4MV motion estimation (not finished yet)

SAD functions rewritten (8x8 support & MMX2 optimizations)
HQ inter/intra decission
msmpeg4 encoding bugfix (MV where too long)

Originally committed as revision 362 to svn://svn.ffmpeg.org/ffmpeg/trunk
This commit is contained in:
Michael Niedermayer 2002-03-27 21:25:22 +00:00
parent 52075cdeaa
commit ba6802debf
9 changed files with 1078 additions and 499 deletions

View File

@ -61,9 +61,14 @@ extern int motion_estimation_method;
#define ME_X1 5
/* encoding support */
/* note not everything is supported yet */
#define CODEC_FLAG_HQ 0x0001 /* high quality (non real time) encoding */
#define CODEC_FLAG_QSCALE 0x0002 /* use fixed qscale */
#define CODEC_FLAG_4MV 0x0004 /* 4 MV per MB allowed */
#define CODEC_FLAG_B 0x0008 /* use B frames */
#define CODEC_FLAG_QPEL 0x0010 /* use qpel MC */
#define CODEC_FLAG_GMC 0x0020 /* use GMC */
/* codec capabilities */

View File

@ -36,6 +36,11 @@ op_pixels_abs_func pix_abs16x16_x2;
op_pixels_abs_func pix_abs16x16_y2;
op_pixels_abs_func pix_abs16x16_xy2;
op_pixels_abs_func pix_abs8x8;
op_pixels_abs_func pix_abs8x8_x2;
op_pixels_abs_func pix_abs8x8_y2;
op_pixels_abs_func pix_abs8x8_xy2;
UINT8 cropTbl[256 + 2 * MAX_NEG_CROP];
UINT32 squareTbl[512];
@ -377,14 +382,14 @@ static void qpel_h_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride,
int i;
for(i=0; i<h; i++)
{
dst[0]= cm[(((src[0]+src[1])*160 - (src[0]+src[2])*48 + (src[1]+src[3])*24 - (src[2]+src[4])*8 + r)>>8)];
dst[1]= cm[(((src[1]+src[2])*160 - (src[0]+src[3])*48 + (src[0]+src[4])*24 - (src[1]+src[5])*8 + r)>>8)];
dst[2]= cm[(((src[2]+src[3])*160 - (src[1]+src[4])*48 + (src[0]+src[5])*24 - (src[0]+src[6])*8 + r)>>8)];
dst[3]= cm[(((src[3]+src[4])*160 - (src[2]+src[5])*48 + (src[1]+src[6])*24 - (src[0]+src[7])*8 + r)>>8)];
dst[4]= cm[(((src[4]+src[5])*160 - (src[3]+src[6])*48 + (src[2]+src[7])*24 - (src[1]+src[8])*8 + r)>>8)];
dst[5]= cm[(((src[5]+src[6])*160 - (src[4]+src[7])*48 + (src[3]+src[8])*24 - (src[2]+src[8])*8 + r)>>8)];
dst[6]= cm[(((src[6]+src[7])*160 - (src[5]+src[8])*48 + (src[4]+src[8])*24 - (src[3]+src[7])*8 + r)>>8)];
dst[7]= cm[(((src[7]+src[8])*160 - (src[6]+src[8])*48 + (src[5]+src[7])*24 - (src[4]+src[6])*8 + r)>>8)];
dst[0]= cm[(((src[0]+src[1])*20 - (src[0]+src[2])*6 + (src[1]+src[3])*3 - (src[2]+src[4]) + r)>>5)];
dst[1]= cm[(((src[1]+src[2])*20 - (src[0]+src[3])*6 + (src[0]+src[4])*3 - (src[1]+src[5]) + r)>>5)];
dst[2]= cm[(((src[2]+src[3])*20 - (src[1]+src[4])*6 + (src[0]+src[5])*3 - (src[0]+src[6]) + r)>>5)];
dst[3]= cm[(((src[3]+src[4])*20 - (src[2]+src[5])*6 + (src[1]+src[6])*3 - (src[0]+src[7]) + r)>>5)];
dst[4]= cm[(((src[4]+src[5])*20 - (src[3]+src[6])*6 + (src[2]+src[7])*3 - (src[1]+src[8]) + r)>>5)];
dst[5]= cm[(((src[5]+src[6])*20 - (src[4]+src[7])*6 + (src[3]+src[8])*3 - (src[2]+src[8]) + r)>>5)];
dst[6]= cm[(((src[6]+src[7])*20 - (src[5]+src[8])*6 + (src[4]+src[8])*3 - (src[3]+src[7]) + r)>>5)];
dst[7]= cm[(((src[7]+src[8])*20 - (src[6]+src[8])*6 + (src[5]+src[7])*3 - (src[4]+src[6]) + r)>>5)];
dst+=dstStride;
src+=srcStride;
}
@ -405,14 +410,14 @@ static void qpel_v_lowpass(UINT8 *dst, UINT8 *src, int dstStride, int srcStride,
const int src6= src[6*srcStride];
const int src7= src[7*srcStride];
const int src8= src[8*srcStride];
dst[0*dstStride]= cm[(((src0+src1)*160 - (src0+src2)*48 + (src1+src3)*24 - (src2+src4)*8 + r)>>8)];
dst[1*dstStride]= cm[(((src1+src2)*160 - (src0+src3)*48 + (src0+src4)*24 - (src1+src5)*8 + r)>>8)];
dst[2*dstStride]= cm[(((src2+src3)*160 - (src1+src4)*48 + (src0+src5)*24 - (src0+src6)*8 + r)>>8)];
dst[3*dstStride]= cm[(((src3+src4)*160 - (src2+src5)*48 + (src1+src6)*24 - (src0+src7)*8 + r)>>8)];
dst[4*dstStride]= cm[(((src4+src5)*160 - (src3+src6)*48 + (src2+src7)*24 - (src1+src8)*8 + r)>>8)];
dst[5*dstStride]= cm[(((src5+src6)*160 - (src4+src7)*48 + (src3+src8)*24 - (src2+src8)*8 + r)>>8)];
dst[6*dstStride]= cm[(((src6+src7)*160 - (src5+src8)*48 + (src4+src8)*24 - (src3+src7)*8 + r)>>8)];
dst[7*dstStride]= cm[(((src7+src8)*160 - (src6+src8)*48 + (src5+src7)*24 - (src4+src6)*8 + r)>>8)];
dst[0*dstStride]= cm[(((src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4) + r)>>5)];
dst[1*dstStride]= cm[(((src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5) + r)>>5)];
dst[2*dstStride]= cm[(((src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6) + r)>>5)];
dst[3*dstStride]= cm[(((src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7) + r)>>5)];
dst[4*dstStride]= cm[(((src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8) + r)>>5)];
dst[5*dstStride]= cm[(((src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8) + r)>>5)];
dst[6*dstStride]= cm[(((src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7) + r)>>5)];
dst[7*dstStride]= cm[(((src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6) + r)>>5)];
dst++;
src++;
}
@ -485,38 +490,38 @@ static void qpel_mc00_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcS
static void qpel_mc10_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
{\
UINT8 half[64];\
qpel_h_lowpass(half, src, 8, srcStride, 8, 128-r);\
qpel_h_lowpass(half, src, 8, srcStride, 8, 16-r);\
avg2_block(dst, src, half, dstStride, srcStride, 1-r);\
}\
\
static void qpel_mc20_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
{\
qpel_h_lowpass(dst, src, dstStride, srcStride, 8, 128-r);\
qpel_h_lowpass(dst, src, dstStride, srcStride, 8, 16-r);\
}\
\
static void qpel_mc30_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
{\
UINT8 half[64];\
qpel_h_lowpass(half, src, 8, srcStride, 8, 128-r);\
qpel_h_lowpass(half, src, 8, srcStride, 8, 16-r);\
avg2_block(dst, src+1, half, dstStride, srcStride, 1-r);\
}\
\
static void qpel_mc01_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
{\
UINT8 half[64];\
qpel_v_lowpass(half, src, 8, srcStride, 8, 128-r);\
qpel_v_lowpass(half, src, 8, srcStride, 8, 16-r);\
avg2_block(dst, src, half, dstStride, srcStride, 1-r);\
}\
\
static void qpel_mc02_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
{\
qpel_v_lowpass(dst, src, dstStride, srcStride, 8, 128-r);\
qpel_v_lowpass(dst, src, dstStride, srcStride, 8, 16-r);\
}\
\
static void qpel_mc03_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
{\
UINT8 half[64];\
qpel_v_lowpass(half, src, 8, srcStride, 8, 128-r);\
qpel_v_lowpass(half, src, 8, srcStride, 8, 16-r);\
avg2_block(dst, src+srcStride, half, dstStride, srcStride, 1-r);\
}\
static void qpel_mc11_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
@ -524,9 +529,9 @@ static void qpel_mc11_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcS
UINT8 halfH[72];\
UINT8 halfV[64];\
UINT8 halfHV[64];\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 128-r);\
qpel_v_lowpass(halfV, src, 8, srcStride, 8, 128-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 128-r);\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
qpel_v_lowpass(halfV, src, 8, srcStride, 8, 16-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
avg4_block(dst, src, halfH, halfV, halfHV, dstStride, srcStride, 2-r);\
}\
static void qpel_mc31_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
@ -534,9 +539,9 @@ static void qpel_mc31_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcS
UINT8 halfH[72];\
UINT8 halfV[64];\
UINT8 halfHV[64];\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 128-r);\
qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 128-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 128-r);\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 16-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
avg4_block(dst, src+1, halfH, halfV, halfHV, dstStride, srcStride, 2-r);\
}\
static void qpel_mc13_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
@ -544,9 +549,9 @@ static void qpel_mc13_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcS
UINT8 halfH[72];\
UINT8 halfV[64];\
UINT8 halfHV[64];\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 128-r);\
qpel_v_lowpass(halfV, src, 8, srcStride, 8, 128-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 128-r);\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
qpel_v_lowpass(halfV, src, 8, srcStride, 8, 16-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
avg4_block(dst, src+srcStride, halfH+8, halfV, halfHV, dstStride, srcStride, 2-r);\
}\
static void qpel_mc33_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
@ -554,25 +559,25 @@ static void qpel_mc33_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcS
UINT8 halfH[72];\
UINT8 halfV[64];\
UINT8 halfHV[64];\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 128-r);\
qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 128-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 128-r);\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 16-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
avg4_block(dst, src+srcStride+1, halfH+8, halfV, halfHV, dstStride, srcStride, 2-r);\
}\
static void qpel_mc21_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
{\
UINT8 halfH[72];\
UINT8 halfHV[64];\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 128-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 128-r);\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
avg2_block(dst, halfH, halfHV, dstStride, 8, 1-r);\
}\
static void qpel_mc23_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
{\
UINT8 halfH[72];\
UINT8 halfHV[64];\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 128-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 128-r);\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
avg2_block(dst, halfH+8, halfHV, dstStride, 8, 1-r);\
}\
static void qpel_mc12_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
@ -580,9 +585,9 @@ static void qpel_mc12_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcS
UINT8 halfH[72];\
UINT8 halfV[64];\
UINT8 halfHV[64];\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 128-r);\
qpel_v_lowpass(halfV, src, 8, srcStride, 8, 128-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 128-r);\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
qpel_v_lowpass(halfV, src, 8, srcStride, 8, 16-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
avg2_block(dst, halfV, halfHV, dstStride, 8, 1-r);\
}\
static void qpel_mc32_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
@ -590,16 +595,16 @@ static void qpel_mc32_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcS
UINT8 halfH[72];\
UINT8 halfV[64];\
UINT8 halfHV[64];\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 128-r);\
qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 128-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 128-r);\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
qpel_v_lowpass(halfV, src+1, 8, srcStride, 8, 16-r);\
qpel_v_lowpass(halfHV, halfH, 8, 8, 8, 16-r);\
avg2_block(dst, halfV, halfHV, dstStride, 8, 1-r);\
}\
static void qpel_mc22_c ## name (UINT8 *dst, UINT8 *src, int dstStride, int srcStride, int mx, int my)\
{\
UINT8 halfH[72];\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 128-r);\
qpel_v_lowpass(dst, halfH, dstStride, 8, 8, 128-r);\
qpel_h_lowpass(halfH, src, 8, srcStride, 9, 16-r);\
qpel_v_lowpass(dst, halfH, dstStride, 8, 8, 16-r);\
}\
qpel_mc_func qpel_mc ## name ## _tab[16]={ \
qpel_mc00_c ## name, \
@ -623,12 +628,12 @@ qpel_mc_func qpel_mc ## name ## _tab[16]={ \
QPEL_MC(0, _rnd)
QPEL_MC(1, _no_rnd)
int pix_abs16x16_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h)
int pix_abs16x16_c(UINT8 *pix1, UINT8 *pix2, int line_size)
{
int s, i;
s = 0;
for(i=0;i<h;i++) {
for(i=0;i<16;i++) {
s += abs(pix1[0] - pix2[0]);
s += abs(pix1[1] - pix2[1]);
s += abs(pix1[2] - pix2[2]);
@ -651,12 +656,12 @@ int pix_abs16x16_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h)
return s;
}
int pix_abs16x16_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h)
int pix_abs16x16_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
{
int s, i;
s = 0;
for(i=0;i<h;i++) {
for(i=0;i<16;i++) {
s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
@ -679,13 +684,13 @@ int pix_abs16x16_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h)
return s;
}
int pix_abs16x16_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h)
int pix_abs16x16_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
{
int s, i;
UINT8 *pix3 = pix2 + line_size;
s = 0;
for(i=0;i<h;i++) {
for(i=0;i<16;i++) {
s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
@ -709,13 +714,13 @@ int pix_abs16x16_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h)
return s;
}
int pix_abs16x16_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h)
int pix_abs16x16_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
{
int s, i;
UINT8 *pix3 = pix2 + line_size;
s = 0;
for(i=0;i<h;i++) {
for(i=0;i<16;i++) {
s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1]));
s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2]));
s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3]));
@ -739,6 +744,90 @@ int pix_abs16x16_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size, int h)
return s;
}
int pix_abs8x8_c(UINT8 *pix1, UINT8 *pix2, int line_size)
{
int s, i;
s = 0;
for(i=0;i<8;i++) {
s += abs(pix1[0] - pix2[0]);
s += abs(pix1[1] - pix2[1]);
s += abs(pix1[2] - pix2[2]);
s += abs(pix1[3] - pix2[3]);
s += abs(pix1[4] - pix2[4]);
s += abs(pix1[5] - pix2[5]);
s += abs(pix1[6] - pix2[6]);
s += abs(pix1[7] - pix2[7]);
pix1 += line_size;
pix2 += line_size;
}
return s;
}
int pix_abs8x8_x2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
{
int s, i;
s = 0;
for(i=0;i<8;i++) {
s += abs(pix1[0] - avg2(pix2[0], pix2[1]));
s += abs(pix1[1] - avg2(pix2[1], pix2[2]));
s += abs(pix1[2] - avg2(pix2[2], pix2[3]));
s += abs(pix1[3] - avg2(pix2[3], pix2[4]));
s += abs(pix1[4] - avg2(pix2[4], pix2[5]));
s += abs(pix1[5] - avg2(pix2[5], pix2[6]));
s += abs(pix1[6] - avg2(pix2[6], pix2[7]));
s += abs(pix1[7] - avg2(pix2[7], pix2[8]));
pix1 += line_size;
pix2 += line_size;
}
return s;
}
int pix_abs8x8_y2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
{
int s, i;
UINT8 *pix3 = pix2 + line_size;
s = 0;
for(i=0;i<8;i++) {
s += abs(pix1[0] - avg2(pix2[0], pix3[0]));
s += abs(pix1[1] - avg2(pix2[1], pix3[1]));
s += abs(pix1[2] - avg2(pix2[2], pix3[2]));
s += abs(pix1[3] - avg2(pix2[3], pix3[3]));
s += abs(pix1[4] - avg2(pix2[4], pix3[4]));
s += abs(pix1[5] - avg2(pix2[5], pix3[5]));
s += abs(pix1[6] - avg2(pix2[6], pix3[6]));
s += abs(pix1[7] - avg2(pix2[7], pix3[7]));
pix1 += line_size;
pix2 += line_size;
pix3 += line_size;
}
return s;
}
int pix_abs8x8_xy2_c(UINT8 *pix1, UINT8 *pix2, int line_size)
{
int s, i;
UINT8 *pix3 = pix2 + line_size;
s = 0;
for(i=0;i<8;i++) {
s += abs(pix1[0] - avg4(pix2[0], pix2[1], pix3[0], pix3[1]));
s += abs(pix1[1] - avg4(pix2[1], pix2[2], pix3[1], pix3[2]));
s += abs(pix1[2] - avg4(pix2[2], pix2[3], pix3[2], pix3[3]));
s += abs(pix1[3] - avg4(pix2[3], pix2[4], pix3[3], pix3[4]));
s += abs(pix1[4] - avg4(pix2[4], pix2[5], pix3[4], pix3[5]));
s += abs(pix1[5] - avg4(pix2[5], pix2[6], pix3[5], pix3[6]));
s += abs(pix1[6] - avg4(pix2[6], pix2[7], pix3[6], pix3[7]));
s += abs(pix1[7] - avg4(pix2[7], pix2[8], pix3[7], pix3[8]));
pix1 += line_size;
pix2 += line_size;
pix3 += line_size;
}
return s;
}
/* permute block according so that it corresponds to the MMX idct
order */
#ifdef SIMPLE_IDCT
@ -806,6 +895,10 @@ void dsputil_init(void)
pix_abs16x16_x2 = pix_abs16x16_x2_c;
pix_abs16x16_y2 = pix_abs16x16_y2_c;
pix_abs16x16_xy2 = pix_abs16x16_xy2_c;
pix_abs8x8 = pix_abs8x8_c;
pix_abs8x8_x2 = pix_abs8x8_x2_c;
pix_abs8x8_y2 = pix_abs8x8_y2_c;
pix_abs8x8_xy2 = pix_abs8x8_xy2_c;
av_fdct = jpeg_fdct_ifast;
use_permuted_idct = 1;

View File

@ -66,17 +66,21 @@ extern void (*sub_pixels_tab[4])(DCTELEM *block, const UINT8 *pixels, int line_s
/* motion estimation */
typedef int (*op_pixels_abs_func)(UINT8 *blk1, UINT8 *blk2, int line_size, int h);
typedef int (*op_pixels_abs_func)(UINT8 *blk1, UINT8 *blk2, int line_size);
extern op_pixels_abs_func pix_abs16x16;
extern op_pixels_abs_func pix_abs16x16_x2;
extern op_pixels_abs_func pix_abs16x16_y2;
extern op_pixels_abs_func pix_abs16x16_xy2;
extern op_pixels_abs_func pix_abs8x8;
extern op_pixels_abs_func pix_abs8x8_x2;
extern op_pixels_abs_func pix_abs8x8_y2;
extern op_pixels_abs_func pix_abs8x8_xy2;
int pix_abs16x16_c(UINT8 *blk1, UINT8 *blk2, int lx, int h);
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);
int pix_abs16x16_c(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_x2_c(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_y2_c(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_xy2_c(UINT8 *blk1, UINT8 *blk2, int lx);
static inline int block_permute_op(int j)
{

View File

@ -469,16 +469,10 @@ void h263_encode_mb(MpegEncContext * s,
}
/* encode each block */
if (s->h263_pred) {
for (i = 0; i < 6; i++) {
// mpeg4_encode_block(s, block[i], i);
}
} else {
for (i = 0; i < 6; i++) {
h263_encode_block(s, block[i], i);
}
}
}
void h263_pred_acdc(MpegEncContext * s, INT16 *block, int n)
{
@ -779,7 +773,7 @@ void h263_encode_init(MpegEncContext *s)
// use fcodes >1 only for mpeg4 & h263 & h263p FIXME
if(s->h263_plus) s->fcode_tab= umv_fcode_tab;
else if(s->h263_pred) s->fcode_tab= fcode_tab;
else if(s->h263_pred && !s->h263_msmpeg4) s->fcode_tab= fcode_tab;
}
static void h263_encode_block(MpegEncContext * s, DCTELEM * block, int n)

View File

@ -24,19 +24,34 @@
int mm_flags; /* multimedia extension flags */
int pix_abs16x16_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h);
int pix_abs16x16_sse(UINT8 *blk1, UINT8 *blk2, int lx, int h);
int pix_abs16x16_x2_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h);
int pix_abs16x16_y2_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h);
int pix_abs16x16_xy2_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h);
int pix_abs16x16_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_x2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_y2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_xy2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_mmx2(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_x2_mmx2(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_y2_mmx2(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs16x16_xy2_mmx2(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs8x8_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs8x8_x2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs8x8_y2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs8x8_xy2_mmx(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs8x8_mmx2(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs8x8_x2_mmx2(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs8x8_y2_mmx2(UINT8 *blk1, UINT8 *blk2, int lx);
int pix_abs8x8_xy2_mmx2(UINT8 *blk1, UINT8 *blk2, int lx);
/* external functions, from idct_mmx.c */
void ff_mmx_idct(DCTELEM *block);
void ff_mmxext_idct(DCTELEM *block);
/* pixel operations */
static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001;
static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002;
static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001LL;
static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002LL;
//static const unsigned short mm_wone[4] __attribute__ ((aligned(8))) = { 0x1, 0x1, 0x1, 0x1 };
//static const unsigned short mm_wtwo[4] __attribute__ ((aligned(8))) = { 0x2, 0x2, 0x2, 0x2 };
@ -1039,6 +1054,10 @@ void dsputil_init_mmx(void)
pix_abs16x16_x2 = pix_abs16x16_x2_mmx;
pix_abs16x16_y2 = pix_abs16x16_y2_mmx;
pix_abs16x16_xy2 = pix_abs16x16_xy2_mmx;
pix_abs8x8 = pix_abs8x8_mmx;
pix_abs8x8_x2 = pix_abs8x8_x2_mmx;
pix_abs8x8_y2 = pix_abs8x8_y2_mmx;
pix_abs8x8_xy2= pix_abs8x8_xy2_mmx;
av_fdct = fdct_mmx;
put_pixels_tab[0] = put_pixels_mmx;
@ -1067,10 +1086,16 @@ void dsputil_init_mmx(void)
sub_pixels_tab[3] = sub_pixels_xy2_mmx;
if (mm_flags & MM_MMXEXT) {
pix_abs16x16 = pix_abs16x16_sse;
}
pix_abs16x16 = pix_abs16x16_mmx2;
pix_abs16x16_x2 = pix_abs16x16_x2_mmx2;
pix_abs16x16_y2 = pix_abs16x16_y2_mmx2;
pix_abs16x16_xy2= pix_abs16x16_xy2_mmx2;
pix_abs8x8 = pix_abs8x8_mmx2;
pix_abs8x8_x2 = pix_abs8x8_x2_mmx2;
pix_abs8x8_y2 = pix_abs8x8_y2_mmx2;
pix_abs8x8_xy2= pix_abs8x8_xy2_mmx2;
if (mm_flags & MM_SSE) {
put_pixels_tab[1] = put_pixels_x2_sse;
put_pixels_tab[2] = put_pixels_y2_sse;

View File

@ -16,229 +16,347 @@
* along with this program; if not, write to the Free Software
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
*
* mostly by Michael Niedermayer <michaelni@gmx.at>
*/
#include "../dsputil.h"
#include "mmx.h"
static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001;
static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002;
static const __attribute__ ((aligned(8))) UINT64 round_tab[3]={
0x0000000000000000,
0x0001000100010001,
0x0002000200020002,
};
/* mm7 is accumulator, mm6 is zero */
static inline void sad_add(const UINT8 *p1, const UINT8 *p2)
static inline void sad8_mmx(UINT8 *blk1, UINT8 *blk2, int stride, int h)
{
movq_m2r(*p1, mm0);
movq_m2r(*p2, mm1);
movq_r2r(mm0, mm2);
psubusb_r2r(mm1, mm0);
psubusb_r2r(mm2, mm1);
por_r2r(mm1, mm0); /* mm0 is absolute value */
movq_r2r(mm0, mm1);
punpcklbw_r2r(mm6, mm0);
punpckhbw_r2r(mm6, mm1);
paddusw_r2r(mm0, mm7);
paddusw_r2r(mm1, mm7);
int len= -(stride<<h);
asm volatile(
".balign 16 \n\t"
"1: \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq (%2, %%eax), %%mm2 \n\t"
"movq (%2, %%eax), %%mm4 \n\t"
"addl %3, %%eax \n\t"
"psubusb %%mm0, %%mm2 \n\t"
"psubusb %%mm4, %%mm0 \n\t"
"movq (%1, %%eax), %%mm1 \n\t"
"movq (%2, %%eax), %%mm3 \n\t"
"movq (%2, %%eax), %%mm5 \n\t"
"psubusb %%mm1, %%mm3 \n\t"
"psubusb %%mm5, %%mm1 \n\t"
"por %%mm2, %%mm0 \n\t"
"por %%mm1, %%mm3 \n\t"
"movq %%mm0, %%mm1 \n\t"
"movq %%mm3, %%mm2 \n\t"
"punpcklbw %%mm7, %%mm0 \n\t"
"punpckhbw %%mm7, %%mm1 \n\t"
"punpcklbw %%mm7, %%mm3 \n\t"
"punpckhbw %%mm7, %%mm2 \n\t"
"paddw %%mm1, %%mm0 \n\t"
"paddw %%mm3, %%mm2 \n\t"
"paddw %%mm2, %%mm0 \n\t"
"paddw %%mm0, %%mm6 \n\t"
"addl %3, %%eax \n\t"
" js 1b \n\t"
: "+a" (len)
: "r" (blk1 - len), "r" (blk2 - len), "r" (stride)
);
}
/* convert mm7 to value */
static inline int sad_end(void)
static inline void sad8_mmx2(UINT8 *blk1, UINT8 *blk2, int stride, int h)
{
int res;
movq_r2r(mm7, mm0);
psrlq_i2r(32, mm7);
paddusw_r2r(mm0, mm7);
movq_r2r(mm7, mm0);
psrlq_i2r(16, mm7);
paddusw_r2r(mm0, mm7);
__asm __volatile ("movd %%mm7, %0" : "=a" (res));
return res & 0xffff;
int len= -(stride<<h);
asm volatile(
".balign 16 \n\t"
"1: \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq (%2, %%eax), %%mm2 \n\t"
"psadbw %%mm2, %%mm0 \n\t"
"addl %3, %%eax \n\t"
"movq (%1, %%eax), %%mm1 \n\t"
"movq (%2, %%eax), %%mm3 \n\t"
"psadbw %%mm1, %%mm3 \n\t"
"paddw %%mm3, %%mm0 \n\t"
"paddw %%mm0, %%mm6 \n\t"
"addl %3, %%eax \n\t"
" js 1b \n\t"
: "+a" (len)
: "r" (blk1 - len), "r" (blk2 - len), "r" (stride)
);
}
int pix_abs16x16_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h)
static inline void sad8_2_mmx2(UINT8 *blk1a, UINT8 *blk1b, UINT8 *blk2, int stride, int h)
{
const UINT8 *p1, *p2;
h >>= 1;
p1 = blk1;
p2 = blk2;
pxor_r2r(mm7, mm7); /* mm7 is accumulator */
pxor_r2r(mm6, mm6); /* mm7 is zero constant */
do {
sad_add(p1, p2);
sad_add(p1 + 8, p2 + 8);
p1 += lx;
p2 += lx;
sad_add(p1, p2);
sad_add(p1 + 8, p2 + 8);
p1 += lx;
p2 += lx;
} while (--h);
return sad_end();
int len= -(stride<<h);
asm volatile(
".balign 16 \n\t"
"1: \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq (%2, %%eax), %%mm2 \n\t"
"pavgb %%mm2, %%mm0 \n\t"
"movq (%3, %%eax), %%mm2 \n\t"
"psadbw %%mm2, %%mm0 \n\t"
"addl %4, %%eax \n\t"
"movq (%1, %%eax), %%mm1 \n\t"
"movq (%2, %%eax), %%mm3 \n\t"
"pavgb %%mm1, %%mm3 \n\t"
"movq (%3, %%eax), %%mm1 \n\t"
"psadbw %%mm1, %%mm3 \n\t"
"paddw %%mm3, %%mm0 \n\t"
"paddw %%mm0, %%mm6 \n\t"
"addl %4, %%eax \n\t"
" js 1b \n\t"
: "+a" (len)
: "r" (blk1a - len), "r" (blk1b -len), "r" (blk2 - len), "r" (stride)
);
}
/* please test it ! */
static inline void sad_add_sse(const UINT8 *p1, const UINT8 *p2)
{
movq_m2r(*(p1 + 0), mm0);
movq_m2r(*(p1 + 8), mm1);
psadbw_m2r(*(p2 + 0), mm0);
psadbw_m2r(*(p2 + 8), mm1);
paddusw_r2r(mm0, mm7);
paddusw_r2r(mm1, mm7);
static inline void sad8_4_mmx2(UINT8 *blk1, UINT8 *blk2, int stride, int h)
{ //FIXME reuse src
int len= -(stride<<h);
asm volatile(
".balign 16 \n\t"
"1: \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq (%2, %%eax), %%mm2 \n\t"
"movq 1(%1, %%eax), %%mm1 \n\t"
"movq 1(%2, %%eax), %%mm3 \n\t"
"pavgb %%mm2, %%mm0 \n\t"
"pavgb %%mm1, %%mm3 \n\t"
"pavgb %%mm3, %%mm0 \n\t"
"movq (%3, %%eax), %%mm2 \n\t"
"psadbw %%mm2, %%mm0 \n\t"
"addl %4, %%eax \n\t"
"movq (%1, %%eax), %%mm1 \n\t"
"movq (%2, %%eax), %%mm3 \n\t"
"movq 1(%1, %%eax), %%mm2 \n\t"
"movq 1(%2, %%eax), %%mm4 \n\t"
"pavgb %%mm3, %%mm1 \n\t"
"pavgb %%mm4, %%mm2 \n\t"
"pavgb %%mm1, %%mm2 \n\t"
"movq (%3, %%eax), %%mm1 \n\t"
"psadbw %%mm1, %%mm2 \n\t"
"paddw %%mm2, %%mm0 \n\t"
"paddw %%mm0, %%mm6 \n\t"
"addl %4, %%eax \n\t"
" js 1b \n\t"
: "+a" (len)
: "r" (blk1 - len), "r" (blk1 - len + stride), "r" (blk2 - len), "r" (stride)
);
}
int pix_abs16x16_sse(UINT8 *blk1, UINT8 *blk2, int lx, int h)
static inline void sad8_2_mmx(UINT8 *blk1a, UINT8 *blk1b, UINT8 *blk2, int stride, int h)
{
const UINT8 *p1, *p2;
h >>= 1;
p1 = blk1;
p2 = blk2;
pxor_r2r(mm7, mm7); /* mm7 is accumulator */
do {
sad_add_sse(p1, p2);
p1 += lx;
p2 += lx;
sad_add_sse(p1, p2);
p1 += lx;
p2 += lx;
} while (--h);
return sad_end();
int len= -(stride<<h);
asm volatile(
".balign 16 \n\t"
"1: \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq (%2, %%eax), %%mm1 \n\t"
"movq (%1, %%eax), %%mm2 \n\t"
"movq (%2, %%eax), %%mm3 \n\t"
"punpcklbw %%mm7, %%mm0 \n\t"
"punpcklbw %%mm7, %%mm1 \n\t"
"punpckhbw %%mm7, %%mm2 \n\t"
"punpckhbw %%mm7, %%mm3 \n\t"
"paddw %%mm0, %%mm1 \n\t"
"paddw %%mm2, %%mm3 \n\t"
"movq (%3, %%eax), %%mm4 \n\t"
"movq (%3, %%eax), %%mm2 \n\t"
"paddw %%mm5, %%mm1 \n\t"
"paddw %%mm5, %%mm3 \n\t"
"psrlw $1, %%mm1 \n\t"
"psrlw $1, %%mm3 \n\t"
"packuswb %%mm3, %%mm1 \n\t"
"psubusb %%mm1, %%mm4 \n\t"
"psubusb %%mm2, %%mm1 \n\t"
"por %%mm4, %%mm1 \n\t"
"movq %%mm1, %%mm0 \n\t"
"punpcklbw %%mm7, %%mm0 \n\t"
"punpckhbw %%mm7, %%mm1 \n\t"
"paddw %%mm1, %%mm0 \n\t"
"paddw %%mm0, %%mm6 \n\t"
"addl %4, %%eax \n\t"
" js 1b \n\t"
: "+a" (len)
: "r" (blk1a - len), "r" (blk1b -len), "r" (blk2 - len), "r" (stride)
);
}
#define DUMP(reg) { mmx_t tmp; movq_r2m(reg, tmp); printf(#reg "=%016Lx\n", tmp.uq); }
/* mm7 is accumulator, mm6 is zero */
static inline void sad_add_x2(const UINT8 *p1, const UINT8 *p2, const UINT8 *p3)
static inline void sad8_4_mmx(UINT8 *blk1, UINT8 *blk2, int stride, int h)
{
movq_m2r(*(p2 + 0), mm0);
movq_m2r(*(p3 + 0), mm1);
movq_r2r(mm0, mm2);
movq_r2r(mm1, mm3);
punpcklbw_r2r(mm6, mm0); /* extract 4 bytes low */
punpcklbw_r2r(mm6, mm1);
punpckhbw_r2r(mm6, mm2); /* high */
punpckhbw_r2r(mm6, mm3);
paddusw_r2r(mm1, mm0);
paddusw_r2r(mm3, mm2);
movq_m2r(*(p1 + 0), mm1); /* mm1 : other value */
paddusw_r2r(mm5, mm0); /* + 1 */
paddusw_r2r(mm5, mm2); /* + 1 */
psrlw_i2r(1, mm0);
psrlw_i2r(1, mm2);
packuswb_r2r(mm2, mm0); /* average is in mm0 */
movq_r2r(mm1, mm2);
psubusb_r2r(mm0, mm1);
psubusb_r2r(mm2, mm0);
por_r2r(mm1, mm0); /* mm0 is absolute value */
movq_r2r(mm0, mm1);
punpcklbw_r2r(mm6, mm0);
punpckhbw_r2r(mm6, mm1);
paddusw_r2r(mm0, mm7);
paddusw_r2r(mm1, mm7);
int len= -(stride<<h);
asm volatile(
".balign 16 \n\t"
"1: \n\t"
"movq (%1, %%eax), %%mm0 \n\t"
"movq (%2, %%eax), %%mm1 \n\t"
"movq %%mm0, %%mm4 \n\t"
"movq %%mm1, %%mm2 \n\t"
"punpcklbw %%mm7, %%mm0 \n\t"
"punpcklbw %%mm7, %%mm1 \n\t"
"punpckhbw %%mm7, %%mm4 \n\t"
"punpckhbw %%mm7, %%mm2 \n\t"
"paddw %%mm1, %%mm0 \n\t"
"paddw %%mm2, %%mm4 \n\t"
"movq 1(%1, %%eax), %%mm2 \n\t"
"movq 1(%2, %%eax), %%mm3 \n\t"
"movq %%mm2, %%mm1 \n\t"
"punpcklbw %%mm7, %%mm2 \n\t"
"punpckhbw %%mm7, %%mm1 \n\t"
"paddw %%mm0, %%mm2 \n\t"
"paddw %%mm4, %%mm1 \n\t"
"movq %%mm3, %%mm4 \n\t"
"punpcklbw %%mm7, %%mm3 \n\t"
"punpckhbw %%mm7, %%mm4 \n\t"
"paddw %%mm3, %%mm2 \n\t"
"paddw %%mm4, %%mm1 \n\t"
"movq (%3, %%eax), %%mm3 \n\t"
"movq (%3, %%eax), %%mm4 \n\t"
"paddw %%mm5, %%mm2 \n\t"
"paddw %%mm5, %%mm1 \n\t"
"psrlw $2, %%mm2 \n\t"
"psrlw $2, %%mm1 \n\t"
"packuswb %%mm1, %%mm2 \n\t"
"psubusb %%mm2, %%mm3 \n\t"
"psubusb %%mm4, %%mm2 \n\t"
"por %%mm3, %%mm2 \n\t"
"movq %%mm2, %%mm0 \n\t"
"punpcklbw %%mm7, %%mm0 \n\t"
"punpckhbw %%mm7, %%mm2 \n\t"
"paddw %%mm2, %%mm0 \n\t"
"paddw %%mm0, %%mm6 \n\t"
"addl %4, %%eax \n\t"
" js 1b \n\t"
: "+a" (len)
: "r" (blk1 - len), "r" (blk1 -len + stride), "r" (blk2 - len), "r" (stride)
);
}
int pix_abs16x16_x2_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h)
static inline int sum_mmx()
{
const UINT8 *p1, *p2;
p1 = blk1;
p2 = blk2;
pxor_r2r(mm7, mm7); /* mm7 is accumulator */
pxor_r2r(mm6, mm6); /* mm7 is zero constant */
movq_m2r(mm_wone, mm5); /* one constant */
do {
sad_add_x2(p1, p2, p2 + 1);
sad_add_x2(p1 + 8, p2 + 8, p2 + 9);
p1 += lx;
p2 += lx;
} while (--h);
return sad_end();
int ret;
asm volatile(
"movq %%mm6, %%mm0 \n\t"
"psrlq $32, %%mm6 \n\t"
"paddw %%mm0, %%mm6 \n\t"
"movq %%mm6, %%mm0 \n\t"
"psrlq $16, %%mm6 \n\t"
"paddw %%mm0, %%mm6 \n\t"
"movd %%mm6, %0 \n\t"
: "=r" (ret)
);
return ret&0xFFFF;
}
int pix_abs16x16_y2_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h)
static inline int sum_mmx2()
{
const UINT8 *p1, *p2;
p1 = blk1;
p2 = blk2;
pxor_r2r(mm7, mm7); /* mm7 is accumulator */
pxor_r2r(mm6, mm6); /* mm7 is zero constant */
movq_m2r(mm_wone, mm5); /* one constant */
do {
sad_add_x2(p1, p2, p2 + lx);
sad_add_x2(p1 + 8, p2 + 8, p2 + 8 + lx);
p1 += lx;
p2 += lx;
} while (--h);
return sad_end();
int ret;
asm volatile(
"movd %%mm6, %0 \n\t"
: "=r" (ret)
);
return ret;
}
/* mm7 is accumulator, mm6 is zero */
static inline void sad_add_xy2(const UINT8 *p1, const UINT8 *p2, const UINT8 *p3)
{
movq_m2r(*(p2 + 0), mm0);
movq_m2r(*(p3 + 0), mm1);
movq_r2r(mm0, mm2);
movq_r2r(mm1, mm3);
punpcklbw_r2r(mm6, mm0); /* extract 4 bytes low */
punpcklbw_r2r(mm6, mm1);
punpckhbw_r2r(mm6, mm2); /* high */
punpckhbw_r2r(mm6, mm3);
paddusw_r2r(mm1, mm0);
paddusw_r2r(mm3, mm2);
#define PIX_SAD(suf)\
int pix_abs8x8_ ## suf(UINT8 *blk2, UINT8 *blk1, int stride)\
{\
asm volatile("pxor %%mm7, %%mm7 \n\t"\
"pxor %%mm6, %%mm6 \n\t":);\
\
sad8_ ## suf(blk1, blk2, stride, 3);\
\
return sum_ ## suf();\
}\
\
int pix_abs8x8_x2_ ## suf(UINT8 *blk2, UINT8 *blk1, int stride)\
{\
asm volatile("pxor %%mm7, %%mm7 \n\t"\
"pxor %%mm6, %%mm6 \n\t"\
"movq %0, %%mm5 \n\t"\
:: "m"(round_tab[1]) \
);\
\
sad8_2_ ## suf(blk1, blk2+1, blk2, stride, 3);\
\
return sum_ ## suf();\
}\
\
int pix_abs8x8_y2_ ## suf(UINT8 *blk2, UINT8 *blk1, int stride)\
{\
asm volatile("pxor %%mm7, %%mm7 \n\t"\
"pxor %%mm6, %%mm6 \n\t"\
"movq %0, %%mm5 \n\t"\
:: "m"(round_tab[1]) \
);\
\
sad8_2_ ## suf(blk1, blk1+stride, blk2, stride, 3);\
\
return sum_ ## suf();\
}\
\
int pix_abs8x8_xy2_ ## suf(UINT8 *blk2, UINT8 *blk1, int stride)\
{\
asm volatile("pxor %%mm7, %%mm7 \n\t"\
"pxor %%mm6, %%mm6 \n\t"\
"movq %0, %%mm5 \n\t"\
:: "m"(round_tab[2]) \
);\
\
sad8_4_ ## suf(blk1, blk2, stride, 3);\
\
return sum_ ## suf();\
}\
\
int pix_abs16x16_ ## suf(UINT8 *blk2, UINT8 *blk1, int stride)\
{\
asm volatile("pxor %%mm7, %%mm7 \n\t"\
"pxor %%mm6, %%mm6 \n\t":);\
\
sad8_ ## suf(blk1 , blk2 , stride, 4);\
sad8_ ## suf(blk1+8, blk2+8, stride, 4);\
\
return sum_ ## suf();\
}\
int pix_abs16x16_x2_ ## suf(UINT8 *blk2, UINT8 *blk1, int stride)\
{\
asm volatile("pxor %%mm7, %%mm7 \n\t"\
"pxor %%mm6, %%mm6 \n\t"\
"movq %0, %%mm5 \n\t"\
:: "m"(round_tab[1]) \
);\
\
sad8_2_ ## suf(blk1 , blk1+1, blk2 , stride, 4);\
sad8_2_ ## suf(blk1+8, blk1+9, blk2+8, stride, 4);\
\
return sum_ ## suf();\
}\
int pix_abs16x16_y2_ ## suf(UINT8 *blk2, UINT8 *blk1, int stride)\
{\
asm volatile("pxor %%mm7, %%mm7 \n\t"\
"pxor %%mm6, %%mm6 \n\t"\
"movq %0, %%mm5 \n\t"\
:: "m"(round_tab[1]) \
);\
\
sad8_2_ ## suf(blk1 , blk1+stride, blk2 , stride, 4);\
sad8_2_ ## suf(blk1+8, blk1+stride+8,blk2+8, stride, 4);\
\
return sum_ ## suf();\
}\
int pix_abs16x16_xy2_ ## suf(UINT8 *blk2, UINT8 *blk1, int stride)\
{\
asm volatile("pxor %%mm7, %%mm7 \n\t"\
"pxor %%mm6, %%mm6 \n\t"\
"movq %0, %%mm5 \n\t"\
:: "m"(round_tab[2]) \
);\
\
sad8_4_ ## suf(blk1 , blk2 , stride, 4);\
sad8_4_ ## suf(blk1+8, blk2+8, stride, 4);\
\
return sum_ ## suf();\
}\
movq_m2r(*(p2 + 1), mm1);
movq_m2r(*(p3 + 1), mm3);
movq_r2r(mm1, mm4);
punpcklbw_r2r(mm6, mm1); /* low */
punpckhbw_r2r(mm6, mm4); /* high */
paddusw_r2r(mm1, mm0);
paddusw_r2r(mm4, mm2);
movq_r2r(mm3, mm4);
punpcklbw_r2r(mm6, mm3); /* low */
punpckhbw_r2r(mm6, mm4); /* high */
paddusw_r2r(mm3, mm0);
paddusw_r2r(mm4, mm2);
movq_m2r(*(p1 + 0), mm1); /* mm1 : other value */
paddusw_r2r(mm5, mm0); /* + 2 */
paddusw_r2r(mm5, mm2); /* + 2 */
psrlw_i2r(2, mm0);
psrlw_i2r(2, mm2);
packuswb_r2r(mm2, mm0); /* average is in mm0 */
movq_r2r(mm1, mm2);
psubusb_r2r(mm0, mm1);
psubusb_r2r(mm2, mm0);
por_r2r(mm1, mm0); /* mm0 is absolute value */
movq_r2r(mm0, mm1);
punpcklbw_r2r(mm6, mm0);
punpckhbw_r2r(mm6, mm1);
paddusw_r2r(mm0, mm7);
paddusw_r2r(mm1, mm7);
}
int pix_abs16x16_xy2_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h)
{
const UINT8 *p1, *p2, *p3;
p1 = blk1;
p2 = blk2;
p3 = blk2 + lx;
pxor_r2r(mm7, mm7); /* mm7 is accumulator */
pxor_r2r(mm6, mm6); /* mm7 is zero constant */
movq_m2r(mm_wtwo, mm5); /* one constant */
do {
sad_add_xy2(p1, p2, p2 + lx);
sad_add_xy2(p1 + 8, p2 + 8, p2 + 8 + lx);
p1 += lx;
p2 += lx;
} while (--h);
return sad_end();
}
PIX_SAD(mmx)
PIX_SAD(mmx2)

View File

@ -26,6 +26,7 @@
#include "mpegvideo.h"
#define ABS(a) ((a)>0 ? (a) : -(a))
#define MAX(a,b) ((a) > (b) ? (a) : (b))
#define INTER_BIAS 257
static void halfpel_motion_search(MpegEncContext * s,
@ -164,7 +165,7 @@ static int full_motion_search(MpegEncContext * s,
for (y = y1; y <= y2; y++) {
for (x = x1; x <= x2; x++) {
d = pix_abs16x16(pix, s->last_picture[0] + (y * s->linesize) + x,
s->linesize, 16);
s->linesize);
if (d < dmin ||
(d == dmin &&
(abs(x - xx) + abs(y - yy)) <
@ -228,7 +229,7 @@ static int log_motion_search(MpegEncContext * s,
do {
for (y = y1; y <= y2; y += range) {
for (x = x1; x <= x2; x += range) {
d = pix_abs16x16(pix, s->last_picture[0] + (y * s->linesize) + x, s->linesize, 16);
d = pix_abs16x16(pix, s->last_picture[0] + (y * s->linesize) + x, s->linesize);
if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
dmin = d;
mx = x;
@ -308,7 +309,7 @@ static int phods_motion_search(MpegEncContext * s,
lastx = x;
for (x = x1; x <= x2; x += range) {
d = pix_abs16x16(pix, s->last_picture[0] + (y * s->linesize) + x, s->linesize, 16);
d = pix_abs16x16(pix, s->last_picture[0] + (y * s->linesize) + x, s->linesize);
if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
dminx = d;
mx = x;
@ -317,7 +318,7 @@ static int phods_motion_search(MpegEncContext * s,
x = lastx;
for (y = y1; y <= y2; y += range) {
d = pix_abs16x16(pix, s->last_picture[0] + (y * s->linesize) + x, s->linesize, 16);
d = pix_abs16x16(pix, s->last_picture[0] + (y * s->linesize) + x, s->linesize);
if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
dminy = d;
my = y;
@ -361,7 +362,7 @@ static int phods_motion_search(MpegEncContext * s,
#define CHECK_MV(x,y)\
{\
d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride, 16);\
d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
if(d<dmin){\
best[0]=x;\
@ -372,7 +373,7 @@ static int phods_motion_search(MpegEncContext * s,
#define CHECK_MV_DIR(x,y,new_dir)\
{\
d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride, 16);\
d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
if(d<dmin){\
best[0]=x;\
@ -382,6 +383,30 @@ static int phods_motion_search(MpegEncContext * s,
}\
}
#define CHECK_MV4(x,y)\
{\
d = pix_abs8x8(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
if(d<dmin){\
best[0]=x;\
best[1]=y;\
dmin=d;\
}\
}
#define CHECK_MV4_DIR(x,y,new_dir)\
{\
d = pix_abs8x8(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\
d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\
if(d<dmin){\
best[0]=x;\
best[1]=y;\
dmin=d;\
next_dir= new_dir;\
}\
}
#define check(x,y,S,v)\
if( (x)<(xmin<<(S)) ) printf("%d %d %d %d xmin" #v, (x), (y), s->mb_x, s->mb_y);\
if( (x)>(xmax<<(S)) ) printf("%d %d %d %d xmax" #v, (x), (y), s->mb_x, s->mb_y);\
@ -440,6 +465,32 @@ static inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,
*/
}
static inline int small_diamond_search4MV(MpegEncContext * s, int *best, int dmin,
UINT8 *new_pic, UINT8 *old_pic, int pic_stride,
int pred_x, int pred_y, UINT16 *mv_penalty, int quant,
int xmin, int ymin, int xmax, int ymax, int shift)
{
int next_dir=-1;
for(;;){
int d;
const int dir= next_dir;
const int x= best[0];
const int y= best[1];
next_dir=-1;
//printf("%d", dir);
if(dir!=2 && x>xmin) CHECK_MV4_DIR(x-1, y , 0)
if(dir!=3 && y>ymin) CHECK_MV4_DIR(x , y-1, 1)
if(dir!=0 && x<xmax) CHECK_MV4_DIR(x+1, y , 2)
if(dir!=1 && y<ymax) CHECK_MV4_DIR(x , y+1, 3)
if(next_dir==-1){
return dmin;
}
}
}
static inline int snake_search(MpegEncContext * s, int *best, int dmin,
UINT8 *new_pic, UINT8 *old_pic, int pic_stride,
int pred_x, int pred_y, UINT16 *mv_penalty, int quant,
@ -469,7 +520,7 @@ if(256*256*256*64%point==0)
x+=x_dir[dir];
y+=y_dir[dir];
if(x>=xmin && x<=xmax && y>=ymin && y<=ymax){
d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride, 16);
d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);
d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;
}else{
d = dmin + 10000; //FIXME smarter boundary handling
@ -517,7 +568,7 @@ static int epzs_motion_search(MpegEncContext * s,
new_pic = s->new_picture[0] + pic_xy;
old_pic = s->last_picture[0] + pic_xy;
dmin = pix_abs16x16(new_pic, old_pic, pic_stride, 16);
dmin = pix_abs16x16(new_pic, old_pic, pic_stride);
if(dmin<Z_THRESHOLD){
*mx_ptr= 0;
*my_ptr= 0;
@ -557,8 +608,65 @@ static int epzs_motion_search(MpegEncContext * s,
return dmin;
}
static int epzs_motion_search4(MpegEncContext * s, int block,
int *mx_ptr, int *my_ptr,
int P[6][2], int pred_x, int pred_y,
int xmin, int ymin, int xmax, int ymax)
{
int best[2]={0, 0};
int d, dmin;
UINT8 *new_pic, *old_pic;
const int pic_stride= s->linesize;
const int pic_xy= ((s->mb_y*2 + (block>>1))*pic_stride + s->mb_x*2 + (block&1))*8;
UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
int quant= s->qscale; // qscale of the prev frame
const int shift= 1+s->quarter_sample;
new_pic = s->new_picture[0] + pic_xy;
old_pic = s->last_picture[0] + pic_xy;
dmin = pix_abs8x8(new_pic, old_pic, pic_stride);
/* first line */
if ((s->mb_y == 0 || s->first_slice_line || s->first_gob_line) && block<2) {
CHECK_MV4(P[1][0]>>shift, P[1][1]>>shift)
}else{
CHECK_MV4(P[4][0]>>shift, P[4][1]>>shift)
if(dmin<Z_THRESHOLD){
*mx_ptr= P[4][0]>>shift;
*my_ptr= P[4][1]>>shift;
//printf("M\n");
return dmin;
}
CHECK_MV4(P[1][0]>>shift, P[1][1]>>shift)
CHECK_MV4(P[2][0]>>shift, P[2][1]>>shift)
CHECK_MV4(P[3][0]>>shift, P[3][1]>>shift)
}
CHECK_MV4(P[0][0]>>shift, P[0][1]>>shift)
CHECK_MV4(P[5][0]>>shift, P[5][1]>>shift)
//check(best[0],best[1],0, b0)
dmin= small_diamond_search4MV(s, best, dmin, new_pic, old_pic, pic_stride,
pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, shift);
//check(best[0],best[1],0, b1)
*mx_ptr= best[0];
*my_ptr= best[1];
// printf("%d %d %d \n", best[0], best[1], dmin);
return dmin;
}
#define CHECK_HALF_MV(suffix, x, y) \
d= pix_abs16x16_ ## suffix(pix, ptr+((x)>>1), s->linesize, 16);\
d= pix_abs16x16_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*quant;\
if(d<dminh){\
dminh= d;\
mx= mx1 + x;\
my= my1 + y;\
}
#define CHECK_HALF_MV4(suffix, x, y) \
d= pix_abs8x8_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*quant;\
if(d<dminh){\
dminh= d;\
@ -614,6 +722,7 @@ static inline void halfpel_motion_search(MpegEncContext * s,
CHECK_HALF_MV(xy2, -1, +1)
CHECK_HALF_MV(y2 , 0, +1)
CHECK_HALF_MV(xy2, +1, +1)
}else{
mx= 2*(mx - xx);
my= 2*(my - yy);
@ -623,19 +732,99 @@ static inline void halfpel_motion_search(MpegEncContext * s,
*my_ptr = my;
}
static inline void halfpel_motion_search4(MpegEncContext * s,
int *mx_ptr, int *my_ptr, int dmin,
int xmin, int ymin, int xmax, int ymax,
int pred_x, int pred_y, int block_x, int block_y)
{
UINT16 *mv_penalty= s->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
const int quant= s->qscale;
int pen_x, pen_y;
int mx, my, mx1, my1, d, xx, yy, dminh;
UINT8 *pix, *ptr;
xx = 8 * block_x;
yy = 8 * block_y;
pix = s->new_picture[0] + (yy * s->linesize) + xx;
mx = *mx_ptr;
my = *my_ptr;
ptr = s->last_picture[0] + ((yy+my) * s->linesize) + xx + mx;
dminh = dmin;
if (mx > xmin && mx < xmax &&
my > ymin && my < ymax) {
mx= mx1= 2*mx;
my= my1= 2*my;
if(dmin < Z_THRESHOLD && mx==0 && my==0){
*mx_ptr = 0;
*my_ptr = 0;
return;
}
pen_x= pred_x + mx;
pen_y= pred_y + my;
ptr-= s->linesize;
CHECK_HALF_MV4(xy2, -1, -1)
CHECK_HALF_MV4(y2 , 0, -1)
CHECK_HALF_MV4(xy2, +1, -1)
ptr+= s->linesize;
CHECK_HALF_MV4(x2 , -1, 0)
CHECK_HALF_MV4(x2 , +1, 0)
CHECK_HALF_MV4(xy2, -1, +1)
CHECK_HALF_MV4(y2 , 0, +1)
CHECK_HALF_MV4(xy2, +1, +1)
}else{
mx*=2;
my*=2;
}
*mx_ptr = mx;
*my_ptr = my;
}
static inline void set_mv_tables(MpegEncContext * s, int mx, int my)
{
const int xy= s->mb_x + s->mb_y*s->mb_width;
s->mv_table[0][xy] = mx;
s->mv_table[1][xy] = my;
/* has allready been set to the 4 MV if 4MV is done */
if(!(s->flags&CODEC_FLAG_4MV)){
int mot_xy= s->block_index[0];
s->motion_val[mot_xy ][0]= mx;
s->motion_val[mot_xy ][1]= my;
s->motion_val[mot_xy+1][0]= mx;
s->motion_val[mot_xy+1][1]= my;
mot_xy += s->block_wrap[0];
s->motion_val[mot_xy ][0]= mx;
s->motion_val[mot_xy ][1]= my;
s->motion_val[mot_xy+1][0]= mx;
s->motion_val[mot_xy+1][1]= my;
}
}
#ifndef CONFIG_TEST_MV_ENCODE
int estimate_motion(MpegEncContext * s,
int mb_x, int mb_y,
int *mx_ptr, int *my_ptr)
void estimate_motion(MpegEncContext * s,
int mb_x, int mb_y)
{
UINT8 *pix, *ppix;
int sum, varc, vard, mx, my, range, dmin, xx, yy;
int xmin, ymin, xmax, ymax;
int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
int pred_x=0, pred_y=0;
int P[5][2];
int P[6][2];
const int shift= 1+s->quarter_sample;
int mb_type=0;
range = 8 * (1 << (s->f_code - 1));
/* XXX: temporary kludge to avoid overflow for msmpeg4 */
@ -680,14 +869,13 @@ int estimate_motion(MpegEncContext * s,
case ME_X1:
case ME_EPZS:
{
static const int off[4]= {2, 1, 1, -1};
const int mot_stride = s->mb_width*2 + 2;
const int mot_xy = (s->mb_y*2 + 1)*mot_stride + s->mb_x*2 + 1;
const int mot_stride = s->block_wrap[0];
const int mot_xy = s->block_index[0];
rel_xmin= xmin - s->mb_x*16;
rel_xmax= xmax - s->mb_x*16;
rel_ymin= ymin - s->mb_y*16;
rel_ymax= ymax - s->mb_y*16;
rel_xmin= xmin - mb_x*16;
rel_xmax= xmax - mb_x*16;
rel_ymin= ymin - mb_y*16;
rel_ymax= ymax - mb_y*16;
P[0][0] = s->motion_val[mot_xy ][0];
P[0][1] = s->motion_val[mot_xy ][1];
@ -696,14 +884,14 @@ int estimate_motion(MpegEncContext * s,
if(P[1][0] > (rel_xmax<<shift)) P[1][0]= (rel_xmax<<shift);
/* special case for first line */
if ((s->mb_y == 0 || s->first_slice_line || s->first_gob_line)) {
if ((mb_y == 0 || s->first_slice_line || s->first_gob_line)) {
P[4][0] = P[1][0];
P[4][1] = P[1][1];
} else {
P[2][0] = s->motion_val[mot_xy - mot_stride ][0];
P[2][1] = s->motion_val[mot_xy - mot_stride ][1];
P[3][0] = s->motion_val[mot_xy - mot_stride + off[0] ][0];
P[3][1] = s->motion_val[mot_xy - mot_stride + off[0] ][1];
P[3][0] = s->motion_val[mot_xy - mot_stride + 2 ][0];
P[3][1] = s->motion_val[mot_xy - mot_stride + 2 ][1];
if(P[2][1] > (rel_ymax<<shift)) P[2][1]= (rel_ymax<<shift);
if(P[3][0] < (rel_xmin<<shift)) P[3][0]= (rel_xmin<<shift);
if(P[3][1] > (rel_ymax<<shift)) P[3][1]= (rel_ymax<<shift);
@ -721,11 +909,73 @@ int estimate_motion(MpegEncContext * s,
}
dmin = epzs_motion_search(s, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax);
mx+= s->mb_x*16;
my+= s->mb_y*16;
mx+= mb_x*16;
my+= mb_y*16;
break;
}
if(s->flags&CODEC_FLAG_4MV){
int block;
mb_type|= MB_TYPE_INTER4V;
for(block=0; block<4; block++){
int mx4, my4;
int pred_x4, pred_y4;
int dmin4;
static const int off[4]= {2, 1, 1, -1};
const int mot_stride = s->block_wrap[0];
const int mot_xy = s->block_index[block];
const int block_x= mb_x*2 + (block&1);
const int block_y= mb_y*2 + (block>>1);
const int rel_xmin4= xmin - block_x*8;
const int rel_xmax4= xmax - block_x*8;
const int rel_ymin4= ymin - block_y*8;
const int rel_ymax4= ymax - block_y*8;
P[0][0] = s->motion_val[mot_xy ][0];
P[0][1] = s->motion_val[mot_xy ][1];
P[1][0] = s->motion_val[mot_xy - 1][0];
P[1][1] = s->motion_val[mot_xy - 1][1];
if(P[1][0] > (rel_xmax4<<shift)) P[1][0]= (rel_xmax4<<shift);
/* special case for first line */
if ((mb_y == 0 || s->first_slice_line || s->first_gob_line) && block<2) {
P[4][0] = P[1][0];
P[4][1] = P[1][1];
} else {
P[2][0] = s->motion_val[mot_xy - mot_stride ][0];
P[2][1] = s->motion_val[mot_xy - mot_stride ][1];
P[3][0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
P[3][1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
if(P[2][1] > (rel_ymax4<<shift)) P[2][1]= (rel_ymax4<<shift);
if(P[3][0] < (rel_xmin4<<shift)) P[3][0]= (rel_xmin4<<shift);
if(P[3][1] > (rel_ymax4<<shift)) P[3][1]= (rel_ymax4<<shift);
P[4][0]= mid_pred(P[1][0], P[2][0], P[3][0]);
P[4][1]= mid_pred(P[1][1], P[2][1], P[3][1]);
}
if(s->out_format == FMT_H263){
pred_x4 = P[4][0];
pred_y4 = P[4][1];
}else { /* mpeg1 at least */
pred_x4= P[1][0];
pred_y4= P[1][1];
}
P[5][0]= mx - mb_x*16;
P[5][1]= my - mb_y*16;
dmin4 = epzs_motion_search4(s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4);
halfpel_motion_search4(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4,
pred_x4, pred_y4, block_x, block_y);
s->motion_val[ s->block_index[block] ][0]= mx4;
s->motion_val[ s->block_index[block] ][1]= my4;
}
}
/* intra / predictive decision */
xx = mb_x * 16;
yy = mb_y * 16;
@ -737,7 +987,7 @@ int estimate_motion(MpegEncContext * s,
sum = pix_sum(pix, s->linesize);
#if 0
varc = pix_dev(pix, s->linesize, (sum+128)>>8) + INTER_BIAS;
vard = pix_abs16x16(pix, ppix, s->linesize, 16);
vard = pix_abs16x16(pix, ppix, s->linesize);
#else
sum= (sum+8)>>4;
varc = ((pix_norm1(pix, s->linesize) - sum*sum + 128 + 500)>>8);
@ -752,25 +1002,33 @@ int estimate_motion(MpegEncContext * s,
printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
#endif
if(s->flags&CODEC_FLAG_HQ){
if (vard*2 + 200 > varc)
mb_type|= MB_TYPE_INTRA;
if (varc*2 + 200 > vard){
mb_type|= MB_TYPE_INTER;
halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y);
}
}else{
if (vard <= 64 || vard < varc) {
mb_type|= MB_TYPE_INTER;
if (s->full_search != ME_ZERO) {
halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y);
} else {
mx -= 16 * s->mb_x;
my -= 16 * s->mb_y;
mx -= 16 * mb_x;
my -= 16 * mb_y;
}
// check(mx + 32*s->mb_x, my + 32*s->mb_y, 1, end)
*mx_ptr = mx;
*my_ptr = my;
return 0;
}else{
*mx_ptr = 0;
*my_ptr = 0;
return 1;
mb_type|= MB_TYPE_INTRA;
mx = 0;//mx*2 - 32 * mb_x;
my = 0;//my*2 - 32 * mb_y;
}
}
s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
set_mv_tables(s, mx, my);
}
#else
/* test version which generates valid random vectors */

View File

@ -228,6 +228,8 @@ int MPV_common_init(MpegEncContext *s)
goto fail;
}
s->block= s->intra_block;
s->context_initialized = 1;
return 0;
fail:
@ -295,7 +297,7 @@ int MPV_encode_init(AVCodecContext *avctx)
s->qblur= avctx->qblur;
s->avctx = avctx;
s->aspect_ratio_info= avctx->aspect_ratio_info;
s->hq= (avctx->flags & CODEC_FLAG_HQ);
s->flags= avctx->flags;
if (s->gop_size <= 1) {
s->intra_only = 1;
@ -1078,68 +1080,183 @@ void MPV_decode_mb(MpegEncContext *s, DCTELEM block[6][64])
}
}
the_end:
emms_c();
emms_c(); //FIXME remove
}
static void encode_mb(MpegEncContext *s)
{
int wrap;
const int mb_x= s->mb_x;
const int mb_y= s->mb_y;
UINT8 *ptr;
const int motion_x= s->mv[0][0][0];
const int motion_y= s->mv[0][0][1];
int i;
/* get the pixels */
wrap = s->linesize;
ptr = s->new_picture[0] + (mb_y * 16 * wrap) + mb_x * 16;
get_pixels(s->block[0], ptr, wrap);
get_pixels(s->block[1], ptr + 8, wrap);
get_pixels(s->block[2], ptr + 8 * wrap, wrap);
get_pixels(s->block[3], ptr + 8 * wrap + 8, wrap);
wrap = s->linesize >> 1;
ptr = s->new_picture[1] + (mb_y * 8 * wrap) + mb_x * 8;
get_pixels(s->block[4], ptr, wrap);
wrap = s->linesize >> 1;
ptr = s->new_picture[2] + (mb_y * 8 * wrap) + mb_x * 8;
get_pixels(s->block[5], ptr, wrap);
/* subtract previous frame if non intra */
if (!s->mb_intra) {
int dxy, offset, mx, my;
dxy = ((motion_y & 1) << 1) | (motion_x & 1);
ptr = s->last_picture[0] +
((mb_y * 16 + (motion_y >> 1)) * s->linesize) +
(mb_x * 16 + (motion_x >> 1));
sub_pixels_2(s->block[0], ptr, s->linesize, dxy);
sub_pixels_2(s->block[1], ptr + 8, s->linesize, dxy);
sub_pixels_2(s->block[2], ptr + s->linesize * 8, s->linesize, dxy);
sub_pixels_2(s->block[3], ptr + 8 + s->linesize * 8, s->linesize ,dxy);
if (s->out_format == FMT_H263) {
/* special rounding for h263 */
dxy = 0;
if ((motion_x & 3) != 0)
dxy |= 1;
if ((motion_y & 3) != 0)
dxy |= 2;
mx = motion_x >> 2;
my = motion_y >> 2;
} else {
mx = motion_x / 2;
my = motion_y / 2;
dxy = ((my & 1) << 1) | (mx & 1);
mx >>= 1;
my >>= 1;
}
offset = ((mb_y * 8 + my) * (s->linesize >> 1)) + (mb_x * 8 + mx);
ptr = s->last_picture[1] + offset;
sub_pixels_2(s->block[4], ptr, s->linesize >> 1, dxy);
ptr = s->last_picture[2] + offset;
sub_pixels_2(s->block[5], ptr, s->linesize >> 1, dxy);
}
#if 0
{
float adap_parm;
adap_parm = ((s->avg_mb_var << 1) + s->mb_var[s->mb_width*mb_y+mb_x] + 1.0) /
((s->mb_var[s->mb_width*mb_y+mb_x] << 1) + s->avg_mb_var + 1.0);
printf("\ntype=%c qscale=%2d adap=%0.2f dquant=%4.2f var=%4d avgvar=%4d",
(s->mb_type[s->mb_width*mb_y+mb_x] > 0) ? 'I' : 'P',
s->qscale, adap_parm, s->qscale*adap_parm,
s->mb_var[s->mb_width*mb_y+mb_x], s->avg_mb_var);
}
#endif
/* DCT & quantize */
if (s->h263_msmpeg4) {
msmpeg4_dc_scale(s);
} else if (s->h263_pred) {
h263_dc_scale(s);
} else {
/* default quantization values */
s->y_dc_scale = 8;
s->c_dc_scale = 8;
}
for(i=0;i<6;i++) {
s->block_last_index[i] = dct_quantize(s, s->block[i], i, s->qscale);
}
/* huffman encode */
switch(s->out_format) {
case FMT_MPEG1:
mpeg1_encode_mb(s, s->block, motion_x, motion_y);
break;
case FMT_H263:
if (s->h263_msmpeg4)
msmpeg4_encode_mb(s, s->block, motion_x, motion_y);
else if(s->h263_pred)
mpeg4_encode_mb(s, s->block, motion_x, motion_y);
else
h263_encode_mb(s, s->block, motion_x, motion_y);
break;
case FMT_MJPEG:
mjpeg_encode_mb(s, s->block);
break;
}
}
static void copy_bits(PutBitContext *pb, UINT8 *src, int length)
{
int bytes= length>>3;
int bits= length&7;
int i;
for(i=0; i<bytes; i++) put_bits(pb, 8, src[i]);
put_bits(pb, bits, src[i]>>(8-bits));
}
static void encode_picture(MpegEncContext *s, int picture_number)
{
int mb_x, mb_y, wrap, last_gob, pdif = 0;
UINT8 *ptr;
int i, motion_x, motion_y;
int mb_x, mb_y, last_gob, pdif = 0;
int i;
int bits;
MpegEncContext best_s;
UINT8 bit_buf[4][3000]; //FIXME check that this is ALLWAYS large enogh for a MB
s->picture_number = picture_number;
s->block_wrap[0]=
s->block_wrap[1]=
s->block_wrap[2]=
s->block_wrap[3]= s->mb_width*2 + 2;
s->block_wrap[4]=
s->block_wrap[5]= s->mb_width + 2;
s->last_mc_mb_var = s->mc_mb_var;
/* Reset the average MB variance */
s->avg_mb_var = 0;
s->mc_mb_var = 0;
/* Estimate motion for every MB */
if(s->pict_type == P_TYPE){
for(mb_y=0; mb_y < s->mb_height; mb_y++) {
s->block_index[0]= s->block_wrap[0]*(mb_y*2 + 1) - 1;
s->block_index[1]= s->block_wrap[0]*(mb_y*2 + 1);
s->block_index[2]= s->block_wrap[0]*(mb_y*2 + 2) - 1;
s->block_index[3]= s->block_wrap[0]*(mb_y*2 + 2);
for(mb_x=0; mb_x < s->mb_width; mb_x++) {
int xy= mb_y * s->mb_width + mb_x;
const int mot_stride = s->mb_width*2 + 2;
int mot_xy = (mb_y*2 + 1)*mot_stride + mb_x*2 + 1;
s->mb_x = mb_x;
s->mb_y = mb_y;
s->block_index[0]+=2;
s->block_index[1]+=2;
s->block_index[2]+=2;
s->block_index[3]+=2;
/* compute motion vector and macro block type (intra or non intra) */
motion_x = 0;
motion_y = 0;
if (s->pict_type == P_TYPE) {
s->mb_intra = estimate_motion(s, mb_x, mb_y,
&motion_x,
&motion_y);
} else {
s->mb_intra = 1;
}
/* Store MB type and MV */
s->mb_type[xy] = s->mb_intra;
s->mv_table[0][xy] = motion_x;
s->mv_table[1][xy] = motion_y;
s->motion_val[mot_xy ][0]= motion_x;
s->motion_val[mot_xy ][1]= motion_y;
s->motion_val[mot_xy+1][0]= motion_x;
s->motion_val[mot_xy+1][1]= motion_y;
mot_xy += mot_stride;
s->motion_val[mot_xy ][0]= motion_x;
s->motion_val[mot_xy ][1]= motion_y;
s->motion_val[mot_xy+1][0]= motion_x;
s->motion_val[mot_xy+1][1]= motion_y;
/* compute motion vector & mb_type and store in context */
estimate_motion(s, mb_x, mb_y);
// s->mb_type[mb_y*s->mb_width + mb_x]=MB_TYPE_INTER;
}
}
emms_c();
}else{
/* I-Frame */
//FIXME do we need to zero them?
memset(s->motion_val[0], 0, sizeof(INT16)*(s->mb_width*2 + 2)*(s->mb_height*2 + 2)*2);
memset(s->mv_table[0] , 0, sizeof(INT16)*s->mb_width*s->mb_height);
memset(s->mv_table[1] , 0, sizeof(INT16)*s->mb_width*s->mb_height);
memset(s->mb_type , MB_TYPE_INTRA, sizeof(UINT8)*s->mb_width*s->mb_height);
}
if(s->avg_mb_var < s->mc_mb_var && s->pict_type != B_TYPE){ //FIXME subtract MV bits
int i;
s->pict_type= I_TYPE;
s->picture_in_gop_number=0;
for(i=0; i<s->mb_num; i++){
s->mb_type[i] = 1;
s->mv_table[0][i] = 0;
s->mv_table[1][i] = 0;
}
memset(s->mb_type , MB_TYPE_INTRA, sizeof(UINT8)*s->mb_width*s->mb_height);
//printf("Scene change detected, encoding as I Frame\n");
}
/* find best f_code for ME which do unlimited searches */
@ -1152,7 +1269,7 @@ static void encode_picture(MpegEncContext *s, int picture_number)
for(i=0; i<8; i++) mv_num[i]=0;
for(i=0; i<s->mb_num; i++){
if(s->mb_type[i] == 0){
if(s->mb_type[i] & (MB_TYPE_INTER|MB_TYPE_INTER4V)){
mv_num[ fcode_tab[s->mv_table[0][i] + MAX_MV] ]++;
mv_num[ fcode_tab[s->mv_table[1][i] + MAX_MV] ]++;
//printf("%d %d %d\n", s->mv_table[0][i], fcode_tab[s->mv_table[0][i] + MAX_MV], i);
@ -1181,16 +1298,20 @@ static void encode_picture(MpegEncContext *s, int picture_number)
UINT8 * fcode_tab= s->fcode_tab;
for(i=0; i<s->mb_num; i++){
if(s->mb_type[i] == 0){
if(s->mb_type[i]&MB_TYPE_INTER){
if( fcode_tab[s->mv_table[0][i] + MAX_MV] > f_code
|| fcode_tab[s->mv_table[0][i] + MAX_MV] == 0
|| fcode_tab[s->mv_table[1][i] + MAX_MV] > f_code
|| fcode_tab[s->mv_table[1][i] + MAX_MV] == 0 ){
s->mb_type[i] = 1;
s->mb_type[i] &= ~MB_TYPE_INTER;
s->mb_type[i] |= MB_TYPE_INTRA;
s->mv_table[0][i] = 0;
s->mv_table[1][i] = 0;
}
}
if(s->mb_type[i]&MB_TYPE_INTER4V){
//FIXME
}
}
}
@ -1249,8 +1370,6 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->mb_incr = 1;
s->last_mv[0][0][0] = 0;
s->last_mv[0][0][1] = 0;
s->mv_type = MV_TYPE_16X16;
s->mv_dir = MV_DIR_FORWARD;
/* Get the GOB height based on picture height */
if (s->out_format == FMT_H263 && !s->h263_pred && !s->h263_msmpeg4) {
@ -1264,12 +1383,6 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->avg_mb_var = s->avg_mb_var / s->mb_num;
s->block_wrap[0]=
s->block_wrap[1]=
s->block_wrap[2]=
s->block_wrap[3]= s->mb_width*2 + 2;
s->block_wrap[4]=
s->block_wrap[5]= s->mb_width + 2;
for(mb_y=0; mb_y < s->mb_height; mb_y++) {
/* Put GOB header based on RTP MTU */
/* TODO: Put all this stuff in a separate generic function */
@ -1292,6 +1405,11 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->block_index[4]= s->block_wrap[4]*(mb_y + 1) + s->block_wrap[0]*(s->mb_height*2 + 2);
s->block_index[5]= s->block_wrap[4]*(mb_y + 1 + s->mb_height + 2) + s->block_wrap[0]*(s->mb_height*2 + 2);
for(mb_x=0; mb_x < s->mb_width; mb_x++) {
const int mb_type= s->mb_type[mb_y * s->mb_width + mb_x];
PutBitContext pb;
int d;
int dmin=10000000;
int best=0;
s->mb_x = mb_x;
s->mb_y = mb_y;
@ -1301,125 +1419,79 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->block_index[3]+=2;
s->block_index[4]++;
s->block_index[5]++;
#if 0
/* compute motion vector and macro block type (intra or non intra) */
motion_x = 0;
motion_y = 0;
if (s->pict_type == P_TYPE) {
s->mb_intra = estimate_motion(s, mb_x, mb_y,
&motion_x,
&motion_y);
} else {
s->mv_type = MV_TYPE_16X16;
s->mv_dir = MV_DIR_FORWARD;
if(mb_type & (mb_type-1)){ // more than 1 MB type possible
pb= s->pb;
if(mb_type&MB_TYPE_INTER){
s->mb_intra= 0;
s->mv[0][0][0] = s->mv_table[0][mb_y * s->mb_width + mb_x];
s->mv[0][0][1] = s->mv_table[1][mb_y * s->mb_width + mb_x];
init_put_bits(&s->pb, bit_buf[1], 3000, NULL, NULL);
s->block= s->inter_block;
encode_mb(s);
d= get_bit_count(&s->pb);
if(d<dmin){
flush_put_bits(&s->pb);
dmin=d;
best_s.mv[0][0][0]= s->mv[0][0][0];
best_s.mv[0][0][1]= s->mv[0][0][1];
best_s.mb_intra= 0;
best_s.pb=s->pb;
best_s.block= s->block;
best=1;
for(i=0; i<6; i++)
best_s.block_last_index[i]= s->block_last_index[i];
}
}
if(mb_type&MB_TYPE_INTRA){
s->mb_intra= 1;
s->mv[0][0][0] = 0;
s->mv[0][0][1] = 0;
init_put_bits(&s->pb, bit_buf[0], 3000, NULL, NULL);
s->block= s->intra_block;
encode_mb(s);
d= get_bit_count(&s->pb);
if(d<dmin){
flush_put_bits(&s->pb);
dmin=d;
best_s.mv[0][0][0]= 0;
best_s.mv[0][0][1]= 0;
best_s.mb_intra= 1;
best_s.pb=s->pb;
best_s.block= s->block;
for(i=0; i<6; i++)
best_s.block_last_index[i]= s->block_last_index[i];
best=0;
}
#endif
s->mb_intra = s->mb_type[mb_y * s->mb_width + mb_x];
motion_x = s->mv_table[0][mb_y * s->mb_width + mb_x];
motion_y = s->mv_table[1][mb_y * s->mb_width + mb_x];
/* get the pixels */
wrap = s->linesize;
ptr = s->new_picture[0] + (mb_y * 16 * wrap) + mb_x * 16;
get_pixels(s->block[0], ptr, wrap);
get_pixels(s->block[1], ptr + 8, wrap);
get_pixels(s->block[2], ptr + 8 * wrap, wrap);
get_pixels(s->block[3], ptr + 8 * wrap + 8, wrap);
wrap = s->linesize >> 1;
ptr = s->new_picture[1] + (mb_y * 8 * wrap) + mb_x * 8;
get_pixels(s->block[4], ptr, wrap);
wrap = s->linesize >> 1;
ptr = s->new_picture[2] + (mb_y * 8 * wrap) + mb_x * 8;
get_pixels(s->block[5], ptr, wrap);
/* subtract previous frame if non intra */
if (!s->mb_intra) {
int dxy, offset, mx, my;
dxy = ((motion_y & 1) << 1) | (motion_x & 1);
ptr = s->last_picture[0] +
((mb_y * 16 + (motion_y >> 1)) * s->linesize) +
(mb_x * 16 + (motion_x >> 1));
sub_pixels_2(s->block[0], ptr, s->linesize, dxy);
sub_pixels_2(s->block[1], ptr + 8, s->linesize, dxy);
sub_pixels_2(s->block[2], ptr + s->linesize * 8, s->linesize, dxy);
sub_pixels_2(s->block[3], ptr + 8 + s->linesize * 8, s->linesize ,dxy);
if (s->out_format == FMT_H263) {
/* special rounding for h263 */
dxy = 0;
if ((motion_x & 3) != 0)
dxy |= 1;
if ((motion_y & 3) != 0)
dxy |= 2;
mx = motion_x >> 2;
my = motion_y >> 2;
/* force cleaning of ac/dc if needed ... */
s->mbintra_table[mb_x + mb_y*s->mb_width]=1;
}
s->mv[0][0][0]= best_s.mv[0][0][0];
s->mv[0][0][1]= best_s.mv[0][0][1];
s->mb_intra= best_s.mb_intra;
for(i=0; i<6; i++)
s->block_last_index[i]= best_s.block_last_index[i];
copy_bits(&pb, bit_buf[best], dmin);
s->block= best_s.block;
s->pb= pb;
} else {
mx = motion_x / 2;
my = motion_y / 2;
dxy = ((my & 1) << 1) | (mx & 1);
mx >>= 1;
my >>= 1;
}
offset = ((mb_y * 8 + my) * (s->linesize >> 1)) + (mb_x * 8 + mx);
ptr = s->last_picture[1] + offset;
sub_pixels_2(s->block[4], ptr, s->linesize >> 1, dxy);
ptr = s->last_picture[2] + offset;
sub_pixels_2(s->block[5], ptr, s->linesize >> 1, dxy);
}
emms_c();
#if 0
{
float adap_parm;
adap_parm = ((s->avg_mb_var << 1) + s->mb_var[s->mb_width*mb_y+mb_x] + 1.0) /
((s->mb_var[s->mb_width*mb_y+mb_x] << 1) + s->avg_mb_var + 1.0);
printf("\ntype=%c qscale=%2d adap=%0.2f dquant=%4.2f var=%4d avgvar=%4d",
(s->mb_type[s->mb_width*mb_y+mb_x] > 0) ? 'I' : 'P',
s->qscale, adap_parm, s->qscale*adap_parm,
s->mb_var[s->mb_width*mb_y+mb_x], s->avg_mb_var);
}
#endif
/* DCT & quantize */
if (s->h263_msmpeg4) {
msmpeg4_dc_scale(s);
} else if (s->h263_pred) {
h263_dc_scale(s);
// only one MB-Type possible
if(mb_type&MB_TYPE_INTRA){
s->mb_intra= 1;
s->mv[0][0][0] = 0;
s->mv[0][0][1] = 0;
}else{
/* default quantization values */
s->y_dc_scale = 8;
s->c_dc_scale = 8;
s->mb_intra= 0;
s->mv[0][0][0] = s->mv_table[0][mb_y * s->mb_width + mb_x];
s->mv[0][0][1] = s->mv_table[1][mb_y * s->mb_width + mb_x];
}
for(i=0;i<6;i++) {
s->block_last_index[i] = dct_quantize(s, s->block[i], i, s->qscale);
encode_mb(s);
}
/* huffman encode */
switch(s->out_format) {
case FMT_MPEG1:
mpeg1_encode_mb(s, s->block, motion_x, motion_y);
break;
case FMT_H263:
if (s->h263_msmpeg4)
msmpeg4_encode_mb(s, s->block, motion_x, motion_y);
else if(s->h263_pred)
mpeg4_encode_mb(s, s->block, motion_x, motion_y);
else
h263_encode_mb(s, s->block, motion_x, motion_y);
break;
case FMT_MJPEG:
mjpeg_encode_mb(s, s->block);
break;
}
/* decompress blocks so that we keep the state of the decoder */
s->mv[0][0][0] = motion_x;
s->mv[0][0][1] = motion_y;
MPV_decode_mb(s, s->block);
}
@ -1437,6 +1509,7 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->first_gob_line = 0;
}
}
emms_c();
if (s->h263_msmpeg4 && s->pict_type == I_TYPE)
msmpeg4_encode_ext_header(s);
@ -1454,7 +1527,6 @@ static void encode_picture(MpegEncContext *s, int picture_number)
s->ptr_lastgob = pbBufPtr(&s->pb);
//fprintf(stderr,"\nGOB: %2d size: %d (last)", s->gob_number, pdif);
}
}
static int dct_quantize_c(MpegEncContext *s,

View File

@ -65,7 +65,7 @@ typedef struct MpegEncContext {
int qmax; /* max qscale */
int max_qdiff; /* max qscale difference between frames */
int encoding; /* true if we are encoding (vs decoding) */
int hq; /* set if CODEC_FLAG_HQ is used in AVCodecContext.flags */
int flags; /* AVCodecContext.flags (HQ, MV4, ...) */
/* the following fields are managed internally by the encoder */
/* bit output */
@ -141,8 +141,16 @@ typedef struct MpegEncContext {
int mb_x, mb_y;
int mb_incr;
int mb_intra;
INT16 *mb_var; /* Table for MB variances */
char *mb_type; /* Table for MB type */
UINT16 *mb_var; /* Table for MB variances */
UINT8 *mb_type; /* Table for MB type */
#define MB_TYPE_INTRA 0x01
#define MB_TYPE_INTER 0x02
#define MB_TYPE_INTER4V 0x04
#define MB_TYPE_SKIPED 0x08
#define MB_TYPE_DIRECT 0x10
#define MB_TYPE_FORWARD 0x20
#define MB_TYPE_BACKWAD 0x40
#define MB_TYPE_BIDIR 0x80
int block_index[6];
int block_wrap[6];
@ -295,7 +303,10 @@ typedef struct MpegEncContext {
UINT8 *ptr_last_mb_line;
UINT32 mb_line_avgsize;
DCTELEM block[6][64] __align8;
DCTELEM (*block)[64]; /* points to one of the following blocks */
DCTELEM intra_block[6][64] __align8;
DCTELEM inter_block[6][64] __align8;
DCTELEM inter4v_block[6][64] __align8;
void (*dct_unquantize)(struct MpegEncContext *s,
DCTELEM *block, int n, int qscale);
} MpegEncContext;
@ -311,9 +322,8 @@ void MPV_common_init_mmx(MpegEncContext *s);
/* motion_est.c */
int estimate_motion(MpegEncContext *s,
int mb_x, int mb_y,
int *mx_ptr, int *my_ptr);
void estimate_motion(MpegEncContext *s,
int mb_x, int mb_y);
/* mpeg12.c */
extern INT16 default_intra_matrix[64];