systemml-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Niketan Pansare" <npan...@us.ibm.com>
Subject Re: [DISCUSS] Blocksparse kernels
Date Thu, 10 May 2018 18:51:12 GMT

Hi Janardhan,

>> 1. Can you help me, estimate how much would it take to implement
blocksparse kernels practically.
This is a difficult question to answer as it depends on how comfortable you
are with writing and optimizing sparse kernels. To implement block-sparse
kernel as per your document, one needs to know (each step is progressively
more difficult than the previous step):
1. How to implement and compile a simple CUDA kernel.
2. How to implement a non-block sparse kernel such that results match with
the CuSPARSE code.
3. How to optimize a non-block sparse kernel such that performance match
with the CuSPARSE code.
4. How to optimize a non-block sparse kernel for a given hardware such that
performance match with the CuSPARSE code on that hardware. This requires
working knowledge of different Nvidia devices and how to tweak sass code.
5. How to implement a block sparse kernel such that results match with the
CuSPARSE code.
6. How to optimize a block sparse kernel such that performance match with
the CuSPARSE code.
7. How to optimize a block sparse kernel for a given hardware such that
performance match with the CuSPARSE code on that hardware.

I would recommend before attempting a block-sparse kernel:
- Picking up JIRAs that will help you through steps (1)-(3). May be, you
want to try SYSTEMML-937 and SYSTEMML-2312/SYSTEMML-2313 first.
- You can skip (4) and (7) as it involves maintenance overhead.

For (2)-(3), you can either use SystemML or the CUDA code I sent in the
earlier thread as baseline.

>> 2. Would like to spare some time to review PRs ( ~2 PRs per week).
Sure. You may want to batch similar PRs as it reduces turn-around time.

Thanks,

Niketan Pansare
IBM Almaden Research Center
E-mail: npansar At us.ibm.com
http://researcher.watson.ibm.com/researcher/view.php?person=us-npansar



From:	Janardhan <janardhan@apache.org>
To:	dev@systemml.apache.org, Niketan Pansare <npansar@us.ibm.com>,
            reinwald@us.ibm.com, Nakul Jindal <nakul02@gmail.com>
Date:	05/10/2018 10:34 AM
Subject:	[DISCUSS] Blocksparse kernels



Hi Niketan, Nakul, and Berthold,

1. Can you help me, estimate how much would it take to implement
blocksparse kernels practically.

2. Would like to spare some time to review PRs ( ~2 PRs per week).

a. Relevant Jira: https://issues.apache.org/jira/browse/SYSTEMML-2041

b. My proposal:

https://docs.google.com/document/d/1cgPdyhhG3kQZxeP1VYOnQoZuTVA2216CC_aWdurmTNw/edit?usp=sharing

c. Research paper:
https://s3-us-west-2.amazonaws.com/openai-assets/blocksparse/blocksparsepaper.pdf

Thank you,
Janardhan


Mime
  • Unnamed multipart/related (inline, None, 0 bytes)
View raw message