ASA core modules #641
ASA core modules #641
Conversation
src/sunadjointcheckpointscheme/fixed/sunadjointcheckpointscheme_fixed.c
Outdated
Show resolved
Hide resolved
src/sunadjointcheckpointscheme/fixed/sunadjointcheckpointscheme_fixed.c
Outdated
Show resolved
Hide resolved
src/sunadjointcheckpointscheme/fixed/sunadjointcheckpointscheme_fixed.c
Outdated
Show resolved
Hide resolved
|
|
||
| if (!(step_num % IMPL_MEMBER(self, interval))) | ||
| { | ||
| if (stage_num == 0) { *yes_or_no = SUNTRUE; } |
There was a problem hiding this comment.
This will be z[0] rather than y_n for methods with an implicit first stage.
| { | ||
| SUNFunctionBegin(self->sunctx); | ||
|
|
||
| void* queue = NULL; |
There was a problem hiding this comment.
Here and below, I think a NULL queue will cause an error with SYCL
There was a problem hiding this comment.
Hmmm. Im not sure how to get access to a queue here. I suppose we could check if the vector is a SYCL vector, and if it is, get it from the vector?
There was a problem hiding this comment.
I suppose another option is to get it from the memory helper, but it would simiarly require checking the type of SUNMemoryHelper and adding a get function for the queue.
There was a problem hiding this comment.
I don't see a good way around this at the moment. The memory helper does not store the queue (it expects the queue to be provided as an input to the memory helper functions). We could add a "get execution queue" function to the NVector but not all the functions that call a memory helper function have access to a vector. Looking back we maybe should have added the queue (stream) as an input the memory helper constructor instead of having it in the function signatures.
| SUNMemoryType buffer_mem_type = N_VGetDeviceArrayPointer(v) ? SUNMEMTYPE_DEVICE | ||
| : SUNMEMTYPE_HOST; |
There was a problem hiding this comment.
Here and below, user supplied vectors my not supply N_VGetDeviceArrayPointer but still use device data. I think we'll need a new NVector function to query what memory space the data is stored in.
There was a problem hiding this comment.
I agree the ideal long term solution is to add a function to query the memory space. However, an alternative is that we just require N_VGetDeviceArrayPointer to be provided when using this module. We already do require it when using a GPU-enabled direct SUNLinearSolver.
There was a problem hiding this comment.
GPU-enabled direct SUNLinearSolvers require this because they need the underlying data pointer. In this case that's not required and I could see this a being potentially problematic because it might give the impression that user need to provide a working GetDeviceArrayPointer implementation or we ask them to make a dummy version that does not work.
| return SUN_SUCCESS; | ||
| } | ||
|
|
||
| SUNErrCode SUNAdjointStepper_ReInit(SUNAdjointStepper self, N_Vector y0, |
There was a problem hiding this comment.
[discussion] The current behavior is like ReInit, but we do not have a SUNStepperReInit function. So the user would need to reinitialize the forward and adjoint steppers. We could add SUNStepperReInit with the same signature of SUNStepperReset. This would require some updates to the ARKODE ReInit functions that would be wrapped by a generic ReInit to allow for NULL inputs in order to retain the current RHS functions (and some other changes handle NULL inputs).
|
|
||
| while ((direction == -one && t > tout) || (direction == one && t < tout)) | ||
| { | ||
| SUNCheckCall(SUNAdjointStepper_OneStep(self, tout, sens, tret)); |
There was a problem hiding this comment.
To get interpolated output at the desired time this should use SUNStepper_Evolve (or we'll need to add a SUNStepper function to get interpolated output)
There was a problem hiding this comment.
Good catch, indeed this should have been calling SUNStepper_Evolve.
| self->last_flag = adj_sunstepper->last_flag; | ||
|
|
||
| self->step_idx--; | ||
| self->nst++; |
There was a problem hiding this comment.
Do we need to store nst or can we get this from the adj_sunstepper?
There was a problem hiding this comment.
The SUNStepper does not have a nst counter nor a method to get it from the underlying integrator.
There was a problem hiding this comment.
Maybe we should add a Get function to the SUNStepper since this counter won't get incremented if using SUNAdjointStepper_Evolve.
Co-authored-by: David Gardner <gardner48@llnl.gov>
… into feature/sunadjoint-core-modules
|
|
||
| Function pointer to insert a checkpoint state represented as a :c:type:`N_Vector`. | ||
|
|
||
| .. c:member:: SUNErrCode (*loadVector)(SUNAdjointCheckpointScheme cs, int64_t step_num, int64_t stage_num, sunbooleantype peek, N_Vector* yout, sunrealtype* tout) |
There was a problem hiding this comment.
I don't think we could check that within the checkpoint scheme since we do not know the time step size. We could probably check that in the package-specific code though.
This PR adds the core modules that will support adjoint sensitivity analysis in a package-agnostic way.