English | 简体中文

Introduction

Description:

• For details about the AI Core, see the "Concepts, Principles, and Terms > Hardware Architecture and Data Processing Principles" section in Ascend C Operator Development.

This project provides development and invocation samples for AI Core operators. Refer to the corresponding implementations based on your actual requirements.

Directory Description

├── example
│   ├── add_example                # Directory name of the AddExample operator, generally in lowercase with underscores
│   │   ├── CMakeLists.txt         # Operator compilation configuration file; keep the original file
│   │   ├── examples               # Operator usage examples
│   │   ├── op_graph               # Directory related to operator graph construction
│   │   ├── op_host                # Implementations related to the operator information library, Tiling, and InferShape
│   │   ├── op_kernel              # Operator kernel directory
│   │   └── op_kernel_aicpu        # Operator kernel_aicpu directory
│   ├── attention                  # Examples of attention operators
│   │   └── flash_attn
│   │       └── L0_minimal         # Minimal FlashAttention sample, with only op_kernel and no Host Tiling
│   ├── mc2                        # Examples of communication-computation fusion operators
│   │   ├── all_gather_add         # Directory name of the AllGatherAdd operator, generally in lowercase with underscores
│   │   │   └── ...
│   ├── CMakeLists.txt             # Operator compilation configuration file; keep the original file
│   └── README.md                  # Operator description document

Operator Development Samples

Sample Directory	Sample Description	Operator Development	Operator Invocation
add_example	Operator that adds two tensors.	For the end-to-end operator development process, see the AI Core Operator Development Guide.	For the invocation sample, see README
mc2/all_gather_add	Performs AllGather collective communication first, and then performs element-wise addition.	For the end-to-end operator development process, see the AI Core Operator Development Guide.	For the invocation sample, see README
mc2/ccu_alltoall	CCU direct-call (<<<>>>) AllToAll collective communication + Blaze MXFP8 MatMul dual-stream operator fusion (kernel direct-call sample).	—	For the invocation sample, see README
attention/flash_attn/L0_minimal	Minimal FlashAttention implementation, with no mask, no combine, and no independent Host Tiling.	For the operator implementation details, see FA_IMPL.md.	For the invocation sample, see README