Thanks Jeff,

     The information is eventually there somewhere, the issue is more getting the information in a simple way, automatically, at PETSc configure time that is portable and will never crash. https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__DEVICE.html seems to require compiling a program and running it to get the information, this means invoking nvcc (what sub compiler to use for nvcc  with what flags etc? Not so easy on systems like Summit where their are multiple choices for the sub-compiler). So how complicated and fragile do we want to make PETSc configure be (for each particular piece of hardware) to always get the best information about the current hardware? 

   It looks like Kokkos really only needs the NVIDIA numerical generation information, not the code name, but their API requires both the codename (irrelevant) and the numerical information (relevant) in what is passed to Kokkos. The problem has always been generating the irrelevant part so Kokkos does not complain. We can, with a little pain, possibly automate completely the CUDA device information, the numerical part, but the mapping to code name has been problematic because it is hard to find in a single place the mapping from numerical information to codename. But I think, thanks to Max's input, I now understand the mapping and have put it in PETSc's configure.

    Generically, independent of Kokkos,  ideally I would run a single precompiled NVIDIA program that gave me all the information about the current hardware I was running and that would provide in simple format exactly the information I needed to configure PETSc, Kokkos etc for THAT system. The idea of support a multitude of hardware is important for package management systems, but is not important for 99% of PETSc users who are configuring for exactly the hardware they have on the system they are configuring on, then all they care about it is "give me the best reasonable performance on the machine I am using today". This means the system software should be able to provide in a trivial way what the current hardware is. The problem is not unique to GPUs, of course, it is not always easy in a portable way to get this information for generic CPUs either.


  Barry






On Apr 5, 2021, at 7:32 PM, Jeff Hammond <jeff.science@gmail.com> wrote:

NVCC has supported multi-versioned "fat" binaries since I worked for Argonne.  Libraries should figure out what the oldest hardware they are about is and then compile for everything from that point forward.  Kepler (3.5) is oldest version any reasonable person should be thinking about at this point.  The oldest thing I know of in the DOE HPC fleet is Pascal (6.x).  Volta and Turing are 7.x and Ampere is 8.x.

The biggest architectural changes came with unified memory (https://developer.nvidia.com/blog/unified-memory-in-cuda-6/) and cooperative (https://developer.nvidia.com/blog/cooperative-groups/ in CUDA 9) but Kokkos doesn't use the latter.  Both features can be used on quite old GPU architectures, although the performance is better on newer ones.

I haven't dug into what Kokkos and PETSc are doing but the direct use of this stuff in CUDA is well-documented, certainly as well as the CPU switches for x86 binaries in the Intel compiler are.


Devices with the same major revision number are of the same core architecture. The major revision number is 8 for devices based on the NVIDIA Ampere GPU architecture, 7 for devices based on the Volta architecture, 6 for devices based on the Pascal architecture, 5 for devices based on the Maxwell architecture, 3 for devices based on the Kepler architecture, 2 for devices based on the Fermi architecture, and 1 for devices based on the Tesla architecture.


Programmatic querying can be done with the following (https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__DEVICE.html):

cudaDeviceGetAttribute
The compiler help tells me this, which can be cross-referenced with CUDA documentation above.

$ /usr/local/cuda-10.0/bin/nvcc -h

Usage  : nvcc [options] <inputfile>

...

Options for steering GPU code generation.
=========================================

--gpu-architecture <arch>                  (-arch)                         
        Specify the name of the class of NVIDIA 'virtual' GPU architecture for which
        the CUDA input files must be compiled.
        With the exception as described for the shorthand below, the architecture
        specified with this option must be a 'virtual' architecture (such as compute_50).
        Normally, this option alone does not trigger assembly of the generated PTX
        for a 'real' architecture (that is the role of nvcc option '--gpu-code',
        see below); rather, its purpose is to control preprocessing and compilation
        of the input to PTX.
        For convenience, in case of simple nvcc compilations, the following shorthand
        is supported.  If no value for option '--gpu-code' is specified, then the
        value of this option defaults to the value of '--gpu-architecture'.  In this
        situation, as only exception to the description above, the value specified
        for '--gpu-architecture' may be a 'real' architecture (such as a sm_50),
        in which case nvcc uses the specified 'real' architecture and its closest
        'virtual' architecture as effective architecture values.  For example, 'nvcc
        --gpu-architecture=sm_50' is equivalent to 'nvcc --gpu-architecture=compute_50
        --gpu-code=sm_50,compute_50'.
        Allowed values for this option:  'compute_30','compute_32','compute_35',
        'compute_37','compute_50','compute_52','compute_53','compute_60','compute_61',
        'compute_62','compute_70','compute_72','compute_75','sm_30','sm_32','sm_35',
        'sm_37','sm_50','sm_52','sm_53','sm_60','sm_61','sm_62','sm_70','sm_72',
        'sm_75'.

--gpu-code <code>,...                      (-code)                         
        Specify the name of the NVIDIA GPU to assemble and optimize PTX for.
        nvcc embeds a compiled code image in the resulting executable for each specified
        <code> architecture, which is a true binary load image for each 'real' architecture
        (such as sm_50), and PTX code for the 'virtual' architecture (such as compute_50).
        During runtime, such embedded PTX code is dynamically compiled by the CUDA
        runtime system if no binary load image is found for the 'current' GPU.
        Architectures specified for options '--gpu-architecture' and '--gpu-code'
        may be 'virtual' as well as 'real', but the <code> architectures must be
        compatible with the <arch> architecture.  When the '--gpu-code' option is
        used, the value for the '--gpu-architecture' option must be a 'virtual' PTX
        architecture.
        For instance, '--gpu-architecture=compute_35' is not compatible with '--gpu-code=sm_30',
        because the earlier compilation stages will assume the availability of 'compute_35'
        features that are not present on 'sm_30'.
        Allowed values for this option:  'compute_30','compute_32','compute_35',
        'compute_37','compute_50','compute_52','compute_53','compute_60','compute_61',
        'compute_62','compute_70','compute_72','compute_75','sm_30','sm_32','sm_35',
        'sm_37','sm_50','sm_52','sm_53','sm_60','sm_61','sm_62','sm_70','sm_72',
        'sm_75'.

--generate-code <specification>,...        (-gencode)                      
        This option provides a generalization of the '--gpu-architecture=<arch> --gpu-code=<code>,
        ...' option combination for specifying nvcc behavior with respect to code
        generation.  Where use of the previous options generates code for different
        'real' architectures with the PTX for the same 'virtual' architecture, option
        '--generate-code' allows multiple PTX generations for different 'virtual'
        architectures.  In fact, '--gpu-architecture=<arch> --gpu-code=<code>,
        ...' is equivalent to '--generate-code arch=<arch>,code=<code>,...'.
        '--generate-code' options may be repeated for different virtual architectures.
        Allowed keywords for this option:  'arch','code'.

On Mon, Apr 5, 2021 at 1:19 PM Satish Balay via petsc-dev <petsc-dev@mcs.anl.gov> wrote:
This is nvidia mess-up. Why isn't there a command that give me these values [if they insist on this interface for nvcc]

I see Barry want configure to do something here - but whatever we do - we would be shifting the problem around.
[even if we detect stuff - build box might not have the GPU used for runs.]

We have --with-cuda-arch - which I tried to remove from configure - but its come back in a different form (--with-cuda-gencodearch)

And I see other packages:

  --with-kokkos-cuda-arch

Wrt spack - I'm having to do:

spack install xsdk+cuda ^magma cuda_arch=60

[magma uses CudaPackage() infrastructure in spack]

Satish

On Mon, 5 Apr 2021, Mills, Richard Tran via petsc-dev wrote:

> You raise a good point, Barry. I've been completely mystified by what some of these names even mean. What does "PASCAL60" vs. "PASCAL61" even mean? Do you know of where this is even documented? I can't really find anything about it in the Kokkos documentation. The only thing I can really find is an issue or two about "hey, shouldn't our CMake stuff figure this out automatically" and then some posts about why it can't really do that. Not encouraging.
>
> --Richard
>
> On 4/3/21 8:42 PM, Barry Smith wrote:
>
>
>   It would be very nice to NOT require PETSc users to provide this flag, how the heck will they know what it should be when we cannot automate it ourselves?
>
>   Any ideas of how this can be determined based on the current system? NVIDIA does not help since these "advertising" names don't seem to trivially map to information you can get from a particular GPU when you logged into it. For example nvidia-smi doesn't use these names directly. Is there some mapping from nvidia-smi  to these names we could use? If we are serious about having a non-trivial number of users utilizing GPUs, which we need to be for future, we cannot have this absurd demands in our installation process.
>
>   Barry
>
> Does spack have some magic for this we could use?
>
>
>
>



--
Jeff Hammond
jeff.science@gmail.com
http://jeffhammond.github.io/