Restore rparWith behavior #36
Conversation
|
Now that I understand runEval (r0 a) = a
runEval (rpar a) = a
runEval (rseq a) = awe need rparWith r0 = r0
rparWith rpar = rpar
rparWith rseq = rpar
rparWith (rparWith strat) = rparWith strat |
Control/Parallel/Strategies.hs
Outdated
| -- rparWith strat x | ||
| -- @ | ||
| -- | ||
| -- will spark a thread to perform @strat x@ in parallel. Note |
There was a problem hiding this comment.
let's not use the terminology "thread" here to avoid confusion. How about "will spark @strat x@".
| -- WHNF, while @rparWith strat@ need not. | ||
| -- | ||
| -- > rparWith r0 = r0 | ||
| -- > rparWith rpar = rpar |
There was a problem hiding this comment.
The LHS creates two sparks, I'm not sure I'd call them equal. It depends what equality you're using, but as I argued elsewhere the only sensible notions of equality here are ones that care about sparks.
There was a problem hiding this comment.
rparWith rpar is actually a quite weird thing: it creates a spark, who creates a spark, who evaluates to WHNF.
However one can probably replace that law with rparWith rseq = rpar.
There was a problem hiding this comment.
How does one replace the other? What we probably actually want is some sort of equivalence relation that equates "essentially equivalent" things. rpar x >>= rseq is effectively the same as rseq x, but I don't know how to express that.
There was a problem hiding this comment.
@simonmar, my intuition suggests that we're never going to find all the relations (Eval is a very large structure), but that we should document some of the more useful ones. It's easy to think you're getting more parallelism than you are when composing things. Speaking of relations,
runEval (m >>= rpar) = runEval m
rpar x >>= \x' -> rpar y >>= f = rpar y >>= \y' -> rpar x >>= f
-- when x' is not free in y and y' is not free in x.
|
I don't think it pays too much to distinguish too much between sparks that
are (almost) guaranteed to fizzle and sparks that don't spark, especially
since I think it reasonable to use rewrite rules to turn the former into
the latter.
|
Control/Parallel/Strategies.hs
Outdated
|
|
||
| data Lift a = Lift a | ||
| #else | ||
| rparWith s a = do l <- rpar (s a); return (case l of Done x -> x) |
There was a problem hiding this comment.
AFAIK, we can also use the GHC >= 7.2 version here, saving some code dupling.
| -- `rseq` in `rparWith`. In particular, we want rparWith r0 = r0, not | ||
| -- rparWith r0 = rpar. | ||
| rparWith s a = do | ||
| l <- rpar r |
There was a problem hiding this comment.
This creates (nearly) always a fizzled spark, because l is immediately pattern matched. A lazy pattern match is needed.
There was a problem hiding this comment.
The pattern match is lazy (it's in the argument to return). Try your test!
There was a problem hiding this comment.
Moreover, if you inline everything manually (including the runRW#) you'll end up with the same thing as the old implementation. We just shouldn't do that because inlining the realWorld# is like asking the optimizer to completely mangle our intentions.
There was a problem hiding this comment.
(FYI, search the bytestring source code for accursedUnutterablePerformIO if you want a more detailed explanation.)
There was a problem hiding this comment.
Sorry I was wrong with my comment.
It does not pass with the current implementation of `rparWith`.
* Lift the result to avoid always reducing to WHNF. * Rewrite the documentation of `rparWith`. Fixes haskell#35
3 participants
rparWith.Fixes #35