cuda_std_macros/

lib.rs

use proc_macro::TokenStream;
use proc_macro2::Span;
use quote::{quote_spanned, ToTokens};
use syn::{
    parse::Parse, parse_macro_input, parse_quote, punctuated::Punctuated, spanned::Spanned, Error,
    FnArg, Ident, ItemFn, ReturnType, Stmt, Token,
};

/// Registers a function as a gpu kernel.
///
/// This attribute must always be placed on gpu kernel functions.
///
/// This attribute does a couple of things:
/// - Tells `rustc_codegen_nvvm` to mark this as a gpu kernel and to not remove it from the ptx file.
/// - Marks the function as `no_mangle`.
/// - Errors if the function is not unsafe.
/// - Makes sure function parameters are all [`Copy`].
/// - Makes sure the function doesn't return anything.
///
/// Note that this does not cfg the function for nvptx(64), that is explicit so that rust analyzer is able to
/// offer intellisense by default.
#[proc_macro_attribute]
pub fn kernel(input: proc_macro::TokenStream, item: proc_macro::TokenStream) -> TokenStream {
    let cloned = input.clone();
    let _ = parse_macro_input!(input as KernelHints);
    let input = parse_macro_input!(cloned as proc_macro2::TokenStream);
    let mut item = parse_macro_input!(item as ItemFn);
    let no_mangle = parse_quote!(#[no_mangle]);
    item.attrs.push(no_mangle);
    let internal = parse_quote!(#[cfg_attr(any(target_arch="nvptx", target_arch="nvptx64"), nvvm_internal(kernel(#input)))]);
    item.attrs.push(internal);

    // used to guarantee some things about how params are passed in the codegen.
    item.sig.abi = Some(parse_quote!(extern "C"));

    let check_fn = parse_quote! {
        fn assert_kernel_parameter_is_copy<T: Copy>() {}
    };
    item.block.stmts.insert(0, check_fn);

    for param in &item.sig.inputs {
        let ty = match param {
            FnArg::Receiver(_) => quote_spanned! {
                param.span() => ::core::compile_error!("Kernel functions may not be struct methods");
            },
            FnArg::Typed(ty) => ty.ty.to_token_stream(),
        };
        let call = parse_quote! {
            assert_kernel_parameter_is_copy::<#ty>();
        };
        item.block.stmts.insert(0, call);
    }

    let ret = item.sig.output.clone();
    if let ReturnType::Type(_, _) = ret {
        let err = quote_spanned! {
            ret.span() => ::core::compile_err!("Kernel functions should not return anything");
        }
        .into();
        item.block.stmts.insert(0, parse_macro_input!(err as Stmt));
    }

    if item.sig.unsafety.is_none() {
        let err = quote_spanned! {
            item.span() => ::core::compile_error!("Kernel functions must be marked as unsafe");
        }
        .into();
        item.block.stmts.insert(0, parse_macro_input!(err as Stmt));
    }

    item.to_token_stream().into()
}

#[derive(Debug, Clone, Copy, PartialEq)]
enum Dimension {
    Dim1,
    Dim2,
    Dim3,
}

impl Parse for Dimension {
    fn parse(input: syn::parse::ParseStream) -> syn::Result<Self> {
        let val = Ident::parse(input)?;
        let val = val.to_string();
        match val.as_str() {
            "1d" | "1D" => Ok(Self::Dim1),
            "2d" | "2D" => Ok(Self::Dim2),
            "3d" | "3D" => Ok(Self::Dim3),
            _ => Err(syn::Error::new(Span::call_site(), "Invalid dimension")),
        }
    }
}

enum KernelHint {
    GridDim(Dimension),
    BlockDim(Dimension),
}

impl Parse for KernelHint {
    fn parse(input: syn::parse::ParseStream) -> syn::Result<Self> {
        let name = Ident::parse(input)?;
        let key = name.to_string();
        <Token![=]>::parse(input)?;
        match key.as_str() {
            "grid_dim" => {
                let dim = Dimension::parse(input)?;
                Ok(Self::GridDim(dim))
            }
            "block_dim" => {
                let dim = Dimension::parse(input)?;
                Ok(Self::BlockDim(dim))
            }
            _ => Err(Error::new(Span::call_site(), "Unrecognized option")),
        }
    }
}

#[derive(Debug, Default, Clone, PartialEq)]
struct KernelHints {
    grid_dim: Option<Dimension>,
    block_dim: Option<Dimension>,
}

impl Parse for KernelHints {
    fn parse(input: syn::parse::ParseStream) -> syn::Result<Self> {
        let iter = Punctuated::<KernelHint, Token![,]>::parse_terminated(input)?;
        let hints = iter
            .into_pairs()
            .map(|x| x.into_value())
            .collect::<Vec<_>>();

        let mut out = KernelHints::default();

        for hint in hints {
            match hint {
                KernelHint::GridDim(dim) => out.grid_dim = Some(dim),
                KernelHint::BlockDim(dim) => out.block_dim = Some(dim),
            }
        }

        Ok(out)
    }
}

// derived from rust-gpu's gpu_only

/// Creates a cpu version of the function which panics and cfg-gates the function for only nvptx/nvptx64.
#[proc_macro_attribute]
pub fn gpu_only(_attr: proc_macro::TokenStream, item: proc_macro::TokenStream) -> TokenStream {
    let syn::ItemFn {
        attrs,
        vis,
        sig,
        block,
    } = syn::parse_macro_input!(item as syn::ItemFn);

    let mut cloned_attrs = attrs.clone();
    cloned_attrs.retain(|a| {
        !a.path
            .get_ident()
            .map(|x| *x == "nvvm_internal")
            .unwrap_or_default()
    });

    let fn_name = sig.ident.clone();

    let sig_cpu = syn::Signature {
        abi: None,
        ..sig.clone()
    };

    let output = quote::quote! {
        #[cfg(not(any(target_arch="nvptx", target_arch="nvptx64")))]
        #[allow(unused_variables)]
        #(#cloned_attrs)* #vis #sig_cpu {
            unimplemented!(concat!("`", stringify!(#fn_name), "` can only be used on the GPU with rustc_codegen_nvvm"))
        }

        #[cfg(any(target_arch="nvptx", target_arch="nvptx64"))]
        #(#attrs)* #vis #sig {
            #block
        }
    };

    output.into()
}

/// Notifies the codegen that this function is externally visible and should not be
/// removed if it is not used by a kernel. Usually used for linking with other PTX/cubin files.
///
/// # Panics
///
/// Panics if the function is not also no_mangle.
#[proc_macro_attribute]
pub fn externally_visible(
    _attr: proc_macro::TokenStream,
    item: proc_macro::TokenStream,
) -> TokenStream {
    let mut func = syn::parse_macro_input!(item as syn::ItemFn);

    assert!(
        func.attrs.iter().any(|a| a.path.is_ident("no_mangle")),
        "#[externally_visible] function should also be #[no_mangle]"
    );

    let new_attr = parse_quote!(#[cfg_attr(target_os = "cuda", nvvm_internal(used))]);
    func.attrs.push(new_attr);

    func.into_token_stream().into()
}

/// Notifies the codegen to put a `static`/`static mut` inside of a specific memory address space.
/// This is mostly for internal use and/or advanced users, as the codegen and `cuda_std` handle address space placement
/// implicitly. **Improper use of this macro could yield weird or undefined behavior**.
///
/// This macro takes a single argument which can either be `global`, `shared`, `constant`, or `local`.
///
/// This macro does nothing on the CPU.
#[proc_macro_attribute]
pub fn address_space(attr: proc_macro::TokenStream, item: proc_macro::TokenStream) -> TokenStream {
    let mut global = syn::parse_macro_input!(item as syn::ItemStatic);
    let input = syn::parse_macro_input!(attr as Ident);

    let addrspace_num = match input.to_string().as_str() {
        "global" => 1,
        // what did you do to address space 2 libnvvm??
        "shared" => 3,
        "constant" => 4,
        "local" => 5,
        addr => panic!("Invalid address space `{}`", addr),
    };

    let new_attr =
        parse_quote!(#[cfg_attr(target_os = "cuda", nvvm_internal(addrspace(#addrspace_num)))]);
    global.attrs.push(new_attr);

    global.into_token_stream().into()
}