System.Numerics.Tests.Perf_VectorConvert.Convert_ulong_double has regressed a lot on x86

[adamsitnik]

System.Numerics.Tests.Perf_VectorConvert.Convert_ulong_double has regressed x25 on x86, but improved for x64.

Reporting system:

Repro:

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f net6.0 net7.0 --filter System.Numerics.Tests.Perf_VectorConvert.Convert_ulong_double --architecture x86

Result	Base	Diff	Ratio	Operating System	Bit
Faster	1617.02	574.60	2.81	Windows 11	X64
Faster	890.92	410.82	2.17	Windows 11	X64
Faster	3121.80	1263.22	2.47	Windows 11	X64
Faster	1445.69	725.92	1.99	Windows 10	X64
Faster	2120.73	809.30	2.62	Windows 11	X64
Faster	2465.62	1009.49	2.44	Windows 11	X64
Faster	953.99	423.17	2.25	ubuntu 18.04	X64
Faster	966.86	425.70	2.27	ubuntu 20.04	X64
Faster	2308.55	965.59	2.39	ubuntu 18.04	X64
Faster	2771.01	1333.80	2.08	ubuntu 18.04	X64
Faster	2147.64	991.60	2.17	pop 20.04	X64
Faster	1263.20	736.57	1.71	alpine 3.13	X64
Faster	1259.47	730.80	1.72	debian 11	X64
Same	585.29	579.48	1.01	macOS Monterey 12.2.1	Arm64
Same	676.53	676.12	1.00	Windows 10	Arm64
Same	678.29	678.43	1.00	Windows 11	Arm64
Slower	3176.11	43751.84	0.07	Windows 10	X86
Slower	1597.47	32929.64	0.05	Windows 10	X86
Slower	1446.79	33592.24	0.04	Windows 10	X86
Faster	53217.69	34251.08	1.55	Windows 10	Arm
Faster	2875.52	1047.20	2.75	macOS Big Sur 11.6.3	X64
Faster	2455.51	887.62	2.77	macOS Monterey 12.2.1	X64
Faster	2373.55	984.52	2.41	macOS Monterey 12.2.1	X64

cc @tannergooding

Most likely related to #65189

[ghost]

Tagging subscribers to this area: @dotnet/area-system-numerics
See info in area-owners.md if you want to be subscribed.

Issue Details

---

System.Numerics.Tests.Perf_VectorConvert.Convert_ulong_double has regressed x25 on x86, but improved for x64.

Reporting system:

Repro:

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f net6.0 net7.0 --filter System.Numerics.Tests.Perf_VectorConvert.Convert_ulong_double --architecture x86

Result	Base	Diff	Ratio	Operating System	Bit
Faster	1617.02	574.60	2.81	Windows 11	X64
Faster	890.92	410.82	2.17	Windows 11	X64
Faster	3121.80	1263.22	2.47	Windows 11	X64
Faster	1445.69	725.92	1.99	Windows 10	X64
Faster	2120.73	809.30	2.62	Windows 11	X64
Faster	2465.62	1009.49	2.44	Windows 11	X64
Faster	953.99	423.17	2.25	ubuntu 18.04	X64
Faster	966.86	425.70	2.27	ubuntu 20.04	X64
Faster	2308.55	965.59	2.39	ubuntu 18.04	X64
Faster	2771.01	1333.80	2.08	ubuntu 18.04	X64
Faster	2147.64	991.60	2.17	pop 20.04	X64
Faster	1263.20	736.57	1.71	alpine 3.13	X64
Faster	1259.47	730.80	1.72	debian 11	X64
Same	585.29	579.48	1.01	macOS Monterey 12.2.1	Arm64
Same	676.53	676.12	1.00	Windows 10	Arm64
Same	678.29	678.43	1.00	Windows 11	Arm64
Slower	3176.11	43751.84	0.07	Windows 10	X86
Slower	1597.47	32929.64	0.05	Windows 10	X86
Slower	1446.79	33592.24	0.04	Windows 10	X86
Faster	53217.69	34251.08	1.55	Windows 10	Arm
Faster	2875.52	1047.20	2.75	macOS Big Sur 11.6.3	X64
Faster	2455.51	887.62	2.77	macOS Monterey 12.2.1	X64
Faster	2373.55	984.52	2.41	macOS Monterey 12.2.1	X64

cc @tannergooding

Most likely related to #65189

Author:	adamsitnik
Assignees:	-
Labels:	area-System.Numerics, tenet-performance
Milestone:	-

[tannergooding]

The previous algorithm was incorrect for x86 and producing incorrect results. This was an intentional change and something that is non-trivial to fix without further work elsewhere.

[dakersnar]

Closing as dup of #65189