8x8 integer dct from x264 as cmp function (under CONFIG_GPL)

if this gives better quality then SATD then someone should port the x86 code too or maybe we could even just call it from libx264
the 4x4 one could be tried too ...


git-svn-id: file:///var/local/repositories/ffmpeg/trunk@4811 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b

libavcodec/avcodec.h

@@ -1402,6 +1402,7 @@ typedef struct AVCodecContext {
 #define FF_CMP_W53  11
 #define FF_CMP_W97  12
 #define FF_CMP_DCTMAX 13
+#define FF_CMP_DCT264 14
 #define FF_CMP_CHROMA 256
 
     /**

libavcodec/dsputil.c

@@ -3115,6 +3115,9 @@ void ff_set_cmp(DSPContext* c, me_cmp_func *cmp, int type){
         case FF_CMP_DCT:
             cmp[i]= c->dct_sad[i];
             break;
+        case FF_CMP_DCT264:
+            cmp[i]= c->dct264_sad[i];
+            break;
         case FF_CMP_DCTMAX:
             cmp[i]= c->dct_max[i];
             break;
@@ -3341,6 +3344,59 @@ static int dct_sad8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2
     return sum;
 }
 
+#ifdef CONFIG_GPL
+#define DCT8_1D {\
+    const int s07 = SRC(0) + SRC(7);\
+    const int s16 = SRC(1) + SRC(6);\
+    const int s25 = SRC(2) + SRC(5);\
+    const int s34 = SRC(3) + SRC(4);\
+    const int a0 = s07 + s34;\
+    const int a1 = s16 + s25;\
+    const int a2 = s07 - s34;\
+    const int a3 = s16 - s25;\
+    const int d07 = SRC(0) - SRC(7);\
+    const int d16 = SRC(1) - SRC(6);\
+    const int d25 = SRC(2) - SRC(5);\
+    const int d34 = SRC(3) - SRC(4);\
+    const int a4 = d16 + d25 + (d07 + (d07>>1));\
+    const int a5 = d07 - d34 - (d25 + (d25>>1));\
+    const int a6 = d07 + d34 - (d16 + (d16>>1));\
+    const int a7 = d16 - d25 + (d34 + (d34>>1));\
+    DST(0,  a0 + a1     ) ;\
+    DST(1,  a4 + (a7>>2)) ;\
+    DST(2,  a2 + (a3>>1)) ;\
+    DST(3,  a5 + (a6>>2)) ;\
+    DST(4,  a0 - a1     ) ;\
+    DST(5,  a6 - (a5>>2)) ;\
+    DST(6, (a2>>1) - a3 ) ;\
+    DST(7, (a4>>2) - a7 ) ;\
+}
+
+static int dct264_sad8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
+    MpegEncContext * const s= (MpegEncContext *)c;
+    int16_t dct[8][8];
+    int i;
+    int sum=0;
+
+    s->dsp.diff_pixels(dct, src1, src2, stride);
+
+#define SRC(x) dct[i][x]
+#define DST(x,v) dct[i][x]= v
+    for( i = 0; i < 8; i++ )
+        DCT8_1D
+#undef SRC
+#undef DST
+
+#define SRC(x) dct[x][i]
+#define DST(x,v) sum += ABS(v)
+    for( i = 0; i < 8; i++ )
+        DCT8_1D
+#undef SRC
+#undef DST
+    return sum;
+}
+#endif
+
 static int dct_max8x8_c(/*MpegEncContext*/ void *c, uint8_t *src1, uint8_t *src2, int stride, int h){
     MpegEncContext * const s= (MpegEncContext *)c;
     uint64_t __align8 aligned_temp[sizeof(DCTELEM)*64/8];
@@ -3587,6 +3643,7 @@ static int vsse16_c(/*MpegEncContext*/ void *c, uint8_t *s1, uint8_t *s2, int st
 WARPER8_16_SQ(hadamard8_diff8x8_c, hadamard8_diff16_c)
 WARPER8_16_SQ(hadamard8_intra8x8_c, hadamard8_intra16_c)
 WARPER8_16_SQ(dct_sad8x8_c, dct_sad16_c)
+WARPER8_16_SQ(dct264_sad8x8_c, dct264_sad16_c)
 WARPER8_16_SQ(dct_max8x8_c, dct_max16_c)
 WARPER8_16_SQ(quant_psnr8x8_c, quant_psnr16_c)
 WARPER8_16_SQ(rd8x8_c, rd16_c)
@@ -3870,6 +3927,7 @@ void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->hadamard8_diff[4]= hadamard8_intra16_c;
     SET_CMP_FUNC(dct_sad)
     SET_CMP_FUNC(dct_max)
+    SET_CMP_FUNC(dct264_sad)
     c->sad[0]= pix_abs16_c;
     c->sad[1]= pix_abs8_c;
     c->sse[0]= sse16_c;

libavcodec/dsputil.h

@@ -170,6 +170,7 @@ typedef struct DSPContext {
     me_cmp_func w53[5];
     me_cmp_func w97[5];
     me_cmp_func dct_max[5];
+    me_cmp_func dct264_sad[5];
 
     me_cmp_func me_pre_cmp[5];
     me_cmp_func me_cmp[5];
@@ -366,6 +367,7 @@ static inline int get_penalty_factor(int lambda, int lambda2, int type){
     case FF_CMP_W97:
         return (2*lambda)>>(FF_LAMBDA_SHIFT);
     case FF_CMP_SATD:
+    case FF_CMP_DCT264:
         return (2*lambda)>>FF_LAMBDA_SHIFT;
     case FF_CMP_RD:
     case FF_CMP_PSNR: