large gridsize bug when using cuda11

mumax / 3

GPU-accelerated micromagnetic simulator

Other

447 stars 150 forks source link

large gridsize bug when using cuda11 #268

Closed JeroenMulkers closed 3 years ago

JeroenMulkers commented 3 years ago

Bug description:

The following input script

N := 512
setGridSize(N,N,1)
setCellSize(1,1,1)
save(B_demag)

gives the following error

panic: CUDA_ERROR_INVALID_VALUE
github.com/mumax/3/cuda/cu.MemsetD32Async(0xb1c000000, 0xc000000000, 0x100000, 0x0)
        C:/Users/jmulkers/GO/src/github.com/mumax/3/cuda/cu/memset.go:25 +0x126

when mumax3.10 is build against cuda11. Simulations with smaller grid sizes (e.g. for N:=256) run fine. This error does not occur in builds against older cuda versions.

Environment

System: Windows 10
Nvidia driver: 451.22
GPU: Titan RTX
CUDA version: 11.0

godsic commented 3 years ago

@JeroenMulkers I ran bench.mx3 on 970, 980 Ti, 1080 and Titan Xp but the issue never showed up. Have you reproduced with most recent CUDA 11.0 Update 1 and on Linux?

JeroenMulkers commented 3 years ago

This recent update for CUDA11.0 seems to resolve the issue on windows! @godsic thank you for pointing out that there has been an update

@JLeliaert , can you confirm that this windows build does not have the bug that you reported to me?

JLeliaert commented 3 years ago

confirmed

JeroenMulkers commented 3 years ago

Building mumax3.10 against cuda11.0.3 (update 1) instead of cuda11.0.1 solves this issue on both windows and linux.

The binaries on the mumax.ugent.be are updated, and the download links are active again.

askeyjw commented 3 years ago

Hello, I am using CUDA driver 11.2, driver version 461.40 and windows 10 but am getting this error even after uninstalling and re-installing mumax3.10. Not quite sure what I'm doing wrong!