Use NEON and preloading for 16 bit small / medium sized image blits.

This gives a nice speedup for blitting of small and medium sized
images by using preloading and avoiding function call overhead to
memcpy for each scanline. For larger image widths memcpy becomes more
efficient.

Speedups of up to 40 % for 64 pixel wide images were measured.
For image widths between 2 and 16 the speedup ranges between 12 %
and 28 %.

Task-number: QT-3401
Reviewed-by: Benjamin Poulain <benjamin.poulain@nokia.com>

1 files changed, 98 insertions, 0 deletions

diff --git a/src/gui/painting/qdrawhelper_neon.cpp b/src/gui/painting/qdrawhelper_neon.cpp
index ed15c5c..0afd077 100644
--- a/src/gui/painting/qdrawhelper_neon.cpp
+++ b/src/gui/painting/qdrawhelper_neon.cpp
@@ -167,6 +167,14 @@ pixman_composite_scanline_over_asm_neon (int32_t         w,
                                          const uint32_t *dst,
                                          const uint32_t *src);
 
+extern "C" void
+pixman_composite_src_0565_0565_asm_neon (int32_t   w,
+                                         int32_t   h,
+                                         uint16_t *dst,
+                                         int32_t   dst_stride,
+                                         uint16_t *src,
+                                         int32_t   src_stride);
+
 // qblendfunctions.cpp
 void qt_blend_argb32_on_rgb16_const_alpha(uchar *destPixels, int dbpl,
                                           const uchar *srcPixels, int sbpl,
@@ -200,6 +208,96 @@ void qt_blend_rgb16_on_argb32_neon(uchar *destPixels, int dbpl,
     pixman_composite_src_0565_8888_asm_neon(w, h, dst, dbpl, src, sbpl);
 }
 
+// qblendfunctions.cpp
+void qt_blend_rgb16_on_rgb16(uchar *dst, int dbpl,
+                             const uchar *src, int sbpl,
+                             int w, int h,
+                             int const_alpha);
+
+template <int N>
+static inline void scanLineBlit16(quint16 *dst, quint16 *src, int dstride)
+{
+    if (N >= 2) {
+        ((quint32 *)dst)[0] = ((quint32 *)src)[0];
+        __builtin_prefetch(dst + dstride, 1, 0);
+    }
+    for (int i = 1; i < N/2; ++i)
+        ((quint32 *)dst)[i] = ((quint32 *)src)[i];
+    if (N & 1)
+        dst[N-1] = src[N-1];
+}
+
+template <int Width>
+static inline void blockBlit16(quint16 *dst, quint16 *src, int dstride, int sstride, int h)
+{
+    union {
+        quintptr address;
+        quint16 *pointer;
+    } u;
+
+    u.pointer = dst;
+
+    if (u.address & 2) {
+        while (h--) {
+            // align dst
+            dst[0] = src[0];
+            if (Width > 1)
+                scanLineBlit16<Width-1>(dst + 1, src + 1, dstride);
+            dst += dstride;
+            src += sstride;
+        }
+    } else {
+        while (h--) {
+            scanLineBlit16<Width>(dst, src, dstride);
+
+            dst += dstride;
+            src += sstride;
+        }
+    }
+}
+
+void qt_blend_rgb16_on_rgb16_neon(uchar *destPixels, int dbpl,
+                                  const uchar *srcPixels, int sbpl,
+                                  int w, int h,
+                                  int const_alpha)
+{
+    // testing show that the default memcpy is faster for widths 150 and up
+    if (const_alpha != 256 || w >= 150) {
+        qt_blend_rgb16_on_rgb16(destPixels, dbpl, srcPixels, sbpl, w, h, const_alpha);
+        return;
+    }
+
+    int dstride = dbpl / 2;
+    int sstride = sbpl / 2;
+
+    quint16 *dst = (quint16 *) destPixels;
+    quint16 *src = (quint16 *) srcPixels;
+
+    switch (w) {
+#define BLOCKBLIT(n) case n: blockBlit16<n>(dst, src, dstride, sstride, h); return;
+    BLOCKBLIT(1);
+    BLOCKBLIT(2);
+    BLOCKBLIT(3);
+    BLOCKBLIT(4);
+    BLOCKBLIT(5);
+    BLOCKBLIT(6);
+    BLOCKBLIT(7);
+    BLOCKBLIT(8);
+    BLOCKBLIT(9);
+    BLOCKBLIT(10);
+    BLOCKBLIT(11);
+    BLOCKBLIT(12);
+    BLOCKBLIT(13);
+    BLOCKBLIT(14);
+    BLOCKBLIT(15);
+#undef BLOCKBLIT
+    default:
+        break;
+    }
+
+    pixman_composite_src_0565_0565_asm_neon (w, h, dst, dstride, src, sstride);
+}
+
 extern "C" void blend_8_pixels_argb32_on_rgb16_neon(quint16 *dst, const quint32 *src, int const_alpha);
 
 void qt_blend_argb32_on_rgb16_neon(uchar *destPixels, int dbpl,