[RFC/WIP] Intrinsics for NTuple{N, VecElement}
#18470
Conversation
| @@ -936,6 +937,24 @@ STATIC_INLINE int jl_is_vecelement_type(jl_value_t* t) | |||
| ((jl_datatype_t*)(t))->name == jl_vecelement_typename); | |||
| } | |||
|
|
|||
| STATIC_INLINE int jl_is_vec_type(jl_value_t* t) | |||
| { | |||
| if (!jl_is_tuple_type(t)) | |||
There was a problem hiding this comment.
4 space (also one case in the test)
|
So the big show-stopper for this are our runtime intrinsics. I am slightly tempted to punt this down the road till somebody actually needs it and for now implement |
StefanKarpinski
added
speculative
| jl_errorf("%s: type is not a bitstype", name); | ||
| if (!(jl_is_bitstype(aty) || jl_is_vec_type(aty))) | ||
| jl_errorf("%s: value is not a bitstype or vector type", name); | ||
| if (!(jl_is_bitstype(ty) || jl_is_vec_type(aty))) |
There was a problem hiding this comment.
Did you mean ty instead of aty here?
* Use `jl_datatype_size` instead of `dt->size`. `jl_datatype_size` is always in bytes, whereas the underlying implementation of `dt->size` might change. * do not use .size directly (that is just asking for trouble) * Introduce `jl_datatype_nbits` that gives nbits Instead of calculating it per hand everywhere. This allows for interesting experiments (and eventually not restricting `bitstype` to multiplies of 8).
|
How's this going? |
|
Stalled because it needs a big refactor of |
|
Cool. You'll probably want to talk/collaborate with @vtjnash on any intrinsics work so you two aren't working across purposes. |
|
This doesn't look like it's going to be ready in time for 0.6. |
|
No this won't be ready in time and is somewhat dependent on the runtime-library julep. |
|
Seems like a feature that can be post-1.0. |
This fixes #18445 by handling
NTuple{N, VecElement}in the lowering of our intrinsics. This should make it easier to implement the fundamentals ofSIMD.jlwithoutllvmcalland maybe in the long-term enable us to have SIMD types in base julia.Todo:
jl_is_vec_typecheck thatVecElementcontains ajl_is_bitstypeIf needed I can separate out b2f6b1b and fb12337 into a separate PR.
cc @eschnett