Revision as of 14:59, 22 August 2014

Computational species "bw filter simplified less" (CID=45741e3fbcf4024b:1db78910464c9d05)

Notes

Some cost-aware experiments (execution time, size, energy, compilation time) performed by Grigori Fursin using Collective Mind framework and artifacts from thepublic repository to be reproducible. It supports our research on continuous performance tracking, code optimization and compiler benchmarking (regression detection).

This computation species (kernel) is a threshold filter - it is used in image processing and neuron activation functions (part of artificial neural networks).

Used artifacts

• Datasets:
  • D1 = grayscale image 1 (CID=8a7141c59cd335f5:c8848a1b1fb1775e), size=1536x1536~2.4E6 (pixels or neurons)
  • D2 = grayscale image 2 (CID=8a7141c59cd335f5:0045c9b59e84318b), size=1536x1536~2.4E6 (pixels or neurons)

• Systems:
  • S1 = Dell Laptop Latitude E6320, Processor=P1, Memory = 8Gb, Storage=256Gb (SSD), Max power consumption=52W, Cost (year of purchase 2011)~1200 euros (CID=cb7e6b406491a11c:0d84339816de0271)
  • S2 = Samsung Mobile Galaxy Duos GT-S6312, Processor=P2, Memory = 0.8Gb, Storage=4Gb, Battery=1300 mAh / 3.9V / up to 250 hours, Max power consumption~5W, Cost (year of purchase 2013)~200 euros (CID=cb7e6b406491a11c:a9740acbe06bcd1e)
  • S3 = Polaroid Tablet Executive 9" MID0927, Processor=P3, Memory=1Gb, Storage=16Gb, Battery=3500 mAh / 3.9V / up to 80 hours, Max power consumption~13W, Cost (year of purchase 2014)~100 euros (CID=cb7e6b406491a11c:3419444faf22f3d0)
  • S4 = Semiconductor neural network, PSpice simulation (year of development = 1993-1997)

• Processors:
  • P1 = Intel Core i5-2540M, 2.60GHz, 2 cores (CID=54cd38490124ef51:425ae4e3483c82e8)
  • P2 = Qualcomm MSM7625A FFA, ARM Cortex A5, ARMv7, 1 GHz, 1 core (CID=54cd38490124ef51:ae17889f40209ae7)
  • P3 = Allwinner A20 (sun7i), Dual-Core ARM Cortex A7, ARMv7, 1.6GHz, Mali400 GPU, 2 core (CID=54cd38490124ef51:fc020ce2e4d44f3d)
  • P4 = NVidia Quadro NVS 135M, 16 cores, 400MHz (TBD)
• Processor mode:
  • W1 = 32 bit
  • W2 = 64 bit

• OSs:
  • O1 = Windows 7 Pro SP1,  cost~170 euros (CID=c4d3ce728f46eea2:10c4f7484446b689)
  • O2 = O1, MinGW32
  • O2 = OpenSuse 12.1, Kernel 3.1.10, cost=free (CID=c4d3ce728f46eea2:29ce89f1a1446e89)
  • O3 = Android 4.1.2, Kernel 3.4.0, cost=free (CID=c4d3ce728f46eea2:e734c48d5a5824c1)
  • O4 = Android 4.2.2, Kernel 3.3.0, cost=free (CID=c4d3ce728f46eea2:d3e9b97f6994444b)

• Compilers:
  • X1 = GCC 4.1.1
  • X2 = GCC 4.4.1
  • X3 = GCC 4.4.4
  • X4 = GCC 4.6.3
  • X5 = GCC 4.7.2
  • X6 = GCC 4.8.3
  • X7 = GCC 4.9.1
  • X8 = LLVM 3.1
  • X9 = LLVM 3.3.2
  • X10 = Open64 5.0
  • X11 = PathScale 2.3.1
  • X12 = NVidia CUDA Toolkit 5.0
  • X13 = Microsoft Visual Studio 2013, cost = has free minimal version
  • X14 = Intel Composer XE 2011, cost =

• Compiler optimization level:
  • O1 = Performance (usually -O3)
  • O2 = Size (usually -Os)

• Number of run-time code repetitions (for example, processing steps in neural networks):
  • R1 = 4000
  • R2 = 1000
  • R3 = 400

• Total number of computations (processed neurons or pixels)
  • T1 ~ 9.6E9
  • T2 ~ 2.4E9
  • T3 ~ 1.0E9

• Costs
  • C1= Execution time
  • C2 = Energy
  • C3 = Code size
  • C4 = Compilation time
  • C5 = System size
  • C6 = Hardware price
  • C7 = Software price
  • C8 = (Auto-)tuning price
  • C9 = Development time
  • C10 = Validation and testing time

Notes

Energy: 1Wh = 3600 joules

W = mAh * V / 1000 = 1300 * 3.9 / 1000 ~ 5W