# spmm_sparse_tree

**Repository Path**: wangyaoyuu/spmm_sparse_tree

## Basic Information

- **Project Name**: spmm_sparse_tree
- **Description**: 1. rl for sparse program (include gpu and cpu)

2. currently, just testing
- **Primary Language**: Unknown
- **License**: Apache-2.0
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2025-04-21
- **Last Updated**: 2026-06-08

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README



## Project overview

### 文件夹描述

- rl_src: for rl
- sparse_src: for gpu
- spmv-forge: for cpu
- v1: origin version of DeepSparse (also called RLSparse) -- very complicated


### 运行例子

#### sparse_gpu (generalsparse)
- ---------------学习运行------------------
- cd sparse_src
- make rl_select -j16
- cd ..
- python  -m rl_src.main  -t sparse_gpu -p train -m transformer -r mcts 
- ----------尝试运行完-------------会有一部分存储数据，记得删除
- cd sparse_src/data_source/
- rm -rf *
- ----------- mark already merge---------------

## 环境设定
### 环境 for gpu (A100/V100/...)
- conda create -n spenv python=3.9.16 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
- pip install torch -i https://pypi.tuna.tsinghua.edu.cn/simple

### 环境 for arm (huawei)
- conda create -n gensp python=3.9.16 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
- pip install scikit-learn==1.4.1.post1 h5py==3.8.0 einops==0.7.0 tqdm==4.66.2 transformers==4.20.1 -i https://pypi.tuna.tsinghua.edu.cn/simple
- pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cpu


## 轻量化接入其他系统

### Search space (搜索空间设定)

#### **command** (根据**selection.json**的可执行文件) 
假设程序运行的指令为
`OMP_NUM_THREADS=1 OMP_PROC_BIND=close OMP_PLACES=cores ./local/bin/kernel_run -f ./sparse_mm_file/ACTIVSg2000.mtx -format SCS -backend ImplFunc -isa Sve -variant SveV2C2 -logdir ./log`

- 环境设定`OMP_NUM_THREADS=1 OMP_PROC_BIND=close OMP_PLACES=cores`
- 程序运行`./local/bin/kernel_run`
- 输入的矩阵`-f ./sparse_mm_file/ACTIVSg2000.mtx`
- 各种搜索参数 (与**selection.json**文件对应)
    - `-format SCS`
    - `-backend ImplFunc` 
    - `-isa Sve` 
    - `-variant SveV2C2`
- 输出的性能数据所在的目录
    - `-logdir ./log`

#### 与**command**对应的**selection.json**构成
- 以json形式存在 (详细见cpu样例)

#### valid_path.json [TODO]
- 合法的路径搜索
- 详细见样例

#### unvalid_path.json [TODO]
- 剪枝掉的路径搜索
- 暂无样例

#### 输出的性能数据所在的目录（假定为log文件夹）
- log文件夹下有运行程序得到的性能记录
- 命名规则：log文件名要与**selection.json**和**command**中的设定保持一致，从而能够从文件名中反映出搜索路径，定位到性能数据
- 样例`ACTIVSg2000_SCS_ImplFunc_Sve_SveV2C2_1.log`，log内有表示性能的数字
- [TODO] need to change for generalsparse for rules