nvcc fatal : Unsupported gpu architecture 'compute_20'

Please help me. I want compile MatConvNet with GPU. Windows10 + nvidia quadro p400 + matlab2017b + cuda10 + visual studio 2017. But there is an error, it tells me this:

nvcc fatal : Unsupported gpu architecture 'compute_20'

How can I fix this error?

vl_compilenn:  CUDA: MEX config file: 'C:\Program Files\MATLAB\R2017b\toolbox\distcomp\gpu\extern\src\mex\win64\mex_CUDA_win64.xml'
  nvcc fatal   : Unsupported gpu architecture 'compute_20' 
  Error using vl_compilenn>nvcc_compile (line 540)
  Command "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0\bin\nvcc" -c
  "D:\MATLAB\GAN\dcgan-matconvnet-master\matconvnet-1.0-beta24\matlab\src\bits\data.cu" -DNDEBUG
  -DENABLE_GPU -DENABLE_DOUBLE -D__SSSE3__ -gencode=arch=compute_20,code=\"sm_20,compute_20\"
  -gencode=arch=compute_30,code=\"sm_30,compute_30\" -I"C:\Program
  Files\MATLAB\R2017b\extern\include" -I"C:\Program
  Files\MATLAB\R2017b\toolbox\distcomp\gpu\extern\include"
  -gencode=arch=compute_20,code=\"sm_20,compute_20\"
  -gencode=arch=compute_30,code=\"sm_30,compute_30\" -O3 -Xcompiler /MD --compiler-bindir
  "C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\bin" -o
  "D:\MATLAB\GAN\dcgan-matconvnet-master\matconvnet-1.0-beta24\matlab\mex\.build\bits\data.obj"
  failed.

[Walter Roberson]

Which NVIDIA drivers and toolkit do you have installed?

[Karthik K]

cuda 10..

[Karthik K]

screenshots of installed drivers

[Walter Roberson]

See https://stackoverflow.com/questions/50247154/nvcc-fatal-unsupported-gpu-architecture-compute-20-while-compiling-matlab

You might need to go up to compute_62 cm_62,compute_62

[Karthik K]

I did not get you sir. Can you please explain..

[Walter Roberson]

Find the file that contains

-gencode=arch=compute_30,code=\"sm_30,compute_30\"

add to it

-gencode=arch=compute_62,code=\"sm_62,compute_62\"

[Karthik K]

Not able to solve. I still have the same error.

[Walter Roberson]

You might need to use the config file at https://github.com/vlfeat/matconvnet/blob/master/matlab/src/config/mex_CUDA_glnxa64.xml

Unfortunately I do not have a Windows development system with a GPU attached (my Windows development system is a virtual machine that cannot access my GPU), and because of the details of how Microsoft licenses, it would be rather expensive for me to put together such a system. Microsoft licensing really only recognizes home users (5 or less lifetime Windows licenses) and full for-profit business licenses (yearly subscription at thousands of dollars per year.) As I started using PCs in the late 80s, I have worn out enough PCs to exceed the lifetime cap on "home" licenses. :(

[Priyanka Bharadwaj]

I tried adding

opts.defCudaArch = [...

'-gencode=arch=compute_20,code=\"sm_20,compute_20\" '...

'-gencode=arch=compute_30,code=\"sm_30,compute_30\"'...

'-gencode=arch=compute_62,code=\"sm_62,compute_62\"'...

'-gencode=arch=compute_61,code=\"sm_61,compute_61\"'];

Later I got the following error

Error using mex

nvcc fatal : redefinition of keyword 'code'

Could you please let me know how I can resolve this?

I am using CUDA_8.0 on ubuntu 16.04 with matlab 2019b

[Walter Roberson]

It appears that you might need to trim it down to just one -gencode line

[Priyanka Bharadwaj]

I tried that and now the error still persists

opts.defCudaArch = [...

'-gencode=arch=compute_61,code=\"sm_61,compute_61\" ' '-gencode=arch=compute_62,code=\"sm_62,compute_62\" ' '-gencode=arch=compute_30,code=\"sm_30,compute_30\"'];

Error using mex

nvcc fatal : Unsupported gpu architecture 'compute_20'

[Walter Roberson]

Sorry, I do not know.

Answer 1

[Joss Knight]

Strictly speaking, for R2017b you need to be using an older version of the CUDA toolkit. MATLAB is expecting you to be using a version which still supports Fermi architecture.

The easiest fix when calling MEX or MEXCUDA is probably to set the NVCCFLAGS variable directly for your Pascal card:

mexcuda NVCCFLAGS="-gencode=arch=compute_60,code=sm_60" ...

but I'm not quite sure how MEX is invoked by the MatConvNet compiler.

[Priyanka Bharadwaj]

Hi I did modify

opts.defCudaArch = [...

'-gencode=arch=compute_61,code=\"sm_61,compute_61\" ' '-gencode=arch=compute_62,code=\"sm_62,compute_62\" ' '-gencode=arch=compute_30,code=\"sm_30,compute_30\"'];

aloong with modifying the

mexcuda NVCCFLAGS="-gencode=arch=compute_60,code=sm_60" ...

But I still get the following error. Could you please suggest me how to rectify this. I am using matlab2019b along with Cuda 8.0 on ubuntu 16.04

nvcc fatal : Unsupported gpu architecture 'compute_20'