CODEWITHSUNDEEP
cudaneural-networkdeep-learning
angubenko
angubenko

Reputation: 274

Darknet framework fails to start with GPU acceleration using CUDA

I was trying to run Darknet with GPU acceleration using CUDA API. So I followed instructions from here, changed GPU=1 in Makefile and started make. When I'm trying to run test it fails due to CUDA Error.

./darknet yolo test cfg/yolo.cfg yolo.weights data/dog.jpg CUDA Error: unknown error darknet: ./src/cuda.c:21: check_error: Assertion `0' failed.

I'm using Ubuntu 14.04, CUDA 7.5 and my NVIDIA-SMI 352.93 and Driver Version: 352.93 on Titan X I'm pretty sure that my CUDA works fine and driver's version is up to date, because I'm using it to accelerate Caffe. My guess is that Darknet cannot locate CUDA directory.

Can anyone help me with that issue?

Upvotes: 4

Views: 10792

Answers (6)

jaylee
jaylee

Reputation: 416

I had the same problem. Modifying the ARCH in Makefile solved my problem. My gpu was GTX 1080 (I checked it out using nvidia-smi)

> $nvidia-smi

I found the setting suitable for my gpu here.

http://arnon.dk/matching-sm-architectures-arch-and-gencode-for-various-nvidia-cards/

I changed my arch setting in the Makefile as follows.

ARCH= -gencode arch=compute_61,code=[sm61,compute_61] \

Upvotes: 2

Peter
Peter

Reputation: 91

I just encountered this problem and solved it. My solution was

sudo rm -rf ~/.nv

and then reboot.

Upvotes: 0

Marcus
Marcus

Reputation: 121

Maybe u should modify the ARCH in Makefile.Ur GPU score can be found in this site: https://developer.nvidia.com/cuda-gpus

Upvotes: 0

hjl240
hjl240

Reputation: 11

Agree with “user6568204”. in addition, you can modify the value of configure parameter 'ARCH' in \Makefile, and you can find other 'ARCH' value in this website: http://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#virtual-architecture-feature-list or look this picture virtual-architecture-feature-list

Upvotes: 1

user6568204
user6568204

Reputation: 66

you should modify the value of configure parameter 'ARCH'. the default value is --gpu-architecture=compute_52, --gpu-code=compute_52. my setting is --gpu-architecture=compute_30, --gpu-code=compute_30, and it works. it depends on your actual gpu architecture. more detail is in cuda toolkit documentation.

Upvotes: 5

kangshiyin
kangshiyin

Reputation: 9779

You could find out the reason by reading the code in ./src/cuda.c at line 21, and checking what assertion is failed, as indicated by the error message. 

./src/cuda.c:21: check_error: Assertion `0' failed.

Upvotes: 0

Related Questions