coffee/fast_float
1
0
Fork 0
mirror of https://github.com/fastfloat/fast_float.git synced 2025-12-06 08:46:49 +08:00
Code Issues Packages Projects Releases Wiki Activity

* Upgrade manually vectorized code to SSE4.2 for FASTFLOAT_64BIT because it's already used SSE4.1 instructions.

Browse Source
This commit is contained in:
IRainman 2025-10-22 00:27:36 +03:00
parent 1f0e2819e0
commit 58b5dc04c3
1 changed files with 28 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

35
include/fast_float/ascii_number.h
View File

@ -74,9 +74,10 @@ read8_to_u64(UC const *chars) {
fastfloat_really_inline uint64_t simd_read8_to_u64(__m128i const &data) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
// _mm_packus_epi16 is SSE4.1+, converts 8×u16 → 8×u8
__m128i const packed = _mm_packus_epi16(data, data);
#ifdef FASTFLOAT_64BIT
return uint64_t(_mm_cvtsi128_si64(packed));
return static_cast<uint64_t>(_mm_cvtsi128_si64(packed));
#else
uint64_t value;
// Visual Studio + older versions of GCC don't support _mm_storeu_si64
@ -109,7 +110,7 @@ fastfloat_really_inline uint64_t simd_read8_to_u64(char16_t const *chars) {
FASTFLOAT_SIMD_RESTORE_WARNINGS
}
#endif // FASTFLOAT_SSE2
#endif
// MSVC SFINAE is broken pre-VS2017
#if defined(_MSC_VER) && _MSC_VER <= 1900
@ -164,20 +165,40 @@ simd_parse_if_eight_digits_unrolled(char16_t const *chars,
}
#ifdef FASTFLOAT_SSE2
FASTFLOAT_SIMD_DISABLE_WARNINGS
// Load 8 UTF-16 characters (16 bytes)
__m128i const data =
_mm_loadu_si128(reinterpret_cast<__m128i const *>(chars));
// (x - '0') <= 9
// http://0x80.pl/articles/simd-parsing-int-sequences.html
__m128i const t0 = _mm_add_epi16(data, _mm_set1_epi16(32720));
__m128i const t1 = _mm_cmpgt_epi16(t0, _mm_set1_epi16(-32759));
#ifdef FASTFLOAT_64BIT
// --- Digit range check using SSE4.2 comparisons ---
// Validate: '0' (0x30) ≤ x ≤ '9' (0x39)
const __m128i ascii0 = _mm_set1_epi16(u'0');
const __m128i ascii9 = _mm_set1_epi16(u'9');
if (_mm_movemask_epi8(t1) == 0) {
__m128i below0 = _mm_cmplt_epi16(data, ascii0); // x < '0'
__m128i above9 = _mm_cmpgt_epi16(data, ascii9); // x > '9'
__m128i invalid = _mm_or_si128(below0, above9);
// Check if any invalid byte exists
if (_mm_testz_si128(invalid, invalid)) { // SSE4.1/4.2: zero flag test
#else
// Branchless "are all digits?" trick from Lemire:
// (x - '0') <= 9 <=> (x + 32720) <= 32729
// encoded as signed comparison: (x + 32720) > -32759 ? not digit : digit
// http://0x80.pl/articles/simd-parsing-int-sequences.html
__m128i const adjust = _mm_set1_epi16(32720);
__m128i const cutoff = _mm_set1_epi16(-32759);
__m128i const t0 = _mm_add_epi16(data, adjust);
__m128i const mask = _mm_cmpgt_epi16(t0, cutoff);
// If mask == 0 → all digits valid.
if (_mm_movemask_epi8(mask) == 0) {
i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data));
return true;
} else
return false;
FASTFLOAT_SIMD_RESTORE_WARNINGS
#endif
#elif defined(FASTFLOAT_NEON)
FASTFLOAT_SIMD_DISABLE_WARNINGS
uint16x8_t const data = vld1q_u16(reinterpret_cast<uint16_t const *>(chars));