Field gang #16
Field gang #16
Conversation
…m_parent_module.fypp
…in field_constants_module.F90
…ld_RANKSUFF_gather_module.fypp & field_gathscat_type_module.fypp
…_data_module.fypp
CMakeLists.txt
Outdated
|
|
||
| ## preprocess fypp files | ||
| foreach (SUFF IN ITEMS IM RM RB RD LM) | ||
| string (TOLOWER ${SUFF} suff) | ||
| foreach (RANK RANGE 2 5) | ||
| foreach (FUNC IN ITEMS "" _gathscat _access _util _array_util) | ||
| add_custom_command (OUTPUT field_${RANK}${suff}${FUNC}_module.F90 | ||
| COMMAND ${FYPP} -n -DRANK=${RANK} -DSUFF='${SUFF}' -m os -M ${CMAKE_CURRENT_SOURCE_DIR} -m fieldType |
There was a problem hiding this comment.
Were the fypp line annotations removed intentionally? (the -n arg)
There was a problem hiding this comment.
Yes and no. Much easier to look at real Fortran code when debugging, but we can got back to fypp line indexing if you do not like it.
field_RANKSUFF_module.fypp
Outdated
| ! Mark data dirty on the device: the pointer returned by GET_VIEW may be rw | ||
| IF (IAND (SELF%ISTATUS, NDEVFRESH)==NDEVFRESH) THEN | ||
| SELF%ISTATUS = IAND(SELF%ISTATUS, NOT(NDEVFRESH)) | ||
| CALL SELF%SET_DEVICE_DIRTY () |
There was a problem hiding this comment.
I could be mistaken, but I think here we do not mark the children as dirty on the device if GET_VIEW is called on the gang. I think the children should also be marked as dirty.
There was a problem hiding this comment.
The code has moved, but I think SET_DEVICE_DIRTY will call SET_STATUS, which will also update children status.
|
|
||
| ELSE | ||
|
|
||
| SELECT CASE (IRAND (1, 134)) |
There was a problem hiding this comment.
I'm not sure I understand the use of IRAND here. Are we selecting the case to test based on an RNG, or am I missing something here?
There was a problem hiding this comment.
Yes, we generate random (but reproducible) sequences of operations on the 4D fields and its 3D children. Note it is not feasible to explore all the combinations.
There was a problem hiding this comment.
For the record, I don't think RNG-based testing is a good idea, and I still don't fully understand the need for this here. However, fixing this probably requires a rethink of the test base implementation, which is beyond this PR.
There was a problem hiding this comment.
Thanks a lot @pmarguinaud for your implementation of the FieldGang, along with all the other useful additions! Thanks also for adding the missing CUDAHOSTUNREGISTER. Once the suggested ifdef guards for that have been added, I am happy to merge this contribution.
| DATA = C_LOC (HST (${ ', '.join (map (lambda i: 'LBOUND (HST, ' + str (i) + ')', range (1, ft.rank+1))) }$)) | ||
|
|
||
| IF (PINNED) THEN |
There was a problem hiding this comment.
| IF (PINNED) THEN | |
| #:if defined('CUDA') | |
| IF (PINNED) THEN |
| IF (ISTAT /= 0) THEN | ||
| CALL FIELD_ABORT ("${ft.name}$_OWNER: FAILED TO UNREGISTER PAGE-LOCKED MEMORY") | ||
| ENDIF | ||
| ENDIF |
There was a problem hiding this comment.
| ENDIF | |
| ENDIF | |
| #:endif |
|
|
||
| ELSE | ||
|
|
||
| SELECT CASE (IRAND (1, 134)) |
There was a problem hiding this comment.
For the record, I don't think RNG-based testing is a good idea, and I still don't fully understand the need for this here. However, fixing this probably requires a rethink of the test base implementation, which is beyond this PR.
This branch implements two new classes designed to handle multi-parameter fields such as aerosols. Its purpose is to provide alternative data representation for such fields in the dynamics and in the physics. This is best described in the following presentation:
fieldbuffer.pdf
The two new classes handle :
The implementation relies on a parent/children relationship between a master field of dimension N and children fields of dimensions N-1. All children fields delegate the management of DEVPTR to their parent; the STATUS of the children is forced to be coherent with the status of the parent.
Other minor cleaning/features are included: