summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorThiago Macieira <thiago.macieira@nokia.com>2011-03-18 17:02:08 (GMT)
committerThiago Macieira <thiago.macieira@nokia.com>2011-03-22 14:51:53 (GMT)
commit152f40fa41bae551dffd7b0ed72ed2008d7ff642 (patch)
tree17c567dcf476489fef0ea1465f959feddfa9f604
parentaa7543e735c5c0fe61049c432b264ad1f2cfd598 (diff)
downloadQt-152f40fa41bae551dffd7b0ed72ed2008d7ff642.zip
Qt-152f40fa41bae551dffd7b0ed72ed2008d7ff642.tar.gz
Qt-152f40fa41bae551dffd7b0ed72ed2008d7ff642.tar.bz2
Add an SSE4 version using PMOVZXBW and PSRLDQ
-rw-r--r--tests/benchmarks/corelib/tools/qstring/main.cpp22
1 files changed, 22 insertions, 0 deletions
diff --git a/tests/benchmarks/corelib/tools/qstring/main.cpp b/tests/benchmarks/corelib/tools/qstring/main.cpp
index c365e5a..4e73f14 100644
--- a/tests/benchmarks/corelib/tools/qstring/main.cpp
+++ b/tests/benchmarks/corelib/tools/qstring/main.cpp
@@ -1618,6 +1618,27 @@ void fromLatin1_sse2_withprolog(ushort *dst, const char *str, int size)
fromLatin1_epilog(dst, str, size);
}
+void fromLatin1_sse4_pmovzxbw(ushort *dst, const char *str, int size)
+{
+ while (size >= 16) {
+ __m128i chunk = _mm_loadu_si128((__m128i*)str); // load
+
+ // unpack the first 8 bytes, padding with zeros
+ const __m128i firstHalf = _mm_cvtepu8_epi16(chunk);
+ _mm_storeu_si128((__m128i*)dst, firstHalf); // store
+
+ // unpack the last 8 bytes, padding with zeros
+ chunk = _mm_srli_si128(chunk, 8);
+ const __m128i secondHalf = _mm_cvtepu8_epi16(chunk);
+ _mm_storeu_si128((__m128i*)(dst + 8), secondHalf); // store
+
+ str += 16;
+ dst += 16;
+ size -= 16;
+ }
+ fromLatin1_epilog(dst, str, size);
+}
+
void fromLatin1_prolog_sse4_overcommit(ushort *dst, const char *str, int)
{
// load 8 bytes and zero-extend them to 16
@@ -1636,6 +1657,7 @@ void tst_QString::fromLatin1Alternatives_data() const
QTest::newRow("sse2-with-prolog-unrolled") << &fromLatin1_sse2_withprolog<&fromLatin1_prolog_unrolled>;
QTest::newRow("sse2-with-prolog-sse2-overcommit") << &fromLatin1_sse2_withprolog<&fromLatin1_prolog_sse2_overcommit>;
QTest::newRow("sse2-with-prolog-sse4-overcommit") << &fromLatin1_sse2_withprolog<&fromLatin1_prolog_sse4_overcommit>;
+ QTest::newRow("sse4-pmovzxbw") << &fromLatin1_sse4_pmovzxbw;
}
extern StringData fromLatin1Data;