Cédric Augonnet
|
5a708531e7
add some legend to the perfmodel debug output
|
16 years ago |
Cédric Augonnet
|
b5bbb50551
Use candlesticks to represent the actual dispersion of the measures, first
|
16 years ago |
Cédric Augonnet
|
a3b4f68e51
Even if the debug mode is not enabled, it is possible to request what would be
|
16 years ago |
Cédric Augonnet
|
27d7b4b275
Do no generate intermediate files when generating the "distribution" profile
|
16 years ago |
Cédric Augonnet
|
414a67a449
add a new script to automatically generate a "nice" output of the model debug
|
16 years ago |
Cédric Augonnet
|
c66ad02803
- The starpu-perfmodel tool may also give the path to the file containing the
|
16 years ago |
Cédric Augonnet
|
f6527fd6da
it is possible to directly request a parameter now, eg:
|
16 years ago |
Cédric Augonnet
|
ec636ae9ca
add a new utility to read the performance models
|
16 years ago |
Cédric Augonnet
|
a51fd80a69
use dynamic code loading mechanisms instead of some broken hard-coded stuff.
|
16 years ago |
Cédric Augonnet
|
63ec9fa7cf
Since we are going to need the "null" gordon kernel often, let's create a small
|
16 years ago |
Cédric Augonnet
|
c1f00e8c91
update the StarPU Gordon driver to the recent changes in Gordon
|
16 years ago |
Cédric Augonnet
|
554cb8bc74
dynamically load an empty kernel for the microbench that measure overhead
|
16 years ago |
Cédric Augonnet
|
76b022ddf0
do not forget to include ppm-downscaler.h in the distribw
|
16 years ago |
Cédric Augonnet
|
c5e951e16a
we need to get rid of .sampling/* between every iteration, not only at the beginning !
|
16 years ago |
Cédric Augonnet
|
be54ec551f
rework the gnuplot output of the "error-model" test
|
16 years ago |
Cédric Augonnet
|
3521786f90
add a *terribly* ugly script to measure the accuracy of performance prediction.
|
16 years ago |
Cédric Augonnet
|
95f9c1a844
When the "verbose" option is enabled, create a log file. Its content is filled
|
16 years ago |
Cédric Augonnet
|
c59fcbc747
do not display debug information after pthread_exit
|
16 years ago |
Cédric Augonnet
|
72528f57d7
Measure the difference between predicted execution time and measured execution
|
16 years ago |
Cédric Augonnet
|
01ed5d79fd
- use a performance models based on history
|
16 years ago |
Cédric Augonnet
|
d5f7fb9d95
add a new benchmark to measure the performance of sgemm on square matrices (to
|
16 years ago |
Samuel Thibault
|
153901a43e
typos
|
16 years ago |
Cédric Augonnet
|
81375feb05
add a new example of matrix multiplication with no filter, only one BLAS ops is
|
16 years ago |
Cédric Augonnet
|
91bc53fb3e
Include the worker identifier in the performance model debug logs.
|
16 years ago |
Cédric Augonnet
|
a729df228a
it is possible to specify where to bind the different worker using the
|
16 years ago |
Cédric Augonnet
|
8a02f335ee
underline that task submission is asynchronous
|
16 years ago |
Cédric Augonnet
|
c37b910143
document the basic examples so that they should be self-sufficient
|
16 years ago |
Cédric Augonnet
|
232d016ed7
implement a very simple matrix multiplication that uses filters
|
16 years ago |
Cédric Augonnet
|
e6c184c3dc
specify what is described/explained in the hello-world example
|
16 years ago |
Cédric Augonnet
|
1d8b7066e9
- put a short description of the goal of the vector-scal example
|
16 years ago |