Do not use a macro to clip integer to uint8 (2x faster YUVA->RV32)

[vlc] / modules / video_filter / blend.c

diff --git a/modules/video_filter/blend.c b/modules/video_filter/blend.c
index 96813781b117cfdcefeaf79fe9ee36f8c92baf6c..f680f04415200c3ba5623403f7702bc0d4aa3d3b 100644 (file)
--- a/modules/video_filter/blend.c
+++ b/modules/video_filter/blend.c
@@ -25,7 +25,12 @@
 /*****************************************************************************
  * Preamble
  *****************************************************************************/
-#include <vlc/vlc.h>
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <vlc_common.h>
+#include <vlc_plugin.h>
 #include <vlc_vout.h>
 #include "vlc_filter.h"
 
@@ -92,7 +97,7 @@ static void BlendRGBAR24( filter_t *, picture_t *, picture_t *, picture_t *,
  * Module descriptor
  *****************************************************************************/
 vlc_module_begin();
-    set_description( _("Video pictures blending") );
+    set_description( N_("Video pictures blending") );
     set_capability( "video blending", 100 );
     set_callbacks( OpenFilter, CloseFilter );
 vlc_module_end();
@@ -129,10 +134,7 @@ static int OpenFilter( vlc_object_t *p_this )
     /* Allocate the memory needed to store the decoder's structure */
     if( ( p_filter->p_sys = p_sys =
           (filter_sys_t *)malloc(sizeof(filter_sys_t)) ) == NULL )
-    {
-        msg_Err( p_filter, "out of memory" );
-        return VLC_EGENERIC;
-    }
+        return VLC_ENOMEM;
 
     /* Misc init */
     p_filter->pf_video_blend = Blend;
@@ -141,7 +143,6 @@ static int OpenFilter( vlc_object_t *p_this )
              (char *)&p_filter->fmt_in.video.i_chroma,
              (char *)&p_filter->fmt_out.video.i_chroma );
 
-
     return VLC_SUCCESS;
 }
 
@@ -178,7 +179,7 @@ static void Blend( filter_t *p_filter, picture_t *p_dst,
     if( i_width <= 0 || i_height <= 0 ) return;
 
 #if 0
-    printf( "chroma: %4.4s -> %4.4s\n",
+    msg_Dbg( p_filter, "chroma: %4.4s -> %4.4s\n",
              (char *)&p_filter->fmt_in.video.i_chroma,
              (char *)&p_filter->fmt_out.video.i_chroma );
 #endif
@@ -311,6 +312,15 @@ static void Blend( filter_t *p_filter, picture_t *p_dst,
 /***********************************************************************
  * Utils
  ***********************************************************************/
+static inline uint8_t vlc_uint8( int v )
+{
+    if( v > 255 )
+        return 255;
+    else if( v < 0 )
+        return 0;
+    return v;
+}
+
 static inline void yuv_to_rgb( int *r, int *g, int *b,
                                uint8_t y1, uint8_t u1, uint8_t v1 )
 {
@@ -318,7 +328,6 @@ static inline void yuv_to_rgb( int *r, int *g, int *b,
 #   define SCALEBITS 10
 #   define ONE_HALF  (1 << (SCALEBITS - 1))
 #   define FIX(x)    ((int) ((x) * (1<<SCALEBITS) + 0.5))
-#   define CLAMP( x ) (((x) > 255) ? 255 : ((x) < 0) ? 0 : (x));
 
     int y, cb, cr, r_add, g_add, b_add;
 
@@ -329,9 +338,12 @@ static inline void yuv_to_rgb( int *r, int *g, int *b,
             - FIX(0.71414*255.0/224.0) * cr + ONE_HALF;
     b_add = FIX(1.77200*255.0/224.0) * cb + ONE_HALF;
     y = (y1 - 16) * FIX(255.0/219.0);
-    *r = CLAMP((y + r_add) >> SCALEBITS);
-    *g = CLAMP((y + g_add) >> SCALEBITS);
-    *b = CLAMP((y + b_add) >> SCALEBITS);
+    *r = vlc_uint8( (y + r_add) >> SCALEBITS );
+    *g = vlc_uint8( (y + g_add) >> SCALEBITS );
+    *b = vlc_uint8( (y + b_add) >> SCALEBITS );
+#undef FIX
+#undef ONE_HALF
+#undef SCALEBITS
 }
 
 static inline void rgb_to_yuv( uint8_t *y, uint8_t *u, uint8_t *v,
@@ -356,7 +368,7 @@ static void BlendI420( filter_t *p_filter, picture_t *p_dst,
     uint8_t *p_src1_v, *p_src2_v, *p_dst_v;
     uint8_t *p_trans;
     int i_x, i_y, i_trans = 0;
-    vlc_bool_t b_even_scanline = i_y_offset % 2;
+    bool b_even_scanline = i_y_offset % 2;
 
     i_dst_pitch = p_dst->p[Y_PLANE].i_pitch;
     p_dst_y = p_dst->p[Y_PLANE].p_pixels + i_x_offset +
@@ -591,11 +603,12 @@ static void BlendR24( filter_t *p_filter, picture_t *p_dst_pic,
 #define TRANS_BITS  8
 
     if( (i_pix_pitch == 4)
-     && (((((int)p_dst)|((int)p_src1)|i_dst_pitch|i_src1_pitch) & 3) == 0) )
+     && (((((intptr_t)p_dst)|((intptr_t)p_src1)|i_dst_pitch|i_src1_pitch)
+          & 3) == 0) )
     {
         /*
-        ** if picture pixels are 32 bits long and lines addresses are 32 bit aligned,
-        ** optimize rendering
+        ** if picture pixels are 32 bits long and lines addresses are 32 bit
+        ** aligned, optimize rendering
         */
         uint32_t *p32_dst = (uint32_t *)p_dst;
         uint32_t i32_dst_pitch = (uint32_t)(i_dst_pitch>>2);
@@ -743,7 +756,7 @@ static void BlendR24( filter_t *p_filter, picture_t *p_dst_pic,
                     /* Completely transparent. Don't change pixel */
                     continue;
                 }
-                else 
+                else
                 {
                     int i_pos = i_x * i_pix_pitch;
                     if( i_trans == MAX_TRANS )
@@ -798,7 +811,7 @@ static void BlendYUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
     uint8_t *p_src2_u, *p_src2_v;
     uint8_t *p_trans;
     int i_x, i_y, i_pix_pitch, i_trans = 0;
-    vlc_bool_t b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
+    bool b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
     int i_l_offset = 0, i_u_offset = 0, i_v_offset = 0;
 
     if( p_filter->fmt_out.video.i_chroma == VLC_FOURCC('Y','U','Y','2') )
@@ -936,7 +949,7 @@ static void BlendI420I420( filter_t *p_filter, picture_t *p_dst,
     uint8_t *p_src1_u, *p_src2_u, *p_dst_u;
     uint8_t *p_src1_v, *p_src2_v, *p_dst_v;
     int i_x, i_y;
-    vlc_bool_t b_even_scanline = i_y_offset % 2;
+    bool b_even_scanline = i_y_offset % 2;
 
     i_dst_pitch = p_dst->p[Y_PLANE].i_pitch;
     p_dst_y = p_dst->p[Y_PLANE].p_pixels + i_x_offset +
@@ -1047,7 +1060,7 @@ static void BlendI420I420_no_alpha( filter_t *p_filter, picture_t *p_dst,
     uint8_t *p_src1_u, *p_src2_u, *p_dst_u;
     uint8_t *p_src1_v, *p_src2_v, *p_dst_v;
     int i_y;
-    vlc_bool_t b_even_scanline = i_y_offset % 2;
+    bool b_even_scanline = i_y_offset % 2;
 
     i_dst_pitch = p_dst->p[Y_PLANE].i_pitch;
     p_dst_y = p_dst->p[Y_PLANE].p_pixels + i_x_offset +
@@ -1096,7 +1109,7 @@ static void BlendI420I420_no_alpha( filter_t *p_filter, picture_t *p_dst,
          p_src2_y += i_src2_pitch )
     {
         /* Completely opaque. Completely overwrite underlying pixel */
-        p_filter->p_libvlc->pf_memcpy( p_dst_y, p_src2_y, i_width );
+        vlc_memcpy( p_dst_y, p_src2_y, i_width );
         if( b_even_scanline )
         {
             p_dst_u  += i_dst_pitch/2;
@@ -1106,8 +1119,8 @@ static void BlendI420I420_no_alpha( filter_t *p_filter, picture_t *p_dst,
         }
         else
         {
-            p_filter->p_libvlc->pf_memcpy( p_dst_u, p_src2_u, i_width/2 );
-            p_filter->p_libvlc->pf_memcpy( p_dst_v, p_src2_v, i_width/2 );
+            vlc_memcpy( p_dst_u, p_src2_u, i_width/2 );
+            vlc_memcpy( p_dst_v, p_src2_v, i_width/2 );
         }
         b_even_scanline = !b_even_scanline;
         if( i_y%2 == 1 )
@@ -1291,7 +1304,7 @@ static void BlendI420YUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
     uint8_t *p_dst, *p_src1, *p_src2_y;
     uint8_t *p_src2_u, *p_src2_v;
     int i_x, i_y, i_pix_pitch;
-    vlc_bool_t b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
+    bool b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
     int i_l_offset = 0, i_u_offset = 0, i_v_offset = 0;
 
     if( p_filter->fmt_out.video.i_chroma == VLC_FOURCC('Y','U','Y','2') )
@@ -1408,7 +1421,7 @@ static void BlendPalI420( filter_t *p_filter, picture_t *p_dst,
     uint8_t *p_src1_u, *p_dst_u;
     uint8_t *p_src1_v, *p_dst_v;
     int i_x, i_y, i_trans;
-    vlc_bool_t b_even_scanline = i_y_offset % 2;
+    bool b_even_scanline = i_y_offset % 2;
 
     i_dst_pitch = p_dst->p[Y_PLANE].i_pitch;
     p_dst_y = p_dst->p[Y_PLANE].p_pixels + i_x_offset +
@@ -1513,7 +1526,7 @@ static void BlendPalYUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
     int i_src1_pitch, i_src2_pitch, i_dst_pitch;
     uint8_t *p_src1, *p_src2, *p_dst;
     int i_x, i_y, i_pix_pitch, i_trans;
-    vlc_bool_t b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
+    bool b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
     int i_l_offset = 0, i_u_offset = 0, i_v_offset = 0;
 
     if( p_filter->fmt_out.video.i_chroma == VLC_FOURCC('Y','U','Y','2') )
@@ -1742,7 +1755,7 @@ static void BlendRGBAI420( filter_t *p_filter, picture_t *p_dst,
     int i_x, i_y, i_trans;
     uint8_t y, u, v;
 
-    vlc_bool_t b_even_scanline = i_y_offset % 2;
+    bool b_even_scanline = i_y_offset % 2;
 
     i_dst_pitch = p_dst->p[Y_PLANE].i_pitch;
     p_dst_y = p_dst->p[Y_PLANE].p_pixels + i_x_offset +
@@ -2012,7 +2025,7 @@ static void BlendRGBAYUVPacked( filter_t *p_filter, picture_t *p_dst_pic,
     uint8_t *p_dst, *p_src1, *p_src2;
     uint8_t *p_trans;
     int i_x, i_y, i_pix_pitch, i_trans;
-    vlc_bool_t b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
+    bool b_even = !((i_x_offset + p_filter->fmt_out.video.i_x_offset)%2);
     int i_l_offset = 0, i_u_offset = 0, i_v_offset = 0;
     uint8_t y, u, v;