Optimize depthwise convolution 3x3 for arm64 platform

mindspore/lite/nnacl/int8/conv_depthwise_int8.c

@@ -208,8 +208,13 @@ void ConvDw3x3Int8Block(int8_t *output, const int8_t *buffer, const int16_t *wei
                         int32_t out_zp, int out_multiplier, int left_shift, int right_shift, int32_t acc_min,
                         int32_t acc_max, int stride) {
   for (; start_c <= end_c - 8; start_c += 8) {
+#ifdef ENABLE_ARM64
+    ConvDw3x3Int8Neon64(output, buffer, weight, bias, col_size, row_size, channel, output_h, output_w, in_zp, out_zp,
+                        out_multiplier, left_shift, right_shift, acc_min, acc_max);
+#else
     ConvDw3x3Int8Window(output, buffer, weight, bias, col_size, row_size, channel, output_h, output_w, in_zp, out_zp,
                         out_multiplier, left_shift, right_shift, acc_min, acc_max, stride);
+#endif
     output += 8;
     buffer += 8;
     weight += 8;

mindspore/lite/nnacl/int8/conv_depthwise_int8.h

@@ -43,6 +43,14 @@ void ConvDwSWInt8(int8_t *output_data, const int8_t *input_data, const int16_t *
 void DeconvDwInt8(int8_t *output_data, int32_t *output_buffer, const int16_t *input_data, const int16_t *weight_data,
                   const int32_t *bias_data, const ConvParameter *conv_param, const SlidingWindowParam *sliding,
                   int task_id);
+
+#ifdef ENABLE_ARM64
+void ConvDw3x3Int8Neon64(int8_t *output, const int8_t *input, const int16_t *weight, const int32_t *bias,
+                         int input_col_size, int input_row_size, int channel, int output_h, int output_w, int8_t in_zp,
+                         int32_t out_zp, int out_multiplier, int left_shift, int right_shift, int32_t acc_min,
+                         int32_t acc_max);
+#endif
+
 #ifdef __cplusplus
 }
 #endif