Use SSSE3 to convert from RGB888 to RGB32

Converting from RGB encoded on 24bits to RGB encoded on 32 bits is
quite inefficient.
This type of conversion is common for some image format.

The patch implement the conversion with SSSE3. This reduce by 3 the
number of instructions, pushing the bottleneck to memory bandwidth.

On Atom N450, the new benchmark is 40% faster for scanlines of
2000 pixels, 30% faster for scanlines of 32 pixels, and 15% slower
for small images (3 and 8px).

Reviewed-by: Olivier Goffart

3 files changed, 160 insertions, 1 deletions

diff --git a/src/gui/image/image.pri b/src/gui/image/image.pri
index 3a02d56..b20a04f 100644
--- a/src/gui/image/image.pri
+++ b/src/gui/image/image.pri
@@ -95,3 +95,4 @@ contains(QT_CONFIG, gif):include($$PWD/qgifhandler.pri)
 
 # SIMD
 SSE2_SOURCES += image/qimage_sse2.cpp
+SSSE3_SOURCES += image/qimage_ssse3.cpp
diff --git a/src/gui/image/qimage.cpp b/src/gui/image/qimage.cpp
index 713e765..30cf758 100644
--- a/src/gui/image/qimage.cpp
+++ b/src/gui/image/qimage.cpp
@@ -3354,7 +3354,7 @@ CONVERT_DECL(qargb4444, quint32)
 
 
 // first index source, second dest
-static const Image_Converter converter_map[QImage::NImageFormats][QImage::NImageFormats] =
+static Image_Converter converter_map[QImage::NImageFormats][QImage::NImageFormats] =
 {
     {
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
@@ -3761,6 +3761,14 @@ void qInitImageConversions()
         inplace_converter_map[QImage::Format_ARGB32][QImage::Format_ARGB32_Premultiplied] = convert_ARGB_to_ARGB_PM_inplace_sse2;
     }
 #endif
+#ifdef QT_HAVE_SSSE3
+    if (features & SSSE3) {
+        extern void convert_RGB888_to_RGB32_ssse3(QImageData *dest, const QImageData *src, Qt::ImageConversionFlags);
+        converter_map[QImage::Format_RGB888][QImage::Format_RGB32] = convert_RGB888_to_RGB32_ssse3;
+        converter_map[QImage::Format_RGB888][QImage::Format_ARGB32] = convert_RGB888_to_RGB32_ssse3;
+        converter_map[QImage::Format_RGB888][QImage::Format_ARGB32_Premultiplied] = convert_RGB888_to_RGB32_ssse3;
+    }
+#endif
 }
 
 /*!
diff --git a/src/gui/image/qimage_ssse3.cpp b/src/gui/image/qimage_ssse3.cpp
new file mode 100644
index 0000000..1c664f2
--- /dev/null
+++ b/src/gui/image/qimage_ssse3.cpp
@@ -0,0 +1,150 @@
+/****************************************************************************
+**
+** Copyright (C) 2010 Nokia Corporation and/or its subsidiary(-ies).
+** All rights reserved.
+** Contact: Nokia Corporation (qt-info@nokia.com)
+**
+** This file is part of the QtGui module of the Qt Toolkit.
+**
+** $QT_BEGIN_LICENSE:LGPL$
+** No Commercial Usage
+** This file contains pre-release code and may not be distributed.
+** You may use this file in accordance with the terms and conditions
+** contained in the Technology Preview License Agreement accompanying
+** this package.
+**
+** GNU Lesser General Public License Usage
+** Alternatively, this file may be used under the terms of the GNU Lesser
+** General Public License version 2.1 as published by the Free Software
+** Foundation and appearing in the file LICENSE.LGPL included in the
+** packaging of this file.  Please review the following information to
+** ensure the GNU Lesser General Public License version 2.1 requirements
+** will be met: http://www.gnu.org/licenses/old-licenses/lgpl-2.1.html.
+**
+** In addition, as a special exception, Nokia gives you certain additional
+** rights.  These rights are described in the Nokia Qt LGPL Exception
+** version 1.1, included in the file LGPL_EXCEPTION.txt in this package.
+**
+** If you have questions regarding the use of this file, please contact
+** Nokia at qt-info@nokia.com.
+**
+**
+**
+**
+**
+**
+**
+**
+** $QT_END_LICENSE$
+**
+****************************************************************************/
+
+#include <qimage.h>
+#include <private/qimage_p.h>
+#include <private/qsimd_p.h>
+
+#ifdef QT_HAVE_SSSE3
+
+#include <stdio.h>
+QT_BEGIN_NAMESPACE
+
+// Convert a scanline of RGB888 (src) to RGB32 (dst)
+// src must be at least len * 3 bytes
+// dst must be at least len * 4 bytes
+inline void convert_rgb888_to_rgb32_ssse3(quint32 *dst, const uchar *src, int len)
+{
+    quint32 *const end = dst + len;
+
+    // Prologue, align dst to 16 bytes. The alignement is done on dst because it has 4 store()
+    // for each 3 load() of src.
+    const int offsetToAlignOn16Bytes = (4 - ((reinterpret_cast<quintptr>(dst) >> 2) & 0x3)) & 0x3;
+    const int prologLength = qMin(len, offsetToAlignOn16Bytes);
+
+    for (int i = 0; i < prologLength; ++i) {
+        *dst++ = qRgb(src[0], src[1], src[2]);
+        src += 3;
+    }
+
+    // Mask the 4 first colors of the RGB888 vector
+    const __m128i shuffleMask = _mm_set_epi8(0xff, 9, 10, 11, 0xff, 6, 7, 8, 0xff, 3, 4, 5, 0xff, 0, 1, 2);
+
+    // Mask the 4 last colors of a RGB888 vector with an offset of 1 (so the last 3 bytes are RGB)
+    const __m128i shuffleMaskEnd = _mm_set_epi8(0xff, 13, 14, 15, 0xff, 10, 11, 12, 0xff, 7, 8, 9, 0xff, 4, 5, 6);
+
+    // Mask to have alpha = 0xff
+    const __m128i alphaMask = _mm_set1_epi32(0xff000000);
+
+    __m128i *inVectorPtr = (__m128i *)src;
+    __m128i *dstVectorPtr = (__m128i *)dst;
+
+    const int simdRoundCount = (len - prologLength) / 16; // one iteration in the loop converts 16 pixels
+    for (int i = 0; i < simdRoundCount; ++i) {
+        /*
+         RGB888 has 5 pixels per vector, + 1 byte from the next pixel. The idea here is
+         to load vectors of RGB888 and use palignr to select a vector out of two vectors.
+
+         After 3 loads of RGB888 and 3 stores of RGB32, we have 4 pixels left in the last
+         vector of RGB888, we can mask it directly to get a last store or RGB32. After that,
+         the first next byte is a R, and we can loop for the next 16 pixels.
+
+         The conversion itself is done with a byte permutation (pshufb).
+         */
+        __m128i firstSrcVector = _mm_lddqu_si128(inVectorPtr);
+        __m128i outputVector = _mm_shuffle_epi8(firstSrcVector, shuffleMask);
+        _mm_store_si128(dstVectorPtr, _mm_or_si128(outputVector, alphaMask));
+        ++inVectorPtr;
+        ++dstVectorPtr;
+
+        // There are 4 unused bytes left in srcVector, we need to load the next 16 bytes
+        // and load the next input with palignr
+        __m128i secondSrcVector = _mm_lddqu_si128(inVectorPtr);
+        __m128i srcVector = _mm_alignr_epi8(secondSrcVector, firstSrcVector, 12);
+        outputVector = _mm_shuffle_epi8(srcVector, shuffleMask);
+        _mm_store_si128(dstVectorPtr, _mm_or_si128(outputVector, alphaMask));
+        ++inVectorPtr;
+        ++dstVectorPtr;
+        firstSrcVector = secondSrcVector;
+
+        // We now have 8 unused bytes left in firstSrcVector
+        secondSrcVector = _mm_lddqu_si128(inVectorPtr);
+        srcVector = _mm_alignr_epi8(secondSrcVector, firstSrcVector, 8);
+        outputVector = _mm_shuffle_epi8(srcVector, shuffleMask);
+        _mm_store_si128(dstVectorPtr, _mm_or_si128(outputVector, alphaMask));
+        ++inVectorPtr;
+        ++dstVectorPtr;
+
+        // There are now 12 unused bytes in firstSrcVector.
+        // We can mask them directly, almost there.
+        outputVector = _mm_shuffle_epi8(secondSrcVector, shuffleMaskEnd);
+        _mm_store_si128(dstVectorPtr, _mm_or_si128(outputVector, alphaMask));
+        ++dstVectorPtr;
+    }
+    src = (uchar *)inVectorPtr;
+    dst = (quint32 *)dstVectorPtr;
+
+    while (dst != end) {
+        *dst++ = qRgb(src[0], src[1], src[2]);
+        src += 3;
+    }
+}
+
+void convert_RGB888_to_RGB32_ssse3(QImageData *dest, const QImageData *src, Qt::ImageConversionFlags)
+{
+    Q_ASSERT(src->format == QImage::Format_RGB888);
+    Q_ASSERT(dest->format == QImage::Format_RGB32 || dest->format == QImage::Format_ARGB32 || dest->format == QImage::Format_ARGB32_Premultiplied);
+    Q_ASSERT(src->width == dest->width);
+    Q_ASSERT(src->height == dest->height);
+
+    const uchar *src_data = (uchar *) src->data;
+    quint32 *dest_data = (quint32 *) dest->data;
+
+    for (int i = 0; i < src->height; ++i) {
+        convert_rgb888_to_rgb32_ssse3(dest_data, src_data, src->width);
+        src_data += src->bytes_per_line;
+        dest_data = (quint32 *)((uchar*)dest_data + dest->bytes_per_line);
+    }
+}
+
+QT_END_NAMESPACE
+
+#endif // QT_HAVE_SSSE3