Use CUDA unified memory for Random123 state #595
Conversation
|
Logfiles from GitLab pipeline #11298 (:no_entry:) have been uploaded here! Status and direct links:
|
olupton
force-pushed
the
olupton/random123-unified-memory
branch
from
ce4bc03 to
8576f8d
Compare
|
Logfiles from GitLab pipeline #11336 (:no_entry:) have been uploaded here! Status and direct links:
|
|
Logfiles from GitLab pipeline #11367 (:white_check_mark:) have been uploaded here! Status and direct links:
|
olupton
force-pushed
the
olupton/random123-unified-memory
branch
from
5375bbd to
f8cc054
Compare
|
Logfiles from GitLab pipeline #11394 (:white_check_mark:) have been uploaded here! Status and direct links:
|
Codecov Report
@@ Coverage Diff @@
## master #595 +/- ##
==========================================
- Coverage 56.46% 56.45% -0.02%
==========================================
Files 99 100 +1
Lines 8021 8030 +9
==========================================
+ Hits 4529 4533 +4
- Misses 3492 3497 +5
Continue to review full report at Codecov.
|
olupton
force-pushed
the
olupton/random123-unified-memory
branch
from
f8cc054 to
fdc8aaf
Compare
|
Logfiles from GitLab pipeline #11415 (:no_entry:) have been uploaded here! Status and direct links:
|
pramodk
left a comment
pramodk
left a comment
There was a problem hiding this comment.
Once channel-benchmark is validated to run with a single exe on CPU as well as GPU, this is good to go!
|
Retest this please |
|
It was tested locally that Note that a bug in the NVHPC compiler version 21.3 and 21.5 means version 21.2 (or older?) or 21.7 (or hopefully newer!) must be used. |
|
Logfiles from GitLab pipeline #11636 (:no_entry:) have been uploaded here! Status and direct links:
|
olupton
force-pushed
the
olupton/random123-unified-memory
branch
from
fdc8aaf to
769352a
Compare
|
Logfiles from GitLab pipeline #11650 (:white_check_mark:) have been uploaded here! Status and direct links:
|
|
Logfiles from GitLab pipeline #11694 (:no_entry:) have been uploaded here! Status and direct links:
|
|
Replayed rxd-mpi tests and they passed - https://bbpgitlab.epfl.ch/hpc/coreneuron/-/jobs/38730 (in previous CI there was timeout) |
* Unified Memory in Random123. * Use OMP_Mutex instead of std::atomic, drop detail. * Load gcc module for CUDA in Jenkins CI. CoreNEURON Repo SHA: BlueBrain/CoreNeuron@864b712
Description
This modifies how we allocate state variables for Random123 in GPU-enabled builds. This should simplify things, and enable use-cases like in #528 where we need to manipulate this state on the device from CPU code in ways that the existing API did not allow.
Because the state is now accessible from both CPU and GPU code, this also closes #345.
How to test this?
Build and run with/without GPU support enabled.
Test System
Use certain branches for the SimulationStack CI
CI_BRANCHES:NEURON_BRANCH=olupton/random123-unified-memory,