AuroraMiddleware/fmtlegacy
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Optimize grisu using uint128_t

Browse Source
This commit is contained in:
Victor Zverovich 2019-04-13 13:04:27 -07:00
parent 41fbaeb3b1
commit bade46aae5
2 changed files with 15 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
include/fmt/format-inl.h
View File

@ -440,6 +440,13 @@ inline fp operator-(fp x, fp y) {
// with half-up tie breaking, r.e = x.e + y.e + 64. Result may not be
// normalized.
FMT_FUNC fp operator*(fp x, fp y) {
int exp = x.e + y.e + 64;
#if FMT_USE_INT128
auto product = static_cast<__uint128_t>(x.f) * y.f;
auto f = static_cast<uint64_t>(product >> 64);
if ((static_cast<uint64_t>(product) & (1ULL << 63)) != 0) ++f;
return fp(f, exp);
#else
// Multiply 32-bit parts of significands.
uint64_t mask = (1ULL << 32) - 1;
uint64_t a = x.f >> 32, b = x.f & mask;
@ -447,7 +454,8 @@ FMT_FUNC fp operator*(fp x, fp y) {
uint64_t ac = a * c, bc = b * c, ad = a * d, bd = b * d;
// Compute mid 64-bit of result and round.
uint64_t mid = (bd >> 32) + (ad & mask) + (bc & mask) + (1U << 31);
return fp(ac + (ad >> 32) + (bc >> 32) + (mid >> 32), x.e + y.e + 64);
return fp(ac + (ad >> 32) + (bc >> 32) + (mid >> 32), exp);
#endif
}
// Returns cached power (of 10) c_k = c_k.f * pow(2, c_k.e) such that its

8
include/fmt/format.h
View File

@ -176,8 +176,12 @@ FMT_END_NAMESPACE
# define FMT_USE_TRAILING_RETURN 0
#endif
#ifndef FMT_USE_INT128
# define FMT_USE_INT128 (__SIZEOF_INT128__ != 0)
#ifdef FMT_USE_INT128
// Do nothing.
#elif defined(__SIZEOF_INT128__)
# define FMT_USE_INT128 1
#else
# define FMT_USE_INT128 0
#endif
// __builtin_clz is broken in clang with Microsoft CodeGen: