[Rust][IRModule] Flesh out IRModule methods #6741
Conversation
rust/tvm-macros/src/external.rs
Outdated
| use quote::quote; | ||
| use syn::parse::{Parse, ParseStream, Result}; | ||
|
|
||
| use syn::{FnArg, Generics, Ident, Lit, Meta, NestedMeta, Pat, ReturnType, TraitItemMethod, Type}; | ||
| use syn::{FnArg, Signature, Attribute, token::Semi, Visibility, Generics, Ident, Lit, Meta, NestedMeta, Pat, ReturnType, Type}; |
There was a problem hiding this comment.
I probably just haven't run it yet, since everything is automatic I tend to just run it when I know I'm gtg.
| use syn::{FnArg, Generics, Ident, Lit, Meta, NestedMeta, Pat, ReturnType, TraitItemMethod, Type}; | ||
| use syn::{FnArg, Signature, Attribute, token::Semi, Visibility, Generics, Ident, Lit, Meta, NestedMeta, Pat, ReturnType, Type}; | ||
|
|
||
| struct ExternalItem { |
There was a problem hiding this comment.
nit: brief rustdoc? Some comments would help me get the story of what ExternalItem and External are for.
| let derives = quote! { | ||
| impl std::hash::Hash for #ref_id { | ||
| fn hash<H: std::hash::Hasher>(&self, state: &mut H) { | ||
| self.0.hash(state) |
There was a problem hiding this comment.
Possibly dumb comment (I'm not familiar with this macro) - is self.0 always the right thing to hash? Or is #ref_id sometimes a tuple with multiple fields, where field .1 and .2 etc. may be part of what distinguishes objects when hashing?
There was a problem hiding this comment.
No we generate a wrapper new type over the ObjectPtr which contains 1 field always. ObjectPtr is currently dispatching to TVM's C++ hashing so we have cross language consistency.
| @@ -83,6 +83,10 @@ impl DataType { | |||
| DataType::new(DL_FLOAT_CODE, bits, lanes) | |||
| } | |||
|
|
|||
| pub const fn float32() -> DataType { | |||
There was a problem hiding this comment.
if this is float32, should the old float be renamed to float64 (to avoid clashing with float vs. double size intuition?)
There was a problem hiding this comment.
We don't use float/double terminology in TVM due to the fact that both integers and floating point values are both parametric by vector length and bit width. The float method here takes both bit width and vector length, unless I'm missing something.
Its pretty easy to get started, @adelbertc has not repro'd running them if you are interested. I will finish adding tests and ping you guys again. |
jroesch
force-pushed
the
rust-improve-ir-mod
branch
from
9608878
to
d871c53
Compare
jroesch
force-pushed
the
rust-improve-ir-mod
branch
from
d871c53
to
726a01b
Compare
| use syn::{FnArg, Generics, Ident, Lit, Meta, NestedMeta, Pat, ReturnType, TraitItemMethod, Type}; | ||
| use syn::{FnArg, Signature, Attribute, token::Semi, Visibility, Generics, Ident, Lit, Meta, NestedMeta, Pat, ReturnType, Type}; | ||
|
|
||
| struct ExternalItem { |
|
@adelbertc @imalsogreg I will try and ship docs for proc macro after landing this branch if that's cool with you guys, just so I can unblock some other branches I have. |
* WIP * WIP * WIP * WIP * Disable WASM and fix rebase * Work on finishing tests * Make entire object system printable * Write some more tests for IRModule * All tests pass * Format * Restore module.cc * Bump syn
* WIP * WIP * WIP * WIP * Disable WASM and fix rebase * Work on finishing tests * Make entire object system printable * Write some more tests for IRModule * All tests pass * Format * Restore module.cc * Bump syn
* WIP * WIP * WIP * WIP * Disable WASM and fix rebase * Work on finishing tests * Make entire object system printable * Write some more tests for IRModule * All tests pass * Format * Restore module.cc * Bump syn
This PR fully expand the IRModule interface in Rust and fixes a whole host of sharp edges and improves the procedural macros for generating the boilerplate.
cc @mwillsey @altanh @imalsogreg @gussmith23