Installation

This page prepares the GTaP source tree and builds the GTaP-enabled Clang.

Tested environment

GTaP has been tested on an NVIDIA GH200 node of the Miyabi-G supercomputer.

Component	Tested configuration
GPU	NVIDIA GH200
Compute capability	9.0 (sm_90)
GTaP-enabled Clang	Based on LLVM 21.1.8
CUDA Toolkit	12.9
Linux kernel	5.14.0-427.13.1.el9_4.aarch64

Other NVIDIA GPU and CUDA configurations may work, but this is the currently verified environment.

1. Clone GTaP

Clone the repository and its compiler submodule:

git clone https://github.com/yukim0359/GTaP.git --recursive
cd GTaP

If the repository was cloned without --recursive, initialize the submodule separately:

git submodule update --init --recursive

2. Build the GTaP-enabled Clang

GTaP programs are compiled with the GTaP-enabled Clang fork in clang-gtap. The following commands assume that the current directory is the root of the cloned GTaP repository.

Install the build requirements

The compiler build requires:

• CMake 3.20 or later
• Ninja
• a host C/C++ compiler, such as Clang or GCC

The CUDA Toolkit is required later to compile and run GTaP programs, but not to build the GTaP-enabled Clang itself.

Check that the build tools are available:

cmake --version
ninja --version
clang --version

Select the LLVM host target

GTaP needs the LLVM backend for the host CPU and the NVPTX backend for CUDA device code. Check the host architecture:

uname -m

Set LLVM_HOST_TARGET from the result:

uname -m	LLVM target
aarch64	AArch64
x86_64	X86

For example, use the following setting on an NVIDIA GH200 system:

export LLVM_HOST_TARGET=AArch64

Configure Clang

Configure an optimized build in clang-gtap/build:

cmake -G Ninja \
  -S clang-gtap/llvm \
  -B clang-gtap/build \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_C_COMPILER=clang \
  -DCMAKE_CXX_COMPILER=clang++ \
  -DLLVM_ENABLE_PROJECTS="clang" \
  -DLLVM_TARGETS_TO_BUILD="${LLVM_HOST_TARGET};NVPTX"

Using GCC as the host compiler

CMAKE_C_COMPILER and CMAKE_CXX_COMPILER select the compiler used to build the GTaP-enabled Clang. They do not change the resulting compiler: the output is still Clang with GTaP support. Replace clang and clang++ above with gcc and g++ when Clang is not available on the host.

Build and verify Clang

Build the clang target:

ninja -C clang-gtap/build clang

The resulting compiler is placed at clang-gtap/build/bin/clang, which is also the default compiler path used by the example Makefiles. Verify the build with:

./clang-gtap/build/bin/clang --version

The command should report a Clang version based on LLVM 21.1.8. The installation is now ready; continue to Quickstart.