dsputil.c

arm平台下的H264编码和解码源代码

        dst+=dstStride;
        src+=srcStride;
    }
}
static inline void copy_block8(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
{
    int i;
    for(i=0; i<h; i++)
    {
		//ST32(dst   , LD32(src   ));
        //ST32(dst+4 , LD32(src+4 ));
		memcpy(dst,src,sizeof(src));
		memcpy(dst+4,src+4,sizeof(src));
        dst+=dstStride;
        src+=srcStride;
    }
}
static inline void copy_block16(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
{
    int i;
    for(i=0; i<h; i++)
    {
        //ST32(dst   , LD32(src   ));
        //ST32(dst+4 , LD32(src+4 ));
        //ST32(dst+8 , LD32(src+8 ));
        //ST32(dst+12, LD32(src+12));
		memcpy(dst,src,sizeof(src));
		memcpy(dst+4,src+4,sizeof(src));
		memcpy(dst+8,src+8,sizeof(src));
		memcpy(dst+12,src+12,sizeof(src));
        dst+=dstStride;
        src+=srcStride;
    }
}
static inline void copy_block17(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
{
    int i;
    for(i=0; i<h; i++)
    {
        //ST32(dst   , LD32(src   ));
        //ST32(dst+4 , LD32(src+4 ));
        //ST32(dst+8 , LD32(src+8 ));
        //ST32(dst+12, LD32(src+12));
		//ST32(dst   , LD32(src   ));
        //ST32(dst+4 , LD32(src+4 ));
        //ST32(dst+8 , LD32(src+8 ));
        //ST32(dst+12, LD32(src+12));
		memcpy(dst,src,sizeof(src));
		memcpy(dst+4,src+4,sizeof(src));
		memcpy(dst+8,src+8,sizeof(src));
		memcpy(dst+12,src+12,sizeof(src));
        dst[16]= src[16];
        dst+=dstStride;
        src+=srcStride;
    }
}
static inline void copy_block9(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h)
{
    int i;
    for(i=0; i<h; i++)
    {
        //ST32(dst   , LD32(src   ));
        //ST32(dst+4 , LD32(src+4 ));
		memcpy(dst,src,sizeof(src));
		memcpy(dst+4,src+4,sizeof(src));
        dst[8]= src[8];
        dst+=dstStride;
        src+=srcStride;
    }
}
#define QPEL_MC(r, OPNAME, RND, OP) \
static void OPNAME ## mpeg4_qpel8_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
    uint8_t *cm = cropTbl + MAX_NEG_CROP;\
    int i;\
    for(i=0; i<h; i++)\
    {\
        OP(dst[0], (src[0]+src[1])*20 - (src[0]+src[2])*6 + (src[1]+src[3])*3 - (src[2]+src[4]));\
        OP(dst[1], (src[1]+src[2])*20 - (src[0]+src[3])*6 + (src[0]+src[4])*3 - (src[1]+src[5]));\
        OP(dst[2], (src[2]+src[3])*20 - (src[1]+src[4])*6 + (src[0]+src[5])*3 - (src[0]+src[6]));\
        OP(dst[3], (src[3]+src[4])*20 - (src[2]+src[5])*6 + (src[1]+src[6])*3 - (src[0]+src[7]));\
        OP(dst[4], (src[4]+src[5])*20 - (src[3]+src[6])*6 + (src[2]+src[7])*3 - (src[1]+src[8]));\
        OP(dst[5], (src[5]+src[6])*20 - (src[4]+src[7])*6 + (src[3]+src[8])*3 - (src[2]+src[8]));\
        OP(dst[6], (src[6]+src[7])*20 - (src[5]+src[8])*6 + (src[4]+src[8])*3 - (src[3]+src[7]));\
        OP(dst[7], (src[7]+src[8])*20 - (src[6]+src[8])*6 + (src[5]+src[7])*3 - (src[4]+src[6]));\
        dst+=dstStride;\
        src+=srcStride;\
    }\
}\
\
static void OPNAME ## mpeg4_qpel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    const int w=8;\
    uint8_t *cm = cropTbl + MAX_NEG_CROP;\
    int i;\
    for(i=0; i<w; i++)\
    {\
        const int src0= src[0*srcStride];\
        const int src1= src[1*srcStride];\
        const int src2= src[2*srcStride];\
        const int src3= src[3*srcStride];\
        const int src4= src[4*srcStride];\
        const int src5= src[5*srcStride];\
        const int src6= src[6*srcStride];\
        const int src7= src[7*srcStride];\
        const int src8= src[8*srcStride];\
        OP(dst[0*dstStride], (src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4));\
        OP(dst[1*dstStride], (src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5));\
        OP(dst[2*dstStride], (src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6));\
        OP(dst[3*dstStride], (src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7));\
        OP(dst[4*dstStride], (src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8));\
        OP(dst[5*dstStride], (src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8));\
        OP(dst[6*dstStride], (src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7));\
        OP(dst[7*dstStride], (src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6));\
        dst++;\
        src++;\
    }\
}\
\
static void OPNAME ## mpeg4_qpel16_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
    uint8_t *cm = cropTbl + MAX_NEG_CROP;\
    int i;\
    \
    for(i=0; i<h; i++)\
    {\
        OP(dst[ 0], (src[ 0]+src[ 1])*20 - (src[ 0]+src[ 2])*6 + (src[ 1]+src[ 3])*3 - (src[ 2]+src[ 4]));\
        OP(dst[ 1], (src[ 1]+src[ 2])*20 - (src[ 0]+src[ 3])*6 + (src[ 0]+src[ 4])*3 - (src[ 1]+src[ 5]));\
        OP(dst[ 2], (src[ 2]+src[ 3])*20 - (src[ 1]+src[ 4])*6 + (src[ 0]+src[ 5])*3 - (src[ 0]+src[ 6]));\
        OP(dst[ 3], (src[ 3]+src[ 4])*20 - (src[ 2]+src[ 5])*6 + (src[ 1]+src[ 6])*3 - (src[ 0]+src[ 7]));\
        OP(dst[ 4], (src[ 4]+src[ 5])*20 - (src[ 3]+src[ 6])*6 + (src[ 2]+src[ 7])*3 - (src[ 1]+src[ 8]));\
        OP(dst[ 5], (src[ 5]+src[ 6])*20 - (src[ 4]+src[ 7])*6 + (src[ 3]+src[ 8])*3 - (src[ 2]+src[ 9]));\
        OP(dst[ 6], (src[ 6]+src[ 7])*20 - (src[ 5]+src[ 8])*6 + (src[ 4]+src[ 9])*3 - (src[ 3]+src[10]));\
        OP(dst[ 7], (src[ 7]+src[ 8])*20 - (src[ 6]+src[ 9])*6 + (src[ 5]+src[10])*3 - (src[ 4]+src[11]));\
        OP(dst[ 8], (src[ 8]+src[ 9])*20 - (src[ 7]+src[10])*6 + (src[ 6]+src[11])*3 - (src[ 5]+src[12]));\
        OP(dst[ 9], (src[ 9]+src[10])*20 - (src[ 8]+src[11])*6 + (src[ 7]+src[12])*3 - (src[ 6]+src[13]));\
        OP(dst[10], (src[10]+src[11])*20 - (src[ 9]+src[12])*6 + (src[ 8]+src[13])*3 - (src[ 7]+src[14]));\
        OP(dst[11], (src[11]+src[12])*20 - (src[10]+src[13])*6 + (src[ 9]+src[14])*3 - (src[ 8]+src[15]));\
        OP(dst[12], (src[12]+src[13])*20 - (src[11]+src[14])*6 + (src[10]+src[15])*3 - (src[ 9]+src[16]));\
        OP(dst[13], (src[13]+src[14])*20 - (src[12]+src[15])*6 + (src[11]+src[16])*3 - (src[10]+src[16]));\
        OP(dst[14], (src[14]+src[15])*20 - (src[13]+src[16])*6 + (src[12]+src[16])*3 - (src[11]+src[15]));\
        OP(dst[15], (src[15]+src[16])*20 - (src[14]+src[16])*6 + (src[13]+src[15])*3 - (src[12]+src[14]));\
        dst+=dstStride;\
        src+=srcStride;\
    }\
}\
\
static void OPNAME ## mpeg4_qpel16_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
    uint8_t *cm = cropTbl + MAX_NEG_CROP;\
    int i;\
    const int w=16;\
    for(i=0; i<w; i++)\
    {\
        const int src0= src[0*srcStride];\
        const int src1= src[1*srcStride];\
        const int src2= src[2*srcStride];\
        const int src3= src[3*srcStride];\
        const int src4= src[4*srcStride];\
        const int src5= src[5*srcStride];\
        const int src6= src[6*srcStride];\
        const int src7= src[7*srcStride];\
        const int src8= src[8*srcStride];\
        const int src9= src[9*srcStride];\
        const int src10= src[10*srcStride];\
        const int src11= src[11*srcStride];\
        const int src12= src[12*srcStride];\
        const int src13= src[13*srcStride];\
        const int src14= src[14*srcStride];\
        const int src15= src[15*srcStride];\
        const int src16= src[16*srcStride];\
        OP(dst[ 0*dstStride], (src0 +src1 )*20 - (src0 +src2 )*6 + (src1 +src3 )*3 - (src2 +src4 ));\
        OP(dst[ 1*dstStride], (src1 +src2 )*20 - (src0 +src3 )*6 + (src0 +src4 )*3 - (src1 +src5 ));\
        OP(dst[ 2*dstStride], (src2 +src3 )*20 - (src1 +src4 )*6 + (src0 +src5 )*3 - (src0 +src6 ));\
        OP(dst[ 3*dstStride], (src3 +src4 )*20 - (src2 +src5 )*6 + (src1 +src6 )*3 - (src0 +src7 ));\
        OP(dst[ 4*dstStride], (src4 +src5 )*20 - (src3 +src6 )*6 + (src2 +src7 )*3 - (src1 +src8 ));\
        OP(dst[ 5*dstStride], (src5 +src6 )*20 - (src4 +src7 )*6 + (src3 +src8 )*3 - (src2 +src9 ));\
        OP(dst[ 6*dstStride], (src6 +src7 )*20 - (src5 +src8 )*6 + (src4 +src9 )*3 - (src3 +src10));\
        OP(dst[ 7*dstStride], (src7 +src8 )*20 - (src6 +src9 )*6 + (src5 +src10)*3 - (src4 +src11));\
        OP(dst[ 8*dstStride], (src8 +src9 )*20 - (src7 +src10)*6 + (src6 +src11)*3 - (src5 +src12));\
        OP(dst[ 9*dstStride], (src9 +src10)*20 - (src8 +src11)*6 + (src7 +src12)*3 - (src6 +src13));\
        OP(dst[10*dstStride], (src10+src11)*20 - (src9 +src12)*6 + (src8 +src13)*3 - (src7 +src14));\
        OP(dst[11*dstStride], (src11+src12)*20 - (src10+src13)*6 + (src9 +src14)*3 - (src8 +src15));\
        OP(dst[12*dstStride], (src12+src13)*20 - (src11+src14)*6 + (src10+src15)*3 - (src9 +src16));\
        OP(dst[13*dstStride], (src13+src14)*20 - (src12+src15)*6 + (src11+src16)*3 - (src10+src16));\
        OP(dst[14*dstStride], (src14+src15)*20 - (src13+src16)*6 + (src12+src16)*3 - (src11+src15));\
        OP(dst[15*dstStride], (src15+src16)*20 - (src14+src16)*6 + (src13+src15)*3 - (src12+src14));\
        dst++;\
        src++;\
    }\
}\
\
static void OPNAME ## qpel8_mc00_c (uint8_t *dst, uint8_t *src, int stride){\
    OPNAME ## pixels8_c(dst, src, stride, 8);\
}\
\
static void OPNAME ## qpel8_mc10_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t half[64];\
    put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
    OPNAME ## pixels8_l2(dst, src, half, stride, stride, 8, 8);\
}\
\
static void OPNAME ## qpel8_mc20_c(uint8_t *dst, uint8_t *src, int stride){\
    OPNAME ## mpeg4_qpel8_h_lowpass(dst, src, stride, stride, 8);\
}\
\
static void OPNAME ## qpel8_mc30_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t half[64];\
    put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
    OPNAME ## pixels8_l2(dst, src+1, half, stride, stride, 8, 8);\
}\
\
static void OPNAME ## qpel8_mc01_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t half[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
    OPNAME ## pixels8_l2(dst, full, half, stride, 16, 8, 8);\
}\
\
static void OPNAME ## qpel8_mc02_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    copy_block9(full, src, 16, stride, 9);\
    OPNAME ## mpeg4_qpel8_v_lowpass(dst, full, stride, 16);\
}\
\
static void OPNAME ## qpel8_mc03_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t half[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
    OPNAME ## pixels8_l2(dst, full+16, half, stride, 16, 8, 8);\
}\
void ff_ ## OPNAME ## qpel8_mc11_old_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfV[64];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l4(dst, full, halfH, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
}\
static void OPNAME ## qpel8_mc11_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
}\
void ff_ ## OPNAME ## qpel8_mc31_old_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfV[64];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l4(dst, full+1, halfH, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
}\
static void OPNAME ## qpel8_mc31_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
}\
void ff_ ## OPNAME ## qpel8_mc13_old_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfV[64];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full, 8, 16);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l4(dst, full+16, halfH+8, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
}\
static void OPNAME ## qpel8_mc13_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2(halfH, halfH, full, 8, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
}\
void ff_ ## OPNAME ## qpel8_mc33_old_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfV[64];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full  , 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfV, full+1, 8, 16);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l4(dst, full+17, halfH+8, halfV, halfHV, stride, 16, 8, 8, 8, 8);\
}\
static void OPNAME ## qpel8_mc33_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2(halfH, halfH, full+1, 8, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
}\
static void OPNAME ## qpel8_mc21_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2(dst, halfH, halfHV, stride, 8, 8, 8);\
}\
static void OPNAME ## qpel8_mc23_c(uint8_t *dst, uint8_t *src, int stride){\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2(dst, halfH+8, halfHV, stride, 8, 8, 8);\
}\
void ff_ ## OPNAME ## qpel8_mc12_old_c(uint8_t *dst, uint8_t *src, int stride){\