new IDCT code by Michael Niedermayer (michaelni@gmx.at) - #define SIMPLE_IDCT to enable

Originally committed as revision 234 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/Makefile

@@ -6,7 +6,7 @@ LDFLAGS= -g
 OBJS= common.o utils.o mpegvideo.o h263.o jrevdct.o jfdctfst.o \
       mpegaudio.o ac3enc.o mjpeg.o resample.o dsputil.o \
       motion_est.o imgconvert.o imgresample.o msmpeg4.o \
-      mpeg12.o h263dec.o rv10.o mpegaudiodec.o pcm.o
+      mpeg12.o h263dec.o rv10.o mpegaudiodec.o pcm.o simple_idct.o
 ASM_OBJS=
 
 # currently using libac3 for ac3 decoding
@@ -20,7 +20,8 @@ endif
 ifeq ($(TARGET_MMX),yes)
 OBJS += i386/fdct_mmx.o i386/cputest.o \
 	i386/dsputil_mmx.o i386/mpegvideo_mmx.o \
-        i386/idct_mmx.o i386/motion_est_mmx.o
+        i386/idct_mmx.o i386/motion_est_mmx.o \
+	i386/simple_idct_mmx.o
 endif
 
 # armv4l specific stuff

libavcodec/dsputil.c

@@ -20,6 +20,7 @@
 #include <stdio.h>
 #include "avcodec.h"
 #include "dsputil.h"
+#include "simple_idct.h"
 
 void (*ff_idct)(DCTELEM *block);
 void (*get_pixels)(DCTELEM *block, const UINT8 *pixels, int line_size);
@@ -388,6 +389,27 @@ int pix_abs16x16_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h)
 
 /* permute block according so that it corresponds to the MMX idct
    order */
+#ifdef SIMPLE_IDCT
+void block_permute(INT16 *block)
+{
+	int i;
+	INT16 temp[64];
+
+//	for(i=0; i<64; i++) temp[i] = block[ block_permute_op(i) ];
+	for(i=0; i<64; i++) temp[ block_permute_op(i) ] = block[i];
+
+	for(i=0; i<64; i++) block[i] = temp[i];
+/*
+	for(i=0; i<64; i++)
+	{
+		if((i&7)==0) printf("\n");
+		printf("%2d ", block[i]);
+	}
+*/
+}
+
+#else
+
 void block_permute(INT16 *block)
 {
     int tmp1, tmp2, tmp3, tmp4, tmp5, tmp6;
@@ -409,6 +431,7 @@ void block_permute(INT16 *block)
         block += 8;
     }
 }
+#endif
 
 void dsputil_init(void)
 {
@@ -425,7 +448,11 @@ void dsputil_init(void)
         squareTbl[i] = (i - 256) * (i - 256);
     }
 
+#ifdef SIMPLE_IDCT
+    ff_idct = simple_idct;
+#else
     ff_idct = j_rev_dct;
+#endif
     get_pixels = get_pixels_c;
     put_pixels_clamped = put_pixels_clamped_c;
     add_pixels_clamped = add_pixels_clamped_c;
@@ -449,6 +476,10 @@ void dsputil_init(void)
     use_permuted_idct = 0;
 #endif
 
+#ifdef SIMPLE_IDCT
+    if(ff_idct == simple_idct) use_permuted_idct=0;
+#endif
+
     if (use_permuted_idct) {
         /* permute for IDCT */
         for(i=0;i<64;i++) {

libavcodec/dsputil.h

@@ -67,10 +67,33 @@ int pix_abs16x16_x2_c(UINT8 *blk1, UINT8 *blk2, int lx, int h);
 int pix_abs16x16_y2_c(UINT8 *blk1, UINT8 *blk2, int lx, int h);
 int pix_abs16x16_xy2_c(UINT8 *blk1, UINT8 *blk2, int lx, int h);
 
+#if defined (SIMPLE_IDCT) && defined (HAVE_MMX)
+static inline int block_permute_op(int j)
+{
+static const int table[64]={
+	0x00, 0x08, 0x01, 0x09, 0x04, 0x0C, 0x05, 0x0D,
+	0x10, 0x18, 0x11, 0x19, 0x14, 0x1C, 0x15, 0x1D,
+	0x02, 0x0A, 0x03, 0x0B, 0x06, 0x0E, 0x07, 0x0F,
+	0x12, 0x1A, 0x13, 0x1B, 0x16, 0x1E, 0x17, 0x1F,
+	0x20, 0x28, 0x21, 0x29, 0x24, 0x2C, 0x25, 0x2D,
+	0x30, 0x38, 0x31, 0x39, 0x34, 0x3C, 0x35, 0x3D,
+	0x22, 0x2A, 0x23, 0x2B, 0x26, 0x2E, 0x27, 0x2F,
+	0x32, 0x3A, 0x33, 0x3B, 0x36, 0x3E, 0x37, 0x3F,
+};
+
+	return table[j];
+}
+#elif defined (SIMPLE_IDCT)
+static inline int block_permute_op(int j)
+{
+    return j;
+}
+#else
 static inline int block_permute_op(int j)
 {
     return (j & 0x38) | ((j & 6) >> 1) | ((j & 1) << 2);
 }
+#endif
 
 void block_permute(INT16 *block);
 

libavcodec/h263.c

@@ -554,7 +554,7 @@ static int mpeg4_pred_dc(MpegEncContext * s, int n, UINT16 **dc_val_ptr, int *di
     return pred;
 }
 
-void mpeg4_pred_ac(MpegEncContext * s, INT16 *block, int n, 
+void mpeg4_pred_ac(MpegEncContext * s, INT16 *block, int n,
                    int dir)
 {
     int x, y, wrap, i;
@@ -579,22 +579,22 @@ void mpeg4_pred_ac(MpegEncContext * s, INT16 *block, int n,
             /* left prediction */
             ac_val -= 16;
             for(i=1;i<8;i++) {
-                block[i*8] += ac_val[i];
+                block[block_permute_op(i*8)] += ac_val[i];
             }
         } else {
             /* top prediction */
             ac_val -= 16 * wrap;
             for(i=1;i<8;i++) {
-                block[i] += ac_val[i + 8];
+                block[block_permute_op(i)] += ac_val[i + 8];
             }
         }
     }
     /* left copy */
     for(i=1;i<8;i++)
-        ac_val1[i] = block[i * 8];
+        ac_val1[i] = block[block_permute_op(i * 8)];
     /* top copy */
     for(i=1;i<8;i++)
-        ac_val1[8 + i] = block[i];
+        ac_val1[8 + i] = block[block_permute_op(i)];
 }
 
 static inline void mpeg4_encode_dc(MpegEncContext * s, int level, int n, int *dir_ptr)

libavcodec/i386/dsputil_mmx.c

@@ -20,6 +20,7 @@
  */
 
 #include "../dsputil.h"
+#include "../simple_idct.h"
 
 int mm_flags; /* multimedia extension flags */
 
@@ -1047,5 +1048,9 @@ void dsputil_init_mmx(void)
         } else {
             ff_idct = ff_mmx_idct;
         }
+#ifdef SIMPLE_IDCT
+//	ff_idct = simple_idct;
+	ff_idct = simple_idct_mmx;
+#endif
     }
 }