Ascend Conversion Tool Description

View Source On Gitee

Introduction

This article introduces the related features of the cloud-side inference model conversion tool in Ascend back-end, such as profile options, dynamic shape, AOE, custom operators.

Configuration File

Table 1: Configure [ascend_context] parameter

Parameters

Attributes

Functions Description

Types

Values Description

input_format

Optional

Specify the model input format.

String

Options: "NCHW", "NHWC", and "ND"

input_shape

Optional

Specify the model input Shape. input_name must be the input name in the network model before conversion, in the order of the inputs, separated by ;. Only works for dynamic BatchSize. For static BatchSize, use converter_lite command to specify inputShape parameter.

String

Such as "input1:[1,64,64,3];input2:[1,256,256,3]"

dynamic_dims

Optional

Specify the dynamic BatchSize and dynamic resolution parameters.

String

见Dynamic shape configuration

precision_mode

Optional

Configure the model accuracy mode.

String

Options: "enforce_fp32", "preferred_fp32", "enforce_fp16", "enforce_origin" or "preferred_optimal". Default: "enforce_fp16"

op_select_impl_mode

Optional

Configure the operator selection mode.

String

Optioans: "high_performance", and "high_precision". Default: "high_performance".

output_type

Optional

Specify the data type of network output

String

Options: "FP16", "FP32", "UINT8"

fusion_switch_config_file_path

Optional

Configure the Fusion Switch Configuration File file path and file name.

String

Specify the configuration file for the fusion switch

insert_op_config_file_path

Optional

Model insertion AIPP operator

String

Path of AIPP configuration file

aoe_mode

Optional

AOE auto-tuning mode

String

Options: "subgraph tuning", "operator tuning" or "subgraph tuning, operator tuning". Default: Not enabled

plugin_custom_ops

Optional

Enable Ascend backend fusion optimization to generate custom operators

String

The available options are All, None, FlashAttention, LayerNormV3, GeGluV2, GroupNormSilu, FFN, AddLayerNorm, MatMulAllReduce and BatchMatmulToMatmul, where All means enabling FlashAttention, LayerNormV3, GeGluV2 and GroupNormSilu, default None means not enabled

custom_fusion_pattern

Optional

Specify custom operator structures in the enabling model

String

custom operator type: original operator name in the model: enabled or disabled, which can be taken as enable or disable

op-attrs

Optional

Specify custom operator attributes for fusion

String

Custom operator name:Attribute:Value. Currently, the operator only supports FlashAttention, which supports three optional configuration attributes: input_layout, seq_threshold, inner_precise, which respectively determine whether FlashAttention is fused in the form of BNSD (default), BSH or BNSD_BSND(BNSD means FlashAttention input and output layout is BNSD, BSH means input and output are BSH, BNSD_BSND means input is BNSD, output is BSND), the seq threshold (default 0), and high-performance (default) or high-precision for FlashAttention fusion

Table 2: Configure [acl_init_options] parameter

Parameters

Attributes

Functions Description

Types

Values Description

ge.engineType

Optional

Set the core type used by the network model.

String

Options: "VectorCore", "AiCore"

ge.socVersion

Optional

Version of the Ascend AI processor.

String

Options: "Ascend310", "Ascend710", "Ascend910"

ge.bufferOptimize

Optional

Data cache optimization switch.

String

Options: "l1_optimize", "l2_optimize", "off_optimize". Default: "l2_optimize"

ge.enableCompressWeight

Optional

Data compression can be performed on Weight to improve performance.

String

Options: "true", "false"

compress_weight_conf

Optional

The path of the configuration file for the node list to be compressed is mainly composed of the conv operator and the fc operator.

String

path of config file

ge.exec.precision_mode

Optional

Select the operator precision mode.

String

Options: "force_fp32", "force_fp16", "allow_fp32_to_fp16", "must_keep_origin_dtype", "allow_mix_precision", Default: "force_fp16"

ge.exec.disableReuseMemory

Optional

Memory reuse switch.

String

Options: "0", "１"

ge.enableSingleStream

Optional

Whether to enable a model to use only one stream.

String

Options: "true", "false"

ge.aicoreNum

Optional

Set the number of AI cores used during compilation.

String

Default: "１0"

ge.fusionSwitchFile

Optional

Fusion configuration file path.

String

path of config file

ge.enableSmallChannel

Optional

Whether to enable the optimization of small channel.

String

Options: "0", "1"

ge.opSelectImplmode

Optional

Select the operator implementation mode.

String

Options: "high_precision", "high_performance"

ge.optypelistForImplmode

Optional

A list of operators that uses the mode specified by the ge.opSelectImplmode parameter.

String

Operator type

ge.op_compiler_cache_mode

Optional

Configure operator to compile Disk buffer mode.

String

Options: "enable", "force", "disable"

ge.op_compiler_cache_dir

Optional

Configure operator mutation Disk buffer directory.

String

Options: $HOME/atc_data

ge.debugDir

Optional

Configure the path to save debugging related process files generated by operator compilation.

String

Default Generate Current Path

ge.opDebugLevel

Optional

Operator debug function switch.

String

Options: "0", "1"

ge.exec.modify_mixlist

Optional

Configure a mixed precision list.

String

path of config file

ge.enableSparseMatrixWeight

Optional

Enable global sparsity characteristics.

String

Options: "1", "0"

ge.externalWeight

Optional

Whether to save the weights of constant nodes separately in a file.

String

Options: "１", "0"

ge.deterministic

Optional

Whether to enable deterministic calculation.

String

Options: "１", "0"

ge.host_env_os

Optional

Support for inconsistency between the compilation environment operating system and the runtime environment.

String

Options: "linux"

ge.host_env_cpu

Optional

Support for inconsistency between the operating system architecture and the runtime environment in the compilation environment.

String

Options: "aarch64", "x86_64"

ge.virtual_type

Optional

Whether the offline model is supported to run on virtual devices generated by the Ascend virtualization instance feature.

String

Options: "0", "1"

ge.compressionOptimizeConf

Optional

Compression optimization feature configuration file path.

String

path of config file

Table 3: Configure [acl_build_options] parameter

Parameters

Attributes

Functions Description

Types

Values Description

input_format

Optional

Specify the model input format.

String

Options: "NCHW", "NHWC", "ND"

input_shape

Optional

Specify the model input shape. After the model is converted, it can be obtained using the Model.get_model_info ("input_shape") method. This parameter is consistent with the command line input_shape.

String

For example: input1:1,3,512,512;input2:1,3,224,224

input_shape_rang

Optional

Specify the model shape rang.

String

For example: input1:[1-10,3,512,512];input2:[1-10,3,224,224]

op_name_map

Optional

Extension operator mapping configuration file path.

String

path of config file

ge.dynamicBatchSize

Optional

Set dynamic batch gear parameters.

String

This parameter needs to be used in conjunction with the input_shape parameter

ge.dynamicImageSize

Optional

Set dynamic resolution parameters for input images.

String

This parameter needs to be used in conjunction with the input_shape parameter

ge.dynamicDims

Optional

Set the gear of dynamic dimensions in ND format. After the model is converted, it can be obtained using the Model.get_model_info ("dynamic_dims") method

String

This parameter needs to be used in conjunction with the input_shape parameter

ge.inserOpFile

Optional

Enter the configuration file path for the preprocessing operator.

String

path of config file

ge.exec.precision_mode

Optional

Enter the configuration file path for the preprocessing operator.

String

Options: "force_fp32", "force_fp16", "allow_fp32_to_fp16", "must_keep_origin_dtype", "allow_mix_precision"，Default: "force_fp16"

ge.exec.disableReuseMemory

Optional

Memory reuse switch.

String

Options: "0", "１"

ge.outputDataType

Optional

Network output data type.

String

Options: "FP32", "UINT8", "FP16"

ge.outputNodeName

Optional

Specify output nodes.

String

For example: "node_name1:0;node_name1:1;node_name2:0"

ge.INPUT_NODES_SET_FP16

Optional

Specify the input node name with input data type FP16.

String

"node_name1;node_name2"

log

Optional

Set Log Level.

String

Options: "debug", "info", "warning", "error"

ge.op_compiler_cache_mode

Optional

Configure operator to compile disk buffer mode.

String

Options: "enable", "force", "disable"

ge.op_compiler_cache_dir

Optional

Configure operator mutation disk buffer directory.

String

Options: $HOME/atc_data

ge.debugDir

Optional

Configure the path to save debugging related process files generated by operator compilation.

String

Default Generate Current Path

ge.opDebugLevel

Optional

Operator debug function switch.

String

Options: "0", "1"

ge.mdl_bank_path

Optional

Path to custom knowledge base after loading model tuning.

String

This parameter needs to be used in conjunction with the ge.bufferOptimize parameter

ge.op_bank_path

Optional

Customizing the knowledge base path after loading operator tuning.

String

Knowledge Base Path

ge.exec.modify_mixlist

Optional

Configure mixed precision list.

String

path of config file

ge.exec.op_precision_mode

Optional

Set the precision mode of a specific operator and use this parameter to set the configuration file path.

String

path of config file

ge.shape_generalized_build_mode

Optional

Shape compilation method during image compilation.

String

Options: "shape_generalized", "shape_precise"

op_debug_config

Optional

Memory detection function switch.

String

path of config file

ge.externalWeight

Optional

Do you want to save the weights of constant nodes separately in a file.

String

Options: "１", "0"

ge.exec.exclude_engines

Optional

Set the network model not to use one or some acceleration engines.

String

Options: "AiCore", "AiVec", "AiCpu"

Dynamic Shape Configuration

In some inference scenarios, such as detecting a target and then executing the target recognition network, the number of targets is not fixed resulting in a variable input BatchSize for the target recognition network. If each inference is computed at the maximum BatchSize or maximum resolution, it will result in wasted computational resources. Therefore, it needs to support dynamic BatchSize and dynamic resolution scenarios during inference. Lite inference on Ascend supports dynamic BatchSize and dynamic resolution scenarios. The dynamic_dims dynamic parameter in [ascend_context] is configured via configFile in the convert phase, and the model Resize is used during inference, to change the input shape.

Dynamic Batch Size

  • Parameter Name

    dynamic_dims

  • Functions

    Set the dynamic batch profile parameter for scenarios where the number of images processed at a time is not fixed during inference. This parameter needs to be used in conjunction with input_shape, and the position of -1 in input_shape is the dimension where the dynamic batch is located.

  • Value

    Support up to 100 profiles configuration. Each profile is separated by English comma. The value limit of each profile: [1~2048]. For example, the parameters in the configuration file are configured as follows:

    [ascend_context]
input_shape=input:[-1,64,64,3]
dynamic_dims=[1],[2]

    "-1" in input_shape means setting dynamic batch, and the profile can take the value of "1,2", that is, support profile 0: [1,64,64,3], profile 1: [2,64,64,3].

    If more than one input exists, the profiles corresponding to the different inputs needs to be the same and separated by ;.

    [ascend_context]
input_shape=input1:[-1,64,64,3];input2:[-1,256,256,3]
dynamic_dims=[1],[2];[1],[2]

  • converter

    ./converter_lite --fmk=ONNX --modelFile=${model_name}.onnx --configFile=./config.txt --optimize=ascend_oriented --outputFile=${model_name}

    Note: When enabling dynamic BatchSize, you do not need to specify the inputShape parameter, and only need to configure the [ascend_context] dynamic batch size through configFile, that is, the configuration content in the previous section.

  • Inference

    Enable dynamic BatchSize. When the model inference is performed, the input shape can only choose the set value of the profile at the time of the converter. If you want to switch to the input shape corresponding to another profile, use the model Resize function.

  • Precautions

    1. If the user performs inference operations and the number of images processed is not fixed at a time, this parameter can be configured to dynamically allocate the number of images processed at a time. For example, if a user needs to process 2, 4, or 8 images each time to perform inference, it can be configured as 2, 4, and 8. Once the profile is requested, memory will be requested based on the actual profile during model inference.

    2. If the profile value set by the user is too large or the profiles are too many, it may cause model compilation failure, in which case the user is advised to reduce the profiles or turn down the profile value.

    3. If the profile value set by the user is too large or the profiles are too many, when performing inference in the runtime environment, it is recommended that the swapoff -a command be executed to turn off the swap interval as memory, to avoid that the swap space is continued to be called as memory, resulting in an unusually slow running environment due to the lack of memory.

Dynamic Resolution

  • Parameter Name

    dynamic_dims

  • Function

    Set the dynamic resolution parameter of the input image. For scenarios where the width and height of the image are not fixed each time during inference. This parameter needs to be used in conjunction with input_shape, and the position of -1 in input_shape is the dimension where the dynamic resolution is located.

  • Value

    Support up to 100 profiles configuration. Each profile is separated by English comma, such as "[imagesize1_height,imagesize1_width],[imagesize2_height,imagesize2_width]". For example, the parameters in the configuration file are configured as follows:

    [ascend_context]
input_format=NHWC
input_shape=input:[1,-1,-1,3]
dynamic_dims=[64,64],[19200,960]

    "-1" in input_shape means setting the dynamic resolution, i.e., it supports profile 0: [1,64,64,3] and profile 1: [1,19200,960,3].

  • converter

    ./converter_lite --fmk=ONNX --modelFile=${model_name}.onnx --configFile=./config.txt --optimize=ascend_oriented --outputFile=${model_name}

    Note: When enabling dynamic BatchSize, you do not need to specify the inputShape parameter, and only need to configure the [ascend_context] dynamic resolution through configFile, that is, the configuration content in the previous section.

  • Inference

    By enabling dynamic resolution, when model inference is performed, the input shape can only select the set profile value at the time of the converter. If you want to switch to the input shape corresponding to another profile, use the model Resize function.

  • Precautions

    1. If the resolution value set by the user is too large or the profiles are too many, it may cause model compilation failure, in which case the user is advised to reduce the profiles or turn down the profile value.

    2. If the user sets a dynamic resolution, the size of the dataset images used for the actual inference needs to match the specific resolution used.

    3. If the resolution value set by the user is too large or the profiles are too many, when performing inference in the runtime environment, it is recommended that the swapoff -a command be executed to turn off the swap interval as memory, to avoid that the swap space is continued to be called as memory, resulting in an unusually slow running environment due to the lack of memory.

Dynamic dimension

  • Parameter Name

    ge.dynamicDims

  • Function

    Set the gear of the dynamic dimension input in ND format. Applicable to scenarios where any dimension is processed each time reasoning is performed, This parameter needs to be used in conjunction with input_shape, and the position of -1 in input_shape is the dimension where the dynamic dim is located.

  • Value

    Up to 100 configurations are supported, each separated by an English comma. For example, the parameters in the configuration file are configured as follows:

    [acl_build_options]
input_format="ND"
input_shape="input1:1,-1,-1;input2:1,-1"
ge.dynamicDims="32,32,24;64,64,36"

    The "-1" in the shape indicates the setting of dynamic dimensions, which supports gear 0: input1:1,32,32; input2:1,24, gear 1:1, 64,64; input2:1,36.

  • converter

    ./converter_lite --fmk=ONNX --modelFile=${model_name}.onnx --configFile=./config.txt --optimize=ascend_oriented --outputFile=${model_name}

    Note: When enabling dynamic dimension, input_format must be set to ND.

  • Inference

    By enabling dynamic dimension, when model inference is performed, the input shape can only select the set profile value at the time of the converter. If you want to switch to the input shape corresponding to another profile, use the model Resize function.

  • Precautions

    1. If the resolution value set by the user is too large or the profiles are too many, it may cause model compilation failure, in which case the user is advised to reduce the profiles or turn down the profile value.

    2. If the user sets a dynamic dimension, the dimension of the inputs for the actual inference needs to match the specific resolution used.

    3. If the resolution value set by the user is too large or the profiles are too many, when performing inference in the runtime environment, it is recommended that the swapoff -a command be executed to turn off the swap interval as memory, to avoid that the swap space is continued to be called as memory, resulting in an unusually slow running environment due to the lack of memory.

AOE Auto-tuning

AOE is a computational graph performance auto-tuning tool built specifically for the Davinci platform. Lite enables AOE ability to integrate the AOE offline executable in the converter phase, to perform performance tuning of the graph, generate a knowledge base, and save the offline model. This function supports subgraph tuning and operator tuning. The function supports subgraph tuning and operator tuning. The specific use process is as follows:

AOE Tool Tuning

  1. Configure environment variables

    ${LOCAL_ASCEND} is the path where the Ascend package is installed

    export LOCAL_ASCEND=/usr/local/Ascend
source ${LOCAL_ASCEND}/latest/bin/setenv.bash

    Confirm that the AOE executable program can be found and run in the environment:

    aoe -h

  2. Specify the knowledge base path

    AOE tuning generates an operator knowledge base. The default path:

    ${HOME}/Ascend/latest/data/aoe/custom/graph(op)/${soc_version}

    (Optional) You can also customize the knowledge base path with the export TUNE_BANK_PATH environment variable.

  3. Clear the cache

    In order for the model compilation to get the knowledge base generated by AOE, it is best to delete the compilation cache before AOE is enabled to avoid cache reuse. Taking Atlas inference series environment with user as root for example, delete /root/atc_data/kernel_cache/Ascend310P3 and /root/atc_data/fuzzy_kernel_cache/Ascend310P3 directories.

  4. Specified options of the configuration file

    Specify the AOE tuning mode in the [ascend_context] configuration file of the conversion tool config. In the following example, the subgraph tuning will be executed first, and then the operator tuning.

    [ascend_context]
aoe_mode="subgraph tuning, operator tuning"

  • The performance improvements will vary from environment to environment, and the actual latency reduction percentage is not exactly the same as the results shown in the tuning logs.

  • AOE tuning generates aoe_workspace directory in the current directory where the task is executed, which is used to save the models before and after tuning for performance improvement comparison, as well as the process data and result files necessary for tuning. This directory will occupy additional disk space, e.g., 2~10GB for a 500MB raw model, depending on the model size, operator type structure, input shape size and other factors. Therefore, it is recommended to reserve enough disk space, otherwise it may lead to tuning failure.

  • The aoe_workspace directory needs to be deleted manually to free up disk space.

AOE API Tuning

For Ascend inference, when the runtime specifies provider as ge, multiple models within one device can share weights, and some the weights in the model can be updated, that is, variables. Currently, only AOE API tuning supports variables exists in the model, and the default AOE tool tuning does not support that. The environment variables, setting and use of knowledge base paths, and AOE tuning cache are consistent with AOE tool tuning in the previous section. For details, please refer to AOE tuning.

AOE API tuning needs to be done through converter tool. When optimize=ascend_oriented, in the configuration file, there is provider=ge in [ascend_context], and there is a valid aoe_mode in [ascend_context] or acl_option_cfg_param], or there is a valid job_type in [aoe_global_options], AOE API tuning will be performed. AOE API tuning only generates a knowledge base and does not generate an optimized model.

  1. Specify provider as ge

    [ascend_context]
provider=ge

  2. AOE options

    The options in [aoe_global_options] will be passed through to the global options of the AOE API. The options in [aoe_tuning_options] will be passed through to the tuning options of the AOE API.

    We will extract the options in sections [acl_option_cfg_param], [ascend_context], [ge_session_options] and [ge_graph_options] and convert them into AOE options to avoid the need for users to manually convert these options. The extracted options include input_format, input_shape, dynamic_dims and precision_mode. When the same option exists in multiple configuration sections at the same time, the priority ranges from low to high, with options in [aoe_global_options] and [aoe_tuning_options] having the highest priority. It is recommended to use [ge_graph_options] and aoe_uning_options.

  3. AOE tuning mode

    The aoe_mode is currently limited to subgraph tuning or operator tuning. Currently, subgraph tuning, operator tuning is not supported, which means that subgraph and operator tuning is not supported in the same tuning process. If necessary, subgraph and operator tuning can be performed separately.

    In [aoe_global_options], when the value of job_type is 1, it means subgraph tuning, and when the value is 2, it means operator tuning.

    [ascend_context]
aoe_mode="operator tuning"

    [acl_option_cfg_param]
aoe_mode="operator tuning"

    [aoe_global_options]
job_type=2

  4. Dynamic dimension profiles

    Dynamic dimension profiles can be set in [acl_option_cfg_param], [ascend_context], [ge_graph_options], [aoe_tuning_options], with priority ranging from low to high. The following settings are equivalent. Setting the dynamic dimension profiles in [ascend_context] can refer to Dynamic Shape Configuration. Setting the dynamic dimension profiles in [acl_option_cfg_param], [ge_graph_options] and [aoe_tuning_options] can refer to dynamic_dims, dynamic_batch_size, dynamic_image_size. Note that the [ge_graph_options] only supports the ge.dynamicDims and does not support the forms of dynamic_batch_size and dynamic_image_size. input_format is used to specify the input dimension layout for dynamic profiles. When using dynamic_image_size, it is necessary to specify input_format as NCHW or NHWC to indicate the location of the H and W dimensions.

    [ascend_context]
input_shape=x1:[-1,3,224,224];x2:[-1,3,1024,1024]
dynamic_dims=[1],[2],[3],[4];[1],[2],[3],[4]

    [acl_option_cfg_param]
input_shape=x1:-1,3,224,224;x2:-1,3,1024,1024
dynamic_dims=1,1;2,2;3,3;4,4

    [ge_graph_options]
ge.inputShape=x1:-1,3,224,224;x2:-1,3,1024,1024
ge.dynamicDims=1,1;2,2;3,3;4,4

    [aoe_tuning_options]
input_shape=x1:-1,3,224,224;x2:-1,3,1024,1024
dynamic_dims=1,1;2,2;3,3;4,4

  5. Precision mode

    Precision mode can be set in [acl_option_cfg_param], [ascend_context], [ge_graph_options], [aoe_tuning_options], with priority ranging from low to high. The following settings are equivalent. Setting the precision mode in [ascend_context] and [acl_option_cfg_param] can refer to ascend_context - precision_mode. Setting the precision mode in [ge_graph_options] and [aoe_tuning_options] can refer to precision_mode.

    [ascend_context]
precision_mode=preferred_fp32

    [acl_option_cfg_param]
precision_mode=preferred_fp32

    [ge_graph_options]
precision_mode=allow_fp32_to_fp16

    [aoe_tuning_options]
precision_mode=allow_fp32_to_fp16

Deploying Ascend Custom Operators

MindSpore Lite converter supports converting models with MindSpore Lite custom Ascend operators to MindSpore Lite models. Custom operators can be used to optimize model inference performance in special scenarios, such as using custom MatMul to achieve higher matrix multiplication, using the transformer fusion operators provided by MindSpore Lite to improve transformer model performance (to be launched) and using the AKG graph fusion operator to automatically fuse models to improve inference performance.

If MindSpore Lite converts Ascend models with custom operators, user needs to deploy the custom operators to the ACL operator library before calling the converter in order to complete the conversion properly. The following describes the key steps to deploy Ascend custom operators:

  1. Configure environment variables

    ${ASCEND_OPP_PATH} is the operator library path of Ascend software CANN package, usually under Ascend software installation path. The default is usually /usr/local/Ascend/latest/opp.

    export ASCEND_OPP_PATH=/usr/local/Ascend/latest/opp

  2. Obtain Ascend custom operator package

    MindSpore Lite cloud-side inference package will contain Ascend custom operator package directory whose relative directory is ${LITE_PACKAGE_PATH}/tools/custom_kernels/ascend. After unzip the Mindspore Lite cloud-side inference package, enter the corresponding directory.

    tar zxf mindspore-lite-{version}-linux-{arch}.tar.gz
cd tools/custom_kernels/ascend

  3. Run install.sh script to deploy custom operator

    Run the installation script in the operator package directory to deploy the custom operator.

    bash install.sh

  4. Check the Ascend library directory to see if the installation is successful

    After deploying the custom operator, go to the Ascend operator library directory /usr/local/Ascend/latest/opp/vendors/ and check whether there are corresponding custom operator files in the directory. At present, we mainly provide the basic operator sample and the AKG graph fusion operator implementation. The specific file structure is as follows:

    /usr/local/Ascend/latest/opp/vendors/
├── config.ini                                                     # Custom operator vendor configuration file, define the priority between different vendors, which needs to have vendor configuration of mslite
└── mslite                                                         # Custom operator directory provided by mslite
    ├── framework                                                  # Third-party framework adaptation configuration
    │    └── tensorflow                                            # tensorflow adaptation configuration, not required
    │       └── npu_supported_ops.json
    ├── op_impl                                                    # Custom operator implementation directory
    │   ├── ai_core                                                # Run operator implementation directory in ai_core
    │   │   └── tbe                                                # tbe operator implementation directory
    │   │       ├── config                                         # Operator configurations for different chips
    │   │       │   ├── ascend310                                  # Operator configuration of Atlas 200/300/500 inference product chip
    │   │       │       └── aic_ascend310-ops-info.json
    │   │       │   ├── ascend310p                                 # Operator configuration of Atlas inference series chip
    │   │       │       └── aic_ascend310p-ops-info.json
    │   │       │   ├── ascend910                                  # Operator configuration of Atlas training series chip
    │   │       │       └── aic_ascend910-ops-info.json
    │   │       └── mslite_impl                                    # Implementation logic directory of operators
    │   │           ├── add_dsl.py                                 # add sample logic implementation file based on dsl development
    │   │           ├── add_tik.py                                 # add sample logic implementation file based on tik development
    │   │           ├── compiler.py                                # Operator compilation logic file needed for akg graph
    │   │           ├── custom.py                                  # akg custom operator implementation file
    │   │           ├── matmul_tik.py                              # matmul sample logic implementation file based on tik development
    │   ├── cpu                                                    # aicpu custom operator subdirectory, not required
    │   │   └── aicpu_kernel
    │   │       └── impl
    │   └── vector_core                                            # Run operator implementation directory in vector_core
    │       └── tbe                                                # tbe operator implementation directory
    │           └── mslite_impl                                    # Implementation logic directory of operators
    │               ├── add_dsl.py                                 # add sample logic implementation file based on dsl development
    │               ├── add_tik.py                                 # add sample logic implementation file based on tik development
    │               └── matmul_tik.py                              # matmul sample logic implementation file based on tik development
    └── op_proto                                                   # Operator prototype definition package directory
        └── libcust_op_proto.so                                    # operator prototype definition so file. akg custom operator is registered by default, and do not need this file