OpenMP GPU porting.

.github/workflows/run-one.yml

@@ -22,7 +22,7 @@ jobs:
             flags: "FFLAGS_OPT=1 OPENMP=1 LIBS+=-lfftw3_threads"
       fail-fast: false # ensures that jobs run in parallel
     env:
-      SKIP_BUILD: ${{ inputs.compiler != 'GNU' && matrix.mode.name == 'opt-omp' && false }} # skipped combinations (disabled)
+      SKIP_BUILD: ${{ matrix.mode.name == 'opt-omp' }} # skipped combinations
     name: build-${{ matrix.mode.name }}
     steps:
     - name: Checkout code

README.md

@@ -23,6 +23,8 @@ P. Costa. *A FFT-based finite-difference solver for massively-parallel direct nu
 ### _Major Update:_ `CaNS 3.0` _is out!_ :tada:
 See the [Release Notes](https://github.com/CaNS-World/CaNS/releases/tag/v3.0.0) for more details.
 
+**[11/08/2025]:** Support for an OpenMP GPU backend is available! See the updated `[docs/INFO_COMPILING.md](docs/INFO_COMPILING.md)` for more details. Note: OpenMP support for shared-memory parallelization on multiple CPUs is not supported anymore. 
+
 **[06/08/2025]:** Support for running on AMD-based supercomputers and new GPU communication backend available! CaNS has been ported using to other platforms using HIP and thanks to the recently developed [diezDecomp library](https://github.com/Rafael10Diez/diezDecomp). See the updated `[docs/INFO_COMPILING.md](docs/INFO_COMPILING.md)` for more details.
 
 **[10/04/2025]:** The writing of checkpoint files has changed. To allow for more flexibility, CaNS now writes one file per scalar field, where each velocity component, pressure, and scalar fields is stored in a different checkpoint file.

configs/defaults/flags-default.mk

@@ -6,16 +6,33 @@ FFLAGS_MOD_DIR := -module # extra space
 endif
 ifeq ($(strip $(FCOMP)),NVIDIA)
 FFLAGS_MOD_DIR := -module # extra space
+override FFLAGS += -cuda -gpu=cc70,cc80,cc90
 ifeq ($(strip $(GPU)),1)
-override FFLAGS += -acc -cuda -Minfo=accel -gpu=cc60,cc70,cc80
+ifneq ($(filter $(GPU_BACKEND),OACC OMP),$(GPU_BACKEND))
+override FFLAGS += -acc -Minfo=accel
+endif
+ifeq ($(strip $(GPU_BACKEND)),OACC)
+override FFLAGS += -acc -Minfo=accel
+endif
+ifeq ($(strip $(GPU_BACKEND)),OMP)
+override FFLAGS += -mp=gpu -Minfo=mp
+endif
 endif
 endif
 ifeq ($(strip $(FCOMP)),CRAY)
 FFLAGS_MOD_DIR := -I./build -ef -J
 ifeq ($(strip $(GPU)),1)
-override FFLAGS += -hacc
+ifneq ($(filter $(GPU_BACKEND),OACC OMP),$(GPU_BACKEND))
+override FFLAGS += -hacc -hnoomp
+endif
+ifeq ($(strip $(GPU_BACKEND)),OACC)
+override FFLAGS += -hacc -hnoomp
+endif
+ifeq ($(strip $(GPU_BACKEND)),OMP)
+override FFLAGS += -homp -hnoacc
+endif
 else
-override FFLAGS += -hnoacc
+override FFLAGS += -hnoacc -hnoomp
 endif
 endif
 ifeq ($(strip $(FCOMP)),FUJITSU)

docs/INFO_COMPILING.md

@@ -24,7 +24,8 @@ In this file, `FCOMP` can be one of `GNU` (`gfortran`), `INTEL` (`ifort`), `NVID
 
  * `SINGLE_PRECISION` : calculation will be carried out in single precision (the default precision is double)
  * `GPU`              : enable GPU accelerated runs
- * `USE_DIEZDECOMP`   : use [diezDecomp](https://github.com/Rafael10Diez/diezDecomp) as GPU communication backend instead of the default (cuDecomp), e.g., for portability in different accelerators. While diezDecomp supports CPU-CPU communication, this is not used in CaNS yet.
+ * `USE_DIEZDECOMP`   : use [diezDecomp](https://github.com/Rafael10Diez/diezDecomp) as GPU communication backend instead of the default (cuDecomp), e.g., for portability in different accelerators; while diezDecomp supports CPU-CPU communication, this is not used in CaNS yet
+ * `GPU_BACKEND`      : backend for directives-based GPU offloading: `OACC` for OpenACC (default), and `OMP` for OpenMP (OpenMP support for shared-memory parallelization on CPU is not supported anymore)
 
 Typing `make libs` will build the 2DECOMP&FFT/cuDecomp/diezDecomp libraries; then typing `make` will compile the code and copy the executable `cans` to a `run/` folder; `make run` will also copy the default input files `*.in` under `src/` to the same `run/` folder.