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.
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
Add a new omp
GPU layer that uses libomptarget
#20186
Conversation
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
…e compiler Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
This reverts commit cb80cee.
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
Signed-off-by: Engin Kayraklioglu <e-kayrakli@users.noreply.github.com>
For now, we want to first modularize the GPU runtime, so we can add new layers such as one that uses libomptarget as the middleware. We will also be looking at writing our own layer for targeting AMD devices in the short term. PR that will refactor the runtime for easy layer addition: #20358 |
This is a WIP PR that creates a new GPU runtime layer that uses libomptarget.
Currently, for this to work, you'll need the
openmp
project inllvm-src/projects
. The main reason this PR is WIP is that I currently don't have a good strategy for packaging that or finding system's version if it exists.Approach
With this PR, the user can set
CHPL_GPU
toomp
. In that case, we'll uselibomptarget
's__tgt_rtl_*
interface which wraps around different plugins (ie, cuda, amdgpu etc). This interface normally compiles into a shared library, which we dynamically load at runtime.This interface does not expose everything that we were using from the CUDA Driver API. To cover for the missing parts, this PR adds supporting cuda headers which has a couple of
static inline
functions. These headers are included from bothcuda
andomp
layers as appropriate.Runtime structure with two GPU layers
The main GPU interface of the runtime is still in
runtime/includes/chpl-gpu.h
which is implemented inruntime/src/chpl-gpu.c
. However, with this PR, the functions there will call new functions that are inchpl_gpu_impl_*
"namespace". These functions implement core functionality and implemented by different layers whereas Chapel functions like debugging output and diagnostics are still maintained inchpl-gpu.c
, and not duplicated in different layers.