61   v8i16 in8, in9, in10, in11, in12, in13, in14, in15;  in fdct8x32_1d_column_even_store()  local
71   LD_SH4(input + 64, 8, in8, in9, in10, in11);  in fdct8x32_1d_column_even_store()
72   BUTTERFLY_8(in4, in5, in6, in7, in8, in9, in10, in11, vec4, vec5, vec6, vec7,  in fdct8x32_1d_column_even_store()
73               in8, in9, in10, in11);  in fdct8x32_1d_column_even_store()
104   ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);  in fdct8x32_1d_column_even_store()
118   SUB2(in9, vec2, in14, vec5, vec2, vec5);  in fdct8x32_1d_column_even_store()
257   v8i16 in8, in9, in10, in11, in12, in13, in14, in15;  in fdct8x32_1d_row_load_butterfly()  local
261   LD_SH8(temp_buff + 24, 32, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_load_butterfly()
264   TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9,  in fdct8x32_1d_row_load_butterfly()
266   BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,  in fdct8x32_1d_row_load_butterfly()
268                step6, step7, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_load_butterfly()
270   ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, (output + 24 * 8), 8);  in fdct8x32_1d_row_load_butterfly()
274   LD_SH8(temp_buff + 16, 32, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_load_butterfly()
277   TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9,  in fdct8x32_1d_row_load_butterfly()
279   BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,  in fdct8x32_1d_row_load_butterfly()
281                step6, step7, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_load_butterfly()
284   ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, (output + 16 * 8), 8);  in fdct8x32_1d_row_load_butterfly()
290   v8i16 in8, in9, in10, in11, in12, in13, in14, in15;  in fdct8x32_1d_row_even_4x()  local
299   LD_SH8(input + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_even_4x()
301   BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,  in fdct8x32_1d_row_even_4x()
303                vec7, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_even_4x()
305   ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, interm_ptr + 64, 8);  in fdct8x32_1d_row_even_4x()
360   LD_SH8(interm_ptr + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_even_4x()
363   ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);  in fdct8x32_1d_row_even_4x()
377   SUB2(in9, vec2, in14, vec5, vec2, vec5);  in fdct8x32_1d_row_even_4x()
394   v8i16 in8, in9, in10, in11, in12, in13, in14, in15;  in fdct8x32_1d_row_even()  local
400   LD_SH8(temp + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_even()
402   BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,  in fdct8x32_1d_row_even()
404                vec7, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_even()
435   ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);  in fdct8x32_1d_row_even()
449   SUB2(in9, vec2, in14, vec5, vec2, vec5);  in fdct8x32_1d_row_even()
697   v8i16 in8, in9, in10, in11, in12, in13, in14, in15;  in fdct8x32_1d_row_even_rd()  local
703   LD_SH8(temp + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_even_rd()
705   BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,  in fdct8x32_1d_row_even_rd()
707                vec7, in8, in9, in10, in11, in12, in13, in14, in15);  in fdct8x32_1d_row_even_rd()
712   FDCT_POSTPROC_2V_NEG_H(in8, in9);  in fdct8x32_1d_row_even_rd()
747   ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);  in fdct8x32_1d_row_even_rd()
759   SUB2(in9, vec2, in14, vec5, vec2, vec5);  in fdct8x32_1d_row_even_rd()