65     const int32_t input_ch = input_dims->c;  in arm_depthwise_conv_s4_opt()  local
69     if (input_ch != output_ch)  in arm_depthwise_conv_s4_opt()
104     const int32_t ch_loop = (input_ch + (S4_CH_IN_BLOCK_MVE - 1)) / S4_CH_IN_BLOCK_MVE;  in arm_depthwise_conv_s4_opt()
129                                           input_slice + (i_ker_y * input_x + i_ker_x) * input_ch,  in arm_depthwise_conv_s4_opt()
145                                                   input_ch,  in arm_depthwise_conv_s4_opt()
155                     out += (4 * input_ch);  in arm_depthwise_conv_s4_opt()
170             out = out_base + (i_buf * input_ch);  in arm_depthwise_conv_s4_opt()
182                 if (input_ch % 2)  in arm_depthwise_conv_s4_opt()
211                         row_0 += (input_ch >> 1) + get_low_nibble;  in arm_depthwise_conv_s4_opt()
229                         row_0 += input_ch >> 1;  in arm_depthwise_conv_s4_opt()
278 …              memset(&col_buffer[index], 0, (kernel_x * input_ch) * ker_y_start * sizeof(int16_t));  in arm_depthwise_conv_s4_opt()
279                 index += (kernel_x * input_ch) * ker_y_start;  in arm_depthwise_conv_s4_opt()
291                         memset(&col_buffer[index], 0, input_ch * sizeof(int16_t));  in arm_depthwise_conv_s4_opt()
295 …                  arm_q7_to_q15_with_offset((int8_t *)input + (idx_y * input_x + idx_x) * input_ch,  in arm_depthwise_conv_s4_opt()
297                                                   input_ch,  in arm_depthwise_conv_s4_opt()
300                     index += input_ch;  in arm_depthwise_conv_s4_opt()
307                 memset(&col_buffer[index], 0, (kernel_x * input_ch) * diff * sizeof(int16_t));  in arm_depthwise_conv_s4_opt()
352                         read_and_pad_s4_uneven(row_pos + (input_ch >> 1), &ip_a1, &ip_b2);  in arm_depthwise_conv_s4_opt()
355                         op_b = arm_nn_read_s16x2(col_pos + input_ch);  in arm_depthwise_conv_s4_opt()
367                         op_b = arm_nn_read_s16x2(col_pos + input_ch + 2);  in arm_depthwise_conv_s4_opt()
392                         ker11 = row_pos[1 + (input_ch >> 1)];  in arm_depthwise_conv_s4_opt()
393                         ker0 = row_pos[0 + (input_ch >> 1)] >> 4;  in arm_depthwise_conv_s4_opt()
396                         ker3 = (int8_t)(row_pos[2 + (input_ch >> 1)] << 4) >> 4;  in arm_depthwise_conv_s4_opt()
398                         sum += ker0 * col_pos[0 + input_ch];  in arm_depthwise_conv_s4_opt()
399                         sum_2 += ker1 * col_pos[1 + input_ch];  in arm_depthwise_conv_s4_opt()
400                         sum_3 += ker2 * col_pos[2 + input_ch];  in arm_depthwise_conv_s4_opt()
401                         sum_4 += ker3 * col_pos[3 + input_ch];  in arm_depthwise_conv_s4_opt()
404                         row_pos += (input_ch);  in arm_depthwise_conv_s4_opt()
405                         col_pos += input_ch << 1;  in arm_depthwise_conv_s4_opt()
430                         row_pos += input_ch >> 1;  in arm_depthwise_conv_s4_opt()
431                         col_pos += input_ch;  in arm_depthwise_conv_s4_opt()
480                         int8_t rhs = row_pos[i * (input_ch >> 1) + col_index];  in arm_depthwise_conv_s4_opt()
482                         int16_t lhs0 = col_pos[i * input_ch];  in arm_depthwise_conv_s4_opt()
552                         read_and_pad_s4(row_pos + (input_ch >> 1), &ip_b2, &ip_a1);  in arm_depthwise_conv_s4_opt()
555                         op_b = arm_nn_read_s16x2(col_pos + input_ch);  in arm_depthwise_conv_s4_opt()
567                         op_b = arm_nn_read_s16x2(col_pos + input_ch + 2);  in arm_depthwise_conv_s4_opt()
577                         row_pos += (input_ch);  in arm_depthwise_conv_s4_opt()
578                         col_pos += input_ch << 1;  in arm_depthwise_conv_s4_opt()
605                         row_pos += input_ch >> 1;  in arm_depthwise_conv_s4_opt()
606                         col_pos += input_ch;  in arm_depthwise_conv_s4_opt()
653                         int8_t rhs = row_pos[i * (input_ch >> 1)];  in arm_depthwise_conv_s4_opt()
658                         int16_t lhs0 = col_pos[i * input_ch];  in arm_depthwise_conv_s4_opt()
659                         int16_t lhs1 = col_pos_2[i * input_ch];  in arm_depthwise_conv_s4_opt()