Generate facts on externals by inspecting
packages.yaml. Added rules in concretize.lp
Added extra logic so that external specs
disregard any conflict encoded in the
package.
In ASP this would be a simple addition to
an integrity constraint:
:- c1, c2, c3, not external(pkg)
Using the the Backend API from Python it
requires some scaffolding to obtain a default
negated statement.
Conflict rules from packages are added as integrity
constraints in the ASP formulation. Most of the code
to generate them has been reused from PyclingoDriver.rules
The new concretizer and the old concretizer solve constraints
in a different way. Here we ensure that a SpackError is raised,
instead of a specific error that made sense in the old concretizer
but probably not in the new.
Instead of python callbacks, use cardinality constraints for package
versions. This is slightly faster and has the advantage that it can be
written to an ASP program to be executed *outside* of Spack. We can use
this in the future to unify the pyclingo driver and the clingo text
driver.
This makes use of add_weight_rule() to implement cardinality constraints.
add_weight_rule() only has a lower bound parameter, but you can implement
a strict "exactly one of" constraint using it. In particular, wee want to
define:
1 {v1; v2; v3; ...} 1 :- version_satisfies(pkg, constraint).
version_satisfies(pkg, constraint) :- 1 {v1; v2; v3; ...} 1.
And we do that like this, for every version constraint:
atleast1(pkg, constr) :- 1 {version(pkg, v1); version(pkg, v2); ...}.
morethan1(pkg, constr) :- 2 {version(pkg, v1); version(pkg, v2); ...}.
version_satisfies(pkg, constr) :- atleast1, not morethan1(pkg, constr).
:- version_satisfies(pkg, constr), morethan1.
:- version_satisfies(pkg, constr), not atleast1.
v1, v2, v3, etc. are computed on the Python side by comparing every
possible package version with the constraint.
Computing things like this has the added advantage that if v1, v2, v3,
etc. comprise *all* possible versions of a package, we can just omit the
rules for the constraint under consideration. This happens pretty
frequently in the Spack mainline.
- [x] Solver now uses the Python interface to clingo
- [x] can extract unsatisfiable cores from problems when things go wrong
- [x] use Python callbacks for versions instead of choice rules (this may
ultimately hurt performance)
There are now three parts:
- `SpackSolverSetup`
- Spack-specific logic for generating constraints. Calls methods on
`AspTextGenerator` to set up the solver with a Spack problem. This
shouln't change much from solver backend to solver backend.
- ClingoDriver
- The solver driver provides methods for SolverSetup to generates an ASP
program, send it to `clingo` (run as an external tool), and parse the
output into function tuples suitable for `SpecBuilder`.
- The interface is generic and should not have to change much for a
driver for, say, the Clingo Python interface.
- SpecBuilder
- Builds Spack specs from function tuples parsed by the solver driver.
The original implementation was difficult to read, as it only had
single-letter variable names. This converts all of them to descriptive
names, e.g., P -> Package, V -> Virtual/Version/Variant, etc.
To handle unknown compilers propely in tests (and elsewhere), we need to
add unknown compilers from the spec to the list of possible compilers.
Rework how the compiler list is generated and includes compilers from
specs if the existence check is disabled.
Specs like hdf5 ^mpi were unsatisfiable because we added a requierment
for `node("mpi").`. This can't be resolved because "mpi" is not a
package.
- [x] Introduce `virtual_node()`, which says *some* provider must be in
the DAG.
This adds compiler flags to the ASP solve so that we can have conditions
based on them in the solve. But, it keeps order out of the solve to
avoid unneeded complexity and combinatorial explosions.
The solver determines which flags are on a spec, but the order is
determined by DAG precedence (childrens' flags take precedence over
parents' and are added on the right) and order (order flags were
specified on the command line is respected).
The solver is responsible for determining when to propagate flags, when
to inheit them from other nodes, when to take them from compiler
preferences, etc.
Weight microarchitectures and prefers more rercent ones. Also disallow
nodes where the compiler does not support the selected target.
We should revisit this at some point as it seems like if I play around
with the compiler support for different architectures, the solver runs
very slowly. See notes in comments -- the bad case was gcc supporting
broadwell and skylake with clang maxing out at haswell.
We didn't have a cardinality constraint for multi-valued variants, so the
solver wasn't filling them in.
- [x] add a requirement for at least one value for multi-valued variants
Variants like `cpu_target` on `openblas` don't have defineed values, but
they have a default. Ensure that the default is always a possible value
for the solver.
Spack was generating the same dependency connstraints twice in the output ASP:
```
declared_dependency("abinit", "hdf5", "link")
:- node("abinit"),
variant_value("abinit", "mpi", "True"),
variant_value("abinit", "mpi", "True").
```
This was because `AspFunction` was modifying itself when called.
- [x] fix `AspFunction` so that every call returns a new object
- [x] Add support for packages.yaml and command-line compiler preferences.
- [x] Rework compiler version propagation to use optimization rather than
hard logic constraints
Technically the ASP output order does not matter, but it's hard to diff
two different solve fomulations unless we order it.
- [x] make sure ASP output is emitted in a deterministic order (by
sorting all hash keys)
This needs more thought, as I am pretty sure the weights are not correct.
Or, at least, I'm not convinced that they do what we want in all cases.
See note in concretize.lp.
Solver now prefers newer versions like the old concretizer. Prefer
package preferences from packages.yaml, preferred=True, package
definition, and finally each version itself.
Competition output only prints out one model, so we do not have to
unnecessarily parse all the non-optimal models. We'll just look at the
best model and bring that in.
In practice, this saves a lot of JSON parsing and spec construction time.
Clingo actually has an option to output JSON -- use that instead of
parsing the raw otuput ourselves.
This also allows us to pick the best answer -- modify the parser to
*only* construct a spec for that one rather than building all of them
like we did before.
- Instead of using default logic, handle variant defaults by minimizing
the number of non-default variants in the solution.
- This actually seems to be pretty fast, and it fixes the long-standing
issue that writing this:
spack install hdf5 ^mpich
will fail if you don't specify hdf5+mpi. With optimization and
allowing enums to be enumerated, the solver seems to be able to quickly
discover that +mpi is the only way hdf5 can depend on mpich, and it
forces the switch to be thrown.
