Copyright (c) 2015 The University of Edinburgh.
This software was developed as part of the
EC FP7 funded project Adept (Project ID: 610490)
http://www.adept-project.eu
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
This README describes the UPC parallel kernel benchmarks.
PLEASE NOTE that currently all UPC benchmarks are implemented to use dynamic compilation only (i.e. the number of THREADS that will be used at runtime is unknown at compile time). This can result in poor performance because the data layout follows the default cyclic distribution. In future releases of these benchmarks, we hope to add options for static compilation as well. (16/09/2015)
If you would like to cite this work, please cite: Nick Johnson et al., "Adept Deliverable D2.3 - Updated Report on Adept Benchmarks", September 2015. available at http://www.adept-project.eu/images/Deliverables/Adept%20D2.3.pdf
In our BLAS-type benchmarks we implement a few of the most common linear algebra computations.
This benchmark takes two vectors x and y, and the scalar a, and computes:
y = a * x + y
The user can choose the length (number of elements) of the vectors, as well as their data type (int, float or double).
The dot product benchmark multiplies two vectors x and y of length n and returns a scalar:
result = x_0 y0 + x_1 y_1 + ... x_n y_n
The user can choose the length (number of elements) of the vectors, as well as their data type (int, float or double).
Thise benchmark scales the vector x by a fixed scalar a:
x = a * x
The user can choose the length (number of elements) of the vectors, as well as their data type (int, float or double).
This benchmarks computes for Euclidean norm of vector x:
|| x || = sqrt ( |x_1|^2 + |x_2|^2 + ... |x_n|^2 )
The user can choose the length (number of elements) of the vectors, as well as their data type (int, float or double).
This benchmarks multiplies a square dense matrix A with a vector x to compute vector y:
y = A * x
Both A and x are randomly generated. The user can choose the size of the data structures (where size*size equals the number of elements in the matrix), as well as their data type (int, float or double).
This benchmarks multiplies a square sparse matrix A with a vector x to compute vector y:
y = A * x
A is represented in CSR format and read from an input file. The vector x is randomly generated. The size of the matrix is fixed by the input file (which the user can substitute for a different matrix). The user can choose the data type to be used (float or double).
This benchmarks multiplies two square sparse matrices A and B to compute matrix C:
C = A * B
A and B are both represented in CSR format and read from an input file. The size of the matrices is fixed by the input file (which the user can substitute for a different matrix). The user can choose the data type to be used (float or double).
The stencil benchmarks compute values for each element in a 2D or 3D grid based on the values of their nearest neighbours.
On a 2D grid, the 5-point stencil computes the value of A[i][j] by taking the values from left, right, up and down from the current position, and scale them with a constant. The 9-point stencil is similar, but also includes the diagonals. The user can choose the data type to be used in the grid (int, float or double).
The 19-point and 27-point stencils are analogous to the 5 and 9 point stencil, but they operate in a 3D space. The user can choose the data type to be used in the grid (int, float or double).