[feature request] Add ffi interface for logp-gradient function

[aseyboldt]

The current shared object build for each stan model includes a function for computing the logp gradient, but calling this function is relatively slow as it converts the dataset each time. This PR adds a couple of functions that make it possible to call the gradient function from a C or numba library without any python overhead:

model=import_services_extension_module(model_id)
ctx=model.new_logp_ctx(data)
function_address=model.logp_func(ctx)
# later, after we are done# model.free_logp_ctx(ctx)

function_address is now a pointer to a C function:

intlogp_gradient(size_tndim, constdouble*unconstrained_parameters, double*gradient, double*logp, void*ctx)

Which is way faster than the previously exported gradient function.

This is useful for implementing different samplers based on stan models (for example the sampler implemented in rust that I'm working on ;-) ).

[riddell-stan]

Feature requests need to be proposed and discussed on the Stan Forums first. The issue tracker is for bugs and confirmed features only.

[aseyboldt]

@riddell-stan Sorry, I didn't realize that.
I created a topic here: https://discourse.mc-stan.org/t/feature-proposal-expose-fast-logp-gradient-function-in-httpstan/27856

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.