Update tutorial to mention local barrier instructions.

See merge request !152

Add a comment about when spill and reload is necessary; add an example about how loopy complains about a missing global barrier.

(Kind of) fix an update_persistent_hash() performance regression

See merge request !150

small fix in get slab decomposition

See merge request !151

regression.

The key builder (701dc4ed) added field
names to the hash key which ended up slowing things down.

Example (in pytential test suite, loads ~110 times from cache):

PYOPENCL_TEST=port \
time -p python test_layer_pot_eigenvalues.py \
"test_ellipse_eigenvalues(cl._csc, 1, 5, 3, False)"

* Prior to 701dc4ed:

    real 21.68
    user 21.06
    sys 2.03

* As of 701dc4ed:

    real 27.11
    user 26.59
    sys 1.94

* As of this commit:

    real 22.29
    user 21.90
    sys 1.86

This commit doesn't make everything as fast as it used to be because
pymbolic expressions are still stringified before hashing. However,
caching (#67) will eventually help with that too.

Replace defaultdict(lambda: set()) with defaultdict(set).

See merge request !149

Towards lazy instruction lists: implement a comparison key generator for instructions

See merge request !144

Remove automatic uniquification of instruction ids in the kernel

Closes #89

See merge request !148

Fix cache collision warnings in test suite (closes #88).

Closes #88

See merge request !147

  constructor.
* Check that all instruction ids are unique on kernel creation.
* Add a uniquify_instruction_ids() transform.

This closes #89.

Include target type in persistent hash (Fixes #88 on gitlab)

Closes #88

See merge request !146

This adds a class for building a key for equality comparison purposes.
In addition, it stringifies pymbolic expressions to reduce the
overhead of pymbolic expression comparison. This class can also be
used for generating persistent hash keys.

This change also converts instructions to use this class for
persistent hashing.

Related: #67.

Kernel constructor: Add an option to control whether instruction IDs get uniquified or not.

See merge request !142

IDs get uniquified or not.

Related: #67

Allow names, exprs, None, and lp.auto for array offsets

See merge request !114

Allow parameters to be fixed at kernel creation time.

See merge request !141

Among other things, passing fixed parameters can improve the accuracy
of temp shape inference, so that fallback can occur less often.

The motivation for this is to remove the temp_shape_fallback warnings
when constructing sumpy kernels.

Lazy data structures

See merge request !139

Adjust the recursionlimit for complexity in number of inames

Closes #78

See merge request !136

This changes implements a LazyList and
LazyListWithEqAndPersistentHashing data structure. These data
structures lazily unpickle their contents.

This change also renames LazilyUnpicklingDictionary to LazyDict.