BUG: Iris AreaWeighted regrid float tolerance bug

[cpelley]

Floating point equality without tolerance within iris.experimental.regrid._within_bounds produces masked results under circumstances where different longitude ranges are wrapped by wrap_lons here when we expect them to be equal (i.e. there is a loss of precision).

Here is a break point at within_bounds for the case where the source longitude extent is [-91, 0] and the target longitude extent is [270, 360]:

(Pdb) src_bounds
array([[-1.58824962,  0.        ]])
(Pdb) tgt_bounds
array([[ -1.57079633e+00,   6.66133815e-16]]

You will note the 6.66133815e-16 value which makes the equality checks fail, since 0 != 6.66133815e-16.

See the added unittest for the full working example.

[cpelley]

I have started an iris dev discussion on the wider issue of dealing with floats in iris here.

[cpelley]

Two exceptions raised.

One under python3.4 test target: default:

======================================================================
ERROR: test_simple (iris.tests.test_mapping.TestUnmappable)
----------------------------------------------------------------------
Traceback (most recent call last):
...
ValueError: Image comparison failed. Message: Actual SHA1 020e0ca45a9fbd5caef915e59c02276725919889 not in expected ['6c139dab88d9e5884d313f1a607de5e9c4fe5a93', 'a14231d0f6fe19de89dfd200d25f1d7dbc31b2ff'] for test iris.tests.test_mapping.TestUnmappable.test_simple.0.

The second under python3.4 test target: example:

======================================================================
ERROR: test_atlantic_profiles (example_tests.test_atlantic_profiles.TestAtlanticProfiles)
----------------------------------------------------------------------
Traceback (most recent call last):
...
ValueError: Image comparison failed. Message: Actual SHA1 f25dc9acc16a887eb224b4970d3042f5bcfd9743 not in expected ['0bde47fbbc4a15498c0ecb8bd3eb5e63339e5121', 'b8365851d0b4b5589289026eda692c643499c7f3', 'c3a2ccc6873ddee7ea0dafd25498b743ffece4fa', '4cbda27eb65fc360d08dde0ccf40b951fdfe4453', '3a2f43bc6b1bd4397a8eb6877a128ab7a4f051ff', 'da5cd8269bcc96533c8c2f4a788e2c980381dfef'] for test example_tests.test_atlantic_profiles.TestAtlanticProfiles.test_atlantic_profiles.0.

Anyone might know what this might specifically effect only python3.4?
Cheers

[marqh]

Anyone might know what this might specifically effect only python3.4?

@cpelley unfortunately, lots of things

we have moved swiftly onto perceptual hashing, to let us handle issues such as this

please rebase and see how this affects your change, it should help with image comparison issues

[cpelley]

Thanks @marqh. All tests pass. I have made an alternate PR (#2237) which points to the new v1.11.x branch now that it exists.
We would have a preference to this making it to v1.11.x so that we don't have to make our users dependent on iris master.

Cheers

[cpelley]

Closing this as it is replaced by #2237 which points to v1.11.x