Add missing optimization for loading RGB32 to RGBA64 using NEON

The rest of the RGB64 routines were optimized, but the loading of RGB32 was not as it was originally not used much, but with ARGB32 using the RGB64 backend, it is essential for decent performance. Task-number: QTBUG-69724 Change-Id: I1c02411ed29d3d993427afde44dfa83689d117e0 Reviewed-by: Lars Knoll <[email protected]>
Tanvir-Aunjum-Sunny · Aug 3, 2018 · 1c8f9eb · 1c8f9eb
1 parent 66be544
commit 1c8f9eb
Showing 1 changed file with 61 additions and 0 deletions.
diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp
@@ -618,13 +618,62 @@ static inline void qConvertARGB32PMToARGB64PM_sse2(QRgba64 *buffer, const uint *
         *buffer++ = QRgba64::fromArgb32(s);
     }
 }
+#elif defined(__ARM_NEON__)
+template<bool RGBA, bool maskAlpha>
+static inline void qConvertARGB32PMToRGBA64PM_neon(QRgba64 *buffer, const uint *src, int count)
+{
+    if (count <= 0)
+        return;
+
+    const uint32x4_t amask = vdupq_n_u32(0xff000000);
+#if defined(Q_PROCESSOR_ARM_64)
+    const uint8x16_t rgbaMask  = { 2, 1, 0, 3, 6, 5, 4, 7, 10, 9, 8, 11, 14, 13, 12, 15};
+#else
+    const uint8x8_t rgbaMask  = { 2, 1, 0, 3, 6, 5, 4, 7 };
+#endif
+    int i = 0;
+    for (; i < count-3; i += 4) {
+        uint32x4_t vs32 = vld1q_u32(src);
+        src += 4;
+        if (maskAlpha)
+            vs32 = vorrq_u32(vs32, amask);
+        uint8x16_t vs8 = vreinterpretq_u8_u32(vs32);
+        if (!RGBA) {
+#if defined(Q_PROCESSOR_ARM_64)
+            vs8 = vqtbl1q_u8(vs8, rgbaMask);
+#else
+            // no vqtbl1q_u8
+            const uint8x8_t vlo = vtbl1_u8(vget_low_u8(vs8), rgbaMask);
+            const uint8x8_t vhi = vtbl1_u8(vget_high_u8(vs8), rgbaMask);
+            vs8 = vcombine_u8(vlo, vhi);
+#endif
+        }
+        uint8x16x2_t v = vzipq_u8(vs8, vs8);
+
+        vst1q_u16((uint16_t *)buffer, vreinterpretq_u16_u8(v.val[0]));
+        buffer += 2;
+        vst1q_u16((uint16_t *)buffer, vreinterpretq_u16_u8(v.val[1]));
+        buffer += 2;
+    }
+
+    SIMD_EPILOGUE(i, count, 3) {
+        uint s = *src++;
+        if (maskAlpha)
+            s = s | 0xff000000;
+        if (RGBA)
+            s = RGBA2ARGB(s);
+        *buffer++ = QRgba64::fromArgb32(s);
+    }
+}
 #endif
 
 static const QRgba64 *QT_FASTCALL convertRGB32ToRGB64(QRgba64 *buffer, const uint *src, int count,
                                                       const QVector<QRgb> *, QDitherInfo *)
 {
 #ifdef __SSE2__
     qConvertARGB32PMToARGB64PM_sse2<false, true>(buffer, src, count);
+#elif defined(__ARM_NEON__)
+    qConvertARGB32PMToRGBA64PM_neon<false, true>(buffer, src, count);
 #else
     for (int i = 0; i < count; ++i)
         buffer[i] = QRgba64::fromArgb32(0xff000000 | src[i]);
@@ -639,6 +688,10 @@ static const QRgba64 *QT_FASTCALL convertARGB32ToARGB64PM(QRgba64 *buffer, const
     qConvertARGB32PMToARGB64PM_sse2<false, false>(buffer, src, count);
     for (int i = 0; i < count; ++i)
         buffer[i] = buffer[i].premultiplied();
+#elif defined(__ARM_NEON__)
+    qConvertARGB32PMToRGBA64PM_neon<false, false>(buffer, src, count);
+    for (int i = 0; i < count; ++i)
+        buffer[i] = buffer[i].premultiplied();
 #else
     for (int i = 0; i < count; ++i)
         buffer[i] = QRgba64::fromArgb32(src[i]).premultiplied();
@@ -651,6 +704,8 @@ static const QRgba64 *QT_FASTCALL convertARGB32PMToARGB64PM(QRgba64 *buffer, con
 {
 #ifdef __SSE2__
     qConvertARGB32PMToARGB64PM_sse2<false, false>(buffer, src, count);
+#elif defined(__ARM_NEON__)
+    qConvertARGB32PMToRGBA64PM_neon<false, false>(buffer, src, count);
 #else
     for (int i = 0; i < count; ++i)
         buffer[i] = QRgba64::fromArgb32(src[i]);
@@ -665,6 +720,10 @@ static const QRgba64 *QT_FASTCALL convertRGBA8888ToARGB64PM(QRgba64 *buffer, con
     qConvertARGB32PMToARGB64PM_sse2<true, false>(buffer, src, count);
     for (int i = 0; i < count; ++i)
         buffer[i] = buffer[i].premultiplied();
+#elif defined(__ARM_NEON__)
+    qConvertARGB32PMToRGBA64PM_neon<true, false>(buffer, src, count);
+    for (int i = 0; i < count; ++i)
+        buffer[i] = buffer[i].premultiplied();
 #else
     for (int i = 0; i < count; ++i)
         buffer[i] = QRgba64::fromArgb32(RGBA2ARGB(src[i])).premultiplied();
@@ -677,6 +736,8 @@ static const QRgba64 *QT_FASTCALL convertRGBA8888PMToARGB64PM(QRgba64 *buffer, c
 {
 #ifdef __SSE2__
     qConvertARGB32PMToARGB64PM_sse2<true, false>(buffer, src, count);
+#elif defined(__ARM_NEON__)
+    qConvertARGB32PMToRGBA64PM_neon<true, false>(buffer, src, count);
 #else
     for (int i = 0; i < count; ++i)
         buffer[i] = QRgba64::fromArgb32(RGBA2ARGB(src[i]));