77         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
79               src_lines_intertwined_4x[2 * cell + outer].val[inner]);  in Pack()
83         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
85               src_lines_intertwined_4x[2 * cell + outer].val[inner]);  in Pack()
96         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
97           sums_of_2_cells[cell][i] = vaddl_u8(  in Pack()
99                   src_lines_intertwined_4x[2 * cell + outer].val[inner]),  in Pack()
101                   src_lines_intertwined_4x[2 * cell + outer].val[inner]));  in Pack()
107       for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
108         sums_of_4_cells[cell][i] = vreinterpretq_s32_u32(  in Pack()
109             vaddl_u16(vget_low_u16(sums_of_2_cells[cell][i]),  in Pack()
110                       vget_high_u16(sums_of_2_cells[cell][i])));  in Pack()
114     for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
116           vaddq_s32(sums_of_4_cells[cell][0], sums_of_4_cells[cell][1]);  in Pack()
118           vaddq_s32(sums_of_4_cells[cell][2], sums_of_4_cells[cell][3]);  in Pack()
121           dst->sums_of_each_slice() + start_width + 4 * cell;  in Pack()
192         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
194               src_lines_intertwined_4x[2 * cell + outer].val[inner]));  in Pack()
198         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
200               src_lines_intertwined_4x[2 * cell + outer].val[inner]));  in Pack()
211         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
212           sums_of_2[cell][i] = vpaddlq_u8(vreinterpretq_u8_u16(  in Pack()
213               src_lines_intertwined_4x[2 * cell + outer].val[inner]));  in Pack()
219       for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
220         sums_of_4[cell][i] =  in Pack()
221             vaddq_u16(sums_of_2[cell][2 * i], sums_of_2[cell][2 * i + 1]);  in Pack()
225     for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
226       sums_of_8[cell] = vaddq_u16(sums_of_4[cell][0], sums_of_4[cell][1]);  in Pack()
230     for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
231       sums_of_16[cell] = vadd_u16(vget_low_u16(sums_of_8[cell]),  in Pack()
232                                   vget_high_u16(sums_of_8[cell]));  in Pack()
235     for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
236       int32x4_t s = vreinterpretq_s32_u32(vmovl_u16(sums_of_16[cell]));  in Pack()
238           dst->sums_of_each_slice() + start_width + 4 * cell;  in Pack()