gst/deinterlace2/tvtime/greedy.c: Add plain MMX implementation for the greedyl method.

Original commit message from CVS:
* gst/deinterlace2/tvtime/greedy.c:
(deinterlace_greedy_packed422_scanline_mmx),
(deinterlace_greedy_packed422_scanline):
Add plain MMX implementation for the greedyl method.

2 files changed, 111 insertions, 0 deletions

diff --git a/ChangeLog b/ChangeLog
index 81c7d765..eda3bad3 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,12 @@
 2008-06-24  Sebastian Dröge  <sebastian.droege@collabora.co.uk>
 
+	* gst/deinterlace2/tvtime/greedy.c:
+	(deinterlace_greedy_packed422_scanline_mmx),
+	(deinterlace_greedy_packed422_scanline):
+	Add plain MMX implementation for the greedyl method.
+
+2008-06-24  Sebastian Dröge  <sebastian.droege@collabora.co.uk>
+
 	* gst/deinterlace2/Makefile.am:
 	Move the assembly includes to noinst_HEADERS where they belong.
 
diff --git a/gst/deinterlace2/tvtime/greedy.c b/gst/deinterlace2/tvtime/greedy.c
index 66b8799d..e1a9b1cf 100644
--- a/gst/deinterlace2/tvtime/greedy.c
+++ b/gst/deinterlace2/tvtime/greedy.c
@@ -111,6 +111,107 @@ deinterlace_greedy_packed422_scanline_c (GstDeinterlace2 * object,
 
 #ifdef HAVE_CPU_I386
 #include "mmx.h"
+static void
+deinterlace_greedy_packed422_scanline_mmx (GstDeinterlace2 * object,
+    uint8_t * m0, uint8_t * t1, uint8_t * b1, uint8_t * m2, uint8_t * output,
+    int width)
+{
+  mmx_t MaxComb;
+
+  // How badly do we let it weave? 0-255
+  MaxComb.ub[0] = GreedyMaxComb;
+  MaxComb.ub[1] = GreedyMaxComb;
+  MaxComb.ub[2] = GreedyMaxComb;
+  MaxComb.ub[3] = GreedyMaxComb;
+  MaxComb.ub[4] = GreedyMaxComb;
+  MaxComb.ub[5] = GreedyMaxComb;
+  MaxComb.ub[6] = GreedyMaxComb;
+  MaxComb.ub[7] = GreedyMaxComb;
+
+  // L2 == m0
+  // L1 == t1
+  // L3 == b1
+  // LP2 == m2
+
+  for (; width > 7; width -= 8) {
+    movq_m2r (*t1, mm1);        // L1
+    movq_m2r (*m0, mm2);        // L2
+    movq_m2r (*b1, mm3);        // L3
+    movq_m2r (*m2, mm0);        // LP2
+
+    // average L1 and L3 leave result in mm4
+    movq_r2r (mm1, mm4);        // L1
+    movq_r2r (mm3, mm5);        // L3
+    psrlw_i2r (1, mm4);         // L1/2
+    psrlw_i2r (1, mm5);         // L3/2
+    paddusb_r2r (mm5, mm4);     // (L1 + L3) / 2
+
+    // get abs value of possible L2 comb
+    movq_r2r (mm2, mm7);        // L2
+    psubusb_r2r (mm4, mm7);     // L2 - avg
+    movq_r2r (mm4, mm5);        // avg
+    psubusb_r2r (mm2, mm5);     // avg - L2
+    por_r2r (mm7, mm5);         // abs(avg-L2)
+
+
+    // get abs value of possible LP2 comb
+    movq_r2r (mm0, mm7);        // LP2
+    psubusb_r2r (mm4, mm7);     // LP2 - avg
+    psubusb_r2r (mm0, mm4);     // avg - LP2
+    por_r2r (mm7, mm4);         // abs(avg-LP2)
+
+    // use L2 or LP2 depending upon which makes smaller comb
+    psubusb_r2r (mm5, mm4);     // see if it goes to zero
+    psubusb_r2r (mm5, mm5);     // 0
+    pcmpeqb_r2r (mm5, mm4);     // if (mm4=0) then FF else 0
+    pcmpeqb_r2r (mm4, mm5);     // opposite of mm4
+
+    // if Comb(LP2) <= Comb(L2) then mm4=ff, mm5=0 else mm4=0, mm5 = 55
+    pand_r2r (mm2, mm5);        // use L2 if mm5 == ff, else 0
+    pand_r2r (mm0, mm4);        // use LP2 if mm4 = ff, else 0
+    por_r2r (mm5, mm4);         // may the best win
+
+    // Now lets clip our chosen value to be not outside of the range
+    // of the high/low range L1-L3 by more than abs(L1-L3)
+    // This allows some comb but limits the damages and also allows more
+    // detail than a boring oversmoothed clip.
+
+    movq_r2r (mm1, mm2);        // copy L1
+    psubusb_r2r (mm3, mm2);     // - L3, with saturation
+    paddusb_r2r (mm3, mm2);     // now = Max(L1,L3)
+
+    pcmpeqb_r2r (mm7, mm7);     // all ffffffff
+    psubusb_r2r (mm1, mm7);     // - L1 
+    paddusb_r2r (mm7, mm3);     // add, may sat at fff..
+    psubusb_r2r (mm7, mm3);     // now = Min(L1,L3)
+
+    // allow the value to be above the high or below the low by amt of MaxComb
+    paddusb_m2r (MaxComb, mm2); // increase max by diff
+    psubusb_m2r (MaxComb, mm3); // lower min by diff
+
+    psubusb_r2r (mm3, mm4);     // best - Min
+    paddusb_r2r (mm3, mm4);     // now = Max(best,Min(L1,L3)
+
+    pcmpeqb_r2r (mm7, mm7);     // all ffffffff
+    psubusb_r2r (mm4, mm7);     // - Max(best,Min(best,L3) 
+    paddusb_r2r (mm7, mm2);     // add may sat at FFF..
+    psubusb_r2r (mm7, mm2);     // now = Min( Max(best, Min(L1,L3), L2 )=L2 clipped
+
+    movq_r2m (mm2, *output);    // move in our clipped best
+
+    // Advance to the next set of pixels.
+    output += 8;
+    m0 += 8;
+    t1 += 8;
+    b1 += 8;
+    m2 += 8;
+  }
+  emms ();
+  if (width > 0)
+    deinterlace_greedy_packed422_scanline_c (object, m0, t1, b1, m2, output,
+        width);
+}
+
 #include "sse.h"
 
 static void
@@ -214,6 +315,9 @@ deinterlace_greedy_packed422_scanline (GstDeinterlace2 * object,
   if (object->cpu_feature_flags & OIL_IMPL_FLAG_MMXEXT) {
     deinterlace_greedy_packed422_scanline_mmxext (object, data->m0, data->t1,
         data->b1, data->m2, output, 2 * object->frame_width);
+  } else if (object->cpu_feature_flags & OIL_IMPL_FLAG_MMX) {
+    deinterlace_greedy_packed422_scanline_mmx (object, data->m0, data->t1,
+        data->b1, data->m2, output, 2 * object->frame_width);
   } else {
     deinterlace_greedy_packed422_scanline_c (object, data->m0, data->t1,
         data->b1, data->m2, output, 2 * object->frame_width);