videomixer: optimize blend code some more

Use more efficient formula that uses less multiplies.
Reduce the amount of scalar code, use MMX to calculate the desired
alpha value.
Unroll and handle 2 pixels in one iteration for improved pairing.

diff --git a/gst/videomixer/blend_ayuv.c b/gst/videomixer/blend_ayuv.c
index af5f1743018668c29336403b3afda76025fd2566..43704a23cdbfdef65132328b4d0e7b124da3e13b 100644 (file)
--- a/gst/videomixer/blend_ayuv.c
+++ b/gst/videomixer/blend_ayuv.c
@@ -185,16 +185,16 @@ gst_videomixer_blend_ayuv_ayuv (guint8 * src, gint xpos, gint ypos,
     gint src_width, gint src_height, gdouble src_alpha,
     guint8 * dest, gint dest_width, gint dest_height)
 {
-  gint alpha, s_alpha;
+  guint s_alpha, alpha;
   gint i, j;
   gint src_stride, dest_stride;
   gint src_add, dest_add;
-  gint Y, U, V;
 
   src_stride = src_width * 4;
   dest_stride = dest_width * 4;
 
   s_alpha = CLAMP ((gint) (src_alpha * 256), 0, 256);
+  //g_print ("%f %d\n", src_alpha, s_alpha);
 
   /* adjust src pointers for negative sizes */
   if (xpos < 0) {
@@ -222,6 +222,20 @@ gst_videomixer_blend_ayuv_ayuv (guint8 * src, gint xpos, gint ypos,
 
   for (i = 0; i < src_height; i++) {
     for (j = 0; j < src_width; j++) {
+#if 0
+      gint Y, U, V;
+
+      alpha = (src[0] * s_alpha) >> 8;
+      Y = dest[1];
+      U = dest[2];
+      V = dest[3];
+      dest[0] = 0xff;
+      dest[1] = (((src[1] - Y) * alpha) >> 8) + Y;
+      dest[2] = (((src[2] - U) * alpha) >> 8) + U;
+      dest[3] = (((src[3] - V) * alpha) >> 8) + V;
+#else
+      gint Y, U, V;
+
       alpha = (src[0] * s_alpha) >> 8;
       BLEND_MODE (dest[1], dest[2], dest[3], src[1], src[2], src[3],
           alpha, Y, U, V);
@@ -229,6 +243,7 @@ gst_videomixer_blend_ayuv_ayuv (guint8 * src, gint xpos, gint ypos,
       dest[1] = Y;
       dest[2] = U;
       dest[3] = V;
+#endif
 
       src += 4;
       dest += 4;
@@ -246,7 +261,7 @@ gst_videomixer_blend_ayuv_ayuv_mmx (guint8 * src, gint xpos, gint ypos,
     gint src_width, gint src_height, gdouble src_alpha,
     guint8 * dest, gint dest_width, gint dest_height)
 {
-  gint b_alpha;
+  gint s_alpha;
   gint i;
   gint src_stride, dest_stride;
   gint src_add, dest_add;
@@ -254,7 +269,7 @@ gst_videomixer_blend_ayuv_ayuv_mmx (guint8 * src, gint xpos, gint ypos,
   src_stride = src_width * 4;
   dest_stride = dest_width * 4;
 
-  b_alpha = CLAMP ((gint) (src_alpha * 255), 0, 255);
+  s_alpha = CLAMP ((gint) (src_alpha * 256), 0, 256);
 
   /* adjust src pointers for negative sizes */
   if (xpos < 0) {
@@ -283,47 +298,104 @@ gst_videomixer_blend_ayuv_ayuv_mmx (guint8 * src, gint xpos, gint ypos,
   for (i = 0; i < src_height; i++) {
     gulong old_ebx;
 
-      /* *INDENT-OFF* */
-      __asm__ __volatile__ (
-          "movl       %%ebx ,      %6   \n\t"
-          "pxor       %%mm7 ,   %%mm7   \n\t"   /* mm7 = 0 */
-          "pcmpeqd    %%mm6 ,   %%mm6   \n\t"   /* mm6 = 0xffff... */
-          "punpcklbw  %%mm7 ,   %%mm6   \n\t"   /* mm6 = 0x00ff00ff00ff... */
-          "pcmpeqd    %%mm5 ,   %%mm5   \n\t"   /* mm5 = 0xffff... */
-          "psrlq        $56 ,   %%mm5   \n\t"   /* mm5 = 0x0...0ff */
-          "xor        %%ebx ,   %%ebx   \n\t"   /* ebx = 0 */
-          "1:                           \n\t"
-          "movzbl      (%2) ,   %%eax   \n\t"   /* eax == source alpha */
-          "imul          %4 ,   %%eax   \n\t"   /* eax = source alpha * alpha */
-          "sar           $8 ,   %%eax   \n\t"   /* eax = (source alpha * alpha) / 256 */
-          "movd       %%eax ,   %%mm0   \n\t"   /* mm0 = apply alpha */
-          "movd        (%2) ,   %%mm2   \n\t"   /* mm2 = src */
-          "movd        (%3) ,   %%mm1   \n\t"   /* mm1 = dest */
-          "punpcklwd  %%mm0 ,   %%mm0   \n\t"
-          "punpckldq  %%mm0 ,   %%mm0   \n\t"   /* mm0 == 0a 0a 0a 0a */
-          "punpcklbw  %%mm7 ,   %%mm1   \n\t"   /* mm1 == dv du dy da */
-          "punpcklbw  %%mm7 ,   %%mm2   \n\t"   /* mm2 == sv su sy sa */
-          "pmullw     %%mm0 ,   %%mm2   \n\t"   /* mm2 == a * s */
-          "pandn      %%mm6 ,   %%mm0   \n\t"   /* mm0 == 255 - a */
-          "pmullw     %%mm0 ,   %%mm1   \n\t"   /* mm1 == (255 - a) * d */
-          "paddusw    %%mm2 ,   %%mm1   \n\t"   /* mm1 == s + d */
-          "psrlw         $8 ,   %%mm1   \n\t"
-          "packuswb   %%mm7 ,   %%mm1   \n\t"
-          "por        %%mm5 ,   %%mm1   \n\t"   /* mm1 = 0x.....ff */
-          "movd       %%mm1 ,    (%3)   \n\t"   /* dest = mm1 */
-          "add           $4 ,     %1    \n\t"
-          "add           $4 ,     %0    \n\t"
-          "add           $1 ,   %%ebx   \n\t"
-          "cmp        %%ebx ,      %5   \n\t"
-          "jne                     1b   \n\t"
-          "movl          %6 ,   %%ebx   \n\t"
-          :"=r" (src), "=r" (dest)
-          :"0" (src), "1" (dest), "r" (b_alpha), "r" (src_width), "m" (old_ebx)
-          :"%eax", "memory"
+    /*      (P1 * (256 - A) + (P2 * A)) / 256
+     * =>   (P1 * 256 - P1 * A + P2 * A) / 256
+     * =>   (P1 * 256 + A * (P2 - P1) / 256
+     * =>   P1 + (A * (P2 - P1)) / 256
+     */
+    /* *INDENT-OFF* */
+    __asm__ __volatile__ (
+        " movl       %%ebx ,      %6   \n\t"
+
+        " pcmpeqd    %%mm5 ,   %%mm5   \n\t"   /* mm5 = 0xffff... */
+        " psrld        $24 ,   %%mm5   \n\t"   /* mm5 = 00 00 00 ff 00 00 00 0ff, selector for alpha */
+        " mov           %4 ,   %%eax   \n\t"   /* eax = s_alpha */
+        " movd       %%eax ,   %%mm6   \n\t"   /* mm6 = s_alpha */
+        " punpckldq  %%mm6 ,   %%mm6   \n\t"   /* mm6 = 00 00 00 aa 00 00 00 aa, alpha scale factor */
+
+        " movl          %5 ,   %%ebx   \n\t"   /* ebx = src_width */
+        " test          $1 ,   %%ebx   \n\t"   /* check odd pixel */
+        " je                      1f   \n\t"
+
+        /* do odd pixel */
+        " movd        (%2) ,   %%mm2   \n\t"   /* mm2 = src,  00 00 00 00 sv su sy sa */
+        " movd        (%3) ,   %%mm1   \n\t"   /* mm1 = dest, 00 00 00 00 dv du dy da */
+        " movq       %%mm2 ,   %%mm0   \n\t"
+        " punpcklbw  %%mm7 ,   %%mm2   \n\t"   /* mm2 = 00 sv 00 su 00 sy 00 sa */
+        " pand       %%mm5 ,   %%mm0   \n\t"   /* mm0 = 00 00 00 00 00 00 00 sa, get alpha component  */
+        " punpcklbw  %%mm7 ,   %%mm1   \n\t"   /* mm1 = 00 dv 00 du 00 dy 00 da */
+        " pmullw     %%mm6 ,   %%mm0   \n\t"   /* mult with scale */
+        " psubw      %%mm1 ,   %%mm2   \n\t"   /* mm2 = mm2 - mm1 */
+        " punpcklwd  %%mm0 ,   %%mm0   \n\t"
+        " punpckldq  %%mm0 ,   %%mm0   \n\t"   /* mm0 == 00 aa 00 aa 00 aa 00 aa */
+        " psrlw         $8 ,   %%mm0   \n\t"
+        " pmullw     %%mm0 ,   %%mm2   \n\t"   /* mm2 == a * mm2 */
+        " psllw         $8 ,   %%mm1   \n\t"   /* scale up */
+        " paddw      %%mm1 ,   %%mm2   \n\t"   /* mm2 == mm2 + mm1 */
+        " psrlw         $8 ,   %%mm2   \n\t"   /* scale down */
+        " por        %%mm5 ,   %%mm2   \n\t"   /* set alpha to ff */
+        " packuswb   %%mm2 ,   %%mm2   \n\t" 
+        " movd       %%mm2 ,    (%3)   \n\t"   /* dest = mm1 */
+        " add           $4 ,     %1    \n\t"
+        " add           $4 ,     %0    \n\t"
+
+        "1:                            \n\t"
+        " sar           $1 ,   %%ebx   \n\t"   /* prepare for 2 pixel per loop */
+        " cmp           $0 ,   %%ebx   \n\t"
+        " je                      3f   \n\t"
+        "2:                            \n\t"
+
+        /* do even pixels */
+        " movq        (%2) ,   %%mm2   \n\t"   /* mm2 = src,  sv1 su1 sy1 sa1  sv0 su0 sy0 sa0 */
+        " movq        (%3) ,   %%mm1   \n\t"   /* mm1 = dest, dv1 du1 dy1 da1  dv0 du0 dy0 da0 */
+        " movq       %%mm2 ,   %%mm4   \n\t"
+        " movq       %%mm1 ,   %%mm3   \n\t"
+        " movq       %%mm2 ,   %%mm0   \n\t"   /* copy for doing the alpha */
+
+        " pxor       %%mm7 ,   %%mm7   \n\t"  
+        " punpcklbw  %%mm7 ,   %%mm2   \n\t"   /* mm2 = 00 sv0  00 su0  00 sy0  00 sa0 */
+        " punpckhbw  %%mm7 ,   %%mm4   \n\t"   /* mm4 = 00 sv1  00 su1  00 sy1  00 sa1 */
+        " punpcklbw  %%mm7 ,   %%mm1   \n\t"   /* mm1 = 00 dv0  00 du0  00 dy0  00 da0 */
+        " punpckhbw  %%mm7 ,   %%mm3   \n\t"   /* mm2 = 00 dv1  00 du1  00 dy1  00 da1 */
+
+        " pand       %%mm5 ,   %%mm0   \n\t"   /* mm0 = 00 00 00 sa1  00 00 00 sa0 */
+        " psubw      %%mm1 ,   %%mm2   \n\t"   /* mm2 = mm2 - mm1 */
+        " pmullw     %%mm6 ,   %%mm0   \n\t"   /* mult with scale */
+        " psubw      %%mm3 ,   %%mm4   \n\t"   /* mm4 = mm4 - mm3 */
+        " psrlw         $8 ,   %%mm0   \n\t"   /* scale back */
+        " movq       %%mm0 ,   %%mm7   \n\t"   /* save copy */
+        " punpcklwd  %%mm0 ,   %%mm0   \n\t"   /* mm0 = 00 00   00 00   00 aa0  00 aa0 */
+        " punpckhwd  %%mm7 ,   %%mm7   \n\t"   /* mm7 = 00 00   00 00   00 aa1  00 aa1 */
+        " punpckldq  %%mm0 ,   %%mm0   \n\t"   /* mm0 = 00 aa0  00 aa0  00 aa0  00 aa0 */
+        " punpckldq  %%mm7 ,   %%mm7   \n\t"   /* mm7 = 00 aa1  00 aa1  00 aa1  00 aa1 */
+
+        " pmullw     %%mm0 ,   %%mm2   \n\t"   /* mm2 == aa * mm2 */
+        " pmullw     %%mm7 ,   %%mm4   \n\t"   /* mm2 == aa * mm2 */
+        " psllw         $8 ,   %%mm1   \n\t"
+        " psllw         $8 ,   %%mm3   \n\t"
+        " paddw      %%mm1 ,   %%mm2   \n\t"   /* mm2 == mm2 + mm1 */
+        " paddw      %%mm3 ,   %%mm4   \n\t"   /* mm2 == mm2 + mm1 */
+
+        " psrlw         $8 ,   %%mm2   \n\t"
+        " psrlw         $8 ,   %%mm4   \n\t"
+        " packuswb   %%mm4 ,   %%mm2   \n\t"
+        " por        %%mm5 ,   %%mm2   \n\t"   /* set alpha to ff */
+        " movq       %%mm2 ,    (%3)   \n\t"
+
+        " add           $8 ,     %1    \n\t"
+        " add           $8 ,     %0    \n\t"
+        " dec           %%ebx          \n\t"
+        " jne                     2b   \n\t"
+
+        "3:                            \n\t"
+        " movl          %6 ,   %%ebx   \n\t"
+        :"=r" (src), "=r" (dest)
+        :"0" (src), "1" (dest), "m" (s_alpha), "m" (src_width), "m" (old_ebx)
+        :"%eax", "memory"
 #ifdef __MMX__
-          , "mm0", "mm1", "mm2", "mm5", "mm6", "mm7"
+        , "mm0", "mm1", "mm2", "mm5", "mm6", "mm7"
 #endif
-      );
+    );
       /* *INDENT-ON* */
     src += src_add;
     dest += dest_add;