mirror of
https://github.com/matrix-construct/construct
synced 2024-11-25 16:22:35 +01:00
280 lines
5.4 KiB
C++
280 lines
5.4 KiB
C++
// The Construct
|
|
//
|
|
// Copyright (C) The Construct Developers, Authors & Contributors
|
|
// Copyright (C) 2016-2020 Jason Volk <jason@zemos.net>
|
|
//
|
|
// Permission to use, copy, modify, and/or distribute this software for any
|
|
// purpose with or without fee is hereby granted, provided that the above
|
|
// copyright notice and this permission notice is present in all copies. The
|
|
// full license for this software is available in the LICENSE file.
|
|
|
|
#pragma once
|
|
#define HAVE_IRCD_SIMD_REDUCE_H
|
|
|
|
namespace ircd::simd
|
|
{
|
|
/// Perform a horizontal left-reduce among lanes. The operation is given
|
|
/// by the caller who supplies a functor template like `std::bit_or` or
|
|
/// `std::plus` etc. The result resides in lane[0] of the return vector
|
|
/// while all other lanes of the return vector are undefined/junk as far
|
|
/// as the caller is concerned.
|
|
///
|
|
/// This operation is intended to "reduce" or "collapse" a vector to a
|
|
/// scalar value generally to make some control transfer etc. It does not
|
|
/// necessitate a scalar result so it can be integrated into a sequence of
|
|
/// vector operations without loss of purity. But of course, this operation
|
|
/// is not efficient (crossing lanes never really is) and this template
|
|
/// will output some log2 number of instructions. Using larger lane widths
|
|
/// (i.e u64 rather than u8) can decrease the number of operations.
|
|
///
|
|
template<template<class>
|
|
class op,
|
|
class T>
|
|
T reduce(T) noexcept = delete;
|
|
|
|
template<template<class> class op> u8x16 reduce(u8x16) noexcept;
|
|
template<template<class> class op> u8x32 reduce(u8x32) noexcept;
|
|
template<template<class> class op> u8x64 reduce(u8x64) noexcept;
|
|
|
|
template<template<class> class op> u16x8 reduce(u16x8) noexcept;
|
|
template<template<class> class op> u16x16 reduce(u16x16) noexcept;
|
|
template<template<class> class op> u16x32 reduce(u16x32) noexcept;
|
|
|
|
template<template<class> class op> u32x4 reduce(u32x4) noexcept;
|
|
template<template<class> class op> u32x8 reduce(u32x8) noexcept;
|
|
template<template<class> class op> u32x16 reduce(u32x16) noexcept;
|
|
|
|
template<template<class> class op> u64x2 reduce(u64x2) noexcept;
|
|
template<template<class> class op> u64x4 reduce(u64x4) noexcept;
|
|
template<template<class> class op> u64x8 reduce(u64x8) noexcept;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u64x8
|
|
ircd::simd::reduce(u64x8 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u64x4 b, c;
|
|
for(size_t i(0); i < 4; ++i)
|
|
b[i] = a[i], c[i] = a[i + 4];
|
|
|
|
b = oper(b, c);
|
|
b = reduce<op>(b);
|
|
a[0] = b[0];
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u64x4
|
|
ircd::simd::reduce(u64x4 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u64x2 b, c;
|
|
for(size_t i(0); i < 2; ++i)
|
|
b[i] = a[i], c[i] = a[i + 2];
|
|
|
|
b = oper(b, c);
|
|
b = reduce<op>(b);
|
|
a[0] = b[0];
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u64x2
|
|
ircd::simd::reduce(u64x2 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
const u64x2 b
|
|
{
|
|
a[1], a[0]
|
|
};
|
|
|
|
a = oper(a, b);
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u32x16
|
|
ircd::simd::reduce(u32x16 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u32x8 b, c;
|
|
for(size_t i(0); i < 8; ++i)
|
|
b[i] = a[i], c[i] = a[i + 8];
|
|
|
|
b = oper(b, c);
|
|
b = reduce<op>(b);
|
|
a[0] = b[0];
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u32x8
|
|
ircd::simd::reduce(u32x8 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u32x4 b, c;
|
|
for(size_t i(0); i < 4; ++i)
|
|
b[i] = a[i], c[i] = a[i + 4];
|
|
|
|
b = oper(b, c);
|
|
b = reduce<op>(b);
|
|
a[0] = b[0];
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u32x4
|
|
ircd::simd::reduce(u32x4 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u32x4 b
|
|
{
|
|
a[2], a[3], 0, 0,
|
|
};
|
|
|
|
a = oper(a, b);
|
|
b[0] = a[1];
|
|
a = oper(a, b);
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u16x32
|
|
ircd::simd::reduce(u16x32 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u16x16 b, c;
|
|
for(size_t i(0); i < 16; ++i)
|
|
b[i] = a[i], c[i] = a[i + 16];
|
|
|
|
b = oper(b, c);
|
|
b = reduce<op>(b);
|
|
a[0] = b[0];
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u16x16
|
|
ircd::simd::reduce(u16x16 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u16x8 b, c;
|
|
for(size_t i(0); i < 8; ++i)
|
|
b[i] = a[i], c[i] = a[i + 8];
|
|
|
|
b = oper(b, c);
|
|
b = reduce<op>(b);
|
|
a[0] = b[0];
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u16x8
|
|
ircd::simd::reduce(u16x8 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u16x8 b
|
|
{
|
|
a[4], a[5], a[6], a[7]
|
|
};
|
|
|
|
a = oper(a, b);
|
|
b[0] = a[2];
|
|
b[1] = a[3];
|
|
a = oper(a, b);
|
|
b[0] = a[1];
|
|
a = oper(a, b);
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u8x64
|
|
ircd::simd::reduce(u8x64 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u8x32 b, c;
|
|
for(size_t i(0); i < 32; ++i)
|
|
b[i] = a[i], c[i] = a[i + 32];
|
|
|
|
b = oper(b, c);
|
|
b = reduce<op>(b);
|
|
a[0] = b[0];
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u8x32
|
|
ircd::simd::reduce(u8x32 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u8x16 b, c;
|
|
for(size_t i(0); i < 16; ++i)
|
|
b[i] = a[i], c[i] = a[i + 16];
|
|
|
|
b = oper(b, c);
|
|
b = reduce<op>(b);
|
|
a[0] = b[0];
|
|
return a;
|
|
}
|
|
|
|
template<template<class>
|
|
class op>
|
|
inline ircd::u8x16
|
|
ircd::simd::reduce(u8x16 a)
|
|
noexcept
|
|
{
|
|
static const op oper;
|
|
|
|
u8x16 b
|
|
{
|
|
a[0x8], a[0x9], a[0xa], a[0xb], a[0xc], a[0xd], a[0xe], a[0xf],
|
|
};
|
|
|
|
a = oper(a, b);
|
|
b = u8x16
|
|
{
|
|
a[0x4], a[0x5], a[0x6], a[0x7],
|
|
};
|
|
|
|
a = oper(a, b);
|
|
b[0x0] = a[0x2];
|
|
b[0x1] = a[0x3];
|
|
a = oper(a, b);
|
|
b[0x0] = a[0x1];
|
|
a = oper(a, b);
|
|
return a;
|
|
}
|