 Cédric Augonnet
|
0fde7850af
Add a small script to automate the performance measurement on the AC cluster.
|
лет назад: 16 |
 Samuel Thibault
|
1774042480
TODO: prefix
|
лет назад: 16 |
|
Cédric Augonnet
|
cfc80b374c
Don't display debug messages by default.
|
лет назад: 16 |
|
Cédric Augonnet
|
1232517b40
Remove this barrier which was only used for debugging purpose.
|
лет назад: 16 |
|
Cédric Augonnet
|
5556836c9e
- Use libnuma if possible
|
лет назад: 16 |
|
Cédric Augonnet
|
f9ecd9b4a7
Test if libnuma is available at configure time.
|
лет назад: 16 |
|
Cédric Augonnet
|
b3caf41960
Get rid of mpi/examples/Makefile.am to avoid a dependency issue.
|
лет назад: 16 |
|
Samuel Thibault
|
e9acd82e11
double-buffer 12 and 21 contributions
|
лет назад: 16 |
|
Cédric Augonnet
|
7cec004adf
Do not wait for the termination of TAG*_SAVE when the data was processed localy
|
лет назад: 16 |
|
Cédric Augonnet
|
722ba734ee
Allocate pinned memory in an existing context rather than from the application
|
лет назад: 16 |
|
Cédric Augonnet
|
13dd236182
Keep this function public
|
лет назад: 16 |
 Nathalie Furmento
|
3c7a6bade8
Rename cuda functions
|
лет назад: 16 |
|
Samuel Thibault
|
fd11fd4590
TODO: use hwloc for binding for cuda perfs
|
лет назад: 16 |
|
Cédric Augonnet
|
f838d1aae0
Use an array of temporary buffers for "task 11" instead of a single buffer.
|
лет назад: 16 |
|
Cédric Augonnet
|
1d37506834
by default, disable debugging messages
|
лет назад: 16 |
|
Samuel Thibault
|
a378191763
debugging message according to C convention
|
лет назад: 16 |
|
Samuel Thibault
|
d22f7b3aef
fix tag dependency, seems to work better now
|
лет назад: 16 |
|
Samuel Thibault
|
34e12b640f
more readable debugging
|
лет назад: 16 |
|
Samuel Thibault
|
2c2170c9c8
initialize thread support, warn if no thread support is available
|
лет назад: 16 |
|
Samuel Thibault
|
fd84a2abd0
cosmetic changes
|
лет назад: 16 |
|
Samuel Thibault
|
31ec191f32
fix parameter name coherencey
|
лет назад: 16 |
|
Samuel Thibault
|
c5cc01e6d5
Fix MPI scatter
|
лет назад: 16 |
|
Cédric Augonnet
|
417c14b7f0
use the good parameters for CUDA calls
|
лет назад: 16 |
|
Samuel Thibault
|
c05e4565d4
only the owner of the block should send it to node 0
|
лет назад: 16 |
|
Cédric Augonnet
|
578c319b4a
Fix bug and add error checking
|
лет назад: 16 |
|
Cédric Augonnet
|
4408a54de9
add an experiment to measure the latency between 2 GPUs
|
лет назад: 16 |
|
Cédric Augonnet
|
c1273c8f02
- add a (pretty naive) test to measure the latency between 2 GPUs
|
лет назад: 16 |
|
Cédric Augonnet
|
516411357c
forgot to commit the public headers in the previous commit
|
лет назад: 16 |
|
Cédric Augonnet
|
3ea8462620
- Use events instead of streams to check whether a data transfer is terminated
|
лет назад: 16 |
|
Cédric Augonnet
|
b308e85842
The user events are typically not executed by a worker, so in case the event is
|
лет назад: 16 |
