System.Numerics.Tests.Perf_VectorConvert.Convert_double_long has regressed

[adamsitnik]

System.Numerics.Tests.Perf_VectorConvert.Convert_double_long has regressed on 64 bit systems and improved on 32 bit.

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_double_long

It seems that @performanceautofiler has missed the regression (cc @DrewScoggins ): https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_Windows%2010.0.18362%2fSystem.Numerics.Tests.Perf_VectorConvert.Convert_double_long.html

System.Numerics.Tests.Perf_VectorConvert.Convert_double_long

Result	Base	Diff	Ratio	Operating System	Bit	Processor Name
Slower	1641.19	5656.11	0.29	Windows 11	X64	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Slower	1568.26	4675.84	0.34	Windows 11	X64	AMD Ryzen 9 5900X
Slower	2917.81	6127.01	0.48	Windows 10	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	3950.44	8210.00	0.48	Windows 11	X64	Intel Core i5-4300U CPU 1.90GHz (Haswell)
Slower	2511.54	5407.97	0.46	Windows 10	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Slower	2391.06	4968.04	0.48	Windows 11	X64	Intel Core i7-8700 CPU 3.20GHz (Coffee Lake)
Slower	2467.05	5512.31	0.45	Windows 11	X64	Intel Core i9-9900T CPU 2.10GHz
Slower	3604.17	8129.05	0.44	Windows 11	X64	Unknown processor
Slower	3086.74	6799.54	0.45	Windows 11	X64	Unknown processor
Slower	1630.15	6176.46	0.26	ubuntu 20.04	X64	AMD Ryzen 9 5900X
Slower	2881.27	7115.21	0.40	ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	2282.50	11087.64	0.21	centos 7	X64	Intel Xeon CPU E5530 2.40GHz
Slower	1626.48	9297.52	0.17	ubuntu 18.04	X64	Intel Core i7-2720QM CPU 2.20GHz (Sandy Bridge)
Slower	2449.92	5405.11	0.45	alpine 3.13	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Slower	982.59	6879.18	0.14	ubuntu 18.04	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Slower	2599.96	5395.83	0.48	ubuntu 20.04	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Same	679.37	675.39	1.01	Windows 10	Arm64	Microsoft SQ1 3.0 GHz
Faster	20996.44	12145.39	1.73	Windows 10	X86	Intel Xeon CPU E5-1650 v4 3.60GHz
Faster	54109.68	30804.49	1.76	Windows 10	Arm	Microsoft SQ1 3.0 GHz
Slower	3566.29	7145.91	0.50	macOS Big Sur 11.6.3	X64	Intel Core i5-4278U CPU 2.60GHz (Haswell)
Slower	3369.04	9991.33	0.34	macOS Big Sur 11.4	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)

cc @tannergooding

[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_double_long has regressed on 64 bit systems and improved on 32 bit.

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_double_long

It seems that @performanceautofiler has missed the regression (cc @DrewScoggins ): https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_Windows%2010.0.18362%2fSystem.Numerics.Tests.Perf_VectorConvert.Convert_double_long.html

System.Numerics.Tests.Perf_VectorConvert.Convert_double_long

Result	Base	Diff	Ratio	Operating System	Bit	Processor Name
Slower	1641.19	5656.11	0.29	Windows 11	X64	AMD Ryzen Threadripper PRO 3945WX 12-Cores
Slower	1568.26	4675.84	0.34	Windows 11	X64	AMD Ryzen 9 5900X
Slower	2917.81	6127.01	0.48	Windows 10	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	3950.44	8210.00	0.48	Windows 11	X64	Intel Core i5-4300U CPU 1.90GHz (Haswell)
Slower	2511.54	5407.97	0.46	Windows 10	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Slower	2391.06	4968.04	0.48	Windows 11	X64	Intel Core i7-8700 CPU 3.20GHz (Coffee Lake)
Slower	2467.05	5512.31	0.45	Windows 11	X64	Intel Core i9-9900T CPU 2.10GHz
Slower	3604.17	8129.05	0.44	Windows 11	X64	Unknown processor
Slower	3086.74	6799.54	0.45	Windows 11	X64	Unknown processor
Slower	1630.15	6176.46	0.26	ubuntu 20.04	X64	AMD Ryzen 9 5900X
Slower	2881.27	7115.21	0.40	ubuntu 18.04	X64	Intel Xeon CPU E5-1650 v4 3.60GHz
Slower	2282.50	11087.64	0.21	centos 7	X64	Intel Xeon CPU E5530 2.40GHz
Slower	1626.48	9297.52	0.17	ubuntu 18.04	X64	Intel Core i7-2720QM CPU 2.20GHz (Sandy Bridge)
Slower	2449.92	5405.11	0.45	alpine 3.13	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Slower	982.59	6879.18	0.14	ubuntu 18.04	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Slower	2599.96	5395.83	0.48	ubuntu 20.04	X64	Intel Core i7-7700 CPU 3.60GHz (Kaby Lake)
Same	679.37	675.39	1.01	Windows 10	Arm64	Microsoft SQ1 3.0 GHz
Faster	20996.44	12145.39	1.73	Windows 10	X86	Intel Xeon CPU E5-1650 v4 3.60GHz
Faster	54109.68	30804.49	1.76	Windows 10	Arm	Microsoft SQ1 3.0 GHz
Slower	3566.29	7145.91	0.50	macOS Big Sur 11.6.3	X64	Intel Core i5-4278U CPU 2.60GHz (Haswell)
Slower	3369.04	9991.33	0.34	macOS Big Sur 11.4	X64	Intel Core i7-5557U CPU 3.10GHz (Broadwell)

cc @tannergooding

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

[tannergooding]

This one is somewhat by design.

The previously logic wasn't doing the right stuff since x86/x64 doesn't have a single vectorized instruction to convert double <-> long/ulong and so the new logic is now correct.

There is a pending work item to go back and accelerate the logic using SIMD intrinsics

[adamsitnik]

This one is somewhat by design.
There is a pending work item to go back and accelerate the logic using SIMD intrinsics

Should we close the issue then or would you prefer to wait until the pending work is finished?

[tannergooding]

We can leave it open until the pending work is finished.

[dakersnar]

@tannergooding We spotted this in the RC2 vs 6.0 perf report, is the pending work you referenced still ongoing?

[tannergooding]

This was by design where the previous logic was incorrect. AVX-512 hardware have a dedicated instruction that helps with the codegen on such hardware.