 Samuel Thibault
|
0423864961
port r11516 from 1.1: considerably reduce the amount of requests submitted at the same time. Pushing more does not seem to really improve performance, and on the contrary increases the latency of possibly urging requests
|
il y a 12 ans |
|
Samuel Thibault
|
f4c7eb0c8e
port r11378 from 1.1: Do not allocate data on the critical path
|
il y a 12 ans |
|
Samuel Thibault
|
e9c70aac17
port r11376 from 1.1: Introduce rwlock_tryrdlock and trywrlock variants. Use them in the progression hooks to avoid getting stuck due to idling workers.
|
il y a 12 ans |
|
Samuel Thibault
|
5f1af2ec79
port r11140 from 1.1: When prefetching, do not try too hard to lock a data handle, better do something else instead
|
il y a 12 ans |
|
Samuel Thibault
|
b217a5399a
data_requests[i] may change, it is still fine for _starpu_handle_node_data_requests to read it unprotected
|
il y a 12 ans |
|
Samuel Thibault
|
e5bea6f293
port r11205 from trunk: Do not push too many data requests at the same time, CUDA would actually enter synchronous mode...
|
il y a 12 ans |
|
Samuel Thibault
|
edce30c7f0
port r11194 from 1.1: Also disable checking _head of renewed request lists
|
il y a 12 ans |
|
Samuel Thibault
|
aea35461df
Document each racy spot normally reported by helgrind
|
il y a 12 ans |
|
Samuel Thibault
|
0d1b0375b3
backport r11155 from 1.1: Disabling access to a variable locally does not actually seem to prevent helgrind from raising the race issues, on the contrary it prevents it from recording where the race comes from... Simply just disable checking these variables completely.
|
il y a 12 ans |
|
Samuel Thibault
|
54460ab919
port r11138 from 1.1: Do not lie to helgrind about having taken a mutex, it produces erroneous races inside mutexes. Rather make helgrind ignore the precise memory accesses we know we do not need to care about
|
il y a 12 ans |
 Thibaud Lambert
|
3b7abcc2cf
mic: Fix bugs
|
il y a 12 ans |
|
Samuel Thibault
|
423529392d
Do not let StarPU queue a request on a node which doesn't have any worker
|
il y a 12 ans |
|
Samuel Thibault
|
fa47d65684
Make sure we don't post disk-related requests to a disk node. This would have happened for a gpu-disk transfer.
|
il y a 12 ans |
|
Samuel Thibault
|
a3f27f08ad
Make some warnings clearer. Avoid a very odd warning when the codelet has more parameters than allowed
|
il y a 12 ans |
 Nathalie Furmento
|
30860cac02
Renaming of diverse functions and datatypes.
|
il y a 12 ans |
|
Nathalie Furmento
|
786b8fb4b7
move _STARPU_PTHREAD_XXX macros to public API starpu_thread_util and rename them to STARPU_PTHREAD_XXX
|
il y a 12 ans |
|
Nathalie Furmento
|
7f792e3082
Define new public API starpu_pthread which is similar to the pthread API.
|
il y a 13 ans |
|
Samuel Thibault
|
d95e1c8398
explain helgrind about some races we are fine with
|
il y a 13 ans |
|
Samuel Thibault
|
add756f1af
Note about racyness. We can't really add a suppression for this case, it would hide potential other races
|
il y a 13 ans |
|
Samuel Thibault
|
55a6fa0d9e
complete lock comment
|
il y a 13 ans |
|
Samuel Thibault
|
38d8bb35e3
Add missing support for mpi_insert_task cases
|
il y a 13 ans |
|
Nathalie Furmento
|
c7321a76ff
fix type for node to always be unsigned (instead of uint32_t)
|
il y a 13 ans |
|
Nathalie Furmento
|
b770541f52
src: memory_nodes function renaming
|
il y a 13 ans |
|
Nathalie Furmento
|
eb32a9b3d4
src: fix coding style
|
il y a 13 ans |
|
Nathalie Furmento
|
edc236df55
memory stats: group all functions in new file src/datawizard/memstats.c and rename them to follow pattern starpu_memory_...
|
il y a 13 ans |
|
Samuel Thibault
|
242b6116cf
pass pthread types through typedefs
|
il y a 13 ans |
|
Nathalie Furmento
|
4ba3bb5eab
memory stats: rename memory_status objects to memory_stats objects as what is displayed are statistics
|
il y a 13 ans |
|
Samuel Thibault
|
e0242fa9ce
optimize holding request list locks
|
il y a 13 ans |
|
Samuel Thibault
|
9badf4f117
Move unregistering tasks from implicit dependencies lists from
|
il y a 13 ans |
|
Samuel Thibault
|
26d2804cd4
Check that the handle spinlock is held when manipulating refcnts
|
il y a 14 ans |
