Add internal ctz/clz implementation.

This commit is contained in:
Antonio Sánchez
2023-12-11 21:03:09 +00:00
parent 454f89af9d
commit 75e273afcc
3 changed files with 227 additions and 9 deletions

View File

@@ -628,6 +628,149 @@ struct meta_floor_log2<n, lower, upper, meta_floor_log2_bogus> {
// no value, error at compile time
};
template <typename BitsType, typename EnableIf = void>
struct count_bits_impl {
static_assert(std::is_integral<BitsType>::value && std::is_unsigned<BitsType>::value,
"BitsType must be an unsigned integer");
static EIGEN_DEVICE_FUNC inline int clz(BitsType bits) {
int n = CHAR_BIT * sizeof(BitsType);
int shift = n / 2;
while (bits > 0 && shift > 0) {
BitsType y = bits >> shift;
if (y > 0) {
n -= shift;
bits = y;
}
shift /= 2;
}
if (shift == 0) {
--n;
}
return n;
}
static EIGEN_DEVICE_FUNC inline int ctz(BitsType bits) {
int n = CHAR_BIT * sizeof(BitsType);
int shift = n / 2;
while (bits > 0 && shift > 0) {
BitsType y = bits << shift;
if (y > 0) {
n -= shift;
bits = y;
}
shift /= 2;
}
if (shift == 0) {
--n;
}
return n;
}
};
// Count leading zeros.
template <typename BitsType>
EIGEN_DEVICE_FUNC inline int clz(BitsType bits) {
return count_bits_impl<BitsType>::clz(bits);
}
// Count trailing zeros.
template <typename BitsType>
EIGEN_DEVICE_FUNC inline int ctz(BitsType bits) {
return count_bits_impl<BitsType>::ctz(bits);
}
#if EIGEN_COMP_GNUC || EIGEN_COMP_CLANG
template <typename BitsType>
struct count_bits_impl<BitsType, std::enable_if_t<sizeof(BitsType) <= sizeof(unsigned int)>> {
static constexpr int kNumBits = static_cast<int>(sizeof(BitsType) * CHAR_BIT);
static_assert(std::is_integral<BitsType>::value, "BitsType must be a built-in integer");
static EIGEN_DEVICE_FUNC inline int clz(BitsType bits) {
static constexpr int kLeadingBitsOffset = (sizeof(unsigned int) - sizeof(BitsType)) * CHAR_BIT;
return bits == 0 ? kNumBits : __builtin_clz(static_cast<unsigned int>(bits)) - kLeadingBitsOffset;
}
static EIGEN_DEVICE_FUNC inline int ctz(BitsType bits) {
return bits == 0 ? kNumBits : __builtin_ctz(static_cast<unsigned int>(bits));
}
};
template <typename BitsType>
struct count_bits_impl<
BitsType, std::enable_if_t<sizeof(unsigned int) < sizeof(BitsType) && sizeof(BitsType) <= sizeof(unsigned long)>> {
static constexpr int kNumBits = static_cast<int>(sizeof(BitsType) * CHAR_BIT);
static_assert(std::is_integral<BitsType>::value, "BitsType must be a built-in integer");
static EIGEN_DEVICE_FUNC inline int clz(BitsType bits) {
static constexpr int kLeadingBitsOffset = (sizeof(unsigned long) - sizeof(BitsType)) * CHAR_BIT;
return bits == 0 ? kNumBits : __builtin_clzl(static_cast<unsigned long>(bits)) - kLeadingBitsOffset;
}
static EIGEN_DEVICE_FUNC inline int ctz(BitsType bits) {
return bits == 0 ? kNumBits : __builtin_ctzl(static_cast<unsigned long>(bits));
}
};
template <typename BitsType>
struct count_bits_impl<BitsType, std::enable_if_t<sizeof(unsigned long) < sizeof(BitsType) &&
sizeof(BitsType) <= sizeof(unsigned long long)>> {
static constexpr int kNumBits = static_cast<int>(sizeof(BitsType) * CHAR_BIT);
static_assert(std::is_integral<BitsType>::value, "BitsType must be a built-in integer");
static EIGEN_DEVICE_FUNC inline int clz(BitsType bits) {
static constexpr int kLeadingBitsOffset = (sizeof(unsigned long long) - sizeof(BitsType)) * CHAR_BIT;
return bits == 0 ? kNumBits : __builtin_clzll(static_cast<unsigned long long>(bits)) - kLeadingBitsOffset;
}
static EIGEN_DEVICE_FUNC inline int ctz(BitsType bits) {
return bits == 0 ? kNumBits : __builtin_ctzll(static_cast<unsigned long long>(bits));
}
};
#elif EIGEN_COMP_MSVC
template <typename BitsType>
struct count_bits_impl<BitsType, std::enable_if_t<sizeof(BitsType) <= sizeof(unsigned long)>> {
static constexpr int kNumBits = static_cast<int>(sizeof(BitsType) * CHAR_BIT);
static_assert(std::is_integral<BitsType>::value, "BitsType must be a built-in integer");
static EIGEN_DEVICE_FUNC inline int clz(BitsType bits) {
static constexpr int kLeadingBitsOffset = static_cast<int>((sizeof(unsigned long) - sizeof(BitsType)) * CHAR_BIT);
unsigned long out;
_BitScanReverse(&out, static_cast<unsigned long>(bits));
return bits == 0 ? kNumBits : static_cast<int>(out - kLeadingBitsOffset);
}
static EIGEN_DEVICE_FUNC inline int ctz(BitsType bits) {
unsigned long out;
_BitScanForward(&out, static_cast<unsigned long>(bits));
return bits == 0 ? kNumBits : static_cast<int>(out);
}
};
#ifdef _WIN64
template <typename BitsType>
struct count_bits_impl<
BitsType, std::enable_if_t<sizeof(unsigned long) < sizeof(BitsType) && sizeof(BitsType) <= sizeof(__int64)>> {
static constexpr int kNumBits = static_cast<int>(sizeof(BitsType) * CHAR_BIT);
static_assert(std::is_integral<BitsType>::value, "BitsType must be a built-in integer");
static EIGEN_DEVICE_FUNC inline int clz(BitsType bits) {
static constexpr int kLeadingBitsOffset = static_cast<int>((sizeof(__int64) - sizeof(BitsType)) * CHAR_BIT);
unsigned long out;
_BitScanReverse64(&out, static_cast<unsigned __int64>(bits));
return bits == 0 ? kNumBits : static_cast<int>(out - kLeadingBitsOffset);
}
static EIGEN_DEVICE_FUNC inline int ctz(BitsType bits) {
unsigned long out;
_BitScanForward64(&out, static_cast<unsigned __int64>(bits));
return bits == 0 ? kNumBits : static_cast<int>(out);
}
};
#endif // _WIN64
#endif // EIGEN_COMP_GNUC || EIGEN_COMP_CLANG
template <typename Scalar>
struct random_default_impl<Scalar, false, true> {
static inline Scalar run(const Scalar& x, const Scalar& y) {