adding blackfin optimized yuyvtoyv12 converter

Originally committed as revision 23831 to svn://svn.mplayerhq.hu/mplayer/trunk/libswscale

libswscale/internal_bfin.S

@@ -531,3 +531,73 @@ DEFUN(uyvytoyv12, mL3,  (const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8
         unlink;
         rts;
 DEFUN_END(uyvytoyv12)
+
+DEFUN(yuyvtoyv12, mL3,  (const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
+                         long width, long height,
+                         long lumStride, long chromStride, long srcStride)):
+        link 0;
+        [--sp] = (r7:4,p5:4);
+
+        p0 = r1;       // Y top even
+
+        i2 = r2; // *u
+        r2 = [fp + ARG_vdst];
+        i3 = r2; // *v
+
+        r1 = [fp + ARG_srcStride];
+        r2 = r0 + r1;
+        r1 += -8;  // i0,i1 is pre read need to correct
+        m0 = r1;
+
+        i0 = r0;  // uyvy_T even
+        i1 = r2;  // uyvy_B odd
+
+        p2 = [fp + ARG_lumStride];
+        p1 = p0 + p2;  // Y bot odd
+
+        p5 = [fp + ARG_width];
+        p4 = [fp + ARG_height];
+        r0 = p5;
+        p4 = p4 >> 1;
+        p5 = p5 >> 2;
+
+        r2 = [fp + ARG_chromStride];
+        r0 = r0 >> 1;
+        r2 = r2 - r0;
+        m1 = r2;
+
+        /*   I0,I1 - src input line pointers
+         *   p0,p1 - luma output line pointers
+         *   I2    - dstU
+         *   I3    - dstV
+         */
+
+        lsetup (0f, 1f) lc1 = p4;   // H/2
+0:        r0 = [i0++] || r2 = [i1++];
+          r1 = [i0++] || r3 = [i1++];
+          r4 = bytepack(r0, r1);
+          r5 = bytepack(r2, r3);
+          lsetup (2f, 3f) lc0 = p5; // W/4
+2:          r0 = r0 >> 8(v) || [p0++] = r4;  // yyyy-even
+            r1 = r1 >> 8(v) || [p1++] = r5;  // yyyy-odd
+            r2 = r2 >> 8(v);
+            r3 = r3 >> 8(v);
+            r4 = byteop1p(r1:0, r3:2);
+            r5 = byteop1p(r1:0, r3:2) (r);
+            r6 = pack(r5.l, r4.l);
+            r7 = pack(r5.h, r4.h)         ||  r0 = [i0++] || r2 = [i1++];
+            r6 = bytepack(r6, r7)         ||  r1 = [i0++] || r3 = [i1++];
+            r4 = bytepack(r0, r1)         ||  w[i2++] = r6.l; // uu
+3:          r5 = bytepack(r2, r3)         ||  w[i3++] = r6.h; // vv
+
+          i0 += m0;
+          i1 += m0;
+          i2 += m1;
+          i3 += m1;
+          p0 = p0 + p2;
+1:        p1 = p1 + p2;
+
+        (r7:4,p5:4) = [sp++];
+        unlink;
+        rts;
+DEFUN_END(yuyvtoyv12)