TVM Building on Windows

Preparation

1. Visual Studio: Download Link, Community version is enough.
2. CMake: Download Link, choose latest release (binary distribution is convenient).
3. Anaconda: Download Link, used to manage python enveriment.
4. Git: Download Link, used to pull tvm repo.

Compile LLVM

Suppose you want to build TVM in directory E:\tvm\

Download and compile LLVM project: Getting Started with the LLVM System — LLVM 20.0.0git documentation, following the guide of Getting the Source Code and Building LLVM. (I can’t build tvm from pre-built binary LLVM.)

在 CMake 中，-S 参数指定了 CMakeLists.txt 文件所在的目录，-B 参数指定了构建文件的生成目录。

cmake -S <source_dir> -B <build_dir>

其中：

• <source_dir>：包含 CMakeLists.txt 文件的源代码目录。
• <build_dir>：用于存放构建文件的目录。

# Check out LLVM (including subprojects like Clang):
git clone --config core.autocrlf=false https://github.com/llvm/llvm-project.git
# change directory
cd llvm-project
# build
cmake -S llvm -B build -G Visual Studio
cmake --build build

If build process is successful, you will find build/release/bin/llvm-config.exe

Compile TVM

Following the official guidance: Install from Source — tvm 0.18.dev0 documentation (apache.org)

Get Source from Github

# clone 
git clone --recursive https://github.com/dmlc/tvm tvm
# check submodule
git submodule init 
git submodule update

Modify config.cmake file

Modify /cmake/config.cmake

Modify set(USE_LLVM OFF) to set(USE_LLVM "/path/to/llvm-config --link-static"), /path/to/llvm-config is the path of llvm-config.exe in the last step.

If want to use CUDA: set(USE_CUDA ON)

Build Anaconda Environment

Install all necessary build dependencies.

cd ./tvm
# Create a conda environment with the dependencies specified by the yaml
conda env create --file conda/build-environment.yaml
# Activate the created environment
conda activate tvm-build

Build Shared Library

TVM support build via MSVC using cmake.

cmake --build build --config Release -- /m 用于在 build 目录中构建项目，并使用 Release 配置，同时启用多线程构建。

• cmake --build build：告诉 CMake 在 build 目录中执行构建操作。
• --config Release：指定构建配置为 Release。这通常意味着启用优化选项，以生成更小的可执行文件并提高性能。
• -- /m：这是 Windows 平台上的一个参数，用于启用多线程构建。它告诉 CMake 使用多个处理器核心来加速构建过程。

mkdir build
cd build
cmake -A x64 -Thost=x64 ..
cd ..
cmake --build build --config Release -- /m

If build process is successful, you will find tvm.dll, tvm_runtime.dll .etc in tvm/build/Release.

Import tvm package into python

Open anaconda prompt in tvm/python.

conda activate tvm-build
python setup.py install

Verify whether the package is installed or not.

python
>>import tvm
>>print(tvm.__version__)