Implement avx512bf16 intrinsics (#998)

library/stdarch/ci/docker/i586-unknown-linux-gnu/Dockerfile

@@ -1,4 +1,4 @@
-FROM ubuntu:18.04
+FROM ubuntu:20.04
 RUN apt-get update && apt-get install -y --no-install-recommends \
   gcc-multilib \
   libc6-dev \

library/stdarch/ci/docker/i686-unknown-linux-gnu/Dockerfile

@@ -1,4 +1,4 @@
-FROM ubuntu:18.04
+FROM ubuntu:20.04
 RUN apt-get update && apt-get install -y --no-install-recommends \
   gcc-multilib \
   libc6-dev \

library/stdarch/ci/docker/x86_64-unknown-linux-gnu-emulated/Dockerfile

@@ -1,4 +1,4 @@
-FROM ubuntu:18.04
+FROM ubuntu:20.04
 RUN apt-get update && apt-get install -y --no-install-recommends \
   gcc \
   libc6-dev \
@@ -10,4 +10,5 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 
 RUN wget https://github.com/gnzlbg/intel_sde/raw/master/sde-external-8.35.0-2019-03-11-lin.tar.bz2
 RUN tar -xjf sde-external-8.35.0-2019-03-11-lin.tar.bz2
+ENV SKIP_TESTS="avx512bf16"
 ENV CARGO_TARGET_X86_64_UNKNOWN_LINUX_GNU_RUNNER="/sde-external-8.35.0-2019-03-11-lin/sde64 -rtm_mode full --"

library/stdarch/ci/docker/x86_64-unknown-linux-gnu/Dockerfile

@@ -1,4 +1,4 @@
-FROM ubuntu:18.04
+FROM ubuntu:20.04
 RUN apt-get update && apt-get install -y --no-install-recommends \
   gcc \
   libc6-dev \

library/stdarch/ci/run.sh

@@ -56,6 +56,9 @@ cargo_test() {
             ;;
     esac
 
+    if [ "$SKIP_TESTS" != "" ]; then
+        cmd="$cmd --skip "$SKIP_TESTS
+    fi
     $cmd
 }
 

library/stdarch/crates/core_arch/src/x86/mod.rs

@@ -296,6 +296,37 @@ types! {
     /// suffixed with "pd" (or otherwise contain "pd"). Not to be confused with
     /// "ps" which is used for `__m512`.
     pub struct __m512d(f64, f64, f64, f64, f64, f64, f64, f64);
+
+    /// 128-bit wide set of eight 'u16' types, x86-specific
+    ///
+    /// This type is representing a 128-bit SIMD register which internally is consisted of
+    /// eight packed `u16` instances. It's purpose is for bf16 related intrinsic
+    /// implementations.
+    pub struct __m128bh(u16, u16, u16, u16, u16, u16, u16, u16);
+
+    /// 256-bit wide set of 16 'u16' types, x86-specific
+    ///
+    /// This type is the same as the `__m128bh` type defined by Intel,
+    /// representing a 256-bit SIMD register which internally is consisted of
+    /// 16 packed `u16` instances. It's purpose is for bf16 related intrinsic
+    /// implementations.
+    pub struct __m256bh(
+        u16, u16, u16, u16, u16, u16, u16, u16,
+        u16, u16, u16, u16, u16, u16, u16, u16
+    );
+
+    /// 512-bit wide set of 32 'u16' types, x86-specific
+    ///
+    /// This type is the same as the `__m128bh` type defined by Intel,
+    /// representing a 512-bit SIMD register which internally is consisted of
+    /// 32 packed `u16` instances. It's purpose is for bf16 related intrinsic
+    /// implementations.
+    pub struct __m512bh(
+        u16, u16, u16, u16, u16, u16, u16, u16,
+        u16, u16, u16, u16, u16, u16, u16, u16,
+        u16, u16, u16, u16, u16, u16, u16, u16,
+        u16, u16, u16, u16, u16, u16, u16, u16
+    );
 }
 
 /// The `__mmask64` type used in AVX-512 intrinsics, a 64-bit integer
@@ -602,6 +633,105 @@ impl m512dExt for __m512d {
     }
 }
 
+#[allow(non_camel_case_types)]
+#[unstable(feature = "stdsimd_internal", issue = "none")]
+pub(crate) trait m128bhExt: Sized {
+    fn as_m128bh(self) -> __m128bh;
+
+    #[inline]
+    fn as_u16x8(self) -> crate::core_arch::simd::u16x8 {
+        unsafe { transmute(self.as_m128bh()) }
+    }
+
+    #[inline]
+    fn as_i16x8(self) -> crate::core_arch::simd::i16x8 {
+        unsafe { transmute(self.as_m128bh()) }
+    }
+
+    #[inline]
+    fn as_u32x4(self) -> crate::core_arch::simd::u32x4 {
+        unsafe { transmute(self.as_m128bh()) }
+    }
+
+    #[inline]
+    fn as_i32x4(self) -> crate::core_arch::simd::i32x4 {
+        unsafe { transmute(self.as_m128bh()) }
+    }
+}
+
+impl m128bhExt for __m128bh {
+    #[inline]
+    fn as_m128bh(self) -> Self {
+        self
+    }
+}
+
+#[allow(non_camel_case_types)]
+#[unstable(feature = "stdsimd_internal", issue = "none")]
+pub(crate) trait m256bhExt: Sized {
+    fn as_m256bh(self) -> __m256bh;
+
+    #[inline]
+    fn as_u16x16(self) -> crate::core_arch::simd::u16x16 {
+        unsafe { transmute(self.as_m256bh()) }
+    }
+
+    #[inline]
+    fn as_i16x16(self) -> crate::core_arch::simd::i16x16 {
+        unsafe { transmute(self.as_m256bh()) }
+    }
+
+    #[inline]
+    fn as_u32x8(self) -> crate::core_arch::simd::u32x8 {
+        unsafe { transmute(self.as_m256bh()) }
+    }
+
+    #[inline]
+    fn as_i32x8(self) -> crate::core_arch::simd::i32x8 {
+        unsafe { transmute(self.as_m256bh()) }
+    }
+}
+
+impl m256bhExt for __m256bh {
+    #[inline]
+    fn as_m256bh(self) -> Self {
+        self
+    }
+}
+
+#[allow(non_camel_case_types)]
+#[unstable(feature = "stdsimd_internal", issue = "none")]
+pub(crate) trait m512bhExt: Sized {
+    fn as_m512bh(self) -> __m512bh;
+
+    #[inline]
+    fn as_u16x32(self) -> crate::core_arch::simd::u16x32 {
+        unsafe { transmute(self.as_m512bh()) }
+    }
+
+    #[inline]
+    fn as_i16x32(self) -> crate::core_arch::simd::i16x32 {
+        unsafe { transmute(self.as_m512bh()) }
+    }
+
+    #[inline]
+    fn as_u32x16(self) -> crate::core_arch::simd::u32x16 {
+        unsafe { transmute(self.as_m512bh()) }
+    }
+
+    #[inline]
+    fn as_i32x16(self) -> crate::core_arch::simd::i32x16 {
+        unsafe { transmute(self.as_m512bh()) }
+    }
+}
+
+impl m512bhExt for __m512bh {
+    #[inline]
+    fn as_m512bh(self) -> Self {
+        self
+    }
+}
+
 mod eflags;
 pub use self::eflags::*;
 
@@ -725,3 +855,6 @@ pub use self::rtm::*;
 
 mod f16c;
 pub use self::f16c::*;
+
+mod avx512bf16;
+pub use self::avx512bf16::*;

library/stdarch/crates/stdarch-verify/src/lib.rs

@@ -137,12 +137,15 @@ fn to_type(t: &syn::Type) -> proc_macro2::TokenStream {
         syn::Type::Path(ref p) => match extract_path_ident(&p.path).to_string().as_ref() {
             // x86 ...
             "__m128" => quote! { &M128 },
+            "__m128bh" => quote! { &M128BH },
             "__m128d" => quote! { &M128D },
             "__m128i" => quote! { &M128I },
             "__m256" => quote! { &M256 },
+            "__m256bh" => quote! { &M256BH },
             "__m256d" => quote! { &M256D },
             "__m256i" => quote! { &M256I },
             "__m512" => quote! { &M512 },
+            "__m512bh" => quote! { &M512BH },
             "__m512d" => quote! { &M512D },
             "__m512i" => quote! { &M512I },
             "__mmask8" => quote! { &MMASK8 },

library/stdarch/crates/stdarch-verify/tests/x86-intel.rs

@@ -45,12 +45,15 @@ static ORDERING: Type = Type::Ordering;
 
 static M64: Type = Type::M64;
 static M128: Type = Type::M128;
+static M128BH: Type = Type::M128BH;
 static M128I: Type = Type::M128I;
 static M128D: Type = Type::M128D;
 static M256: Type = Type::M256;
+static M256BH: Type = Type::M256BH;
 static M256I: Type = Type::M256I;
 static M256D: Type = Type::M256D;
 static M512: Type = Type::M512;
+static M512BH: Type = Type::M512BH;
 static M512I: Type = Type::M512I;
 static M512D: Type = Type::M512D;
 static MMASK8: Type = Type::MMASK8;
@@ -75,12 +78,15 @@ enum Type {
     ConstPtr(&'static Type),
     M64,
     M128,
+    M128BH,
     M128D,
     M128I,
     M256,
+    M256BH,
     M256D,
     M256I,
     M512,
+    M512BH,
     M512D,
     M512I,
     MMASK8,
@@ -493,6 +499,9 @@ fn matches(rust: &Function, intel: &Intrinsic) -> Result<(), String> {
             // The intrinsics guide calls `f16c` `fp16c` in disagreement with
             // Intel's architecture manuals.
             "fp16c" => String::from("f16c"),
+            "avx512_bf16" => String::from("avx512bf16"),
+            // The XML file names VNNI as "avx512_bf16", while Rust calls
+            // it "avx512bf16".
             _ => cpuid,
         };
         let fixed_cpuid = fixup_cpuid(cpuid);
@@ -693,12 +702,15 @@ fn equate(t: &Type, intel: &str, intrinsic: &str, is_const: bool) -> Result<(),
         (&Type::PrimUnsigned(8), "unsigned char") => {}
         (&Type::M64, "__m64") => {}
         (&Type::M128, "__m128") => {}
+        (&Type::M128BH, "__m128bh") => {}
         (&Type::M128I, "__m128i") => {}
         (&Type::M128D, "__m128d") => {}
         (&Type::M256, "__m256") => {}
+        (&Type::M256BH, "__m256bh") => {}
         (&Type::M256I, "__m256i") => {}
         (&Type::M256D, "__m256d") => {}
         (&Type::M512, "__m512") => {}
+        (&Type::M512BH, "__m512bh") => {}
         (&Type::M512I, "__m512i") => {}
         (&Type::M512D, "__m512d") => {}
         (&Type::MMASK64, "__mmask64") => {}
@@ -726,12 +738,15 @@ fn equate(t: &Type, intel: &str, intrinsic: &str, is_const: bool) -> Result<(),
         (&Type::MutPtr(&Type::PrimUnsigned(64)), "__mmask64*") => {}
         (&Type::MutPtr(&Type::M64), "__m64*") => {}
         (&Type::MutPtr(&Type::M128), "__m128*") => {}
+        (&Type::MutPtr(&Type::M128BH), "__m128bh*") => {}
         (&Type::MutPtr(&Type::M128I), "__m128i*") => {}
         (&Type::MutPtr(&Type::M128D), "__m128d*") => {}
         (&Type::MutPtr(&Type::M256), "__m256*") => {}
+        (&Type::MutPtr(&Type::M256BH), "__m256bh*") => {}
         (&Type::MutPtr(&Type::M256I), "__m256i*") => {}
         (&Type::MutPtr(&Type::M256D), "__m256d*") => {}
         (&Type::MutPtr(&Type::M512), "__m512*") => {}
+        (&Type::MutPtr(&Type::M512BH), "__m512bh*") => {}
         (&Type::MutPtr(&Type::M512I), "__m512i*") => {}
         (&Type::MutPtr(&Type::M512D), "__m512d*") => {}
 
@@ -754,12 +769,15 @@ fn equate(t: &Type, intel: &str, intrinsic: &str, is_const: bool) -> Result<(),
         (&Type::ConstPtr(&Type::PrimUnsigned(32)), "void const*") => {}
         (&Type::ConstPtr(&Type::M64), "__m64 const*") => {}
         (&Type::ConstPtr(&Type::M128), "__m128 const*") => {}
+        (&Type::ConstPtr(&Type::M128BH), "__m128bh const*") => {}
         (&Type::ConstPtr(&Type::M128I), "__m128i const*") => {}
         (&Type::ConstPtr(&Type::M128D), "__m128d const*") => {}
         (&Type::ConstPtr(&Type::M256), "__m256 const*") => {}
+        (&Type::ConstPtr(&Type::M256BH), "__m256bh const*") => {}
         (&Type::ConstPtr(&Type::M256I), "__m256i const*") => {}
         (&Type::ConstPtr(&Type::M256D), "__m256d const*") => {}
         (&Type::ConstPtr(&Type::M512), "__m512 const*") => {}
+        (&Type::ConstPtr(&Type::M512BH), "__m512bh const*") => {}
         (&Type::ConstPtr(&Type::M512I), "__m512i const*") => {}
         (&Type::ConstPtr(&Type::M512D), "__m512d const*") => {}
         (&Type::ConstPtr(&Type::PrimUnsigned(32)), "__mmask32*") => {}