Firedrake meeting 2021 02 10

Date and time 2021-02-10 16:00UTC (16:00GMT)

Action Items

1. Pick Chair and Minuter.
2. ALL: (ongoing) triage the open issues and confirm if they are indeed still open (and perhaps provide labels)

Minutes

Present:

DH, LM, PK, CW, DS, JB, NB, RNH, RK, SK, SV, KS

Apologies:

Stash assembly_callable on the form? (CW)

• This would eliminate a lot of the symbolic overhead occurring in assemble() but may annoy James.

RK observed 45% assembly time saving by using create_assembly_callable.

PyOP2 design question:

• Need nice separation: things that you need to execute and data to execute those on.
• ParLoop objects not to hold DataCarrier objects at all: in principle we just need shapes.
• args (shape, READ/WRITE, Map) are now in code, but these should actually be attached to the kernel separately from data.

Short term hack?

• Don't introduce public API change
• Modify create_asembly_callable so that it would take output tensor?

Realated (pointwise assignment):

Pointwise assignement is inefficient for the same reason, but this is harder as there is no persistent object lying around (we make new expression every time).

-> Lift expressions to make them persistent objects.

Related (time integration):

Define reusable Block, and allow for:

while t < ntime:
    timestep.replay() # run through stashed parloop

CW may or may not work on these.

FEnICS 2021

What to present

RK: irksome

India: dual space

NB : external operator

KS: subspace (may or may not) -> write abstract

Deadline: 20/02/2021

RNH: TSFC "bug fixes"

See https://github.com/firedrakeproject/tsfc/pull/237 - source of some discussion (Miklos doesn't think the PR is necessary) that seems to get at some core GEM design philosophy questions.

Two possibilities:

1. A[i,j]=B[j] -> implicitly broadcast across i
2. A[i,j]=Broadcast(B[j], i) or something (explicit broadcast node makes it more type-safe)

Merge PRs:

AOB

Date of next meeting

2021-02-17 16:00UTC (16:00GMT)