Schedule bug with descript_scheduler

[yadej]

With that python file

import xtc.graphs.xtc.op as O
from xtc.backends.mlir import Backend
from xtc.schedules.descript import descript_scheduler

i, j, k, dtype = 200, 224, 240, "float32"
a = O.tensor((i, k), dtype, name="A")
b = O.tensor((k, j), dtype, name="B")

with O.graph(name="matmul") as gb:
    p = O.matmul(a, b, name="C")
graph = gb.graph
print(graph)

backend = Backend(graph)
sch = backend.get_scheduler()
descript_scheduler(
    scheduler=sch, 
    node_name="C", 
    abstract_axis=["i", "j", "k"],
    spec={
        'k': {},
        'j': {},
        'i[0:18]':
            {'i#18': {},
             'k#80': {},
             'k#16': {},
             'i#9': {'unroll': 9},
             'j#32': {'unroll': 32},
             'j#16': {'vectorize': None}},
        'i[18:200]': {
            'i#182': {},
            'k#80': {},
            'k#16': {},
            'i#14': {'unroll': 14},
            'j#32': {'unroll': 32},
            'j#16': {'vectorize': None}
        }
    }
)

I got the error

  File "/home/ruicesista/Documents/xtc/xtc/src/xtc/schedules/descript.py", line 246, in _interpret_spec
    raise ScheduleInterpretError(
        f"Splitting of {axis} unachieved (stops at {cut})."
    )
xtc.schedules.descript.ScheduleInterpretError: Splitting of i unachieved (stops at 200).

The i dimension goes to 200 but it says that the splitting of i is unachieved even thought in this case is complete

[guillon]

@qaco should the indexes on i[18:200] be rewritten i[18:199] instead? I do not remember if the semantic assumes last included or last excluded.
@yadej did you try this? Also you may write i[18:] I suppose as a work around.

[yadej]

With i[18:199] I have the same error as i[18:200]

But with i[18:]

I got this error

File [~/Documents/xtc/xtc/src/xtc/schedules/descript.py:615](http://localhost:8888/home/ruicesista/Documents/xtc/xtc/src/xtc/schedules/descript.py#line=614), in LoopNest._must_be_smaller_routine(new_size, current_sizes, loop_name, axis)
    613 old_size = current_sizes[axis]
    614 if old_size is not None and new_size > old_size:
--> 615     raise ScheduleValidationError(
    616         f"""
    617         Inner loop {loop_name} on axis {axis} must be smaller than outer loop.
    618         """
    619     )

ScheduleValidationError: 
                Inner loop i0 on axis i must be smaller than outer loop.

Even though the schedule should be valid

[qaco]

[18:] is the right way to do it (for now).

According to the ScheduleValidationError, I guess something goes wrong with the offsets computations.

Could you:

• try 181 for the first level of tiling on i ?
• write a minimal failing example to put it in the unit tests ?

[yadej]

I tried with i[0:182] and i[182:] and does work

[guillaumeiooss]

[18:] is the right way to do it (for now).

Does it work if there are loops on "i" above it?

For example, if the problem size on dimension "I" is 200, and we want to have a loop on i with 2 iterations before splitting i (on the remaining 100 iterations), is this a valid syntax ?

'i': {},
  'i[0:9]': {
    blablabla
  },
  'i[9:]': {
    bliblibli
  }

Indeed, we might be missing a size info on the split (given by i[9:100]) there.