[petsc-dev] Makefile.user and CUDA

[Jed Brown]

Matthew Knepley <knepley at gmail.com> writes:

> On Mon, Oct 12, 2020 at 3:03 PM Patrick Sanan <patrick.sanan at gmail.com>
> wrote:
>
>>
>>
>> Am 12.10.2020 um 20:11 schrieb Matthew Knepley <knepley at gmail.com>:
>>
>> On Mon, Oct 12, 2020 at 3:47 AM Patrick Sanan <patrick.sanan at gmail.com>
>> wrote:
>>
>>> I have a toy application code built on PETSc which needs to compile and
>>> link a .cu file.
>>>
>>> I'd love to be able to configure PETSc (with CUDA), and then use a
>>> modified version of share/petsc/Makefile.user to compile and link my code,
>>> using a consistent set of compilers, libraries, and flags.  Makefile.user
>>> uses petsc.pc (via pkg-config) and implicit GNU make rules to do almost
>>> everything for you for C, C++, and Fortran.
>>>
>>> However, I don't think it currently supports CUDA, and I'm not familiar
>>> enough with BuildSystem or pkg-config to quickly add support myself, so I
>>> resort to the "old" way, including things like this in my Makefile:
>>>
>>
>> Okay, here is where the pkgconf file gets generated, and in fact where the
>> language sections are
>>
>>
>> https://gitlab.com/petsc/petsc/-/blob/master/config/PETSc/Configure.py#L161
>>
>> I think you can just put a "CUDA" section in. I do not know what names to
>> use, so I have not done it.
>>
>>
>> That's where I got stuck as well :D
>>
>> I can follow the pattern to get the compiler (CUDAC) and some of the
>> flags, but that doesn't seem to have all the information that CUDAC,
>> CUDAC_FLAGS, CUDA_INCLUDE, and CUDA_LIB provide in petscvariables.
>>
>
> I can dig around and find all these for you.
>
>
>> Naively I would assume I'd need to dig around to figure out how those
>> variables are populated, and get the same info into petsc.pc
>>
>> But, I worry that I'm misunderstanding fundamentals about pkg-config and
>> how it is supposed to work. Am I free to put whatever fields I want in
>> there, or is there some authority on what's "standard"?
>
> I do not understand anything about pkg-config (and think it is a
> fundamentally misguided mechanism). Jed, how should CUDA work with this?

I have no experience, but here are some examples.  I think the caller is responsible for using nvcc if you have *.cu source.

$ cat /usr/lib/pkgconfig/cublas.pc 
cudaroot=/opt/cuda
libdir=${cudaroot}/targets/x86_64-linux/lib
includedir=${cudaroot}/targets/x86_64-linux/include

Name: cublas
Description: CUDA BLAS Library
Version: 11.0
Libs: -L${libdir} -lcublas
Cflags: -I${includedir}


$ cat /usr/lib/pkgconfig/nvrtc.pc 
cudaroot=/opt/cuda
libdir=${cudaroot}/targets/x86_64-linux/lib
includedir=${cudaroot}/targets/x86_64-linux/include

Name: nvrtc
Description: A runtime compilation library for CUDA C++
Version: 11.0
Libs: -L${libdir} -lnvrtc
Cflags: -I${includedir}