Nathalie Furmento
|
e13a68dcbb
tests: remove second definition of FPRINTF
|
14 years ago |
Nathalie Furmento
|
13cd230e8c
tests/: improve error checking
|
14 years ago |
Nathalie Furmento
|
ec44ab2026
tests/: improve error checking
|
14 years ago |
Nathalie Furmento
|
cca5379613
tests/loader.c: print message when execution times out
|
14 years ago |
Nathalie Furmento
|
f00330a558
src/core/workers.c: display message only if environment variable STARPU_SILENT is unset
|
14 years ago |
Nathalie Furmento
|
8afb82815b
src/core/perfmodel/perfmodel_bus.c: improve error checking for cudaHostAlloc
|
14 years ago |
Nathalie Furmento
|
2732132e82
examples/cholesky/cholesky_models.c: include cholesky.h to define FPRINTF
|
14 years ago |
Nathalie Furmento
|
b98e434d3b
src/core/perfmodel/perfmodel_history.c: fix type for variable handle
|
14 years ago |
Nathalie Furmento
|
2a767d273b
src/datawizard/coherency.c: include math.h to define INFINITY
|
14 years ago |
Samuel Thibault
|
1bcc606097
TODO to fix memcpy3DPeer
|
14 years ago |
Samuel Thibault
|
f336e6a37f
fix spacing
|
14 years ago |
Samuel Thibault
|
fc9d76bf52
Fix not choosing RAM when it actually does not have a valid copy
|
14 years ago |
Samuel Thibault
|
f428415792
properly terminate starpu even on error
|
14 years ago |
Samuel Thibault
|
24fd5851a4
do not crash on error, only return 1
|
14 years ago |
Samuel Thibault
|
0fc0871ae4
Permit to try direct GPU-GPU matrix transfers even if cudaMemcpy3DPeer is broken, when matrices have no stride
|
14 years ago |
Samuel Thibault
|
ccf09d0cfa
Fix assertion
|
14 years ago |
Samuel Thibault
|
15e1b57cd1
drop spurious commit
|
14 years ago |
Samuel Thibault
|
09d5790083
Take transfer time into account when choosing the source. Also check that direct transfer is possible. Add GPU-GPU bus to account for statistics
|
14 years ago |
Samuel Thibault
|
29801fa627
fix warnings
|
14 years ago |
Samuel Thibault
|
e72f3acf6c
really use the best implementation we have computed
|
14 years ago |
Samuel Thibault
|
d50fd7e681
fix warnings
|
14 years ago |
Samuel Thibault
|
907bfdd621
avoid measuring transfer inside GPU
|
14 years ago |
Samuel Thibault
|
14f2b383d8
Fix bandwidth measurement when gpu sizes are differing and small: make timing µs per byte instead of just µs
|
14 years ago |
Samuel Thibault
|
2eba67b128
Add GPU-GPU transfer measurement
|
14 years ago |
Samuel Thibault
|
7d283d3f14
Do not let glpk try to solve a supposedly-impossible problem because there is no performance model for some tasks
|
14 years ago |
Samuel Thibault
|
a5661257d0
do not let bound computation equations half-empty when some tasks are not calibrated.
|
14 years ago |
Samuel Thibault
|
0925502c69
Do not unload OpenCL programs when they haven't been loaded due to no worker (see r4310)
|
14 years ago |
Nathalie Furmento
|
011cf9daa0
OpenCL driver: get the correct number of OpenCL workers
|
14 years ago |
Samuel Thibault
|
2b6c0e9af5
Make sure the topology is initialized when benchmarking the bus performance
|
14 years ago |
Samuel Thibault
|
88a22759c2
Warn when some debugging configuration option was used
|
14 years ago |