AuroraMiddleware/qt5base-lts
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

QUtf8Codec: Use one 32-byte load instead of two 16-byte ones on AVX2

Browse Source 
The number of instructions is the same. But if the CPU can issue
32-byte-wide loads, this will be faster. For CPUs that would do two
16-byte loads, this is no worse than current code.

Change-Id: I8f261579aad648fdb4f0fffd1553d060b4fc852f
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
This commit is contained in:
Thiago Macieira 2018-09-12 18:10:11 -07:00
parent a0907e6ac1
commit 570ef11c28
1 changed files with 6 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
src/corelib/codecs/qutfcodec.cpp
View File

@ -70,9 +70,14 @@ static inline bool simdEncodeAscii(uchar *&dst, const ushort *&nextAscii, const
{
// do sixteen characters at a time
for ( ; end - src >= 16; src += 16, dst += 16) {
# ifdef __AVX2__
__m256i data = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src));
__m128i data1 = _mm256_castsi256_si128(data);
__m128i data2 = _mm256_extracti128_si256(data, 1);
# else
__m128i data1 = _mm_loadu_si128((const __m128i*)src);
__m128i data2 = _mm_loadu_si128(1+(const __m128i*)src);
# endif
// check if everything is ASCII
// the highest ASCII value is U+007F