Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

CUDA exception handling #4095

Merged
merged 22 commits into from
Jan 29, 2021
Merged

CUDA exception handling #4095

merged 22 commits into from
Jan 29, 2021

Conversation

jngrad
Copy link
Member

@jngrad jngrad commented Jan 15, 2021 •
edited
Loading

Description of changes:

  • convert CUDA error codes into runtime errors using the message associated to the error code
  • properly handle CUDA errors by halting the flow of the program instead of ignoring them (Barnes-Hut, LB GPU)
  • remove superfluous CUDA global variables (partial fix for Remove global variables #2628)
  • restore cuda_gather_gpus() functionality and convert the device_list() getter to a regular function (API change)
  • remove unused libcuda dependency (fixes -lcuda needs to be removed for cuda-11.2 #4085)

@jngrad
python: Cleanup Cython bindings to CUDA functions
e2e7952
@jngrad
core: Make Barnes-Hut fields private
537206d
@jngrad
core: Rename CUDA header file
5b8e575 
This header file can only be used in CUDA source files.
@jngrad
core: Introduce exception mechanism for CUDA code
11d9489 
Utility CUDA functions now throw a runtime error wrapping the CUDA
error message when a primitive CUDA function returns an error code.

The LB GPU checking function gpu_init_particle_comm() was rewritten
to check the currently selected GPU instead of the one with ID 0.
The warning about insufficient compute capability was converted
to an error message. The warning about default GPU ID was removed.
The Python-specific message was removed. The function now calls
std::abort() upon any error since we cannot recover from it in
the Python interface (the list of actors cannot be cleared due to
a mismatch between `Actor` and `Actors` regarding active actors).

The dipolar Barnes-Hut code now exits early upon any CUDA error to
avoid undefined behavior in memory allocation on the GPU.
@jngrad
core: Remove unused CUDA global variables
fe8ddaa
@jngrad
core: Simplify code involving EspressoGpuDevice
d8aeff0 
Also fix the type of the memory field to avoid integer overflow.
@jngrad
python: Add bindings for cuda_gather_gpus()
5003b3d
@jngrad
core: Modernize MPI function
a915aa8
@jngrad
core: Improve doxygen documentation
c49f7a6 
Convert regular comments to doxygen comments where appropriate,
document function arguments, remove duplicate doxygen blocks.
@jngrad jngrad added this to the Espresso 4.2 milestone Jan 15, 2021
@jngrad jngrad removed the BugFix label Jan 15, 2021
KaiSzuttor
KaiSzuttor reviewed Jan 25, 2021
View reviewed changes
src/core/cuda_init.cpp Show resolved Hide resolved
src/core/cuda_utils.cuh Outdated Show resolved Hide resolved
src/core/cuda_utils.cuh Outdated Show resolved Hide resolved
src/python/espressomd/cuda_init.pyx Show resolved Hide resolved
testsuite/python/gpu_availability.py Outdated Show resolved Hide resolved
@jngrad jngrad self-assigned this Jan 29, 2021
@jngrad jngrad added the automerge Merge with kodiak label Jan 29, 2021
@kodiakhq kodiakhq bot merged commit e4d6ecc into espressomd:python Jan 29, 2021
@jngrad jngrad deleted the cuda-exception-handling branch January 29, 2021 14:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@KaiSzuttor KaiSzuttor KaiSzuttor approved these changes

Assignees

@jngrad jngrad

Labels
ApiChange automerge Merge with kodiak CodeQuality Core Documentation Python
Projects
None yet
Milestone
Espresso 4.2.0
Development

Successfully merging this pull request may close these issues.

-lcuda needs to be removed for cuda-11.2
2 participants
@jngrad @KaiSzuttor