minor mmx2 optimization if the dct

Originally committed as revision 2423 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/dsputil.h

@@ -41,6 +41,7 @@ void ff_jpeg_fdct_islow (DCTELEM *data);
 void j_rev_dct (DCTELEM *data);
 
 void ff_fdct_mmx(DCTELEM *block);
+void ff_fdct_mmx2(DCTELEM *block);
 
 /* encoding scans */
 extern const uint8_t ff_alternate_horizontal_scan[64];

libavcodec/i386/dsputil_mmx.c

@@ -1603,8 +1603,13 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
         const int idct_algo= avctx->idct_algo;
 
 #ifdef CONFIG_ENCODERS
-        if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX)
+        if(dct_algo==FF_DCT_AUTO || dct_algo==FF_DCT_MMX){
+            if(mm_flags & MM_MMXEXT){
+                c->fdct = ff_fdct_mmx2;
+            }else{
                 c->fdct = ff_fdct_mmx;
+            }
+        }
 #endif //CONFIG_ENCODERS
 
         if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_SIMPLEMMX){

libavcodec/i386/fdct_mmx.c

@@ -210,14 +210,19 @@ static always_inline void fdct_col(const int16_t *in, int16_t *out, int offset)
     movq_r2m(mm3, *(out + offset + 7 * 8));
 }
 
-static always_inline void fdct_row(const int16_t *in, int16_t *out, const int16_t *table)
+static always_inline void fdct_row(const int16_t *in, int16_t *out, const int16_t *table, int mmx2)
 { 
+  if(mmx2){
+    pshufw_m2r(*(in + 4), mm5, 0x1B);
+    movq_m2r(*(in + 0), mm0);
+  }else{
     movd_m2r(*(in + 6), mm5);
     punpcklwd_m2r(*(in + 4), mm5);
     movq_r2r(mm5, mm2);
     psrlq_i2r(0x20, mm5);
     movq_m2r(*(in + 0), mm0);
     punpcklwd_r2r(mm2, mm5);
+  }
     movq_r2r(mm0, mm1);	
     paddsw_r2r(mm5, mm0);
     psubsw_r2r(mm5, mm1);
@@ -283,7 +288,30 @@ void ff_fdct_mmx(int16_t *block)
     table = tab_frw_01234567;
     out = block;
     for(i=8;i>0;i--) {
-        fdct_row(block1, out, table);
+        fdct_row(block1, out, table, 0);
+        block1 += 8;
+        table += 32;
+        out += 8;
+    }
+}
+
+void ff_fdct_mmx2(int16_t *block)
+{
+    int64_t align_tmp[16] ATTR_ALIGN(8);
+    int16_t * const block_tmp= (int16_t*)align_tmp;
+    int16_t *block1, *out;
+    const int16_t *table;
+    int i;
+
+    block1 = block_tmp;
+    fdct_col(block, block1, 0);
+    fdct_col(block, block1, 4);
+
+    block1 = block_tmp;
+    table = tab_frw_01234567;
+    out = block;
+    for(i=8;i>0;i--) {
+        fdct_row(block1, out, table, 1);
         block1 += 8;
         table += 32;
         out += 8;

libavcodec/i386/mpegvideo_mmx.c

@@ -490,11 +490,13 @@ static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w)
 
 #undef HAVE_MMX2
 #define RENAME(a) a ## _MMX
+#define RENAMEl(a) a ## _mmx
 #include "mpegvideo_mmx_template.c"
 
 #define HAVE_MMX2
 #undef RENAME
 #define RENAME(a) a ## _MMX2
+#define RENAMEl(a) a ## _mmx2
 #include "mpegvideo_mmx_template.c"
 
 void MPV_common_init_mmx(MpegEncContext *s)

libavcodec/i386/mpegvideo_mmx_template.c

@@ -43,7 +43,7 @@ static int RENAME(dct_quantize)(MpegEncContext *s,
     assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
 
     //s->fdct (block);
-    ff_fdct_mmx (block); //cant be anything else ...
+    RENAMEl(ff_fdct) (block); //cant be anything else ...
 
     if (s->mb_intra) {
         int dummy;