# AscendC_910B_Operators

**Repository Path**: liujing0000/AscendC_910B_Operators

## Basic Information

- **Project Name**: AscendC_910B_Operators
- **Description**: [蓝区开发] 个人蓝区->黄区同步用仓库
AscendC 基于910B芯片 开发的算子
- **Primary Language**: C++
- **License**: Apache-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 5
- **Created**: 2024-04-25
- **Last Updated**: 2024-04-25

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# AscendC 910B Operators

#### 介绍
蓝区->黄区 同步仓

#### 软件架构
当前基于CANN版本[7.0.T4.0.B114:7.0.T2] - 海思分包

#### 算子清单
**====== FlashAttentionMix ======**

Input	query	Tensor	FP16	ND	[batch, q_seqLen, heads, embed]

Input	key	Tensor	FP16	ND	[batch, kv_seqLen, heads, embed]

Input	value	Tensor	FP16	ND	[batch, kv_seqLen, heads, embed]

Output	gmOutput	Tensor	FP16	ND	[batch, q_seqLen, heads, embed]

Output	gmOutputS	Tensor	INT32	ND	[batch * heads * maxseqlen * maxseqlen * 2]

Output	gmOutputP	Tensor	FP16	ND	[batch * heads * maxseqlen * maxseqlen * 2]

Output	oTmp	Tensor	FP16	ND	[batch * heads * maxseqlen * maxseqlen * 2]

tiling参数	tilingParam	Array	Uint32_t	ND	HARDWARE_AICORE_NUM * TILING_PARA_SIZE



**====== UnpadAddLayerNorm ======**

Input	hiddenStates	Tensor	FP16		ND	[batch, seqlen, hiddenSize]

Input	residual	Tensor	FP16		ND	[batch, seqlen, hiddenSize]

Input	epsilon	Tensor	FP32		ND	1

Input	weight	Tensor	FP16		ND	[hiddenSize]

Input	bias	Tensor	FP16		ND	[hiddenSize]

Output	outStates	Tensor	FP16		ND	[batch, seqlen, hiddenSize]




#### 安装教程

git clone https://gitee.com/chenchuw/AscendC_910B_Operators.git

#### 使用说明

cd build

bash build_ops.sh

./custom_project/build_out/custom_opp_${OS}_aarch64.run