summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorBenjamin Poulain <benjamin.poulain@nokia.com>2010-02-25 11:33:53 (GMT)
committerBenjamin Poulain <benjamin.poulain@nokia.com>2010-02-25 12:32:30 (GMT)
commit107a7aed0d64971136637e8ac26bf112108ec3bf (patch)
tree5e869f0aec0d63079c1262ec5f4382052ad4c743
parentdadb99ea2c59d7d0f7a83134b7df5aaaaf80a995 (diff)
downloadQt-107a7aed0d64971136637e8ac26bf112108ec3bf.zip
Qt-107a7aed0d64971136637e8ac26bf112108ec3bf.tar.gz
Qt-107a7aed0d64971136637e8ac26bf112108ec3bf.tar.bz2
Implement toLatin1_helper with Neon
Implement toLatin1 with neon to process 8 characters at a time. Using Neon improve the speed it there is at least 16 characters, there is no improvement when using it for 8 characters. Speed difference on a N900: -8 characters: no change 100% -16 characters: 126% -1000 characters: 361% -10000 characters: 423% Reviewed-by: Samuel Rødal
-rw-r--r--src/corelib/tools/qstring.cpp22
1 files changed, 22 insertions, 0 deletions
diff --git a/src/corelib/tools/qstring.cpp b/src/corelib/tools/qstring.cpp
index 995e4cf..cce25f4 100644
--- a/src/corelib/tools/qstring.cpp
+++ b/src/corelib/tools/qstring.cpp
@@ -3524,6 +3524,28 @@ static QByteArray toLatin1_helper(const QChar *data, int length)
}
length = length % 16;
}
+#elif QT_HAVE_NEON
+ // this use eactly the same method as for SSE except the packing
+ // which is done to 64 bits (8 x 8bits component).
+ // Refer to the documentation of the SSE2 implementation
+ if (length >= 16) {
+ const int chunkCount = length >> 3; // divided by 8
+ const uint16x8_t questionMark = vdupq_n_u16('?'); // set
+ const uint16x8_t thresholdMask = vdupq_n_u16(0xff); // set
+ for (int i = 0; i < chunkCount; ++i) {
+ uint16x8_t chunk = vld1q_u16((uint16_t *)src); // load
+ src += 8;
+
+ const uint16x8_t offLimitMask = vcgeq_u16(chunk, thresholdMask); // ==
+ const uint16x8_t offLimitQuestionMark = vandq_u16(offLimitMask, questionMark); // offLimitMask & questionMark
+ const uint16x8_t correctBytes = vbicq_u16(chunk, offLimitMask); // !offLimitMask & chunk
+ chunk = vorrq_u16(correctBytes, offLimitQuestionMark); // correctBytes | offLimitQuestionMark
+ const uint8x8_t result = vmovn_u16(chunk); // narrowing move->packing
+ vst1_u8(dst, result); // store
+ dst += 8;
+ }
+ length = length % 8;
+ }
#endif
while (length--) {
*dst++ = (*src>0xff) ? '?' : (uchar) *src;