CPredict

From cTuning.org

You are warmly welcome to quickly self-register at this website and our community mailing lists (developments, announcements) to join our collaborative effort, follow and update cTuning news, update wiki pages, add and rank optimization cases, use web-services and participate in developments!

All developments moved to cKnowledge.org in 2015!


Machine learning based optimization prediction using program similarities based on static and dynamic program features using continously collected knowledge in the cDatabase as described in [FKMP2011].If you have any questions or comments, don't hesitate to contact Grigori Fursin or cTuning discussions mailing list.

cTuning Web-service is now available to predict multi-objective optimizations (execution time vs code size vs compilation time) on the fly during compilation. It is now connected with MILEPOST GCC (with ICI plugin system and feature extractor) and CCC framework to optimize programs for better execution time and code size on the fly during compilation. This is a part of our statistical collective optimization concept to convert compilers into modular, self-tuning intelligent adaptive infrastructure. It allows substitution of the compiler default optimization heuristic with the external plugin that communicates with our machine learning based online predictor and improves optimization predictions continuosly for a given architecture, program and a dataset leveraging experience of multiple users.

Sample program:
/* Sample C program to demonstrate optimization prediction */ #include<stdio.h> int main(void) { int i; float x[100]; for (i=0; i<10; i++) x[i]=i*1.2; printf("x[5]=%f", x[5]); return 0; }

Compiler with feature extractor:
MILEPOST GCC 4.4.0 ICI 2.0 ML 2.0\n

Static program features (Description of current program features from the MILEPOST GCC):
Dynamic program features (hardware counters):
On-going work

Predict optimizations:

Find optimization cases (that improve execution time, code size, power, etc) from the Collective Optimization Database for programs with similar static features.
This is an on-going work - we continue working to improve prediction mechanisms (automatic feature generation, quality and importance of features, etc).

Note: Prune the search until you find your configuration:

Database:	fursinne_coptcases fursinne_ccomparisons fursinne_coptcasestest fursinne_coptcases - main database with optimization cases fursinne_ccomparisons - database with comparison of different compilers (GCC, LLVM, Open64, Intel, etc) fursinne_coptcasestest - database for testing cTools and cDatabase fursinne_cfeatures - (WILL BE ADDED SOON) database with static and dynamic program features to enable optimization prediction based on machine learning
Architectures:	AMD Athlon 64 3700+ AMD Athlon XP 2800+ AMD Opteron 2218 Core2Duo T8300 Intel Core 2 Duo T9300 Intel Xeon 2800MHz Intel Xeon EM64T Itanium 2 1300MHz
Environments:	Debian Sid Linux v1.1 GNU/Linux 2.6.28-16-generic Ubuntu SMP Linux Debian 2.6.10-1-mckinley Linux Mandriva 2.6.17-10alchemy Linux Mandriva 2.6.22-18-desktop-1mdv Linux Mandriva 2.6.22-9-desktop-1mdv Ubuntu 9.10 Ubuntu GNU Linux 9.10 SMP
Compilers:	GCC4CIL 4.2 Intel 11.0 LLVM 2.5 Open64 4.2.1 PathScale 3.2 gcc 4.2.2-ICI-0.9.6 gcc 4.4.0 gcc 4.4.0 ICI 2.0 milepost gcc 4.4.0-ICI-2.0

Development history

---

Support and development:

Retrieved from "https://ctuning.org/wiki/index.php/Special:CPredict"