fix[cartesian]: Fix minimal k-range computation

[twicki]

Description

The computation of the minimal k-ranges that happen during the vaildate-args step is allowing for inconsistent computation to happen. This PR stiffens the requirements on fields:

• intervals [START+X ,END+y) are now also considered:
  • interval(3,-1) requires a minimal size of 5 for the interval to not be empty
  • interval(3,None) now requires a minimal size of 4
• intervals [START+X, START+Y) and [END+X,END+Y) are not affected.
• empty intervals are still allowed to have a 0-domain as to accomodate 2-dimensional fields
  • interval(...) still requires no k-size

Requirements

• All fixes and/or new features come with corresponding tests.

[romanc]

I am confused. In my mental model, min_k_size is the minimal height of the vertical column. Why do

def stencil_no_extent_1(field_a: gs.Field[float], field_b: gs.Field[float]):
    with computation(PARALLEL), interval(0, 2):
        field_a = field_b[0, 0, 0]

def stencil_no_extent_2(field_a: gs.Field[float], field_b: gs.Field[float]):
    with computation(PARALLEL), interval(1, 2):
        field_a = field_b[0, 0, 0]

both have a min_k_size of 2? Both have the same upper bound of 2, but the second one starts at the first level. I'd thus expect 2 for the first one and 1 for the second one. This was already before your changes, so maybe I don't understand correctly. What's wrong with my thinking?

[twicki]

I am confused. In my mental model, min_k_size is the minimal height of the vertical column. Why do

def stencil_no_extent_1(field_a: gs.Field[float], field_b: gs.Field[float]):
    with computation(PARALLEL), interval(0, 2):
        field_a = field_b[0, 0, 0]

def stencil_no_extent_2(field_a: gs.Field[float], field_b: gs.Field[float]):
    with computation(PARALLEL), interval(1, 2):
        field_a = field_b[0, 0, 0]

both have a min_k_size of 2? Both have the same upper bound of 2, but the second one starts at the first level. I'd thus expect 2 for the first one and 1 for the second one. This was already before your changes, so maybe I don't understand correctly. What's wrong with my thinking?

Sor for the second stencil to make sense, the domain in k need to have size 2. Otherwise the second interval would be empty.
Yes, the extent in which computation happens is only of size 1, but these checks are here to understand if the arguments to the call are compatible with the stencil definition, and this is why we need to check domain- and field-sizes

[romanc]

Alright, I understand the meaning of min_k_size now and this looks good from what I understand.

I don't understand the k_bounds thingy to be honest, hence the question if the docstring was wrong to begin with. Anyways, since we don't have changes there I guess that's okay for now and I don't need to understand.

[romanc]

Suggestion: also change the comment in

gt4py/tests/cartesian_tests/integration_tests/multi_feature_tests/stencil_definitions.py

Line 298 in 4b566d7

with interval(6, -10): # this stage will only run if field has more than 16 elements

I think it's technically still valid because this interval demands a size of 16 and the extra one comes from the interval(-10, None) below. I just - I don't like the wording because - to me - "will only run" sounds more like it will be skipped rather than throwing an error.

[romanc]

all good from my side

[FlorianDeconinck]

cscs-ci run

[FlorianDeconinck]

cscs-ci run

[FlorianDeconinck]

cscs-ci run

[FlorianDeconinck]

Gitlab ci timed out on SLURM - re-launch