https://github.com/lanl/compacthash
A compact hash algorithm for CPUs and GPUs using OpenCL
https://github.com/lanl/compacthash
Last synced: about 1 year ago
JSON representation
A compact hash algorithm for CPUs and GPUs using OpenCL
- Host: GitHub
- URL: https://github.com/lanl/compacthash
- Owner: lanl
- License: other
- Created: 2014-02-15T15:27:23.000Z (over 12 years ago)
- Default Branch: master
- Last Pushed: 2020-09-26T15:07:41.000Z (over 5 years ago)
- Last Synced: 2024-06-27T00:20:06.687Z (almost 2 years ago)
- Language: C
- Size: 1.3 MB
- Stars: 14
- Watchers: 9
- Forks: 5
- Open Issues: 0
-
Metadata Files:
- Readme: README
- License: LICENSE
Awesome Lists containing this project
README
Compact Hash Algorithms for Computational Meshes: Data from Neighbors 2D Testbed
This code is a set of hash functions applied to a sample mesh. It supports the paper "Compact Hash Algorithms for Computational Meshes," submitted to the SIAM Journal of Scientific Computing.
Authors:
Bob Robey XCP-2 (brobey@lanl.gov)
Peter Ahrens XCP-2 (peter.ahrens@lanl.gov, ptrahrens@gmail.com)
Sara Hartse XCP-2 (sara@lanl.gov, sara.hartse@gmail.com)
Rebecka Tumblin (rtumblin@lanl.gov, rebeckatumblin@gmail.com)
On a machine with at least one OpenCL compliant device:
Enter the 'parallel_hash' directory
./configure
make
./neigh2d -o
Configure options
--enable-detailed-timing
--disable-write-mem-usage
=================================================================
Builds with and without OpenCL
=================================================================
By default, OpenCL is included in the build. This can be problematic
on some architectures. If OpenCL is not found, it will try to complete
the build without it and only the CPU routines will be available. You can
build without OpenCL with a configure option:
./configure --enable-opencl=no
or
./configure --disable-opencl
If you want OpenCL in the build and the location of the OpenCL library
is causing the problem, try setting the environment variables OPENCL_INCLUDE
and OPENCL_LIBS to the locations of these files.
If the build fails, you may have to go into the hash subdirectory
and force hash.c and hash.h to rebuild. Try "make hash.c hash.h" or
remove the hash.c and hash.h files and make them.
=================================================================
On Mac OSx install gnu-indent. The Mac version does not recognize some of the options that are used.
'./neigh2d -h' gives further information on option keywords.
The option -r randomizes the cell data (default is not random)
The option -n sets the min mesh size and -N the max mesh size [ default is 256 min to 256 max ]
The option -l sets the min levmx and -L the max levmx [ default is 1 min to 5 max ]
The -t option refers to threshold for refinement of cells and can range from 1 to 100. Cells
are randomly refined and a lower threshold means fewer will be refined. The default value
is 20 and that was what was used in most of our trials.
The keywords for corresponding options are as follows:
br Brute
kd kdtree
hc Hash CPU
hlc Hash Lib CPU
hc1 Hash CPU 1
hlc1 Hash Lib CPU 1
hc2 Hash CPU 2
hlc2 Hash Lib CPU 2
hc3 Hash CPU 3
hlc3 Hash Lib CPU 3
holc3 Hash OldLib CPU 3
hg Hash GPU
hg1 Hash GPU 1
hg2 Hash GPU 2
hg3 Hash GPU 3
hlg Hash Lib GPU
hlg1 Hash Lib GPU 1
hlg2 Hash Lib GPU 2
hlg3 Hash Lib GPU 3
holg3 Hash OldLib GPU 3
Our figures:
Memory optimizations figure (each run on a Kepler)
Use the -r randomization flag
hg
hg1
hg2
hg3
holg3
Compact versus perfect figure (run on various cards)
with and without the -r randomization flag
./neigh2d -r -t 2 -L 6 -o hg3 hlg3 and ./neigh2d -t 2 -L 6 -o hg3 hlg3
hg3
holg3
Table construction (each run on Kepler)
No randomization of cell data in adaptiveMeshConstructorWi -- don't use -r
hc
hg
holg3
Output file features:
Hardware detection report - an overview of the CL devices and their information
Local work size - the size of the workgroup size
Size - side length of mesh at coarsest level
ncells - total number of cells, accounting for refined cells
compression - The number of possible cells at most refined level/ncells