Hi, I'm the maintainer of upbge-git which is basically a blatant copy of your package, slightly modified for UPBGE.
I just found that recent builds giving me intermittent CUDA errors like "Misaligned address" and it looks like your package shows the same symptom.
In comparison, the experimental build from the Blender homepage doesn't seem to be affected.
It happens in a few of my scenes with large textures but not in others.
I tried to download CUDA but it didn't make any difference.
If you have an idea what might be causing this issue, or what I can do to track down the cause, please let me know. Thanks!
Pinned Comments
bartus commented on 2020-01-16 16:01 (UTC) (edited on 2021-11-01 14:05 (UTC) by bartus)
Current packages state (non-conflicting ones):
bartus commented on 2019-04-10 11:40 (UTC) (edited on 2024-03-01 09:41 (UTC) by bartus)
Note: This is a non-conflicting version of blender-git package with GPU architecture detected at build time.
This package is also hosted on GitHub.
issues
andpatches
at GitHub.Use env vars to control the build process:
DISABLE_CUDA=1
to skip cuda kernel build (cycles+compositor+optix,requires:cuda)DISABLE_HIP=1
to skip amdgpu kernel build (requires:hip-runtime-amd)ENABLE_PYTHON_INSTALL
include bundled python (default:off)DISABLE_NINJA=1
to switch the build system to make (default:ninja)DISABLE_OPTIX=1
to skip Optix device build.DISABLE_DRACO=1
to skip gltf exporter draco mesh compression extension.FRAGMENT="#{commit,tag,branch}=..."
for making bisect build.CUDA_ARCH="sm_xx sm_yy"
to build for a specific Cuda arch, supports multiple values.HIP_ARCH="gfx900 gfx1103"
to build for a specific amdgpu arch, supports multiple values.MAKEFLAGS="xxx"
to override default make flags (check oom-killer disclaimer below)MAKEFLAGS="xxx:yyy"
to override default make flags (colon-separated list)Usage cases:
export DISABLE_CUDA=1
before buildDISABLE_CUDA=1 ~your-aur-helper~
makepkg DISABLE_CUDA=1
yay -S blender-develop-git --mflags "DISABLE_CUDA=1 DISABLE_HIP=1"
Yay
users warning:Since yay relies solely on
aur-rpc
to solve dependencies instead of callingmakepkg --sync
it can't follow dependencies logic defined inPKGBUILD
. That's why it's up to the user to provide dependencies to enable desired features. Optional features includes cuda/optix/usd.tl;dr; If you want to prebuild cycles kernels, you have to install CUDA beforehand. If you don't have CUDA installed,
PKGBUILD
will fail unless you disable cuda with DISABLE_CUDA=1, same for optix and usd.Out of memory killer.
You may use ninja-mem to prevent oom-killer when building on system with low memory to core ratio.
To activate use MAKEFLAGS+=" -m75" where 75 is the percentage upper threshold of memory usage when new build jobs is postponed.
Or simply lower the make jobs count in "MAKEFLAGS" variable, but this will prolong your build process.