 Samuel Thibault
|
206b3adf49
port r13269 from 1.1: free data outside critical section
|
10 年之前 |
|
Samuel Thibault
|
f43bf86292
Fix complexity of implicit task/data dependency, from quadratic to linear
|
10 年之前 |
 Nathalie Furmento
|
e1fafd8cb1
src/datawizard/interfaces/data_interface.c: display warning when starpu_shutdown is called while there is some unregistered data handles (thanks to Benoit Lize for suggesting the feature)
|
11 年之前 |
|
Nathalie Furmento
|
733a26726d
backport branches/starpu-1.1@13065: src/datawizard/interfaces/data_interface.c: only display message once when calling starpu_data_set_rank and starpu_data_set_tag
|
11 年之前 |
|
Samuel Thibault
|
b24213f35d
Add a way to register an existing on-GPU buffer to be used by a handle
|
11 年之前 |
|
Samuel Thibault
|
4a9e46d095
Reduce the memory footprint of the "requested" bitfield
|
11 年之前 |
|
Samuel Thibault
|
bbbf1610c4
Allocate dynamically an array used only while reducing data
|
11 年之前 |
|
Samuel Thibault
|
f255d2c95a
port r12703 from 1.1: Allow data to be uninitialized but the application provide an init_cl to initialize it
|
11 年之前 |
|
Samuel Thibault
|
ea5d8f596b
Make a copy of the interface to the memchunk only when the latter gets detached from the data, and thus the interface code will not work on it. Drop the copy when the memchunk gets reattached. This allows interfaces to modify pointers in the interface in unpack, notably
|
11 年之前 |
 Andra Hugo
|
aed609a645
rollback r12518: don't allow unregister in callback or task codelets
|
11 年之前 |
|
Andra Hugo
|
a78e7bdc6a
allow unregisters in callbacks and codelets when we have only CPUs
|
11 年之前 |
|
Nathalie Furmento
|
f08f953c72
src/datawizard/interfaces/data_interface.c: modify _starpu_data_unregister() to not fetch the data if there
|
11 年之前 |
|
Samuel Thibault
|
8e3d4676f4
Make _starpu_data_wait_until_available take a synchronization name, so that the node in the DAG becomes understandable
|
11 年之前 |
|
Samuel Thibault
|
1a62c13496
reformat comment
|
11 年之前 |
|
Samuel Thibault
|
b59aff6ad5
Fix starpu_data_unregister_submit and starpu_data_unregister_no_coherency, to make them properly wait for pending tasks
|
11 年之前 |
|
Samuel Thibault
|
24f00257b5
Fix acquisition of data on node -1, re-enable using it in data_invalidate
|
11 年之前 |
|
Samuel Thibault
|
c8c8eeeb31
More fixes for acquiring node -1 case. Still not working apparently, so not using it for data_invalidate for now.
|
11 年之前 |
|
Samuel Thibault
|
515b80ef3f
Avoid making starpu_data_invalidate actually allocate RAM memory, by passing node -1 to acquire_on_node, and making the latter not allocate in that case. This fixes overzealous memory consumption by MPI layer
|
11 年之前 |
|
Samuel Thibault
|
092378f2cc
Avoid releasing/retaking the spinlock for non-MPI data
|
11 年之前 |
|
Nathalie Furmento
|
0e69cb4725
src/datawizard/interfaces/data_interface.c: need to release the lock before calling the unregister_hook
|
11 年之前 |
|
Nathalie Furmento
|
2a9ebe7478
src: rename field to unregister_hook
|
11 年之前 |
|
Nathalie Furmento
|
d60dbc4cf1
src: advise not to call starpu_data_set_rank and starpu_data_set_tag but starpu_mpi_data_register (not yet committed)
|
11 年之前 |
|
Nathalie Furmento
|
84b8576862
src: add hook to struct _starpu_data_state which will be called when unregistering the data
|
11 年之前 |
|
Nathalie Furmento
|
42c8e7dc8e
turn function starpu_data_get_data_handle_from_tag private and hence rename it as _starpu_data_get_data_handle_from_tag
|
11 年之前 |
|
Samuel Thibault
|
2f62b5cd10
Use big fat abortions when one tries to make a task or callback sleep, instead of just returning EDEADLCK which few people will test
|
11 年之前 |
|
Samuel Thibault
|
6b5492002c
Drop unused variable
|
11 年之前 |
|
Nathalie Furmento
|
31f6ba2030
use STARPU_MAIN_RAM when appropriate
|
11 年之前 |
|
Nathalie Furmento
|
5f35f4c737
src/datawizard/interfaces/data_interface.c: rename internal function
|
11 年之前 |
|
Samuel Thibault
|
245bdafca2
port r11388 from 1.1: Allocate RAM memory immediately, without using requests
|
11 年之前 |
|
Samuel Thibault
|
3c7645aeff
port r11168 from 1.1: Also do not lie about spinlocks either, rather ignore accesses to busy_count. Also actually fix a missing check in perfmodel.c
|
11 年之前 |
