compiler: Add fission-for-pressure subpass #2114
Conversation
Codecov Report
@@ Coverage Diff @@
## master #2114 +/- ##
===========================================
- Coverage 87.77% 61.96% -25.81%
===========================================
Files 221 222 +1
Lines 39003 39107 +104
Branches 5067 5081 +14
===========================================
- Hits 34233 24232 -10001
- Misses 4208 14163 +9955
- Partials 562 712 +150
... and 129 files with indirect coverage changes Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. |
| continue | ||
|
|
||
| # At this point we know we want to fission. But can we? | ||
| for dep in c.scope.d_flow.independent(): |
There was a problem hiding this comment.
Why not check this first and avoid all the retrieve/sets/...
| ns = len(functions_shared) | ||
|
|
||
| if not ns: | ||
| ns = .001 |
There was a problem hiding this comment.
I guess safe to assume we won't have over 1k function
There was a problem hiding this comment.
Maybe add some comment for this, even though not really necessary? Just to justify this option
| ispace = c.ispace.lift(d) | ||
| processed.append(c.rebuild(exprs=g1, ispace=ispace)) | ||
|
|
||
| break |
There was a problem hiding this comment.
So at most one fission? Could check the remainder g can be split too not sure of what would want
processed.append(c.rebuild(exprs=g0))
processed.append(fission_for_pressure(c.rebuild(exprs=g1, ispace=ispace)))
| pass | ||
|
|
||
|
|
||
| class IntelGoldenCode(Intel64): |
| ns = len(functions_shared) | ||
|
|
||
| if not ns: | ||
| ns = .001 |
There was a problem hiding this comment.
Maybe add some comment for this, even though not really necessary? Just to justify this option
| t = grid.stepping_dim | ||
| def test_nofission_as_unprofitable(self): | ||
| """ | ||
| Test there's no fission if not gonna increase number of collapsable loops. |
| assert_structure(op, ['t,x,y', 't,x,y'], 't,x,y,y') | ||
| def test_fission_partial(self): | ||
| """ | ||
| Test there's no fission if not gonna increase number of collapsable loops. |
So far, this isn't enabled by default. More experimentation required
This PR also adds a new compiler flag to GNUCompiler when running on Skylake and later processors. I've tested it in a few places, and it seems to generally improve things. I'd like people to reproduce it. Note that you need to ensure DEVITO_PLATFORM is properly set to skx, clk, ... the autodetection mechanism, with the recent proliferation of intel architectures, doesn't seem to be as good as it was.