Samuel Thibault
|
bfcac01ad1
Avoid emitting progress probes repeatedly, allowing to re-enable them in the trace
|
8 år sedan |
Samuel Thibault
|
9d562f013b
Fix handling memnode state: when several transfers are queued, states should be pushed and poped, otherwise on termination of the first transfer the trace shows the memnode as idle
|
8 år sedan |
Samuel Thibault
|
82cc951ab9
Smooth gflop computations a bit
|
8 år sedan |
Samuel Thibault
|
a2e38ce59c
Set the device ID when running an initialization codelet
|
8 år sedan |
Samuel Thibault
|
ecac7ec55c
Also always set the current CUDA devid when freeing, it is not very costly and makes things more flexible
|
8 år sedan |
Samuel Thibault
|
922ff28a1e
Make datawizard_progress use worker 0 of worker sets as reference for determining which memory nodes to make progress
|
8 år sedan |
Samuel Thibault
|
384397546c
Label the CUDA driver just 'CUDA' one using just one thread
|
8 år sedan |
Samuel Thibault
|
283d125846
fix build without CUDA
|
8 år sedan |
Samuel Thibault
|
7cdeadd447
Add STARPU_CUDA_THREAD_PER_DEV environment variable to support driving all
|
8 år sedan |
Samuel Thibault
|
2ea92eb0fb
Always set the current CUDA devid when allocating, it is not very costly and makes things more flexible
|
8 år sedan |
Samuel Thibault
|
43a3c904ed
directly lookup destination node instead of relying on the node of the current worker
|
8 år sedan |
Samuel Thibault
|
f37e65a6f2
Drop the memory_node thread key, we do not really need it since we can get it from the worker key, and it makes switching more costly
|
8 år sedan |
Samuel Thibault
|
c71f05d00b
fix warning
|
8 år sedan |
Samuel Thibault
|
62fb4d67b8
comment
|
8 år sedan |
Nathalie Furmento
|
422cad39ec
New function starpu_worker_display_names to display the names of all the workers of a specified type.
|
8 år sedan |
Nathalie Furmento
|
ca777699be
tests/main/driver_api/init_run_deinit.c: fix number of workers
|
8 år sedan |
Samuel Thibault
|
a2d3350e43
Fix file leak
|
8 år sedan |
Nathalie Furmento
|
8e8b686bb5
doc: fix text
|
8 år sedan |
Samuel Thibault
|
5a5eb593d1
Always reset the local worker key to worker0 for datawizard to get the proper wait queue when using multiple workers in the same driver thread
|
8 år sedan |
Samuel Thibault
|
a41080d2a6
Fix setting the proper CUDA stream when using several streams from the same driver thread
|
8 år sedan |
Samuel Thibault
|
52d97c6ae9
Fix cublas initialization/shutdown when using one thread per stream with multistream
|
8 år sedan |
Samuel Thibault
|
27cfafb119
Move setting worker_set->nworkers to topology, where it makes more sense
|
8 år sedan |
Samuel Thibault
|
6dbd5add56
Fix indent
|
8 år sedan |
Samuel Thibault
|
d3d32d6451
Fix priorities for schedulers which don't have unbound priority values
|
8 år sedan |
Samuel Thibault
|
ace9f6b23e
Use better priorities in cholesky algorithm
|
8 år sedan |
Samuel Thibault
|
5278ab7d1b
typo
|
8 år sedan |
Samuel Thibault
|
14492fb09a
typo
|
8 år sedan |
Nathalie Furmento
|
461ca6a408
mpi/src/starpu_mpi_cache_stats.c: remove internal array which is not needed
|
8 år sedan |
Nathalie Furmento
|
d8961fb495
mpi: move cache data in the starpu_data_handle_t
|
8 år sedan |
Samuel Thibault
|
6a9541027d
comments
|
8 år sedan |