dc_patterns

Standalone, dependency-free do concurrent portability reproducers. Each mini-app is built from ≥2 separately-compiled TUs and links nothing but the Fortran + offload runtime. for the design.

Concept

do concurrent does the compute in every kernel. Only the data-mapping directive layer changes per compiler, selected by DATA:

DATA	layer	for
omp	!$omp target enter data / declare target	NVHPC, AMD (portable default)
acc	!$acc enter data / routine seq	Cray/CCE; NVHPC cross-check
none	none — bare DC on the host	gfortran/ifx serial reference

Build & run

Edit config.mk for your machine (FC, FC_FLAGS, DATA, MODFLAG) — the NVIDIA/NVHPC block is pre-filled — or override on the CLI:

# one pattern, default (NVHPC, omp data):
make -C patterns/p01_acc_routine_xtu run

# NVHPC, OpenACC data layer:
make -C patterns/p01_acc_routine_xtu DATA=acc \
     FC_FLAGS='-O2 -Kieee -stdpar=gpu -gpu=cc70 -acc' run

# gfortran host reference:
make -C patterns/p01_acc_routine_xtu FC=gfortran FC_FLAGS=-O2 \
     MODFLAG='-J build' DATA=none run

# Cray/CCE: fill the ftn block in config.mk, then:
make -C patterns/p01_acc_routine_xtu FC=ftn FC_FLAGS='...' \
     MODFLAG='-J build' DATA=acc run

# everything:
make run            # uses config.mk

Each exe prints one line:

RESULT <name> data=<omp|acc|host> n=.. iters=.. status=PASS|FAIL max_abs_err=.. mcells_per_s=..

PASS ⇒ exit 0; FAIL ⇒ non-zero (so a matrix runner / CI can gate).

Catalog

• patterns/ — single-hazard reproducers (p01 = cross-module device routine, …).
• models/ — representative miniatures of real kernel classes (continuity, remap, hor_visc, vert_visc, kappa_shear, barotropic).