该渲染由 Shiro API 生成，可能存在排版问题，最佳体验请前往：https://hyli.xyz/posts/notes/transformer-fundamentals

Multi-Head Self-Attention
-> Add & LayerNorm
-> Position-wise Feed-Forward Network
-> Add & LayerNorm

token embedding + positional encoding

token ids
-> embedding + positional encoding
-> N 层 Encoder blocks

input representation = token embedding + positional encoding

SubLayer -> Add -> LayerNorm

当前前缀
-> 预测下一个 token
-> 把这个 token 接回前缀
-> 继续预测

看完了？说点什么呢

linear / relu / sgd -> tokenization -> LLM workflow -> inference / deployment

SGD、Linear、ReLU

神经网络为什么能工作

神经网络本质上不是神秘黑盒，而是很多简单函数的组合。

最基础的一条链可以写成：

x -> Linear -> ReLU -> Linear -> loss -> backward -> SGD step

Linear 是什么

nn.Linear 做的是一个仿射变换。 它会把输入特征映射到新的特征空间。

例如：

• 输入是 in_features
• 输出是 out_features

对应代码

import torch
from torch import nn

linear = nn.Linear(4, 2)
x = torch.randn(3, 4)
y = linear(x)

print("x shape:", x.shape)
print("y shape:", y.shape)

需要记住的点

• Linear 层本身有参数
• 这些参数通常包括 weight 和 bias
• 训练时被更新的，就是这些参数

为什么只有线性层不够

如果一个模型里只有线性层，那么无论叠多少层，整体仍然只是一个线性变换。 这意味着模型表达能力有限，不能表示复杂函数。

所以，神经网络真正变强的关键，不只是“层数变多”，而是：

在线性层之间加入非线性。

ReLU 是什么

ReLU(x) = max(0, x)

也就是：

• 正数保留
• 负数变成 0

为什么需要 ReLU

ReLU 是非线性激活函数。 加入 ReLU 后，模型就不再只是简单线性变换的堆叠，而可以表达更复杂的函数关系。

对应代码

import torch
from torch import nn

relu = nn.ReLU()

x = torch.tensor([[-1.0, 0.5, 2.0]])
y = relu(x)

print("before relu:", x)
print("after relu:", y)

SGD 是什么

SGD（Stochastic Gradient Descent）做的事情是：

1. 用当前 batch 计算 loss
2. 根据 loss 反向传播得到梯度
3. 按学习率更新参数

也就是：

param = param - lr * grad

为什么叫 stochastic

因为它通常不是每次用整个数据集精确计算梯度， 而是只用当前一个 batch 来近似当前梯度。

学习率为什么重要

学习率过大：

• 参数更新步子太大
• 容易在某个范围内震荡
• 甚至可能直接发散

学习率过小：

• loss 下降太慢
• 训练时间过长
• 可能很久都看不到明显改善

需要记住的点

• 梯度决定方向
• 学习率决定步长

Tokenization

tokenization 是什么

语言模型不能直接处理原始文本。 文本在进入模型之前，必须先被转换成模型能处理的离散单位和数字表示。

这个过程就是 tokenization。

最小视角下，tokenizer 做两件事：

1. 把文本切成 token
2. 把 token 转成 token id

token 是什么

token 是模型输入的基本单位。 在现代 LLM 中，token 往往既不是完整单词，也不是单个字符，而更常是 subword（子词）。

为什么不用纯单词切分

如果按完整单词切分：

• 词表会非常大
• 遇到新词、罕见词时处理很差
• 语言泛化能力不够好

为什么不用纯字符切分

如果按单个字符切分：

• 序列会变得很长
• 训练和推理成本会变高
• 对模型来说不一定是更好的表示方式

所以现代 tokenizer 通常在两者之间取折中，使用 subword。

subword tokenization

需要认识 3 个名字：

• BPE
• WordPiece
• Unigram

它们的共同目标都是：

在控制词表大小的同时，尽可能保留有意义的文本片段。

对 CJK 的处理

对于中文、日文、韩文，tokenizer 通常不会简单按空格切分， 而更可能按字符边界、子词规则，或者直接在原始文本上学习切分方式。

所以：

• 中文里常常接近按字或短片段切
• 日文和韩文也常结合 subword 方法处理
• 同一句 CJK 文本，在不同 tokenizer 下可能切分结果不同

special tokens 是什么

除了普通文本 token，tokenizer 通常还会管理一些特殊 token，例如：

• BOS
• EOS
• CLS
• SEP
• PAD

这说明模型输入不只是正文 token，还包含一些控制信息。

token 数为什么重要

token 数会直接影响：

• 输入序列长度
• 上下文窗口占用
• 推理成本
• 延迟和吞吐
• 系统性能

所以 tokenizer 不是一个无关紧要的文本前处理步骤， 它实际上是模型系统的一部分。

最小例子

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

text = "I like machine learning."
encoded = tokenizer(text)

print("input_ids:", encoded["input_ids"])
print("tokens:", tokenizer.convert_ids_to_tokens(encoded["input_ids"]))

LLM 的最小全景图

LLM 是什么

LLM（Large Language Model）是在大量文本数据上训练出来的语言模型。 它可以理解和生成文本，并且在不做任务专门训练的情况下，处理很多语言任务。

这里要注意：

• 不是所有任务都能直接做好
• 也不是永远不需要进一步适配
• 但相比传统 NLP，LLM 的泛化能力更强

training / fine-tuning / inference

training / pretraining：

• 从大量数据中训练模型
• 学习通用语言模式
• 得到一个基础模型

fine-tuning：

• 在已有预训练模型基础上继续训练
• 用更具体的数据做适配
• 让模型更适合某个任务、领域或风格

inference：

• 使用现有模型做预测或生成
• 不再更新模型参数
• 只做前向计算

deployment / serving：

• 把 inference 包装成一个可调用的服务
• 对外提供接口或应用
• 让模型真正可用

为什么“做 LLM”不等于“只看模型结构”

做 LLM 不只是看 transformer 结构或者参数量。

一个真实的 LLM 工作流还包括：

• data
• tokenizer
• model
• training / fine-tuning
• inference
• serving
• 工具链与模型管理

也就是说：

模型只是系统的一部分，不是全部。

最小流程图

data -> tokenizer -> model -> training/fine-tuning -> inference -> serving -> agent

Deployment Awareness

为什么训练脚本不等于部署

训练脚本的职责是：

• 加载数据
• 计算 loss
• 反向传播
• 更新参数

而部署需要的是：

• 一个稳定的推理入口
• 接受输入
• 返回预测结果
• 对外提供服务

所以训练代码和部署代码不是一回事。

inference script 最小需要什么

一个最小推理脚本通常需要：

1. 加载模型结构
2. 加载模型权重
3. 切到 model.eval()
4. 接受输入
5. 做前向计算
6. 返回预测结果

如果是文本模型，还通常需要：

• 加载 tokenizer

为什么 model.eval() 在部署里很重要

推理阶段一般需要稳定、可复现的输出。 如果不切到 eval()：

• Dropout 可能还在随机丢弃神经元
• BatchNorm 可能还在使用当前 batch 的统计量

这会导致推理结果不稳定。

所以部署时通常会写：

model.eval()
with torch.no_grad():
    output = model(x)

最小 Web demo 的核心抽象

最小 Web demo 的核心不是“服务器”，而是：

input -> predict_function -> output

也就是说，本质上你只需要先写一个预测函数，然后再把它包装成界面或接口。

一个最小 Gradio 例子

import gradio as gr

def predict(text):
    return f"you entered: {text}"

demo = gr.Interface(
    fn=predict,
    inputs="text",
    outputs="text"
)

demo.launch()

这里最关键的是：

• fn：预测函数
• inputs：输入类型
• outputs：输出类型

最重要的结论

1. 线性层负责做特征映射。
2. 只有线性层不够，因为整体仍然是线性函数。
3. ReLU 这样的非线性激活函数让模型具备更强表达能力。
4. SGD 用 batch 的梯度近似来更新参数。
5. tokenizer 负责把文本切成 token，并转换成 token id。
6. 现代模型常使用 subword，而不是纯单词或纯字符。
7. token 数会影响上下文长度、推理成本和系统性能。
8. LLM 是在大规模文本上训练出来的语言模型。
9. training、fine-tuning、inference、deployment 是不同阶段。
10. 做 LLM 工程不是只看模型结构，而是整条工作流。
11. 训练代码负责更新参数，推理代码负责接受输入并输出预测。
12. 最小 Web demo 的核心抽象是：input -> prediction function -> output。

到这里，应该已经能做到

• 理解 Linear 层在做什么
• 理解为什么神经网络需要非线性激活函数
• 理解 ReLU 的作用
• 理解 SGD 为什么能让 loss 下降
• 理解学习率对训练速度和稳定性的影响
• 理解 token、word、character 的区别
• 理解 tokenizer 的最小职责
• 知道常见 subword 算法名字
• 知道 token 数为什么影响 LLM 系统表现
• 理解 LLM 的基本定义和工作流
• 区分 training / fine-tuning / inference / deployment
• 理解训练脚本和推理脚本的区别
• 理解为什么部署前必须 model.eval()
• 理解最小 Web demo 的抽象方式

References

从神经网络基础到 LLM 与部署

• fast.ai Lesson 3
• PyTorch nn.Linear Documentation
• PyTorch nn.ReLU Documentation
• PyTorch SGD Documentation
• Hugging Face Tokenization Algorithms
• Hugging Face LLM Course - Tokenizers
• Hugging Face LLM Course - Introduction
• Hugging Face LLM Course - NLP and LLMs
• fast.ai Lesson 2
• Gradio Quickstart

核心目标不是学会很多 API，而是先把最小训练闭环真正串起来：

dataset -> dataloader -> model -> forward -> loss -> backward -> optimizer.step() -> eval -> save/load

最小训练闭环是什么

训练一个神经网络，本质上是在重复下面几步：

1. 从 dataloader 中取出一个 batch 的数据
2. 把数据送进 model 做前向计算，得到预测值
3. 用 loss function 比较预测值和真实标签，得到 loss
4. 对 loss 调用 backward()，计算梯度
5. 调用 optimizer.step()，根据梯度更新参数
6. 调用 optimizer.zero_grad()，清空旧梯度

需要理解的概念

• batch：一次送进模型的一批样本
• step：处理一个 batch 并更新一次参数
• epoch：完整遍历整个训练集一次
• loss：模型当前预测得有多差
• gradient：loss 对参数的变化率
• learning rate (lr)：每次更新参数走多大一步

为什么参数更新时用减号

因为梯度指向的是 loss 增大的方向。
训练的目标是让 loss 下降，所以参数更新要沿着 负梯度方向 走：

w = w - lr * w.grad

Dataset 和 DataLoader

Dataset 是什么

Dataset 定义的是：第 i 个样本是什么。

它通常负责：

• 读取样本
• 返回输入和标签
• 支持按索引取数据

DataLoader 是什么

DataLoader 定义的是：这些样本怎么一批一批取出来。

它通常负责：

• 按 batch 组织样本
• 是否打乱顺序（shuffle=True/False）
• 多进程加载（num_workers）

两者关系

• Dataset：单个样本级别
• DataLoader：batch 级别

对应代码

from torch.utils.data import Dataset, DataLoader
import torch

class ToyDataset(Dataset):
    def __init__(self):
        self.X = torch.tensor([
            [0.0, 0.0],
            [0.0, 1.0],
            [1.0, 0.0],
            [1.0, 1.0],
        ])
        self.y = torch.tensor([0, 1, 1, 0])

    def __len__(self):
        return len(self.X)

    def __getitem__(self, idx):
        return self.X[idx], self.y[idx]

dataset = ToyDataset()
dataloader = DataLoader(dataset, batch_size=2, shuffle=True)

for X, y in dataloader:
    print("X shape:", X.shape)
    print("y shape:", y.shape)
    break

Model、forward 和 autograd

nn.Module 是什么

在 PyTorch 中，模型通常继承自 nn.Module。

一般会在：

• __init__() 中定义层
• forward() 中定义数据如何流过这些层

forward() 是什么

forward() 描述的是：输入如何变成输出。

例如：

• 输入先经过线性层
• 再经过激活函数
• 最后输出 logits

autograd 是什么

PyTorch 在前向计算时会自动构建计算图。 调用 loss.backward() 后，会自动按链式法则计算梯度，并把结果存到参数的 .grad 里。

对应代码

import torch
from torch import nn

class SimpleNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = nn.Linear(2, 8)
        self.relu = nn.ReLU()
        self.linear2 = nn.Linear(8, 2)

    def forward(self, x):
        x = self.linear1(x)
        x = self.relu(x)
        x = self.linear2(x)
        return x

model = SimpleNet()

X = torch.tensor([[1.0, 0.0]])
logits = model(X)

print("logits:", logits)
print("shape:", logits.shape)

最小 autograd 例子

import torch

x = torch.tensor([[1.0, 2.0]])
y = torch.tensor([[1.0]])

w = torch.tensor([[0.5], [-1.0]], requires_grad=True)
b = torch.tensor([0.1], requires_grad=True)

z = x @ w + b
loss = (z - y).pow(2).mean()

print("loss:", loss.item())

loss.backward()

print("w.grad:", w.grad)
print("b.grad:", b.grad)

需要记住的点

• requires_grad=True 表示要跟踪梯度
• loss.backward() 会把梯度写到参数的 .grad 上
• 梯度默认会累积

optimizer、train/eval、checkpoint

optimizer 是什么

optimizer 负责根据梯度更新参数。

例如：

optimizer = torch.optim.SGD(model.parameters(), lr=0.1)

这里：

• model.parameters() 表示要训练哪些参数
• lr 表示学习率

train 和 eval 的区别

• model.train()：训练模式
• model.eval()：评估/推理模式

它们本身不会自动更新参数，也不会自动计算准确率。 它们的作用是切换某些层的行为，例如：

• Dropout
• BatchNorm

为什么推理前要 eval()

因为如果模型里有 Dropout / BatchNorm，而你没有切到 eval()：

• Dropout 还会随机丢弃神经元
• BatchNorm 还会用当前 batch 的统计量

这样同一个输入可能得到不稳定的输出。

checkpoint 是什么

checkpoint 就是把训练好的模型参数保存下来，之后再加载恢复。

推荐方式是保存 state_dict()。

最小 train / test loop

import torch
from torch import nn
from torch.utils.data import DataLoader, Dataset

class ToyDataset(Dataset):
    def __init__(self):
        self.X = torch.tensor([
            [0.0, 0.0],
            [0.0, 1.0],
            [1.0, 0.0],
            [1.0, 1.0],
        ])
        self.y = torch.tensor([0, 1, 1, 0])

    def __len__(self):
        return len(self.X)

    def __getitem__(self, idx):
        return self.X[idx], self.y[idx]

class SimpleNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = nn.Linear(2, 8)
        self.relu = nn.ReLU()
        self.linear2 = nn.Linear(8, 2)

    def forward(self, x):
        return self.linear2(self.relu(self.linear1(x)))

def train_loop(dataloader, model, loss_fn, optimizer):
    model.train()
    for X, y in dataloader:
        pred = model(X)
        loss = loss_fn(pred, y)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

    return loss.item()

def test_loop(dataloader, model, loss_fn):
    model.eval()
    total_loss = 0.0
    total_correct = 0
    total_samples = 0

    with torch.no_grad():
        for X, y in dataloader:
            pred = model(X)
            total_loss += loss_fn(pred, y).item()
            total_correct += (pred.argmax(dim=1) == y).sum().item()
            total_samples += y.size(0)

    avg_loss = total_loss / len(dataloader)
    acc = total_correct / total_samples
    return avg_loss, acc

dataset = ToyDataset()
dataloader = DataLoader(dataset, batch_size=2, shuffle=True)

model = SimpleNet()
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)

for epoch in range(10):
    train_loss = train_loop(dataloader, model, loss_fn, optimizer)
    test_loss, test_acc = test_loop(dataloader, model, loss_fn)
    print(f"epoch={epoch+1}, train_loss={train_loss:.4f}, test_loss={test_loss:.4f}, test_acc={test_acc:.2%}")

保存和加载模型

import torch

# save
torch.save(model.state_dict(), "model_weights.pth")

# load
new_model = SimpleNet()
new_model.load_state_dict(torch.load("model_weights.pth", weights_only=True))
new_model.eval()

最重要的结论

1. 训练闭环的核心是： forward -> loss -> backward -> optimizer.step()

2. Dataset 和 DataLoader 的区别是：

   • Dataset 负责单个样本
   • DataLoader 负责 batch

3. 模型通常继承 nn.Module，并在 forward() 中定义前向计算。

4. autograd 会自动求梯度，loss.backward() 后梯度会写到参数的 .grad 上。

5. optimizer 根据梯度更新参数，学习率控制更新步长。

6. 推理时要用 model.eval()，并通常配合 torch.no_grad()。

7. 推荐保存 state_dict()，加载时先重新创建模型再 load_state_dict()。

到这里，应该已经能做到：

• 看懂一个最小 PyTorch 训练脚本
• 理解 batch / step / epoch 的关系
• 理解 loss、gradient、learning rate 的基本作用
• 写出一个最小 nn.Module
• 写出最小 train/test loop
• 保存并加载模型参数

References

最小训练闭环

• PyTorch Learn the Basics
• PyTorch Quickstart

Dataset 和 DataLoader

• PyTorch Datasets & DataLoaders
• torch.utils.data.DataLoader Documentation

Model、forward 和 autograd

• Build the Neural Network
• Automatic Differentiation with torch.autograd

optimizer、train/eval、checkpoint

• Optimizing Model Parameters
• Save and Load the Model

Optional

• PyTorch nn.Module Documentation
• PyTorch Dropout Documentation
• PyTorch BatchNorm2d Documentation

该渲染由 Shiro API 生成，可能存在排版问题，最佳体验请前往：https://hyli.xyz/posts/docs/prompt-engineering_v7

from langchain.agents import load_tools
from langchain.agents import initialize_agent
from langchain.agents import AgentType
from langchain.llms import VertexAI

prompt = "How many kids do the band members of Metallica have?"

llm = VertexAI(temperature=0.1)
tools = load_tools(["serpapi"], llm=llm)

agent = initialize_agent(tools, llm,
agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION, verbose=True)
agent.run(prompt)

> Entering new AgentExecutor chain...  
Metallica has 4 members.  
Action: Search  
Action Input: How many kids does James Hetfield have?  
Observation: three children  
Thought: 1/4 Metallica band members have 3 children  
Action: Search  
Action Input: How many kids does Lars Ulrich have?  
Observation: 3  
Thought: 2/4 Metallica band members have 6 children  
Action: Search  
Action Input: How many kids does Kirk Hammett have?  
Observation: Hammett has been married to his second wife Lani since 1998. They have two sons, Angel (b. September 29, 2006) and Vincenzo (b. June 28, 2008).  
Thought: 3/4 Metallica band members have 8 children  
Action: Search  
Action Input: How many kids does Robert Trujillo have?  
Observation: 2  
Thought: 4/4 Metallica band members have 10 children  
Final Answer: 10

import os
import shutil

folder_name = input("Enter the folder name: ")
prefix = input("Enter the string to prepend to the filename: ")
text = toUpperCase(prefix)

if not os.path.isdir(folder_name):
    print("Folder does not exist.")
    exit(1)

files = os.listdir(folder_name)

for file in files:
    new_filename = f"{text}_{file}"

shutil.move(os.path.join(folder_name, file),
os.path.join(folder_name, new_file_name))

print("Files renamed successfully.")

Traceback (most recent call last):
  File "/Users/leeboonstra/Documents/test_folder/rename_files.py", line 7,
in <module>
    text = toUpperCase(prefix)
NameError: name 'toUpperCase' is not defined

{
  "type": "object",
  "properties": {
    "name": { "type": "string", "description": "Product name" },
    "category": { "type": "string", "description": "Product category" },
    "price": { "type": "number", "format": "float", "description": "Product price" },
    "features": {
      "type": "array",
      "items": { "type": "string" },
      "description": "Key features of the product"
    },
    "release_date": { "type": "string", "format": "date", "description": "Date the product was released"}
  },
}

{
  "name": "Wireless Headphones",
  "category": "Electronics",
  "price": 99.99,
  "features": ["Noise cancellation", "Bluetooth 5.0", "20-hour battery life"],
  "release_date": "2023-10-27"
}

看完了？说点什么呢

Prompting guide 101

A quick-start handbook for effective prompts

October 2024 edition

Writing effective prompts

From the very beginning, Google Workspace was built to allow you to collaborate in real time with other people. Now, you can also collaborate with AI using Gemini for Google Workspace to help boost your productivity and creativity without sacrificing privacy or security. The embedded generative AI-powered features can help you write, organize information, create images, accelerate workflows, have richer meetings, and much more, all while using your favorite apps like Gmail, Google Docs, Google Drive, Google Sheets, Google Meet, Google Slides, and Gemini Advanced (the standalone chat experience available at gemini.google.com with enterprise-grade security). Gemini is accessible right where you are doing your work — with access to your personal knowledge base in Drive, Docs, Gmail, and more — so you can enhance and create powerful workflows across the Workspace apps with less tab switching and interruption.

This guide provides you with the foundational skills to write effective prompts when using Gemini for Workspace. You can think of a prompt as a conversation starter with your AI-powered assistant. You might write several prompts as the conversation progresses. While the possibilities are virtually endless, you can put consistent best practices to work today.

The four main areas to consider when writing an effective prompt are:

• Persona
• Task
• Context
• Format

Here is an example of a prompt using all four areas that could work well in Gmail and Google Docs:

You are a program manager in [industry].Draft an executive summary email to[persona] based on [details about relevant program docs]. Limit to bullet points.

You don't need to use all four in every prompt, but using a few will help! Always remember to include a verb or command as part of your task; this is the most important component of a prompt.

Here are quick tips to get you started with Gemini for Google Workspace:

1. Use natural language. Write as if you're speaking to another person. Express complete thoughts in full sentences.

2. Be specific and iterate. Tell Gemini what you need it to do (summarize, write, change the tone, create). Provide as much context as possible.

3. Be concise and avoid complexity. State your request in brief — but specific — language. Avoid jargon.

4. Make it a conversation. Fine-tune your prompts if the results don’t meet your expectations or if you believe there’s room for improvement. Use follow-up prompts and an iterative process of review and refinement to yield better results.

5. Use your documents. Personalize Gemini's output with information from your own files in Google Drive.

6. Make Gemini your prompt editor. When using Gemini Advanced, start your prompts with: “Make this a power prompt: [original prompt text here].” Gemini will make suggestions on how to improve your prompt. Ensure it says what you need, and then paste it back into Gemini Advanced to get an output.

Prompting is a skill we can all learn. You don't have to be a prompt engineer to use generative AI. However, you will likely need to try a few different approaches for your prompt if you don't get your desired outcome the first time. Based on what we've learned from our users so far, the most fruitful prompts average around 21 words with relevant context, yet the prompts people try are usually less than nine words.

Generative AI and all of its possibilities are exciting, but it’s still new. Even though our models are getting better every day, prompts can sometimes have unpredictable responses.

Before putting an output from Gemini for Workspace into action, review it to ensure clarity, relevance, and accuracy. And of course, keep the most important thing in mind: Generative AI is meant to help humans but the final output is yours.

The example prompts in this guide are meant for illustrative purposes.

Table of contents

• Writing effective prompts
• Introduction
• Administrative support
• Communications
• Customer service
• Executives
• Frontline management
• Human resources
• Marketing
• Project management
• Sales
• Small business owners and entrepreneurs
• Startup leaders
• Leveling up your prompt writing

Introduction

Gemini for Google Workspace: Prompting 101

Gemini for Workspace is your AI-powered assistant integrated into the apps you use every day — Gmail, Google Docs, Google Sheets, Google Meet, Google Slides, and Gemini Advanced (the standalone chat experience available at gemini.google.com with enterprise-grade security). This means the apps you know and use will work together smoothly so you can collaborate with Gemini right where you are. You can have fewer interruptions to your focus and workflow, helping you complete tasks and do things you might not have initially known how to do.

You can access the features of Gemini for Workspace in multiple ways. Engaging with Gemini in the side panel of your Workspace apps allows you to create highly personalized generative AI outputs that are based on your own files and documents — even if they aren’t Google Docs. You can generate personalized emails in seconds referencing your own Docs to pull in relevant context, generate Slides that are based on information directly from your own briefs or reports, and so much more.

Understanding what makes an effective prompt and learning to craft prompts on the fly can boost your productivity and creativity. Gemini for Workspace can help you:

• Improve your writing

• Organize data

• Create original images

• Summarize information and surface insights

• Have better meetings with automatic note taking

• Research unfamiliar topics easily

• Spot trends, synthesize information, and identify business opportunities

For 25 years, Google has built helpful, secure products that give users choice and control over their data. It’s a bedrock principle for us. This was the case back when we first launched Gmail in 2004, and it remains true in the era of generative AI. This means your data is your data and does not belong to Google. Your data stays in your Workspace environment. Your privacy is protected. Your content is never used for targeting ads or to train or improve Gemini or any other generative AI models.

How to use this prompt guide

This guide introduces you to prompting with Gemini for Workspace. It includes strong prompt design examples to help you get started. Additionally, it covers scenarios for different personas, use cases, and potential prompts.

You will notice a variety of prompt styles. Some prompts have brackets, which indicate where you would fill in specific details or tag your own personal files by typing @file name. Other prompts are presented without variables highlighted to show you what a full prompt could look like. All of the prompts in this guide are meant to inspire you, but ultimately they will need to be customized to help you with your specific work.

To get started, use the role-specific suggested prompts as inspiration to help you unlock a new and powerful way of working.

Administrative support

As an administrative support professional, you are responsible for keeping teams on track. You’re required to stay organized and efficient — even under pressure — while juggling many priority tasks.

This section provides you with simple ways to integrate prompts in your daily tasks.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Google Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

NEW Use case: Plan agendas (offsite, meetings, and more)

You're planning a three-day offsite meeting. To build an agenda, you brainstorm with Gemini Advanced.
You type:

I am an executive administrator to a team director.Our newly formed team now consists of content marketers, digital marketers, and product marketers. We are gathering for the first time at a three-day offsite in Washington, DC.Plan activities for each day that include team bonding activities and time for deeper strategic work.Create a sample agenda for me.

• Persona
• Task
• Context
• Format

This is a helpful start to your planning. You need to generate specific ideas for the team bonding activities.

You type:

Suggest three different icebreaker activitiesthat encourage people to learn about their teammates’ preferred working styles, strengths, and goals. Make sure the icebreaker ideas are engaging and can be completed by a group of 25 people in 30 minutes or less.

You are happy with the agenda as a starting point. You now want to reformat Gemini’s response into a table.
You type:

Organize this agenda in a table format. Include one of your suggested icebreakers for each day.

You select Export to Docs. You open the newly created Doc. Now, you want to bring in detailed summaries for the strategy sessions using your existing files in Google Drive to provide more context for what will be discussed. You prompt Gemini in Docs and tag your relevant files by typing @file name.

Use @[2024 H2 Team Vision] to generate a summary for the opening remarks on Day 1 of this agenda.

Example use cases

Executive administrators and executive business partners

NEW Use case: Manage multiple email inboxes

After returning from vacation, you have many unread, unsorted emails. You prompt Gemini in the Gmail side panel. You type:

Summarize emails from [manager] from the last week.

Gemini returns short summaries of each message. To directly access a message, you click on Sources and see tiles that bring you to specific emails. You select the most important one. Once the email thread opens, you see that many messages were exchanged. You prompt Gemini in Gmail:

Summarize this email thread and list all action items and deadlines.

You owe a response to a question, which you believe is best answered by a document in your Drive. You prompt Gemini in the Gmail side panel. You type:

Generate a response to this email and use @[file name] to describe how the [initiative] can complement the workflow outlined in [colleague’s name]’s message.

Gemini in Gmail returns a suggested email that pulls directly from your own Doc. After reading it over, you select the Copy icon in the side panel and paste it directly into your message.

NEW Use case: Plan business travel

Your manager has an upcoming meeting that is out of town. You are responsible for booking travel arrangements and creating a personalized itinerary. You need to research places to eat. You brainstorm with Gemini Advanced. You type:

I am an executive assistant. I need to create an itinerary for a two-day business trip in [location] during [dates]. My manager is staying at [hotel]. Suggest different options for breakfast and dinner within a 10-minute walk of the hotel, and find one entertainment option such as a movie theater, a local art show, or a popular tourist attraction. Put it in a table for me.

You continue your conversation until you are happy with the itinerary. Before you make reservations, you want to share the draft with your manager. You select Share & export and select Draft in Gmail. Once the drafted email is created, you put the final touches on the message and send.

NEW Use case: Track travel and entertainment budget

You want to create a spreadsheet to keep track of all of the travel expenses incurred. You open a new Google Sheet and prompt Gemini in the Sheets side panel. You type:

Create a budget tracker for business travel. It should include columns for: date, expense type (meal, entertainment, transportation), vendor name, and a description.

Gemini returns a tracker that is now ready for you to enter data.

Communications

As a communications professional, you are responsible for ensuring your business is well understood by the public. You have to stay up to date with the trends, communicate clearly and effectively with many stakeholders, and build compelling narratives.

This section provides you with simple ways to integrate prompts in your daily tasks.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Google Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

NEW Use case: Create a press release

You are in charge of public relations at a company in the personal care industry. The company you work for has just acquired a smaller brand, and you need to craft a press release. You’ve completed interviews with your company’s CEO, CFO, and the acquired company’s CEO. You’ve stored all of the most important quotes in one Doc. You also have a Doc with all of the information about the acquired brand, its vision, how it got started, and stats. You open a new Doc and prompt Gemini in the Docs side panel and type @file name to reference your relevant files. You type:

I'm a PR manager.I need to create a press release with a catchy title.Include quotes from@[VIP Quotes Acquisition].

• Persona
• Task
• Context
• Format

[Gemini returns a response]

Now you have a starting place for the press release, but you want to include more details about the brand that is being acquired and its founder. This information is stored in your Drive in another file. In the press release Doc, you prompt Gemini in the Docs side panel. You type:

Use @[Biography and Mission Statement] to add more information about the company that is being acquired, its mission, and how it got started.

The generated paragraphs are a good starting place, so you select Insert to add them into your draft, and you begin making edits to the press release.

Example use cases

Analyst and public relations

NEW Use case: Prepare for analyst or press briefings

You need to create a brief to prepare a spokesperson for an upcoming meeting with analysts and the media for a new product launch. You open a new Doc and prompt Gemini in the Docs side panel. You type:

Generate a brief template to prepare [spokesperson] for an upcoming media and analyst briefing for @[Product Launch]. Include space for a synopsis, key messages, and supporting data.

This gives you a starting point to pull in additional information from your files. You prompt Gemini in the Docs side panel and tag your relevant files by typing @file name. You type:

Craft a synopsis of the product launch in three main points using @[Product Launch - Notes].

You click Insert before repeating the process to fill out the rest of the briefing document. Next, you need to create a spreadsheet of media and analyst contacts. You open a new Google Sheet and prompt Gemini in the Sheets side panel. You type:

Organize my media and analyst contacts from @[Analyst and Journalist Contact Notes] for a new product briefing. I need to keep track of their names, type of contact (analyst or journalist), focus area, the name of the outlet, agency or firm that they work for, and a place where I can indicate the priority level of their attendance at this briefing (low, medium, high).

Gemini in Sheets returns a spreadsheet, and you can go through and indicate priority level for each contact. Next, you want to create a slideshow to use during the briefing. You open a new Google Slide and prompt Gemini in the Slides side panel. You tag relevant files by typing @file name in the prompt. You type:

Create a slide describing what [product] is from @[Product Launch - Notes]. Make sure it is short and easily understood by a broad audience.

Gemini returns a Slide. You continue to build your presentation by using this method to generate additional Slides.

NEW Use case: Create mock interview questions to prepare spokespeople

Now, you need to prepare your company's spokesperson for interviews that will follow the briefing. To generate a list of mock interview questions, you decide to chat with Gemini Advanced. You type:

I am a [PR/AR] manager at [company name]. We just launched [product] and had a briefing where we discussed [key messages]. I am preparing [spokesperson and role/title] for interviews. Generate a list of mock interview questions to help [spokesperson] prepare. Include a mixture of easy and hard questions, with some asking about the basics of [product] and some asking about the long-term vision of [product].

Gemini returns a list of questions that can help you prepare your company's spokesperson. You refine the suggested questions by continuing the conversation with Gemini. Then you select Share & export and Export to Docs. You open the newly created Doc, prompt Gemini in the Docs side panel, and tag relevant files by typing @file name. You type:

Use @[Product Launch Notes] to write suggested answers for these questions. Write the talking points as if you are [title of spokesperson] at [company].

Gemini in Docs returns suggested talking points, and you select Insert to add them into your draft. Now you're ready to continue tweaking the interview prep for your spokesperson.

Communications manager

NEW Use case: Craft internal communications

Your company has redesigned its intranet to be more user friendly. You’re in charge of internal communications for the launch. You want help drafting this message. You open a new Google Doc and prompt Gemini in the Docs side panel. You type:

I need to draft a company-wide memo unveiling our relaunched intranet. The [new page] addresses [common feedback we heard from employees] and aims to create a more user friendly experience. Draft an upbeat memo announcing [the new site] using @[Intranet Launch Plan Notes].

Gemini in Docs returns a drafted memo. You refine and edit the text to be exactly as you need it.

Customer service

As a customer service professional, you strive to deliver service that’s effortlessly efficient, consistently delightful, and powered by a proactive, helpful team. This section provides you with simple ways to integrate prompts in your daily tasks.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Google Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

Use case: Draft customer communications

You're a customer service representative, and you're responsible for responding directly to customer inquiries and concerns. You just received an email from a customer who received damaged goods. You open a new Google Doc and click on Help me write to prompt Gemini in Docs. Type the following:

Help me craft an empathetic email response.I am a customer service representative, and I need to create a response to a customer complaint.The customer ordered a pair of headphones that arrived damaged. They’ve already contacted us via email and provided pictures of the damage. I’ve offered a replacement, but they’re requesting an expedited shipping option that isn’t typically included with their order.Include a paragraph that acknowledges their frustration and three bullet points with potential resolutions.

• Persona
• Task
• Context
• Format

You like the email that Gemini in Docs created so you select Insert. But you want to brainstorm ways to resolve the issue without offering expedited shipping. You prompt by selecting Help me write. You type:

Suggest 10 alternative options in place of expedited shipping to resolve the customer’s frustration about receiving the damaged package.

Example use cases

Customer Service Manager or Representative

NEW Use case: Respond to complex customer issues using FAQ documents

A customer has reached out with a multi-part, complex question. You need to find and use information that is spread across multiple documents in order to respond accurately. You prompt Gemini in the Drive side panel.

You type:

Summarize information about [product name] including the product’s specific [return policy], [ingredients], and [certifications].

Gemini returns a summary and links to relevant files, which you can directly click into from the side panel. You read the information before returning to your email to generate a response to the customer. You open the message and prompt Gemini in the Gmail side panel and tag relevant files by typing @file name. You type:

Generate a response to the customer question about our [return policy] and [product certifications] based on @[Customer FAQ Document]. Use a helpful and professional tone.

Use case: Standardize communication frameworks

You're a customer service team manager. You need to create scalable resources to standardize your team's communications. You open a new Google Doc. You brainstorm by prompting Gemini in the Docs side panel. You type:

Draft templates for three different types of customer communication. Create templates for apology emails, order confirmation messages, and thank you notes for loyal customers. Keep each template to one paragraph and use a friendly tone.

The suggested templates offer a starting point for you to begin editing and personalizing with elements consistent with your company’s brand and policies. Now you want to outline your team’s communication best practices for onboarding. You open a new Doc and prompt Gemini in Docs. You type:

Craft a list of customer communication best practices that can be used to train new team members. Outline three sections, including how to handle happy customer inquiries, neutral customer inquiries, and dissatisfied customer inquiries.

You also want to support the team with standardized language that they can use when interacting with customers on phone calls. You prompt Gemini Advanced:

I am a [customer service manager]. I am trying to create standardized language that the team can use when interacting with customers on phone calls. Generate templates for common call openings, greetings, and closures for a customer service representative at a retail store. These templates should allow for personalization with customer details. The goal is to ensure consistency and professionalism while allowing for differentiation with specific customer information.

Use case: Improve customer service

You've noticed an uptick in customer complaints. You need to collaborate across departments to address recurring issues. You prompt Gemini in Gmail. You type:

Draft an email to my colleagues proposing a meeting to discuss customer experience improvement initiatives. Request that marketing, sales, and product stakeholders meet in the next week to get a clear sense of roles and responsibilities.

You edit the email and send it to your colleagues. Now you want to create a spreadsheet that you can use to track progress on this cross-departmental initiative. You open a Google Sheet and prompt Gemini in the Sheets side panel. You type:

Create a table to track the progress and impact of different customer experience improvement tactics using relevant metrics, including support ticket volume and priority level (high, medium, low).

Customer Support Specialist

NEW Use case: Analyze customer feedback

You have a spreadsheet that tracks customer feedback. You want to analyze it and brainstorm potential reasons for the trends. You chat with Gemini Advanced. You upload the file and type:

I am a customer support specialist. Using the attached spreadsheet, identify trends and patterns in our [customer feedback] by [category] over [time period]. Identify areas where [customer outreach] has increased significantly and investigate potential reasons.

Use case: Enable customer self-service

Customer feedback has consistently said your return policy guidelines are unclear. You open a Doc that states the return, refund, and store credit policies. You prompt Gemini in Docs by selecting Help me write. You type:

Summarize this content to write a clear and concise product return policy and outline 5 steps for customers to take in sequential order.

You like how simple the steps are. You repeat the process for your refund policy and store credit policy. Now, you want to use the newly simplified content to create a blog post for customers. Using your Google Doc with the newly written guidance, you prompt Gemini in Google Docs. You type:

Take this content and turn it into a short blog with the title “Resolve Common Issues Without Agent Assistance.” Have separate sections for our return policy, our refund policy, and our store credit policy.

Now you want to create an email template that the team can use when they receive customer questions around these three areas. You open a new Google Doc and prompt Gemini in Docs using Help me write. You type:

Draft an email template to a customer that highlights self-service resources referencing [blog link] for [support issues]. Thank the customer for their business and assure them of our commitment to meeting their needs.

Use case: Conduct voice of the customer research

You want to email a dissatisfied customer to attempt to make things right. You open an email that includes a customer complaint. You prompt Gemini in Gmail by selecting Help me write. You type:

Request a follow-up conversation on [date] at [time] with this customer who provided negative feedback to understand their concern and offer resolutions. Include example solutions.

The drafted response is a nice start, but you want to refine the language. You iterate by prompting Gemini in Gmail using Refine and Elaborate. Next, you want to create a short survey that you can send after each follow-up customer call. You open a new Google Doc and prompt Gemini in Docs. You type:

Create five different questions to customers who have just spoken to an agent on the phone. Questions should gauge how effective the call was, if the customer’s concern was addressed, and if they would recommend our business to others.

Executives

As an executive, your time is incredibly constrained. Every decision you make can impact growth, innovation, and the trajectory of your business. Understanding your market and making informed, strategic decisions is paramount, and so is getting urgent tasks done while you’re on the go.

This section provides you with simple ways to integrate prompts in your daily tasks.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Google Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

Use case: Communicate on the go

You are an executive about to board a long flight, and you just received an invitation for the next board meeting with an agenda. You have a couple of comments, and you want to propose adding a few topics to the agenda. You open Gmail, and you prompt Gemini in Gmail. You type:

Draft an emailconfirming that I will be at the board meeting.Ask if we can adjust the agenda to give 15 minutes to [urgent topics].

• Persona
• Task
• Context
• Format

Gemini in Gmail: [Drafts an email]

The email looks good, but you want to make sure the tone is as formal as possible. You select Refine and Formalize.

Gemini in Gmail: [Formalizes tone]

You read the email and select Insert. Before sending it, you make a light edit to thank the team for keeping you on track while traveling.

Example use cases

Chief Executive Officer

NEW Use case: Enhance personal productivity and time management

You have important email threads that have numerous responses. You need to quickly catch up. You open the message in Gmail and read the automatically generated summary from Gemini in the Gmail side panel. To respond, you prompt Gemini in the Gmail side panel and tag relevant files by typing @file name. You type:

Generate a response to [person] about [topic]. Include details on [deliverable] and [timeline] using @[Project A Status Report].

NEW Use case: Create outlines of presentations in seconds

Your team will pull together a presentation for you, and you want to provide an outline to get them started. You want to generate an outline using Gemini Advanced. You select the microphone icon and use your voice to prompt. You say:

I'm the CEO giving a presentation to [audience] at [event], and I want to create a detailed outline for my team to get started. I want to include a few important topics, including [areas of focus] and how our company is innovating with [company initiatives]. I'm envisioning time for a customer Q&A to end the presentation. Include suggested questions we could ask of a customer from the [industry] industry about how they are using our [product] to achieve [business outcome].

Chief Operating Officer

Use case: Prepare challenging employee communications

You're hosting a quarterly town hall meeting with the entire company. You want to write uplifting remarks to open the meeting. You open a new Doc and prompt Gemini in the Docs side panel. You type:

Write two uplifting paragraphs for employees who have just finished a challenging quarter. Acknowledge [difficulties] and emphasize [positives] for the upcoming quarter. Use a tone that is motivating, optimistic, and fosters a sense of unity and collaboration.

You want to brainstorm and practice how you will respond empathetically to potentially tough questions. You go to Gemini Advanced and type:

I'm the COO of a mid-sized company. I am hosting a quarterly town hall meeting with the entire company. I want to brainstorm and practice how I will respond to potentially tough questions. Help me write challenging questions that employees may ask at the upcoming town hall about [URL of company announcement]. Generate potential answers for each question that use a confident but firm tone. The responses should acknowledge the concern and let the employees know that we are striving to do our best for the entire company.

Use case: Streamline responses on the go

Your plans have changed, and you can't attend a meeting. You need to provide the team with answers on a few key items. You open Gmail and use a voice command to prompt Gemini in Gmail. You say:

Draft an email to [project lead] letting them know I will not be in the meeting due to an urgent matter. Ask them to take detailed notes and to ensure the team arrives at a decision on [key topic] in addition to assigning ownership of the postmortem report to [colleague].

Chief Marketing Officer

NEW Use case: Perform market research and campaign planning

You're starting annual planning. You want to conduct research on your target audience. You chat with Gemini Advanced. You type:

I'm a marketing leader conducting analysis in preparation for next year's [launch]. Define my target audiences [audiences], for my new line of [product]. Include interests, relevant marketing channels, and top trends that drive their consideration and purchase behavior.

Next, you export your findings to a Doc by selecting Share & export and Export to Docs. Now, you want to pull in relevant data from your own files by typing @file name. You prompt Gemini in the Docs side panel. You type:

Brainstorm value props for my [target audiences] based on features from @[Product Requirements Document]. Include a section on campaign learnings from @[Campaign Performance].

Use case: Brainstorm content and thought leadership

You finished a meeting with your social media team leads. You took notes in a Doc about what resonates with your audience, trending topics, target audience data, and keywords that are effective in driving engagement with your brand. You want to brainstorm potential thought leadership pieces using these insights. You prompt Gemini in the Docs side panel. You type:

Generate a list of four relevant and engaging thought leadership blog post ideas for [company] based on trending topics, target audience analysis, and brand keywords.

During the same conversation, the team discussed launching a new brand campaign. You know that your customers value your reliable and unique services, and your company has a long history of delivering for customers. You need help getting started with ideas on a new campaign tagline. You open a new Google Doc and select Help me write. You type:

Generate three options for a new slogan emphasizing reliability, innovation, and a long history of popularity for [company].

The slogans help you get started with the creative process. You have upcoming events that could be the perfect place to test elements of a new campaign. You want to mock up ideas for booth graphics for your events team. You open a new presentation in Google Slides and select Create image with Gemini. You type:

Create an image of a trade show booth using orange and blue colors. The booth should be modern and showcase interactive computer stations.

Use case: Conduct competitive analysis

Your team is considering expanding into a new line of business. To research, you go to Gemini Advanced, and you type:

I am a CMO conducting a competitive analysis. My company is considering expanding into [a new line of business]. Generate a list of the top five competitors in the [industry] industry and include their pricing, strengths, weaknesses, and target audience.

After going deeper in your research, you decide to create a five-year strategy to see what this could look like for the company. You type:

Okay, I am going to try to convince my CEO that we should expand into [line of business]. Draft a concise, competitive strategy outline for the next five years for the [industry] industry across North America markets with potential goals, strategies, and tactics.

After iterating to generate an appropriate outline, you fill in additional details and thoughts to make the document your own.

Chief Technology Officer

Use case: Summarize emerging technology trends

You need to catch up on emerging technology trends as the landscape is shifting quickly. You open Gemini Advanced, and you type:

I am the CTO of [company] in [industry]. I want to understand emerging technology trends. Summarize the top five emerging technologies with the most significant potential impact on [industry]. For each technology, list its potential benefits and challenges, and suggest how it could impact [company] in the next two to three years.

You want to dig deeper on specific topics, so you continue the conversation by typing:

Recommend three areas where [my company] can take proactive steps to stay ahead of the curve on [specific areas].

Chief Information Officer

NEW Use case: Communicate technical topics to non-technical audiences

You're making the case to digitally transform your company by adopting generative AI solutions. You need to present to the CEO and other leadership. You want help in communicating technical topics to non-technical audiences. You chat with Gemini Advanced. You type:

I am the CIO at [company], and I am trying to build the case to [adopt generative AI solutions]. I need to explain the technical concept of generative AI to a non-technical audience (the CEO and board). Help me write talking points that will help me convey what generative AI is, ways it could help us digitally transform, and why it’s important to our growth as a company. Include details about how it could potentially refocus our technical talent on more strategic work, help enhance our workforce’s productivity, and help us better serve our global workforce and customers.

Gemini provides suggested ways to discuss the topic. You continue your brainstorm and then export your conversation by clicking Share & export and Export to Docs. Then, to build a presentation, you open a new Google Slide and prompt Gemini in the Slides side panel and tag relevant files by typing @file name. You type:

I need to build a presentation to explain a technical topic to a non-technical audience. Generate an [introduction slide] that [describes what generative AI is] using @[Gen AI Explanation Notes].

You continue to use the same prompt, adjusting the topic to generate more slides for your presentation based on your notes.

NEW Use case: Research vendor products, services, and features

You're working on a report to make a vendor recommendation. You visit Gemini Advanced and type:

I am the CIO at [company]. We are currently evaluating vendor options to [replatform our intranet]. Right now, we use [vendor], but we are looking to switch because [we are unhappy with limited functionality and account support]. Suggest additional vendor options to consider and include descriptions of their product and services and key features.

Use case: Develop technical summaries

Your team just provided a lengthy technical report. You need to summarize it for your CEO. You open the Google Doc with the full report, and you prompt Gemini in the Docs side panel. You type:

Summarize the key findings and implications of this report for [audience]. Focus on the main [vulnerabilities] identified and the recommended actions to address them. Use a formal tone. You make light edits to the summary and include it as an executive summary.

Use case: Track IT assets

Your company needs a quick way to track software access for new hires. You open a new Google Sheet and prompt Gemini in the Sheets side panel. You type:

Create a tracker of software licenses for employees and include columns for license types, usage rights, and renewal dates.

Chief Human Resources Officer

Use case: Demonstrate employee appreciation

You want to set up a new program to help everyone feel included, appreciated, and acknowledged across the organization. To brainstorm, you open a new Google Doc and prompt Gemini in the Docs side panel. You type:

Brainstorm 10 employee appreciation ideas based on diverse employee interests such as cooking, gardening, sports, reading, and traveling.

Gemini in Docs kick-starts your creativity, and now you have ideas for employee interest clubs and events. You also want to ensure your leadership team is regularly encouraging managers to recognize talent on their teams, so you create email templates they can use as inspiration. You prompt Gemini in Docs by selecting Help me write, and you type:

Draft an email template that thanks [employee] for their hard work and [recent accomplishments]. Offer them an extra perk for their dedication, such as [a coffee gift card]. Use an upbeat and professional tone.

Use case: Assess employee satisfaction

You want to draft an anonymous survey that allows people to openly and honestly assess how they are feeling. To draft questions, you open a new Google Doc and prompt Gemini in the Docs side panel. You type:

Draft an anonymous employee satisfaction survey with questions and answer options that touch upon key areas like workload, work-life balance, compensation, and career growth opportunities. Ensure the questions are clear, concise, and avoid leading answers.

You received feedback from 15 senior leaders, and you’ve gathered all of the anonymous results in a Doc.
You want to create a summary that you can use in your next call. You prompt Gemini in the Docs side panel.
You type:

Summarize the results of the employee feedback to identify key themes.

Frontline management

As a frontline worker manager, your team’s work is indispensable to your organization — your team may not primarily complete its day’s work on a computer, but communication and collaboration remains key.

This section provides you with simple ways to integrate prompts in your daily tasks.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Google Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

NEW Use case: Find accurate information quickly

A customer just approached you with a question about an ongoing sale. You could use help navigating the numerous files you have access to so that you find the right information quickly. You prompt Gemini in the Drive side panel. You type:

Find the documentthat details the [company name]'s [holiday] sale details.

• Persona
• Task
• Context
• Format

Gemini in Drive returns suggested relevant files. From the side panel, you can directly summarize the files or you can click into a specific document. You open a suggested Doc to help answer the question. You prompt Gemini in the Docs side panel. You type:

How much can customers save on [product type] during this sale?

Gemini returns a response, which helps you answer your customer's question in a timely manner.

Example use cases

Retail associate

NEW Use case: Improve team collaboration by finding and sharing information easily

Your store recently updated its return and exchange policies. To find the information, you prompt Gemini in the Drive side panel. You type:

Find the document that discusses our new return and exchange policies.

Gemini returns suggested files that are related to the new policies. You directly click into the relevant file. Now, you want to send an email summarizing the document for your colleagues’ future reference. You open your email and prompt Gemini in the Gmail side panel. You type:

Write an email to my new colleagues summarizing @[Updated Return and Exchange Policy H2 2024].

You select Insert and further personalize the message before sending it.

NEW Use case: Streamline task management

You have a list of opening and closing duties that you must perform depending on what shift you are working. You want to keep yourself organized, so you create a tracker using the duties listed in your onboarding Doc. You open a new Google Sheet and prompt Gemini in the Sheets side panel and tag relevant files by typing @file name. You type:

I am a retail manager and I need to create a checklist for my opening and closing duties. Create a template with columns for [opening and closing duties] from @[Onboarding New Hire Information].

Gemini creates a spreadsheet. As you go through your day, you mark different tasks as complete. You have to leave your shift early, but you first need to communicate to the rest of the team what still needs to be done. You open your Gmail and prompt Gemini in the Gmail side panel and tag the spreadsheet you just created. You type:

Write an email to the team telling them what still needs to be done from the AM shift from @[Opening and Closing Duties Tracker].

Warehouse worker

NEW Use case: Manage inventory

A customer wants to place a bulk order. You need to check the store's inventory to see if you have enough to fulfill it. You open your inventory spreadsheet that tracks this information and prompt Gemini in the Sheets side panel. You type:

How many [units] of [product] do we have left in our inventory?

NEW Use case: Manage audits

Your warehouse is undergoing an inventory audit, and you’re in charge of verifying any numbers that are misaligned between your inventory tracker product total and what was counted during the audit. You prompt Gemini in the Sheets side panel. You type:

Create a formula that helps me calculate the difference between two columns. Which items have a discrepancy in [the total number counted] versus [the quantity on hand]?

You verify Gemini’s response that there are only a few items whose count did not align to your inventory tracker’s total. You need to write a message to your supervisor telling them that you’re looking into the issue. You open your Gmail and prompt Gemini in the Gmail side panel. You type:

I'm a warehouse worker managing an audit. Write a message to my supervisor to let them know that I am looking into the products whose counts are incorrect.

The drafted email looks good to go, so you hit send after reviewing.

Human resources

As an HR professional, you're the backbone of your organization, and you deal with a large volume of confidential and sensitive information. You shape company culture, find and nurture talent, and ensure a positive employee experience. These are no small feats.

This section provides you with simple ways to integrate prompts in your day-to-day tasks.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Google Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

Use case: Welcome new employees

You're an HR manager working on a presentation script. You have a Google Doc full of notes, bullet points, and topics that you would like to cover. You begin by opening your Google Doc with notes, and you prompt Gemini in Docs.

I am an HR manager,and I am developing a script for my presentation for new hires.I need to create the script for an onboarding presentation about our company’s commitment to employee development and well-being.Help me draft talking points that showcase why employee mentorship and development are core values for our company using @[Mission Statement and Core Values].

• Persona
• Task
• Context
• Format

Gemini in Docs: [Drafts talking points]

You select Insert. Now, you want to add more targeted talking points. You type:

Add four talking pointsfor a new section of the presentation script that explains how we support our employees’ development.Mention our training and certification programs and mentorship opportunities using@[Learning and Development Paths],and write a strong closing statementabout our expectation that everyone contributes to a respectful and welcoming workplace.Use a professional tone.

Gemini in Docs: [Adds talking points]

You add in more details and then you’re ready to create a draft of the Google Slides that will accompany your talking points.

Example use cases

Recruiter

NEW Use case: Report on recruitment metrics

The business is growing, and you have a large hiring effort underway. You want to see a holistic view of how your hiring efforts are going. You open your Google Sheet and prompt Gemini in the Sheets side panel. You type:

Help me create a formula to calculate the total total number of [hires] by [department].

You continue your conversation by prompting additional questions. You type:

In what month did we hire the most people?

You continue with your line of questions until you feel ready to write your report.

Use case: Manage the recruiting process

You want to brainstorm potential ways the company can better manage the recruiting process. You open the team’s Google Doc with recruiting strategies. You prompt Gemini in the Docs side panel. You type:

Create a list of strategies our recruiters can use to improve our existing recruiting process and identify potential job candidates.

After creating a short recommendation for leadership on how the team will improve existing recruiting processes, the team receives guidance for a job opening for a content marketing manager. You open a new Doc and prompt Gemini in Docs. You type:

I am opening a new job position on the marketing team. Write a compelling role description for a content marketing manager. Highlight key responsibilities [insert] and requirements, including B2B and B2C content creation, a minimum of five years experience, and a portfolio of writing examples.

Use case: Manage the interview process

You want to prepare questions for phone screen interviews. You decide to prepare by using Gemini Advanced.
You upload the relevant file and type:

I am a recruiter, and I am preparing for candidate interviews. Using the job description in the file I'm uploading, write a list of 20 open-ended interview questions that I can use to screen candidates.

Use case: Communicate with candidates

The team has made its hiring decisions. You open the Google Doc with notes on each candidate. You prompt Gemini in Docs by selecting Help me write. You type:

I am writing an email to a job candidate who just finished the interview process. Create a template for an offer letter for the [selected candidate] for the [position] with a request to schedule a call to discuss benefits, compensation, and start date.

Now, you want to generate personalized, empathetic email copy to send to the job candidates who will not receive an offer. You prompt Gemini in Docs by selecting Help me write. You type:

I am writing an email to job candidates who finished the interview process, but who were not selected. Help me write a rejection letter for [candidate] for the [position]. Use an empathetic tone.

HR Manager

NEW Use case: Conduct employee engagement and satisfaction surveys

You are in charge of building a survey that will go out to all employees. You want to brainstorm ideas on questions to ask. You visit Gemini Advanced and type:

I am an HR manager in charge of running our enterprise-wide survey at [company] to gauge employee engagement and satisfaction. Generate a list of questions I can use to build the survey.

Your company has completed its annual employee engagement and satisfaction survey. Now, you want to clean up the data before you analyze it. You go to Gemini Advanced, upload the relevant file, and type:

Help me clean my employee survey spreadsheet. Specifically, fill any blank values in the name column with “Anonymous,” and if the region column shows Headquarters, replace that with HQ. Finally, remove any rows where the satisfaction column is blank. Please generate a new file for me with my cleaned data.

NEW Use case: Create individualized learning and development plans

You have all of your company's learning resources stored in your Google Drive. For each new hire, you want to create a tailored learning and development plan. To do this, you prompt Gemini in the Drive side panel. You type:

Create a personalized learning and development plan for a new hire who needs to learn about [topic]. Organize it by day and suggest relevant files.

Use case: Onboard employees

The recruiters have just filled the company’s two open roles. Now, you’re in charge of ensuring the candidates have a smooth onboarding experience. You need help in structuring information for the new hires, so you open a Google Sheet and prompt Gemini in the Sheets side panel. You type:

Create a table that outlines a new employee's first-week schedule, including key meetings, training sessions, and introductions. Provide a column for key contacts and priority level (low, medium, high) for each activity.

Gemini in Sheets returns a formatted Google Sheet that you can now fill in with key contacts, meetings, and activities. The conditional formatting makes it easy for you to sort tasks by priority level with color-coded visual cues. Next, you need to create ways for the team to bond. You open a new Google Doc and prompt Gemini in the Docs side panel. You type:

Design a team-bonding activity, such as an office scavenger hunt, to have team members work together during their team meeting.

Gemini in Docs provides suggestions that help you brainstorm about the scavenger hunt. You tweak the outputs and get the idea approved by the team lead. Now, you need to communicate with the new hires about their first day when they will meet the team. You open Gmail and prompt Gemini in Gmail by selecting Help me write. You type:

Draft an email to the new employees on the [team] to meet the rest of their team and explain the team-building purposes of the meeting.

Use case: Communicate key findings and draft follow-up surveys

Now that you’ve finished onboarding new employees, you need to focus on ensuring that the latest company research data is easily understood by leadership. You’re committed to creating a welcoming environment for all employees where they can develop their skills. You open the Google Doc with the finalized report. You prompt Gemini in Docs by selecting Help me write. You type:

Draft an email to senior leadership that summarizes the key findings from our [report]. Include a short introductory paragraph with bullet points on the most important findings.

Gemini in Docs returns a summary with bullet points. You edit it and then use it to email the leadership team. As a follow-up action, you want to understand how changes made to company policies impact the employee experience. You open Gemini in Docs to begin drafting a survey. You select Help me write and type:

Draft an anonymous employee survey with questions and answer options to monitor company progress on [topics].

Marketing

As a marketing professional, you're the creative force behind captivating campaigns, brand experiences, lead generation, and more. You understand the power of data-driven insights, compelling messaging, and connecting with your audience on a deeper level.

This section provides you with simple ways to integrate prompts in your day-to-day tasks. For chief marketing officer (CMO) use cases, visit the Executives section of the guide.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Google Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

Use case: Develop a visual identity

You own your own consulting business and are often hired to function as a brand manager for your clients. You help businesses in a variety of industries. Your customer is getting ready to launch a coffee shop and video game cafe, and you need to kick-start the creative process by developing a visual identity. You want to ideate and provide early thoughts to the rest of the team. You decide to chat with Gemini Advanced. You type:

Generate ideas for a creative and eye-catching logofor my new business, a coffee shop combined with a video game cafe. Generate a logo considering the following:

Dual Concept: The logo needs to clearly signal both the coffee and gaming aspects of the business without being too cluttered.

Target Audience: Appeal to a wide range of gamers (casual and enthusiast), as well as coffee lovers seeking a unique hangout spot.

Style Options: I'm open to these approaches — let's get a few examples in each of these three styles to compare: Modern and Playful: Bold colors, fun graphics, maybe a pixel art aesthetic. Retro-Cool: Think classic arcade style — chunky lettering, neon color inspiration. Sleek and Minimalist: Clean lines, geometric shapes, a more subtle nod to both themes.

• Persona
• Task
• Context
• Format

You like the retro-cool options. You continue your conversation and you type:

I like the retro-cool options.Can you provide three morein that same style?

Now that you have a sense of what the logo could look like, you want to brainstorm names. You type:

Write a tagline and 10 potential namesfor the business to go with these logos.

Example use cases

Brand Manager

NEW Use case: Develop brand strategy

You are working to refresh your company's brand architecture and messaging frameworks. To get started, you visit Gemini Advanced and type:

I am a brand manager at [company]. Help me define a clear and effective brand architecture for [company], considering its diverse products and service portfolio. For additional context, here is our current brand portfolio: [List all existing brands, products, and services]. Here is our company mission and vision: [Provide a brief overview of the company’s mission and vision]. And these are our target audience(s): [describe target audience(s)]. Our desired brand positioning is [explain how the company wants to be perceived in the market].

NEW Use case: Brainstorm brand partnerships

You are working on a new brand campaign. You want to identify influencers or complementary brands you could partner with as part of the social amplification plan. You visit Gemini Advanced and type:

I am a [brand manager] at [company] working to launch a new campaign focused on [topic]. Identify potential types of influencers and complementary brands that [company] could partner with to amplify the [campaign] on social media channels. The goal is to reach a wide audience of [audiences], while building credibility and driving engagement.

Use case: Conduct market research and identify trends

The landscape in your industry is rapidly changing, and you need to conduct market research to better identify and understand emerging trends. You go to Gemini Advanced, and you type:

I need to do market research on [industry] industry to identify new trends. Use [URLs] to uncover emerging trends and shifting consumer preferences.

After completing your research, you and the team have new messaging that you want to A/B test. You want to generate multiple variations of ad copy using Gemini Advanced. You type:

I need to A/B test new messaging. Here is our messaging: [messaging]. Generate three different variations of ad copy.

Use case: Create and manage content and distribution

A customer has exciting organizational changes underway. You need to create content to shape the brand narrative of the company as it enters its next era. You open a Google Doc to get started on a blog draft. You prompt Gemini in Docs by selecting Help me write. You type:

Create a blog draft announcing that [name] is joining [company] as [position]. [Share two or three details from their bio, such as their previous position and company, their involvement in professional organizations, etc.].

You also want a way to efficiently track how and where this content is amplified, so you open a Google Sheet.
You prompt Gemini in the Sheets side panel. You type:

Create a project tracker for content amplification and include columns for channel, owner, URL, and priority level (low, medium, high).

Marketing Specialist

NEW Use case: Improve collaboration with customers, agencies, and teams

You are hosting a meeting discussing an upcoming project with multiple teams and an agency that will complete the project’s design work. You use Gemini in Google Meet and select Take notes with Gemini so that all participants can give their undivided attention to the conversation. After the meeting, Gemini provides a summary of the discussion and pulls out action items to keep the team on track.

From the generated Doc with call notes, you want to create a spreadsheet to help keep the team on track. You open a new Google Sheet and prompt Gemini in the Sheets side panel and tag relevant files by typing @file name. You type:

Generate a project tracker using the action items from @[Meeting Notes from Gemini].

NEW Use case: Analyze social media trends and other data to reduce time to market

You want to analyze different data sources and collate findings to help you reduce your time to market.
You open Gemini Advanced and type:

I am a [marketing specialist] at [company]. We are working on our [go to market] plans for [type of product]. Help me research social media trends around [topics]. Be specific about trending keywords, top influencer voices, and common themes in popular content.

You verify Gemini's response by selecting the Double-check response option beneath Gemini's response.

Now, you want to review a report you've commissioned that surveyed customers from different industries. You continue your conversation with Gemini. You upload the relevant file and type:

Analyze the findings in this [report]. I am especially interested in any common themes about [topic] that stand out to you that will help me better position [marketing materials] for [product] for [target audience].

(Gemini Advanced) ✦

NEW Use case: Perform audience research and develop personas

You need to refresh your audience research and persona development as the team updates webpage copy, pitch decks, and other marketing assets. You brainstorm and research using Gemini Advanced. You type:

I am a marketing specialist focused on [area] at [company]. I need to conduct in-depth audience research so that I can develop convincing marketing artifacts for [personas]. To start, help me generate a comprehensive profile of [target audience]. Include core demographics and psychographics, online platforms they frequent, key pain points [product] could solve, and language and messaging that resonates with them.

Digital Marketing Manager

NEW Use case: Create and optimize copy for search engine marketing (SEM)

You want to create a robust list of keywords and long-tail keywords and phrases to uncover new opportunities for SEM targeting. You go to Gemini Advanced and type:

I am a digital marketing manager at [company]. I am working on SEM ads for [product]. Here are my seed keywords: [list keywords]. Help me generate a list of additional keywords and long-tail keywords and phrases that can help me maximize ad performance.

After you finish brainstorming your keywords list, you want to generate a few variations of ad copy. You type:

For my SEM campaign, use these keywords as inspiration to generate multiple ad copy variations with different headlines, descriptions, and calls to action for [product]. Use a [tone] tone in the copy.

You want to further refine the text according to different audiences, so you type:

Do the same thing, except write new options for [audience], adjust the tone to be [tone] and focus the copy on highlighting [feature] of [product].

Use case: Draft customer acquisition communications

Email is one of your company's main channels of direct communication with prospects and customers. You want help getting started with copy for a new email campaign. You open a new Google Doc, and you prompt Gemini in Docs by selecting Help me write. You type:

Write three different email subject lines that reference [audience segments] and our [product]. Make them catchy but professional.

Now you want to share the proposed email subject lines with the copywriting team. You open Gmail, and you select Help me write. You type:

Write an email proposing [suggested email subject lines] to the copywriting team. Keep the email short and simple and request feedback by the end of week. Thank them for their help.

Use case: Generate inbound marketing campaigns

The team created a new ebook on best practices for executives using our new solution. You’re creating a landing page to house the gated asset, and you need engaging copy. You open a new Google Doc and select Help me write. You type:

Create compelling copy for a landing page promoting a new [ebook/webinar/free trial and details] designed for an executive target audience. Highlight key benefits and encourage conversions with persuasive calls to action.

The webpage launched, and you’re now running an inbound marketing campaign. You need to nurture the leads that downloaded your latest ebook. You open a new Google Doc, and you prompt Gemini in Docs by selecting Help me write. You type:

Generate copy for a sequence of five automated emails to nurture leads after they download the ebook on [topic]. Personalize emails and encourage further engagement [with other valuable resources or offers].

Content Marketing Manager

NEW Use case: Deliver personalized content to customers at scale

You want to create copy for a five-step email nurture cadence for your new product. You open a new Google Doc and prompt Gemini in the Docs side panel and tag relevant files by typing @file name. You type:

Create a 5-step nurture email cadence to [prospective customers] who have signed up for [our newsletter], with the goal of getting them to [purchase] [product] using @[Product Specific Notes] and @[Product FAQ].

NEW Use case: Create visuals for ad campaigns

You want to create visuals to help your creative agency better understand the team’s direction for an upcoming campaign. You open a new Google Slide and prompt Gemini in Slides. You type:

Help me create inspirational images for a marketing campaign for [type of product]. Images should use [colors] and [natural elements, such as clouds]. Use a [photorealistic] style.

Use case: Generate inspiration for your blog

You work for a travel company as the content marketing manager for the company’s blog channel. You need to kick-start the brainstorming process for a new blog post. You decide to gather ideas by collaborating with Gemini Advanced. You type:

Suggest blog post topics that would be interesting for people passionate about travel and the tourism industry. Here’s what I want you to focus on: Make the topics unique. There are lots of tourism blogs out there — let’s come up with fresh angles that would stand out. Keep the topics relevant. Tap into current trends or recent challenges/innovations within the tourism industry when brainstorming. I’d like each topic to include:

Target audience: Who would this topic specifically appeal to?

Content outline: A few bullet points with the main ideas the blog post would discuss.

Call to action: Suggest one way to engage the reader at the end of the post.

You love the initial ideas you were able to create. You also need to focus on generating creative imagery to accompany the copy in the blog. You type:

Create an image of a plane flying above the clouds over mountains and rivers during sunrise that I can use in the marketing campaign to promote my travel company.

Use case: Create social media posts

You're focused on creating content that is optimized for social media channels. You need to gather ideas for content targeted to distinct audiences. You open a new Google Doc and prompt Gemini in Docs by selecting Help me write. You type:

Write three engaging social media posts about [product/service/topic] that would appeal to [target audience]. Keep each social media post to two sentences and include a call to action to visit [our website].

You also need to craft social media posts to drive registration for an upcoming event targeting recent grads.
You open a new Google Doc and you prompt Gemini in Docs by selecting Help me write. You type:

Create a social media post promoting our upcoming [event name]. Include attention-grabbing language and relevant hashtags for [audience].

Use case: Create a strategic marketing plan

Your company is launching a new app. You need a robust marketing plan, but you want ideas to get started.
You chat with Gemini Advanced. You type:

I'm developing a marketing plan for a new app that provides [functionality]. My target audience is [audience]. Help me create a plan with a focus on [marketing channels]. Here's what I'd like you to cover: competitor analysis, ideal marketing channel mix with rationale, budget recommendations, key messaging ideas, and proposed campaign timeline with KPIs.

The responses from your chat are helpful in shaping your marketing plan. You need to get the high-level details to your chief marketing officer (CMO). You open Gmail and prompt Gemini in Gmail by selecting Help me write. You type:

Draft an email to the CMO telling them that I will provide a one-pager with a strategic marketing plan for the new app launch project by [date], and it will include an executive summary, overview of the competitive landscape, top marketing channels, and the target demographic for all South American markets.

Project management

As the conductor of complex, ever-evolving projects, your mission is to navigate timelines, coordinate teams, and ensure your programs deliver the intended impact.

This section provides you with simple ways to integrate prompts in your daily tasks.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Google Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

Use case: Generate user acceptance tests

Your team completed the registration form for a new website, and now you need to generate user acceptance tests (UATs). To start, you visit Gemini Advanced and type:

Create a table with 10 user acceptance tests (UAT)for the website registration form.

• Persona
• Task
• Context
• Format

You think the results are a helpful starting point, so you copy the results to a Google Sheet before drafting an email to your colleague who is running the UATs. You want to explain what they need to do. You continue your conversation with Gemini Advanced. You type:

Draft an emailto [my colleague] who is running this UAT and explain what they need to do next.

The drafted email provides a helpful starting point, so you export the results to Gmail, and you make edits directly before sending the message to your colleague.

Example use cases

Project Manager

Use case: Report on project status

You just had a lengthy call with all of your project stakeholders, and now you want to summarize what was discussed and follow up with assigned action items. In the Google Doc with the meeting transcript, you prompt Gemini in Docs. You type:

Summarize this call transcript in a short paragraph. In bullet points, highlight the action items, decisions made, and owners for each item based off of [call transcript].

You need to update your manager based on the activity from the last call. You want to templatize how your project status updates are delivered. You open a new Google Doc, and you prompt Gemini in Docs by selecting Help me write. You type:

Draft a project status update email template to send to my manager. Include sections for a summary of key accomplishments this week, any challenges faced, and the top three priorities for next week.

The team just hit its key milestones an entire week early. It’s been a challenging project, so you want to gather everyone to celebrate together. You open Gmail and prompt Gemini in Gmail by selecting Help me write.

You type:

Write an invitation for a team lunch to celebrate the progress made on a project and include [date, time, and location]. Thank them for all of their hard work and acknowledge that this has been a challenging project.

Use case: Create a project retrospective

You've just wrapped the project, and your senior leadership team needs a project retrospective. To kick-start the process of gathering feedback, you open a Google Doc and prompt Gemini in Docs by selecting Help me write. You type:

I need to write a report detailing the successes, failures, and lessons learned from [project]. Draft a list of 20 questions to guide a cross-team process investigation. Include questions to uncover what worked, what didn't, specific process breakdowns, technical issues, communication gaps, or any other potential contributing factors to the problem or success of the project.

The questions give you a great starting place. You edit them before sharing with the team for their input. After you gather everyone’s feedback, you want help structuring the report. You prompt Gemini in Docs by selecting Help me write. You type:

Summarize this document in two paragraphs. Include high-level information about the project’s goals, the main contributors, the outcomes of the project, and any key successes or failures.

Use case: Develop an issue tracker and related communications

You need to create a project issue tracker to keep track of risks and solve them in a timely manner. You want to create a template quickly, so you open a new Google Sheet and prompt Gemini in the Sheets side panel. You type:

Create a spreadsheet to track project issues, including descriptions, status, assigned owner, and action items for resolution.

Before the project fully kicks off, you want to have standardized communication templates at your disposal. For example, you want an email that can be used if an issue arises. You open a new Google Doc and prompt Gemini in Docs by selecting Help me write. You type:

Draft an email template to announce when an issue arises and include causes, solutions, and timelines to resolve it.

You like the template that Gemini in Docs creates, and you want to create an additional, slightly different email template. In the same Google Doc, you prompt Gemini in Docs by selecting Help me write. You type:

Draft an email template to a stakeholder to escalate a critical project issue, outlining the impact and proposed solution.

Technical Project Manager

NEW Use case: Create a workback schedule

You are the technical project manager for a software release. You already have the scope of the project documented. Now, you want to get started on building a workstream tracker and workback schedule. You go to Gemini Advanced and type:

I am a [technical project manager] at [company] overseeing [project and brief project description].
The project has the following scope: [scope]. Our project goals are: [project goals]. Our project deliverables are: [project deliverables]. Our budget is [budget], and our delivery date is [delivery date].
Help me create a workback schedule to keep the team on track. Include dates for key milestones and demos.

Sales

Understanding your customers inside and out is your ticket to success. You’re in charge of maintaining critical relationships, deciphering buying signals, crafting tailored solutions, driving revenue for the business, and more.

This section provides you with simple ways to integrate prompts in your daily tasks.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Google Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

Use case: Conduct customer research

You're an account executive, and you've just been assigned to a new customer. You need a research assistant. You will need to get to know key contacts at the account to begin building trust between your teams, but first, you want to send an introductory email, so you open Gmail and prompt Gemini in Gmail by selecting Help me write. You type:

Write an email to [name], the new [title] at [company].Congratulate them on their new role. Introduce me as their contact point at [company name]. Invite them to lunch next week and check if they prefer Monday or Tuesday.

• Persona
• Task
• Context
• Format

Gemini in Gmail: [Drafts email]

This provides a helpful starting point, but you want to try getting an even better response. You click Refine and Formalize.

You're happy with the email, so you click Insert. You read the message one last time, make final light edits directly, and then send the message. Now, you want to learn more about the customer and how it markets itself. To research, you visit Gemini Advanced and type:

I am an account executive in charge of a new account, [customer name].I need to do initial research. What is the market strategy of [customer]?

Gemini provides you with useful information to get started in your research. You continue your research by first focusing on news announcements. You gather a list of URLs, and you paste them into your conversation with Gemini Advanced. You type:

[URLs] Summarize these articles.Provide key insights and contextualize why these announcements are important.

Now you have a clear summary of what was announced, why the news is important, and additional insights. Next, you want to better understand the executive who will be your main point of contact. You find a recorded interview featuring the executive. You paste the YouTube URL into your conversation with Gemini Advanced and type:

[YouTube URL] Summarize this interview and tell me more about [executive name].What does [executive] care about?

You continue the conversation with additional lines of questioning to build familiarity with your key contact and the account. You prompt:

Tell me how [company] can help [customer company] with achieving their goals.

Once you wrap up your conversation, you export your results into a Google Doc. You open the Google Doc and prompt Gemini in Docs. You type:

Create an email draft for [customer] explaining why [your company] is the perfect partner for them to achieve their market goals.

Example use cases

Customer Success Manager

NEW Use case: Map customer journeys

It's your first time onboarding a new customer, and you realize you could benefit from creating custom-tailored assets. You open a new Doc and prompt Gemini in the Docs side panel and tag relevant files by typing @file name. You type:

Create personalized onboarding materials for [customer]. Use @[Standard Onboarding Documents] and @[New Customer Migration Notes] to personalize the assets.

Sales Manager

NEW Use case: Manage the request for proposal (RFP) process

You've just received an RFP, and you want to quickly ingest the request as part of your information gathering process. First, you want to do basic research on the company that issued the request. You visit Gemini Advanced and you type:

I just received an RFP from [company]. Before I dive into the RFP, I want your help in conducting research. Give me a business profile of the company including all of the basics (where they are located, what they provide for customers, who their target audience is, any recent news from the company). Be as detailed as possible as I want to see a full view of [the company].

Once you finish your research on the company, you want to summarize the RFP. You continue your conversation with Gemini. You type:

[URL or uploaded file] I am a sales manager at [company], and this is the RFP we’ve received from [company]. Summarize this content in a few paragraphs. What is the customer seeking, what is the budget, and when is a response due by?

NEW Use case: Access information and tools on your phone while on the go

You are working remotely from your phone. From the mobile app, you open a thread in Gmail and select the Gemini chip to Summarize this email. Gemini quickly provides you with a summary of the back and forth so that you can focus on the most important points.

Now, you want to generate a response acknowledging the latest developments. You prompt Gemini in Gmail.

You type:

Write a response to this email letting [them] know that I’ve received the message and will take [action] by [Friday].

Use case: Develop customer relationships

Your annual conference is coming up, and your most important prospects will be there. You want to personally invite them to a happy hour. You open Gmail and prompt Gemini in Gmail by selecting Help me write. You type:

Write an email inviting people interested in [focus area] to our happy hour taking place on [date, time] at [trade show event]. Include that we specialize in [focus area].

Now that the event is over, you want to follow up with customers who came to the happy hour. You open Gmail and prompt Gemini in Gmail by selecting Help me write. You type:

Draft an email thanking customers for their time at the happy hour on [date, time, location]. End with an invitation to continue the conversations in the next few weeks. Use a friendly tone.

You want to check in with the customers who attended workshops at the conference because their early feedback is important. You prompt Gemini in Docs. You type:

Draft 10 questions that I can use to survey customers about their recent experience with our [product/ service]. Include questions to gauge how useful [the product] is, what they liked, and what they thought could use improvement.

Use case: Support the sales team

You need to contact all of your team leads in the Southeast region to provide immediate guidance on how to proactively reach out to customers about an ongoing issue. You open Gmail and prompt Gemini in Gmail by selecting Help me write. You type:

Draft an email to all Southeast region sales leads. Inform them of [issues]. Advise them to communicate with their teams to contact their customers and offer a 20% discount on a future order as an apology.

Now, you need to email all of the regional team members. You open Gmail and prompt Gemini in Gmail by selecting Help me write. You type:

Draft an email to the regional sales representatives about an urgent meeting that needs to take place next week about the [issues]. Ask them to provide availability on Monday or Tuesday.

Use case: Coach and train the sales team

You've heard from many team members that they want more learning opportunities. You're organizing a half-day learning program to support this request. You need to create a schedule, so you open a new Google Doc and

prompt Gemini in Docs by selecting Help me write. You type:

Create a half-day agenda for an educational session on our latest technology [products] for sales teams. Include time for the product development team to present and include time for lunch.

As a follow up to the team meeting, you want to highlight different learning opportunities available. You open Google Sheets and prompt Gemini in the Sheets side panel. You type:

Create a spreadsheet that tracks online courses for sellers. Include columns for the course's main topic, price, duration, and priority level.

Account Manager and Account Executive

NEW Use case: Improve collaboration and execution by customizing sales materials

You are having an important meeting with a customer. From Google Meet, you turn on Transcription and activate Gemini in Meet by selecting Take notes with Gemini. The transcript provides an unedited Doc of what was said. The Take notes with Gemini file will generate notes recapping the meeting, important topics discussed, and action items. Now, you can fully engage with the customer conversation. (Gemini in Meet)

After the call, you want to send a recap message to the customer. You open a new message and prompt Gemini in the Gmail side panel and tag relevant files by typing @file name. You type:

Write a message to [customer] thanking them for their time at our last [meeting]. Provide a quick summary of the meeting and acknowledge any pain points discussed. Ask for additional time to discuss our [solution] using @[Customer Meeting Gemini Notes].

In preparation for your next meeting, you want to use the transcript and your existing sales materials to generate a customized asset that showcases how your company’s product solves the customer’s pain points mentioned during the call. To do this, you open a new Doc and prompt Gemini in the Docs side panel and tag relevant files by typing @file name. You type:

I am an [account manager] and I just finished a call with [customer]. I want to summarize the [pain points] mentioned by [customer] during our last meeting. Provide a list of direct quotes from @[Customer Call Transcript] where [customer] discusses what they are trying to solve.

You read through the summary of pain points and see that they capture what was discussed. You click Insert from the side panel. Then, you want to use your existing files to generate custom responses to each of their pain points. You prompt Gemini in the Docs side panel again and tag relevant files. You type:

I need to create convincing reasons why [customer] should adopt [product] to solve for [their pain points]. Write specific reasons why [product] from [company] could help them achieve their [business goals] using @[Product Sales Kit Full Assets].

Use case: Build customer relationships

You just had a great call with a customer and now you want to use the notes you took from the meeting in Google Docs to draft an email to the customer. In the Google Doc with your notes, you prompt Gemini in Docs by selecting Help me write. You type:

Compose a personalized follow-up email to [client] following an initial conversation. Summarize the key points we discussed and address any outstanding questions.

The account has just adopted one of the company's service offerings and you need to ensure that they feel supported during the onboarding process. You want to make sure you check in on how things are progressing once a week, but you want to explore what the emails could look like. You open a new Google Doc and prompt Gemini in Docs by selecting Help me write. You type:

Draft four email templates to check in on my customer weekly now that they have purchased our new [service]. Use one value proposition (cost, ease of use, security, availability, and customization) as the main topic for each email, and include [call to action] in each message.

Use case: Prepare for new customer calls

You have an upcoming call with a prospect. This is a brand new use case for you, and you need help preparing for the call. You visit Gemini Advanced, and you type:

Draft a customized script for me to follow during my sales call with a prospect. The call will happen over a video call and is set to last 30 minutes. Make sure to add the following in the script: how [company products/solutions] can help address potential customer’s pain points, how [company]’s delivery system guarantees seamless and timely delivery, competitive pricing and volume-discount table, and space for a customer reference in the [customer’s industry] industry.

Now that you’ve done initial research, you export your findings to a new Google Doc. You open the Google Doc and continue working. Now, you want to create a tailored pitch. Using the Google Doc with all of your research notes, you prompt Gemini in Docs by selecting Help me write. You type:

Generate an elevator pitch for [product name] and include key benefits, competitive points of differentiation, and the pain points that [product name] solves for.

You have a great start to your elevator pitch and short talking points. You want to use this to further anticipate how the customer call might go. You resume your meeting preparation by returning to Gemini Advanced.

You type:

I have an upcoming call with a prospect. [Use case] is a new use case for me, and I need help preparing for the call. List the most likely objections [customer] might have for me during a sales call, with suggestions on how to respond to them. I work in [industry], and I am trying to sell [product]. Also provide ideas on how to handle objections and suggest ways to respond.

Business Development Manager

Use case: Nurture relationships, personalized outreach, and thought leadership

You're hoping to build deeper relationships with prospective customers that you recently met. You want to draft a template that you can customize for multiple contacts. You open a new Google Doc, and you prompt Gemini in Docs. You type:

Draft an outreach email template to industry influencers. Express gratitude that we connected at [event], and propose collaboration opportunities such as [opportunities].

After having a successful call with prospective customers, you want to follow up with thought leadership content from your founder that they may find interesting. You open the Google Doc with the blog post, and prompt Gemini in Docs by selecting Help me write. You type:

Summarize this blog content in bullet points and generate three ideas for follow-up questions I can ask my customers about their thoughts.

Use case: Generate personalized customer appreciation materials

You want to personally thank your customers and check in. You open Gmail and prompt Gemini in Gmail by selecting Help me write. You type:

Generate a personalized email for [customer] on their one-month anniversary working with [company]. Thank them for being a customer. Ask them if they have any questions. Include information about [other product].

You also want to send these customers a gift to thank them. You open a Google Sheet and prompt Gemini in the Sheets side panel. You type:

Give me a list of gifts to send new clients that are under $200 and can be shipped to offices.

Small business owners and entrepreneurs

As the owner of a business, getting the most out of your working hours is critical when you’re juggling multiple roles and responsibilities. Understanding your market, delivering for your customers, and staying on top of many competing priorities is critical.

This section introduces you to AI prompts designed to simplify complex choices with AI data analysis, streamline your email inbox, and help you stand out with creative marketing tactics. Discover how Gemini for Google Workspace can help you unlock deep insights, foster collaboration, and help propel your company to new heights.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

Use case: Create pricing comparison

You are the owner of a local spa. You are evaluating offers you’ve received from two different cleaning companies. You want to find a company with the right price, flexibility, and level of service. You open a new Doc and prompt Gemini in the Docs side panel and tag relevant files by typing @file name in your prompt. You type:

I'm a business ownerand I'm trying to determine the right cleaning vendor using@[Company A Proposal] and @[Company B Proposal]. I need someone to come twice a week, and I want them to vacuum, mop, dust, clean the windows, and wipe down all surfaces. If available, include information about the booking and cancellation policy.Create a comparison table between the two companies’ proposals.

• Persona
• Task
• Context
• Format

Gemini returns a formatted table comparing the two proposals. After you make your decision, you go to your email and prompt Gemini in the Gmail side panel. You type:

Write an email to CompanyA thanking them for their time and their proposal.Ask for a few times to meet to schedule cleanings.

Example use cases

Owner

Use case: Enhance personal productivity

You have many important email messages to catch up on. You open your email and select an important thread. You open Gemini in the Gmail side panel, and it automatically summarizes the content.

Use case: Brainstorm and generate marketing content

As the business owner, you are also responsible for marketing your services via your social channels, your email-based newsletter, and email marketing. You aren’t sure where to start, so you chat with Gemini Advanced. You type:

I own a [type of business] in [location]. I am working on marketing materials to advertise [event/sale] on [services]. I want to focus on using this sale to bring in repeat customers who haven’t purchased in a while and new customers alike. I want the social posts to feel [inspirational] and [fun]. Suggest some social copy I can use on [social platform] with relevant hashtags, suggested newsletter copy, and two email drafts (one for existing customers and one for new customers).

You like the suggestions Gemini provided, so you select Share & export and Export to Docs. You want to continue your brainstorm, so you ask Gemini:

What are some other effective [event/sale] tactics I can use to bring in new customers? I don’t always want to offer discounts. Are there other incentives I am overlooking?

You continue your conversation with Gemini and are able to create a solid list of marketing tactics to try.

Use case: Develop a competitive analysis

You started a company, and your online business is gaining traction. You have always dreamed of opening a brick-and-mortar store, and now might be the perfect time. You want a thought partner to help you better understand the current landscape. You open Gemini Advanced, and you type:

I am an online business owner. I am considering opening a brick-and-mortar store. Conduct an analysis into the competitive landscape focusing on [focus area]. Provide the strengths and weaknesses of [key competitors] in this area, including their specific strategies, tactics, and results. Identify actionable insights and recommendations for how [my company] can improve its approach and gain a competitive advantage.

You gathered useful information from your discussion with Gemini Advanced. You want to go deeper in your brainstorming around two competitors in particular. You type:

Generate a competitive analysis of [company] versus [competitor] within the current market landscape.

You select Share & export and Export to Docs.

Use case: Conduct fundraising and investor relations

You're ready to reach out to potential investors to make your brick-and-mortar store a reality. You want help getting started on an email to investors, so in the same Google Doc with your competitive analysis research, you prompt Gemini in Docs. You type:

Draft a personalized email template to potential investors, highlighting [company's] unique value proposition and recent progress on [initiatives]. Request a time to meet to discuss opportunities to collaborate in the next month.

The email template gives you a starting place. You tweak the draft and continue to add a few personal touches before sending the email to the potential investors. After a successful meeting with them a month later, you want to draft a thank you message. You open your Google Doc with the meeting transcript and notes. You prompt Gemini in Docs to help you write an email draft. You type:

Draft an email thanking a potential investor for the call and ask for time to schedule a follow-up meeting to address [questions and concerns].

Use case: Manage time off policies and tracking

You have a lengthy handbook detailing all of your company’s policies and procedures. You want to make the time-off request policy easily digestible for new hires. You open the Google Doc with the handbook. You prompt Gemini in Docs by selecting Help me write. You type:

Generate a step-by-step checklist summarizing the company’s time-off request policy. Ensure it is written in plain language and easy for employees to understand.

You need a quick way to track staffing each week because many of your employees are shift-based. You open Gemini in the Sheets side panel. You type:

Create a table that tracks weekly staffing. Create columns for date, name, shift (AM or PM), and notes.

Startup leaders

You thrive in fast-paced, dynamic environments where you can wear many hats and make a tangible impact. You’re driven by a passion for innovation, a desire to learn and grow, and a tolerance for risk. Your work is unique in its variety, its potential for high reward, and its direct connection to the company’s success. You’re not just executing tasks; you’re building something from the ground up, shaping the future of your company, and potentially disrupting entire industries.

Gemini for Google Workspace can help you redefine productivity and foster meaningful connections with investors, customers, and coworkers. This section provides practical prompts and real-world use cases designed specifically for you and your team. Learning to write effective prompts with Gemini for Workspace will help improve your productivity and streamline your everyday tasks, giving you more time to focus on strategic work.

Getting started

First, review the general prompt-writing tips on page 2 and the Prompting 101 section at the beginning of this guide.

Each prompt below is presented with an accompanying scenario to serve as inspiration for how you can collaborate with Gemini for Workspace. The prompt iteration example shows how you could write follow-up prompts to build on the initial generated response.

Prompt iteration example

Use case: Brainstorm business and strategy

You just had a productive planning and strategy brainstorming session with colleagues and you took many notes physically on a whiteboard. You snap a quick image with your phone and upload it directly to Gemini Advanced.
You type:

I am a founder at a startup focused on [industry].I was brainstorming with colleagues about [topic], and we took notes on this whiteboard. Turn these notes into text.

• Persona
• Task
• Context
• Format

Now you want to proactively continue brainstorming before you recap all of the ideas and notes for the group in a follow-up email. You continue the conversation and type:

Suggest follow-up items we could discuss for our [topic of brainstorm session].What was not covered that could have been, and what are we potentially missing?

You save all of your notes by clicking Share & export and Export to Docs. You are ready to send the recap message to the team, so you open your email and prompt Gemini in the Gmail side panel and tag the relevant file of notes by typing @file name. You type:

Use @[Brainstorm Notes and Ideas 9/1/24]to write a meeting recap to the teamusing an upbeat and friendly tone.Share some of the ideas I have for our next meeting to discuss [topic].

Example use cases

Founder

Use case: Create an elevator pitch (speech to text)

You're scheduled to present to a group of prospective investors. This will be your first time discussing your business with this audience. You need to work on your elevator pitch, so you chat with Gemini Advanced using your voice to prompt. You select the microphone icon and say:

I'm the founder of [startup] in [industry], and I need help creating a short elevator pitch for [company and product description]. I need to make the pitch relevant to [audience] and I want to especially highlight [key features of product] because I want them to [take this action]. Include a compelling hook and anticipate questions an investor might have. Make the tone professional but relaxed and confident.

Use case: Develop your personal brand

As your company grows, you're working on increasing your social media presence, so you want to define and hone your personal brand. To brainstorm, you turn to Gemini Advanced. You type:

Help me grow my personal brand. I am the founder of [a startup] in [industry]. I am passionate about [topics]. I want to inspire [audience] with business tips and lessons I’ve learned from starting my own company. My goals are to build a following so that I can [generate more media] for the business.

Gemini returns insights into how you can begin to build messaging and content that aligns to your personal brand and that can help you achieve your goals.

Head of Operations

Use case: Communicate and negotiate with vendors

You've received a quote from two different manufacturers to create packaging for the company's new product. You want to compare and contrast the offers before you negotiate. You open a new Doc and prompt Gemini in the Docs side panel and reference relevant files by typing @file name. You type:

I need to make a vendor decision for packaging manufacturing. Create a table that compares the two proposals I’ve received @[Company A’s Proposal] and @[Company B’s Proposal].

Gemini creates a table comparing the two different proposals. You make a decision, but now you want to see if you can negotiate with your preferred vendor. You go to your inbox and start a new email draft. You prompt Gemini in the Gmail side panel. You type:

Create an email draft to [selected vendor] telling them that I’ve decided to move forward with them as the [packaging] vendor, but I would like to negotiate [a bulk pricing discount]. Use a collaborative tone.

Gemini in Gmail returns a drafted message that is ready to send. You select Insert and send the email.

Use case: Plan and track budgets

You're in planning mode and you first want to understand where previous years' budgets were spent. You have all of this data in a Sheet. You decide to chat with Gemini Advanced. You upload the Sheet and prompt Gemini by typing:

Using the attached spreadsheet, identify trends and patterns in our expenses by category over the last three years. Identify areas where costs have increased significantly and investigate potential reasons.

Gemini returns a response that helps inform your budget proposal for next year.

Head of Product

Use case: Develop a product launch plan

Your team is creating a new product, and you want to conduct research to inform your launch plan in collaboration with the marketing team. Using Gemini, you want to simulate different launch scenarios based on factors like pricing, marketing strategies, and target audience. You go to Gemini Advanced to conduct research and type:

I am head of product at [startup] in [industry] industry. We are building a product launch plan for [product]. I want to brainstorm a few different scenarios. We are considering offering the [product] at two different price points [A and B] and we are considering launching in [December or January]. Provide pros and cons of each scenario and suggest different ideas we may not have considered.

You want to continue market research brainstorming. You type:

How do these prices compare to [competitor products'] prices? Detail what pricing strategies [competitors] use for [products], and list any common tactics they use (such as free trials, discounts, etc.). Summarize how they position the product to [audience]. Cite your sources.

Your research helps you refine your pricing structure and go-to-market strategy for your most important target audience.

Use case: Develop product strategy and roadmap

You want to refine your product strategy and roadmap. You’ve collected user feedback in a spreadsheet, and you want to clean it up so that it is ready for deeper analysis. You chat with Gemini Advanced and upload a file. You type:

Help me clean my [user feedback] survey spreadsheet. Specifically, fill any blank values in the name column with “Anonymous,” then if the [recommend] column shows [Yes], replace that with [Y]. Finally, remove any rows where the satisfaction column is blank. Please generate a new file for me with my cleaned data.

Gemini returns a clean file for you to conduct deeper analysis on, and from this file, you notice a few trends. You have alignment from the team on features to address recurring user feedback, and now you want to build a high-level roadmap that you can use as a starting point. You continue your conversation with Gemini Advanced. You type:

I am head of product at [startup] in the [industry] industry. We are adding [features] to our [product] to address recurring user feedback, including [feedback trends]. Build a high-level roadmap that will keep us on track for a Q4 delivery. Put it in a table format.

Gemini returns a helpful starting point. You want to save the work so you click Export to Docs.

Leveling up your prompt writing

This guide is meant to serve as inspiration, and the possibilities are nearly endless with Gemini for Google Workspace. Build on your prompt-writing skills using these tips.

• Break it up. If you want Gemini for Workspace to perform several related tasks, break them into separate prompts.

• Give constraints. To generate specific results, include details in your prompt such as character count limits or the number of options you’d like to generate.

• Assign a role. To encourage creativity, assign a role. You can do this by starting your prompt with language like: "You are the head of a creative department for a leading advertising agency ..."

• Ask for feedback. In your conversation with Gemini Advanced, tell it that you’re giving it a project, include all the relevant details, and then describe the output you want. Continue the conversation by asking questions like, “What questions do you have for me that would help you provide the best output?”

• Consider tone. Tailor your prompts to suit your intended audience. Ask for outputs to have a specific tone, such as formal, informal, technical, creative, or casual.

• Say it another way. Fine-tune your prompts if the results don’t meet your expectations or if you believe there’s room for improvement. An iterative process of review and refinement often yields better results.

Generative AI and all of its possibilities are exciting, but it’s still new. Even though our models are getting better every day, prompts can sometimes have unpredictable responses.

Before putting an output from Gemini for Workspace into action, review it to ensure clarity, relevance, and accuracy. And of course, keep the most important thing in mind: Generative AI is meant to help humans, but the final output is yours.

The example prompts in this guide are meant for illustrative purposes.

基于 CANN 社区版 8.2.RC1 制作

Ascend C 算子开发

• 1. Ascend C简介
• 2. 环境准备
• 3. 快速入门
     • 3.1. HelloWorld
     • 3.2. 基于Kernel直调工程的算子开发
     • 3.3. 基于自定义算子工程的算子开发
• 4. 抽象硬件架构
• 5. 编程模型
     • 5.1. SPMD模型
     • 5.2. 核函数
     • 5.3. 编程范式
     • 5.4. 编程接口概述
• 6. 算子实现
     • 6.1. 概述
     • 6.2. 矢量编程
       • 6.2.1. 概述
       • 6.2.2. 基础矢量算子
       • 6.2.3. TBuf的使用
       • 6.2.4. 多核&Tiling切分
         • 6.2.4.1. 概述
         • 6.2.4.2. 多核Tiling
         • 6.2.4.3. 尾块Tiling
         • 6.2.4.4. 尾核Tiling
         • 6.2.4.5. 尾核&尾块
       • 6.2.5. DoubleBuffer场景
       • 6.2.6. Broadcast场景
       • 6.2.7. 非对齐场景
     • 6.3. 矩阵编程（高阶API）
       • 6.3.1. 基础知识
       • 6.3.2. 算子实现
       • 6.3.3. 特性场景
         • 6.3.3.1. Matmul特性介绍
         • 6.3.3.2. 多核对齐切分
         • 6.3.3.3. 多核非对齐切分
         • 6.3.3.4. 异步场景处理
         • 6.3.3.5. 矩阵乘输出的量化/反量化
         • 6.3.3.6. 矩阵乘输出的Channel拆分
         • 6.3.3.7. 矩阵向量乘
         • 6.3.3.8. 4:2稀疏矩阵乘
         • 6.3.3.9. TSCM输入的矩阵乘
         • 6.3.3.10. 矩阵乘输出的N方向对齐
         • 6.3.3.11. 单次矩阵乘局部输出
         • 6.3.3.12. AIC和AIV独立运行机制
         • 6.3.3.13. Batch Matmul基础功能
     • 6.4. 矩阵编程（基础API）
       • 6.4.1. 耦合架构
       • 6.4.2. 分离架构
     • 6.5. 融合算子编程
       • 6.5.1. 基础知识
       • 6.5.2. 算子实现
     • 6.6. Kernel直调算子开发
       • 6.6.1. 概述
       • 6.6.2. Kernel直调
       • 6.6.3. Pybind调用
     • 6.7. 工程化算子开发
       • 6.7.1. 概述
       • 6.7.2. 创建算子工程
       • 6.7.3. 算子原型定义
       • 6.7.4. Kernel侧算子实现
       • 6.7.5. Host侧tiling实现
       • 6.7.6. 算子包编译
         • 6.7.6.1. 算子工程编译
         • 6.7.6.2. 算子包部署
       • 6.7.7. 算子动态库编译
       • 6.7.8. 单算子API调用
• 7. 算子调试调优
     • 7.1. 孪生调试简介
     • 7.2. CPU域调试
     • 7.3. NPU域调试调优
• 8. 算子入图（GE图）开发
     • 8.1. 概述
     • 8.2. 开发流程
     • 8.3. Tiling下沉
     • 8.4. 图编译和图执行
• 9. AI框架算子适配
     • 9.1. 概述
     • 9.2. PyTorch框架
     • 9.3. ONNX框架
       • 9.3.1. 适配插件开发
       • 9.3.2. 调用样例
     • 9.4. TensorFlow框架
• 10. 概念原理和术语
      • 10.1. 术语表
      • 10.2. 神经网络和算子
        • 10.2.1. 算子基本概念
        • 10.2.2. 数据排布格式
      • 10.3. 硬件架构与数据处理原理
        • 10.3.1. 基本架构
        • 10.3.2. 计算单元
        • 10.3.3. 存储单元
        • 10.3.4. 控制单元
      • 10.4. 编程模型设计原理
      • 10.5. 性能优化技术原理
        • 10.5.1. DoubleBuffer
• 11. 常用操作
      • 11.1. 如何开发动态输入算子
      • 11.2. 如何在矢量编程时使能Vector Core
      • 11.3. 如何开发通算融合算子
      • 11.4. 如何使用Tensor高维切分计算API
      • 11.5. 如何使用归约指令
      • 11.6. 如何使用掩码操作API
      • 11.7. 如何使用workspace
      • 11.8. 如何进行Tiling调测
      • 11.9. 如何使用SPM Buffer
      • 11.10. 如何使用更底层编程方式提升算子性能
      • 11.11. 如何使用Tensor原地操作提升算子性能
• 12. 附录
      • 12.1. 简易自定义算子工程
      • 12.2. showkerneldebug_data工具
      • 12.3. msobjdump工具
• 13. FAQ
      • 13.1. 核函数运行验证时算子存在精度问题
      • 13.2. 运行验证时AllocTensor/FreeTensor失败
      • 13.3. kernel侧获取Tiling信息不正确
      • 13.4. Kernel编译时报错“error: out of jump/jumpc imm range”
      • 13.5. 使用跨版本的自定义算子包时，含有Matmul高阶API的算子存在编译或执行报错
      • 13.6. 含有Matmul高阶API的算子精度问题
      • 13.7. 算子工程编译时出现文件名过长报错
      • 13.8. 调用算子时出现无法打开config.ini的报错
      • 13.9. 算子包部署时出现权限不足报错

Ascend C API

• 1. Ascend C API列表
• 2. 通用约束
• 3. 数据类型定义
     • 3.1. LocalTensor
       • 3.1.1. 简介
       • 3.1.2. 构造函数
       • 3.1.3. SetValue
       • 3.1.4. GetValue
       • 3.1.5. operator[]
       • 3.1.6. operator()
       • 3.1.7. SetSize
       • 3.1.8. GetSize
       • 3.1.9. SetUserTag
       • 3.1.10. GetUserTag
       • 3.1.11. ReinterpretCast
       • 3.1.12. GetPhyAddr
       • 3.1.13. GetPosition
       • 3.1.14. GetLength
       • 3.1.15. SetShapeInfo
       • 3.1.16. GetShapeInfo
       • 3.1.17. SetAddrWithOffset
       • 3.1.18. SetBufferLen
       • 3.1.19. ToFile
       • 3.1.20. Print
     • 3.2. GlobalTensor
       • 3.2.1. 简介
       • 3.2.2. 构造函数
       • 3.2.3. SetGlobalBuffer
       • 3.2.4. GetPhyAddr
       • 3.2.5. GetValue
       • 3.2.6. operator()
       • 3.2.7. SetValue
       • 3.2.8. GetSize
       • 3.2.9. operator[]
       • 3.2.10. SetShapeInfo
       • 3.2.11. GetShapeInfo
       • 3.2.12. SetL2CacheHint
     • 3.3. ShapeInfo
     • 3.4. ListTensorDesc
     • 3.5. TensorDesc
       • 3.5.1. 简介
       • 3.5.2. 构造和析构函数
       • 3.5.3. SetShapeAddr
       • 3.5.4. GetDim
       • 3.5.5. GetIndex
       • 3.5.6. GetShape
       • 3.5.7. GetDataPtr
       • 3.5.8. GetDataObj
     • 3.6. TensorTrait
     • 3.7. Layout
       • 3.7.1. 简介
       • 3.7.2. 构造函数
       • 3.7.3. layout
       • 3.7.4. GetShape
       • 3.7.5. GetStride
       • 3.7.6. MakeShape
       • 3.7.7. MakeStride
       • 3.7.8. MakeLayout
       • 3.7.9. is_layout
     • 3.8. LocalMemAllocator
       • 3.8.1. 简介
       • 3.8.2. 构造函数
       • 3.8.3. GetCurAddr
       • 3.8.4. Alloc
     • 3.9. UnaryRepeatParams
     • 3.10. BinaryRepeatParams
• 4. 基础API
     • 4.1. 标量计算
       • 4.1.1. ScalarGetCountOfValue
       • 4.1.2. ScalarCountLeadingZero
       • 4.1.3. ScalarCast
       • 4.1.4. CountBitsCntSameAsSignBit
       • 4.1.5. ScalarGetSFFValue
       • 4.1.6. ToBfloat16
       • 4.1.7. ToFloat
     • 4.2. 矢量计算
       • 4.2.1. 单目指令
         • 4.2.1.1. Exp
         • 4.2.1.2. Ln
         • 4.2.1.3. Abs
         • 4.2.1.4. Reciprocal
         • 4.2.1.5. Sqrt
         • 4.2.1.6. Rsqrt
         • 4.2.1.7. Not
         • 4.2.1.8. Relu
         • 4.2.1.9. VectorPadding(ISASI)
         • 4.2.1.10. 更多样例
       • 4.2.2. 双目指令
         • 4.2.2.1. Add
         • 4.2.2.2. Sub
         • 4.2.2.3. Mul
         • 4.2.2.4. Div
         • 4.2.2.5. Max
         • 4.2.2.6. Min
         • 4.2.2.7. And
         • 4.2.2.8. Or
         • 4.2.2.9. AddRelu
         • 4.2.2.10. AddReluCast
         • 4.2.2.11. AddDeqRelu
         • 4.2.2.12. SubRelu
         • 4.2.2.13. SubReluCast
         • 4.2.2.14. MulAddDst
         • 4.2.2.15. MulCast
         • 4.2.2.16. FusedMulAdd
         • 4.2.2.17. FusedMulAddRelu
         • 4.2.2.18. BilinearInterpolation(ISASI)
         • 4.2.2.19. 更多样例
       • 4.2.3. 标量双目指令
         • 4.2.3.1. Adds
         • 4.2.3.2. Muls
         • 4.2.3.3. Maxs
         • 4.2.3.4. Mins
         • 4.2.3.5. ShiftLeft
         • 4.2.3.6. ShiftRight
         • 4.2.3.7. LeakyRelu
         • 4.2.3.8. 更多样例
       • 4.2.4. 标量三目指令
         • 4.2.4.1. Axpy
         • 4.2.4.2. 更多样例
       • 4.2.5. 比较指令
         • 4.2.5.1. Compare
         • 4.2.5.2. Compare（结果存入寄存器）
         • 4.2.5.3. CompareScalar
         • 4.2.5.4. GetCmpMask(ISASI)
         • 4.2.5.5. SetCmpMask(ISASI)
       • 4.2.6. 选择指令
         • 4.2.6.1. Select
         • 4.2.6.2. GatherMask
       • 4.2.7. 精度转换指令
         • 4.2.7.1. Cast
         • 4.2.7.2. CastDeq
       • 4.2.8. 归约指令
         • 4.2.8.1. ReduceMax
         • 4.2.8.2. ReduceMin
         • 4.2.8.3. ReduceSum
         • 4.2.8.4. WholeReduceMax
         • 4.2.8.5. WholeReduceMin
         • 4.2.8.6. WholeReduceSum
         • 4.2.8.7. BlockReduceMax
         • 4.2.8.8. BlockReduceMin
         • 4.2.8.9. BlockReduceSum
         • 4.2.8.10. PairReduceSum
         • 4.2.8.11. RepeatReduceSum
         • 4.2.8.12. GetAccVal(ISASI)
         • 4.2.8.13. GetReduceMaxMinCount(ISASI)
       • 4.2.9. 数据转换
         • 4.2.9.1. Transpose
         • 4.2.9.2. TransDataTo5HD
       • 4.2.10. 数据填充
         • 4.2.10.1. Duplicate
         • 4.2.10.2. Brcb
         • 4.2.10.3. CreateVecIndex
       • 4.2.11. 排序组合(ISASI)
         • 4.2.11.1. ProposalConcat
         • 4.2.11.2. ProposalExtract
         • 4.2.11.3. RpSort16
         • 4.2.11.4. MrgSort4
         • 4.2.11.5. Sort32
         • 4.2.11.6. MrgSort
         • 4.2.11.7. GetMrgSortResult
       • 4.2.12. 数据分散/数据收集
         • 4.2.12.1. Gather
         • 4.2.12.2. Gatherb(ISASI)
         • 4.2.12.3. Scatter(ISASI)
       • 4.2.13. 掩码操作
         • 4.2.13.1. SetMaskCount
         • 4.2.13.2. SetMaskNorm
         • 4.2.13.3. SetVectorMask
         • 4.2.13.4. ResetMask
       • 4.2.14. 量化设置
         • 4.2.14.1. SetDeqScale
     • 4.3. 数据搬运
       • 4.3.1. DataCopy
         • 4.3.1.1. 普通数据搬运
         • 4.3.1.2. 增强数据搬运
         • 4.3.1.3. 切片数据搬运
         • 4.3.1.4. 随路格式转换
       • 4.3.2. Copy
       • 4.3.3. DataCopyPad(ISASI)
       • 4.3.4. SetPadValue(ISASI)
       • 4.3.5. SetFixPipeConfig(ISASI)
       • 4.3.6. SetFixpipeNz2ndFlag(ISASI)
       • 4.3.7. SetFixpipePreQuantFlag(ISASI)
       • 4.3.8. SetFixPipeClipRelu(ISASI)
       • 4.3.9. SetFixPipeAddr(ISASI)
     • 4.4. 内存管理与同步控制
       • 4.4.1. TPipe
         • 4.4.1.1. TPipe构造函数
         • 4.4.1.2. Init
         • 4.4.1.3. Destroy
         • 4.4.1.4. InitBuffer
         • 4.4.1.5. Reset
         • 4.4.1.6. AllocEventID
         • 4.4.1.7. ReleaseEventID
         • 4.4.1.8. FetchEventID
         • 4.4.1.9. GetBaseAddr
         • 4.4.1.10. InitBufPool
       • 4.4.2. GetTPipePtr
       • 4.4.3. TBufPool
         • 4.4.3.1. 简介
         • 4.4.3.2. 构造函数
         • 4.4.3.3. InitBufPool
         • 4.4.3.4. InitBuffer
         • 4.4.3.5. Reset
       • 4.4.4. 自定义TBufPool
         • 4.4.4.1. EXTERNIMPLBUFPOOL宏
         • 4.4.4.2. Reset
         • 4.4.4.3. Init
         • 4.4.4.4. GetBufHandle
         • 4.4.4.5. SetCurAddr
         • 4.4.4.6. GetCurAddr
         • 4.4.4.7. SetCurBufSize
         • 4.4.4.8. GetCurBufSize
       • 4.4.5. TQue
         • 4.4.5.1. 简介
         • 4.4.5.2. AllocTensor
         • 4.4.5.3. FreeTensor
         • 4.4.5.4. EnQue
         • 4.4.5.5. DeQue
         • 4.4.5.6. VacantInQue
         • 4.4.5.7. HasTensorInQue
         • 4.4.5.8. GetTensorCountInQue
         • 4.4.5.9. HasIdleBuffer
       • 4.4.6. TQueBind
         • 4.4.6.1. 简介
         • 4.4.6.2. 构造函数
         • 4.4.6.3. AllocTensor
         • 4.4.6.4. FreeTensor
         • 4.4.6.5. EnQue
         • 4.4.6.6. DeQue
         • 4.4.6.7. VacantInQue
         • 4.4.6.8. HasTensorInQue
         • 4.4.6.9. GetTensorCountInQue
         • 4.4.6.10. HasIdleBuffer
         • 4.4.6.11. FreeAllEvent
         • 4.4.6.12. InitBufHandle
         • 4.4.6.13. InitStartBufHandle
       • 4.4.7. TBuf
         • 4.4.7.1. TBuf简介
         • 4.4.7.2. 构造函数
         • 4.4.7.3. Get
         • 4.4.7.4. GetWithOffset
       • 4.4.8. SPM Buffer
         • 4.4.8.1. InitSpmBuffer
         • 4.4.8.2. WriteSpmBuffer
         • 4.4.8.3. ReadSpmBuffer
       • 4.4.9. workspace
         • 4.4.9.1. GetSysWorkSpacePtr
         • 4.4.9.2. SetSysWorkSpace
         • 4.4.9.3. GetUserWorkspace
       • 4.4.10. 核内同步
         • 4.4.10.1. TQueSync
           • 4.4.10.1.1. 同步控制简介
           • 4.4.10.1.2. 模板参数
           • 4.4.10.1.3. SetFlag/WaitFlag
         • 4.4.10.2. SetFlag/WaitFlag(ISASI)
         • 4.4.10.3. PipeBarrier(ISASI)
         • 4.4.10.4. DataSyncBarrier(ISASI)
       • 4.4.11. 核间同步
         • 4.4.11.1. IBSet
         • 4.4.11.2. IBWait
         • 4.4.11.3. SyncAll
         • 4.4.11.4. CrossCoreSetFlag(ISASI)
         • 4.4.11.5. CrossCoreWaitFlag(ISASI)
         • 4.4.11.6. InitDetermineComputeWorkspace
         • 4.4.11.7. NotifyNextBlock
         • 4.4.11.8. WaitPreBlock
       • 4.4.12. 任务间同步
         • 4.4.12.1. SetNextTaskStart
         • 4.4.12.2. WaitPreTaskEnd
       • 4.4.13. TPosition
     • 4.5. 缓存处理
       • 4.5.1. DataCachePreload
       • 4.5.2. DataCacheCleanAndInvalid
       • 4.5.3. ICachePreLoad(ISASI)
       • 4.5.4. GetICachePreloadStatus(ISASI)
     • 4.6. 系统变量访问
       • 4.6.1. GetBlockNum
       • 4.6.2. GetBlockIdx
       • 4.6.3. GetDataBlockSizeInBytes
       • 4.6.4. GetArchVersion
       • 4.6.5. GetTaskRation
       • 4.6.6. InitSocState
       • 4.6.7. GetProgramCounter(ISASI)
       • 4.6.8. GetSubBlockNum(ISASI)
       • 4.6.9. GetSubBlockIdx(ISASI)
       • 4.6.10. GetSystemCycle(ISASI)
     • 4.7. 原子操作
       • 4.7.1. SetAtomicAdd
       • 4.7.2. SetAtomicType
       • 4.7.3. SetAtomicNone
       • 4.7.4. SetAtomicMax(ISASI)
       • 4.7.5. SetAtomicMin(ISASI)
       • 4.7.6. SetStoreAtomicConfig(ISASI)
       • 4.7.7. GetStoreAtomicConfig(ISASI)
     • 4.8. 矩阵计算(ISASI)
       • 4.8.1. InitConstValue
       • 4.8.2. LoadData
       • 4.8.3. LoadDataWithTranspose
       • 4.8.4. SetAippFunctions
       • 4.8.5. LoadImageToLocal
       • 4.8.6. LoadUnzipIndex
       • 4.8.7. LoadDataUnzip
       • 4.8.8. LoadDataWithSparse
       • 4.8.9. SetFmatrix
       • 4.8.10. SetLoadDataBoundary
       • 4.8.11. SetLoadDataRepeat
       • 4.8.12. SetLoadDataPaddingValue
       • 4.8.13. Mmad
       • 4.8.14. MmadWithSparse
       • 4.8.15. Fixpipe
       • 4.8.16. SetHF32Mode
       • 4.8.17. SetHF32TransMode
       • 4.8.18. SetMMLayoutTransform
       • 4.8.19. CheckLocalMemoryIA
       • 4.8.20. Conv2D
       • 4.8.21. Gemm
     • 4.9. 资源管理(ISASI)
       • 4.9.1. CubeResGroupHandle
         • 4.9.1.1. 使用说明
         • 4.9.1.2. CubeResGroupHandle构造函数
         • 4.9.1.3. CreateCubeResGroup
         • 4.9.1.4. AssignQueue
         • 4.9.1.5. AllocMessage
         • 4.9.1.6. PostMessage
         • 4.9.1.7. PostFakeMsg
         • 4.9.1.8. SetQuit
         • 4.9.1.9. Wait
         • 4.9.1.10. FreeMessage
         • 4.9.1.11. SetSkipMsg
       • 4.9.2. GroupBarrier
         • 4.9.2.1. 使用说明
         • 4.9.2.2. GroupBarrier构造函数
         • 4.9.2.3. Arrive
         • 4.9.2.4. Wait
         • 4.9.2.5. GetWorkspaceLen
       • 4.9.3. KfcWorkspace
         • 4.9.3.1. 构造函数与析构函数
         • 4.9.3.2. UpdateKfcWorkspace
         • 4.9.3.3. GetKfcWorkspace
     • 4.10. Kernel Tiling
       • 4.10.1. GETTILINGDATA
       • 4.10.2. GETTILINGDATAWITHSTRUCT
       • 4.10.3. GETTILINGDATA_MEMBER
       • 4.10.4. TILINGKEYIS
       • 4.10.5. REGISTERTILINGDEFAULT
       • 4.10.6. REGISTERTILINGFOR_TILINGKEY
       • 4.10.7. 设置kernel类型
• 5. 高阶API
     • 5.1. 数学库
       • 5.1.1. Tanh
         • 5.1.1.1. Tanh
         • 5.1.1.2. GetTanhMaxMinTmpSize
         • 5.1.1.3. GetTanhTmpBufferFactorSize
       • 5.1.2. Asin
         • 5.1.2.1. Asin
         • 5.1.2.2. GetAsinMaxMinTmpSize
         • 5.1.2.3. GetAsinTmpBufferFactorSize
       • 5.1.3. Sin
         • 5.1.3.1. Sin
         • 5.1.3.2. GetSinMaxMinTmpSize
         • 5.1.3.3. GetSinTmpBufferFactorSize
       • 5.1.4. Acos
         • 5.1.4.1. Acos
         • 5.1.4.2. GetAcosMaxMinTmpSize
         • 5.1.4.3. GetAcosTmpBufferFactorSize
       • 5.1.5. Cos
         • 5.1.5.1. Cos
         • 5.1.5.2. GetCosMaxMinTmpSize
         • 5.1.5.3. GetCosTmpBufferFactorSize
       • 5.1.6. Log
         • 5.1.6.1. Log
         • 5.1.6.2. GetLogMaxMinTmpSize
         • 5.1.6.3. GetLogTmpBufferFactorSize
       • 5.1.7. Atan
         • 5.1.7.1. Atan
         • 5.1.7.2. GetAtanMaxMinTmpSize
         • 5.1.7.3. GetAtanTmpBufferFactorSize
       • 5.1.8. Power
         • 5.1.8.1. Power
         • 5.1.8.2. GetPowerMaxMinTmpSize
         • 5.1.8.3. GetPowerTmpBufferFactorSize
       • 5.1.9. Sinh
         • 5.1.9.1. Sinh
         • 5.1.9.2. GetSinhMaxMinTmpSize
         • 5.1.9.3. GetSinhTmpBufferFactorSize
       • 5.1.10. Cosh
         • 5.1.10.1. Cosh
         • 5.1.10.2. GetCoshMaxMinTmpSize
         • 5.1.10.3. GetCoshTmpBufferFactorSize
       • 5.1.11. Tan
         • 5.1.11.1. Tan
         • 5.1.11.2. GetTanMaxMinTmpSize
         • 5.1.11.3. GetTanTmpBufferFactorSize
       • 5.1.12. Trunc
         • 5.1.12.1. Trunc
         • 5.1.12.2. GetTruncMaxMinTmpSize
         • 5.1.12.3. GetTruncTmpBufferFactorSize
       • 5.1.13. Frac
         • 5.1.13.1. Frac
         • 5.1.13.2. GetFracMaxMinTmpSize
         • 5.1.13.3. GetFracTmpBufferFactorSize
       • 5.1.14. Erf
         • 5.1.14.1. Erf
         • 5.1.14.2. GetErfMaxMinTmpSize
         • 5.1.14.3. GetErfTmpBufferFactorSize
       • 5.1.15. Erfc
         • 5.1.15.1. Erfc
         • 5.1.15.2. GetErfcMaxMinTmpSize
         • 5.1.15.3. GetErfcTmpBufferFactorSize
       • 5.1.16. Sign
         • 5.1.16.1. Sign
         • 5.1.16.2. GetSignMaxMinTmpSize
         • 5.1.16.3. GetSignTmpBufferFactorSize
       • 5.1.17. Atanh
         • 5.1.17.1. Atanh
         • 5.1.17.2. GetAtanhMaxMinTmpSize
         • 5.1.17.3. GetAtanhTmpBufferFactorSize
       • 5.1.18. Asinh
         • 5.1.18.1. Asinh
         • 5.1.18.2. GetAsinhMaxMinTmpSize
         • 5.1.18.3. GetAsinhTmpBufferFactorSize
       • 5.1.19. Acosh
         • 5.1.19.1. Acosh
         • 5.1.19.2. GetAcoshMaxMinTmpSize
         • 5.1.19.3. GetAcoshTmpBufferFactorSize
       • 5.1.20. Floor
         • 5.1.20.1. Floor
         • 5.1.20.2. GetFloorMaxMinTmpSize
         • 5.1.20.3. GetFloorTmpBufferFactorSize
       • 5.1.21. Ceil
         • 5.1.21.1. Ceil
         • 5.1.21.2. GetCeilMaxMinTmpSize
         • 5.1.21.3. GetCeilTmpBufferFactorSize
       • 5.1.22. Clamp
         • 5.1.22.1. ClampMax
         • 5.1.22.2. ClampMin
         • 5.1.22.3. GetClampMaxMinTmpSize
         • 5.1.22.4. GetClampTmpBufferFactorSize
       • 5.1.23. Round
         • 5.1.23.1. Round
         • 5.1.23.2. GetRoundMaxMinTmpSize
         • 5.1.23.3. GetRoundTmpBufferFactorSize
       • 5.1.24. Axpy
         • 5.1.24.1. Axpy
         • 5.1.24.2. GetAxpyMaxMinTmpSize
         • 5.1.24.3. GetAxpyTmpBufferFactorSize
       • 5.1.25. Exp
         • 5.1.25.1. Exp
         • 5.1.25.2. GetExpMaxMinTmpSize
         • 5.1.25.3. GetExpTmpBufferFactorSize
       • 5.1.26. Lgamma
         • 5.1.26.1. Lgamma
         • 5.1.26.2. GetLgammaMaxMinTmpSize
         • 5.1.26.3. GetLgammaTmpBufferFactorSize
       • 5.1.27. Digamma
         • 5.1.27.1. Digamma
         • 5.1.27.2. GetDigammaMaxMinTmpSize
         • 5.1.27.3. GetDigammaTmpBufferFactorSize
       • 5.1.28. Xor
         • 5.1.28.1. Xor
         • 5.1.28.2. GetXorMaxMinTmpSize
         • 5.1.28.3. GetXorTmpBufferFactorSize
       • 5.1.29. CumSum
         • 5.1.29.1. CumSum
         • 5.1.29.2. GetCumSumMaxMinTmpSize
       • 5.1.30. Fmod
         • 5.1.30.1. Fmod
         • 5.1.30.2. GetFmodMaxMinTmpSize
         • 5.1.30.3. GetFmodTmpBufferFactorSize
       • 5.1.31. 更多样例
     • 5.2. Matmul
       • 5.2.1. Matmul
         • 5.2.1.1. 使用说明
         • 5.2.1.2. Matmul模板参数
         • 5.2.1.3. MatmulConfig
         • 5.2.1.4. MatmulCallBackFunc
         • 5.2.1.5. MatmulPolicy
         • 5.2.1.6. Matmul特性场景
         • 5.2.1.7. GetNormalConfig
         • 5.2.1.8. GetMDLConfig
         • 5.2.1.9. GetSpecialMDLConfig
         • 5.2.1.10. GetIBShareNormConfig
         • 5.2.1.11. GetBasicConfig
         • 5.2.1.12. GetSpecialBasicConfig
         • 5.2.1.13. GetMMConfig
         • 5.2.1.14. REGISTMATMULOBJ
         • 5.2.1.15. Init
         • 5.2.1.16. SetTensorA
         • 5.2.1.17. SetTensorB
         • 5.2.1.18. SetBias
         • 5.2.1.19. DisableBias
         • 5.2.1.20. GetBatchTensorC
         • 5.2.1.21. Iterate
         • 5.2.1.22. GetTensorC
         • 5.2.1.23. IterateAll
         • 5.2.1.24. WaitIterateAll
         • 5.2.1.25. IterateBatch
         • 5.2.1.26. WaitIterateBatch
         • 5.2.1.27. IterateNBatch
         • 5.2.1.28. End
         • 5.2.1.29. SetHF32
         • 5.2.1.30. SetTail
         • 5.2.1.31. SetBatchNum
         • 5.2.1.32. SetQuantScalar
         • 5.2.1.33. SetQuantVector
         • 5.2.1.34. SetOrgShape
         • 5.2.1.35. SetSingleShape
         • 5.2.1.36. SetLocalWorkspace
         • 5.2.1.37. SetWorkspace
         • 5.2.1.38. SetAntiQuantScalar
         • 5.2.1.39. SetAntiQuantVector
         • 5.2.1.40. WaitGetTensorC
         • 5.2.1.41. GetOffsetC
         • 5.2.1.42. AsyncGetTensorC
         • 5.2.1.43. SetUserDefInfo
         • 5.2.1.44. SetSelfDefineData
         • 5.2.1.45. GetMatmulApiTiling
         • 5.2.1.46. ClearBias
         • 5.2.1.47. GetBatchC
         • 5.2.1.48. SetSparseIndex
       • 5.2.2. Matmul Tiling
         • 5.2.2.1. Matmul Tiling类
           • 5.2.2.1.1. 使用说明
           • 5.2.2.1.2. 构造函数
           • 5.2.2.1.3. TCubeTiling结构体
           • 5.2.2.1.4. SetAType
           • 5.2.2.1.5. SetBType
           • 5.2.2.1.6. SetCType
           • 5.2.2.1.7. SetBiasType
           • 5.2.2.1.8. SetSingleShape
           • 5.2.2.1.9. SetShape
           • 5.2.2.1.10. SetOrgShape
           • 5.2.2.1.11. SetFixSplit
           • 5.2.2.1.12. EnableMultiCoreSplitK
           • 5.2.2.1.13. EnableBias
           • 5.2.2.1.14. SetBufferSpace
           • 5.2.2.1.15. SetTraverse
           • 5.2.2.1.16. SetMadType
           • 5.2.2.1.17. SetSplitRange
           • 5.2.2.1.18. SetDoubleBuffer
           • 5.2.2.1.19. GetBaseM
           • 5.2.2.1.20. GetBaseN
           • 5.2.2.1.21. GetBaseK
           • 5.2.2.1.22. GetTiling
           • 5.2.2.1.23. SetDim
           • 5.2.2.1.24. SetSingleRange
           • 5.2.2.1.25. GetSingleShape
           • 5.2.2.1.26. GetCoreNum
           • 5.2.2.1.27. SetAlignSplit
           • 5.2.2.1.28. SetDequantType
           • 5.2.2.1.29. SetALayout
           • 5.2.2.1.30. SetBLayout
           • 5.2.2.1.31. SetCLayout
           • 5.2.2.1.32. SetBatchNum
           • 5.2.2.1.33. SetBatchInfoForNormal
           • 5.2.2.1.34. SetMatmulConfigParams
           • 5.2.2.1.35. SetBias
           • 5.2.2.1.36. SetSplitK
           • 5.2.2.1.37. SetSparse
         • 5.2.2.2. 获取Matmul计算所需空间
           • 5.2.2.2.1. MultiCoreMatmulGetTmpBufSize
           • 5.2.2.2.2. MultiCoreMatmulGetTmpBufSizeV2
           • 5.2.2.2.3. BatchMatmulGetTmpBufSize
           • 5.2.2.2.4. BatchMatmulGetTmpBufSizeV2
           • 5.2.2.2.5. MatmulGetTmpBufSize
           • 5.2.2.2.6. MatmulGetTmpBufSizeV2
     • 5.3. 激活函数
       • 5.3.1. SoftMax
         • 5.3.1.1. SoftMax
         • 5.3.1.2. SimpleSoftMax
         • 5.3.1.3. SoftmaxFlash
         • 5.3.1.4. SoftmaxGrad
         • 5.3.1.5. SoftmaxFlashV2
         • 5.3.1.6. SoftmaxFlashV3
         • 5.3.1.7. SoftmaxGradFront
         • 5.3.1.8. AdjustSoftMaxRes
         • 5.3.1.9. SoftMax Tiling使用说明
         • 5.3.1.10. SoftMax/SimpleSoftMax Tiling
         • 5.3.1.11. SoftmaxFlash Tiling接口
         • 5.3.1.12. SoftmaxGrad Tiling接口
         • 5.3.1.13. SoftmaxFlashV2 Tiling接口
         • 5.3.1.14. SoftmaxFlashV3 Tiling接口
         • 5.3.1.15. IsBasicBlockInSoftMax
       • 5.3.2. LogSoftMax
         • 5.3.2.1. LogSoftMax
         • 5.3.2.2. LogSoftMax Tiling
       • 5.3.3. Gelu
         • 5.3.3.1. Gelu
         • 5.3.3.2. FasterGelu
         • 5.3.3.3. FasterGeluV2
         • 5.3.3.4. GetGeluMaxMinTmpSize
       • 5.3.4. SwiGLU
         • 5.3.4.1. SwiGLU
         • 5.3.4.2. GetSwiGLUMaxMinTmpSize
         • 5.3.4.3. GetSwiGLUTmpBufferFactorSize
       • 5.3.5. Silu
         • 5.3.5.1. Silu
         • 5.3.5.2. GetSiluTmpSize
       • 5.3.6. Swish
         • 5.3.6.1. Swish
         • 5.3.6.2. GetSwishTmpSize
       • 5.3.7. GeGLU
         • 5.3.7.1. GeGLU
         • 5.3.7.2. GetGeGLUMaxMinTmpSize
         • 5.3.7.3. GetGeGLUTmpBufferFactorSize
       • 5.3.8. ReGlu
         • 5.3.8.1. ReGlu
         • 5.3.8.2. GetReGluMaxMinTmpSize
       • 5.3.9. Sigmoid
         • 5.3.9.1. Sigmoid
         • 5.3.9.2. GetSigmoidMaxMinTmpSize
     • 5.4. 数据归一化
       • 5.4.1. LayerNorm
         • 5.4.1.1. LayerNorm
         • 5.4.1.2. LayerNormGrad
         • 5.4.1.3. LayerNormGradBeta
       • 5.4.2. LayerNorm Tiling
         • 5.4.2.1. LayerNorm Tiling
         • 5.4.2.2. LayerNormGrad Tiling
         • 5.4.2.3. LayerNormGradBeta Tiling
       • 5.4.3. RmsNorm
       • 5.4.4. RmsNorm Tiling
       • 5.4.5. BatchNorm
       • 5.4.6. BatchNorm Tiling
       • 5.4.7. DeepNorm
       • 5.4.8. DeepNorm Tiling
       • 5.4.9. Normalize
       • 5.4.10. Normalize Tiling
       • 5.4.11. WelfordUpdate
       • 5.4.12. WelfordUpdate Tiling
       • 5.4.13. WelfordFinalize
       • 5.4.14. WelfordFinalize Tiling
       • 5.4.15. GroupNorm
       • 5.4.16. GroupNorm Tiling
     • 5.5. 量化反量化
       • 5.5.1. AscendQuant
       • 5.5.2. GetAscendQuantMaxMinTmpSize
       • 5.5.3. AscendDequant
       • 5.5.4. GetAscendDequantMaxMinTmpSize
       • 5.5.5. AscendAntiQuant
       • 5.5.6. GetAscendAntiQuantMaxMinTmpSize
     • 5.6. 归约操作
       • 5.6.1. Sum
         • 5.6.1.1. Sum
         • 5.6.1.2. GetSumMaxMinTmpSize
       • 5.6.2. Mean
         • 5.6.2.1. Mean
         • 5.6.2.2. GetMeanMaxMinTmpSize
         • 5.6.2.3. GetMeanTmpBufferFactorSize
       • 5.6.3. ReduceXorSum
         • 5.6.3.1. ReduceXorSum
         • 5.6.3.2. GetReduceXorSumMaxMinTmpSize
       • 5.6.4. ReduceSum
         • 5.6.4.1. ReduceSum
         • 5.6.4.2. GetReduceSumMaxMinTmpSize
       • 5.6.5. ReduceMean
         • 5.6.5.1. ReduceMean
         • 5.6.5.2. GetReduceMeanMaxMinTmpSize
       • 5.6.6. ReduceMax
         • 5.6.6.1. ReduceMax
         • 5.6.6.2. GetReduceMaxMaxMinTmpSize
       • 5.6.7. ReduceMin
         • 5.6.7.1. ReduceMin
         • 5.6.7.2. GetReduceMinMaxMinTmpSize
       • 5.6.8. ReduceAny
         • 5.6.8.1. ReduceAny
         • 5.6.8.2. GetReduceAnyMaxMinTmpSize
       • 5.6.9. ReduceAll
         • 5.6.9.1. ReduceAll
         • 5.6.9.2. GetReduceAllMaxMinTmpSize
       • 5.6.10. ReduceProd
         • 5.6.10.1. ReduceProd
         • 5.6.10.2. GetReduceProdMaxMinTmpSize
     • 5.7. 排序
       • 5.7.1. TopK
       • 5.7.2. TopK Tiling
       • 5.7.3. Sort
         • 5.7.3.1. Concat
         • 5.7.3.2. GetConcatTmpSize
         • 5.7.3.3. Extract
         • 5.7.3.4. Sort
         • 5.7.3.5. GetSortTmpSize
         • 5.7.3.6. GetSortOffset
         • 5.7.3.7. GetSortLen
         • 5.7.3.8. MrgSort
     • 5.8. 数据填充
       • 5.8.1. Pad
       • 5.8.2. Pad Tiling
       • 5.8.3. UnPad
       • 5.8.4. UnPad Tiling
       • 5.8.5. Broadcast
       • 5.8.6. GetBroadCastMaxMinTmpSize
     • 5.9. 索引操作
       • 5.9.1. ArithProgression
       • 5.9.2. GetArithProgressionMaxMinTmpSize
     • 5.10. 比较选择
       • 5.10.1. SelectWithBytesMask
       • 5.10.2. GetSelectWithBytesMaskMaxMinTmpSize
     • 5.11. 数据过滤
       • 5.11.1. DropOut
       • 5.11.2. GetDropOutMaxMinTmpSize
     • 5.12. 变形
       • 5.12.1. ConfusionTranspose
       • 5.12.2. ConfusionTranspose Tiling
       • 5.12.3. TransData
       • 5.12.4. GetTransDataMaxMinTmpSize
     • 5.13. Hccl
       • 5.13.1. Hccl
         • 5.13.1.1. 使用说明
         • 5.13.1.2. Hccl模板参数
         • 5.13.1.3. Init
         • 5.13.1.4. SetCcTiling
         • 5.13.1.5. AllReduce
         • 5.13.1.6. AllGather
         • 5.13.1.7. ReduceScatter
         • 5.13.1.8. AlltoAll
         • 5.13.1.9. AlltoAllV
         • 5.13.1.10. BatchWrite
         • 5.13.1.11. Commit
         • 5.13.1.12. Wait
         • 5.13.1.13. Iterate
         • 5.13.1.14. Query
         • 5.13.1.15. InterHcclGroupSync
         • 5.13.1.16. Finalize
         • 5.13.1.17. GetWindowsInAddr
         • 5.13.1.18. GetWindowsOutAddr
         • 5.13.1.19. GetRankId
         • 5.13.1.20. GetRankDim
       • 5.13.2. Hccl Tiling
         • 5.13.2.1. 使用说明
         • 5.13.2.2. TilingData结构体
         • 5.13.2.3. 构造函数
         • 5.13.2.4. GetTiling
         • 5.13.2.5. SetOpType
         • 5.13.2.6. SetGroupName
         • 5.13.2.7. SetAlgConfig
         • 5.13.2.8. SetReduceType
         • 5.13.2.9. SetStepSize
         • 5.13.2.10. SetSkipLocalRankCopy
         • 5.13.2.11. SetSkipBufferWindowCopy
         • 5.13.2.12. SetDebugMode
         • 5.13.2.13. v1版本TilingData（废弃）
         • 5.13.2.14. v2版本TilingData（废弃）
       • 5.13.3. Hccl Context
         • 5.13.3.1. 简介
         • 5.13.3.2. SetHcclContext
         • 5.13.3.3. GetHcclContext
     • 5.14. 工具类
       • 5.14.1. InitGlobalMemory
     • 5.15. 卷积正向
       • 5.15.1. Conv3D
         • 5.15.1.1. Conv3D
           • 5.15.1.1.1. 使用说明
           • 5.15.1.1.2. Conv3D模板参数
           • 5.15.1.1.3. Init
           • 5.15.1.1.4. SetInput
           • 5.15.1.1.5. SetWeight
           • 5.15.1.1.6. SetBias
           • 5.15.1.1.7. SetSingleOutputShape
           • 5.15.1.1.8. SetInputStartPosition
           • 5.15.1.1.9. IterateAll
           • 5.15.1.1.10. End
         • 5.15.1.2. Conv3D Tiling
           • 5.15.1.2.1. 使用说明
           • 5.15.1.2.2. 构造函数
           • 5.15.1.2.3. TConv3DApiTiling结构体
           • 5.15.1.2.4. GetTiling
           • 5.15.1.2.5. SetOrgWeightShape
           • 5.15.1.2.6. SetOrgInputShape
           • 5.15.1.2.7. SetSingleWeightShape
           • 5.15.1.2.8. SetSingleOutputShape
           • 5.15.1.2.9. SetInputType
           • 5.15.1.2.10. SetWeightType
           • 5.15.1.2.11. SetBiasType
           • 5.15.1.2.12. SetOutputType
           • 5.15.1.2.13. SetPadding
           • 5.15.1.2.14. SetDilation
           • 5.15.1.2.15. SetStride
           • 5.15.1.2.16. SetGroups
     • 5.16. 卷积反向
       • 5.16.1. Conv3DBackpropInput
         • 5.16.1.1. Conv3DBackpropInput
           • 5.16.1.1.1. 使用说明
           • 5.16.1.1.2. Init
           • 5.16.1.1.3. SetGradOutput
           • 5.16.1.1.4. SetWeight
           • 5.16.1.1.5. SetSingleShape
           • 5.16.1.1.6. SetStartPosition
           • 5.16.1.1.7. Iterate
           • 5.16.1.1.8. GetTensorC
           • 5.16.1.1.9. End
         • 5.16.1.2. Conv3DBackpropInput Tiling
           • 5.16.1.2.1. 使用说明
           • 5.16.1.2.2. 构造函数
           • 5.16.1.2.3. TConv3DBackpropInputTiling结构体
           • 5.16.1.2.4. GetTiling
           • 5.16.1.2.5. SetWeightShape
           • 5.16.1.2.6. SetInputShape
           • 5.16.1.2.7. SetGradOutputShape
           • 5.16.1.2.8. SetWeightType
           • 5.16.1.2.9. SetInputType
           • 5.16.1.2.10. SetGradOutputType
           • 5.16.1.2.11. SetPadding
           • 5.16.1.2.12. SetStride
           • 5.16.1.2.13. SetDilation
           • 5.16.1.2.14. SetOutputPadding
       • 5.16.2. Conv3DBackpropFilter
         • 5.16.2.1. Conv3DBackpropFilter
           • 5.16.2.1.1. 使用说明
           • 5.16.2.1.2. Init
           • 5.16.2.1.3. SetInput
           • 5.16.2.1.4. SetGradOutput
           • 5.16.2.1.5. SetSingleShape
           • 5.16.2.1.6. SetStartPosition
           • 5.16.2.1.7. Iterate
           • 5.16.2.1.8. GetTensorC
           • 5.16.2.1.9. End
         • 5.16.2.2. Conv3DBackpropFilter Tiling
           • 5.16.2.2.1. 使用说明
           • 5.16.2.2.2. 构造函数
           • 5.16.2.2.3. TConv3DBpFilterTiling结构体
           • 5.16.2.2.4. GetTiling
           • 5.16.2.2.5. SetWeightShape
           • 5.16.2.2.6. SetInputShape
           • 5.16.2.2.7. SetGradOutputShape
           • 5.16.2.2.8. SetWeightType
           • 5.16.2.2.9. SetInputType
           • 5.16.2.2.10. SetGradOutputType
           • 5.16.2.2.11. SetPadding
           • 5.16.2.2.12. SetStride
           • 5.16.2.2.13. SetDilation
     • 5.17. 模板库函数
       • 5.17.1. max
       • 5.17.2. min
       • 5.17.3. index_sequence
       • 5.17.4. 容器函数
         • 5.17.4.1. tuple
         • 5.17.4.2. get
         • 5.17.4.3. make_tuple
       • 5.17.5. typetraits
         • 5.17.5.1. is_convertible
         • 5.17.5.2. isbaseof
         • 5.17.5.3. is_same
         • 5.17.5.4. enable_if
         • 5.17.5.5. [conditional](https://www.hiascend.com/doc
         center/source/zh/CANNCommunityEdition/82RC1/API/ascendcopapi/atlasascendcapi07_10118.html)
• 6. Host API
     • 6.1. 原型注册与管理
       • 6.1.1. 原型注册接口（OP_ADD）
       • 6.1.2. OpDef
         • 6.1.2.1. Input
         • 6.1.2.2. Output
         • 6.1.2.3. Attr
         • 6.1.2.4. SetInferShape
         • 6.1.2.5. SetInferShapeRange
         • 6.1.2.6. SetInferDataType
         • 6.1.2.7. AICore
         • 6.1.2.8. MC2
         • 6.1.2.9. EnableFallBack
         • 6.1.2.10. Comment
         • 6.1.2.11. FormatMatchMode
       • 6.1.3. OpParamDef
         • 6.1.3.1. ParamType
         • 6.1.3.2. DataType
         • 6.1.3.3. DataTypeForBinQuery
         • 6.1.3.4. DataTypeList
         • 6.1.3.5. Format
         • 6.1.3.6. FormatForBinQuery
         • 6.1.3.7. FormatList
         • 6.1.3.8. UnknownShapeFormat（废弃）
         • 6.1.3.9. ValueDepend
         • 6.1.3.10. IgnoreContiguous
         • 6.1.3.11. AutoContiguous
         • 6.1.3.12. Scalar
         • 6.1.3.13. ScalarList
         • 6.1.3.14. To
         • 6.1.3.15. Version
         • 6.1.3.16. InitValue
         • 6.1.3.17. OutputShapeDependOnCompute
         • 6.1.3.18. Follow
         • 6.1.3.19. Comment
       • 6.1.4. OpAttrDef
         • 6.1.4.1. OpAttrDef
         • 6.1.4.2. Comment
       • 6.1.5. OpAICoreDef
         • 6.1.5.1. SetTiling
         • 6.1.5.2. SetCheckSupport
         • 6.1.5.3. SetOpSelectFormat
         • 6.1.5.4. AddConfig
       • 6.1.6. OpAICoreConfig
         • 6.1.6.1. Input
         • 6.1.6.2. Output
         • 6.1.6.3. DynamicCompileStaticFlag
         • 6.1.6.4. DynamicFormatFlag
         • 6.1.6.5. DynamicRankSupportFlag
         • 6.1.6.6. DynamicShapeSupportFlag
         • 6.1.6.7. NeedCheckSupportFlag
         • 6.1.6.8. PrecisionReduceFlag
         • 6.1.6.9. ExtendCfgInfo
       • 6.1.7. OpMC2Def
         • 6.1.7.1. 简介
         • 6.1.7.2. 构造函数
         • 6.1.7.3. HcclGroup
         • 6.1.7.4. operator=
     • 6.2. Tiling数据结构注册
       • 6.2.1. TilingData结构定义
       • 6.2.2. TilingData结构注册
       • 6.2.3. ContextBuilder
         • 6.2.3.1. 简介
         • 6.2.3.2. 构造函数
         • 6.2.3.3. KernelRunContextHolder结构定义
         • 6.2.3.4. Inputs
         • 6.2.3.5. Outputs
         • 6.2.3.6. BuildKernelRunContext
         • 6.2.3.7. NodeIoNum
         • 6.2.3.8. SetOpNameType
         • 6.2.3.9. IrInstanceNum
         • 6.2.3.10. AddInputTd
         • 6.2.3.11. AddOutputTd
         • 6.2.3.12. AddAttr
         • 6.2.3.13. CompileInfo
         • 6.2.3.14. PlatformInfo
         • 6.2.3.15. AddPlatformInfo
         • 6.2.3.16. TilingData
         • 6.2.3.17. Workspace
         • 6.2.3.18. BuildTilingContext
       • 6.2.4. OpTilingRegistry
         • 6.2.4.1. 简介
         • 6.2.4.2. 构造和析构函数
         • 6.2.4.3. GetTilingFunc
         • 6.2.4.4. LoadTilingLibrary
       • 6.2.5. 模板参数定义
       • 6.2.6. GETTPLTILING_KEY
       • 6.2.7. ASCENDCTPLSEL_PARAM
       • 6.2.8. DEVICEIMPLOP_OPTILING
     • 6.3. 平台信息获取
       • 6.3.1. PlatformAscendC
         • 6.3.1.1. 简介
         • 6.3.1.2. 构造及析构函数
         • 6.3.1.3. GetCoreNum
         • 6.3.1.4. GetSocVersion
         • 6.3.1.5. GetCoreNumAic
         • 6.3.1.6. GetCoreNumAiv
         • 6.3.1.7. GetCoreNumVector
         • 6.3.1.8. CalcTschBlockDim
         • 6.3.1.9. GetCoreMemSize
         • 6.3.1.10. GetCoreMemBw
         • 6.3.1.11. GetLibApiWorkSpaceSize
         • 6.3.1.12. GetResGroupBarrierWorkSpaceSize
         • 6.3.1.13. GetResCubeGroupWorkSpaceSize
         • 6.3.1.14. ReserveLocalMemory
       • 6.3.2. PlatformAscendCManager
     • 6.4. 单算子API执行相关接口
       • 6.4.1. 单算子API执行相关接口列表
       • 6.4.2. 框架能力接口
         • 6.4.2.1. 常用宏和类
           • 6.4.2.1.1. ADDTOLAUNCHERLISTAICORE
           • 6.4.2.1.2. ADDTOLAUNCHERLISTAICPU
           • 6.4.2.1.3. CREATE_EXECUTOR
           • 6.4.2.1.4. DFX_IN
           • 6.4.2.1.5. DFX_OUT
           • 6.4.2.1.6. INFER_SHAPE
           • 6.4.2.1.7. L0_DFX
           • 6.4.2.1.8. L2DFXPHASE_1
           • 6.4.2.1.9. L2DFXPHASE_2
           • 6.4.2.1.10. OP_ATTR
           • 6.4.2.1.11. OPATTRNAMES
           • 6.4.2.1.12. OPEMPTYARG
           • 6.4.2.1.13. OP_INPUT
           • 6.4.2.1.14. OP_MODE
           • 6.4.2.1.15. OP_OUTPUT
           • 6.4.2.1.16. OP_OUTSHAPE
           • 6.4.2.1.17. OP_OPTION
           • 6.4.2.1.18. OPTYPEREGISTER
           • 6.4.2.1.19. OP_WORKSPACE
           • 6.4.2.1.20. OpImplMode
           • 6.4.2.1.21. OpExecMode
         • 6.4.2.2. bfloat16
         • 6.4.2.3. commontypes
           • 6.4.2.3.1. GetStorageShape
           • 6.4.2.3.2. SetStorageShape
           • 6.4.2.3.3. GetOriginalShape
           • 6.4.2.3.4. SetOriginalShape
           • 6.4.2.3.5. GetViewShape
           • 6.4.2.3.6. SetViewShape
           • 6.4.2.3.7. GetStorageFormat
           • 6.4.2.3.8. SetStorageFormat
           • 6.4.2.3.9. GetOriginalFormat
           • 6.4.2.3.10. SetOriginalFormat
           • 6.4.2.3.11. GetViewFormat
           • 6.4.2.3.12. SetViewFormat
           • 6.4.2.3.13. GetViewStrides
           • 6.4.2.3.14. SetViewStrides
           • 6.4.2.3.15. Numel
           • 6.4.2.3.16. GetDataType
           • 6.4.2.3.17. IsEmpty
           • 6.4.2.3.18. SetData
           • 6.4.2.3.19. SetDataType
           • 6.4.2.3.20. SetBoolData
           • 6.4.2.3.21. SetIntData
           • 6.4.2.3.22. SetFloatData
           • 6.4.2.3.23. SetFp16Data
           • 6.4.2.3.24. SetBf16Data
           • 6.4.2.3.25. CheckOverflows
           • 6.4.2.3.26. [预留接口](https://www.hiascend.com/doc
           center/source/zh/CANNCommunityEdition/82RC1/API/ascendcopapi/atlasascendcapi07_1093.html)
         • 6.4.2.4. datatypeutils
           • 6.4.2.4.1. IsBasicType
           • 6.4.2.4.2. IsNumberType
           • 6.4.2.4.3. IsRealNumberType
           • 6.4.2.4.4. TypeSize
           • 6.4.2.4.5. ToOpDataType
           • 6.4.2.4.6. ToAclDataType
           • 6.4.2.4.7. 预留接口
         • 6.4.2.5. formatutils
           • 6.4.2.5.1. IsPrivateFormat
           • 6.4.2.5.2. GetPrimaryFormat
           • 6.4.2.5.3. ToOpFormat
           • 6.4.2.5.4. ToAclFormat
           • 6.4.2.5.5. [预留接口](https://www.hiascend.com/doc
           center/source/zh/CANNCommunityEdition/82RC1/API/ascendcopapi/atlasascendcapi07_1107.html)
         • 6.4.2.6. fp16_t
         • 6.4.2.7. frameworkop
           • 6.4.2.7.1. CopyToNpu
           • 6.4.2.7.2. CopyToNpuSync
           • 6.4.2.7.3. [CopyNpuToNpu](https://www.hiascend.com/doc
           center/source/zh/CANNCommunityEdition/82RC1/API/ascendcopapi/atlasascendcapi07_1112.html)
         • 6.4.2.8. object
         • 6.4.2.9. opargdef
         • 6.4.2.10. op_cache
         • 6.4.2.11. opcachecontainer
         • 6.4.2.12. op_config
         • 6.4.2.13. op_def
         • 6.4.2.14. op_dfx
         • 6.4.2.15. opexecutor
           • 6.4.2.15.1. AllocTensor
           • 6.4.2.15.2. AllocHostTensor
           • 6.4.2.15.3. AllocIntArray
           • 6.4.2.15.4. AllocFloatArray
           • 6.4.2.15.5. AllocBoolArray
           • 6.4.2.15.6. AllocTensorList
           • 6.4.2.15.7. AllocScalarList
           • 6.4.2.15.8. AllocScalar
           • 6.4.2.15.9. ConvertToTensor
           • 6.4.2.15.10. CommonOpExecutorRun
           • 6.4.2.15.11. ReleaseTo
           • 6.4.2.15.12. Getworkspacesize
           • 6.4.2.15.13. AbandonCache
           • 6.4.2.15.14. [预留接口](https://www.hiascend.com/doc
           center/source/zh/CANNCommunityEdition/82RC1/API/ascendcopapi/atlasascendcapi07_1133.html)
         • 6.4.2.16. op_log
         • 6.4.2.17. platform
           • 6.4.2.17.1. GetCurrentPlatformInfo
           • 6.4.2.17.2. GetSocVersion
           • 6.4.2.17.3. GetSocLongVersion
           • 6.4.2.17.4. GetBlockSize
           • 6.4.2.17.5. 预留接口
         • 6.4.2.18. pool_allocator
         • 6.4.2.19. shapeutils
           • 6.4.2.19.1. ToShape
           • 6.4.2.19.2. ToShapeVector
           • 6.4.2.19.3. ToContiguousStrides
           • 6.4.2.19.4. CheckBroadcastShape
           • 6.4.2.19.5. BroadcastInferShape
           • 6.4.2.19.6. [预留接口](https://www.hiascend.com/doc
           center/source/zh/CANNCommunityEdition/82RC1/API/ascendcopapi/atlasascendcapi07_1148.html)
         • 6.4.2.20. small_vector
         • 6.4.2.21. tensorviewutils
           • 6.4.2.21.1. IsContiguous
           • 6.4.2.21.2. 预留接口
         • 6.4.2.22. opcommon相关接口
           • 6.4.2.22.1. data
           typeutils
           • 6.4.2.22.1.1. IsComplexType
           • 6.4.2.22.1.2. IsFloatingType
           • 6.4.2.22.1.3. IsIntegralType
           • 6.4.2.22.1.4. IsIntegralType（含bool）
           • 6.4.2.22.1.5. CanCast
           • 6.4.2.22.1.6. [PromoteType](https://www.hiascend.com/doc
           center/source/zh/CANNCommunityEdition/82RC1/API/ascendcopapi/atlasascendcapi07_1160.html)
         • 6.4.2.23. aicpu相关接口
           • 6.4.2.23.1. AicpuTask类
           • 6.4.2.23.2. AicpuTaskSpace类
           • 6.4.2.23.3. aicpuargshandler
           • 6.4.2.23.4. aicpuextinfo_handle
           • 6.4.2.23.5. aicpu_uitls
           • 6.4.2.23.6. aicpu_task
       • 6.4.3. 基础核函数接口
         • 6.4.3.1. Cast
         • 6.4.3.2. Contiguous
         • 6.4.3.3. ViewCopy
         • 6.4.3.4. Pad
         • 6.4.3.5. Reshape
         • 6.4.3.6. Slice
         • 6.4.3.7. Transpose
         • 6.4.3.8. TransData
         • 6.4.3.9. TransDataSpecial
         • 6.4.3.10. ReFormat
         • 6.4.3.11. IsNullptr
     • 6.5. 内部关联接口
• 7. 算子调测API
     • 7.1. DumpTensor
     • 7.2. printf
     • 7.3. assert
     • 7.4. DumpAccChkPoint
     • 7.5. Trap
     • 7.6. PrintTimeStamp
     • 7.7. GmAlloc
     • 7.8. ICPURUNKF
     • 7.9. ICPUSETTILING_KEY
     • 7.10. GmFree
     • 7.11. SetKernelMode
     • 7.12. TRACE_START
     • 7.13. TRACE_STOP
     • 7.14. MetricsProfStart
     • 7.15. MetricsProfStop
• 8. 预留接口

Ascend C 最佳实践

• 1. 前言
• 2. 异构计算
• 3. 功能调试
     • 3.1. 运行正常
     • 3.2. 精度正常
     • 3.3. 算子调试
• 4. 性能分析
     • 4.1. 获取性能数据
     • 4.2. 分析性能数据
• 5. 性能优化
     • 5.1. 优化建议总览表
     • 5.2. 搬运优化
       • 5.2.1. 尽量一次搬运较大的数据块
       • 5.2.2. GM地址尽量512B对齐
       • 5.2.3. 高效的使用搬运API
     • 5.3. 内存优化
       • 5.3.1. 算子与高阶API共享临时Buffer
       • 5.3.2. 限制TilingData结构大小
       • 5.3.3. 通过缩减Tensor ShapeInfo维度，优化栈空间
       • 5.3.4. 通过Unified Buffer融合实现连续vector计算
       • 5.3.5. 通过BT Buffer实现高效的bias计算
       • 5.3.6. 通过FP Buffer存放量化参数实现高效随路量化
       • 5.3.7. 通过L0C Buffer数据暂存实现高效的矩阵乘结果累加
       • 5.3.8. 较小矩阵长驻L1 Buffer，仅分次搬运较大矩阵
       • 5.3.9. 优化bank分配以提升读写性能
     • 5.4. API使用优化
       • 5.4.1. 纯搬运类算子VECIN和VECOUT建议复用
       • 5.4.2. 避免TPipe在对象内创建和初始化
       • 5.4.3. Matmul使能AtomicAdd选项
       • 5.4.4. Vector算子灵活运用Counter模式
       • 5.4.5. 针对不同场景合理使用归约指令
     • 5.5. 流水优化
       • 5.5.1. 使能double buffer
       • 5.5.2. 使能Iterate或IterateAll异步接口避免AIC/AIV同步依赖
     • 5.6. Tiling优化
       • 5.6.1. L2 Cache切分
       • 5.6.2. 核间负载均衡
• 6. 优秀实践
     • 6.1. FlashAttention算子性能调优案例
     • 6.2. Matmul算子性能调优案例
     • 6.3. GroupedMatmul算子性能调优案例
     • 6.4. MC²算子性能调优案例
     • 6.5. Matmul高阶API使能IBShare性能提升案例
     • 6.6. Matmul常量化算子性能提升案例

算子开发工具

• 1. 简介
• 2. 环境准备
• 3. 算子设计（msKPP）
     • 3.1. 工具概述
     • 3.2. 使用前准备
     • 3.3. 性能建模
       • 3.3.1. 原理概述
       • 3.3.2. 算子特性建模
       • 3.3.3. 算子计算搬运规格分析
       • 3.3.4. 极限性能分析
       • 3.3.5. 算子tiling初步设计
       • 3.3.6. 对外接口使用说明
         • 3.3.6.1. 接口列表
         • 3.3.6.2. 基础功能接口
           • 3.3.6.2.1. Chip
           • 3.3.6.2.2. Core
           • 3.3.6.2.3. Tensor
           • 3.3.6.2.4. Tensor.load
         • 3.3.6.3. 同步类指令接口
           • 3.3.6.3.1. set_flag
           • 3.3.6.3.2. wait_flag
         • 3.3.6.4. 指令接口
           • 3.3.6.4.1. mmad
           • 3.3.6.4.2. vadd
           • 3.3.6.4.3. vbrcb
           • 3.3.6.4.4. vconv
           • 3.3.6.4.5. vconv_deq
           • 3.3.6.4.6. vconv_vdeq
           • 3.3.6.4.7. vector_dup
           • 3.3.6.4.8. vexp
           • 3.3.6.4.9. vln
           • 3.3.6.4.10. vmax
           • 3.3.6.4.11. vmul
           • 3.3.6.4.12. vmuls
           • 3.3.6.4.13. vsub
           • 3.3.6.4.14. vdiv
           • 3.3.6.4.15. vcadd
           • 3.3.6.4.16. vabs
           • 3.3.6.4.17. vaddrelu
           • 3.3.6.4.18. vaddreluconv
           • 3.3.6.4.19. vadds
           • 3.3.6.4.20. vand
           • 3.3.6.4.21. vaxpy
           • 3.3.6.4.22. vbitsort
           • 3.3.6.4.23. vcgadd
           • 3.3.6.4.24. vcgmax
           • 3.3.6.4.25. vcgmin
           • 3.3.6.4.26. vcmax
           • 3.3.6.4.27. vcmin
           • 3.3.6.4.28. vcmp_xxx
           • 3.3.6.4.29. vcmpv_xxx
           • 3.3.6.4.30. vcmpvs_xxx
           • 3.3.6.4.31. vcopy
           • 3.3.6.4.32. vcpadd
           • 3.3.6.4.33. vgather
           • 3.3.6.4.34. vgatherb
           • 3.3.6.4.35. vlrelu
           • 3.3.6.4.36. vmadd
           • 3.3.6.4.37. vmaddrelu
           • 3.3.6.4.38. vmaxs
           • 3.3.6.4.39. vmin
           • 3.3.6.4.40. vmins
           • 3.3.6.4.41. vmla
           • 3.3.6.4.42. vmrgsort
           • 3.3.6.4.43. vmulconv
           • 3.3.6.4.44. vnot
           • 3.3.6.4.45. vor
           • 3.3.6.4.46. vrec
           • 3.3.6.4.47. vreduce
           • 3.3.6.4.48. vreducev2
           • 3.3.6.4.49. vrelu
           • 3.3.6.4.50. vrsqrt
           • 3.3.6.4.51. vsel
           • 3.3.6.4.52. vshl
           • 3.3.6.4.53. vshr
           • 3.3.6.4.54. vsqrt
           • 3.3.6.4.55. vsubrelu
           • 3.3.6.4.56. vsubreluconv
           • 3.3.6.4.57. vtranspose
     • 3.4. 调用msOpGen算子工程
       • 3.4.1. 功能介绍
       • 3.4.2. 调用示例
       • 3.4.3. 接口列表
         • 3.4.3.1. mskpp.tiling_func
         • 3.4.3.2. mskpp.getkernelfrom_binary
     • 3.5. 自动调优
       • 3.5.1. 功能介绍
       • 3.5.2. 快速入门
       • 3.5.3. 自动调优示例
       • 3.5.4. 接口列表
         • 3.5.4.1. autotune
         • 3.5.4.2. code_gen
         • 3.5.4.3. compile
         • 3.5.4.4. autotune_v2
         • 3.5.4.5. compile_executable
       • 3.5.5. 附录
         • 3.5.5.1. basicmatmulautotune.py
         • 3.5.5.2. jit_build.sh
         • 3.5.5.3. basicmatmulexecutable_autotune.py
         • 3.5.5.4. jitbuildexecutable.sh
     • 3.6. FAQ
       • 3.6.1. 运行Kernel时提示权限错误
• 4. 算子工程创建（msOpGen）
     • 4.1. 工具概述
     • 4.2. 使用前准备
     • 4.3. 创建算子工程
     • 4.4. 算子开发
     • 4.5. 算子编译部署
     • 4.6. 查看算子仿真流水图
     • 4.7. 典型案例
       • 4.7.1. Ascend C自定义算子开发实践
• 5. 算子测试（msOpST）
     • 5.1. 工具概述
     • 5.2. 使用前准备
     • 5.3. 生成测试用例定义文件
     • 5.4. 生成/执行测试用例
     • 5.5. 生成单算子上板测试框架
     • 5.6. 典型案例
       • 5.6.1. 测试用例定义文件
• 6. 异常检测（msSanitizer）
     • 6.1. 工具概述
     • 6.2. 使用前准备
     • 6.3. 内存检测
     • 6.4. 竞争检测
     • 6.5. 未初始化检测
     • 6.6. 典型案例
       • 6.6.1. 检测内核调用符方式的Ascend C算子
       • 6.6.2. 检测API调用的单算子
       • 6.6.3. 检测PyTorch接口调用的算子
       • 6.6.4. 检测CANN软件栈的内存
     • 6.7. FAQ
       • 6.7.1. msSanitizer工具异常报告中未显示正确的文件名和行号
       • 6.7.2. msSanitizer工具使用"--cce-enable-sanitizer -g"编译算子时出现"InputSection too large"错误
       • 6.7.3. msSanitizer工具提示--cache-size异常
     • 6.8. 对外接口使用说明
       • 6.8.1. 接口列表
       • 6.8.2. sanitizer接口
         • 6.8.2.1. sanitizerRtMalloc
         • 6.8.2.2. sanitizerRtMallocCached
         • 6.8.2.3. sanitizerRtFree
         • 6.8.2.4. sanitizerRtMemset
         • 6.8.2.5. sanitizerRtMemsetAsync
         • 6.8.2.6. sanitizerRtMemcpy
         • 6.8.2.7. sanitizerRtMemcpyAsync
         • 6.8.2.8. sanitizerRtMemcpy2d
         • 6.8.2.9. sanitizerRtMemcpy2dAsync
         • 6.8.2.10. sanitizerReportMalloc
         • 6.8.2.11. sanitizerReportFree
       • 6.8.3. 扩展接口（mstx）
         • 6.8.3.1. mstx接口简介
         • 6.8.3.2. mstxDomainCreateA
         • 6.8.3.3. mstxMemHeapRegister
         • 6.8.3.4. mstxMemRegionsRegister
         • 6.8.3.5. mstxMemRegionsUnregister
         • 6.8.3.6. mstxMemHeapUnregister
• 7. 算子调试（msDebug）
     • 7.1. 工具概述
     • 7.2. 使用前准备
     • 7.3. 指定Device ID（通算融合算子场景）
     • 7.4. 断点设置
     • 7.5. 内存与变量打印
     • 7.6. 单步调试
     • 7.7. 中断运行
     • 7.8. 核切换
     • 7.9. 读取寄存器
     • 7.10. 调试信息展示
     • 7.11. 解析异常算子dump文件
     • 7.12. 典型案例
       • 7.12.1. 上板调试vector算子
       • 7.12.2. 调用Ascend CL单算子
       • 7.12.3. 调试PyTorch接口调用的算子
       • 7.12.4. 上板调试模板库的算子
     • 7.13. FAQ
       • 7.13.1. msDebug工具使用"-O0 -g"编译算子时，stack frame size超出限制
       • 7.13.2. msDebug工具打印Tensor变量功能不可用，提示“unavailable”或“memory read failed”
       • 7.13.3. msDebug工具在容器环境中调试运行失败，提示需安装HDK驱动包
       • 7.13.4. msDebug工具断点设置在核函数内，命中断点后执行continue命令，算子运行失败
       • 7.13.5. msDebug工具在docker中执行"run"命令运行程序后，提示“'A' packet returned an error: 8”
• 8. 算子调优（msProf）
     • 8.1. 工具概述
     • 8.2. 使用前准备
     • 8.3. 工具使用
     • 8.4. 计算内存热力图
     • 8.5. Roofline瓶颈分析图
     • 8.6. Cache热力图
     • 8.7. 通算流水图
     • 8.8. 指令流水图
     • 8.9. 算子代码热点图
     • 8.10. 内存通路吞吐率波形图
     • 8.11. 性能数据文件
       • 8.11.1. msprof op
         • 8.11.1.1. ArithmeticUtilization（cube及vector类型指令耗时和占比）
         • 8.11.1.2. L2Cache（L2 Cache命中率）
         • 8.11.1.3. Memory（内存读写带宽速率）
         • 8.11.1.4. MemoryL0（L0读写带宽速率）
         • 8.11.1.5. MemoryUB（UB读写带宽速率）
         • 8.11.1.6. OpBasicInfo（算子基础信息）
         • 8.11.1.7. PipeUtilization（计算单元和搬运单元耗时占比）
         • 8.11.1.8. ResourceConflictRatio（资源冲突占比）
       • 8.11.2. msprof op simulator
         • 8.11.2.1. 代码行耗时数据文件
         • 8.11.2.2. 代码指令信息文件
     • 8.12. Json配置文件说明
     • 8.13. 典型案例
       • 8.13.1. 采集Ascend C算子的性能数据
       • 8.13.2. 通过指令流水图优化算子
       • 8.13.3. 采集MC2算子的性能数据
     • 8.14. 扩展接口（mstx）
       • 8.14.1. mstx接口简介
       • 8.14.2. mstxRangeStartA
       • 8.14.3. mstxRangeEnd
• 9. 附录
     • 9.1. TBE&AI CPU算子开发场景
       • 9.1.1. 基于msOpGen工具创建算子工程
       • 9.1.2. 算子编译部署
         • 9.1.2.1. 简介
         • 9.1.2.2. 算子工程编译
         • 9.1.2.3. 算子交付件独立编译
         • 9.1.2.4. 算子包部署
       • 9.1.3. 基于msOpST工具进行算子ST测试
         • 9.1.3.1. 简介
         • 9.1.3.2. 生成测试用例定义文件
         • 9.1.3.3. 生成/执行测试用例
         • 9.1.3.4. 测试用例定义文件配置样例

该渲染由 Shiro API 生成，可能存在排版问题，最佳体验请前往：https://hyli.xyz/notes/8

看完了？说点什么呢

该渲染由 Shiro API 生成，可能存在排版问题，最佳体验请前往：https://hyli.xyz/posts/tutorial/cloudflare-images-workers-image-compression

export default {
  async fetch(request) {
    if (/image-resizing/.test(request.headers.get("via"))) {
      return fetch(request)
    }

    let options = { cf: { image: {} } };
    
    const accept = request.headers.get("Accept");
    if (/image\/webp/.test(accept)) {
      options.cf.image.format = "webp";
    } else if (/image\/avif/.test(accept)) {
      options.cf.image.format = "avif";
    }

    const imageRequest = new Request(request.url, {
      headers: request.headers,
    });

    let transformRequests = fetch(imageRequest, options);

    if ((await transformRequests).status == 200) {
      return transformRequests;
    } else {
      return fetch(request);
    }
  }
};

看完了？说点什么呢

该渲染由 Shiro API 生成，可能存在排版问题，最佳体验请前往：https://hyli.xyz/notes/7

看完了？说点什么呢

该渲染由 Shiro API 生成，可能存在排版问题，最佳体验请前往：https://hyli.xyz/notes/9

看完了？说点什么呢