Build error

[nvlabs]

Build error

[Tom94]

I believe I found the problem (CUDA 10.2 only supports constexpr __device__ functions and not variables) and am currently working towards a fix over there: #9

Fingers crossed!

In the interim, if you are able to update your CUDA version to 11.4/.5, that should do the trick

[mmalex]

may i ask what version of cuda you have installed?

[phongnhhn92]

I am using cuda 10.2

[mmalex]

thanks! ive not seen these errors before but they made me feel like it might be a too-old cuda compiler or perhaps a too-old or mis-detected gpu architecture. do you know what architecture was detected during the initial cmake step? iirc it is printed out. if building on a machine with multiple or no gpus, you may need to explicitly set a cuda architecture via environment variable, as mentioned in the readme. but i admit im clutching at straws at this moment.

[afiaka87]

@mmalex I receive the same error when building on an RTX 2070 Super (has tensor cores). The CUDA architecture is correctly detected as 75 in the prior build step.

cmake --version
cmake version 3.22.1

gmake --version
GNU Make 4.3
Built for x86_64-pc-linux-gnu

nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2020 NVIDIA Corporation
Built on Mon_Nov_30_19:08:53_PST_2020
Cuda compilation tools, release 11.2, V11.2.67
Build cuda_11.2.r11.2/compiler.29373293_0

neofetch 
----------------                                           
OS: Pop!_OS 21.10 x86_64                                           
Host: B450M DS3H                                           
Kernel: 5.15.8-76051508-generic                                         
Shell: bash 5.1.8                                                        
CPU: AMD Ryzen 9 3900X (24) @ 3.800GH                                           
GPU: NVIDIA GeForce RTX 2070 SUPER                                           
Memory: 3392MiB / 32113MiB      

nvidia-smi
Fri Jan 14 19:29:09 2022       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.86       Driver Version: 470.86       CUDA Version: 11.4     |

Manually setting TCNN_CUDA_ARCHITECTURES to 75 gives the same error

TCNN_CUDA_ARCHITECTURES=75 cmake --build build --config RelWithDebInfo -j 16

[aoliao12138]

I solve this error by using cuda 11.4.

[MultiPath]

Hi I am also having the same error "error: A device variable cannot be marked constexpr"
My GPU is V100 and I set the architecture to 70. Is it too old for this code?

Thanks

[MultiPath]

And to add, my current cuda version is 11.2 but still seem to have similar problems.. Although it is a bit hard to upgrade to 11.4 on my current server though

[Tom94]

Please feel free to re-open this issue if the compile error persists despite the attempted fix I just merged.

FWIW, CI successfully compiles CUDA 10.2 on Ubuntu 18.04 now.

[afiaka87]

Confirmed working, thanks @Tom94