2.80 and cuda kernel 3.0

hi,

For some time I notice building/x64 WIn MSVC 2017/ error for cubin 3.0 /I have GTX760/

(…)

73>kernel_split_avx.cpp
78>Generating filter_sm_30.cubin
78>CUSTOMBUILD : error : cuew init fialed (0xffffffff)
78>
78>CUSTOMBUILD : error : initialization error, exiting
78>Kompilowanie projektu „cycles_kernel_cuda.vcxproj” wykonane — NIEPOWODZENIE.
80>------ Kompilacja rozpoczęta: Projekt: buildinfoobj, Konfiguracja: Release x64 ------
81>------ Kompilacja rozpoczęta: Projekt: python_numpy, Konfiguracja: Release x64 ------
77>cycles_util.vcxproj -> D:\Programowanie\blender_build\lib\Release\cycles_util.lib

Support for 3.0 kernel abandoned?

the sm_30 kernel is still supported, but the build process has issues loading the cuda dlls required for building it., if i had to guess i’d say you have cuda 9.2 installed? downgrade to 9.1 and you should be good to go again!

Thx for reply.

Actually I have CUDA 10 installed but some weeks ago I had no issues with building 2.80…with it.
Put it still on development?

Hmm strange.

It should not have tried to use cuew to build with cuda 10, can be a couple of things, outdated master code, i added support for cuda 10 a couple of weeks ago, if you haven’t pulled since then you might miss the code needed.

second option is the WITH_CYCLES_CUBIN_COMPILER is set to ON, check your cmakecache.txt file to validate.

This all being said, cuda10 is broken for sm_30 it makes broken render kernels, and things will look wonky as soon as glass shaders come into play (which is one of the reasons we still ship with 9.1)

1 Like

I update/check both blender and library git on daily basis so trhay have to be up to date.;
Is there any date set to implement glass shader ?

it’s not a problem on the blender side, it’s a cuda bug, it’s been dragging on since cuda 9.2 which came out in may sooo… it doesn’t seem like nvidia is in a rush to fix it.

Great ((( :wink:

Btw. Would it matter to me /cuda kernel 3.0-gtx760 + Blender/ if I switch back from CUDA 10 to 9.1 anyhow ?
I mean: would I loose any efficiency and/or features with my graphics ?
There were some problems with installing 9.1 afaik…

it’s the recommended thing to do for your card, you won’t lose any features and the broken things will just start working.

1 Like

I noticed that while I can use CUDA 10 to compile on Windows (which renders without visual problems on a 1080ti), it runs slower than when I use CUDA 9.1.
I don’t know if this is true under Linux as well, but at the moment it looks like there’s nothing to be gained from a CUDA version later than 9.1.

1 Like

the only thing v10 has going for it is RTX20xx support…

A bit offtop but… Blender should benefit greatly on hardware accelerated raytracing ?

…another “feature” of CUDA 10 is native support for Visual Studio 2017 and later and support for gcc 7.3.0 in Ubuntu 18.04 / Mint 19 :slight_smile:

This post by Lukas Stockner might explain it a little: Cycles Development Updates

1 Like

We resolved the compiler hostage situation quite a while ago, there’s a cmake option WITH_CYCLES_CUBIN_COMPILER if you turn it on, it’ll use nvrtc to build the kernels, relieving you from being forced into a compiler version nvcc approves of.

1 Like

When will buildbot have 20xx support

unknown, it’s mostly depended on nvidia putting out a cuda compiler that’ll work for all cards.

So never? :smiley:

v9.2 was broken for all cards, v10 is seemingly only broken for sm_30, so i have high hopes for v10.1/v11/vWhateverisnext