 Samuel Thibault
|
ab9b3fcfbd
forwardport r6997 from 1.0: fix bus calibration for more than 32 cpus
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
f33693b5d4
Add actual bus measurements in starpu_machine_display output
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
f30a1170aa
Replace starpu_force_bus_sampling with bus_calibrate configuration option, as calibration should done carefully during starpu initialization, and not after initialization
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
09dab6fd3e
Drop spurious nbsp
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
c39f87d9cd
clearly separate CUDA and OpenCL
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
473617f4c0
Do not print CPU%d, it is misleading, explain that it is CPUs in preferrence order, instead
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
75ad46f0de
document temporary buffer optimizations
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
a7aa151601
revert r6895, it perturbs measurements
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
226d8973c7
Do not use starpu_cuda_set_device while calibrating the bus: since we clear things with cudaThreadExit, opengl interoperability can be enabled at proper starpu initialization
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
9b373e7f9c
Fix initialization for CUDA bus calibration: we use cudaThreadExit to clean our environment before starting starpu
|
%!s(int64=13) %!d(string=hai) anos |
 Nathalie Furmento
|
e5545db0a7
mpi stats: there is no need to have a 2D array to store comm stats as we only store amount of data sent from the current node
|
%!s(int64=13) %!d(string=hai) anos |
|
Nathalie Furmento
|
99f71d5c50
mpi: The cache of starpu_mpi_insert_task is fixed and thus now enabled by default.
|
%!s(int64=13) %!d(string=hai) anos |
|
Nathalie Furmento
|
d03319d5a0
mpi/starpu_mpi_insert_task.c: fix indice for looking for data in hashtable
|
%!s(int64=13) %!d(string=hai) anos |
|
Nathalie Furmento
|
52064847b0
mpi: only declare cache related stuff when needed
|
%!s(int64=13) %!d(string=hai) anos |
|
Nathalie Furmento
|
6ad304e82a
mpi: remove files which are no longer needed - clearing cache is now done directly and no longer through a task
|
%!s(int64=13) %!d(string=hai) anos |
|
Nathalie Furmento
|
e96c40fa4a
mpi: fix cache mechanism for starpu_mpi_insert_task
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
3ed6d73180
update changelog
|
%!s(int64=13) %!d(string=hai) anos |
|
Nathalie Furmento
|
8d892f8554
mpi: Deal with new parameter STARPU_DATA_ARRAY in starpu_mpi_insert_task
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
438839d828
Enable GPU-GPU direct transfers
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
af21f804df
Fix choosing the right CPU for driving GPU
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
5abb0b8256
Fix macro for use directly in an if
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
e0d408b020
reduce memory use by deallocating useless data buffers
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
902b5befd6
Add starpu_data_invalidate_submit
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
a42a37bf8e
rename starpu_data_unregister_lazy into starpu_data_unregister_submit for naming coherency
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
34212a9b9f
mention that tasks must be finished before unpartitioning
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
f44d649ebf
Add starpu_data_acquire_on_node
|
%!s(int64=13) %!d(string=hai) anos |
|
Samuel Thibault
|
b4dedc2d59
more explicit
|
%!s(int64=13) %!d(string=hai) anos |
|
Nathalie Furmento
|
6ecd809dd0
ChangeLog: fix following the latest commit
|
%!s(int64=13) %!d(string=hai) anos |
|
Nathalie Furmento
|
d6416329c9
Merge 2 functions starpu_insert_task and starpu_insert_task_array.
|
%!s(int64=13) %!d(string=hai) anos |
|
Nathalie Furmento
|
9e98540ad3
ChangeLog: mention starpu_insert_task_array
|
%!s(int64=13) %!d(string=hai) anos |
