dsputil.c

arm平台下的H264编码和解码源代码

        pixels[7] = cm[block[7]];

        pixels += line_size;
        block += 8;
    }
}

static void put_signed_pixels_clamped_c(const DCTELEM *block, 
                                        uint8_t *restrict pixels,
                                        int line_size)
{
    int i, j;

    for (i = 0; i < 8; i++) {
        for (j = 0; j < 8; j++) {
            if (*block < -128)
                *pixels = 0;
            else if (*block > 127)
                *pixels = 255;
            else
                *pixels = (uint8_t)(*block + 128);
            block++;
            pixels++;
        }
        pixels += (line_size - 8);
    }
}

static void add_pixels_clamped_c(const DCTELEM *block, uint8_t *restrict pixels,
                          int line_size)
{
    int i;
    uint8_t *cm = cropTbl + MAX_NEG_CROP;
    
    /* read the pixels */
    for(i=0;i<8;i++) {
        pixels[0] = cm[pixels[0] + block[0]];
        pixels[1] = cm[pixels[1] + block[1]];
        pixels[2] = cm[pixels[2] + block[2]];
        pixels[3] = cm[pixels[3] + block[3]];
        pixels[4] = cm[pixels[4] + block[4]];
        pixels[5] = cm[pixels[5] + block[5]];
        pixels[6] = cm[pixels[6] + block[6]];
        pixels[7] = cm[pixels[7] + block[7]];
        pixels += line_size;
        block += 8;
    }
}

#if 0

#define PIXOP2(OPNAME, OP) \
static void OPNAME ## _pixels(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
{\
    int i;\
    for(i=0; i<h; i++){\
        OP(*((uint64_t*)block), LD64(pixels));\
        pixels+=line_size;\
        block +=line_size;\
    }\
}\
\
static void OPNAME ## _no_rnd_pixels_x2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
{\
    int i;\
    for(i=0; i<h; i++){\
        const uint64_t a= LD64(pixels  );\
        const uint64_t b= LD64(pixels+1);\
        OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
        pixels+=line_size;\
        block +=line_size;\
    }\
}\
\
static void OPNAME ## _pixels_x2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
{\
    int i;\
    for(i=0; i<h; i++){\
        const uint64_t a= LD64(pixels  );\
        const uint64_t b= LD64(pixels+1);\
        OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
        pixels+=line_size;\
        block +=line_size;\
    }\
}\
\
static void OPNAME ## _no_rnd_pixels_y2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
{\
    int i;\
    for(i=0; i<h; i++){\
        const uint64_t a= LD64(pixels          );\
        const uint64_t b= LD64(pixels+line_size);\
        OP(*((uint64_t*)block), (a&b) + (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
        pixels+=line_size;\
        block +=line_size;\
    }\
}\
\
static void OPNAME ## _pixels_y2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
{\
    int i;\
    for(i=0; i<h; i++){\
        const uint64_t a= LD64(pixels          );\
        const uint64_t b= LD64(pixels+line_size);\
        OP(*((uint64_t*)block), (a|b) - (((a^b)&0xFEFEFEFEFEFEFEFEULL)>>1));\
        pixels+=line_size;\
        block +=line_size;\
    }\
}\
\
static void OPNAME ## _pixels_xy2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
{\
        int i;\
        const uint64_t a= LD64(pixels  );\
        const uint64_t b= LD64(pixels+1);\
        uint64_t l0=  (a&0x0303030303030303ULL)\
                    + (b&0x0303030303030303ULL)\
                    + 0x0202020202020202ULL;\
        uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
                   + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
        uint64_t l1,h1;\
\
        pixels+=line_size;\
        for(i=0; i<h; i+=2){\
            uint64_t a= LD64(pixels  );\
            uint64_t b= LD64(pixels+1);\
            l1=  (a&0x0303030303030303ULL)\
               + (b&0x0303030303030303ULL);\
            h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
              + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
            OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
            pixels+=line_size;\
            block +=line_size;\
            a= LD64(pixels  );\
            b= LD64(pixels+1);\
            l0=  (a&0x0303030303030303ULL)\
               + (b&0x0303030303030303ULL)\
               + 0x0202020202020202ULL;\
            h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
              + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
            OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
            pixels+=line_size;\
            block +=line_size;\
        }\
}\
\
static void OPNAME ## _no_rnd_pixels_xy2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h)\
{\
        int i;\
        const uint64_t a= LD64(pixels  );\
        const uint64_t b= LD64(pixels+1);\
        uint64_t l0=  (a&0x0303030303030303ULL)\
                    + (b&0x0303030303030303ULL)\
                    + 0x0101010101010101ULL;\
        uint64_t h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
                   + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
        uint64_t l1,h1;\
\
        pixels+=line_size;\
        for(i=0; i<h; i+=2){\
            uint64_t a= LD64(pixels  );\
            uint64_t b= LD64(pixels+1);\
            l1=  (a&0x0303030303030303ULL)\
               + (b&0x0303030303030303ULL);\
            h1= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
              + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
            OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
            pixels+=line_size;\
            block +=line_size;\
            a= LD64(pixels  );\
            b= LD64(pixels+1);\
            l0=  (a&0x0303030303030303ULL)\
               + (b&0x0303030303030303ULL)\
               + 0x0101010101010101ULL;\
            h0= ((a&0xFCFCFCFCFCFCFCFCULL)>>2)\
              + ((b&0xFCFCFCFCFCFCFCFCULL)>>2);\
            OP(*((uint64_t*)block), h0+h1+(((l0+l1)>>2)&0x0F0F0F0F0F0F0F0FULL));\
            pixels+=line_size;\
            block +=line_size;\
        }\
}\
\
CALL_2X_PIXELS(OPNAME ## _pixels16_c    , OPNAME ## _pixels_c    , 8)\
CALL_2X_PIXELS(OPNAME ## _pixels16_x2_c , OPNAME ## _pixels_x2_c , 8)\
CALL_2X_PIXELS(OPNAME ## _pixels16_y2_c , OPNAME ## _pixels_y2_c , 8)\
CALL_2X_PIXELS(OPNAME ## _pixels16_xy2_c, OPNAME ## _pixels_xy2_c, 8)\
CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_x2_c , OPNAME ## _no_rnd_pixels_x2_c , 8)\
CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_y2_c , OPNAME ## _no_rnd_pixels_y2_c , 8)\
CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_xy2_c, OPNAME ## _no_rnd_pixels_xy2_c, 8)
#define op_avg(a, b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEFEFEFEFEULL)>>1) )
#else // 64 bit variant
#define op_avg(a, b) a = rnd_avg32(a, b)
#define op_put(a, b) a = b

static void avg_pixels4_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    int i;
    for(i=0; i<h; i++){
        op_avg(*((uint32_t*)(block  )), LD32(pixels  ));
        pixels+=line_size;
        block +=line_size;
    }
}
static void avg_pixels8_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    int i;
    for(i=0; i<h; i++){
        op_avg(*((uint32_t*)(block  )), LD32(pixels  ));
        op_avg(*((uint32_t*)(block+4)), LD32(pixels+4));
        pixels+=line_size;
        block +=line_size;
    }
}
static inline void avg_pixels8_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h){
    int i;
    for(i=0; i<h; i++){
        uint32_t a,b;
        a= LD32(&src1[i*src_stride1  ]);
        b= LD32(&src2[i*src_stride2  ]);
        op_avg(*((uint32_t*)&dst[i*dst_stride  ]), rnd_avg32(a, b));
        a= LD32(&src1[i*src_stride1+4]);
        b= LD32(&src2[i*src_stride2+4]);
        op_avg(*((uint32_t*)&dst[i*dst_stride+4]), rnd_avg32(a, b));
    }
}

static inline void avg_pixels4_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h){
    int i;
    for(i=0; i<h; i++){
        uint32_t a,b;
        a= LD32(&src1[i*src_stride1  ]);
        b= LD32(&src2[i*src_stride2  ]);
        op_avg(*((uint32_t*)&dst[i*dst_stride  ]), rnd_avg32(a, b));
    }
}

#undef op_avg
#undef op_put 

static void put_pixels4_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    int i;
    for(i=0; i<h; i++){
        //op_put(*((uint32_t*)(block  )), LD32(pixels  ));
		memcpy((block  ), (pixels  ),sizeof(pixels));
        pixels+=line_size;
        block +=line_size;
    }
}

static void put_pixels8_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){
    int i;
    for(i=0; i<h; i++){
        memcpy((block  ), (pixels  ),sizeof(pixels));
        memcpy((block+4), (pixels+4),sizeof(pixels));
        pixels+=line_size;
        block +=line_size;
    }
}

static inline void put_pixels8_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h){
    int i;
	uint32_t a,b,c;
    for(i=0; i<h; i++){        
        //a= LD32(&src1[i*src_stride1  ]);
        //b= LD32(&src2[i*src_stride2  ]);		
		memcpy(&a,&src1[i*src_stride1  ],sizeof(a));
		memcpy(&b,&src2[i*src_stride2  ],sizeof(b));
		c=rnd_avg32(a, b);
        //op_put(*((uint32_t*)&dst[i*dst_stride  ]), rnd_avg32(a, b));
		memcpy(&dst[i*dst_stride  ],&c,sizeof(c));


        //a= LD32(&src1[i*src_stride1+4]);
        //b= LD32(&src2[i*src_stride2+4]);
		memcpy(&a,&src1[i*src_stride1+4],sizeof(a));
		memcpy(&b,&src2[i*src_stride2+4],sizeof(b));
		c=rnd_avg32(a, b);
        //op_put(*((uint32_t*)&dst[i*dst_stride+4]), rnd_avg32(a, b));
		memcpy(&dst[i*dst_stride+4],&c,sizeof(c));
    }
}

static inline void put_pixels4_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h){
    int i;
	uint32_t a,b,c;
    for(i=0; i<h; i++){
        
        //a= LD32(&src1[i*src_stride1  ]);
        //b= LD32(&src2[i*src_stride2  ]);
		memcpy(&a,&src1[i*src_stride1  ],sizeof(a));
		memcpy(&b,&src2[i*src_stride2  ],sizeof(b));
		c=rnd_avg32(a, b);
        //OP(*((uint32_t*)&dst[i*dst_stride  ]), rnd_avg32(a, b));
		memcpy(&dst[i*dst_stride  ],&c,sizeof(c));
    }
}



#define PIXOP2(OPNAME, OP) \
static void OPNAME ## _pixels2_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
    int i;\
    for(i=0; i<h; i++){\
        OP(*((uint16_t*)(block  )), LD16(pixels  ));\
        pixels+=line_size;\
        block +=line_size;\
    }\
}\
static inline void OPNAME ## _no_rnd_pixels8_c(uint8_t *block, const uint8_t *pixels, int line_size, int h){\
    OPNAME ## _pixels8_c(block, pixels, line_size, h);\
}\
\
static inline void OPNAME ## _no_rnd_pixels8_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, \
                                                int src_stride1, int src_stride2, int h){\
    int i;\
    for(i=0; i<h; i++){\
        uint32_t a,b;\
        a= LD32(&src1[i*src_stride1  ]);\
        b= LD32(&src2[i*src_stride2  ]);\
        OP(*((uint32_t*)&dst[i*dst_stride  ]), no_rnd_avg32(a, b));\
        a= LD32(&src1[i*src_stride1+4]);\
        b= LD32(&src2[i*src_stride2+4]);\
        OP(*((uint32_t*)&dst[i*dst_stride+4]), no_rnd_avg32(a, b));\
    }\
}\
\
static inline void OPNAME ## _pixels2_l2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, \
                                                int src_stride1, int src_stride2, int h){\
    int i;\
    for(i=0; i<h; i++){\
        uint32_t a,b;\