109     __m128i sum0_l, sum1_l, sum2_l, sum3_l;                              \
122     LSX_BUTTERFLY_4_W(temp0_r, temp0_l, temp1_l, temp1_r, sum0_r, sum0_l,\
127     sum3_l = sum0_l;                                                     \
135     sum0_l  = __lsx_vadd_w(sum0_l, temp2_l);                             \
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,\
166     sum3_l = sum0_l;                                                     \
174     sum0_l  = __lsx_vadd_w(sum0_l, temp2_l);                             \
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);          \
227         sum1_l = sum0_l;                                                     \
240         sum0_l = __lsx_vadd_w(sum0_l, temp0_l);                              \
253         sum0_l = __lsx_vadd_w(sum0_l, temp0_l);                              \
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
282     tmp1_l = sum0_l;                                                  \
284     sum0_l = __lsx_vadd_w(sum0_l, tmp0_l);                            \
286     __lsx_vst(sum0_l, (input + load_idx * 8), 16);                    \
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
570     sum0_l = __lsx_vdp2_w_h(src0_l, filter0);  in hevc_idct_8x32_column_lsx()
572     sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);  in hevc_idct_8x32_column_lsx()
574     __lsx_vst(sum0_l, tmp_buf_ptr, 16);  in hevc_idct_8x32_column_lsx()
579     sum0_l = __lsx_vdp2_w_h(src0_l, filter0);  in hevc_idct_8x32_column_lsx()
581     sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);  in hevc_idct_8x32_column_lsx()
583     __lsx_vst(sum0_l, tmp_buf_ptr, 48);  in hevc_idct_8x32_column_lsx()
588     sum0_l = __lsx_vdp2_w_h(src0_l, filter0);  in hevc_idct_8x32_column_lsx()
590     sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);  in hevc_idct_8x32_column_lsx()
592     __lsx_vst(sum0_l, tmp_buf_ptr, 80);  in hevc_idct_8x32_column_lsx()
597     sum0_l = __lsx_vdp2_w_h(src0_l, filter0);  in hevc_idct_8x32_column_lsx()
599     sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);  in hevc_idct_8x32_column_lsx()
601     __lsx_vst(sum0_l, tmp_buf_ptr, 112);  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()
610     sum1_l = __lsx_vsub_w(sum0_l, tmp1_l);  in hevc_idct_8x32_column_lsx()
612     sum0_l = __lsx_vadd_w(sum0_l, tmp1_l);  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()
623     sum1_l = __lsx_vsub_w(sum0_l, tmp1_l);  in hevc_idct_8x32_column_lsx()
625     sum0_l = __lsx_vadd_w(sum0_l, tmp1_l);  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()
653         sum0_l = __lsx_vdp2_w_h(src0_l, filter0);  in hevc_idct_8x32_column_lsx()
655         sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);  in hevc_idct_8x32_column_lsx()
657         sum0_l = __lsx_vdp2add_w_h(sum0_l, src2_l, filter2);  in hevc_idct_8x32_column_lsx()
659         sum0_l = __lsx_vdp2add_w_h(sum0_l, src3_l, filter3);  in hevc_idct_8x32_column_lsx()
666         tmp0_l = __lsx_vadd_w(tmp0_l, sum0_l);  in hevc_idct_8x32_column_lsx()
668         tmp1_l = __lsx_vsub_w(tmp1_l, sum0_l);  in hevc_idct_8x32_column_lsx()
714         sum0_l = __lsx_vdp2_w_h(src0_l, filter0);  in hevc_idct_8x32_column_lsx()
716         sum0_l = __lsx_vdp2add_w_h(sum0_l, src1_l, filter1);  in hevc_idct_8x32_column_lsx()
718         sum0_l = __lsx_vdp2add_w_h(sum0_l, src2_l, filter2);  in hevc_idct_8x32_column_lsx()
720         sum0_l = __lsx_vdp2add_w_h(sum0_l, src3_l, filter3);  in hevc_idct_8x32_column_lsx()
722         tmp1_l = sum0_l;  in hevc_idct_8x32_column_lsx()
729         sum0_l = __lsx_vdp2_w_h(src4_l, filter0);  in hevc_idct_8x32_column_lsx()
731         sum0_l = __lsx_vdp2add_w_h(sum0_l, src5_l, filter1);  in hevc_idct_8x32_column_lsx()
733         sum0_l = __lsx_vdp2add_w_h(sum0_l, src6_l, filter2);  in hevc_idct_8x32_column_lsx()
735         sum0_l = __lsx_vdp2add_w_h(sum0_l, src7_l, filter3);  in hevc_idct_8x32_column_lsx()
737         sum0_l = __lsx_vadd_w(sum0_l, tmp1_l);  in hevc_idct_8x32_column_lsx()
744         tmp0_l = __lsx_vadd_w(tmp0_l, sum0_l);  in hevc_idct_8x32_column_lsx()
751         tmp1_l = __lsx_vsub_w(tmp1_l, sum0_l);  in hevc_idct_8x32_column_lsx()