<div dir="ltr"><div>> Jed suggested having any number of "RHS" functions. I don't think we need more than 2, 1 for left hand side</div><div>> and 1 for right. If that ends up being not enough we can change to have any number of them. Just to be clear.</div><div>> I suggest three functions</div><div>> IFunction which defaults to u_t plus TSSetMassMatrix() which changes the default IFunction to M u_t</div><div>> LHS function which defaults to 0, if provided defaults to being treated implicitly by IMEX</div><div>> RHS function which defaults to 0, if provided defaults to being treated explicitly by IMEX</div><div>> Then a TSSetStiffMatrix(ts,Mat L) (horrible name) that provides u_t -Lu = g() + Lu</div><div><br></div><div>At my current level of understanding I get this trifecta (quad-fecta?).</div><div><br></div><div>In my ice sheet case u_t = f(t,u) + g(t,u), I would not set IFunction at all, or I would use a constant. The divergence of flux f(t,u) = div(q) would go into the LHSFunction and the elevation-dependent mass balance g(t,u) would go into the RHSFunction. I would not use TSSetStiffMatrix() at all.</div><div><br></div><div>Barry's suggestion makes sense for this and the few other usage cases I have already experienced ... kind of a lame endorsement, but, on such a weak basis one casts a vote ...<br></div><div><br></div><div>Ed</div><div><br></div><div><br></div><div>PS I would not use TSSetStiffMatrix() because none of the linearizations/simplfications known at the completion of a TS step could "capture" the stiff part without being derived a la the Jacobian anyway. To see what I mean, note that VINEWTONRSLS has a different *dimension* at each Newton iteration. I don't know the (in-active-constraint) dimension of the solution I'll get at each time step, much less how to scale the various stiff-equation-parts for those in-active dimensions. In these problems there is a lot going on in the implicit solve, besides just handling stiffness.</div><div><br></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Tue, Feb 14, 2017 at 6:20 PM, Barry Smith <span dir="ltr"><<a href="mailto:bsmith@mcs.anl.gov" target="_blank">bsmith@mcs.anl.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><span class=""><br>
> On Feb 14, 2017, at 8:56 PM, Emil Constantinescu <<a href="mailto:emconsta@mcs.anl.gov">emconsta@mcs.anl.gov</a>> wrote:<br>
><br>
> On 2/14/17 4:10 PM, Barry Smith wrote:<br>
>> Ok, you don't recompile but forcing that into user code is still disgusting. With my api the user code is<br>
>><br>
>>>>> TSSetRHSFunction(ts,NULL,<wbr>RHSFunction,&ptype[0]);<br>
>>>>> TSSetLHSFunction(ts,NULL,<wbr>LHSFunction,&ptype[0]);<br>
>>>>> TSSetRHSJacobian(ts,Jac,Jac,<wbr>RHSJacobian,&ptype[0]);<br>
>>>>> TSSetLHSJacobian(ts,Jac,Jac,<wbr>LHSJacobian,&ptype[0]);<br>
>> and -ts_type xxx works correctly for ALL methods, implicit, explicit and imex without requiring any special command line options for different methods.<br>
><br>
> Is this a viable solution? Growing the API to fix this situation will just put a burden with each new TS method after we refactor it in the current landscape.<br>
<br>
</span> No just the opposite, the TS implementations will talk to functions who will put things together for it. So All implicit methods will call something like TSBuildImplicitFunction(), all explicit methods will call something like TSBuildExplicitFunction() and then IMEX methods will call both of these. In fact likely we can refactor to make things a little better than today. Depending on options and flagsTSBuildExplicitFunction() would build out of all the user provided functions what it needs etc.<br>
<br>
One problem with the current code is the TS methods call things with the same names as the user API. So implicit methods call TSComputeIFunction() while explicit methods call TSComputeRHSFunction(). This is wrong because implicit methods actually also use the rhs function provided by the user.<br>
<br>
The function below absolutely should not be called TSComputeIFunction()! It does not just compute IFunction()<br>
PetscErrorCode TSComputeIFunction(TS ts,PetscReal t,Vec U,Vec Udot,Vec Y,PetscBool imex)<br>
{<br>
PetscErrorCode ierr;<br>
TSIFunction ifunction;<br>
TSRHSFunction rhsfunction;<br>
<span class=""> void *ctx;<br>
DM dm;<br>
<br>
PetscFunctionBegin;<br>
PetscValidHeaderSpecific(ts,<wbr>TS_CLASSID,1);<br>
PetscValidHeaderSpecific(U,<wbr>VEC_CLASSID,3);<br>
</span> PetscValidHeaderSpecific(Udot,<wbr>VEC_CLASSID,4);<br>
PetscValidHeaderSpecific(Y,<wbr>VEC_CLASSID,5);<br>
<br>
ierr = TSGetDM(ts,&dm);CHKERRQ(ierr);<br>
ierr = DMTSGetIFunction(dm,&<wbr>ifunction,&ctx);CHKERRQ(ierr);<br>
ierr = DMTSGetRHSFunction(dm,&<wbr>rhsfunction,NULL);CHKERRQ(<wbr>ierr);<br>
<span class=""><br>
if (!rhsfunction && !ifunction) SETERRQ(PetscObjectComm((<wbr>PetscObject)ts),PETSC_ERR_<wbr>USER,"Must call TSSetRHSFunction() and / or TSSetIFunction()");<br>
<br>
</span> ierr = PetscLogEventBegin(TS_<wbr>FunctionEval,ts,U,Udot,Y);<wbr>CHKERRQ(ierr);<br>
if (ifunction) {<br>
PetscStackPush("TS user implicit function");<br>
ierr = (*ifunction)(ts,t,U,Udot,Y,<wbr>ctx);CHKERRQ(ierr);<br>
PetscStackPop;<br>
}<br>
if (imex) {<br>
if (!ifunction) {<br>
ierr = VecCopy(Udot,Y);CHKERRQ(ierr);<br>
}<br>
} else if (rhsfunction) {<br>
if (ifunction) {<br>
Vec Frhs;<br>
ierr = TSGetRHSVec_Private(ts,&Frhs);<wbr>CHKERRQ(ierr);<br>
ierr = TSComputeRHSFunction(ts,t,U,<wbr>Frhs);CHKERRQ(ierr);<br>
ierr = VecAXPY(Y,-1,Frhs);CHKERRQ(<wbr>ierr);<br>
} else {<br>
ierr = TSComputeRHSFunction(ts,t,U,Y)<wbr>;CHKERRQ(ierr);<br>
ierr = VecAYPX(Y,-1,Udot);CHKERRQ(<wbr>ierr);<br>
}<br>
}<br>
ierr = PetscLogEventEnd(TS_<wbr>FunctionEval,ts,U,Udot,Y);<wbr>CHKERRQ(ierr);<br>
PetscFunctionReturn(0);<br>
}<br>
<br>
The current code entangles too much of the user API to the methods, this can be fixed.<br>
<span class=""><br>
> If the user experiments with different ways of splitting the solution they would have to define RHS and IF or RHS and LHS in different ways (according to the splittings they experiment with). It may look disgusting, but I don't see another way around it unless you allow for a list of operators to be defined and then the user to assign them to LHS or RHS.<br>
<br>
</span> Jed suggested having any number of "RHS" functions. I don't think we need more than 2, 1 for left hand side and 1 for right. If that ends up being not enough we can change to have any number of them. Just to be clear. I suggest three functions<br>
<br>
IFunction which defaults to u_t plus TSSetMassMatrix() which changes the default IFunction to M u_t<br>
LHS function which defaults to 0, if provided defaults to being treated implicitly by IMEX<br>
RHS function which defaults to 0, if provided defaults to being treated explicitly by IMEX<br>
<br>
Then a TSSetStiffMatrix(ts,Mat L) (horrible name) that provides u_t -Lu = g() + Lu<br>
<br>
None of the TS implementations will every directly know about what the user has provided. They will call the wrapper functions I mention above.<br>
<br>
I think Jed and Emil may be too enamored with the reductionist model of only IFunction() and RHSFunction() to see that though it encompasses everything it may not be the best user API.<br>
<div class="HOEnZb"><div class="h5"><br>
<br>
<br>
><br>
>><br>
>>> Adding all that logic to keep track of left sides and right sides for academic examples is likely not the best development.<br>
>> I don't think it is "just academic examples", it is all examples without a mass matrix.<br>
>><br>
>> Once the user has decided with ts_type to use for production if it is fully implicit or explicit then they can depending on the type selected, write just a left hand side, just a right hand side for higher efficiency (less update of ghost points, fewer iterations over loops etc).<br>
>><br>
>> With a constant mass matrix we can have TSSetMassMatrix() and then TSSetIFunction() is reserved for when it is absolutely needed.<br>
><br>
> As much as I would disagree with growing the API at the level of defining the problem, I think TSSetMassMatrix() would let us do more things in the solvers. Also it would be useful to know if the mass matrix is singular or not for efficiency reasons.<br>
><br>
> Emil<br>
><br>
>> Barry<br>
>><br>
<br>
</div></div></blockquote></div><br><br clear="all"><div><br></div>-- <br><div class="gmail_signature" data-smartmail="gmail_signature">Ed Bueler<br>Dept of Math and Stat and Geophysical Institute<br>University of Alaska Fairbanks<br>Fairbanks, AK 99775-6660<br>301C Chapman and 410D Elvey<br>907 474-7693 and 907 474-7199 (fax 907 474-5394)</div>
</div>