.2327	0.998x	0.0957	0.0943	0.985x
f16 mid 1d	0.2668	0.2709	1.015x	0.1187	0.1252	1.055x	0.0513	0.0485	0.944x
f64 mid 1d	1.1502	1.1415	0.992x	0.2002	0.2035	1.016x	0.1894	0.1844	0.974x
f32 small 1d	0.5836	0.5780	0.990x	0.2330	0.2330	1.000x	0.0958	0.0943	0.985x
f16 noncontig 4096	0.2730	0.2688	0.985x	0.1183	0.1253	1.059x	0.0525	0.0486	0.926x
f32 noncontig 4096	0.5850	0.5714	0.977x	0.2332	0.2331	0.999x	0.0970	0.0946	0.975x
f64 noncontig 2048	0.2683	0.2663	0.993x	0.0523	0.0535	1.024x	0.0522	0.0504	0.966x
f32 noncontig out 4096	0.5930	0.5803	0.979x	0.2327	0.2333	1.003x	0.0969	0.0941	0.971x
f32 permute3d 256x256x128	0.2682	0.2657	0.991x	0.1185	0.1183	0.998x	0.0521	0.0504	0.967x
f32 permute3d out 256x256x128	0.2650	0.2659	1.003x	0.1181	0.1180	0.998x	0.0519	0.0495	0.952x

copysign Performance Cases

Case	NVIDIA ntops	NVIDIA torch	NVIDIA ratio	Iluvatar ntops	Iluvatar torch	Iluvatar ratio	MetaX ntops	MetaX torch	MetaX ratio
f16 large 1d	0.4356	0.4256	0.977x	0.1718	0.1860	1.082x	0.0769	0.0705	0.917x
f32 large 1d	0.8649	0.8607	0.995x	0.3376	0.3462	1.025x	0.1440	0.1390	0.966x
f64 large 1d	1.7334	1.7010	0.981x	0.2062	0.2029	0.984x	0.2733	0.2736	1.001x
f32 large 2d	0.8719	0.8672	0.995x	0.3407	0.3456	1.014x	0.1423	0.1390	0.977x
f16 large 3d	0.4436	0.4322	0.974x	0.1714	0.1852	1.080x	0.0751	0.0704	0.938x
f32 large 3d	0.8687	0.8583	0.988x	0.3380	0.3457	1.023x	0.1424	0.1390	0.976x
f64 large 3d	1.7360	1.6895	0.973x	0.2057	0.2027	0.985x	0.2727	0.2737	1.003x
f32 large out 1d	0.8670	0.8608	0.993x	0.3383	0.3456	1.021x	0.1405	0.1387	0.987x
f64 large out 2d	1.7392	1.7328	0.996x	0.2062	0.2032	0.986x	0.2724	0.2733	1.003x
f16 large out 3d	0.4426	0.4345	0.982x	0.1717	0.1855	1.081x	0.0751	0.0703	0.936x
f32 mid 1d	0.8669	0.8644	0.997x	0.3378	0.3455	1.023x	0.1405	0.1391	0.990x
f16 mid 1d	0.4343	0.4310	0.992x	0.1716	0.1857	1.082x	0.0739	0.0705	0.953x
f64 mid 1d	1.7180	1.6913	0.984x	0.2059	0.2028	0.985x	0.2716	0.2739	1.008x
f32 small 1d	0.8674	0.8591	0.990x	0.3380	0.3466	1.025x	0.1408	0.1389	0.987x
f32 broadcast rect 2048x8192	0.2726	0.2637	0.967x	0.0982	0.1912	1.948x	0.0701	0.2643	3.773x
f32 broadcast 4096	0.2724	0.2630	0.966x	0.0963	0.1911	1.983x	0.0889	0.2219	2.496x
f16 noncontig 4096	0.4444	0.4337	0.976x	0.1716	0.1850	1.078x	0.0748	0.0720	0.962x
f32 noncontig 4096	0.8715	0.8548	0.981x	0.3381	0.3460	1.023x	0.1429	0.1389	0.972x
f64 noncontig 2048	0.4356	0.4243	0.974x	0.0546	0.0532	0.975x	0.0752	0.0732	0.973x
f32 permute3d out 256x256x128	0.4407	0.4288	0.973x	0.1721	0.1761	1.023x	0.0755	0.0713	0.944x

lcm Performance Cases

Case	NVIDIA ntops	NVIDIA torch	NVIDIA ratio	Iluvatar ntops	Iluvatar torch	Iluvatar ratio	MetaX ntops	MetaX torch	MetaX ratio
i32 large 1d	0.8299	0.8783	1.058x	0.7592	1.1980	1.578x	0.5388	0.5212	0.967x
i32 large positive 1d	0.8266	0.8775	1.062x	0.5901	0.9836	1.667x	0.4818	0.4343	0.901x
i32 large 2d	0.8289	0.8758	1.057x	0.7593	1.1989	1.579x	0.5382	0.5213	0.968x
i32 large positive 2d	0.8276	0.8743	1.057x	0.5927	0.9571	1.615x	0.4821	0.4342	0.901x
i32 large 3d	0.8328	0.8746	1.050x	0.7441	1.1598	1.559x	0.5385	0.5215	0.968x
i32 large positive 3d	0.8289	0.8756	1.056x	0.5828	0.9516	1.633x	0.4817	0.4343	0.902x
i32 large out 1d	0.8311	0.8608	1.036x	0.7433	1.1598	1.560x	0.5374	0.5212	0.970x
i32 large out 2d	0.8317	0.8632	1.038x	0.7210	1.1240	1.559x	0.5384	0.5212	0.968x
i32 broadcast 8192	2.2100	2.5649	1.161x	1.6411	4.9779	3.033x	1.1024	2.2396	2.032x
i32 large low 1d	0.8306	0.8747	1.053x	0.4153	0.7198	1.733x	0.3039	0.3471	1.142x
i16 mid 1d	0.4892	0.4483	0.916x	0.5307	0.7479	1.409x	0.5121	0.5542	1.082x
i16 large 1d	0.5039	0.4690	0.931x	0.5313	0.7483	1.408x	0.5153	0.5543	1.076x
i64 mid 1d	1.6634	1.7177	1.033x	1.2172	4.3829	3.601x	0.5994	0.9209	1.536x
i64 large 1d	1.6428	1.7353	1.056x	1.2224	4.3806	3.584x	0.5990	0.9211	1.538x
u8 mid 1d	0.4331	0.3955	0.913x	0.4417	0.5512	1.248x	0.4292	0.3880	0.904x
i8 mid 1d	0.4013	0.3805	0.948x	0.3832	0.7197	1.878x	0.4183	0.5263	1.258x
i32 noncontig 4096	0.8354	0.8741	1.046x	0.7216	1.1239	1.557x	0.5434	0.5218	0.960x
i32 noncontig out 4096	0.8381	0.8628	1.030x	0.7197	1.1234	1.561x	0.5417	0.5212	0.962x
i16 noncontig 6144	1.2626	1.2803	1.014x	1.3421	1.6669	1.242x	1.1479	1.2346	1.076x
i32 permute3d out 256x256x128	0.4263	0.4397	1.031x	0.3654	0.5720	1.566x	0.2778	0.2685	0.967x

lgamma Performance Cases

Case	NVIDIA ntops	NVIDIA torch	NVIDIA ratio	Iluvatar ntops	Iluvatar torch	Iluvatar ratio	MetaX ntops	MetaX torch	MetaX ratio
f16 large 1d	0.2756	0.2679	0.972x	0.3863	0.3691	0.955x	0.2804	0.2962	1.056x
f32 large 1d	0.5803	0.5752	0.991x	0.3870	0.3609	0.932x	0.2774	0.2968	1.070x
f64 large 1d	11.4071	11.2792	0.989x	4.4191	10.0906	2.283x	3.7320	3.6298	0.973x
f32 large 2d	0.5847	0.5757	0.985x	0.3632	0.3389	0.933x	0.2775	0.2965	1.068x
f16 large 3d	0.2783	0.2676	0.962x	0.3623	0.3460	0.955x	0.2807	0.2964	1.056x
f32 large 3d	0.5839	0.5756	0.986x	0.3629	0.3386	0.933x	0.2776	0.2967	1.069x
f64 large 3d	11.3901	11.3198	0.994x	4.2130	9.8332	2.334x	3.7390	3.6237	0.969x
f32 large out 1d	0.5823	0.5831	1.001x	0.3632	0.3388	0.933x	0.2770	0.2965	1.071x
f64 large out 2d	11.3916	11.3031	0.992x	4.2129	9.7558	2.316x	3.7283	3.6147	0.970x
f16 large out 3d	0.2772	0.2645	0.954x	0.3621	0.3465	0.957x	0.2803	0.2965	1.058x
f32 mid 1d	0.5850	0.5757	0.984x	0.3630	0.3386	0.933x	0.2774	0.2964	1.069x
f16 mid 1d	0.2751	0.2684	0.975x	0.3628	0.3461	0.954x	0.2805	0.2963	1.056x
f64 mid 1d	11.3780	11.3018	0.993x	4.2691	9.9091	2.321x	3.7281	3.6198	0.971x
f32 small 1d	0.5815	0.5760	0.991x	0.3631	0.3385	0.932x	0.2773	0.2966	1.069x
f16 noncontig 4096	0.2813	0.2679	0.953x	0.3623	0.3462	0.956x	0.2809	0.2965	1.056x
f32 noncontig 4096	0.5854	0.5754	0.983x	0.3631	0.3386	0.933x	0.2781	0.2967	1.067x
f64 noncontig 2048	2.9838	2.8712	0.962x	1.0582	2.2172	2.095x	0.9696	0.9401	0.970x
f32 noncontig out 4096	0.5884	0.5841	0.993x	0.3631	0.3387	0.933x	0.2776	0.2964	1.068x
f32 permute3d 256x256x128	0.2775	0.2680	0.966x	0.1849	0.1747	0.945x	0.1442	0.1548	1.074x
f32 permute3d out 256x256x128	0.2761	0.2685	0.972x	0.1844	0.1746	0.947x	0.1437	0.1547	1.076x

nextafter Performance Cases

Case	NVIDIA ntops	NVIDIA torch	NVIDIA ratio	Iluvatar ntops	Iluvatar torch	Iluvatar ratio	MetaX ntops	MetaX torch	MetaX ratio
f16 large 1d	0.4218	0.4293	1.018x	0.1737	0.1883	1.084x	0.0791	0.0762	0.964x
f32 large 1d	0.8754	0.8567	0.979x	0.3334	0.3449	1.034x	0.1443	0.1404	0.973x
f64 large 1d	1.7087	1.6975	0.993x	0.7302	0.9705	1.329x	0.2760	0.2753	0.997x
f32 large 2d	0.8341	0.8579	1.029x	0.3647	0.3450	0.946x	0.1427	0.1404	0.983x
f16 large 3d	0.4265	0.4297	1.007x	0.1737	0.1875	1.079x	0.0804	0.0761	0.946x
f32 large 3d	0.8429	0.8556	1.015x	0.3333	0.3445	1.034x	0.1427	0.1405	0.984x
f64 large 3d	1.7130	1.7088	0.998x	0.7304	0.9526	1.304x	0.2765	0.2755	0.996x
f32 large out 1d	0.8335	0.8580	1.029x	0.3338	0.3448	1.033x	0.1419	0.1402	0.988x
f64 large out 2d	1.7086	1.7063	0.999x	0.7296	0.9514	1.304x	0.2756	0.2751	0.998x
f16 large out 3d	0.4351	0.4314	0.991x	0.1734	0.1872	1.080x	0.0803	0.0761	0.948x
f32 mid 1d	0.8506	0.8571	1.008x	0.3335	0.3446	1.033x	0.1420	0.1404	0.988x
f16 mid 1d	0.4249	0.4276	1.006x	0.1735	0.1873	1.079x	0.0789	0.0764	0.968x
f64 mid 1d	1.7145	1.7048	0.994x	0.7293	0.9517	1.305x	0.2758	0.2753	0.998x
f32 small 1d	0.8367	0.8528	1.019x	0.3332	0.3446	1.034x	0.1421	0.1404	0.988x
f32 broadcast rect 2048x8192	0.2683	0.2633	0.981x	0.0990	0.2595	2.622x	0.0698	0.2658	3.808x
f32 broadcast 4096	0.2675	0.2628	0.982x	0.0964	0.2597	2.693x	0.0926	0.2163	2.336x
f16 noncontig 4096	0.4291	0.4282	0.998x	0.1738	0.1874	1.078x	0.0814	0.0792	0.973x
f32 noncontig 4096	0.8486	0.8563	1.009x	0.3331	0.3446	1.035x	0.1436	0.1405	0.979x
f64 noncontig 2048	0.4375	0.4204	0.961x	0.1866	0.2450	1.313x	0.0763	0.0776	1.017x
f32 permute3d out 256x256x128	0.4387	0.4284	0.976x	0.1701	0.1757	1.033x	0.0767	0.0747	0.974x

Notes

• PyTorch is used only as the test reference, not as the runtime implementation.
• Latest Iluvatar nextafter float16 fix is gated by half and iluvatar; NVIDIA paths remain on the existing ntops/NineToothed kernels.
• MetaX C500/MACA performance was measured on the current main PR branch with the submodule hashes listed above; nextafter uses the MetaX 1024/8 configuration before the Iluvatar 256/4 branch.