Arch Linux User Repository

Hi all, I no longer use this package and have no interest in maintaining it. Before disowning it I'd ideally like to make someone a co-maintainer. Please reach out if you're interested. Thanks!

Hi, I've pushed a fix to PKGBUILD, it build now with CUDA 12.5.0.

Thanks @agowa for your analysis, I took your suggestions and applied them to CMake files directly.

Sorry @sidicer, I didn't see your comment sooner, thanks for looking more deeply at LTO. I've added -ffat-lto-objects to the CMake flags.

Using this gcc package: https://aur.archlinux.org/gcc12.git and this line cmake -DCMAKE_CXX_FLAGS="-ffat-lto-objects" -DCUDA_LIB=/usr/local/cuda/lib64/stubs/libcuda.so -DCUDA_TOOLKIT_ROOT_DIR=/opt/cuda -DCMAKE_C_COMPILER=/usr/bin/gcc-13 -DCMAKE_CXX_COMPILER=/usr/bin/g++-13 -S . -B build makes it compile successfully again.

/opt/cuda/bin/gcc and /opt/cuda/bin/g++ appear to no longer exist. For me this cmake command worked: cmake -DCMAKE_CXX_FLAGS="-ffat-lto-objects" -DCUDA_LIB=/usr/local/cuda/lib64/stubs/libcuda.so -DCUDA_TOOLKIT_ROOT_DIR=/opt/cuda -DCMAKE_C_COMPILER=/usr/bin/gcc -DCMAKE_CXX_COMPILER=/usr/bin/g++ -S . -B build

regardless of that the actual build afterward still fails because ArchLinux ships GCC 14.1 and that is apparently incompatible... I get the same error as in this forum post and their resolution is basically "there is none, it's not supported."....

https://forums.developer.nvidia.com/t/cuda-12-4-nvcc-and-gcc-14-1-incompatibility/293295

@hv15 I was building on a GTX780 for testing purposes which uses 11.4 Cuda version, but I installed the latest 12.* Cuda package on the system.
Will remove everything, install 11.* Cuda and update if that worked

But just a heads up as I do not think that was the issue as after building on RTX3080 with latest CUDA i got the same undefined symbol: blake256_hash error.
Hoping for a random missing library or a package 🤞

UPDATE:

Replacing -fno-lto flag with -ffat-lto-objects links libraries but uses LTO which is required for the blake256_hash to build. Currently on RTX3080 machine I was able to build xmrig-cuda and use it with this.
Apologies @hv15 for not investigating more at first

GTX780 not successful

diff --git a/PKGBUILD b/PKGBUILD
index 9afa5d1..33f5ac4 100644
--- a/PKGBUILD
+++ b/PKGBUILD
@@ -24,7 +24,7 @@ prepare () {
 build() {
   cd "${pkgname}-${pkgver}"

-  cmake -DCMAKE_CXX_FLAGS="-fno-lto" -DCMAKE_C_COMPILER=/opt/cuda/bin/gcc -DCMAKE_CXX_COMPILER=/opt/cuda/bin/g++ -S . -B build
+  cmake -DCMAKE_CXX_FLAGS="-ffat-lto-objects" -DCMAKE_C_COMPILER=/opt/cuda/bin/gcc -DCMAKE_CXX_COMPILER=/opt/cuda/bin/g++ -S . -B build
   cmake --build build
 }

@sidicer, thanks for the no-LTO tip, I can confirm it build for me as well. I've pushed a new PKGBUILD. On my system I am able to get it working:

 * CUDA GPU     #0 01:00.0 NVIDIA RTX A3000 Laptop GPU 1485/5501 MHz smx:32 arch:86 mem:5807/5937 MB

I'm not sure about the missing blake256_hash symbol, I can only guess that some lib dependency isn't being met, what is the output of CMake when building xmrig-cuda?

Could not push changes to the repo, so i'll post here: To fix the symbol 'fatbinData' is already defined error - add -DCMAKE_CXX_FLAGS="-fno-lto" to the cmake line in PKGBUILD

EDIT: Replacing -fno-lto flag with -ffat-lto-objects links libraries but uses LTO which is required for the blake256_hash to build.

Same issue like @silkky

Getting the same issue as @SycoLTH, building with CUDA 12.3 in chroot. Anyone find a fix?

{standard input}: Assembler messages:
{standard input}:213135: Error: symbol `fatbinData' is already defined
{standard input}:589064: Error: symbol `fatbinData' is already defined
{standard input}:673670: Error: symbol `fatbinData' is already defined
{standard input}:758439: Error: symbol `fatbinData' is already defined
{standard input}:843031: Error: symbol `fatbinData' is already defined
{standard input}:927660: Error: symbol `fatbinData' is already defined
{standard input}:927844: Error: symbol `fatbinData' is already defined
{standard input}:1012301: Error: symbol `fatbinData' is already defined
{standard input}:1071861: Error: symbol `fatbinData' is already defined
{standard input}:1136826: Error: symbol `fatbinData' is already defined
make[3]: *** [/tmp/ccTx9vXq.mk:2: /tmp/cc4ePyvf.ltrans0.ltrans.o] Error 1
make[3]: *** Waiting for unfinished jobs....
lto-wrapper: fatal error: make returned 2 exit status
compilation terminated.

@SycoLTH Sorry, I can't reproduce the error message... what version of CUDA do you have? Have you tried building in a chroot?