MirrorHub/construct
0
0
Fork 0
mirror of https://github.com/matrix-construct/construct synced 2024-11-29 10:12:39 +01:00
Code Issues Releases Wiki Activity

ircd::simd: Add clz/ctz convenience template suite for now; minor reorg.

Browse source
This commit is contained in:
Jason Volk 2020-06-29 21:05:09 -07:00
parent aec21af0e4
commit 35bee76625
1 changed files with 96 additions and 27 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

123
include/ircd/simd/simd.h
View file

@ -19,7 +19,11 @@
namespace ircd::simd
{
template<class T> T popmask(const T) noexcept;
template<class T> size_t popcount(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;
// xmmx shifter workarounds
template<int bits, class T> T shl(const T &a) noexcept;
@ -37,32 +41,6 @@ namespace ircd
using simd::lane_cast;
}
/// 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 size_t
ircd::simd::popcount(const T a)
noexcept
{
size_t i(0), ret(0);
while(i < lanes(a))
ret += __builtin_popcountll(a[i++]);
return ret;
}
/// 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;
}
#ifdef HAVE_X86INTRIN_H
template<int b>
[[using gnu: always_inline, gnu_inline, artificial]]
@ -126,3 +104,94 @@ noexcept
return _mm256_slli_si256(a, b / 8);
}
#endif
/// Convenience template. Unfortunately this drops to scalar until specific
/// targets and specializations are created.
template<class T>
inline uint
ircd::simd::clz(const T a)
noexcept
{
uint ret(0), i(0); do
{
const auto mask
{
boolmask(uint(ret == sizeof_lane<T>() * 8 * i))
};
if constexpr(sizeof_lane<T>() <= sizeof(u16))
ret += __lzcnt16(__builtin_bswap16(a[i++])) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u32))
ret += __lzcnt32(__builtin_bswap32(a[i++])) & mask;
else
ret += __lzcnt64(__builtin_bswap64(a[i++])) & mask;
}
while(i < lanes(a));
return ret;
}
/// Convenience template. Unfortunately this drops to scalar until specific
/// targets and specializations are created.
template<class T>
inline uint
ircd::simd::ctz(const T a)
noexcept
{
uint ret(0), i(lanes(a)), mask(-1U); do
{
if constexpr(sizeof_lane<T>() <= sizeof(u16))
ret += __lzcnt16(a[--i]) & mask;
else if constexpr(sizeof_lane<T>() <= sizeof(u32))
ret += __lzcnt32(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;
}