 Cédric Augonnet
|
7a58fe13de
Do report the size of the data tranfers in the FxT traces.
|
16 years ago |
|
Cédric Augonnet
|
c63a74a4ca
fix the restart test
|
16 years ago |
|
Cédric Augonnet
|
622da03466
bug fix: we consider that a write-only buffer is cost-less (we still ignore
|
16 years ago |
|
Cédric Augonnet
|
1e41ce878c
minor cleanups to avoid some dereferencing
|
16 years ago |
|
Cédric Augonnet
|
6bb9f7d56a
reset the "beta" value of the dmda policy which was modified by mistake
|
16 years ago |
|
Cédric Augonnet
|
d4c76f3f46
fix debug code
|
16 years ago |
|
Cédric Augonnet
|
07a84c2b9f
While sampling the performance of the bus, we now measure which are the CPUs
|
16 years ago |
|
Cédric Augonnet
|
2b7933a7f4
Simplify code
|
16 years ago |
|
Cédric Augonnet
|
cf4df79338
The footprint method of the data_interface_ops_t structure now only takes a
|
16 years ago |
|
Cédric Augonnet
|
0636d338aa
Add a --enable-fast option where the STARPU_ASSERT are ignored (by default,
|
16 years ago |
|
Cédric Augonnet
|
0281ad5c41
** IMPORTANT **
|
16 years ago |
|
Cédric Augonnet
|
99cc9c0ec0
The starpu_test_if_data_is_allocated_on_node makes it possible to check whether
|
16 years ago |
|
Cédric Augonnet
|
a6b0a7fbee
The filter functions need not return the number of children as the nchildren
|
16 years ago |
|
Cédric Augonnet
|
e081dedd94
- Factorize the code to register a new piece of data so that adding an
|
16 years ago |
|
Cédric Augonnet
|
67ee67d411
There is no need to explicitely set the interface ops field since
|
16 years ago |
|
Cédric Augonnet
|
066a3c3236
Add a test to make sure MPI works well with data in pinned memory.
|
16 years ago |
|
Cédric Augonnet
|
6fee32672b
Include the size of the interface structure in the data_interface_ops_t
|
16 years ago |
|
Cédric Augonnet
|
e8fc456195
follow up of the previous cleanup commit
|
16 years ago |
|
Cédric Augonnet
|
077020f485
Cleanup: get rid of the data_state type which was just a useless alias for the
|
16 years ago |
|
Cédric Augonnet
|
f5b2fe1170
Simplify the code by manipulating "handles" instead of starpu_data_state_t
|
16 years ago |
|
Cédric Augonnet
|
fbc16730b4
Minor cleanups: we expose the starpu_data_handle opaque type rather than the
|
16 years ago |
|
Cédric Augonnet
|
c113a7a008
Remove the dump_data_interface field of the data_interface_ops_t structure as
|
16 years ago |
|
Cédric Augonnet
|
d770a3c10f
fix typo
|
16 years ago |
|
Cédric Augonnet
|
8c86d9e771
Start to get rid of the starpu_data_interface_t enum type: for now, the codelet
|
16 years ago |
|
Cédric Augonnet
|
b99221a3bb
Cleanup the various data interfaces by using the
|
16 years ago |
|
Cédric Augonnet
|
d80fb2d49a
Revert previous commit: as the comment DID tell, there is a problem as we could
|
16 years ago |
|
Cédric Augonnet
|
4288eb2a81
There is no need to keep a local copy of the "interface" within the memory
|
16 years ago |
|
Cédric Augonnet
|
da6206e6cc
Provide a new helper function starpu_get_local_cuda_stream
|
16 years ago |
|
Cédric Augonnet
|
dbbf7a2a93
Simplify the cublas helper functions by the means of the
|
16 years ago |
|
Cédric Augonnet
|
c35710a810
do not hardcode the number of memory nodes anymore
|
16 years ago |
