pgf90-Error-CUDA version 5.0 is not available in this instal

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi,

I am trying to compile and CUDA FORTRAN code however I got some rather odd messages.
Running on a Cray machine, among my modules loaded we find :

cudatoolkit/5.5.20-1.0402.7700.8.1
craype-accel-nvidia35
PrgEnv-pgi/4.2.34

Now, I compile the code :

pgf90 -c  -O3 -r8 -Mpreprocess -Mcuda=cuda5.0,cc30 -ta=tesla,cc30 -Minfo=accel cuf_code.cuf -o  cuf_code.o

And this is what I get:

pgf90-Error-CUDA version 5.0 is not available in this installation

However, nvcc does not agree with that:

nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2013 NVIDIA Corporation
Built on Tue_Jul__9_18:28:15_PDT_2013
Cuda compilation tools, release 5.5, V5.5.0

What is going on here?

Thank you very much

Luiceur

2

Hi Luiceur,

PGI ships with the compilers all CUDA components necessary to build your code. This is independent of the CUDA driver you may have installed.

With each PGI release we will default to using the CUDA version one generation behind the current released version but also include the current version which can be used via the “-ta” flag. For example in PGI 14.10, the default version is CUDA 6.0, with CUDA 6.5 available as well.

Due to the size of the packages, we decided to not ship older CUDA versions. So what’s happening here is that you’re trying to use a older CUDA version but the PGI package did not include it.

Your options are to use a more current CUDA version by either removing “cuda5.0” or changing it to “cuda5.5” (assuming you’re using a PGI version that shipped it) or use an older version of PGI.

Ideally, Cray would update the CUDA driver to a more current version. However, we understand that there’s support and validation reasons why they don’t. Because of this, in next years 15.1 release, we will include CUDA 5.5 as well as 6.0 and 6.5.

  • Mat
3

Hi Matt,

That makes sense. I’ll give it a go with an older version of PGI.
Thanks for your help.

Luiceur

4

I came to the similar yet not the same problem.

[gpuserver@gpuserver gpu_hardware_test]$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Sep__1_21:08:03_CDT_2017
Cuda compilation tools, release 9.0, V9.0.176

[gpuserver@gpuserver gpu_hardware_test]$ pgf90 peakbandwidth.cuf
pgf90-Error-CUDA version 9.0 is not available in this installation. Please read documentation for CUDA_HOME to solve this issue

Any suggestion? Hope to get response.

5

Hi zongchen@nmefc,

Which PGI compiler version are you using?

We typically ship the current version of CUDA plus one or two previous versions at the time the PGI compiler was released. For PGI 19.5, that was CUDA 10.1, 10.0, and 9.2.

For earlier CUDA versions as the message from the compiler indicates, you’ll need to set the environment variable “CUDA_HOME” to the CUDA SDK install directory and the PGI compiler will use that CUDA installation.

Alternatively, you can update your CUDA driver to a newer version.

Full details can be found in our release notes: PGI Release Notes Version 19.5 for x86 and NVIDIA Processors

Hope this helps,
Mat

6

Hi, Mat

PGI version is pgf90 18.10-1 64-bit target on x86-64 Linux -tp haswell.

Following your instruction, I set the CUDA_HOME directory to CUDA SDK install directory. Yes, it works.

Thanks for your reply!

Zongchen

7

Hi Mat,

I have set the CUDA_HOME to CUDA SDK directory and I am using NVIDIA HPC_SDK 20.9, cuda-11.1. But error was like this:
pgfortran-Error-CUDA version 11.1 is not available in this installation. Please read documentation for CUDA_HOME to solve this issue

Any suggestion for this?

Thanks.

Howard

8

Hi Howard,

Since 20.9 was released before CUDA 11.1 was available, it doesn’t support CUDA 11.1. CUDA 11.1 support will be available in the upcoming NVHPC SDK 20.11 release.

-Mat