0
0
Fork 0
mirror of https://github.com/matrix-construct/construct synced 2024-11-25 08:12:37 +01:00

ircd::simd: Split bit manipulations suites to headers; renames.

This commit is contained in:
Jason Volk 2020-07-17 16:41:22 -07:00
parent 291038a014
commit 3a70943b94
5 changed files with 208 additions and 140 deletions

66
include/ircd/simd/lzcnt.h Normal file
View file

@ -0,0 +1,66 @@
// 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_LZCNT_H
namespace ircd::simd
{
template<class T> uint lzcnt(const T) noexcept;
}
/// Convenience template. Unfortunately this drops to scalar until specific
/// targets and specializations are created.
template<class T>
inline uint
__attribute__((target("lzcnt")))
ircd::simd::lzcnt(const T a)
noexcept
{
// The behavior of lzcnt can differ among platforms; when true we expect
// lzcnt to fall back to bsr-like behavior.
constexpr auto bitscan
{
#ifdef __LZCNT__
false
#else
true
#endif
};
uint ret(0), i(0); do
{
const auto mask
{
boolmask(uint(ret == sizeof_lane<T>() * 8 * i))
};
if constexpr(bitscan && sizeof_lane<T>() <= sizeof(u16))
ret += (15 - __lzcnt16(__builtin_bswap16(a[i++]))) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u16))
ret += __lzcnt16(__builtin_bswap16(a[i++])) & mask;
else if constexpr(bitscan && sizeof_lane<T>() <= sizeof(u32))
ret += (31 - __lzcnt32(__builtin_bswap32(a[i++]))) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u32))
ret += __lzcnt32(__builtin_bswap32(a[i++])) & mask;
else if constexpr(bitscan)
ret += (63 - __lzcnt64(__builtin_bswap64(a[i++]))) & mask;
else
ret += __lzcnt64(__builtin_bswap64(a[i++])) & mask;
}
while(i < lanes<T>());
return ret;
}

View file

@ -0,0 +1,59 @@
// 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_POPCNT_H
namespace ircd::simd
{
template<class T> T popmask(const T) noexcept;
template<class T> T boolmask(const T) noexcept;
template<class T> uint popcnt(const T) noexcept;
}
/// Convenience template. Unfortunately this drops to scalar until specific
/// targets and specializations are created.
template<class T>
inline uint
ircd::simd::popcnt(const T a)
noexcept
{
uint ret(0), i(0);
for(; i < lanes<T>(); ++i)
if constexpr(sizeof_lane<T>() <= sizeof(int))
ret += __builtin_popcount(a[i]);
else if constexpr(sizeof_lane<T>() <= sizeof(long))
ret += __builtin_popcountl(a[i]);
else
ret += __builtin_popcountll(a[i]);
return ret;
}
/// Convenience template. Extends a bool value where the lsb is 1 or 0 into a
/// mask value like the result of vector comparisons.
template<class T>
inline T
ircd::simd::boolmask(const T a)
noexcept
{
return ~(popmask(a) - 1);
}
/// Convenience template. Vector compare instructions yield 0xff on equal;
/// sometimes one might need an actual value of 1 for accumulators or maybe
/// some bool-type reason...
template<class T>
inline T
ircd::simd::popmask(const T a)
noexcept
{
return a & 1;
}

View file

@ -16,144 +16,22 @@
#include "lane_cast.h"
#include "broad_cast.h"
#include "shift.h"
#include "popcnt.h"
#include "lzcnt.h"
#include "tzcnt.h"
#include "print.h"
namespace ircd::simd
{
template<class T> T popmask(const T) noexcept;
template<class T> T boolmask(const T) noexcept;
template<class T> uint popcount(const T) noexcept;
template<class T> uint clz(const T) noexcept;
template<class T> uint ctz(const T) noexcept;
}
namespace ircd
{
using simd::lane_cast;
using simd::shl;
using simd::shr;
}
/// Convenience template. Unfortunately this drops to scalar until specific
/// targets and specializations are created.
template<class T>
inline uint
__attribute__((target("lzcnt")))
ircd::simd::clz(const T a)
noexcept
{
constexpr auto uses_bsr
{
#ifndef __LZCNT__
true
#else
false
#endif
};
uint ret(0), i(0); do
{
const auto mask
{
boolmask(uint(ret == sizeof_lane<T>() * 8 * i))
};
if constexpr(sizeof_lane<T>() <= sizeof(u16) && uses_bsr)
ret += (15 - __lzcnt16(__builtin_bswap16(a[i++]))) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u16))
ret += __lzcnt16(__builtin_bswap16(a[i++])) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u32) && uses_bsr)
ret += (31 - __lzcnt32(__builtin_bswap32(a[i++]))) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u32))
ret += __lzcnt32(__builtin_bswap32(a[i++])) & mask;
else if constexpr(uses_bsr)
ret += (63 - __lzcnt64(__builtin_bswap64(a[i++]))) & mask;
else
ret += __lzcnt64(__builtin_bswap64(a[i++])) & mask;
}
while(i < lanes<T>());
return ret;
}
/// Convenience template. Unfortunately this drops to scalar until specific
/// targets and specializations are created.
template<class T>
inline uint
__attribute__((target("lzcnt")))
ircd::simd::ctz(const T a)
noexcept
{
constexpr auto uses_bsr
{
#ifndef __LZCNT__
true
#else
false
#endif
};
uint ret(0), i(lanes<T>()), mask(-1U); do
{
if constexpr(sizeof_lane<T>() <= sizeof(u16) && uses_bsr)
ret += (15 - __lzcnt16(a[--i])) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u16))
ret += __lzcnt16(a[--i]) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u32) && uses_bsr)
ret += (31 - __lzcnt32(a[--i])) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u32))
ret += __lzcnt32(a[--i]) & mask;
else if constexpr(uses_bsr)
ret += (63 - __lzcnt64(a[--i])) & mask;
else
ret += __lzcnt64(a[--i]) & mask;
static const auto lane_bits(sizeof_lane<T>() * 8);
mask &= boolmask(uint(ret % lane_bits == 0));
mask &= boolmask(uint(ret != 0));
}
while(i);
return ret;
}
/// Convenience template. Unfortunately this drops to scalar until specific
/// targets and specializations are created.
template<class T>
inline uint
ircd::simd::popcount(const T a)
noexcept
{
uint ret(0), i(0);
for(; i < lanes<T>(); ++i)
if constexpr(sizeof_lane<T>() <= sizeof(int))
ret += __builtin_popcount(a[i]);
else if constexpr(sizeof_lane<T>() <= sizeof(long))
ret += __builtin_popcountl(a[i]);
else
ret += __builtin_popcountll(a[i]);
return ret;
}
/// Convenience template. Extends a bool value where the lsb is 1 or 0 into a
/// mask value like the result of vector comparisons.
template<class T>
inline T
ircd::simd::boolmask(const T a)
noexcept
{
return ~(popmask(a) - 1);
}
/// Convenience template. Vector compare instructions yield 0xff on equal;
/// sometimes one might need an actual value of 1 for accumulators or maybe
/// some bool-type reason...
template<class T>
inline T
ircd::simd::popmask(const T a)
noexcept
{
return a & 1;
using simd::lzcnt;
using simd::tzcnt;
using simd::popcnt;
using simd::popmask;
using simd::boolmask;
}

65
include/ircd/simd/tzcnt.h Normal file
View file

@ -0,0 +1,65 @@
// 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_TZCNT_H
namespace ircd::simd
{
template<class T> uint tzcnt(const T) noexcept;
}
/// Convenience template. Unfortunately this drops to scalar until specific
/// targets and specializations are created.
template<class T>
inline uint
__attribute__((target("lzcnt")))
ircd::simd::tzcnt(const T a)
noexcept
{
// The behavior of lzcnt/tzcnt can differ among platforms; when false we
// lzcnt/tzcnt to fall back to bsr/bsf-like behavior.
constexpr auto bitscan
{
#ifdef __LZCNT__
false
#else
true
#endif
};
uint ret(0), i(lanes<T>()), mask(-1U); do
{
if constexpr(bitscan && sizeof_lane<T>() <= sizeof(u16))
ret += (15 - __lzcnt16(a[--i])) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u16))
ret += __lzcnt16(a[--i]) & mask;
else if constexpr(bitscan && sizeof_lane<T>() <= sizeof(u32))
ret += (31 - __lzcnt32(a[--i])) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u32))
ret += __lzcnt32(a[--i]) & mask;
else if constexpr(bitscan)
ret += (63 - __lzcnt64(a[--i])) & mask;
else
ret += __lzcnt64(a[--i]) & mask;
static const auto lane_bits(sizeof_lane<T>() * 8);
mask &= boolmask(uint(ret % lane_bits == 0));
mask &= boolmask(uint(ret != 0));
}
while(i);
return ret;
}

View file

@ -3405,7 +3405,7 @@ ircd::json::string_stringify(u8x16 &__restrict__ block,
// counting past the end of the string.
const u64 regular_prefix_count
{
simd::clz(u64x2(is_special | ~block_mask)) / 8
simd::lzcnt(u64x2(is_special | ~block_mask)) / 8
};
// Fast-path; backward branch to count and consume uninteresting characters
@ -3460,7 +3460,7 @@ ircd::json::string_stringify(u8x16 &__restrict__ block,
const u64 match_depth
{
simd::clz(u64x2(match)) / 8
simd::lzcnt(u64x2(match)) / 8
};
// Legitimately escaped single char
@ -3568,7 +3568,7 @@ ircd::json::string_stringify_utf16(u8x16 &__restrict__ block,
const auto surrogates
{
simd::popcount(u64x2(simd::popmask(u8x16(is_surrogate))))
simd::popcnt(u64x2(simd::popmask(u8x16(is_surrogate))))
};
assert(di == length[0] + length[1]);
@ -3647,7 +3647,7 @@ ircd::json::string_serialized(const u8x16 block,
const u64 regular_prefix_count
{
simd::clz(u64x2(is_special | ~block_mask)) / 8
simd::lzcnt(u64x2(is_special | ~block_mask)) / 8
};
// Fast-path; backward branch to count and consume uninteresting characters
@ -3691,7 +3691,7 @@ ircd::json::string_serialized(const u8x16 block,
const u64 match_depth
{
simd::clz(u64x2(match)) / 8
simd::lzcnt(u64x2(match)) / 8
};
// Legitimately escaped single char
@ -3779,7 +3779,7 @@ ircd::json::string_serialized_utf16(const u8x16 block,
const auto surrogates
{
simd::popcount(u64x2(simd::popmask(u8x16(is_surrogate))))
simd::popcnt(u64x2(simd::popmask(u8x16(is_surrogate))))
};
return u64x2
@ -3801,7 +3801,7 @@ ircd::json::string_serialized_ctrl(const u8x16 block,
const u64 ctrl_prefix_count
{
simd::clz(u64x2(~is_ctrl | ~block_mask)) / 8
simd::lzcnt(u64x2(~is_ctrl | ~block_mask)) / 8
};
u64 ret(0);