512 filter nets does not work with Cuda backend on gtx cards

[Kovax007]

BUG REPORT

Describe the bug
A clear and concise description of what the bug is.

Steps to Reproduce

1. only running with 512 size filter nets
   Expected behavior: runing normally, as cudnn with --backend-opts=custom_winograd=false
   Observed behavior: Unhandled exception in worker thread: CUDA error: too many resources requested for launch (C:\lc0\src\neural\cuda\winograd_helper.inc:665), if i reduce manually the kOpInpTransformBlockSize than the error is: Unhandled exception in worker thread: CUDA error: too many resources requested for launch (C:\lc0\src\neural\cuda\winograd_helper.inc:819)

Lc0 version
Lc0 master: (commit: 025105e), Windows 10 with cuda 11.5, compiled binary, cuda backend

Lc0 parameters
With and without: --backend-opts=max_batch=256

Hardware

• gtx1080.
• 16Gb
• i74790k

PS D:\Leela\0.28\ONNX> .\lc0_master.exe
_
| _ | |
|_ |_ || v0.29.0-dev+git.dirty built Mar 2 2022
Detected 4 core(s) and 8 thread(s) in 1 group(s).
Group 0 has 4 core(s) and 8 thread(s).
Found configuration file: D:\Leela\0.28\ONNX/lc0.config
go nodes 1
Loading Syzygy tablebases from Z:/Syzygy/dtz;Z:/Syzygy/wdl
Found 510 WDL, 0 DTM and 510 DTZ tablebase files.
Found pb network file: D:\Leela\0.28\ONNX/512x15-t79_9-swa-256000.pb.gz
Creating backend [cuda]...
CUDA Runtime version: 11.5.0
Latest version of CUDA supported by the driver: 11.6.0
GPU: NVIDIA GeForce GTX 1080
GPU memory: 7.99969 Gb
GPU clock frequency: 1885.5 MHz
GPU compute capability: 6.1
Unhandled exception in worker thread: CUDA error: too many resources requested for launch (C:\lc0\src\neural\cuda\winograd_helper.inc:665)
PS D:\Leela\0.28\ONNX> .\lc0_launchbond.exe
_
| _ | |
| |_ |_| v0.29.0-dev+git.dirty built Mar 7 2022
Detected 4 core(s) and 8 thread(s) in 1 group(s).
Group 0 has 4 core(s) and 8 thread(s).
Found configuration file: D:\Leela\0.28\ONNX/lc0.config
go nodes 1
Loading Syzygy tablebases from Z:/Syzygy/dtz;Z:/Syzygy/wdl
Found 510 WDL, 0 DTM and 510 DTZ tablebase files.
Found pb network file: D:\Leela\0.28\ONNX/512x15-t79_9-swa-256000.pb.gz
Creating backend [cuda]...
CUDA Runtime version: 11.5.0
Latest version of CUDA supported by the driver: 11.6.0
GPU: NVIDIA GeForce GTX 1080
GPU memory: 7.99969 Gb
GPU clock frequency: 1885.5 MHz
GPU compute capability: 6.1
Unhandled exception in worker thread: CUDA error: too many resources requested for launch (C:\lc0\src\neural\cuda\winograd_helper.inc:819)
PS D:\Leela\0.28\ONNX>

[borg323]

@ankan-ban can you take a look at this?