Add state replication and if raising transformations #1639
Conversation
| return False | ||
|
|
||
| # make sure this is not a loop guard | ||
| if len(out_edges) == 2: |
There was a problem hiding this comment.
Application on loops results in the addition of useless states (but the SDFG is still correct). This will not get rid of the loop, meaning that apply_transformations_repeated will never halt.
| if if_guard.is_empty(): | ||
| return False | ||
|
|
||
| # check that condition does not depend on computations in the state |
There was a problem hiding this comment.
might be missing something I do not know about here. Also I don't know if this can result in spurious failures
|
Apart from @acalotoiu 's comment, this looks good to me. However, I want to add: I believe that the presence of more and more niche transformations like this should make us prioritize the idea of a 'community transformation repository' a bit more. |
I really see this as a member of a potential set of "transformation building blocks", towards having composable transformations. This also lies in between a utility method (e.g. in the SDFG class) and an actual transformation |
| sdfg.apply_transformations_repeated([OTFMapFusion]) | ||
|
|
||
| assert len(sdfg.nodes()) == 4 | ||
| assert sdfg.start_state.is_empty() |
There was a problem hiding this comment.
the tests fails here but it works on my machine. Is there something specific to the setup in the pipeline?
pratyai
force-pushed
the
rearrange_sdfg
branch
from
355af5b
to
109eeba
Compare
pratyai
force-pushed
the
rearrange_sdfg
branch
from
109eeba
to
ad097eb
Compare
|
Added one "negative test" (besides the existing one) where the expected behaviour is that the transformation is not applied. Gave each of the two transformations its own test file. PTAL. |
|
Thank you for addressing the comments @pratyai! In general, this looks good to me. However, I would prefer if this PR could be part of a brief discussion in the DaCe weekly this week before it is merged. For that reason I am not pressing 'accept' yet. I would prefer to discuss it due to two key points:
|
This PR adds the StateReplication and IfRaising transformations.
StateReplication performs a normalization step. It takes a state with
nincoming edges and creates an equivalent SDFG withnreplicas of the original state, each with a single incoming edge.IfRaising takes an if guard and 'raises' the evaluation of the conditional. The condition needs to be independent of the computations done in the if guard. The if guard is replicated in each branch and the conditional is evaluated earlier in a new empty state.
The combination of these two transformations unlocks extra dataflow in the case of "diamond" patterns that can be found in some weather codes. In particular, it targets the case where some flag changes what computation is performed.
Example before:
After:
After State, Map, and Tasklet fusion: