* Activation des Sparses idct dans le video parser;
* Utilisation de la mmx pour la dct si l'architechture s'y pr�te
  (modif de Makefile et vdec_idct)
* idctmmx.S avec une forme d'appel identique a vdec_idct a ete ajoute a
  video_decoder.

Makefile

@@ -14,9 +14,9 @@
 #SHELL = /bin/sh
 
 # Video output settings
-#VIDEO=X11
+VIDEO=X11
 #VIDEO=DGA (not yet supported)
-VIDEO=FB
+#VIDEO=FB
 #VIDEO=GGI
 #VIDEO=BEOS (not yet supported)
 
@@ -27,8 +27,8 @@ ARCH=MMX
 #ARCH=SPARC
 
 # Decoder choice - ?? old decoder will be removed soon
-DECODER=old
-#DECODER=new
+#DECODER=old
+DECODER=new
 # !!! don't forget to run this command after changing decoder type !!!
 # touch input/input.c input/input_ctrl.c include/vlc.h include/video_decoder.h
 
@@ -232,8 +232,14 @@ misc_obj =			misc/mtime.o \
 						misc/netutils.o
 
 ifeq ($(ARCH),MMX)
+ifeq ($(DECODER),old)
 ASM_OBJ = 			video_decoder_ref/idctmmx.o \
 						video_output/video_yuv_mmx.o
+else
+ASM_OBJ = 			video_decoder/idctmmx.o \
+						video_output/video_yuv_mmx.o
+endif
+
 endif
 
 C_OBJ = $(interface_obj) \

src/video_decoder/vdec_idct.c

@@ -44,8 +44,6 @@
  * Local prototypes
  */
 
-//extern IDCT_mmx(dctelem_t*);
-
 /* Our current implementation is a fast DCT, we might move to a fast DFT or
  * an MMX DCT in the future. */
 
@@ -80,7 +78,6 @@ void vdec_InitIDCT (vdec_thread_t * p_vdec)
 void vdec_SparseIDCT (vdec_thread_t * p_vdec, dctelem_t * p_block, 
                       int i_sparse_pos)
 {
-    /*debug*/
     short int val;
     int * dp;
     int v;
@@ -92,20 +89,7 @@ void vdec_SparseIDCT (vdec_thread_t * p_vdec, dctelem_t * p_block,
     if ( i_sparse_pos == 0 ) 
     {
 	    dp=(int *)p_block;
-//        v=*p_block;
-/* cuisine a verifier */        
-/*        if (v < 0) 
-        {
-            val=-v;
-            val+=4;
-            val/=8;
-            val=-val;
-        }
-        else
-        {*/
-/*            val= (v + 4) /8; */
         val=RIGHT_SHIFT((*p_block + 4), 3);
-//        }        
         /* Compute int to assign.  This speeds things up a bit */
         v = ((val & 0xffff) | (val << 16));
         dp[0] = v;     dp[1] = v;     dp[2] = v;     dp[3] = v;
@@ -150,9 +134,10 @@ void vdec_SparseIDCT (vdec_thread_t * p_vdec, dctelem_t * p_block,
 /*****************************************************************************
  * vdec_IDCT : IDCT function for normal matrices
  *****************************************************************************/
+
+#ifndef HAVE_MMX
 void vdec_IDCT( vdec_thread_t * p_vdec, dctelem_t * p_block, int i_idontcare )
 {
-//IDCT_mmx(p_block);
 #if 0 
     /* dct classique: pour tester la meilleure entre la classique et la */
     /* no classique */
@@ -372,7 +357,7 @@ void vdec_IDCT( vdec_thread_t * p_vdec, dctelem_t * p_block, int i_idontcare )
     }
 #endif    
 
-#if 1 /*dct avec no classique*/    
+#if 1  /*dct avec non classique*/    
     
     s32 tmp0, tmp1, tmp2, tmp3;
     s32 tmp10, tmp11, tmp12, tmp13;
@@ -389,6 +374,7 @@ void vdec_IDCT( vdec_thread_t * p_vdec, dctelem_t * p_block, int i_idontcare )
 
     dataptr = p_block;
 
+    fprintf( stderr, "normal dct" );
     for (rowctr = DCTSIZE-1; rowctr >= 0; rowctr--) 
     {
         /* Due to quantization, we will usually find that many of the input
@@ -1550,3 +1536,4 @@ void vdec_IDCT( vdec_thread_t * p_vdec, dctelem_t * p_block, int i_idontcare )
     }
 #endif
 }
+#endif

src/video_output/video_yuv_mmx.S

@@ -51,6 +51,8 @@ sixbitu:            .quad 0xc0c0c0c0c0c0c0c0
 #define LocalFrameSize  156
 #define RegisterStorageSize  16
 
+//#define DOUBLE /*double le nombre de colonnes */
+
 /* Arguments: */
 #define YPlane                    LocalFrameSize + RegisterStorageSize +  4
 #define UPlane                    LocalFrameSize + RegisterStorageSize +  8
@@ -476,11 +478,29 @@ do_next_8x2_block:
   por        %mm1,%mm2               /* mm2: low RBG16 */
   psllw      GLeftShift(%esp),%mm3   /* shift high G 5 positions */
   por        %mm3,%mm7               /* mm7: high RBG16 */
+#ifdef DOUBLE
+  movq       %mm2,%mm1
+  movq       %mm7,%mm5
+  movq       %mm2,%mm0
+  movq       %mm7,%mm3
+  punpckhwd  %mm2,%mm1
+  punpckhwd  %mm7,%mm5
+  punpcklwd  %mm2,%mm0
+  punpcklwd  %mm7,%mm3
+  movq       %mm0,(%edi,%eax,)
+  movq       %mm1,8(%edi,%eax,)
+  movq       %mm3,16(%edi,%eax,)
+  movq       %mm5,24(%edi,%eax,)
+  addl       $32,%edi
+  addl       $4,%ebx
+#endif
+#ifndef DOUBLE
   movq       %mm2,(%edi,%eax,)
   movq       %mm7,8(%edi,%eax,)      /* aligned */
   addl       $16,%edi                /* ih take 16 bytes (8 pixels-16 bit) */
   addl       $4,%ebx                 /* ? to take 4 pixels together
                                         instead of 2 */
+#endif					
   jl         do_next_8x2_block
 
   addl       CCOSkipDistance(%esp),%edi /* go to begin of next line */

src/video_parser/vpar_blocks.c

@@ -1244,9 +1244,9 @@ static void vpar_DecodeMPEG2Non( vpar_thread_t * p_vpar, macroblock_t * p_mb, in
                                                         : i_level;
                 break;
             case DCT_EOB:
-                if( i_nc <= 0 )
+                if( i_nc <= 1 )
                 {
-                    p_mb->pf_idct[i_b] = vdec_IDCT;//vdec_SparseIDCT;
+                    p_mb->pf_idct[i_b] = vdec_SparseIDCT;
                     p_mb->pi_sparse_pos[i_b] = i_coef;
                 }
                 else
@@ -1479,7 +1479,7 @@ static void vpar_DecodeMPEG2Intra( vpar_thread_t * p_vpar, macroblock_t * p_mb,
             case DCT_EOB:
                 if( i_nc <= 1 )
                 {
-                    p_mb->pf_idct[i_b] = vdec_IDCT;//vdec_SparseIDCT;
+                    p_mb->pf_idct[i_b] = vdec_SparseIDCT;
                     p_mb->pi_sparse_pos[i_b] = i_coef;
                 }
                 else