Optimize unpremultiply on SSE4.1

Adds an SSE4.1 optimized version of qUnpremultiply and uses it in the
most drawing conversions methods. This gives a speed-up of little over
2x.

Change-Id: Ieb858a94ada1eb86d7af715ac1a100f1587f360d
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>

4 files changed, 96 insertions, 2 deletions

diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp
index ac73d78afe..83372b198d 100644
--- a/src/gui/painting/qdrawhelper.cpp
+++ b/src/gui/painting/qdrawhelper.cpp
@@ -45,6 +45,7 @@
 #include <private/qpaintengine_raster_p.h>
 #include <private/qpainter_p.h>
 #include <private/qdrawhelper_x86_p.h>
+#include <private/qdrawingprimitive_sse2_p.h>
 #include <private/qdrawhelper_neon_p.h>
 #if defined(QT_COMPILER_SUPPORTS_MIPS_DSP) || defined(QT_COMPILER_SUPPORTS_MIPS_DSPR2)
 #include <private/qdrawhelper_mips_dsp_p.h>
@@ -572,6 +573,18 @@ static const uint *QT_FASTCALL convertARGB32FromARGB32PM(uint *buffer, const uin
     return buffer;
 }
 
+#if QT_COMPILER_SUPPORTS_HERE(SSE4_1)
+QT_FUNCTION_TARGET(SSE4_1)
+static const uint *QT_FASTCALL convertARGB32FromARGB32PM_sse4(uint *buffer, const uint *src, int count,
+                                                              const QPixelLayout *, const QRgb *)
+{
+    for (int i = 0; i < count; ++i)
+        buffer[i] = qUnpremultiply_sse4(src[i]);
+    return buffer;
+}
+#endif
+
+
 static const uint *QT_FASTCALL convertRGBA8888PMFromARGB32PM(uint *buffer, const uint *src, int count,
                                                              const QPixelLayout *, const QRgb *)
 {
@@ -588,6 +601,17 @@ static const uint *QT_FASTCALL convertRGBA8888FromARGB32PM(uint *buffer, const u
     return buffer;
 }
 
+#if QT_COMPILER_SUPPORTS_HERE(SSE4_1)
+QT_FUNCTION_TARGET(SSE4_1)
+static const uint *QT_FASTCALL convertRGBA8888FromARGB32PM_sse4(uint *buffer, const uint *src, int count,
+                                                              const QPixelLayout *, const QRgb *)
+{
+    for (int i = 0; i < count; ++i)
+        buffer[i] = ARGB2RGBA(qUnpremultiply_sse4(src[i]));
+    return buffer;
+}
+#endif
+
 static const uint *QT_FASTCALL convertRGBXFromRGB32(uint *buffer, const uint *src, int count,
                                                     const QPixelLayout *, const QRgb *)
 {
@@ -604,6 +628,17 @@ static const uint *QT_FASTCALL convertRGBXFromARGB32PM(uint *buffer, const uint
     return buffer;
 }
 
+#if QT_COMPILER_SUPPORTS_HERE(SSE4_1)
+QT_FUNCTION_TARGET(SSE4_1)
+static const uint *QT_FASTCALL convertRGBXFromARGB32PM_sse4(uint *buffer, const uint *src, int count,
+                                                            const QPixelLayout *, const QRgb *)
+{
+    for (int i = 0; i < count; ++i)
+        buffer[i] = ARGB2RGBA(0xff000000 | qUnpremultiply_sse4(src[i]));
+    return buffer;
+}
+#endif
+
 template<QtPixelOrder PixelOrder>
 static const uint *QT_FASTCALL convertA2RGB30PMToARGB32PM(uint *buffer, const uint *src, int count,
                                                           const QPixelLayout *, const QRgb *)
@@ -6879,10 +6914,15 @@ void qInitDrawhelperAsm()
     }
 #endif // SSSE3
 
-#if QT_COMPILER_SUPPORTS_HERE(SSE4_1) && !defined(__SSE4_1__)
+#if QT_COMPILER_SUPPORTS_HERE(SSE4_1)
     if (qCpuHasFeature(SSE4_1)) {
+#if !defined(__SSE4_1__)
         qPixelLayouts[QImage::Format_ARGB32].convertToARGB32PM = convertARGB32ToARGB32PM_sse4;
         qPixelLayouts[QImage::Format_RGBA8888].convertToARGB32PM = convertRGBA8888ToARGB32PM_sse4;
+#endif
+        qPixelLayouts[QImage::Format_ARGB32].convertFromARGB32PM = convertARGB32FromARGB32PM_sse4;
+        qPixelLayouts[QImage::Format_RGBA8888].convertFromARGB32PM = convertRGBA8888FromARGB32PM_sse4;
+        qPixelLayouts[QImage::Format_RGBX8888].convertFromARGB32PM = convertRGBXFromARGB32PM_sse4;
     }
 #endif
 
diff --git a/src/gui/painting/qdrawingprimitive_sse2_p.h b/src/gui/painting/qdrawingprimitive_sse2_p.h
index 2b4cdc286e..deac31b797 100644
--- a/src/gui/painting/qdrawingprimitive_sse2_p.h
+++ b/src/gui/painting/qdrawingprimitive_sse2_p.h
@@ -236,4 +236,26 @@ QT_END_NAMESPACE
 
 #endif // __SSE2__
 
+QT_BEGIN_NAMESPACE
+#if QT_COMPILER_SUPPORTS_HERE(SSE4_1)
+QT_FUNCTION_TARGET(SSE4_1)
+inline QRgb qUnpremultiply_sse4(QRgb p)
+{
+    const uint alpha = qAlpha(p);
+    const uint invAlpha = qt_inv_premul_factor[alpha];
+    const __m128i via = _mm_set1_epi32(invAlpha);
+    const __m128i vr = _mm_set1_epi32(0x8000);
+    __m128i vl = _mm_unpacklo_epi8(_mm_cvtsi32_si128(p), _mm_setzero_si128());
+    vl = _mm_unpacklo_epi16(vl, _mm_setzero_si128());
+    vl = _mm_mullo_epi32(vl, via);
+    vl = _mm_add_epi32(vl, vr);
+    vl = _mm_srai_epi32(vl, 16);
+    vl = _mm_insert_epi32(vl, alpha, 3);
+    vl = _mm_packus_epi32(vl, _mm_setzero_si128());
+    vl = _mm_packus_epi16(vl, _mm_setzero_si128());
+    return _mm_cvtsi128_si32(vl);
+}
+#endif
+QT_END_NAMESPACE
+
 #endif // QDRAWINGPRIMITIVE_SSE2_P_H
diff --git a/tests/auto/gui/painting/qcolor/qcolor.pro b/tests/auto/gui/painting/qcolor/qcolor.pro
index f7439c243c..44d65bb50b 100644
--- a/tests/auto/gui/painting/qcolor/qcolor.pro
+++ b/tests/auto/gui/painting/qcolor/qcolor.pro
@@ -2,4 +2,4 @@ CONFIG += testcase
 CONFIG += parallel_test
 TARGET = tst_qcolor
 SOURCES  += tst_qcolor.cpp
-QT += testlib
+QT += testlib gui-private core-private
diff --git a/tests/auto/gui/painting/qcolor/tst_qcolor.cpp b/tests/auto/gui/painting/qcolor/tst_qcolor.cpp
index 3b3334ba1f..95f1da1354 100644
--- a/tests/auto/gui/painting/qcolor/tst_qcolor.cpp
+++ b/tests/auto/gui/painting/qcolor/tst_qcolor.cpp
@@ -38,6 +38,7 @@
 
 #include <qcolor.h>
 #include <qdebug.h>
+#include <private/qdrawingprimitive_sse2_p.h>
 
 class tst_QColor : public QObject
 {
@@ -102,6 +103,9 @@ private slots:
 
     void achromaticHslHue();
 
+    void premultiply();
+    void unpremultiply_sse4();
+
 #ifdef Q_DEAD_CODE_FROM_QT4_X11
     void setallowX11ColorNames();
 #endif
@@ -1432,5 +1436,33 @@ void tst_QColor::setallowX11ColorNames()
 }
 #endif
 
+void tst_QColor::premultiply()
+{
+    // Tests that qPremultiply(qUnpremultiply(x)) returns x.
+    for (uint a = 0; a < 256; a++) {
+        for (uint c = 0; c <= a; c++) {
+            QRgb p = qRgba(c, a-c, c, a);
+            QCOMPARE(p, qPremultiply(qUnpremultiply(p)));
+        }
+    }
+}
+
+void tst_QColor::unpremultiply_sse4()
+{
+    // Tests that qUnpremultiply_sse4 returns the same as qUnpremultiply.
+#if QT_COMPILER_SUPPORTS_HERE(SSE4_1)
+    if (qCpuHasFeature(SSE4_1)) {
+        for (uint a = 0; a < 256; a++) {
+            for (uint c = 0; c <= a; c++) {
+                QRgb p = qRgba(c, a-c, c, a);
+                QCOMPARE(qUnpremultiply(p), qUnpremultiply_sse4(p));
+            }
+        }
+        return;
+    }
+#endif
+    QSKIP("SSE4 not supported on this CPU.");
+}
+
 QTEST_MAIN(tst_QColor)
 #include "tst_qcolor.moc"