use arg_enum_proc_macro::ArgEnum;
use std::env;
use std::str::FromStr;
#[derive(Debug, Copy, Clone, PartialEq, Eq, PartialOrd, ArgEnum)]
pub enum CpuFeatureLevel {
RUST,
SSE2,
SSSE3,
#[arg_enum(alias = "sse4.1")]
SSE4_1,
AVX2,
AVX512,
#[arg_enum(alias = "avx512vpclmulqdq")]
AVX512ICL,
}
impl CpuFeatureLevel {
pub const fn len() -> usize {
CpuFeatureLevel::AVX512ICL as usize + 1
}
#[inline(always)]
pub fn as_index(self) -> usize {
self as usize
}
}
impl Default for CpuFeatureLevel {
fn default() -> CpuFeatureLevel {
fn avx512_detected() -> bool {
is_x86_feature_detected!("avx512bw")
&& is_x86_feature_detected!("avx512cd")
&& is_x86_feature_detected!("avx512dq")
&& is_x86_feature_detected!("avx512f")
&& is_x86_feature_detected!("avx512vl")
}
fn avx512icl_detected() -> bool {
avx512_detected()
&& is_x86_feature_detected!("avx512vnni")
&& is_x86_feature_detected!("avx512ifma")
&& is_x86_feature_detected!("avx512vbmi")
&& is_x86_feature_detected!("avx512vbmi2")
&& is_x86_feature_detected!("avx512vpopcntdq")
&& is_x86_feature_detected!("avx512bitalg")
&& is_x86_feature_detected!("avx512gfni")
&& is_x86_feature_detected!("avx512vaes")
&& is_x86_feature_detected!("avx512vpclmulqdq")
}
let detected: CpuFeatureLevel = if avx512icl_detected() {
CpuFeatureLevel::AVX512ICL
} else if avx512_detected() {
CpuFeatureLevel::AVX512
} else if is_x86_feature_detected!("avx2") {
CpuFeatureLevel::AVX2
} else if is_x86_feature_detected!("sse4.1") {
CpuFeatureLevel::SSE4_1
} else if is_x86_feature_detected!("ssse3") {
CpuFeatureLevel::SSSE3
} else if is_x86_feature_detected!("sse2") {
CpuFeatureLevel::SSE2
} else {
CpuFeatureLevel::RUST
};
let manual: CpuFeatureLevel = match env::var("RAV1E_CPU_TARGET") {
Ok(feature) => CpuFeatureLevel::from_str(&feature).unwrap_or(detected),
Err(_e) => detected,
};
if manual > detected {
detected
} else {
manual
}
}
}
macro_rules! cpu_function_lookup_table {
($pub:vis, $name:ident: [$type:ty], default: $empty:expr, [$(($key:ident, $value:expr)),*]) => {
$pub static $name: [$type; crate::cpu_features::CpuFeatureLevel::len()] = {
use crate::cpu_features::CpuFeatureLevel;
#[allow(unused_mut)]
let mut out: [$type; CpuFeatureLevel::len()] = [$empty; CpuFeatureLevel::len()];
#[allow(unused_mut)]
let mut set: [bool; CpuFeatureLevel::len()] = [false; CpuFeatureLevel::len()];
#[allow(unused_imports)]
use CpuFeatureLevel::*;
$(
out[$key as usize] = $value;
set[$key as usize] = true;
)*
cpu_function_lookup_table!(waterfall_cpu_features(out, set, [SSE2, SSSE3, SSE4_1, AVX2, AVX512, AVX512ICL]));
out
};
};
(waterfall_cpu_features($out:ident, $set:ident, [$($cpu:ident),*])) => {
#[allow(unused_assignments)]
let mut best = [$out[0], $out[0]];
$(
best[$set[$cpu as usize] as usize] = $out[$cpu as usize];
$out[$cpu as usize] = best[1];
)*
};
($name:ident: [$type:ty], default: $empty:expr, [$(($key:ident, $value:expr)),*]) => {
cpu_function_lookup_table!(pub(self), $name: [$type], default: $empty, [$(($key, $value)),*]);
};
($pub:vis, $name:ident: [$type:ty], default: $empty:expr, [$($key:ident),*]) => {
paste::item!{
cpu_function_lookup_table!(
$pub, $name: [$type], default: $empty, [$(($key, [<$name _$key>])),*]
);
}
};
($name:ident: [$type:ty], default: $empty:expr, [$($key:ident),*]) => {
cpu_function_lookup_table!(
pub(self), $name: [$type], default: $empty, [$($key),*]
);
};
}