 Cédric Augonnet
|
5763bd2d0c
Implement starpu_get_vector_elemsize that was somehow missing.
|
hace 17 años |
|
Cédric Augonnet
|
2a2758ac3e
Make the "ring" example more challenging as we now have both MPI and CUDA
|
hace 17 años |
|
Cédric Augonnet
|
c3c1b0de02
Introduce the MPI helper lib (enabled with --with-mpi at configure time).
|
hace 17 años |
|
Cédric Augonnet
|
c94a373e97
Fix the starpu_sync_data_with_mem API, it is now required to call
|
hace 17 años |
|
Cédric Augonnet
|
fff603e54c
Make sure that the directory that should contain the performance models is
|
hace 17 años |
|
Cédric Augonnet
|
0363a81ed4
fix compilation in case CUDA is not used
|
hace 17 años |
|
Cédric Augonnet
|
0934f728e3
We don't hardcode the performance of the bus anymore, so that tool is useless
|
hace 17 años |
|
Cédric Augonnet
|
c054754a75
Implement bus performance sampling: this is no longer hardcoded.
|
hace 17 años |
|
Cédric Augonnet
|
a3be10b313
The performande models associated to the codelets are now put in
|
hace 17 años |
|
Cédric Augonnet
|
3cad2c6bbb
The src and dst nodes were inverted in the matrices
|
hace 17 años |
|
Cédric Augonnet
|
e9b8da269e
add some dirty program to measure the bandwith/latency used by the performance models
|
hace 17 años |
|
Cédric Augonnet
|
0b80205b82
It is possible to generate a file with the distribution of the duration of the
|
hace 17 años |
|
Cédric Augonnet
|
bcd9c67fc4
Fix debug assertions, it is possible that no request is associated to a data
|
hace 17 años |
|
Cédric Augonnet
|
e561dd511a
We now assume that the header lock is already taken when calling
|
hace 17 años |
|
Cédric Augonnet
|
eb5c03ad79
BUG FIX: when a data dependency is unlocked, we have to update the current mode
|
hace 17 años |
|
Cédric Augonnet
|
485ce7b82f
cudaMemset only fills one byte at a time, so we use a char array instead of an
|
hace 17 años |
|
Cédric Augonnet
|
20420637e6
starpu_wait_all_tasks and starpu_wait_task are blocking too.
|
hace 17 años |
|
Cédric Augonnet
|
6f21718f48
The helper functions to allocate/liberate pinned memory use tasks internally,
|
hace 17 años |
|
Cédric Augonnet
|
5284a61274
When a data request needs to actually perform a data transfer from a source, we
|
hace 17 años |
|
Cédric Augonnet
|
499cbfed6a
Also check that we don't synchronize with a tag while blocking is prohibited.
|
hace 17 años |
|
Cédric Augonnet
|
02119279fc
We now check whether the following blocking calls are called from a callback or
|
hace 17 años |
|
Cédric Augonnet
|
a9394e1af3
Keep the status of each worker so that we can detect whether a blocking call is
|
hace 17 años |
|
Cédric Augonnet
|
4ac2e92f42
Add another classical synchronization problem to stress data management.
|
hace 17 años |
|
Cédric Augonnet
|
152b6a7329
Add a the classical dining philosopher example: we have to explicitely order
|
hace 17 años |
|
Cédric Augonnet
|
9e1bd3857d
Make it possible to select the maximum number of buffers per task at configure
|
hace 17 años |
|
Cédric Augonnet
|
da38bcd033
Implement async data transfers for the "block" data interface.
|
hace 17 años |
|
Cédric Augonnet
|
d3cfd654a8
Cleanup FxT tool, there is no svg output anymore (Vite is much better in every
|
hace 17 años |
|
Cédric Augonnet
|
d8132d45cf
Make it possible to select the colour of the tasks in the DOT output depending
|
hace 17 años |
|
Cédric Augonnet
|
82828a010d
On some systems (eg. MacOS), it's just not possible to bind threads at all, so
|
hace 17 años |
|
Cédric Augonnet
|
98540ae1f7
don't compile CUDA functions if CUDA is not available
|
hace 17 años |
