 Nathalie Furmento
|
1ae61a95f9
CUDA and OpenCL drivers: when STARPU_CUDA_PIPELINE and STARPU_OPENCL_PIPELINE are set to 0, force synchronous execution of all kernels
|
%!s(int64=10) %!d(string=hai) anos |
 Samuel Thibault
|
7ac2c135ad
protect worker status with the scheduler mutex
|
%!s(int64=10) %!d(string=hai) anos |
|
Samuel Thibault
|
6a0c92daab
Avoid void* arithmetic, icc refuses it on windows...
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
1f1954c43b
Add OpenCL kernel submission pipelining, to overlap costs
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
f313079113
Factorize code to prepare for OpenCL pipelining
|
%!s(int64=11) %!d(string=hai) anos |
|
Nathalie Furmento
|
bafe84486c
fix #13183
|
%!s(int64=11) %!d(string=hai) anos |
|
Nathalie Furmento
|
7e743c9457
backport branches/perfmodel@13175: src/drivers: use a more meaningful name than args for the worker parameter
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
b25760aa94
fix warnings
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
f14e68dc2c
Separate worker and thread state, to see streamed kernel executions
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
778b4c444a
* Add codelet size, footprint and tag id in the paje trace.
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
3903429984
Fix build with simgrid and without opencl
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
7832518f4d
port r12742 from 1.1: Fix OpenCL device number detection when MAXOPENCLDEVS is not big enough
|
%!s(int64=11) %!d(string=hai) anos |
|
Nathalie Furmento
|
7891f63369
src/drivers/opencl/driver_opencl.c: revert #12566 and #12578, and explain why, the function availability would need to be checked at runtime, we do not want to do that
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
d4aeabce96
Also make sure that the header has the declaration
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
d0eab8ef8c
Do not rely on the .h file for functions actually available at link time, do check them explicitly
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
83a38dfa85
clEnqueueMarker is deprecated by clEnqueueMarkerWithWaitList. Be able to use both
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
092f322b1c
Add CUDA concurrent kernel execution support through the STARPU_NWORKER_PER_CUDA environment variable.
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
5a9ffce947
backport r12506: fix disabling out-of-order for data transfers
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
bf8aa6338d
Let the OpenCL driver progress while the GPU is computing
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
8f26d26ad7
Introduce the STARPU_CUDA_ASYNC and STARPU_OPENCL_ASYNC flags. No optimization is done yet for now.
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
8cc1f6c0db
Use separate OpenCL command queues for CPU->GPU, GPU->CPU, and GPU->GPU transfers
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
e4c9c8774a
Fix GPU memory size in the simgrid case
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
5ec6f8b8a7
Drop unused mask parameter of _starpu_fetch_task_input and _starpu_push_task_output
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
b1dfd30389
Store GPU size in platform file
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
e9e4a6361a
Use the same cuda/opencl exclusion mechanism on simgrid, instead of having a bogus assertion
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
6325a34ab3
Move device id lookup into run-driver functions, to avoid paying the cost at each task execution
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
d448281fc3
New STARPU_DISABLE_KERNELS environment variable to disable actual kernel execution.
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
89ac7e59ac
Provide hint for the user about missing func pointer
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
b801a66994
port r 11483 from 1.1: Default to using only 90% of the available GPU memory, to avoid seeing cudaMemset run out of memory..
|
%!s(int64=11) %!d(string=hai) anos |
|
Samuel Thibault
|
4ce5fe9027
port r11357 from 1.1: Fix memleak
|
%!s(int64=11) %!d(string=hai) anos |
