Searched refs:S4_CH_IN_BLOCK_MVE (Results 1 – 3 of 3) sorted by relevance
85 const int8_t *lhs_1 = lhs + row_x_col * S4_CH_IN_BLOCK_MVE + offset; in arm_nn_depthwise_conv_nt_t_s4()86 const int8_t *lhs_2 = lhs + (row_x_col * S4_CH_IN_BLOCK_MVE * 2) + offset; in arm_nn_depthwise_conv_nt_t_s4()87 const int8_t *lhs_3 = lhs + (row_x_col * S4_CH_IN_BLOCK_MVE * 3) + offset; in arm_nn_depthwise_conv_nt_t_s4()126 lhs_0 += S4_CH_IN_BLOCK_MVE; in arm_nn_depthwise_conv_nt_t_s4()127 lhs_1 += S4_CH_IN_BLOCK_MVE; in arm_nn_depthwise_conv_nt_t_s4()128 lhs_2 += S4_CH_IN_BLOCK_MVE; in arm_nn_depthwise_conv_nt_t_s4()129 lhs_3 += S4_CH_IN_BLOCK_MVE; in arm_nn_depthwise_conv_nt_t_s4()160 lhs_0 += S4_CH_IN_BLOCK_MVE; in arm_nn_depthwise_conv_nt_t_s4()161 lhs_1 += S4_CH_IN_BLOCK_MVE; in arm_nn_depthwise_conv_nt_t_s4()162 lhs_2 += S4_CH_IN_BLOCK_MVE; in arm_nn_depthwise_conv_nt_t_s4()[all …]
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()106 int32_t active_ch = MIN(S4_CH_IN_BLOCK_MVE, remaining_ch); in arm_depthwise_conv_s4_opt()107 remaining_ch -= S4_CH_IN_BLOCK_MVE; in arm_depthwise_conv_s4_opt()111 out = output + i_ch * S4_CH_IN_BLOCK_MVE; in arm_depthwise_conv_s4_opt()112 const int8_t *input_slice = input + (i_ch * S4_CH_IN_BLOCK_MVE); in arm_depthwise_conv_s4_opt()132 lhs_buffer += S4_CH_IN_BLOCK_MVE; in arm_depthwise_conv_s4_opt()139 const int32_t block_offset = i_ch * S4_CH_IN_BLOCK_MVE; in arm_depthwise_conv_s4_opt()171 for (int i_loop_cnt = 0, offset = i_ch * S4_CH_IN_BLOCK_MVE; i_loop_cnt < loop_count; in arm_depthwise_conv_s4_opt()174 … const int8_t *col_0 = lhs_buffer + (kernel_size * S4_CH_IN_BLOCK_MVE * i_buf) + (i_loop_cnt * 4); in arm_depthwise_conv_s4_opt()210 col_0 += S4_CH_IN_BLOCK_MVE; in arm_depthwise_conv_s4_opt()[all …]
69 #define S4_CH_IN_BLOCK_MVE (124) macro
