refactor[next-dace]: New Optimization Scheme in Intra-Map Optimization #2457
Conversation
NOT WORKING: 5.7612245082855225
DOES NOT WORK: 5.77982020
NOT WORKING: 5.89192s
SEEMS WORKING: 4.57165s
…tion. But it also has an additional simplify that was present when TF was run in stage 1, but not in the other version. PERFORMANCE: 4.5106589s
…n_map_dataflow_optimization_order_public
philip-paul-mueller
changed the title
|
It seems that this change introduces a 5% performance penalty in 
|
|
It seems that there is a 5% performance penalty in 
|
|
cscs-ci run |
edopao
changed the title
|
cscs-ci run |
| # things simpler or prevent it from doing certain, negative, things). | ||
| # TODO(phimuell): Restrict it to Tasklets only inside Maps. | ||
| # TODO(phimuell): Investigate more. | ||
| sdfg.apply_transformations_repeated( |
There was a problem hiding this comment.
Can we add an option to disable this to avoid getting the regression in the other stencil as an interim solution until we understand the implications of this transformation?
There was a problem hiding this comment.
Yes we should do it the problem is just that we have to coordinate this, because PR#2450 broke compatibility (the metric package in GT4Py was moved, thus we can not simply make a new release.
There was a problem hiding this comment.
FYI: The staging branch (C2SM/icon4py#1009) has the correct import of the metric package.
By default it is off.
| blocking_only_if_independent_nodes: Optional[bool], | ||
| scan_loop_unrolling: bool, | ||
| scan_loop_unrolling_factor: int, | ||
| compact_tasklets: bool, |
There was a problem hiding this comment.
| compact_tasklets: bool, | |
| fuse_tasklets: bool, |
| assume_pointwise: bool = True, | ||
| optimization_hooks: Optional[dict[GT4PyAutoOptHook, GT4PyAutoOptHookFun]] = None, | ||
| demote_fields: Optional[list[str]] = None, | ||
| compact_tasklets: bool = False, |
There was a problem hiding this comment.
| compact_tasklets: bool = False, | |
| fuse_tasklets: bool = False, |
3 participants
This PR changes how Intra-Map dataflow optimization works. Specifically it adds
TaskletFusionwhich, as its name suggests is able to merge Tasklets together.TODO:
TaskletFusionin Auto Optimizer C2SM/icon4py#1016