Sai Praveen Bangaru
Sai Praveen Bangaru
…and added support for differentiating overloaded functions. - Added AST and IR dictionary nodes to track differentiable types. - Added overload and generic function resolution. - Adjusted JVP logic to...
Right now, its possible to accidentally mark a method that relies on global/shared memory side-effects with `[PreferRecompute]`. The additional calls for recomputation will produce inconsistent output, and cause seemingly random...
This PR includes the following: + Add diagnostic for missing diff-member associations. + Automatically create diff member associations if differential type is the same as the primal type. + Move...
Global parameters & variables are by-default not differentiable (even if they are of a differentiable type), because our auto-diff passes do not touch anything outside of function bodies (and will...
Added a draft of all the type system components related to auto-diff. Also added a separate file that discusses all the decorators related to auto-diff and their implementations. Resolves #4215
We have two ways to use slang-torch: `[AutoPyBindCUDA]` which generates the kernel launch and a manual appraoch that allows the user to directly launch kernels through `__dispatch_kernel()` Currently, the capability...
The decorator that points a member in the differential type to its corresponding member is missing here, and that causes a crash when differentiating the `IRMakeStruct` instruction. We likely need...
Currently, nested associated types are expected to be fully resolved rather than treated as existential types. This is a limitation of the current existential type lowering approach that needs some...
Our decoration naming and usage is rather messy, mostly because we rewrote it several times. The most glaring part is where we have a single decoration `ForwardDerivativeDecoration` to hold user-provided...