Winograd F(4x4,3x3) #1643 has decreased performance on my machine

[kuba97531]

After changes done in #1643 I have observed slowdown rather than speedup on my machine.
I have compared builds from commits 6333b66 and 7cfbb72 it seems that on my machine the new version is slower.
System Information:
OS: Windows 2010
GPU: 1080Ti
Build: Release build from visual studio 2017

The table presents the results of benchmark after running the full tuner for both versions.
I run benchmark with two commands:
leelaz.exe -w 153 --benchmark
leelaz.exe -w 153 --benchmark -t 4

build	n/s	n/s (-t 4)
7cfbb72 (new)	446, 451, 451, 452, 451	664, 656, 659, 659, 665
6333b66 (old)	454, 454, 454, 459, 458	676, 695, 700, 691, 699

The results don't differ by much, but all tests I do show a very clear tiny slowdown.

I case it matters I have attached a tuning.
tuning.txt

Is it possible I am doing something wrong?

[kuba97531]

I've also uploaded the two compiled .exe files
https://drive.google.com/open?id=1BtiXRye_UpaU4-hpjmhSD_EX5CY03yQt

[gcp]

These results are, quite frankly, terrible.

This is a GTX 1060, quite a bit less than half the peak TFLOPS of your 1080 Ti:
374 n/s default, 501 n/s t=4

Your card is benchmarking only 20-30% faster.

[tterava]

LZ tends to work best when you use the same -t parameter in tuning as you do in normal use. I get about 950 n/s on gtx 1080 using -t 4. Seems there's some other issue with your system.

[kuba97531]

It seems that the problem might be with my build. The benchmark on the official leelaz.exe 0.15 is 600 and 900 (t -4) so a bit more in line with the expectations.

Do you have any idea what I might be doing wrong regarding the build procedure?
Are official windows binaries build from visual studio or with some other tool?

[gcp]

I build with MSVC2017, latest release. I do use a newer OpenBLAS that is not available in Nuget by building that from source with mingw64 (I just noticed it's missing Ryzen 2 support, so I'll have to update it again). The DLL is in the release package. Maybe not set to optimize or missing a define like NDEBUG or something?

I benchmarked on a GTX 1070 on Windows 10 and got ~530 n/s with t=4. That's not great, but it's still clearly better than the older versions, which did ~488 n/s.

So there's probably something wrong with your build, but comparing to Linux benchmarks is probably going to show some difference too because of Windows' terribly slow GPU computing support.

[kuba97531]

I close the issue until I resolve my local problems and reproduce the issue.