4.4 KiB
Introduction
This is a directory developed by Luca Pepè Sciarria and Simone Staccone froma Tor Vergata University to start to create some unit tests for PSBLAS 3.9, in particular for psb_spmm routine.
Environment
These tests are developed using a linux environment, in particular Rocky Linux 9.5 (Blue Onyx).
The compiler used is:
- gnu 12.2.1
The necessary dependnces are:
- mpich 4.2.2
- psblas 3.9
In order to have the exact same environment used for testing compile PSBALS library using cuda 12.5.
Getting started
Steps to reproduce the tests:
- Compile the code using
make - Insert the matrix files inside the matrix/ directory (or create one if it doesn't exists; psblas3/test/spmm/matrix/)
- Launch the script autotest.sh
- Check the output log file test_log.txt to collect results
Test goal
Check the correctness of the matrix-vector multiplication y = Ax using the psb_spmm routine, checking for all the test suite cases.
Test Suite
Overall Analysys
THe subroutine psb_spmm materialize in three different procedures:
psb_sspmm, psb_sspmv, psb_sspmv_vect. The focus is on testing the psb_sspmv one, so as first step let's analyze the signature of the function to study a satisfyng input parameter space. The signature of the function is:
psb_sspmv(alpha, a, x, beta, y,&
& desc_a, info, trans, work,doswap)
In order to have a black-box approach and to consider only the functionalities of the routine, the parameters work and doswap are not considered during this tests, since they are used in situation that presents internally to the library code. Moreover, info it's an output parameter used to signal some kind of error, therefore it is not part of the input space.
There are seven parameters left. In the following tables are highlighted the values assigned to each parameter. THe test are runned composing a Cartesian product of all the possible values assigned to each parameter.
Parameters Values
x vectors are located in the vectors/ directory. They are generated randomly using the same seed and then saved on different files based on their characteristics. The size of the vector is choosen accordingly to the size of the matrix column space considered for the single test instance.
| Vector | File Name | Coefficients | Coefficients Description |
|---|---|---|---|
x_1 |
x1.txt | x_i> 0, \forall i |
Positive coefficients |
x_2 |
x2.txt | x_i < 0, \forall i |
Negative coefficients |
x_3 |
x3.txt | x_i \ne 0, \forall i |
Random coefficients |
x_4 |
x4.txt | x_i = 0, \forall i |
Null coefficients |
y vectors are located in the vectors/ directory. They are generated randomly using the same seed and then saved on different files based on their characteristics. The size of the vector is choosen accordingly to the size of the matrix rows space considered for the single test instance.
| Vector | File Name | Coefficients | Coefficients Description |
|---|---|---|---|
y_1 |
y1.txt | y_i> 0, \forall i |
Positive coefficients |
y_2 |
y2.txt | y_i < 0, \forall i |
Negative coefficients |
y_3 |
y3.txt | y_i \ne 0, \forall i |
Random coefficients |
y_4 |
y4.txt | y_i = 0, \forall i |
Null coefficients |
\alpha
\alpha |
Value | Coefficients Description |
|---|---|---|
\alpha_1 |
1.0 | Positive value |
\alpha_2 |
-1.0 | Negative value |
\alpha_3 |
0.0 | Null value |
\beta
\alpha |
Value | Coefficients Description |
|---|---|---|
\beta_1 |
1.0 | Positive value |
\beta_2 |
-1.0 | Negative value |
\beta_3 |
0.0 | Null value |
trans
Output
The ouput files generated by the test are automatically compared by the autotest.sh script, but if it is needed to manually run the test here it is the naming convenction used.
The results of the computation will be saved on different files based on the instance of the test considered. In particular the naming conventiona format the output file as sol_m#_x#_y#_a#_b#.mtx, where each # is a number choosen w.r.t. the test instance. (Ex. sol_m1_x1_y1_a1_b1.mtx is the solution computed using the first matrix file, the first x vector file , the first y vector file, alpha = 1.0 and beta = 1.0). Moreover, the files will be saved in the serial/ directory if the program is launched using 1 process or in parrallel/ directory if the program is launched with more than one process.
Notes
For now only integer multiplication is tested and on a single matrix
TODO
- Update docs
- Parametrize also on trans parameter
- Create the script autotest.sh
- Write the I/O output on log file
- ...