gdk: add i386 MMX RGB565 compositing

gdk/gdkpixelops-x86-mmx.c

@@ -114,3 +114,146 @@ _gdk_pixel_ops_composite_0888_mmx (guchar       *src_buf,
     _mm_empty ();
   }
 }
+
+static guint32
+expand_565 (guint16 value)
+{
+  guint32 hi;
+
+  hi = (((guint32) value & 0xf800U) >> 8) |
+       (((guint32) value & 0x07e0U) << 5) |
+       (((guint32) value & 0x001fU) << 19);
+
+  return hi |
+         ((hi >> 5) & 0x00070007U) |
+         ((hi >> 6) & 0x00000300U);
+}
+
+static guint16
+pack_565 (guint32 rgb)
+{
+  return (guint16)
+    (((rgb & 0x000000f8U) << 8) |
+     ((rgb & 0x0000fc00U) >> 5) |
+     ((rgb & 0x00f80000U) >> 19));
+}
+
+static guint32
+blend_565_mmx (__m64   zero,
+               __m64   ff,
+               __m64   bias,
+               guint32 src,
+               guint32 dest)
+{
+  __m64 s;
+  __m64 d;
+  __m64 a;
+  __m64 ia;
+  __m64 t;
+  guint32 result;
+
+  s = _mm_cvtsi32_si64 ((int) src);
+  d = _mm_cvtsi32_si64 ((int) dest);
+
+  s = _mm_unpacklo_pi8 (s, zero);
+  d = _mm_unpacklo_pi8 (d, zero);
+
+  a = _mm_srli_si64 (s, 48);
+  a = _mm_unpacklo_pi16 (a, a);
+  a = _mm_unpacklo_pi32 (a, a);
+
+  ia = _mm_sub_pi16 (ff, a);
+
+  t = _mm_add_pi16 (_mm_mullo_pi16 (s, a),
+                    _mm_mullo_pi16 (d, ia));
+
+  t = _mm_add_pi16 (t, bias);
+  t = _mm_add_pi16 (t, _mm_srli_pi16 (t, 8));
+  t = _mm_srli_pi16 (t, 8);
+  t = _mm_packs_pu16 (t, t);
+
+  __asm__ ("movd %1,%0"
+           : "=r" (result)
+           : "y" (t));
+
+  return result;
+}
+
+G_GNUC_INTERNAL void
+_gdk_pixel_ops_composite_565_mmx (guchar       *src_buf,
+                                  gint          src_rowstride,
+                                  guchar       *dest_buf,
+                                  gint          dest_rowstride,
+                                  GdkByteOrder  dest_byte_order,
+                                  gint          width,
+                                  gint          height)
+{
+  const __m64 zero = _mm_setzero_si64 ();
+  const __m64 ones = _mm_cmpeq_pi16 (zero, zero);
+  const __m64 ff = _mm_srli_pi16 (ones, 8);
+  const __m64 bias = _mm_set1_pi16 (0x80);
+  guchar *src = src_buf;
+  guchar *dest = dest_buf;
+
+  (void) dest_byte_order;
+
+  while (height--)
+    {
+      gint pairs = width >> 1;
+      guchar *p = src;
+      guchar *q = dest;
+
+      while (pairs--)
+        {
+          guint32 sp0;
+          guint32 sp1;
+          guint32 dpair;
+          guint32 dp0;
+          guint32 dp1;
+          guint32 rgb0;
+          guint32 rgb1;
+          guint32 result;
+
+          memcpy (&sp0, p, sizeof sp0);
+          memcpy (&sp1, p + 4, sizeof sp1);
+          memcpy (&dpair, q, sizeof dpair);
+
+          dp0 = expand_565 ((guint16) dpair);
+          dp1 = expand_565 ((guint16) (dpair >> 16));
+
+          rgb0 = blend_565_mmx (zero, ff, bias, sp0, dp0);
+          rgb1 = blend_565_mmx (zero, ff, bias, sp1, dp1);
+
+          result = (guint32) pack_565 (rgb0);
+          result |= (guint32) pack_565 (rgb1) << 16;
+
+          memcpy (q, &result, sizeof result);
+
+          p += 8;
+          q += 4;
+        }
+
+      if (width & 1)
+        {
+          guint32 sp;
+          guint16 dp16;
+          guint16 result;
+          guint32 dp;
+          guint32 rgb;
+
+          memcpy (&sp, p, sizeof sp);
+          memcpy (&dp16, q, sizeof dp16);
+
+          dp = expand_565 (dp16);
+          rgb = blend_565_mmx (zero, ff, bias, sp, dp);
+          result = pack_565 (rgb);
+
+          memcpy (q, &result, sizeof result);
+        }
+
+      src += src_rowstride;
+      dest += dest_rowstride;
+    }
+
+  _mm_empty ();
+}

gdk/gdkpixelops-x86.c

@@ -26,6 +26,15 @@ _gdk_pixel_ops_composite_0888_mmx (guchar       *src_buf,
                                    gint          width,
                                    gint          height);
 
+G_GNUC_INTERNAL void
+_gdk_pixel_ops_composite_565_mmx (guchar       *src_buf,
+                                  gint          src_rowstride,
+                                  guchar       *dest_buf,
+                                  gint          dest_rowstride,
+                                  GdkByteOrder  dest_byte_order,
+                                  gint          width,
+                                  gint          height);
+
 G_GNUC_INTERNAL void
 _gdk_pixel_ops_init (void)
 {
@@ -36,6 +45,10 @@ _gdk_pixel_ops_init (void)
 
   if (__get_cpuid (1, &eax, &ebx, &ecx, &edx) &&
       (edx & bit_MMX) != 0)
-    _gdk_pixel_ops.composite_0888 =
-      _gdk_pixel_ops_composite_0888_mmx;
+    {
+      _gdk_pixel_ops.composite_0888 =
+        _gdk_pixel_ops_composite_0888_mmx;
+      _gdk_pixel_ops.composite_565 =
+        _gdk_pixel_ops_composite_565_mmx;
+    }
 }