Add AVX2 versions of the fast blending functions

This patch adds AVX2 versions of the fast blending functions that we
already have SSE2 versions of.

Change-Id: Ifd1a22f7891b6208cb74929ad26095d12c5a1efb
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>

3 files changed, 340 insertions, 10 deletions

diff --git a/src/corelib/tools/qsimd_p.h b/src/corelib/tools/qsimd_p.h
index 9d8695c0f6..2fd4be00a5 100644
--- a/src/corelib/tools/qsimd_p.h
+++ b/src/corelib/tools/qsimd_p.h
@@ -468,6 +468,9 @@ static inline quint64 qCpuFeatures()
 #define ALIGNMENT_PROLOGUE_16BYTES(ptr, i, length) \
     for (; i < static_cast<int>(qMin(static_cast<quintptr>(length), ((4 - ((reinterpret_cast<quintptr>(ptr) >> 2) & 0x3)) & 0x3))); ++i)
 
+#define ALIGNMENT_PROLOGUE_32BYTES(ptr, i, length) \
+    for (; i < static_cast<int>(qMin(static_cast<quintptr>(length), ((8 - ((reinterpret_cast<quintptr>(ptr) >> 2) & 0x7)) & 0x7))); ++i)
+
 QT_END_NAMESPACE
 
 #endif // QSIMD_P_H
diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp
index c697aceaf3..ea0c5bbb88 100644
--- a/src/gui/painting/qdrawhelper.cpp
+++ b/src/gui/painting/qdrawhelper.cpp
@@ -6548,6 +6548,15 @@ static void qInitDrawhelperFunctions()
 
     qt_fetch_radial_gradient = qt_fetch_radial_gradient_sse2;
 
+    extern void QT_FASTCALL comp_func_SourceOver_sse2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha);
+    extern void QT_FASTCALL comp_func_solid_SourceOver_sse2(uint *destPixels, int length, uint color, uint const_alpha);
+    extern void QT_FASTCALL comp_func_Source_sse2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha);
+    extern void QT_FASTCALL comp_func_Plus_sse2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha);
+    qt_functionForMode_C[QPainter::CompositionMode_SourceOver] = comp_func_SourceOver_sse2;
+    qt_functionForModeSolid_C[QPainter::CompositionMode_SourceOver] = comp_func_solid_SourceOver_sse2;
+    qt_functionForMode_C[QPainter::CompositionMode_Source] = comp_func_Source_sse2;
+    qt_functionForMode_C[QPainter::CompositionMode_Plus] = comp_func_Plus_sse2;
+
 #ifdef QT_COMPILER_SUPPORTS_SSSE3
     if (qCpuHasFeature(SSSE3)) {
         extern void qt_blend_argb32_on_argb32_ssse3(uchar *destPixels, int dbpl,
@@ -6592,24 +6601,39 @@ static void qInitDrawhelperFunctions()
     }
 #endif
 
-#if defined(QT_COMPILER_SUPPORTS_AVX2) && !defined(__AVX2__)
+#if defined(QT_COMPILER_SUPPORTS_AVX2)
     if (qCpuHasFeature(AVX2)) {
+#if !defined(__AVX2__)
         extern const uint *QT_FASTCALL convertARGB32ToARGB32PM_avx2(uint *buffer, const uint *src, int count,
                                                                     const QVector<QRgb> *, QDitherInfo *);
         extern const uint *QT_FASTCALL convertRGBA8888ToARGB32PM_avx2(uint *buffer, const uint *src, int count,
                                                                       const QVector<QRgb> *, QDitherInfo *);
         qPixelLayouts[QImage::Format_ARGB32].convertToARGB32PM = convertARGB32ToARGB32PM_avx2;
         qPixelLayouts[QImage::Format_RGBA8888].convertToARGB32PM = convertRGBA8888ToARGB32PM_avx2;
+#endif
+        extern void qt_blend_rgb32_on_rgb32_avx2(uchar *destPixels, int dbpl,
+                                                 const uchar *srcPixels, int sbpl,
+                                                 int w, int h, int const_alpha);
+        extern void qt_blend_argb32_on_argb32_avx2(uchar *destPixels, int dbpl,
+                                                   const uchar *srcPixels, int sbpl,
+                                                   int w, int h, int const_alpha);
+        qBlendFunctions[QImage::Format_RGB32][QImage::Format_RGB32] = qt_blend_rgb32_on_rgb32_avx2;
+        qBlendFunctions[QImage::Format_ARGB32_Premultiplied][QImage::Format_RGB32] = qt_blend_rgb32_on_rgb32_avx2;
+        qBlendFunctions[QImage::Format_RGB32][QImage::Format_ARGB32_Premultiplied] = qt_blend_argb32_on_argb32_avx2;
+        qBlendFunctions[QImage::Format_ARGB32_Premultiplied][QImage::Format_ARGB32_Premultiplied] = qt_blend_argb32_on_argb32_avx2;
+        qBlendFunctions[QImage::Format_RGBX8888][QImage::Format_RGBX8888] = qt_blend_rgb32_on_rgb32_avx2;
+        qBlendFunctions[QImage::Format_RGBA8888_Premultiplied][QImage::Format_RGBX8888] = qt_blend_rgb32_on_rgb32_avx2;
+        qBlendFunctions[QImage::Format_RGBX8888][QImage::Format_RGBA8888_Premultiplied] = qt_blend_argb32_on_argb32_avx2;
+        qBlendFunctions[QImage::Format_RGBA8888_Premultiplied][QImage::Format_RGBA8888_Premultiplied] = qt_blend_argb32_on_argb32_avx2;
+
+        extern void QT_FASTCALL comp_func_SourceOver_avx2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha);
+        extern void QT_FASTCALL comp_func_solid_SourceOver_avx2(uint *destPixels, int length, uint color, uint const_alpha);
+        extern void QT_FASTCALL comp_func_Source_avx2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha);
+        qt_functionForMode_C[QPainter::CompositionMode_SourceOver] = comp_func_SourceOver_avx2;
+        qt_functionForModeSolid_C[QPainter::CompositionMode_SourceOver] = comp_func_solid_SourceOver_avx2;
+        qt_functionForMode_C[QPainter::CompositionMode_Source] = comp_func_Source_avx2;
     }
 #endif
-    extern void QT_FASTCALL comp_func_SourceOver_sse2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha);
-    extern void QT_FASTCALL comp_func_solid_SourceOver_sse2(uint *destPixels, int length, uint color, uint const_alpha);
-    extern void QT_FASTCALL comp_func_Source_sse2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha);
-    extern void QT_FASTCALL comp_func_Plus_sse2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha);
-    qt_functionForMode_C[QPainter::CompositionMode_SourceOver] = comp_func_SourceOver_sse2;
-    qt_functionForModeSolid_C[QPainter::CompositionMode_SourceOver] = comp_func_solid_SourceOver_sse2;
-    qt_functionForMode_C[QPainter::CompositionMode_Source] = comp_func_Source_sse2;
-    qt_functionForMode_C[QPainter::CompositionMode_Plus] = comp_func_Plus_sse2;
 
 #endif // SSE2
 
diff --git a/src/gui/painting/qdrawhelper_avx2.cpp b/src/gui/painting/qdrawhelper_avx2.cpp
index 35a975c972..01ffd54918 100644
--- a/src/gui/painting/qdrawhelper_avx2.cpp
+++ b/src/gui/painting/qdrawhelper_avx2.cpp
@@ -37,12 +37,14 @@
 **
 ****************************************************************************/
 
-#include <private/qdrawhelper_p.h>
+#include "qdrawhelper_p.h"
+#include "qdrawingprimitive_sse2_p.h"
 
 #if defined(QT_COMPILER_SUPPORTS_AVX2)
 
 QT_BEGIN_NAMESPACE
 
+// Autovectorized premultiply functions:
 const uint *QT_FASTCALL convertARGB32ToARGB32PM_avx2(uint *buffer, const uint *src, int count,
                                                      const QVector<QRgb> *, QDitherInfo *)
 {
@@ -55,6 +57,307 @@ const uint *QT_FASTCALL convertRGBA8888ToARGB32PM_avx2(uint *buffer, const uint
     return qt_convertRGBA8888ToARGB32PM(buffer, src, count);
 }
 
+// Vectorized blend functions:
+
+// See BYTE_MUL_SSE2 for details.
+inline static void BYTE_MUL_AVX2(__m256i &pixelVector, const __m256i &alphaChannel, const __m256i &colorMask, const __m256i &half)
+{
+    __m256i pixelVectorAG = _mm256_srli_epi16(pixelVector, 8);
+    __m256i pixelVectorRB = _mm256_and_si256(pixelVector, colorMask);
+
+    pixelVectorAG = _mm256_mullo_epi16(pixelVectorAG, alphaChannel);
+    pixelVectorRB = _mm256_mullo_epi16(pixelVectorRB, alphaChannel);
+
+    pixelVectorRB = _mm256_add_epi16(pixelVectorRB, _mm256_srli_epi16(pixelVectorRB, 8));
+    pixelVectorAG = _mm256_add_epi16(pixelVectorAG, _mm256_srli_epi16(pixelVectorAG, 8));
+    pixelVectorRB = _mm256_add_epi16(pixelVectorRB, half);
+    pixelVectorAG = _mm256_add_epi16(pixelVectorAG, half);
+
+    pixelVectorRB = _mm256_srli_epi16(pixelVectorRB, 8);
+    pixelVectorAG = _mm256_andnot_si256(colorMask, pixelVectorAG);
+
+    pixelVector = _mm256_or_si256(pixelVectorAG, pixelVectorRB);
+}
+
+// See INTERPOLATE_PIXEL_255_SSE2 for details.
+inline static void INTERPOLATE_PIXEL_255_AVX2(const __m256i &srcVector, __m256i &dstVector, const __m256i &alphaChannel, const __m256i &oneMinusAlphaChannel, const __m256i &colorMask, const __m256i &half)
+{
+    const __m256i srcVectorAG = _mm256_srli_epi16(srcVector, 8);
+    const __m256i dstVectorAG = _mm256_srli_epi16(dstVector, 8);
+    const __m256i srcVectorRB = _mm256_and_si256(srcVector, colorMask);
+    const __m256i dstVectorRB = _mm256_and_si256(dstVector, colorMask);
+    const __m256i srcVectorAGalpha = _mm256_mullo_epi16(srcVectorAG, alphaChannel);
+    const __m256i srcVectorRBalpha = _mm256_mullo_epi16(srcVectorRB, alphaChannel);
+    const __m256i dstVectorAGoneMinusAlpha = _mm256_mullo_epi16(dstVectorAG, oneMinusAlphaChannel);
+    const __m256i dstVectorRBoneMinusAlpha = _mm256_mullo_epi16(dstVectorRB, oneMinusAlphaChannel);
+    __m256i finalAG = _mm256_add_epi16(srcVectorAGalpha, dstVectorAGoneMinusAlpha);
+    __m256i finalRB = _mm256_add_epi16(srcVectorRBalpha, dstVectorRBoneMinusAlpha);
+    finalAG = _mm256_add_epi16(finalAG, _mm256_srli_epi16(finalAG, 8));
+    finalRB = _mm256_add_epi16(finalRB, _mm256_srli_epi16(finalRB, 8));
+    finalAG = _mm256_add_epi16(finalAG, half);
+    finalRB = _mm256_add_epi16(finalRB, half);
+    finalAG = _mm256_andnot_si256(colorMask, finalAG);
+    finalRB = _mm256_srli_epi16(finalRB, 8);
+
+    dstVector = _mm256_or_si256(finalAG, finalRB);
+}
+
+// See BLEND_SOURCE_OVER_ARGB32_SSE2 for details.
+inline static void BLEND_SOURCE_OVER_ARGB32_AVX2(quint32 *dst, const quint32 *src, const int length)
+{
+    const __m256i half = _mm256_set1_epi16(0x80);
+    const __m256i one = _mm256_set1_epi16(0xff);
+    const __m256i colorMask = _mm256_set1_epi32(0x00ff00ff);
+    const __m256i alphaMask = _mm256_set1_epi32(0xff000000);
+    const __m256i offsetMask = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
+    const __m256i alphaShuffleMask = _mm256_set_epi8(char(0xff),15,char(0xff),15,char(0xff),11,char(0xff),11,char(0xff),7,char(0xff),7,char(0xff),3,char(0xff),3,
+                                                     char(0xff),15,char(0xff),15,char(0xff),11,char(0xff),11,char(0xff),7,char(0xff),7,char(0xff),3,char(0xff),3);
+
+    const int minusOffsetToAlignDstOn32Bytes = (reinterpret_cast<quintptr>(dst) >> 2) & 0x7;
+
+    int x = 0;
+    // Prologue to handle all pixels until dst is 32-byte aligned in one step.
+    if (minusOffsetToAlignDstOn32Bytes != 0 && x < (length - 7)) {
+        const __m256i prologueMask = _mm256_sub_epi32(_mm256_set1_epi32(minusOffsetToAlignDstOn32Bytes - 1), offsetMask);
+        const __m256i srcVector = _mm256_maskload_epi32((const int *)&src[x - minusOffsetToAlignDstOn32Bytes], prologueMask);
+        const __m256i prologueAlphaMask = _mm256_blendv_epi8(_mm256_setzero_si256(), alphaMask, prologueMask);
+        if (!_mm256_testz_si256(srcVector, prologueAlphaMask)) {
+            if (_mm256_testc_si256(srcVector, prologueAlphaMask)) {
+                _mm256_maskstore_epi32((int *)&dst[x - minusOffsetToAlignDstOn32Bytes], prologueMask, srcVector);
+            } else {
+                __m256i alphaChannel = _mm256_shuffle_epi8(srcVector, alphaShuffleMask);
+                alphaChannel = _mm256_sub_epi16(one, alphaChannel);
+                __m256i dstVector = _mm256_maskload_epi32((int *)&dst[x - minusOffsetToAlignDstOn32Bytes], prologueMask);
+                BYTE_MUL_AVX2(dstVector, alphaChannel, colorMask, half);
+                dstVector = _mm256_add_epi8(dstVector, srcVector);
+                _mm256_maskstore_epi32((int *)&dst[x - minusOffsetToAlignDstOn32Bytes], prologueMask, dstVector);
+            }
+        }
+        x += (8 - minusOffsetToAlignDstOn32Bytes);
+    }
+
+    for (; x < (length - 7); x += 8) {
+        const __m256i srcVector = _mm256_lddqu_si256((const __m256i *)&src[x]);
+        if (!_mm256_testz_si256(srcVector, alphaMask)) {
+            if (_mm256_testc_si256(srcVector, alphaMask)) {
+                _mm256_store_si256((__m256i *)&dst[x], srcVector);
+            } else {
+                __m256i alphaChannel = _mm256_shuffle_epi8(srcVector, alphaShuffleMask);
+                alphaChannel = _mm256_sub_epi16(one, alphaChannel);
+                __m256i dstVector = _mm256_load_si256((__m256i *)&dst[x]);
+                 BYTE_MUL_AVX2(dstVector, alphaChannel, colorMask, half);
+                dstVector = _mm256_add_epi8(dstVector, srcVector);
+                _mm256_store_si256((__m256i *)&dst[x], dstVector);
+            }
+        }
+    }
+
+    // Epilogue to handle all remaining pixels in one step.
+    if (x < length) {
+        const __m256i epilogueMask = _mm256_add_epi32(offsetMask, _mm256_set1_epi32(x - length));
+        const __m256i srcVector = _mm256_maskload_epi32((const int *)&src[x], epilogueMask);
+        const __m256i epilogueAlphaMask = _mm256_blendv_epi8(_mm256_setzero_si256(), alphaMask, epilogueMask);
+        if (!_mm256_testz_si256(srcVector, epilogueAlphaMask)) {
+            if (_mm256_testc_si256(srcVector, epilogueAlphaMask)) {
+                _mm256_maskstore_epi32((int *)&dst[x], epilogueMask, srcVector);
+            } else {
+                __m256i alphaChannel = _mm256_shuffle_epi8(srcVector, alphaShuffleMask);
+                alphaChannel = _mm256_sub_epi16(one, alphaChannel);
+                __m256i dstVector = _mm256_maskload_epi32((int *)&dst[x], epilogueMask);
+                BYTE_MUL_AVX2(dstVector, alphaChannel, colorMask, half);
+                dstVector = _mm256_add_epi8(dstVector, srcVector);
+                _mm256_maskstore_epi32((int *)&dst[x], epilogueMask, dstVector);
+            }
+        }
+    }
+}
+
+
+// See BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_SSE2 for details.
+inline static void BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_AVX2(quint32 *dst, const quint32 *src, const int length, const int const_alpha)
+{
+    int x = 0;
+
+    ALIGNMENT_PROLOGUE_32BYTES(dst, x, length)
+        blend_pixel(dst[x], src[x], const_alpha);
+
+    const __m256i half = _mm256_set1_epi16(0x80);
+    const __m256i one = _mm256_set1_epi16(0xff);
+    const __m256i colorMask = _mm256_set1_epi32(0x00ff00ff);
+    const __m256i alphaMask = _mm256_set1_epi32(0xff000000);
+    const __m256i alphaShuffleMask = _mm256_set_epi8(char(0xff),15,char(0xff),15,char(0xff),11,char(0xff),11,char(0xff),7,char(0xff),7,char(0xff),3,char(0xff),3,
+                                                     char(0xff),15,char(0xff),15,char(0xff),11,char(0xff),11,char(0xff),7,char(0xff),7,char(0xff),3,char(0xff),3);
+    const __m256i constAlphaVector = _mm256_set1_epi16(const_alpha);
+    for (; x < (length - 7); x += 8) {
+        __m256i srcVector = _mm256_lddqu_si256((const __m256i *)&src[x]);
+        if (!_mm256_testz_si256(srcVector, alphaMask)) {
+            BYTE_MUL_AVX2(srcVector, constAlphaVector, colorMask, half);
+
+            __m256i alphaChannel = _mm256_shuffle_epi8(srcVector, alphaShuffleMask);
+            alphaChannel = _mm256_sub_epi16(one, alphaChannel);
+            __m256i dstVector = _mm256_load_si256((__m256i *)&dst[x]);
+            BYTE_MUL_AVX2(dstVector, alphaChannel, colorMask, half);
+            dstVector = _mm256_add_epi8(dstVector, srcVector);
+            _mm256_store_si256((__m256i *)&dst[x], dstVector);
+        }
+    }
+    for (; x < length; ++x)
+        blend_pixel(dst[x], src[x], const_alpha);
+}
+
+void qt_blend_argb32_on_argb32_avx2(uchar *destPixels, int dbpl,
+                                    const uchar *srcPixels, int sbpl,
+                                    int w, int h,
+                                    int const_alpha)
+{
+    if (const_alpha == 256) {
+        for (int y = 0; y < h; ++y) {
+            const quint32 *src = reinterpret_cast<const quint32 *>(srcPixels);
+            quint32 *dst = reinterpret_cast<quint32 *>(destPixels);
+            BLEND_SOURCE_OVER_ARGB32_AVX2(dst, src, w);
+            destPixels += dbpl;
+            srcPixels += sbpl;
+        }
+    } else if (const_alpha != 0) {
+        const_alpha = (const_alpha * 255) >> 8;
+        for (int y = 0; y < h; ++y) {
+            const quint32 *src = reinterpret_cast<const quint32 *>(srcPixels);
+            quint32 *dst = reinterpret_cast<quint32 *>(destPixels);
+            BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_AVX2(dst, src, w, const_alpha);
+            destPixels += dbpl;
+            srcPixels += sbpl;
+        }
+    }
+}
+
+void qt_blend_rgb32_on_rgb32_avx2(uchar *destPixels, int dbpl,
+                                   const uchar *srcPixels, int sbpl,
+                                   int w, int h,
+                                   int const_alpha)
+{
+    if (const_alpha == 256) {
+        for (int y = 0; y < h; ++y) {
+            const quint32 *src = reinterpret_cast<const quint32 *>(srcPixels);
+            quint32 *dst = reinterpret_cast<quint32 *>(destPixels);
+            ::memcpy(dst, src, w * sizeof(uint));
+            srcPixels += sbpl;
+            destPixels += dbpl;
+        }
+        return;
+    }
+    if (const_alpha == 0)
+        return;
+
+    const __m256i half = _mm256_set1_epi16(0x80);
+    const __m256i colorMask = _mm256_set1_epi32(0x00ff00ff);
+
+    const_alpha = (const_alpha * 255) >> 8;
+    int one_minus_const_alpha = 255 - const_alpha;
+    const __m256i constAlphaVector = _mm256_set1_epi16(const_alpha);
+    const __m256i oneMinusConstAlpha =  _mm256_set1_epi16(one_minus_const_alpha);
+    for (int y = 0; y < h; ++y) {
+        const quint32 *src = reinterpret_cast<const quint32 *>(srcPixels);
+        quint32 *dst = reinterpret_cast<quint32 *>(destPixels);
+        int x = 0;
+
+        // First, align dest to 32 bytes:
+        ALIGNMENT_PROLOGUE_32BYTES(dst, x, w)
+            dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], one_minus_const_alpha);
+
+        // 2) interpolate pixels with AVX2
+        for (; x < (w - 7); x += 8) {
+            const __m256i srcVector = _mm256_lddqu_si256((const __m256i *)&src[x]);
+            if (!_mm256_testc_si256(srcVector, _mm256_setzero_si256())) {
+                __m256i dstVector = _mm256_load_si256((__m256i *)&dst[x]);
+                INTERPOLATE_PIXEL_255_AVX2(srcVector, dstVector, constAlphaVector, oneMinusConstAlpha, colorMask, half);
+                _mm256_store_si256((__m256i *)&dst[x], dstVector);
+            }
+        }
+
+        // 3) Epilogue
+        for (; x < w; ++x)
+            dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], one_minus_const_alpha);
+
+        srcPixels += sbpl;
+        destPixels += dbpl;
+    }
+}
+
+void QT_FASTCALL comp_func_SourceOver_avx2(uint *destPixels, const uint *srcPixels, int length, uint const_alpha)
+{
+    Q_ASSERT(const_alpha < 256);
+
+    const quint32 *src = (const quint32 *) srcPixels;
+    quint32 *dst = (quint32 *) destPixels;
+
+    if (const_alpha == 255)
+        BLEND_SOURCE_OVER_ARGB32_AVX2(dst, src, length);
+    else
+        BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_AVX2(dst, src, length, const_alpha);
+}
+
+void QT_FASTCALL comp_func_Source_avx2(uint *dst, const uint *src, int length, uint const_alpha)
+{
+    if (const_alpha == 255) {
+        ::memcpy(dst, src, length * sizeof(uint));
+    } else {
+        const int ialpha = 255 - const_alpha;
+
+        int x = 0;
+
+        // 1) prologue, align on 32 bytes
+        ALIGNMENT_PROLOGUE_32BYTES(dst, x, length)
+            dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
+
+        // 2) interpolate pixels with AVX2
+        const __m256i half = _mm256_set1_epi16(0x80);
+        const __m256i colorMask = _mm256_set1_epi32(0x00ff00ff);
+        const __m256i constAlphaVector = _mm256_set1_epi16(const_alpha);
+        const __m256i oneMinusConstAlpha =  _mm256_set1_epi16(ialpha);
+        for (; x < length - 7; x += 8) {
+            const __m256i srcVector = _mm256_lddqu_si256((const __m256i *)&src[x]);
+            __m256i dstVector = _mm256_load_si256((__m256i *)&dst[x]);
+            INTERPOLATE_PIXEL_255_AVX2(srcVector, dstVector, constAlphaVector, oneMinusConstAlpha, colorMask, half);
+            _mm256_store_si256((__m256i *)&dst[x], dstVector);
+        }
+
+        // 3) Epilogue
+        for (; x < length; ++x)
+            dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
+    }
+}
+
+void QT_FASTCALL comp_func_solid_SourceOver_avx2(uint *destPixels, int length, uint color, uint const_alpha)
+{
+    if ((const_alpha & qAlpha(color)) == 255) {
+        qt_memfill32(destPixels, color, length);
+    } else {
+        if (const_alpha != 255)
+            color = BYTE_MUL(color, const_alpha);
+
+        const quint32 minusAlphaOfColor = qAlpha(~color);
+        int x = 0;
+
+        quint32 *dst = (quint32 *) destPixels;
+        const __m256i colorVector = _mm256_set1_epi32(color);
+        const __m256i colorMask = _mm256_set1_epi32(0x00ff00ff);
+        const __m256i half = _mm256_set1_epi16(0x80);
+        const __m256i minusAlphaOfColorVector = _mm256_set1_epi16(minusAlphaOfColor);
+
+        ALIGNMENT_PROLOGUE_32BYTES(dst, x, length)
+            destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
+
+        for (; x < length - 7; x += 8) {
+            __m256i dstVector = _mm256_load_si256((__m256i *)&dst[x]);
+            BYTE_MUL_AVX2(dstVector, minusAlphaOfColorVector, colorMask, half);
+            dstVector = _mm256_add_epi8(colorVector, dstVector);
+            _mm256_store_si256((__m256i *)&dst[x], dstVector);
+        }
+        for (; x < length; ++x)
+            destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
+    }
+}
+
 QT_END_NAMESPACE
 
 #endif