Collective Knowledge Aggregator proof-of-concept
Crowd results Raw CK browser Graphs Reports Datasets Models Home

This page is outdated! New version is available here.

Distinct solutions after online classification (auto/crowd-tune GCC compiler flags (minimize execution time))

Scenario UID8289e0cf24346aa7 (experiment.tune.compiler.flags.gcc.e)
Data UID27521a143a596ebc
Discuss (optimizations to improve compilers,
semantic/data set/hardware features
to improve predictions
, etc):
GitHub wiki, Google group
Download:[ All solutions in JSON ], [ Solutions' classification in JSON ]
Reproduce all (with reactions):ck replay 8289e0cf24346aa7:27521a143a596ebc
CompilerGCC 4.9
CPUQualcomm Technologies, Inc MSM8994
Improvement key IK1Main kernel execution time speedup [min]
Improvement key IK2Code size improvement

Improvements (<4% variation) Distinct workload for highest improvement
# Solution UID IK1 IK2 New distinct optimization choices Ref Best species Worst species Touched Iters Program CMD Dataset Dataset file CPU freq (MHz) Cores Platform OS Replay
S1 81f91b0d913d68e0 2.87 0.99 -O3 -fno-branch-target-load-optimize -fconserve-stack -fcprop-registers -fno-cx-limited-range -fno-data-sections -fdevirtualize-speculatively -fno-gcse -findirect-inlining -fno-ipa-pta -fno-ipa-pure-const -fno-ipa-reference -fno-ira-loop-pressure -fmove-loop-invariants -fpeephole -fpeephole2 -fno-omit-frame-pointer -fno-prefetch-loop-arrays -freorder-blocks -fno-rounding-math -fno-sched-critical-path-heuristic -fno-sched-last-insn-heuristic -fno-selective-scheduling -fsel-sched-pipelining-outer-loops -fsingle-precision-constant -fno-split-wide-types -ftree-ch -ftree-loop-if-convert-stores -ftree-loop-ivcanon -fno-tree-switch-conversion -fno-tree-vrp -fno-unsafe-loop-optimizations -fuse-linker-plugin -falign-functions=0 -falign-jumps=0 --param predictable-branch-outcome=41 --param min-inline-recursive-probability=3 --param comdat-sharing-probability=27 --param max-unroll-times=0 --param max-completely-peel-loop-nest-depth=6 --param hot-bb-count-ws-permille=206 --param unlikely-bb-count-fraction=3254 --param align-threshold=5 --param tracer-max-code-growth=24 --param max-goto-duplication-insns=7 --param max-cse-path-length=10 --param omega-max-eqs=246 --param max-reload-search-insns=83 --param sched-spec-prob-cutoff=26 --param selsched-max-sched-times=4 --param max-last-value-rtl=2420 --param min-size-for-stack-sharing=2 --param max-fields-for-field-sensitive=0 --param prefetch-latency=352 --param l1-cache-size=113 --param sccvn-max-scc-size=12832 --param ira-max-loops-num=149 --param ira-loop-reserved-regs=0 --param asan-instrument-writes=1 --param asan-use-after-return=1 -O3 1 0 2 1 milepost-codelet-mibench-automotive-susan-s-src-susan-codelet-1-1 default audio-wav-0001 data.wav 1 LETV LE 1 PRO Android 5.0.2

[ Participated users, platforms, OS, CPU, GPU, GPGPU, NN, NPU ] [ How to participate ] [ Motivation (PPT) (PDF) ] [ Papers 1 , 2 , 3] [ Android app ] [ Collective training set ] [ Unified AI ]
View entry in raw format

Developed by Grigori Fursin           
Implemented as a CK workflow
                      Hosted at