span lists were a failure. sloweer than the current brute force method - and

slower to boot.

ooh.. found i was fuckign up the memcpy's. fixed :) and much faster too :)


SVN revision: 13103

legacy/evas/src/lib/engines/common/evas_blend_pixel_pixel.c

@@ -204,13 +204,13 @@ evas_common_copy_pixels_rgba_to_rgba_mmx(DATA32 *src, DATA32 *dst, int len)
    src_ptr = src;
    dst_ptr = dst;
    dst_end_ptr = dst + len;
-   dst_end_ptr_pre = dst + ((len / 10) * 10);
+   dst_end_ptr_pre = dst + ((len / 16) * 16);
    
    while (dst_ptr < dst_end_ptr_pre)
      {
-	MOVE_10DWORDS_MMX(src_ptr, dst_ptr);
-	src_ptr+=10;
-	dst_ptr+=10;
+	MOVE_16DWORDS_MMX(src_ptr, dst_ptr);
+	src_ptr+=16;
+	dst_ptr+=16;
      }
    while (dst_ptr < dst_end_ptr)
      {
@@ -310,15 +310,14 @@ evas_common_copy_pixels_rgba_to_rgba_sse(DATA32 *src, DATA32 *dst, int len)
    src_ptr = src;
    dst_ptr = dst;
    dst_end_ptr = dst + len;
-   dst_end_ptr_pre = dst + ((len / 10) * 10);
+   dst_end_ptr_pre = dst + ((len / 16) * 16);
 
    while (dst_ptr < dst_end_ptr_pre)
      {
-	prefetch(&src_ptr[128]);
-	prefetch(&dst_ptr[128]);
-	MOVE_10DWORDS_MMX(src_ptr, dst_ptr);
-	src_ptr+=10;
-	dst_ptr+=10;
+	prefetch(&src_ptr[16]);
+	MOVE_16DWORDS_MMX(src_ptr, dst_ptr);
+	src_ptr+=16;
+	dst_ptr+=16;
      }
    while (dst_ptr < dst_end_ptr)
      {
@@ -373,21 +372,21 @@ evas_common_copy_pixels_rev_rgba_to_rgba_mmx(DATA32 *src, DATA32 *dst, int len)
 {
    DATA32 *src_ptr, *dst_ptr, *dst_end_ptr, *dst_end_ptr_pre;
    
-   src_ptr = src + len - 10;
-   dst_ptr = dst + len - 10;
+   src_ptr = src + len - 16;
+   dst_ptr = dst + len - 16;
    dst_end_ptr = dst;
-   dst_end_ptr_pre = dst + len - ((len / 10) * 10);
+   dst_end_ptr_pre = dst + len - ((len / 16) * 16);
    
-   if (len >= 10)
+   if (len >= 16)
      {
 	while (dst_ptr >= dst_end_ptr_pre)
 	  {
-	     MOVE_10DWORDS_MMX(src_ptr, dst_ptr);
-	     src_ptr-=10;
-	     dst_ptr-=10;
+	     MOVE_16DWORDS_MMX(src_ptr, dst_ptr);
+	     src_ptr-=16;
+	     dst_ptr-=16;
 	  }
-	src_ptr+=9;
-	dst_ptr+=9;
+	src_ptr+=15;
+	dst_ptr+=15;
 	while (dst_ptr >= dst_end_ptr)
 	  {
 	     *dst_ptr = *src_ptr;
@@ -415,23 +414,22 @@ evas_common_copy_pixels_rev_rgba_to_rgba_sse(DATA32 *src, DATA32 *dst, int len)
 {
    DATA32 *src_ptr, *dst_ptr, *dst_end_ptr, *dst_end_ptr_pre;
    
-   src_ptr = src + len - 10;
-   dst_ptr = dst + len - 10;
+   src_ptr = src + len - 16;
+   dst_ptr = dst + len - 16;
    dst_end_ptr = dst;
-   dst_end_ptr_pre = dst + len - ((len / 10) * 10);
+   dst_end_ptr_pre = dst + len - ((len / 16) * 16);
 
-   if (len >= 10)
+   if (len >= 16)
      {
 	while (dst_ptr >= dst_end_ptr_pre)
 	  {
-	     prefetch(&src_ptr[-128]);
-	     prefetch(&dst_ptr[-128]);
+	     prefetch(&src_ptr[-16]);
 	     MOVE_10DWORDS_MMX(src_ptr, dst_ptr);
-	     src_ptr-=10;
-	     dst_ptr-=10;
+	     src_ptr-=16;
+	     dst_ptr-=16;
 	  }
-	src_ptr+=9;
-	dst_ptr+=9;
+	src_ptr+=15;
+	dst_ptr+=15;
 	while (dst_ptr >= dst_end_ptr)
 	  {
 	     *dst_ptr = *src_ptr;

legacy/evas/src/lib/engines/common/evas_image_main.c

@@ -76,22 +76,13 @@ evas_common_image_shutdown(void)
 #endif
 }
 
-/* alpha tiles! - asctually span lists - need to do it as span lists */
+#if 0
 void
 evas_common_image_surface_alpha_tiles_calc(RGBA_Surface *is, int tsize)
 {
    int x, y;
    DATA32 *ptr;
 
-#if 1  
-   return;
-#endif   
-   /* hmm i only get about a 15% speedup on my "best cases". the complexity
-    * imho isn't worth the small gain, so i have disabled it here :( (this
-    * is best case scenario - average case will be much less gain)
-    * 
-    * thought for now the only case is 
-    */
    if (is->spans) return;
    if (!(is->im->flags & RGBA_IMAGE_HAS_ALPHA)) return;
    /* FIXME: dont handle alpha only images yet */
@@ -149,26 +140,7 @@ evas_common_image_surface_alpha_tiles_calc(RGBA_Surface *is, int tsize)
 	  }
      }
 }
-
-void
-evas_common_image_surface_alpha_tiles_free(RGBA_Surface *is)
-{
-   int i;
-   
-   if (!is->spans) return;
-   for (i = 0; i < is->h; i++)
-     {
-	while (is->spans[i])
-	  {
-	     RGBA_Image_Span *sp;
-	     
-	     sp = is->spans[i];
-	     is->spans[i] = evas_object_list_remove(sp, sp);
-	     free(sp);
-	  }
-     }
-   free(is->spans);
-}
+#endif   
 
 RGBA_Surface *
 evas_common_image_surface_new(RGBA_Image *im)
@@ -219,7 +191,6 @@ evas_common_image_surface_dealloc(RGBA_Surface *is)
 	free(is->data);
 	is->data = NULL;
      }
-   evas_common_image_surface_alpha_tiles_free(is);
 }
 
 RGBA_Image *
@@ -524,7 +495,6 @@ evas_common_image_dirty(RGBA_Image *im)
 {
    int i;
    
-   if (im->image) evas_common_image_surface_alpha_tiles_free(im->image);
    evas_common_image_unstore(im);
    im->flags |= RGBA_IMAGE_IS_DIRTY;
 }

legacy/evas/src/lib/engines/common/evas_scale_smooth_scaler.c

@@ -209,8 +209,6 @@ SCALE_FUNC(RGBA_Image *src, RGBA_Image *dst,
  * -:- 
  *
  */
-   /* 8x8 tiles - this will incurr about a < 2% memory overhead */
-   evas_common_image_surface_alpha_tiles_calc(src->image, 8);
    
    /* if 1:1 scale */
    if ((dst_region_w == src_region_w) && 

legacy/evas/src/lib/engines/common/evas_scale_smooth_scaler_noscale.c

@@ -32,63 +32,11 @@
 	Gfx_Func_Blend_Src_Dst func;
 	
 	func = evas_common_draw_func_blend_get(src, dst, dst_clip_w);
-#if 0
-	/* part of the spans experiemnt. doesnt seem to help much on top of
-	 * what we already have
-	 */
-	if (src->image->spans)
+	for (y = 0; y < dst_clip_h; y++)
 	  {
-	     int x2, y2;
-	     int xoff, woff;
-	     RGBA_Image_Flags pflags;
-	     Gfx_Func_Blend_Src_Dst func_solid;
-	     
-	     pflags = src->flags;
-	     src->flags &= ~RGBA_IMAGE_HAS_ALPHA;
-	     func_solid = evas_common_draw_func_blend_get(src, dst, dst_clip_w);
-	     src->flags = pflags;
-	     
-	     x2 = (dst_clip_x - dst_region_x) + src_region_x;
-	     y2 = (dst_clip_y - dst_region_y) + src_region_y;
-	     for (y = 0; y < dst_clip_h; y++, y2++)
-	       {
-		  Evas_Object_List *l;
-		  
-		  for (l = src->image->spans[y2]; l; l = l->next)
-		    {
-		       RGBA_Image_Span *sp;
-		       
-		       sp = l;
-		       if ((sp->x + sp->w) > x2)
-			 {
-			    xoff = sp->x - x2;
-			    woff = sp->w;
-			    if (xoff < 0)
-			      {
-				 woff += xoff;
-				 xoff = 0;
-			      }
-			    if ((xoff + woff) > (dst_clip_w))
-			      woff += (dst_clip_w) - (xoff + woff);
-			    if (sp->v == 2)
-			      func_solid(ptr + xoff, dst_ptr + xoff, woff);
-			    else
-			      func(ptr + xoff, dst_ptr + xoff, woff);
-			 }
-		    }
-		  ptr += src_w;
-		  dst_ptr += dst_w;
-	       }
-	  }
-	else
-#endif	  
-	  {
-	     for (y = 0; y < dst_clip_h; y++)
-	       {
-		  func(ptr, dst_ptr, dst_clip_w);
-		  ptr += src_w;
-		  dst_ptr += dst_w;
-	       }
+	     func(ptr, dst_ptr, dst_clip_w);
+	     ptr += src_w;
+	     dst_ptr += dst_w;
 	  }
      }
 }

legacy/evas/src/lib/include/evas_common.h

@@ -291,14 +291,6 @@ struct _RGBA_Surface
    DATA32            *data;
    char               no_free : 1;
    RGBA_Image        *im;
-   RGBA_Image_Span  **spans;
-};
-
-struct _RGBA_Image_Span
-{
-   Evas_Object_List     _list_data;
-   int                  x, w;
-   int                  v;
 };
 
 struct _RGBA_Image
@@ -769,9 +761,6 @@ void evas_common_scale_rgba_in_to_out_clip_sample      (RGBA_Image *src, RGBA_Im
 /****/
 void          evas_common_image_init              (void);
 void          evas_common_image_shutdown          (void);
-
-void          evas_common_image_surface_alpha_tiles_calc(RGBA_Surface *is, int tsize);
-void          evas_common_image_surface_alpha_tiles_free(RGBA_Surface *is);
    
 RGBA_Surface *evas_common_image_surface_new       (RGBA_Image *im);
 void          evas_common_image_surface_free      (RGBA_Surface *is);

legacy/evas/src/lib/include/evas_mmx.h

@@ -573,6 +573,24 @@ typedef	union {
 				 : \
 				 : "r" (var) \
 				 );
+#define prefetch0(var) \
+	__asm__ __volatile__ ( \
+				 "prefetcht0 (%0) \n" \
+				 : \
+				 : "r" (var) \
+				 );
+#define prefetch1(var) \
+	__asm__ __volatile__ ( \
+				 "prefetcht1 (%0) \n" \
+				 : \
+				 : "r" (var) \
+				 );
+#define prefetch2(var) \
+	__asm__ __volatile__ ( \
+				 "prefetcht2 (%0) \n" \
+				 : \
+				 : "r" (var) \
+				 );
 #define pshufw(r1, r2, imm) \
 	__asm__ __volatile__ ( \
 				 "pshufw $" #imm ", %" #r1 ", %" #r2 " \n" \