where are the fortran modules cudafor, cusparse, etc?

Accelerated Computing GPU-Accelerated Libraries
1

Hi guys,

sorry for a dumb question. I’m just starting to look into using cublas and cusparse for our code (written in fortran90, built with intel compilers).

The manuals casually mention that there’s a bunch of fortran system modules, such as cudafor, cusparse, etc, but all I see in my installation of the CUDA-10.0 is a bunch of *.h files and a couple of *.c files with fortran interfaces. The installation is handled by the IT department, so I have no clue whether the installation is complete.

So are the module files missing or are the users supposed to build them themselves?

Thanks in advance.

2

It would help if you indicated which manuals you are referring to.

cudafor is part of the CUDA fortran compiler toolchain, which is provided by PGI:

[url]https://devblogs.nvidia.com/easy-introduction-cuda-fortran/[/url]

It is not included in a standard CUDA toolkit install.

3

Thanks for the quick response.

I see, the manuals I was referring to are indeed by PGI: CUDA Fortran Compiler Information | PGI

It’s a pity that Nvidia does not provide the fortran modules for CUDA libs. Intel MKL, for example, does, and it saves us quite a bit of work. At this point I’m not even sure whether we will want to pursue using cusparse.

4

It has been a few years since I last shelled out big bucks for Intel’s development suite, but as I recall MKL is not given away free of charge. Neither are Intel’s compilers and tools.

I have not checked PGI’s pricing, but would assume that for an amount of money similar to what you would pay Intel, they would be happy to sell you the CUDA Fortran compiler with the modules you mention.

5

Before I hijack a similar thread or post a new question, is it ok to ask people’s experience with PGI when working with CUDA or it should go to PGI’s forums regardless?

@Njuffa, I am using MKL and didn’t need to shell out $$$ for it, but support is paid. BTW, I once had a ICC/IFort license for Linux, but am interested in checking PGI and specially its FlexLM integration.

6

When I last acquired a license, I acquiredt icc/ifort/MKL/VTune as a combo, and shelled out ~$2000 altogether as I recall. I have been using multiple versions of Intel’s toolchain since about 1995, and generally speaking their pricing has gone up substantially over time. May well be justified giving the improved functionality and robustness.

Available combos and pricing may have changed drastically since my last purchase. For one thing, these days Intel has good competition from high-performance GPUs where many tools are free of charge, which may help keep their prices in check.

I am not a fan of fostering a something-for-nothing entitlement mentality when it comes to software, by providing “free beer” software.

7

That’s why I like to hear people from the old school.
Though I enjoyed getting free or hugely discounted licenses in the university (but no free beer), since I could only pay the bus tickets and cheap lunch, today I subscribe to:

I have to check if FlexLM is required just for PGI’s trial licenses or if the compiler “embeds” license checkout code on a release. The company I work for uses FlexLM for its products, but I don’t really know how the license verification is called.
Performance-wise, I will search a bit more on how PGI compares to compilers in the GPU business.