 Corentin Salingue
|
df36b837d5
fix declaration of array
|
8 lat temu |
 Nathalie Furmento
|
a017fcbdeb
src: driver operations are stored in a struct _starpu_driver_ops which are used to implement the public driver API
|
8 lat temu |
 Samuel Thibault
|
43a3c904ed
directly lookup destination node instead of relying on the node of the current worker
|
8 lat temu |
|
Samuel Thibault
|
82ee3ee150
Clean up worker_set vs worker confusion, and CUDA initialization when several threads use the same device
|
8 lat temu |
|
Samuel Thibault
|
2510657bc9
Add code that takes into account PCI topology to determine contention. But disable it for now since it seems to actually decrease performance...
|
8 lat temu |
 Mathieu Lirzin
|
b911350051
src/drivers/cuda/driver_cuda.h: Delete useless headers.
|
10 lat temu |
|
Nathalie Furmento
|
bb61f3808d
update cnrs copyright
|
10 lat temu |
|
Samuel Thibault
|
4b1a16edde
Fix University name after the fusion..
|
10 lat temu |
|
Samuel Thibault
|
7240b27585
port r12759 from 1.1: Make sure we use the local stream, and not the stream of another device. Define separate streams for GPU-GPU transfers, depending on which device emits it
|
11 lat temu |
|
Samuel Thibault
|
f2b1d0f595
port r12584 from 1.1: Re-enable asynchronous transfers from CPUs, now that the issue has been found to be in application code
|
11 lat temu |
|
Samuel Thibault
|
092f322b1c
Add CUDA concurrent kernel execution support through the STARPU_NWORKER_PER_CUDA environment variable.
|
11 lat temu |
|
Samuel Thibault
|
058e113825
port r12533 from 1.1: Do not let non-CUDA workers use non-0 streams, CUDA seems not very threadsafe with that. Make the coherency engine avoid selecting non-CUDA workers to issue transfers, to avoid letting them use the 0 stream.
|
11 lat temu |
|
Samuel Thibault
|
6325a34ab3
Move device id lookup into run-driver functions, to avoid paying the cost at each task execution
|
11 lat temu |
|
Samuel Thibault
|
b4976f3d00
Use different streams for gpu-gpu transfers
|
11 lat temu |
|
Samuel Thibault
|
2b30d49db1
Make it possible to use simgrid without having CUDA/OpenCL installed
|
12 lat temu |
|
Samuel Thibault
|
27960c28bc
Use a separate stream for gpu in/out, and peer, as some cards can actually do some transfers in parallel
|
12 lat temu |
 Cyril Roelandt
|
ace934c823
Clean up src/core/topology.c.
|
12 lat temu |
|
Nathalie Furmento
|
8b00244b60
src: declare related driver functions in appropriate .h file
|
13 lat temu |
 Cédric Augonnet
|
8412f45ae4
Create a stream dedicated to StarPU's async data transfers so that we can
|
14 lat temu |
|
Samuel Thibault
|
0d9f08ae8d
Fix copyrights, thanks Ludo!
|
14 lat temu |
|
Cédric Augonnet
|
952534293d
Update copyright date
|
15 lat temu |
|
Nathalie Furmento
|
4a983beb04
Update copyright
|
15 lat temu |
|
Cédric Augonnet
|
e241148129
Clarify the behaviour of starpu_wait_all_tasks with regenarable tasks. A
|
15 lat temu |
|
Nathalie Furmento
|
807432b12c
Internal functions are named _starpu_XXX
|
15 lat temu |
|
Nathalie Furmento
|
78b0247b46
Prefixing ...
|
15 lat temu |
|
Nathalie Furmento
|
3739cfa184
Prefixing of objects defined in starpu_config_h.in
|
15 lat temu |
|
Samuel Thibault
|
52331b4652
do not systematically include sched.h
|
15 lat temu |
|
Cédric Augonnet
|
13dd236182
Keep this function public
|
15 lat temu |
|
Nathalie Furmento
|
3c7a6bade8
Rename cuda functions
|
15 lat temu |
|
Cédric Augonnet
|
95659a6ed2
prefix more internal functions with _starpu
|
15 lat temu |
