80         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
82               src_lines_intertwined_4x[2 * cell + outer].val[inner]);  in Pack()
86         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
88               src_lines_intertwined_4x[2 * cell + outer].val[inner]);  in Pack()
99         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
100           sums_of_2_cells[cell][i] = vaddl_u8(  in Pack()
102                   src_lines_intertwined_4x[2 * cell + outer].val[inner]),  in Pack()
104                   src_lines_intertwined_4x[2 * cell + outer].val[inner]));  in Pack()
110       for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
111         sums_of_4_cells[cell][i] = vreinterpretq_s32_u32(  in Pack()
112             vaddl_u16(vget_low_u16(sums_of_2_cells[cell][i]),  in Pack()
113                       vget_high_u16(sums_of_2_cells[cell][i])));  in Pack()
117     for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
119           vaddq_s32(sums_of_4_cells[cell][0], sums_of_4_cells[cell][1]);  in Pack()
121           vaddq_s32(sums_of_4_cells[cell][2], sums_of_4_cells[cell][3]);  in Pack()
124           dst->sums_of_each_slice() + start_width + 4 * cell;  in Pack()
195         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
197               src_lines_intertwined_4x[2 * cell + outer].val[inner]));  in Pack()
201         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
203               src_lines_intertwined_4x[2 * cell + outer].val[inner]));  in Pack()
214         for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
215           sums_of_2[cell][i] = vpaddlq_u8(vreinterpretq_u8_u16(  in Pack()
216               src_lines_intertwined_4x[2 * cell + outer].val[inner]));  in Pack()
222       for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
223         sums_of_4[cell][i] =  in Pack()
224             vaddq_u16(sums_of_2[cell][2 * i], sums_of_2[cell][2 * i + 1]);  in Pack()
228     for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
229       sums_of_8[cell] = vaddq_u16(sums_of_4[cell][0], sums_of_4[cell][1]);  in Pack()
233     for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
234       sums_of_16[cell] = vadd_u16(vget_low_u16(sums_of_8[cell]),  in Pack()
235                                   vget_high_u16(sums_of_8[cell]));  in Pack()
238     for (int cell = 0; cell < kCells; cell++) {  in Pack()  local
239       int32x4_t s = vreinterpretq_s32_u32(vmovl_u16(sums_of_16[cell]));  in Pack()
241           dst->sums_of_each_slice() + start_width + 4 * cell;  in Pack()