Optimize RGBA64->RGBA64PM for SSE2/AVX2

And remove the direct conversion so we can get both the SIMD
optimization and threading applied.

Change-Id: Id032ea91cc40c1cbf1c8a1da0386de35aa36cfb5
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>

3 files changed, 45 insertions, 4 deletions

diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp
index 61963a5e61..f28e8a6c59 100644
--- a/src/gui/painting/qdrawhelper.cpp
+++ b/src/gui/painting/qdrawhelper.cpp
@@ -5225,16 +5225,16 @@ static void qInitDrawhelperFunctions()
         qPixelLayouts[QImage::Format_RGBA8888].fetchToARGB32PM = fetchRGBA8888ToARGB32PM_avx2;
         qPixelLayouts[QImage::Format_RGBA8888].convertToARGB32PM = convertRGBA8888ToARGB32PM_avx2;
 
-#if QT_CONFIG(raster_64bit)
-        extern const QRgba64 * QT_FASTCALL convertARGB32ToRGBA64PM_avx2(QRgba64 *, const uint *, int, const QList<QRgb> *, QDitherInfo *);
-        extern const QRgba64 * QT_FASTCALL convertRGBA8888ToRGBA64PM_avx2(QRgba64 *, const uint *, int count, const QList<QRgb> *, QDitherInfo *);
+        extern const QRgba64 *QT_FASTCALL convertARGB32ToRGBA64PM_avx2(QRgba64 *, const uint *, int, const QList<QRgb> *, QDitherInfo *);
+        extern const QRgba64 *QT_FASTCALL convertRGBA8888ToRGBA64PM_avx2(QRgba64 *, const uint *, int count, const QList<QRgb> *, QDitherInfo *);
         extern const QRgba64 *QT_FASTCALL fetchARGB32ToRGBA64PM_avx2(QRgba64 *, const uchar *, int, int, const QList<QRgb> *, QDitherInfo *);
         extern const QRgba64 *QT_FASTCALL fetchRGBA8888ToRGBA64PM_avx2(QRgba64 *, const uchar *, int, int, const QList<QRgb> *, QDitherInfo *);
+        extern const QRgba64 *QT_FASTCALL fetchRGBA64ToRGBA64PM_avx2(QRgba64 *buffer, const uchar *src, int index, int count, const QList<QRgb> *, QDitherInfo *);
         qPixelLayouts[QImage::Format_ARGB32].convertToRGBA64PM = convertARGB32ToRGBA64PM_avx2;
         qPixelLayouts[QImage::Format_RGBX8888].convertToRGBA64PM = convertRGBA8888ToRGBA64PM_avx2;
         qPixelLayouts[QImage::Format_ARGB32].fetchToRGBA64PM = fetchARGB32ToRGBA64PM_avx2;
         qPixelLayouts[QImage::Format_RGBX8888].fetchToRGBA64PM = fetchRGBA8888ToRGBA64PM_avx2;
-#endif
+        qPixelLayouts[QImage::Format_RGBA64].fetchToRGBA64PM = fetchRGBA64ToRGBA64PM_avx2;
     }
 #endif
 
diff --git a/src/gui/painting/qdrawhelper_avx2.cpp b/src/gui/painting/qdrawhelper_avx2.cpp
index 7d85355003..e401a1463f 100644
--- a/src/gui/painting/qdrawhelper_avx2.cpp
+++ b/src/gui/painting/qdrawhelper_avx2.cpp
@@ -1229,6 +1229,38 @@ const QRgba64 *QT_FASTCALL fetchRGBA8888ToRGBA64PM_avx2(QRgba64 *buffer, const u
     return buffer;
 }
 
+const QRgba64 *QT_FASTCALL fetchRGBA64ToRGBA64PM_avx2(QRgba64 *buffer, const uchar *src, int index, int count,
+                                                      const QList<QRgb> *, QDitherInfo *)
+{
+    const QRgba64 *s = reinterpret_cast<const QRgba64 *>(src) + index;
+    int i = 0;
+    const __m256i vh = _mm256_set1_epi32(0x8000);
+    for (; i < count - 3; i += 4) {
+        __m256i vs256 = _mm256_loadu_si256((const __m256i *)(s + i));
+        __m256i va256 = _mm256_shufflelo_epi16(vs256, _MM_SHUFFLE(3, 3, 3, 3));
+        va256 = _mm256_shufflehi_epi16(va256, _MM_SHUFFLE(3, 3, 3, 3));
+        const __m256i vmullo = _mm256_mullo_epi16(vs256, va256);
+        const __m256i vmulhi = _mm256_mulhi_epu16(vs256, va256);
+        __m256i vslo = _mm256_unpacklo_epi16(vmullo, vmulhi);
+        __m256i vshi = _mm256_unpackhi_epi16(vmullo, vmulhi);
+        vslo = _mm256_add_epi32(vslo, _mm256_srli_epi32(vslo, 16));
+        vshi = _mm256_add_epi32(vshi, _mm256_srli_epi32(vshi, 16));
+        vslo = _mm256_add_epi32(vslo, vh);
+        vshi = _mm256_add_epi32(vshi, vh);
+        vslo = _mm256_srli_epi32(vslo, 16);
+        vshi = _mm256_srli_epi32(vshi, 16);
+        vs256 = _mm256_packus_epi32(vslo, vshi);
+        _mm256_storeu_si256((__m256i *)(buffer + i), vs256);
+    }
+    for (; i < count; ++i) {
+        __m128i vs = _mm_loadl_epi64((const __m128i *)(s + i));
+        __m128i va = _mm_shufflelo_epi16(vs, _MM_SHUFFLE(3, 3, 3, 3));
+        vs = multiplyAlpha65535(vs, va);
+        _mm_storel_epi64((__m128i *)(buffer + i), vs);
+    }
+    return buffer;
+}
+
 QT_END_NAMESPACE
 
 #endif
diff --git a/src/gui/painting/qpixellayout.cpp b/src/gui/painting/qpixellayout.cpp
index acaee9f289..8dd62b0c00 100644
--- a/src/gui/painting/qpixellayout.cpp
+++ b/src/gui/painting/qpixellayout.cpp
@@ -1045,8 +1045,17 @@ static const QRgba64 *QT_FASTCALL fetchRGBA64ToRGBA64PM(QRgba64 *buffer, const u
                                                         const QList<QRgb> *, QDitherInfo *)
 {
     const QRgba64 *s = reinterpret_cast<const QRgba64 *>(src) + index;
+#ifdef __SSE2__
+    for (int i = 0; i < count; ++i) {
+        __m128i vs = _mm_loadl_epi64((const __m128i *)(s + i));
+        __m128i va = _mm_shufflelo_epi16(vs, _MM_SHUFFLE(3, 3, 3, 3));
+        vs = multiplyAlpha65535(vs, va);
+        _mm_storel_epi64((__m128i *)(buffer + i), vs);
+    }
+#else
     for (int i = 0; i < count; ++i)
         buffer[i] = QRgba64::fromRgba64(s[i]).premultiplied();
+#endif
     return buffer;
 }