Inheritance diagram for pyfortool.openacc.Openacc:
Public Member Functions | |
| removeACC (self) | |
| removebyPassDOCONCURRENT (self) | |
| craybyPassDOCONCURRENT (self) | |
| allocatetoHIP (self) | |
| addACCData (self) | |
| addACCRoutineSeq (self, stopScopes) | |
Detailed Description
OpenACC directives transformation methods. Provides utilities for adding, removing, and transforming OpenACC directives for GPU acceleration.
Definition at line 11 of file openacc.py.
Member Function Documentation
◆ addACCData()
| pyfortool.openacc.Openacc.addACCData | ( | self | ) |
Add !$acc data directives for GPU data transfer.
For each subroutine, inserts:
- !$acc data present (array1, array2, ...) after declarations
- !$acc end data at the end of the routine
Only affects INTENT arrays (IN, OUT, INOUT).
Examples
--------
>>> pft = PYFT('gpu_code.F90')
>>> pft.addACCData()
Definition at line 244 of file openacc.py.
Here is the call graph for this function:
◆ addACCRoutineSeq()
| pyfortool.openacc.Openacc.addACCRoutineSeq | ( | self, | |
| stopScopes | |||
| ) |
Add !$acc routine seq directive to subroutines.
Parameters
----------
stopScopes : list of str
Scope paths where to stop adding directives.
Examples
--------
>>> pft = PYFT('gpu_code.F90')
>>> pft.addACCRoutineSeq(['module:MOD/sub:SUB'])
# Adds !$acc routine (SUB) seq to SUB subroutine
Definition at line 301 of file openacc.py.
Here is the call graph for this function:
◆ allocatetoHIP()
| pyfortool.openacc.Openacc.allocatetoHIP | ( | self | ) |
Convert ALLOCATE/DEALLOCATE to HIP versions for AMD GPUs.
Transforms memory allocation calls for variables that are:
- Sent to GPU via !$acc enter data copyin
- Created on GPU via !$acc enter data create
Required for AMD MI250X GPU managed memory usage (e.g., Adastra cluster).
Examples
--------
>>> pft = PYFT('gpu_code.F90')
>>> pft.allocatetoHIP()
# ALLOCATE(X) becomes CALL MNH_HIPALLOCATE(X)
Definition at line 148 of file openacc.py.
Here is the call graph for this function:
◆ craybyPassDOCONCURRENT()
| pyfortool.openacc.Openacc.craybyPassDOCONCURRENT | ( | self | ) |
Handle CRAY compiler vectorization issues with GPU directives. On compute kernels with !$acc loop independent collapse(X) directives: - If BR_ functions are used: removes !$acc loop collapse and uses DO CONCURRENT - If !$mnh_undef(OPENACC) is present: removes !$acc loop collapse - If !$mnh_undef(LOOP) is present: converts nested loops to DO CONCURRENT Notes ----- Works around CRAY compiler vectorization bugs with BR_ (bit-reproducibility) functions and OpenACC directives.
Definition at line 53 of file openacc.py.
Here is the call graph for this function:
◆ removeACC()
| pyfortool.openacc.Openacc.removeACC | ( | self | ) |
Remove all OpenACC directives.
Examples
--------
>>> pft = PYFT('gpu_code.F90')
>>> pft.removeACC()
Definition at line 20 of file openacc.py.
Here is the call graph for this function:
◆ removebyPassDOCONCURRENT()
| pyfortool.openacc.Openacc.removebyPassDOCONCURRENT | ( | self | ) |
Remove MNH OpenACC bypass macros for non-Cray compilers. Removes the following directives: - !$mnh_undef(LOOP) - !$mnh_undef(OPENACC) - !$mnh_define(LOOP) - !$mnh_define(OPENACC)
Definition at line 33 of file openacc.py.
Here is the call graph for this function:
The documentation for this class was generated from the following file:
- /home/runner/work/pyfortool/pyfortool/src/pyfortool/openacc.py
