Fix API documentation build problems

Replace deprecated new_hash

…versionning

Co-authored-by: David A. Ham <[email protected]>

Add zero kernel cache

Test code on merge with master

Decorator version 5 changes the API in ways that break the
argument type checking in PyOP2. We're therefore pinning on the
version that works.

Make PETSc and petsc4py for CI use Firedrake not pip

actions: Build docs at end of successful test run

* Change requirements on loopy to point at the loopy sprint branch.

* Replace lp.register_callable_kernel with lp.merge

knl = lp.register_callable_kernel(knl, callee) is now
knl = lp.merge([knl, callee]) .
Reason given for loopy breaking back-compat: earlier they had a notion
of a special "root_kernel" which only permitted one level of nesting
loopy kernel calls and that was too restrictive.

* Loopy callables: adapt to new API for registering.

* Loopy Callables: register callables on a kernel not functions.

* name is static property on SolveCallable and INVCallable

* Loopy callables: register the callables rather than functions and remove the lookup functions.

* Some progress

* Loopy callables: Adapt API of with_types.

* INVCallable has name "inverse" in loopy rep

* Always pass a target when generating loopy kernels, otherwise we cannot merge them with wrapper kernels.

* flake8

* Maybe generate call to inner kernel correctly

Still doesn't work later.

* Revert "Maybe generate call to inner kernel correctly"

This reverts commit 009ca19f3fb133d53affeab53cd016c9b55620a9.

* Revert "Revert "Maybe generate call to inner kernel correctly""

This reverts commit c85b60e6382b4762076b2744227fa5a6d903b830.

* Remove dead code

* rep2loopy: get name of callee kernel for matching the args.

* Add missing code target

* Maybe handle loopy argument access correctly

dat.zero still doesn't work

* Promote dats with shape (n,) to shape (n, 1)

This enables sweeping indices to pick up a direct loop.

* Set lang_version on all make_function calls

* Fix disjoint arguments to C-string kernels

* Resolve review comments

* Rename function_args -> kernel_parameters

* WIP Match args for all kernels in the callables table besides for LACallables? Why?

* Only match args for callables that are CallableKernels.

* Pin decorator

Decorator version 5 changes the API in ways that break the
argument type checking in PyOP2. We're therefore pinning on the
version that works.

* catch up with numpy type changes

* Set within_inames_is_final on instructions

* Squash numpy warning

* Set strides on GlobalArgs

* Caching: produce key for loop Programs.

* Check if kernel is a loopy Program.

* loopy.program.Program -> loopy.Program

* Adapt to loopy interface changes.

* Maintain loopy automatic reshaping (was in loopy)

Everything in loopycompat.py file was in loopy/transform/callable.py
but is removed in inducer/loopy#327.

This commit maintains compatibility but should be phased out.

* Fix loopy.program -> loopy.translation_unit

* dont lint automatic reshaping (was in loopy)

* Lint to pass flake8

* Fix loopy automatic reshaping (was in loopy)

Changes were copied from an out of date branch by accident.

* Fix type case error (recursively)

This mirrors inducer/loopy#326

* ? Maybe fix the last dimension error mismatch. Not sure this will break things in other places.

* Codegen: Don't merge indices of extent 1, so that the instructions are schedulable. This will fix e.g. the failure of tests/extrusion/test_steady_advection_2D_extr.py

* Testing: Update loopy branch

* Squashed commit of the following:

commit 5a89bfc
Author: Connor Ward <[email protected]>
Date:   Mon May 10 10:44:33 2021 +0100

    Use Firedrake PETSc and petsc4py

commit 948cb7d
Author: David Ham <[email protected]>
Date:   Wed Apr 21 13:30:50 2021 +0100

    Pin decorator

    Decorator version 5 changes the API in ways that break the
    argument type checking in PyOP2. We're therefore pinning on the
    version that works.

commit 121c2a6
Author: David Ham <[email protected]>
Date:   Wed Apr 21 13:45:58 2021 +0100

    catch up with numpy type changes

* Revert "Squashed commit of the following:"

This reverts commit bd1916142f883d6a4f0371ca6de2c344f0c7e78a.

* Squashed commit of the following:

commit 9a90c9f
Merge: f1daf21 c8d7ff6
Author: Lawrence Mitchell <[email protected]>
Date:   Tue May 11 12:07:33 2021 +0100

    Merge pull request #618 from OP2/wence/fix/actions-doc-build

    actions: Build docs at end of successful test run

commit c8d7ff6
Author: Lawrence Mitchell <[email protected]>
Date:   Tue May 11 11:41:20 2021 +0100

    actions: Build docs at end of successful test run

commit f1daf21
Merge: d0cc348 5a89bfc
Author: Lawrence Mitchell <[email protected]>
Date:   Mon May 10 20:32:24 2021 +0100

    Merge pull request #617 from connorjward/test-ci

    Make PETSc and petsc4py for CI use Firedrake not pip

commit 5a89bfc
Author: Connor Ward <[email protected]>
Date:   Mon May 10 10:44:33 2021 +0100

    Use Firedrake PETSc and petsc4py

commit 948cb7d
Author: David Ham <[email protected]>
Date:   Wed Apr 21 13:30:50 2021 +0100

    Pin decorator

    Decorator version 5 changes the API in ways that break the
    argument type checking in PyOP2. We're therefore pinning on the
    version that works.

commit 121c2a6
Author: David Ham <[email protected]>
Date:   Wed Apr 21 13:45:58 2021 +0100

    catch up with numpy type changes

* Lint.

* Jenkins

* Make import global.

* Drop package branches.

Co-authored-by: Sophia Vorderwuelbecke <[email protected]>
Co-authored-by: Lawrence Mitchell <[email protected]>
Co-authored-by: Connor Ward <[email protected]>
Co-authored-by: David Ham <[email protected]>
Co-authored-by: Kaushik Kulkarni <[email protected]>

Something like this will be necessary if we want general symmetry
group stuff in the maps. It is also convenient if you want to stage in
a permutation from the canonical FIAT order.

Instead of pattern matching on node names to determine whether nodes
should be relabelled, just use a name generation scheme for
everything. Now we have one less problem.

…versionning

Add PETSc decorator to compile

More loopy interface changes: arg_id_to_val -> arg_id_to_arg

…problem

This previously relied on the old petsc Python package which we no
longer use.

add standard set operations

Fix pyop2-clean

Melina97/fix cache problem

Call apt update in workflow

* Lint fix

* No longer need _copy_parloop

* Unnecessary parloops removed

* Add condition to __iadd__ and __isub__

* Minor correction

* Allow data conversion

* Use _check_shape

* Remove _iop_kernel

* Add casting=unsafe

* Corrections

* Only consider owned indices for zero and copy w. subset

* Add compat checks to zero and copy

Also reverted work on __iadd__, __isub__ etc as these were breaking and
feature incomplete (we did not implement __add__ etc). These should be
covered in a separate PR.

* Add halo optimisations to dat zero and copy

Co-authored-by: Melina Giagiozis <[email protected]>
Co-authored-by: Connor Ward <[email protected]>

Permuted maps

`-march=native` doesn't work on M1 at present.

Fix compilation failures on Mac M1 machines

Major refactor of base.py etc

Make Global.zero accept same kwargs as Dat.zero

This would fail if the provided layers array has zero length (e.g.
for boundary facets on an interior parallel domain) as numpy does not
allow you to call min() on such an array.

Bug fix empty extruded set with variable layers

…in-parallel

Matrix tests run correctly in parallel

If we have a parallel loop with

   map1 = Map(...)
   map2 = PermutedMap(map1, ...)

   par_loop(..., d1(..., map1), d2(..., map2))

The codegen should only require one global argument (for the data from
map1) since map2 is a local permutation of the global argument. This
was being provided by runtime argument passing, but not by the code
generator, which treated map1 and map2 as distinct things.

To fix this, create a new PMap wrapper for Map objects in the codegen
builder that just know how to index themselves through their
permutation. Now we can share the underlying global map data between
the two local map accesses.

Wence/fix/permuted map

…versionning

… into DataCarrier-object-versionning

The fundamental idea behind the changes introduced by this commit are:

- The code generation component of what PyOP2 does is now user-facing
  (via the GlobalKernel object). This permits the user to add their own
  caching for this object.
- Code generation no longer requires any data structures to take place.

Purge parloop data references