compiler: Improve quality of generated code #2263
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
FabioLuporini
force-pushed
the
better-codegen
branch
from
c7f0cb9 to
2cbc4f4
Compare
Codecov ReportAttention:
Additional details and impacted files@@ Coverage Diff @@
## master #2263 +/- ##
==========================================
- Coverage 86.91% 86.80% -0.12%
==========================================
Files 229 229
Lines 42156 42296 +140
Branches 7787 7830 +43
==========================================
+ Hits 36641 36715 +74
- Misses 4865 4921 +56
- Partials 650 660 +10 ☔ View full report in Codecov by Sentry. |
| @@ -1253,3 +1281,16 @@ def generate(self): | |||
| if self.cast: | |||
| tip = '(%s)%s' % (self.cast, tip) | |||
| yield tip | |||
|
|
|||
|
|
|||
| def sorted_efuncs(efuncs): | |||
There was a problem hiding this comment.
Might be simpler to define that in efunc to avoid the local import
There was a problem hiding this comment.
circular import, can't import efunc from visitors
| v = priority[cls] | ||
| break | ||
| else: | ||
| v = 0 |
There was a problem hiding this comment.
WHy no just a priority.get(i.__class__, 0)
There was a problem hiding this comment.
if it's here it's because we can't infer anything from priority?
There was a problem hiding this comment.
No I mean instead of the whole for/else
There was a problem hiding this comment.
oh I see. Basically, simply because we want to hit subclasses too, which is essentially the point of this utility
| @@ -36,7 +36,7 @@ def __init__(self, *functions, op=dv.mpi.MPI.SUM, dtype=None): | |||
| def __enter__(self): | |||
| i = dv.Dimension(name='mri',) | |||
| self.n = dv.Function(name='n', shape=(1,), dimensions=(i,), | |||
| grid=self.grid, dtype=self.dtype) | |||
| grid=self.grid, dtype=self.dtype, space='host') | |||
There was a problem hiding this comment.
because this one really needs to be on the host as it will ultimately be accessed in user-land
There was a problem hiding this comment.
it's not exceptionally elegant, admittedly
| # Attempt simplifcation | ||
| # E.g., `FLOAT(32) -> 32.0` of type `sympy.Float` | ||
| try: | ||
| return sympify(eval(cls._base_typ)(base)) |
There was a problem hiding this comment.
THis will loose accuracy for double but not sure what's the easier way as there is only float and int really defined as python types
There was a problem hiding this comment.
no unless there's an 'F' in C all numbers with digits are double, so we should be good
FabioLuporini
force-pushed
the
better-codegen
branch
from
e0b27ed to
4dfc228
Compare
This PR, and its sibling in PRO, dramatically improve the quality of the generated code when running production-level kernels. This is mostly evident in PRO, when compression, serialization, mpi, and data streaming are all used in combination in the same Operator. A lot of code is now nicely abstracted away into routines, which significantly decreases the final code size.