Floating point cost values with some heuristics for ExhaustiveSynthesis #399
Conversation
| @@ -50,6 +50,8 @@ namespace { | |||
| static cl::opt<bool> EnableBigQuery("souper-exhaustive-synthesis-enable-big-query", | |||
| cl::desc("Enable big query in exhaustive synthesis (default=false)"), | |||
| cl::init(false)); | |||
| static cl::opt<std::string> CostModel("souper-cost-model", cl::desc("Cost Model"), | |||
| cl::init("default")); | |||
There was a problem hiding this comment.
in the description you should list the possible choices
| using CostFunctionType = std::function<float(Inst *, bool)>; | ||
|
|
||
| float defaultCostFunction(Inst* I, bool IgnoreDepsWithExternalUses) { | ||
| return souper::cost(I, IgnoreDepsWithExternalUses); |
There was a problem hiding this comment.
might as well move the code over here, instead of calling out to somewhere else
There was a problem hiding this comment.
The old synthesis still relies on this.
And having it support new cost models is tricky.
(eg; #395)
| @@ -86,6 +88,60 @@ namespace { | |||
| // experiment with synthesizing at reduced bitwidth, then expanding the result | |||
| // aggressively avoid calling into the solver | |||
|
|
|||
| using CostFunctionType = std::function<float(Inst *, bool)>; | |||
There was a problem hiding this comment.
let's use double instead of float everywhere
| return souper::cost(I, IgnoreDepsWithExternalUses); | ||
| } | ||
|
|
||
| // TODO : Derive these randomly to maximize infer rate with test-infer.sh |
There was a problem hiding this comment.
what does "derive these randomly" mean?
There was a problem hiding this comment.
Pick a plausible cost range and a sample frequency for each instruction.
For all possible cost combinations, run test-infer.sh on the solver/alive directory and pick the combinations with the best infer success rate.
There was a problem hiding this comment.
don't explain it here, explain it in the code
| bool IgnoreDepsWithExternalUses) { | ||
| if (!Visited.insert(I).second) | ||
| return 0; | ||
| if (IgnoreDepsWithExternalUses && I != Root && |
There was a problem hiding this comment.
I think you should leave a comment explaining the rationale for counting things with external uses are zero cost
| int &TooExpensive) { | ||
| if (souper::cost(RHS) < MaxCost) | ||
| void addGuess(Inst *RHS, float MaxCost, std::vector<Inst *> &Guesses, | ||
| float &TooExpensive, CostFunctionType &Cost) { |
There was a problem hiding this comment.
rather than passing the cost function around a lot, investigate whether there's a better place to stash this pointer, for example in one of the context objects
| if (souper::cost(RHS) < MaxCost) | ||
| void addGuess(Inst *RHS, float MaxCost, std::vector<Inst *> &Guesses, | ||
| float &TooExpensive, CostFunctionType &Cost) { | ||
| if (Cost(RHS, false) < MaxCost) |
There was a problem hiding this comment.
when passing a boolean constant, follow the convention of including the name of the flag as a comment
| CostFunctionType Cost; | ||
| if (CostModel == "default") { | ||
| Cost = defaultCostFunction; | ||
| } else { |
There was a problem hiding this comment.
the else clause needs to be an error
|
I agree that we want pluggable cost functions, but do you need to add a new one right now? it is very similar to the existing one, so maybe the existing one simply wants to be replaced (while also leaving the pluggable infrastructure so that we can plug in more sophisticated stuff later)? |
| @@ -0,0 +1,11 @@ | |||
| ; REQUIRES: solver, solver-model | |||
|
|
|||
| ; RUN: %souper-check %solver -reinfer-rhs -souper-exhaustive-synthesis -souper-exhaustive-synthesis-num-instructions=1 -souper-cost-model=heuristics %s > %t | |||
There was a problem hiding this comment.
this isn't a great cost model test, since many different cost models would allow this optimization to go.
please try to find a place in our unit tests where you can test the cost model directly, including whether it properly deals with the costs of external uses.
There was a problem hiding this comment.
The default one (without ignoring cost) does not infer this.
There was a problem hiding this comment.
it's not a good test case and wants to be improved
I think it makes sense to keep the existing one because the CEGIS synthesis needs it. |
|
it's fine to leave the old one if you want, but make the other changes |
Implements some infrastructure to support richer cost models.