* Altivec-enabled version of fastmemcpy

* Use ImageBuffer instead of ScreenBuffer in OS X video output (should be
faster)

Makefile

@@ -115,6 +115,7 @@ PLUGINS_TARGETS := a52/a52 \
 		memcpy/memcpymmx \
 		memcpy/memcpymmxext \
 		memcpy/memcpy3dn \
+		memcpy/memcpyaltivec \
 		mga/mga \
 		mga/xmga \
 		motion/motion \

plugins/macosx/vout_macosx.c

@@ -388,7 +388,7 @@ static void vout_Display( vout_thread_t *p_vout, picture_t *p_pic )
                     p_vout->p_sys->i_seq,
                     p_pic->p_sys->p_info,
                     p_pic->p_sys->i_size,                    
-                    codecFlagUseScreenBuffer, &flags, nil ) != noErr ) )
+                    codecFlagUseImageBuffer, &flags, nil ) != noErr ) )
     {
         intf_ErrMsg( "DecompressSequenceFrameS failed: %d", err );
     }
@@ -569,7 +569,7 @@ static int QTCreateSequence( vout_thread_t *p_vout )
                               NULL, NULL,
                               p_vout->p_sys->p_matrix,
                               0, NULL,
-                              codecFlagUseScreenBuffer,
+                              codecFlagUseImageBuffer,
                               codecLosslessQuality,
                               p_vout->p_sys->img_dc ) ) )
     {

plugins/memcpy/Makefile

@@ -2,3 +2,4 @@ memcpy_SOURCES = memcpy.c
 memcpymmx_SOURCES = memcpy.c
 memcpymmxext_SOURCES = memcpy.c
 memcpy3dn_SOURCES = memcpy.c
+memcpyaltivec_SOURCES = memcpy.c

plugins/memcpy/fastmemcpy.h

 /*****************************************************************************
  * fastmemcpy.h : fast memcpy routines
  *****************************************************************************
- * $Id: fastmemcpy.h,v 1.1 2001/12/03 16:18:37 sam Exp $
+ * $Id: fastmemcpy.h,v 1.2 2002/04/02 22:16:07 massiot Exp $
  *
  * Authors: various Linux kernel hackers
  *          various MPlayer hackers
  *          Nick Kurshev <nickols_k@mail.ru>
+ *          Christophe Massiot <massiot@via.ecp.fr> (Altivec)
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -382,3 +383,71 @@ void * _M( fast_memcpy )(void * to, const void * from, size_t len)
 
 
 #endif /* #if defined( HAVE_MMX2 ) || defined( HAVE_3DNOW ) || defined( HAVE_MMX ) */
+
+#ifdef HAVE_ALTIVEC
+#    if defined(CAN_COMPILE_C_ALTIVEC) || defined( __BUILD_ALTIVEC_ASM )
+
+#define vector_s16_t vector signed short
+#define vector_u16_t vector unsigned short
+#define vector_s8_t vector signed char
+#define vector_u8_t vector unsigned char
+#define vector_s32_t vector signed int
+#define vector_u32_t vector unsigned int
+#undef MMREG_SIZE
+#define MMREG_SIZE 16
+
+void * _M( fast_memcpy )(void * _to, const void * _from, size_t len)
+{
+    void * retval = _to;
+    u8 * to = (u8 *)_to;
+    u8 * from = (u8 *)_from;
+
+    if( len > 16 )
+    {
+        /* Align destination to MMREG_SIZE -boundary */
+        register unsigned long int delta;
+
+        delta = ((unsigned long)to)&(MMREG_SIZE-1);
+        if( delta )
+        {
+            delta = MMREG_SIZE - delta;
+            len -= delta;
+            memcpy(to, from, delta);
+            to += delta;
+            from += delta;
+        }
+
+        if( len & ~(MMREG_SIZE-1) )
+        {
+            vector_u8_t perm, ref0, ref1, tmp;
+
+            perm = vec_lvsl( 0, from );
+            ref0 = vec_ld( 0, from );
+            ref1 = vec_ld( 15, from );
+            from += 16;
+            len -= 16;
+            tmp = vec_perm( ref0, ref1, perm );
+            do
+            {
+                ref0 = vec_ld( 0, from );
+                ref1 = vec_ld( 15, from );
+                from += 16;
+                len -= 16;
+                vec_st( tmp, 0, to );
+                tmp = vec_perm( ref0, ref1, perm );
+                to += 16;
+            } while( len & ~(MMREG_SIZE-1) );
+            vec_st( tmp, 0, to );
+        }
+    }
+
+    if( len )
+    {
+        memcpy( to, from, len );
+    }
+
+    return retval;
+}
+
+#   endif
+#endif

plugins/memcpy/memcpy.c

@@ -2,7 +2,7 @@
  * memcpy.c : classic memcpy module
  *****************************************************************************
  * Copyright (C) 2001 VideoLAN
- * $Id: memcpy.c,v 1.5 2002/02/15 13:32:53 sam Exp $
+ * $Id: memcpy.c,v 1.6 2002/04/02 22:16:07 massiot Exp $
  *
  * Authors: Samuel Hocevar <sam@zoy.org>
  *
@@ -34,6 +34,7 @@
 #undef HAVE_SSE
 #undef HAVE_SSE2
 #undef HAVE_3DNOW
+#undef HAVE_ALTIVEC
 
 #if defined( MODULE_NAME_IS_memcpy3dn )
 #   define HAVE_3DNOW
@@ -44,6 +45,9 @@
 #elif defined( MODULE_NAME_IS_memcpymmxext )
 #   define HAVE_MMX2
 #   include "fastmemcpy.h"
+#elif defined( MODULE_NAME_IS_memcpyaltivec )
+#   define HAVE_ALTIVEC
+#   include "fastmemcpy.h"
 #endif
 
 /*****************************************************************************
@@ -87,6 +91,12 @@ MODULE_INIT_START
     ADD_REQUIREMENT( MMXEXT )
     ADD_SHORTCUT( "mmxext" )
     ADD_SHORTCUT( "memcpymmxext" )
+#elif defined( MODULE_NAME_IS_memcpyaltivec )
+    SET_DESCRIPTION( "Altivec memcpy module" )
+    ADD_CAPABILITY( MEMCPY, 100 )
+    ADD_REQUIREMENT( ALTIVEC )
+    ADD_SHORTCUT( "altivec" )
+    ADD_SHORTCUT( "memcpyaltivec" )
 #endif
 MODULE_INIT_STOP
 

src/input/input_ext-plugins.c

@@ -2,7 +2,7 @@
  * input_ext-plugins.c: useful functions for access and demux plug-ins
  *****************************************************************************
  * Copyright (C) 2001, 2002 VideoLAN
- * $Id: input_ext-plugins.c,v 1.3 2002/03/26 23:08:40 gbazin Exp $
+ * $Id: input_ext-plugins.c,v 1.4 2002/04/02 22:16:07 massiot Exp $
  *
  * Authors: Christophe Massiot <massiot@via.ecp.fr>
  *
@@ -536,9 +536,12 @@ ssize_t input_FillBuffer( input_thread_t * p_input )
     p_buf->i_refcount = 1;
 
     if( p_input->p_data_buffer != NULL )
+    {
+        if( i_remains )
         {
             FAST_MEMCPY( (byte_t *)p_buf + sizeof(data_buffer_t),
                          p_input->p_current_data, (size_t)i_remains );
+        }
         ReleaseBuffer( p_input->p_method_data, p_input->p_data_buffer );
     }