John Cavazos

1
Models to Predict Good Compiler Optimizations

• John Cavazos
• Dept of Computer Information Sciences
• University of Delaware

2
Whole-Program Autotuning
STEPS

• Characterize each function
• Prediction models ranks optimization seqs
• Apply top sequences

3
Train and Test Model

• Train on kernels
• Generate training data (inputs, outputs)
• Automatically construct a model
• Can be expensive, but can be done offline
• Supervised learning problem
• Test on whole programs
• Extract features (code characteristics)
• Model predicts optimizations to apply

4
Constructing Models

• Trained on data from Random Search
• 200 evaluations for each benchmark
• Leave-one-out cross validation
• Regression and Support Vector Machines

5
Are models predictive?
6
Solution Overview
Static Program Features Source code or
IR Dynamic Program Features Running time
7
Solution Overview
Static Program Features Source code or
IR Dynamic Program Features Running time
Static Program Features From source code or
IR Dynamic Program Features From runtime

• Sequence Predictor
• Speedup Predictor
• Tournament Predictor

8
Characterization of 181.mcf
9
Characterization of 181.mcf
Problem Greater number of memory accesses per
instruction than average
10
Sequence Predictor
11
Speedup Predictor
12
Tournament Predictor
13
Open64 Optimizations

• Control 63 Open64 optimizations
• Loop optimizations
• Unrolling
• Interchange
• Fussion / Fission
• Prefetching
• Traditional optimizations
• PRE, copy prop, strength reduction, CSE, etc.

14
Experimental Setup

• HPCToolkit / PAPI 3.6
• Intel Quad _at_2GHz with 8GB RAM
• Open64 Compiler version 4.2
• Baseline -Ofast
• 200 randomly generated sequences
• Benchmarks
• 25 hot functions in kernels from Linpack, NAS,
  and Polybench
• 16 programs from MiBench

15
Experimental Setup

• 3 Prediction Models
• Sequence Predictor
• Speedup Predictor
• Tournament Predictor
• Machine Learning Algorithms
• Regression
• Support Vector Machine (SVM)

16
Regression (10 evals)
AVG Sequence 9, Tournament 21, Speedup 25
17
SVM (10 evals)
AVG Sequence 9, Tournament 15, Speedup 23,
18
Future work

• Apply to whole applications
• Applying to different compilers and architectures
• Compare different characterization methods
• Source code vs dynamic characterization

19
Across Different Machines

• Machine 1
• 4 of Intel Core2 Quad CPU Q9650 _at_ 3.00GHz
• RAM 8 GB
• Cache size 6144 KB
• Machine 2
• 4 of Intel Core2 Quad CPU Q9300 _at_ 2.50GHz
• RAM 4 GB
• Cache size 3072 KB
• Machine 3
• 4 of Intel Xeon CPU E5335 _at_ 2.00GHz
• RAM 2 GB
• Cache size 4096 KB

20
MiBench Across Machines
21
Case Study PoCC

• Experimental setup
• Intel Xeon E5620 _at_2.4 GHz
• 16 hardware threads
• Baseline ICC fast
• 768 randomly generated sequences
• PoCC (Polyhedral Compiler Collection)
• Unrolling, Tiling, Loop fuse, Auto
  parallelization
• Polybench (28 kernels)
• Speedup Predictor (Regression / SVM)

22
PoCC (10 evals)
18.26
20.67
14.44
AVERAGE Random 3.1X, SVM 5.8X, LR 5.9X, Best
6.1X
23
Model Evaluation Summary

• Comparison of 3 prediction models with 2 machine
  learning algorithms on kernels
• Newly proposed and evaluated models (speedup
  predictor and tournament predictor) outperformed
  state-of-the-art predictor
• Applying speedup predictor trained with kernels
  to MiBench
• For seen sequences 5.4 for regression, 4.6 for
  SVM
• For unseen sequences 5.1 for regression, 2.1
  for SVM

Sequence Predictor
Speedup Predictor
Tournament Predictor
Regression
8.7
25.0
20.7
SVM
8.8
22.5
14.6
24
Regression/SVM - MiBench (Speedup Predictor / 10
evals)

• Regression seen sequences 5.4, unseen
  sequences 5.1
• SVM seen sequences 4.6, unseen sequences 2.2

25
Dynamic Program Features
26
Optimizations
Phase
List of Optimizations
OPT
align-padding, ptr-opt, swp, unroll-size
WOPT
aggcm, aggstr-reduction, const-pre,
copy-propagate, bdce, dce-aggressive,
dce-global, hoisting,iv-elimination, spre,
value-numbering, dse-aggressive,unroll,
canon-expr, aggcm-threshold, combine, intrinsic,
mem-opnds, fold2const
LNO
optimize-cache, lego, prefetch-stores,prefetch-ahe
ad, interchange, pure, fusion, hoistif,
blocking-size, ecspct,fission, fission-inner-regis
ter-limit, full-unroll, full-unroll-size,
fusion-peeling-limit, outer-unroll-max,
sclrze, outer-unroll-further, max-depth,
outer-unroll-prod-max, shackle, svr, cse,
preferred-doacross-tile size, prefetch-cache-fact
or, vintr, split-tile, olf-ub, unswitch,
lego-local, call-info, prefetch,
apply-illegal-xform-directives
CG
unroll-fully, gcm, loop-opt
IPA
aggr-cprop, cprop, dce, dve
27
PoCC (1 evals)
AVERAGE- SVM 188.9, LR 256.53
28
Whole-Program Autotuning

• Current Solution
• Outline hot functions
• Tune hot function in isolation
• Integrate tuned hot function back in application
• Disadvantage
• Does not account for code interactions

29
Proposed solution

• Tune several hot functions at same time
• Cannot afford random exploration
• Performance model ranks variants
• Apply only predicted best optimizations