static void MC_put_o_8_c (uint8_t *dest, const uint8_t *ref, const int stride, int height)
{
uint32_t ref_aln, ref_rs;
ref_aln = (uint32_t)(ref-stride) & 0xfffffffc;
ref_rs = 4 - ((uint32_t)(ref-stride) & 3);
dest -= stride;
do {
S32LDIV(xr1,ref_aln,stride,0x0);
S32LDD(xr2,ref_aln,0x4);
S32LDD(xr4,ref_aln,0x8);
S32ALN(xr3,xr2,xr1,ref_rs);
S32ALN(xr5,xr4,xr2,ref_rs);
S32SDIV(xr3,dest,stride,0x0);
S32STD(xr5,dest,0x4);
} while (--height);
}
void ff_init_cabac_states_mxu(CABACContext *c){
uint32_t i = 0;
uint8_t *p_lps_range = lps_range[0]-4;
uint8_t *p_mps_state = mps_state-4;
uint8_t *p_lps_state = lps_state-4;
uint8_t *p_ff_h264_mlps_state1 = &ff_h264_mlps_state[0]+124;
uint8_t *p_ff_h264_mlps_state2 = &ff_h264_mlps_state[0]+128;//
uint8_t *p_ff_h264_lps_range = ff_h264_lps_range-128;
S32I2M(xr15,0x1010101);
for(i=0; i<16; i++){
/* load 32 */
S32LDI(xr1,p_lps_range,4);
S32LDI(xr2,p_lps_range,4);
S32LDI(xr3,p_lps_range,4);
S32LDI(xr4,p_lps_range,4);
S32SFL(xr7,xr2,xr1,xr8,0);
S32SFL(xr9,xr4,xr3,xr10,0);
S32SFL(xr11,xr7,xr7,xr12,0);
S32SFL(xr13,xr8,xr8,xr14,0);
S32SFL(xr1,xr9,xr9,xr2,0);
S32SFL(xr3,xr10,xr10,xr4,0);
S32SDI(xr14,p_ff_h264_lps_range,128);
S32SDI(xr4,p_ff_h264_lps_range,4);
S32SDI(xr13,p_ff_h264_lps_range,124);
S32SDI(xr3,p_ff_h264_lps_range,4);
S32SDI(xr12,p_ff_h264_lps_range,124);
S32SDI(xr2,p_ff_h264_lps_range,4);
S32SDI(xr11,p_ff_h264_lps_range,124);
S32SDI(xr1,p_ff_h264_lps_range,4);
p_ff_h264_lps_range-=128*4-4;
/* part2 */
S32LDI(xr1,p_mps_state,4);
Q8ADD_AA(xr3,xr1,xr1);// 2*mps_state[i]+0
Q8ADD_AA(xr4,xr3,xr15);//2*mps_state[i]+1
S32SFL(xr11,xr4,xr3,xr12,0);
S32SDI(xr12,p_ff_h264_mlps_state1,4);// ff_h264_mlps_state[128+2*i+1]= 2*mps_state[i]+0;
S32SDI(xr11,p_ff_h264_mlps_state1,4);
/*part3 */
S32LDI(xr1,p_lps_state,4);
Q8ADD_AA(xr3,xr1,xr1);// 2*mps_state[i]+0
Q8ADD_AA(xr4,xr3,xr15);//2*mps_state[i]+1
S32SFL(xr11,xr3,xr4,xr12,0);
S32ALN(xr11,xr11,xr11,2);
S32ALN(xr12,xr12,xr12,2);
S32SDI(xr12,p_ff_h264_mlps_state2,-4);
S32SDI(xr11,p_ff_h264_mlps_state2,-4);
}
ff_h264_mlps_state[127]= 1;
ff_h264_mlps_state[126]= 0;
}