reviews.txt

----------------------- REVIEW 1 ---------------------
PAPER: 17
TITLE: Autotuning OpenCL Workgroup Sizes
AUTHORS: Chris Cummins


----------- Review -----------
The paper describes the usage of a autotuning framework named Omnitune
to help predict the appropriate working group size for opencl skeleton
code.

The novelty of the work is unclear. There have been many autotuning
papers and some of them have already covered the prediction of working
group size. The paper did not make it clear what unique strengths this
work has compared to prior studies. The paper briefly mentioned the
sharing of tuning results across systems, which could involve some
interesting elements. But the paper fails to expand on it. It instead
puts focus on the group size autotuning results.


----------------------- REVIEW 2 ---------------------
PAPER: 17
TITLE: Autotuning OpenCL Workgroup Sizes
AUTHORS: Chris Cummins


----------- Review -----------
This work describes the use of ML techniques to guide the search-space
when auto-tuning OpenCL workgroup sizes. This is interesting work, but
has been presented before in:

Autotuning OpenCL Workgroup Size for Stencil Patterns. / Cummins,
Chris; Petoumenos, Pavlos; Steuwer, Michel; Leather, Hugh.
International Workshop on Adaptive Self-tuning Computing Systems
(ADAPT) 2016 @ HiPEAC, Prague, Czech Republic, January 18, 2016. 2016.


----------------------- REVIEW 3 ---------------------
PAPER: 17
TITLE: Autotuning OpenCL Workgroup Sizes
AUTHORS: Chris Cummins


----------- Review -----------
This poster abstract proposes incorporating autotuning of workgroup
sizes into OpenCL code derived from algorithmic skeletons, implemented
for the stencil skeleton in SkelCL.  A linear regression approach is
used to construct a predictor for workgroup sizes, based on 102
features extracted at runtime.  Significant speedups are achieved over
non-tuned fixed workgroup sizes.

Related work: If I remember properly, the earliest use of autotuning
of workgroup sizes in a skeleton programming framework (supporting
OpenCL but also other target platforms) was done by Dastgeer:

@inproceedings{Dastgeer:2011:ASM:1984693.1984697,
 author = {Dastgeer, Usman and Enmyren, Johan and Kessler, Christoph W.},
 title = {Auto-tuning SkePU: A Multi-backend Skeleton Programming Framework for multi-GPU Systems},
 booktitle = {Proceedings of the 4th International Workshop on Multicore Software Engineering},
 series = {IWMSE '11},
 year = {2011},
 isbn = {978-1-4503-0577-8},
 location = {Waikiki, Honolulu, HI, USA},
 pages = {25--32},
 numpages = {8},
 url = {http://doi.acm.org/10.1145/1984693.1984697},
 doi = {10.1145/1984693.1984697},
 acmid = {1984697},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {auto-tuning, cuda, data parallelism, gpu, opencl, skeleton programming},
}

Minor issues:

- with an an -> with an

- level of the can -> level can

- Figures: the blue color is too dark, and the axis inscriptions are
  unreadably small.


----------------------- REVIEW 4 ---------------------
PAPER: 17
TITLE: Autotuning OpenCL Workgroup Sizes
AUTHORS: Chris Cummins


----------- Review -----------
This paper proposes using autotuning techniques to optimize the
workgroup size of OpenCL codes that are the stencil skeletons of
SkelCL ("A Portable Skeleton Library for High-Level GPU
Programming"). Autotuning data is accumulated across a network of
cooperating systems and machine learning is used to tune the workgroup
size parameter.

This is an important practical problem, and the results presented are
convincing. However, I am surprised that a single parameter (workgroup
size) is presented as the main target for autotuning. What about using
OpenCL async_prefetch, async_copy and working in local memory, as
advised for CPU-based and DSP-based platforms (eg TI KeyStone II
OpenCL optimization manual)? In my experience of optimizing stencils
on another multicore platform, this is where large performance gains
can be found, rather than workgroup size.

The figures are so reduced it is impossible to read text on the axis.