SGEMM_GPU_kernel_performance

OpenML dataset with id 43871

No author found.

Full work available at URL: https://api.openml.org/data/v1/download/22102748/SGEMM_GPU_kernel_performance.arff

Upload date: 20 April 2022
Copyright license: Creative Commons Attribution 4.0 International

Dataset Characteristics

Number of classes: 0
Number of features: 15 (numeric: 15, symbolic: 0 and in total binary: 0 )
Number of instances: 241,600
Number of instances with missing values: 0
Number of missing values: 0

Description

Dataset description

This data set measures the running time of a matrix-matrix product $A \times B = C$, where all matrices have size 2048 x 2048, using a parameterizable *SGEMM GPU* (Single Precision General Matrix Multiply) kernel with 241600 possible parameter combinations. For each tested combination, 4 runs were performed and their results are reported as the 4 last columns. All times are measured in milliseconds*. There are 14 parameters, the first 10 are ordinal and can only take up to 4 different powers of two values, and the 4 last variables are binary. Out of 1327104 total parameter combinations, only 241600 are feasible (due to various kernel constraints). This data set contains the results for all these feasible combinations. The experiment was run on a desktop workstation running Ubuntu 16.04 Linux with an Intel Core i5 (3.5GHz), 16GB RAM, and a NVidia Geforce GTX 680 4GB GF580 GTX-1.5GB GPU. We use the 'gemm_fast' kernel from the automatic OpenCL kernel tuning library 'CLTune' (https://github.com/CNugteren/CLTune).

\* *Note*: For this kind of data sets it is usually better to work with the logarithm of the running times (see e.g. Falch and Elster, 'Machine learning-based auto-tuning for enhanced performance portability of OpenCL applications', 2015).

Attribute description

Independent variables*

MWG, NWG: per-matrix 2D tiling at workgroup level: {16, 32, 64, 128}

(integer)

KWG: inner dimension of 2D tiling at workgroup level: {16, 32} (integer)
MDIMC, NDIMC: local workgroup size: {8, 16, 32} (integer) 6-7. MDIMA, NDIMB: local memory shape:

{8, 16, 32} (integer)

KWI: kernel loop unrolling factor: {2, 8} (integer)
VWM, VWN: per-matrix vector widths for loading and storing: {1, 2, 4, 8} (integer)
STRM, STRN: enable stride for accessing off-chip memory within a single thread: {0, 1} (categorical)
SA, SB: per-matrix manual caching of the 2D workgroup tile: {0, 1} (categorical) -

Output*

Run1, Run2, Run3, Run4: performance times in milliseconds for 4 independent runs using the same

parameters. They range between 13.25 and 3397.08. Run1 is used as the default target variable.

Related Studies

Rafael Ballester-Ripoll, Enrique G. Paredes, Renato Pajarola. Sobol Tensor Trains for Global Sensitivity Analysis. In arXiv Computer Science / Numerical Analysis e-prints, 2017, https://doi.org/10.1016/j.ress.2018.11.007

Authors

Enrique Paredes and Rafael Ballester-Ripoll. The original data was obtained from the UCI Machine Learning repository Link.

Citation

Please cite one of the following papers:

Rafael Ballester-Ripoll, Enrique G. Paredes, Renato Pajarola.

Sobol Tensor Trains for Global Sensitivity Analysis. In arXiv Computer Science / Numerical Analysis e-prints, 2017, https://arxiv.org/abs/1712.00233

Cedric Nugteren and Valeriu Codreanu.

CLTune: A Generic Auto-Tuner for OpenCL Kernels. In: MCSoC: 9th International Symposium on Embedded Multicore/Many-core Systems-on-Chip. IEEE, 2015, https://doi.org/10.1109/MCSoC.2015.10