108     __m128i sum0_r, sum1_r, sum2_r, sum3_r;                              \
122     LSX_BUTTERFLY_4_W(temp0_r, temp0_l, temp1_l, temp1_r, sum0_r, sum0_l,\
126     sum3_r = sum0_r;                                                     \
134     sum0_r  = __lsx_vadd_w(sum0_r, temp2_r);                             \
139     in0 = __lsx_vssrarni_h_w(sum0_l, sum0_r, shift);                     \
161     LSX_BUTTERFLY_4_W(temp0_r, temp0_l, temp1_l, temp1_r, sum0_r, sum0_l,\
165     sum3_r = sum0_r;                                                     \
173     sum0_r  = __lsx_vadd_w(sum0_r, temp2_r);                             \
178     in1 = __lsx_vssrarni_h_w(sum0_l, sum0_r, shift);                     \
204     __m128i sum0_r, sum1_r, sum2_r, sum3_r, sum0_l, sum1_l, sum2_l;          \
216                   src4_r, filter2, src4_l, filter2,  sum0_r, sum0_l,         \
220         DUP4_ARG3(__lsx_vdp2add_w_h, sum0_r, src1_r, filter1, sum0_l,        \
222                   src5_l, filter3, sum0_r, sum0_l, sum2_r, sum2_l);          \
226         sum1_r = sum0_r;                                                     \
239         sum0_r = __lsx_vadd_w(sum0_r, temp0_r);                              \
252         sum0_r = __lsx_vadd_w(sum0_r, temp0_r);                              \
257         LSX_BUTTERFLY_4_W(sum0_r, sum0_l, sum2_l, sum2_r, res0_r, res0_l,    \
277 #define HEVC_EVEN16_CALC(input, sum0_r, sum0_l, load_idx, store_idx)  \  argument
281     tmp1_r = sum0_r;                                                  \
283     sum0_r = __lsx_vadd_w(sum0_r, tmp0_r);                            \
285     __lsx_vst(sum0_r, (input + load_idx * 8), 0);                     \
548     __m128i sum0_r, sum0_l, sum1_r, sum1_l, tmp0_r, tmp0_l, tmp1_r, tmp1_l;  in hevc_idct_8x32_column_lsx()  local
569     sum0_r = __lsx_vdp2_w_h(src0_r, filter0);  in hevc_idct_8x32_column_lsx()
571     sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);  in hevc_idct_8x32_column_lsx()
573     __lsx_vst(sum0_r, tmp_buf_ptr, 0);  in hevc_idct_8x32_column_lsx()
578     sum0_r = __lsx_vdp2_w_h(src0_r, filter0);  in hevc_idct_8x32_column_lsx()
580     sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);  in hevc_idct_8x32_column_lsx()
582     __lsx_vst(sum0_r, tmp_buf_ptr, 32);  in hevc_idct_8x32_column_lsx()
587     sum0_r = __lsx_vdp2_w_h(src0_r, filter0);  in hevc_idct_8x32_column_lsx()
589     sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);  in hevc_idct_8x32_column_lsx()
591     __lsx_vst(sum0_r, tmp_buf_ptr, 64);  in hevc_idct_8x32_column_lsx()
596     sum0_r = __lsx_vdp2_w_h(src0_r, filter0);  in hevc_idct_8x32_column_lsx()
598     sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);  in hevc_idct_8x32_column_lsx()
600     __lsx_vst(sum0_r, tmp_buf_ptr, 96);  in hevc_idct_8x32_column_lsx()
608               src3_r, filter1, src3_l, filter1, sum0_r, sum0_l, tmp1_r, tmp1_l);  in hevc_idct_8x32_column_lsx()
609     sum1_r = __lsx_vsub_w(sum0_r, tmp1_r);  in hevc_idct_8x32_column_lsx()
611     sum0_r = __lsx_vadd_w(sum0_r, tmp1_r);  in hevc_idct_8x32_column_lsx()
614     HEVC_EVEN16_CALC(tmp_buf_ptr, sum0_r, sum0_l, 0, 7);  in hevc_idct_8x32_column_lsx()
621               src3_r, filter1, src3_l, filter1, sum0_r, sum0_l, tmp1_r, tmp1_l);  in hevc_idct_8x32_column_lsx()
622     sum1_r = __lsx_vsub_w(sum0_r, tmp1_r);  in hevc_idct_8x32_column_lsx()
624     sum0_r = __lsx_vadd_w(sum0_r, tmp1_r);  in hevc_idct_8x32_column_lsx()
627     HEVC_EVEN16_CALC(tmp_buf_ptr, sum0_r, sum0_l, 1, 6);  in hevc_idct_8x32_column_lsx()
652         sum0_r = __lsx_vdp2_w_h(src0_r, filter0);  in hevc_idct_8x32_column_lsx()
654         sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);  in hevc_idct_8x32_column_lsx()
656         sum0_r = __lsx_vdp2add_w_h(sum0_r, src2_r, filter2);  in hevc_idct_8x32_column_lsx()
658         sum0_r = __lsx_vdp2add_w_h(sum0_r, src3_r, filter3);  in hevc_idct_8x32_column_lsx()
665         tmp0_r = __lsx_vadd_w(tmp0_r, sum0_r);  in hevc_idct_8x32_column_lsx()
667         tmp1_r = __lsx_vsub_w(tmp1_r, sum0_r);  in hevc_idct_8x32_column_lsx()
713         sum0_r = __lsx_vdp2_w_h(src0_r, filter0);  in hevc_idct_8x32_column_lsx()
715         sum0_r = __lsx_vdp2add_w_h(sum0_r, src1_r, filter1);  in hevc_idct_8x32_column_lsx()
717         sum0_r = __lsx_vdp2add_w_h(sum0_r, src2_r, filter2);  in hevc_idct_8x32_column_lsx()
719         sum0_r = __lsx_vdp2add_w_h(sum0_r, src3_r, filter3);  in hevc_idct_8x32_column_lsx()
721         tmp1_r = sum0_r;  in hevc_idct_8x32_column_lsx()
728         sum0_r = __lsx_vdp2_w_h(src4_r, filter0);  in hevc_idct_8x32_column_lsx()
730         sum0_r = __lsx_vdp2add_w_h(sum0_r, src5_r, filter1);  in hevc_idct_8x32_column_lsx()
732         sum0_r = __lsx_vdp2add_w_h(sum0_r, src6_r, filter2);  in hevc_idct_8x32_column_lsx()
734         sum0_r = __lsx_vdp2add_w_h(sum0_r, src7_r, filter3);  in hevc_idct_8x32_column_lsx()
736         sum0_r = __lsx_vadd_w(sum0_r, tmp1_r);  in hevc_idct_8x32_column_lsx()
743         tmp0_r = __lsx_vadd_w(tmp0_r, sum0_r);  in hevc_idct_8x32_column_lsx()
750         tmp1_r = __lsx_vsub_w(tmp1_r, sum0_r);  in hevc_idct_8x32_column_lsx()