Request access to the cTuning MiDataSets

Requests are reviewed by the maintainers; a few days may pass.

These inputs were collected from public web sources in 2006-2007 for compiler research; their origins were not recorded and the collection's authors hold no rights over them. They are shared for research and education only, to reproduce or extend published research on program optimization: you agree not to redistribute them, to name the cTuning MiDataSets when you use them, and to remove any input on a rights holder's request (the terms in LICENSE.md).

Log in or Sign Up to review the conditions and access this dataset content.

cTuning MiDataSets: the inputs of the cBench benchmarks

The full sets of inputs of the cBench benchmarks (MiBench-based programs of the cTuning project), known as MiDataSets / cDatasets: about 20 inputs for each program (images, audio, text, archives, encrypted files, numbers ...), collected in 2006-2007 for research on iterative and machine-learning-based compiler optimization, where the best optimization of a program changes with its input (G. Fursin, J. Cavazos, M. O'Boyle, O. Temam. MiDataSets: Creating the Conditions for a More Realistic Evaluation of Iterative Optimization. HiPEAC 2007, LNCS 4367, pp. 245-260). They were kept as Collective Mind (2012) and Collective Knowledge (CK, 2015-2021) dataset entries (the repositories ctuning-datasets and ctuning-datasets-min) and are here so that the benchmarks can be run on all their inputs without the old CK repositories.

Access is on request (the form above) because the origin and the licence of the individual inputs were not recorded when they were collected: see LICENSE.md for the terms, and for how to ask for an input's removal or to tell us its origin.

Layout

One folder per kind of input, one folder per input inside it, the files of the input as the benchmark takes them, and ck-meta.json - the CK meta of the input (its tags, its files in the order a program takes them, the number a "number" input stands for):

adpcm/adpcm-0001/data.adpcm      adpcm/adpcm-0001/ck-meta.json
adpcm/adpcm-0002/data.adpcm      ...
txt/txt-0007/data.txt            txt/txt-0007/data.s.txt
number/number-0003/number.txt

index.json lists every kind and input: the files with size and SHA-256 (file is the path in this repository), the order a program takes them, the number, the tags, the CK UID and date of the input.

Kind Inputs Size Tags Used by (cBench)
adpcm 20 6.8 MB audio, adpcm cbench-telecom-adpcm-d
au 20 27.1 MB audio, au cbench-telecom-gsm
audio-mp3 20 3.1 MB audio, mp3 cbench-consumer-mad
audio-wav 20 32.8 MB audio, wav cbench-consumer-lame
benc 20 11.4 MB crypto, benc cbench-security-blowfish (decode)
bzip2 32 17.9 MB archiver, bzip2 cbench-bzip2 (decode)
cdataset-dijkstra 20 3.6 MB dijkstra cbench-network-dijkstra
cdataset-patricia 20 5.7 MB patricia cbench-network-patricia
cdataset-qsort 20 13.7 MB sorting, qsort cbench-automotive-qsort1
enc 20 11.4 MB crypto, enc cbench-security-blowfish (encode), cbench-security-rijndael (decode)
gsm 20 2.8 MB gsm - (GSM-encoded audio)
image-jpeg 20 3.1 MB image, jpeg, jpg cbench-consumer-jpeg-d
image-pgm 21 9.3 MB image, pgm cbench-automotive-susan
image-pgm-clean-gray-square 8 30.0 MB image, pgm cbench-automotive-susan
image-ppm 20 216.9 MB image, ppm cbench-consumer-jpeg-c
image-tiff 30 36.4 MB image, tiff, tif, orig cbench-consumer-tiff2bw, cbench-consumer-tiff2rgba
image-tiff-bw 30 29.5 MB image, tiff, tif, bw cbench-consumer-tiff2dither, cbench-consumer-tiff2median
image-tiff-nocomp 30 258.2 MB image, tiff, tif, nocomp - (uncompressed TIFF)
number 20 0.0 MB number, cmd cbench-automotive-bitcount
pcm 20 27.1 MB audio, pcm cbench-telecom-adpcm-c, cbench-telecom-crc32
pgp 20 3.9 MB crypto, pgp cbench-security-pgp (decode)
ps 20 5.0 MB doc, ps cbench-bzip2 (encode)
txt 20 11.9 MB doc, txt cbench-office-stringsearch2, cbench-security-sha, cbench-security-rijndael (encode), cbench-security-pgp (encode)
all 491 767 MB

Input 0001 of each kind is the small input that every cBench program takes by default (CK's ctuning-datasets-min); the others are the rest of the set.

Use with cMeta

With cMeta (pip install cmeta, the public cmeta-aops repository and cmeta-ck-legacy), every kind is the dataset ctuning-<kind> and its inputs are the dataset's sub-datasets, fetched on demand from this repository (one kind at a time, checked against index.json; the 0001 input travels with the dataset). After your access request is accepted, log in once (hf auth login, or cx tool run hf -- auth login):

cx program run cbench-consumer-jpeg-c cpu --dataset=ctuning-image-ppm/image-ppm-0007 -q
cx task run test-ck-programs --programs=cbench-* --all_datasets -q      # every program on every input

Without cMeta: the files are as they are; index.json or an input's ck-meta.json tells which file a program takes first.

Provenance

The inputs are those of the CK repositories ctuning-datasets (inputs 0002 and above) and ctuning-datasets-min (0001), copied byte for byte from the CK home of the cTuning server (restored in 2025); the CK meta of every input is kept beside it. Their CK description reads "cDatasets/MiDataSets for cBench" (Collective Mind, 2012-04-12). Left out: the DNN images of ctuning-datasets-min (used by the CK DNN workflows) and the office webcam snapshots of the CK paper, which have a dataset of their own (ctuning/CK-office-snapshots, Zenodo 10.5281/zenodo.24275729).

Downloads last month
33