Skip to content

wrong GPts/s in artificial corner case with multiple overlapping SubDimensions #1552

Open
@FabioLuporini

Description

The following will produce a wrong value for points (and eventually will produce a wrong local per-section GPts/s)

      from devito import *
       
      grid = Grid(shape=(20, 20)) 
        x, y = grid.dimensions
        t = grid.stepping_dim
        thickness = 4
        u = TimeFunction(name='u', save=None, grid=grid, space_order=0, time_order=1)
        v = TimeFunction(name='v', save=None, grid=grid, space_order=0, time_order=1)
        xi = SubDimension.middle(name='xi', parent=x,
                                 thickness_left=thickness, thickness_right=thickness)
        xi2 = SubDimension.middle(name='xi2', parent=x,
                                 thickness_left=thickness, thickness_right=thickness)
        yi = SubDimension.middle(name='yi', parent=y,
                                 thickness_left=thickness, thickness_right=thickness)
        t_in_centre = Eq(u[t+1, xi, yi], 1)
        t2_in_centre = Eq(v[t+1, xi2, yi], u[t+1, xi2+1, yi])
        op = Operator([t_in_centre, t2_in_centre], opt=('advanced', {'blockinner': True}))

this is due to using set rather than (much) more complex calculations (eg distinguishing between different types of Dimensions etc) in Profiler._analyze when the points are calculated

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions