more pack

crates/generate/src/generate.rs

@@ -198,6 +198,7 @@ impl VariableType {
         let (is_signed, elem_width) = match etype {
             "SI8" => (true, 8),
             "SI16" => (true, 16),
+            "SI32" => (true, 32),
             "UI8" => (false, 8),
             "UI16" => (false, 16),
             _ => bail!("unknown element type: {etype}"),

crates/generate/src/main.rs

@@ -88,4 +88,10 @@ fn main() -> Result<()> {
     Ok(())
 }
 
-const INTRINSICS_GENERATE: &[&str] = &["_mm_packus_epi16", "_mm_packs_epi16", "_mm_setr_epi16"];
+const INTRINSICS_GENERATE: &[&str] = &[
+    "_mm_packus_epi16",
+    "_mm_packs_epi16",
+    "_mm_packus_epi32",
+    "_mm_packs_epi32",
+    "_mm_setr_epi16",
+];

crates/intringen/src/x86/generated.rs

@@ -84,6 +84,37 @@ pub trait Intrinsics: super::Core {
         let __tmp = self.saturate8(__tmp);
         self.set_lane___m128i_i8(dst, 15u64, __tmp);
     }
+    fn _mm_packs_epi32(
+        &mut self,
+        dst: &mut Self::__m128i,
+        a: Self::__m128i,
+        b: Self::__m128i,
+    ) {
+        let __tmp = self.get_lane___m128i_i32(a, 0u64);
+        let __tmp = self.saturate16(__tmp);
+        self.set_lane___m128i_i16(dst, 0u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(a, 1u64);
+        let __tmp = self.saturate16(__tmp);
+        self.set_lane___m128i_i16(dst, 1u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(a, 2u64);
+        let __tmp = self.saturate16(__tmp);
+        self.set_lane___m128i_i16(dst, 2u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(a, 3u64);
+        let __tmp = self.saturate16(__tmp);
+        self.set_lane___m128i_i16(dst, 3u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(b, 0u64);
+        let __tmp = self.saturate16(__tmp);
+        self.set_lane___m128i_i16(dst, 4u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(b, 1u64);
+        let __tmp = self.saturate16(__tmp);
+        self.set_lane___m128i_i16(dst, 5u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(b, 2u64);
+        let __tmp = self.saturate16(__tmp);
+        self.set_lane___m128i_i16(dst, 6u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(b, 3u64);
+        let __tmp = self.saturate16(__tmp);
+        self.set_lane___m128i_i16(dst, 7u64, __tmp);
+    }
     fn _mm_packus_epi16(
         &mut self,
         dst: &mut Self::__m128i,
@@ -139,6 +170,37 @@ pub trait Intrinsics: super::Core {
         let __tmp = self.saturate_u8(__tmp);
         self.set_lane___m128i_u8(dst, 15u64, __tmp);
     }
+    fn _mm_packus_epi32(
+        &mut self,
+        dst: &mut Self::__m128i,
+        a: Self::__m128i,
+        b: Self::__m128i,
+    ) {
+        let __tmp = self.get_lane___m128i_i32(a, 0u64);
+        let __tmp = self.saturate_u16(__tmp);
+        self.set_lane___m128i_u16(dst, 0u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(a, 1u64);
+        let __tmp = self.saturate_u16(__tmp);
+        self.set_lane___m128i_u16(dst, 1u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(a, 2u64);
+        let __tmp = self.saturate_u16(__tmp);
+        self.set_lane___m128i_u16(dst, 2u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(a, 3u64);
+        let __tmp = self.saturate_u16(__tmp);
+        self.set_lane___m128i_u16(dst, 3u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(b, 0u64);
+        let __tmp = self.saturate_u16(__tmp);
+        self.set_lane___m128i_u16(dst, 4u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(b, 1u64);
+        let __tmp = self.saturate_u16(__tmp);
+        self.set_lane___m128i_u16(dst, 5u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(b, 2u64);
+        let __tmp = self.saturate_u16(__tmp);
+        self.set_lane___m128i_u16(dst, 6u64, __tmp);
+        let __tmp = self.get_lane___m128i_i32(b, 3u64);
+        let __tmp = self.saturate_u16(__tmp);
+        self.set_lane___m128i_u16(dst, 7u64, __tmp);
+    }
 }
 pub mod soft_arch {
     pub use super::super::soft_arch_types::*;
@@ -163,11 +225,21 @@ pub mod soft_arch {
         super::super::ValueCore._mm_packs_epi16(&mut output, a, b);
         output
     }
+    pub fn _mm_packs_epi32(a: __m128i, b: __m128i) -> __m128i {
+        let mut output = unsafe { std::mem::zeroed() };
+        super::super::ValueCore._mm_packs_epi32(&mut output, a, b);
+        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
     }
+    pub fn _mm_packus_epi32(a: __m128i, b: __m128i) -> __m128i {
+        let mut output = unsafe { std::mem::zeroed() };
+        super::super::ValueCore._mm_packus_epi32(&mut output, a, b);
+        output
+    }
 }
 #[cfg(all(test, target_arch = "x86_64"))]
 pub mod tests {
@@ -190,13 +262,31 @@ pub mod tests {
         }
     }
     #[test]
-    fn _mm_packus_epi16() {
+    fn _mm_packs_epi32() {
         hard_soft_same_128! {
             { let a = _mm_setr_epi16(18077i16, 23617i16, - 9205i16, 21233i16, - 4332i16,
             - 31339i16, 23623i16, - 22080i16); let b = _mm_setr_epi16(- 1436i16, -
             30227i16, 8629i16, 10922i16, - 16731i16, - 1013i16, - 14310i16, 2892i16);
+            _mm_packs_epi32(a, b) }
+        }
+    }
+    #[test]
+    fn _mm_packus_epi16() {
+        hard_soft_same_128! {
+            { let a = _mm_setr_epi16(- 28568i16, 12614i16, 20103i16, 32412i16, -
+            28704i16, - 27930i16, 4197i16, 1829i16); let b = _mm_setr_epi16(9149i16,
+            18759i16, 30885i16, - 3879i16, 21600i16, 24454i16, 23524i16, 10765i16);
             _mm_packus_epi16(a, b) }
         }
     }
+    #[test]
+    fn _mm_packus_epi32() {
+        hard_soft_same_128! {
+            { let a = _mm_setr_epi16(32539i16, 26890i16, - 3892i16, 4386i16, 18704i16,
+            8253i16, - 29217i16, 32013i16); let b = _mm_setr_epi16(7448i16, 2172i16, -
+            14764i16, - 1068i16, - 25463i16, 21215i16, - 31392i16, - 14015i16);
+            _mm_packus_epi32(a, b) }
+        }
+    }
 }
 

crates/intringen/src/x86/mod.rs

@@ -21,13 +21,18 @@ pub trait Core {
 
     fn get_lane___m128i_u16(&mut self, value: Self::__m128i, idx: u64) -> Self::u16;
     fn get_lane___m128i_i16(&mut self, value: Self::__m128i, idx: u64) -> Self::i16;
+    fn get_lane___m128i_u32(&mut self, value: Self::__m128i, idx: u64) -> Self::u32;
+    fn get_lane___m128i_i32(&mut self, value: Self::__m128i, idx: u64) -> Self::i32;
 
     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 set_lane___m128i_i16(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::i16);
 
     fn saturate8(&mut self, elem: Self::i16) -> Self::i8;
     fn saturate_u8(&mut self, elem: Self::i16) -> Self::u8;
+    fn saturate16(&mut self, elem: Self::i32) -> Self::i16;
+    fn saturate_u16(&mut self, elem: Self::i32) -> Self::u16;
 }
 
 pub struct ValueCore;
@@ -57,10 +62,20 @@ impl Core for ValueCore {
     }
 
     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];
+        self.get_lane___m128i_u16(value, idx) as i16
+    }
 
-        ((((first as u16) << 8) as u16) | (second as u16)) as i16
+    fn get_lane___m128i_u32(&mut self, value: Self::__m128i, idx: u64) -> Self::u32 {
+        let first = value[(idx * 4 + 3) as usize];
+        let second = value[(idx * 4 + 2) as usize];
+        let third = value[(idx * 4 + 1) as usize];
+        let fourth = value[(idx * 4) as usize];
+
+        ((first as u32) << 24) | ((second as u32) << 16) | ((third as u32) << 8) | (fourth as u32)
+    }
+
+    fn get_lane___m128i_i32(&mut self, value: Self::__m128i, idx: u64) -> Self::i32 {
+        self.get_lane___m128i_u32(value, idx) as i32
     }
 
     fn set_lane___m128i_u8(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::u8) {
@@ -68,7 +83,7 @@ impl Core for ValueCore {
     }
 
     fn set_lane___m128i_i8(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::i8) {
-        place[idx as usize] = value as u8;
+        self.set_lane___m128i_u8(place, idx, value as u8);
     }
 
     fn set_lane___m128i_u16(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::u16) {
@@ -78,6 +93,10 @@ impl Core for ValueCore {
         place[(idx * 2 + 1) as usize] = second;
     }
 
+    fn set_lane___m128i_i16(&mut self, place: &mut Self::__m128i, idx: u64, value: Self::i16) {
+        self.set_lane___m128i_u16(place, idx, value as u16);
+    }
+
     fn saturate8(&mut self, elem: Self::i16) -> Self::i8 {
         let clamp = elem.clamp(i8::MIN as i16, i8::MAX as i16);
         clamp as i8
@@ -87,6 +106,15 @@ impl Core for ValueCore {
         let clamp = elem.clamp(0, u8::MAX as i16);
         clamp as u8
     }
+
+    fn saturate16(&mut self, elem: Self::i32) -> Self::i16 {
+        let clamp = elem.clamp(i16::MIN as i32, i16::MAX as i32);
+        clamp as i16
+    }
+    fn saturate_u16(&mut self, elem: Self::i32) -> Self::u16 {
+        let clamp = elem.clamp(0, u16::MAX as i32);
+        clamp as u16
+    }
 }
 
 mod soft_arch_types {