View source for Meeting 19 August 2014, Embecosm

<center>'''MAGEEC Meeting 19/08/2014 -Lymington'''</center>

<center>Present: OR, JP, GC, SG, GF, CB, SC, AB, SH</center>


'''Review of action points from last time'''


Blog posts

* '''[ABOpen] '''Oliver's ready to go.
* [JP] Blog post on how not-to measure

SC Update

* '''[SC]''' contact David Malcom at Redhat re abstracting GCC plugin interface to make consistant across releases.
* Profiling of time spent per function now available. Large variation, some large, some too small to measure. Suggest arithmetic power allocation of shorter spans pro-rata for shorter blocks.
** ? How long does it take to train with this granularity, since simulation takes longer than running.

Calibration

* Physical location variation after taking into account placement ~3%
* Need a calibration step with extra start and stop trigger with a known calibration routine. Compiled in with the main benchmark to save flashing time
** ? Can it be at the end of memory so that it does not influence the benchmarked code. We agreed that this is probably not a big issue, but we should keep an eye on not influincing measurement via calibration.
** '''[JP]''' To commit an appropriate calibration benchmark to BEEBS.

'''[JB] '''to chase Joern to see what progress is on flash alignment optimisation.

* [JB] Joern working on it. Progress may be behind schedule. JB to confirm whether deadline of 29<sup>th</sup> will be hit.

'''''[SH+KIE] by next meeting''''' structure of monster MAGEEC journal paper with all authors, describing process of developing a system with machine learning, plus the evaluation. Adds in the software engineering + evaluation. Not a push on the novelty of the machine learning.


'''[JB]''' Update events page

'''[SH]''' Add CASES + Craig's ILP (using BEEBS) to page


Boards:

'''[ABOpen]''' GBP50 target for board, base and cables. ABOpen creating publicity this week for end of September orders.

George:


* Comparison and sampling algorithms – not started.
* Minimal analysis started. Script error → not much data yet.
** Substantiates <3% board/run variation test-to-test bound
** Identifies the three {most significant helping and harming} flags.
** Moving forward: complete analysis runs. Remove outliers by <100ms threshold. Manual review on <1s run time. 
*** Then comparison algorithms

[SC&GF] Some pass combinations give weird results. To examine and see if it's a bug.


Stoil


* GIMPL Tree analysis of instruction distribution of benchmarks (BEEBS V1)
** Counting the different statements.
** TODO: Graph the results
*** What grouping
**** Existing ones + sub-division of ALU to horizonal and vertical (JP to advise)
** We can connect the basic block GIMPL analysis to the feature vector.
*** An extension is to consider a histogram of the number of GIMPL statements (block size approximation) per basic block 
** '''[RQ]''' We could compare the machine code allocations both before and after MAGEEC to see e.g. if MAGEEC learns to substibute memory accesses with ALU ops.

* Weather station case study up and running, with remote data send and receive.
** '''[SG]''' Need to deploy MAGEEC to see the improvement on the weather station.
*** Iterative compilation on weather station hardware.
*** Compare with MAGEEC's output with MAGEEC trained on other platforms.
** [JP] Energytool interactive improvement by review progress.
* Satellite case study needs compilation in of MSP430 support.
* Battery + RTOS on STMDiscovery

'''Greg'''


* PCA results including Energy measurement from -O0
** Seems that there are a small number of features that influence energy a lot.
** Energy itself if the best indicator of a benchmark.
** Features need normalising to prevent over-easy correlation.
** The results set needs more detailed exploration.

* Automatic program generation work
** Ideas, but not going to be able to complete in the remaining time

* Going forward
** How to predict energy in terms of the features we have already captured.
*** Could use WECA to explore how to use supervised learning to predict energy from the features already 

'''BEEBS'''


* Some cross-compilation problems on -Wall for ~11 benchmarks
* '''[SC] '''Add noinline_benchmarks on BEEBS
* '''[AB]''' New super-duper header mechanism to prevent duplicated code bases from diverging and still allow for build mechanisms to work.
* Ignored benchmarks infrastructure to be on a per-test basis
* BEEBS V2.0 ready for release
* Approx 10 have queries on release under GPL. Need to clarify license or find replacements ?resourcing?

[[Category:Meetings]]