Re: [petsc-users] quick question on MatXAIJSetPreallocation

[Jed Brown via petsc-users]

Xiangdong via petsc-users  writes:

> Hello everyone,
>
> It seems that to use MatXAIJSetPreallocation, one has to pass the array of
> the number of nonzero blocks per row, even if this number is same across
> all the local rows.
>
> For the other preallocation functions (say, MatMPIBAIJSetPreallocation),
> they do support both a constant and a array for the number of nonzeros.
>
> I am curious that the reason MatAIJSetPreallocation does not support the
> constant number of nonzeros case.

It was in the interest of limiting the number of arguments and
recognizing that it's easy to call all relevant preallocation functions
if you have constant row lengths, but somewhat involved to convert if
you don't.

Re: [petsc-users] BAIJCUSPARSE?

[Xiangdong via petsc-users]

Thanks for pointing the function MatXAIJSetPreallocation() out. I followed
your suggestions and tried it on my problem:
preallocate the matrix with XAIJ, set its values by
MatSetValuesBlockedLocal, and use this matrix with aijcusparse.

It works!

Thank you.

Xiangdong



On Tue, Oct 29, 2019 at 8:10 PM Mills, Richard Tran  wrote:

> We will let you know when this is ready, Xiangdong.
>
> Let me address a part of your original question that I don't think anyone
> else noticed:
>
> In my current code, the Jacobian matrix preallocated and assembled as BAIJ
> format. Do I have to rewrite this part of code to preallocate and assemble
> the matrix as AIJ in order to use aijcusparse?
>
>
> If you are doing your preallocation via MatXAIJSetPreallocation() and
> setting values via MatSetValuesBlocked() (or its variants), you can change
> the type to AIJ (or AIJCUSPARSE) instead of BAIJ and things should just
> work. (If not, let us know, as this may mean a bug in PETSc.) If you are
> calling MatSetFromOptions(), you should be able to do this on the command
> line (or otherwise through the PETSc options database) -- no code rewrites
> needed.
>
> Best regards,
> Richard
>
>
> On 10/28/19 7:40 AM, Xiangdong wrote:
>
> Thanks for your information. Glad to hear that BAIJ gpu support is on the
> way. Waiting a few weeks is not a issue at all. Once you finish the BAIJ
> GPU interface, could you please make an announcement here or in the change
> log?
>
> Thank you.
>
> Xiangdong
>
> On Sat, Oct 26, 2019 at 12:11 AM Mills, Richard Tran 
> wrote:
>
>> Xiangdong,
>>
>> cuSPARSE does support block compressed sparse row (BAIJ) format, but we
>> don't currently support that cuSPARSE functionality in PETSc. It should be
>> easy to add, but we are currently refactoring the way we interface with
>> third party GPU libraries such as cuSPARSE, and it would probably make more
>> sense to add this support after that refactor is done. Do you need this
>> right away, or could it wait maybe a few weeks until this is completed?
>>
>> Best regards,
>> Richard
>>
>> On Fri, Oct 25, 2019 at 1:50 PM Smith, Barry F. via petsc-users <
>> petsc-users@mcs.anl.gov> wrote:
>>
>>>
>>>   You would need to investigate if the Nvidia cuSPARSE package supports
>>> such a format. If it does then it would be reasonably straightforward for
>>> you to hook up the required interface from PETSc. If it does not then it is
>>> a massive job to provide such code and you should see if any open source
>>> packages provide such CUDA support and then you could hook PETSc up to use
>>> that.
>>>
>>>   Barry
>>>
>>>
>>> > On Oct 25, 2019, at 3:43 PM, Xiangdong via petsc-users <
>>> petsc-users@mcs.anl.gov> wrote:
>>> >
>>> > Can anyone comment on the PETSc's GPU version of Block CSR, say
>>> BAIJCUSPARSE? Does it make sense to have such format on GPU? Is it under
>>> development?
>>> >
>>> > Thank you.
>>> >
>>> > Xiangdong
>>> >
>>> > On Wed, Oct 23, 2019 at 11:36 AM Xiangdong  wrote:
>>> > Hello everyone,
>>> >
>>> > I am wondering whether there is a format BAIJCUSPARSE for Block CSR on
>>> GPU.
>>> >
>>> > In my current code, the Jacobian matrix preallocated and assembled as
>>> BAIJ format. Do I have to rewrite this part of code to preallocate and
>>> assemble the matrix as AIJ in order to use aijcusparse?
>>> >
>>> > Thank you.
>>> >
>>> > Xiangdong
>>> >
>>>
>>>
>