swscale/rgb2rgb_template: Implement shuffle_bytes_0321_c and fix shuffle_bytes_2103_c on BE

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libswscale/rgb2rgb.c

@@ -51,6 +51,7 @@ void (*rgb16to15)(const uint8_t *src, uint8_t *dst, int src_size);
 void (*rgb15to16)(const uint8_t *src, uint8_t *dst, int src_size);
 void (*rgb15to32)(const uint8_t *src, uint8_t *dst, int src_size);
 
+void (*shuffle_bytes_0321)(const uint8_t *src, uint8_t *dst, int src_size);
 void (*shuffle_bytes_2103)(const uint8_t *src, uint8_t *dst, int src_size);
 
 void (*yv12toyuy2)(const uint8_t *ysrc, const uint8_t *usrc,
@@ -333,7 +334,6 @@ void shuffle_bytes_ ## a ## b ## c ## d(const uint8_t *src,             \
     }                                                                   \
 }
 
-DEFINE_SHUFFLE_BYTES(0, 3, 2, 1)
 DEFINE_SHUFFLE_BYTES(1, 2, 3, 0)
 DEFINE_SHUFFLE_BYTES(3, 0, 1, 2)
 DEFINE_SHUFFLE_BYTES(3, 2, 1, 0)

libswscale/rgb2rgb.h

@@ -50,6 +50,7 @@ extern void    (*rgb24to15)(const uint8_t *src, uint8_t *dst, int src_size);
 extern void (*rgb32tobgr16)(const uint8_t *src, uint8_t *dst, int src_size);
 extern void (*rgb32tobgr15)(const uint8_t *src, uint8_t *dst, int src_size);
 
+extern void (*shuffle_bytes_0321)(const uint8_t *src, uint8_t *dst, int src_size);
 extern void (*shuffle_bytes_2103)(const uint8_t *src, uint8_t *dst, int src_size);
 
 void rgb64tobgr48_nobswap(const uint8_t *src, uint8_t *dst, int src_size);
@@ -71,7 +72,6 @@ void rgb15tobgr15(const uint8_t *src, uint8_t *dst, int src_size);
 void rgb12tobgr12(const uint8_t *src, uint8_t *dst, int src_size);
 void    rgb12to15(const uint8_t *src, uint8_t *dst, int src_size);
 
-void shuffle_bytes_0321(const uint8_t *src, uint8_t *dst, int src_size);
 void shuffle_bytes_1230(const uint8_t *src, uint8_t *dst, int src_size);
 void shuffle_bytes_3012(const uint8_t *src, uint8_t *dst, int src_size);
 void shuffle_bytes_3210(const uint8_t *src, uint8_t *dst, int src_size);

libswscale/rgb2rgb_template.c

@@ -328,6 +328,20 @@ static inline void shuffle_bytes_2103_c(const uint8_t *src, uint8_t *dst,
     }
 }
 
+static inline void shuffle_bytes_0321_c(const uint8_t *src, uint8_t *dst,
+                                        int src_size)
+{
+    int idx          = 15  - src_size;
+    const uint8_t *s = src - idx;
+    uint8_t *d       = dst - idx;
+
+    for (; idx < 15; idx += 4) {
+        register unsigned v   = *(const uint32_t *)&s[idx], g = v & 0x00ff00ff;
+        v                    &= 0xff00ff00;
+        *(uint32_t *)&d[idx]  = (v >> 16) + g + (v << 16);
+    }
+}
+
 static inline void rgb24tobgr24_c(const uint8_t *src, uint8_t *dst, int src_size)
 {
     unsigned i;
@@ -929,7 +943,11 @@ static av_cold void rgb2rgb_init_c(void)
     rgb24to15          = rgb24to15_c;
     rgb24to16          = rgb24to16_c;
     rgb24tobgr24       = rgb24tobgr24_c;
-#if !HAVE_BIGENDIAN
+#if HAVE_BIGENDIAN
+    shuffle_bytes_0321 = shuffle_bytes_2103_c;
+    shuffle_bytes_2103 = shuffle_bytes_0321_c;
+#else
+    shuffle_bytes_0321 = shuffle_bytes_0321_c;
     shuffle_bytes_2103 = shuffle_bytes_2103_c;
 #endif
     rgb32tobgr16       = rgb32tobgr16_c;