compare hard and soft

crates/generate/src/generate.rs

@@ -6,17 +6,35 @@ use eyre::{bail, Context, OptionExt, Result};
 
 pub fn generate(intrinsics: &[Intrinsic]) -> Result<()> {
     println!("impl<C: super::Core> Intrinsics for C {{}}");
-    println!("trait Intrinsics: super::Core {{");
+    println!("pub trait Intrinsics: super::Core {{");
     for intr in intrinsics {
         generate_intr(intr).wrap_err_with(|| format!("generating `{}`", intr.name))?;
     }
 
     println!("}}");
 
+    println!();
+    generate_soft_arch_module(intrinsics).context("generating soft_arch module")?;
+
+    Ok(())
+}
+
+fn generate_soft_arch_module(intrinsics: &[Intrinsic]) -> Result<()> {
+    println!("pub mod soft_arch {{");
+    println!("    pub use super::super::soft_arch_types::*;");
+    println!("    use super::Intrinsics;");
+
+    for intr in intrinsics {
+        generate_intr_soft_arch_wrap(intr)
+            .wrap_err_with(|| format!("generating soft_arch `{}`", intr.name))?;
+    }
+
+    println!("}}");
     Ok(())
 }
 
 fn generate_intr(intr: &Intrinsic) -> Result<(), eyre::Error> {
+    eprintln!("generating {}...", intr.name);
     let signature = signature(intr)?;
     println!("    {signature} {{");
     let body = generate_body(intr).wrap_err("generating body")?;
@@ -25,6 +43,37 @@ fn generate_intr(intr: &Intrinsic) -> Result<(), eyre::Error> {
     Ok(())
 }
 
+fn generate_intr_soft_arch_wrap(intr: &Intrinsic) -> Result<(), eyre::Error> {
+    eprintln!("generating soft_arch wrapper {}...", intr.name);
+    let signature = signature_soft_arch(intr)?;
+    println!("    {signature} {{");
+    let body = generate_body_soft_arch(intr).wrap_err("generating body")?;
+    println!("{}", indent(&body, 8));
+    println!("    }}");
+    Ok(())
+}
+
+fn generate_body_soft_arch(intr: &Intrinsic) -> Result<String> {
+    let mut rust_stmts = Vec::<String>::new();
+
+    rust_stmts.push("let mut output = unsafe { std::mem::zeroed() };".into());
+
+    let name = &intr.name;
+    let args = intr
+        .parameter
+        .iter()
+        .map(|param| param.varname.as_deref().ok_or_eyre("parameter has no name"))
+        .collect::<Result<Vec<_>>>()?
+        .join(", ");
+    rust_stmts.push(format!(
+        "super::super::ValueCore.{name}(&mut output, {args});"
+    ));
+
+    rust_stmts.push("output".into());
+
+    Ok(rust_stmts.join("\n"))
+}
+
 fn indent(input: &str, indent: usize) -> String {
     let replace = format!("\n{}", " ".repeat(indent));
     let mut s = " ".repeat(indent);
@@ -34,6 +83,7 @@ fn indent(input: &str, indent: usize) -> String {
 
 struct VariableType {
     is_signed: bool,
+    rawtype_signed: bool,
     elem_width: u64,
     #[allow(dead_code)]
     full_width: u64,
@@ -42,20 +92,23 @@ struct VariableType {
 
 impl VariableType {
     fn of(etype: &str, ty: &str) -> Result<Self> {
-        let full_width = match ty {
+        let (rawtype_signed, full_width) = match map_type_to_rust(ty) {
-            "__m128i" => 128,
+            "__m128i" => (false, 128),
+            "i16" => (true, 16),
             _ => bail!("unknown type: {ty}"),
         };
         let (is_signed, elem_width) = match etype {
             "SI16" => (true, 16),
             "UI8" => (false, 8),
+            "UI16" => (false, 16),
             _ => bail!("unknown element type: {etype}"),
         };
         Ok(Self {
             is_signed,
+            rawtype_signed,
             full_width,
             elem_width,
-            raw_type: ty.to_owned(),
+            raw_type: map_type_to_rust(ty).to_owned(),
         })
     }
 
@@ -144,7 +197,18 @@ fn generate_expr_tmp(
 ) -> Result<String> {
     let result = match expr {
         Expr::Int(int) => int.to_string(),
-        Expr::Ident(_) => todo!(),
+        Expr::Ident(ident) => {
+            let ty = type_of_ident(&ident)?;
+            if ty.is_signed != ty.rawtype_signed {
+                let from = &ty.raw_type;
+                let to = ty.rust_type();
+                let stmt = format!("let __tmp = self.cast_sign_{from}_{to}({ident});");
+                rust_stmts.push(stmt);
+                "__tmp".into()
+            } else {
+                ident
+            }
+        }
         Expr::Index { lhs, idx } => {
             let Expr::Ident(ident) = *lhs else {
                 bail!("lhs of indexing must be identifier");
@@ -215,16 +279,44 @@ fn signature(intr: &Intrinsic) -> Result<String> {
             format!(
                 "{}: Self::{}",
                 param.varname.as_ref().unwrap(),
-                param.r#type.as_ref().unwrap()
+                map_type_to_rust(param.r#type.as_ref().unwrap())
             )
         })
         .collect::<Vec<_>>()
         .join(", ");
 
     let ret_name = intr.ret.varname.as_ref().unwrap();
-    let ret_ty = intr.ret.r#type.as_ref().unwrap();
+    let ret_ty = map_type_to_rust(intr.ret.r#type.as_ref().unwrap());
 
     Ok(format!(
         "fn {name}(&mut self, {ret_name}: &mut Self::{ret_ty}, {args})"
     ))
 }
+
+fn signature_soft_arch(intr: &Intrinsic) -> Result<String> {
+    let name = &intr.name;
+
+    let args = intr
+        .parameter
+        .iter()
+        .map(|param| {
+            format!(
+                "{}: {}",
+                param.varname.as_ref().unwrap(),
+                map_type_to_rust(param.r#type.as_ref().unwrap())
+            )
+        })
+        .collect::<Vec<_>>()
+        .join(", ");
+
+    let ret_ty = map_type_to_rust(intr.ret.r#type.as_ref().unwrap());
+
+    Ok(format!("pub fn {name}({args}) -> {ret_ty}"))
+}
+
+fn map_type_to_rust(ty: &str) -> &str {
+    match ty {
+        "short" => "i16",
+        ty => ty,
+    }
+}

crates/generate/src/main.rs

@@ -76,7 +76,7 @@ fn main() -> Result<()> {
     let list = list
         .into_iter()
         .filter(|intr| intr.cpuid.iter().any(|cpu| !cpu.value.contains("AVX512")))
-        .filter(|intr| intr.name == "_mm_packus_epi16")
+        .filter(|intr| INTRINSICS_GENERATE.contains(&intr.name.as_str()))
         .collect::<Vec<_>>();
 
     eprintln!("filtered: {}", list.len());
@@ -85,3 +85,5 @@ fn main() -> Result<()> {
 
     Ok(())
 }
+
+const INTRINSICS_GENERATE: &[&str] = &["_mm_packus_epi16", "_mm_setr_epi16"];

crates/intringen/src/x86/generated.rs

@@ -1,5 +1,23 @@
 impl<C: super::Core> Intrinsics for C {}
-trait Intrinsics: super::Core {
+pub trait Intrinsics: super::Core {
+    fn _mm_setr_epi16(&mut self, dst: &mut Self::__m128i, e7: Self::i16, e6: Self::i16, e5: Self::i16, e4: Self::i16, e3: Self::i16, e2: Self::i16, e1: Self::i16, e0: Self::i16) {
+        let __tmp = self.cast_sign_i16_u16(e7);
+        self.set_lane___m128i_u16(dst, 0, __tmp);
+        let __tmp = self.cast_sign_i16_u16(e6);
+        self.set_lane___m128i_u16(dst, 1, __tmp);
+        let __tmp = self.cast_sign_i16_u16(e5);
+        self.set_lane___m128i_u16(dst, 2, __tmp);
+        let __tmp = self.cast_sign_i16_u16(e4);
+        self.set_lane___m128i_u16(dst, 3, __tmp);
+        let __tmp = self.cast_sign_i16_u16(e3);
+        self.set_lane___m128i_u16(dst, 4, __tmp);
+        let __tmp = self.cast_sign_i16_u16(e2);
+        self.set_lane___m128i_u16(dst, 5, __tmp);
+        let __tmp = self.cast_sign_i16_u16(e1);
+        self.set_lane___m128i_u16(dst, 6, __tmp);
+        let __tmp = self.cast_sign_i16_u16(e0);
+        self.set_lane___m128i_u16(dst, 7, __tmp);
+    }
     fn _mm_packus_epi16(&mut self, dst: &mut Self::__m128i, a: Self::__m128i, b: Self::__m128i) {
         let __tmp = self.get_lane___m128i_i16(a, 0);
         let __tmp = self.saturate_u8(__tmp);
@@ -51,3 +69,18 @@ trait Intrinsics: super::Core {
         self.set_lane___m128i_u8(dst, 15, __tmp);
     }
 }
+
+pub mod soft_arch {
+    pub use super::super::soft_arch_types::*;
+    use super::Intrinsics;
+    pub fn _mm_setr_epi16(e7: i16, e6: i16, e5: i16, e4: i16, e3: i16, e2: i16, e1: i16, e0: i16) -> __m128i {
+        let mut output = unsafe { std::mem::zeroed() };
+        super::super::ValueCore._mm_setr_epi16(&mut output, e7, e6, e5, e4, e3, e2, e1, e0);
+        output
+    }
+    pub fn _mm_packus_epi16(a: __m128i, b: __m128i) -> __m128i {
+        let mut output = unsafe { std::mem::zeroed() };
+        super::super::ValueCore._mm_packus_epi16(&mut output, a, b);
+        output
+    }
+}

crates/intringen/src/x86/mod.rs

@@ -2,48 +2,90 @@
 
 mod generated;
 
+pub use generated::soft_arch;
+
 pub trait Core {
+    type u8: Copy;
+    type u16: Copy;
+    type u32: Copy;
+    type u64: Copy;
+
+    type i8: Copy;
+    type i16: Copy;
+    type i32: Copy;
+    type i64: Copy;
+
     type __m128i: Copy;
 
-    fn get_lane___m128i_u16(&mut self, value: Self::__m128i, idx: u64) -> u16;
+    fn cast_sign_i16_u16(&mut self, value: Self::i16) -> Self::u16;
-    fn get_lane___m128i_i16(&mut self, value: Self::__m128i, idx: u64) -> i16;
 
-    fn set_lane___m128i_u8(&mut self, place: &mut Self::__m128i, idx: u64, value: u8);
+    fn get_lane___m128i_u16(&mut self, value: Self::__m128i, idx: u64) -> Self::u16;
-    fn set_lane___m128i_i8(&mut self, place: &mut Self::__m128i, idx: u64, value: i8);
+    fn get_lane___m128i_i16(&mut self, value: Self::__m128i, idx: u64) -> Self::i16;
 
-    fn saturate_u8(&mut self, elem: i16) -> u8;
+    fn set_lane___m128i_u8(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::u8);
+    fn set_lane___m128i_i8(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::i8);
+    fn set_lane___m128i_u16(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::u16);
+
+    fn saturate_u8(&mut self, elem: Self::i16) -> Self::u8;
 }
 
 pub struct ValueCore;
 
 impl Core for ValueCore {
+    type u8 = u8;
+    type u16 = u16;
+    type u32 = u32;
+    type u64 = u64;
+
+    type i8 = i8;
+    type i16 = i16;
+    type i32 = i32;
+    type i64 = i64;
+
     type __m128i = [u8; 16];
 
-    fn get_lane___m128i_u16(&mut self, value: Self::__m128i, idx: u64) -> u16 {
+    fn cast_sign_i16_u16(&mut self, value: Self::i16) -> Self::u16 {
-        let first = value[(idx * 2) as usize];
+        value as _
-        let second = value[(idx * 2 + 1) as usize];
-        // todo: le? be?
-        ((first << 8) as u16) | (second as u16)
     }
 
-    fn get_lane___m128i_i16(&mut self, value: Self::__m128i, idx: u64) -> i16 {
+    fn get_lane___m128i_u16(&mut self, value: Self::__m128i, idx: u64) -> Self::u16 {
-        let first = value[(idx * 2) as usize];
+        let first = value[(idx * 2 + 1) as usize];
-        let second = value[(idx * 2 + 1) as usize];
+        let second = value[(idx * 2) as usize];
-        // todo: le? be?
+
-        (((first << 8) as u16) | (second as u16)) as i16
+        ((first as u16) << 8) | (second as u16)
     }
 
-    fn set_lane___m128i_u8(&mut self, place: &mut Self::__m128i, idx: u64, value: u8) {
+    fn get_lane___m128i_i16(&mut self, value: Self::__m128i, idx: u64) -> Self::i16 {
+        let first = value[(idx * 2 + 1) as usize];
+        let second = value[(idx * 2) as usize];
+
+        ((((first as u16) << 8) as u16) | (second as u16)) as i16
+    }
+
+    fn set_lane___m128i_u8(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::u8) {
         place[idx as usize] = value;
     }
 
-    fn set_lane___m128i_i8(&mut self, place: &mut Self::__m128i, idx: u64, value: i8) {
+    fn set_lane___m128i_i8(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::i8) {
         place[idx as usize] = value as u8;
     }
 
-    fn saturate_u8(&mut self, elem: i16) -> u8 {
+    fn set_lane___m128i_u16(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::u16) {
+        let first = (value & 0xFF) as u8;
+        let second = (value >> 8) as u8;
+        place[(idx * 2) as usize] = first;
+        place[(idx * 2 + 1) as usize] = second;
+    }
+
+    fn saturate_u8(&mut self, elem: Self::i16) -> Self::u8 {
         let clamp = elem.clamp(0, u8::MAX as i16);
         clamp as u8
     }
 }
-pub trait Lanes<const N: usize, Elem> {}
+
+mod soft_arch_types {
+    pub type __m128i = [u8; 16];
+}
+
+#[cfg(test)]
+mod tests;

crates/intringen/src/x86/tests.rs

@@ -0,0 +1,34 @@
+#[cfg(target_arch = "x86_64")]
+mod x86_compare {
+    macro_rules! hard_soft_same_128 {
+        ($($stmt:tt)*) => {
+            let soft = {
+                use crate::x86::soft_arch::*;
+                $($stmt)*
+            };
+            let hard = unsafe {
+                std::mem::transmute::<_, [u8; 16]>({
+                    use core::arch::x86_64::*;
+                    $($stmt)*
+                })
+            };
+            assert_eq!(soft, hard);
+        };
+    }
+
+    #[test]
+    fn _mm_setr_epi16() {
+        hard_soft_same_128! {
+            _mm_setr_epi16(0, -1, 100, 3535, 35, 2, i16::MIN, i16::MAX)
+        };
+    }
+
+    #[test]
+    fn _mm_packus_epi16() {
+        hard_soft_same_128! {
+            let a = _mm_setr_epi16(0x100, -1, 0, 0, 0, 0, 0, 0);
+            let b = _mm_setr_epi16(0, 0, 0, 0, 0, 0, -1, 0x100);
+            _mm_packus_epi16(a, b)
+        }
+    }
+}