 Cédric Augonnet
|
d72cd0b547
Avoid to use a quadratic algorithm when matching MPI communications within
|
15 년 전 |
|
Cédric Augonnet
|
89890c1716
fix the computation of the trace offset
|
15 년 전 |
|
Cédric Augonnet
|
e64bb38448
display the MPI transfers between multiple nodes
|
15 년 전 |
|
Cédric Augonnet
|
0824f64dbb
Generate a unique key during starpu_mpi_init so that we can make sure that all
|
15 년 전 |
 Samuel Thibault
|
a4f14ca856
MPI types should always be committed before use
|
15 년 전 |
|
Cédric Augonnet
|
6983263fed
Add a synchronization point during the initialization of the MPI lib so that we
|
15 년 전 |
|
Cédric Augonnet
|
ca41b1537d
We do not wait until the termination of the program (with atexit) to actually
|
15 년 전 |
|
Cédric Augonnet
|
c003625624
Forgot one file using STARPU_MAXCUDADEVS too.
|
15 년 전 |
|
Cédric Augonnet
|
88147be9d1
Rename the MAXCUDADEVS constant into STARPU_MAXCUDADEVS and add an option to
|
15 년 전 |
|
Cédric Augonnet
|
d1820bf537
fix previous commit
|
15 년 전 |
|
Cédric Augonnet
|
1ba632d181
Small cleanups in the public headers
|
15 년 전 |
|
Cédric Augonnet
|
1fd39a00de
Make it possible to generate a single Paje trace with multiple FxT traces by
|
15 년 전 |
|
Cédric Augonnet
|
f844d8e3e6
forgot to add a new file
|
15 년 전 |
|
Cédric Augonnet
|
6b3be68abf
Start to cleanup the tool to generate Paje traces out of FxT traces
|
15 년 전 |
|
Cédric Augonnet
|
4765071b06
Remove dead code, we don't need to generate traces in SVG format since Vite is
|
15 년 전 |
|
Cédric Augonnet
|
95659a6ed2
prefix more internal functions with _starpu
|
15 년 전 |
|
Cédric Augonnet
|
d3acc00120
To avoid problems when linking dynamically, we prefix some internal functions
|
15 년 전 |
|
Cédric Augonnet
|
15831ca9a1
In the case of a non blocking operation, we don't use a local variable to store
|
15 년 전 |
|
Cédric Augonnet
|
f17de349e8
- Export the headers of the MPI lib when installing
|
15 년 전 |
|
Cédric Augonnet
|
d4c070f3ee
Add a --with-mpicc option to specify which compiler should be used with MPI.
|
15 년 전 |
|
Cédric Augonnet
|
d10b14ab3c
Callbacks are sometimes executed directly by the application threads.
|
15 년 전 |
|
Cédric Augonnet
|
9ee0246b6e
- Bug fix in the block interface (the block size of nx*ny*nz, and not nx*nx*ny)
|
15 년 전 |
|
Cédric Augonnet
|
a95033f93a
Replace cublas calls with plain cuda functions in the BLAS, CSR and BCSR
|
15 년 전 |
|
Cédric Augonnet
|
2927c638b3
Don't use CUBLAS function to perform data transfers with the vector interface.
|
15 년 전 |
|
Cédric Augonnet
|
e86adf65df
Having a "request" associated to a detached call does not make much sense as it
|
15 년 전 |
|
Cédric Augonnet
|
dde612872a
Provide helper functions which tell the MPI lib to release a tag when an
|
15 년 전 |
|
Cédric Augonnet
|
14c8458f90
Provide helpers for the "block" interface.
|
15 년 전 |
|
Cédric Augonnet
|
f635a0a3db
Do no liberate the job structure too early.
|
15 년 전 |
|
Cédric Augonnet
|
d924344bf6
bug fix: that value was used uninitialized
|
15 년 전 |
|
Cédric Augonnet
|
be6b0b16e5
Correct some path which was modified by mistake during previous commit.
|
15 년 전 |
