Cuda rebase - #6

Open

eddy16112 wants to merge 68 commits into

ICLDisco:mainfrom

eddy16112:cuda-rebase

Commits on Oct 18, 2016

Add GPU packing and unpacking

rolfv
authored and
eddy16112
committed
7a03ffb
indexed datatype new, bonus stask support.
eddy16112
committed
55badab
RDMA send is now working.
eddy16112
committed
ef41551
Add support for vector datatype. Add pipeline.
eddy16112
committed
9bbed91
unrestricted GPU. Instead of forcing everything to go on

bosilca
authored and
eddy16112
committed
fe03183
Using globally defined indexes lead to several synchronization

bosilca
authored and
eddy16112
committed
26f2237
Generate the Makefile. It will now be placed in the bindir

bosilca
authored and
eddy16112
committed
63e10df
Add a patch from Rolf fixing 2 issues:

bosilca
authored and
eddy16112
committed
cf7e185
big changes, now pack is driven by receiver by active message
eddy16112
committed
fd91bb9
intel test working
eddy16112
committed
6d837d1
Upon datatype commit create a list of iovec representing a single

bosilca
authored and
eddy16112
committed
10a6932
contiguous vs non-contiguous is working
eddy16112
committed
44a64e1
now we are able to pack directly to remote buffer if receiver is
eddy16112
committed
f0e8bff
add ddt_benchmark
eddy16112
committed
cf95914
modify for matrix transpose
eddy16112
committed
7a4a10d
enable vector
eddy16112
committed
bc80b3e
receiver now will send msg back to sender for buffer reuse
eddy16112
committed
4d6ebb3
offset instead of actual addess, and lots of clean up for unused
eddy16112
committed
3525419
re-write pipeline
eddy16112
committed
9b9f783
opal_datatype is chnaged, so we need more space

Wei Wu
authored and
eddy16112
committed
e77302e
remove smcuda btl calls from pml ob1
eddy16112
committed
b57f1e5
cuda ddt support is able to turn itself off. Make it support
eddy16112
committed
7cdf09e
move ddt kernel support function pointer into opal_datatype_cuda.c
eddy16112
committed
e3c6d36
support caching datatype

bosilca
authored and
eddy16112
committed
66d30fb
apply loop unroll for pack and unpack kernels
eddy16112
committed
4e3c5d6
fix a cuda event bug. cudaStreamWaitEvent is not blocking call.
eddy16112
committed
f77c382
new vector kernel
eddy16112
committed
5a97315
fix a if CUDA_41 error
eddy16112
committed
8868914
use cuda event to track the completion of pack and unpack
eddy16112
committed
99f03b3
make openib support multi-stream
eddy16112
committed
e214c53
create a btl function to register convertor to registration handle, now we dont need ob1 header in smcuda btl
eddy16112
committed
9a22660
fix a bug: we should also track the completion of unpack operation, but only track D2D.
eddy16112
committed
9b7d28a
use multiple cuda stream for P2P, it allows multiple send/recv working simutaniously
eddy16112
committed
e56bd4f
fix the renaming issue after merge
eddy16112
committed
acc3647
fix a GPU memory leak issue.
eddy16112
committed
dcbd75e
put ompi_datatype_t back to 512 byte, clean up printf and unused functions
eddy16112
committed
612d77e
convertor should be async
eddy16112
committed
489cc8d
revert the ddt_test, will have a separate cuda test later
eddy16112
committed
0017991
set the default of ddt pipeline size to 4M
eddy16112
committed
9b9d5b0
bug fix, set gpu buffer to NULL when init
eddy16112
committed
545091b
fix configuration and slient warnings of datatype padding
eddy16112
committed
052c1ce
minor fix in makefile
eddy16112
committed
ac184dc
more fix in makefile
eddy16112
committed
5819a55
clean up printf
eddy16112
committed
b0a3000
disable ddt cuda test
eddy16112
committed
e016bcb
roll back not use multiple ipc stream
eddy16112
committed
8b85c3d

Commits on Oct 19, 2016

remove unused functions
eddy16112
committed
39ec7ae
more cleanup
eddy16112
committed
f7e0c5f
add a printf
eddy16112
committed
b2b69e5

Commits on Oct 20, 2016

A lot of minor changes.
bosilca
committed
ce34259
minor fix to make it work again
eddy16112
committed
ad4198d

Commits on Oct 21, 2016

add comment for opal_datatype_cuda.cuh, cached device id, remove commented code
eddy16112
committed
42c5a4d
this function is no longer needed
eddy16112
committed
5c2ce9f
merge pack unpack put sig function into one
eddy16112
committed
99ffb84
recheck the opal_datatype_cuda_kernel_support
eddy16112
committed
66ac26a
OPAL_DATATYPE_IOV_UNIFIED_MEM is no longer needed
eddy16112
committed
b47a5cd
clean up cu files
eddy16112
committed
023f489
Remove useless #define. Other minor cleanups.
bosilca
committed
f00ab75
Do not modify the PMIx files (there should be no need).
bosilca
committed
d4a48d1

Commits on Oct 24, 2016

use OPAL_VERBOSE instead of my own DEBUG print
eddy16112
committed
2006b86
Small updates.
bosilca
committed
269067c

Commits on Oct 25, 2016

clean up comments and remove unused define
eddy16112
committed
12f5f83

Commits on Oct 26, 2016

remove NB_GPUS; use mca to set cuda buffer size; use mca to set cuda datatype support enabled or not; check cuda calls.
eddy16112
committed
d30cc73

Commits on Oct 28, 2016

add some protection for the case there is no mem for pack/unpack
eddy16112
committed
fcc9ccb

Commits on Oct 31, 2016

clean up testing
eddy16112
committed
1d91384

Commits on Nov 1, 2016

Commits on Nov 15, 2016

mca_cuda_convertor_init is called in MPI_Init if using pre-connect, so do not init kernel support until confirming buffer is gpu buffer.
eddy16112
committed
7063d19