72                 for (int mult_tile = 0; mult_tile < ch_mult; mult_tile += 4)  in depthwise_conv_s16_mult_4_s16()  local
74                     int32_t out_buff32[4] = {REDUCE_MULTIPLIER(output_mult[out_ch + 0 + mult_tile]),  in depthwise_conv_s16_mult_4_s16()
75                                              REDUCE_MULTIPLIER(output_mult[out_ch + 1 + mult_tile]),  in depthwise_conv_s16_mult_4_s16()
76                                              REDUCE_MULTIPLIER(output_mult[out_ch + 2 + mult_tile]),  in depthwise_conv_s16_mult_4_s16()
77 …                                           REDUCE_MULTIPLIER(output_mult[out_ch + 3 + mult_tile])};  in depthwise_conv_s16_mult_4_s16()
83                         out_buff[0] = bias[out_ch + 0 + mult_tile];  in depthwise_conv_s16_mult_4_s16()
84                         out_buff[1] = bias[out_ch + 1 + mult_tile];  in depthwise_conv_s16_mult_4_s16()
85                         out_buff[2] = bias[out_ch + 2 + mult_tile];  in depthwise_conv_s16_mult_4_s16()
86                         out_buff[3] = bias[out_ch + 3 + mult_tile];  in depthwise_conv_s16_mult_4_s16()
102                             out_buff[0] += in_val * kernel[ker_idx + 0 + mult_tile];  in depthwise_conv_s16_mult_4_s16()
103                             out_buff[1] += in_val * kernel[ker_idx + 1 + mult_tile];  in depthwise_conv_s16_mult_4_s16()
104                             out_buff[2] += in_val * kernel[ker_idx + 2 + mult_tile];  in depthwise_conv_s16_mult_4_s16()
105                             out_buff[3] += in_val * kernel[ker_idx + 3 + mult_tile];  in depthwise_conv_s16_mult_4_s16()
110 …           arm_nn_requantize_s64(out_buff[0], out_buff32[0], output_shift[out_ch + 0 + mult_tile]);  in depthwise_conv_s16_mult_4_s16()
112 …           arm_nn_requantize_s64(out_buff[1], out_buff32[1], output_shift[out_ch + 1 + mult_tile]);  in depthwise_conv_s16_mult_4_s16()
114 …           arm_nn_requantize_s64(out_buff[2], out_buff32[2], output_shift[out_ch + 2 + mult_tile]);  in depthwise_conv_s16_mult_4_s16()
116 …           arm_nn_requantize_s64(out_buff[3], out_buff32[3], output_shift[out_ch + 3 + mult_tile]);  in depthwise_conv_s16_mult_4_s16()