Installation Guide for AccelProf

This guide walks you through the installation of AccelProf, a customizable GPU profiling framework. It covers system requirements, build instructions, environment setup, and a basic test to verify successful installation.

Requirements

AccelProf requires both hardware and software components to be correctly configured before use.

Hardware Requirements

  • At least one GPU (NVIDIA or AMD)

  • x86-based systems are tested and supported (ARM-based systems may also work, though untested)

Software Requirements

  • CUDA Toolkit (Tested on version 12.1; older versions may work)

  • ROCm Toolkit (Tested on version 6.3; older versions may work)

  • PyTorch (Tested on version 2.1; older versions may work)

Build Instructions

You can either use the integrated build script or perform a manual step-by-step build to gain finer control and debug support.

Step 1: Download the Code

git clone --recursive https://github.com/AccelProf/AccelProf.git

cd AccelProf
export ACCEL_PROF_DIR=$(pwd)

Step 3: Step-by-Step Build (Advanced Users)

Use these steps if you want to inspect or debug the compilation process.

1. Build libbacktrace

cd $ACCEL_PROF_DIR/sanalyzer/cpp_trace
if [ ! -d "$ACCEL_PROF_DIR/build/backtrace" ]; then
    source ${ACCEL_PROF_DIR}/bin/utils/build_libbacktrace
fi

2. Build cpp_trace

cd $ACCEL_PROF_DIR/sanalyzer/cpp_trace
make -j install DEBUG=$DEBUG BACKTRACE_DIR=$ACCEL_PROF_DIR/build/backtrace                              INSTALL_DIR=$ACCEL_PROF_DIR/build/sanalyzer/cpp_trace

3. Build py_frame

cd $ACCEL_PROF_DIR/sanalyzer/py_frame
make -j install DEBUG=$DEBUG PYBIND11_DIR=$ACCEL_PROF_DIR/third_party/pybind11                              INSTALL_DIR=$ACCEL_PROF_DIR/build/sanalyzer/py_frame

4. Build sanalyzer

cd $ACCEL_PROF_DIR/sanalyzer
make -j install DEBUG=$DEBUG SANITIZER_TOOL_DIR=$ACCEL_PROF_DIR/nv-compute                 NV_NVBIT_DIR=$ACCEL_PROF_DIR/nv-nvbit                 CPP_TRACE_DIR=$ACCEL_PROF_DIR/build/sanalyzer/cpp_trace                 PY_FRAME_DIR=$ACCEL_PROF_DIR/build/sanalyzer/py_frame                 INSTALL_DIR=$ACCEL_PROF_DIR/build/sanalyzer

5. Build tensor_scope

cd $ACCEL_PROF_DIR/tensor_scope
make -j install DEBUG=$DEBUG INSTALL_DIR=$ACCEL_PROF_DIR/build/tensor_scope

6. Build nv-compute

cd $ACCEL_PROF_DIR/nv-compute
make -j DEBUG=$DEBUG SANALYZER_DIR=$ACCEL_PROF_DIR/build/sanalyzer                 TORCH_SCOPE_DIR=$ACCEL_PROF_DIR/build/tensor_scope                 PATCH_SRC_DIR=$ACCEL_PROF_DIR/nv-compute/gpu_src

cp $ACCEL_PROF_DIR/nv-compute/lib/*.so $ACCEL_PROF_DIR/lib/

7. Build nv-nvbit

cd $ACCEL_PROF_DIR/nv-nvbit
make -j DEBUG=$DEBUG SANALYZER_DIR=$ACCEL_PROF_DIR/build/sanalyzer                      TORCH_SCOPE_DIR=$ACCEL_PROF_DIR/build/tensor_scope

cp $ACCEL_PROF_DIR/nv-nvbit/lib/*.so $ACCEL_PROF_DIR/lib/

8. Build amd-rocm (Optional – requires ROCm dev environment)

cd $ACCEL_PROF_DIR/amd-rocm
make -j DEBUG=$DEBUG SANALYZER_DIR=$ACCEL_PROF_DIR/build/sanalyzer

cp $ACCEL_PROF_DIR/amd-rocm/build/lib/*.so $ACCEL_PROF_DIR/lib/

Configuration

To use AccelProf from the terminal or within scripts, set up the environment:

export ACCEL_PROF_HOME=$(pwd)/AccelProf
export PATH=${ACCEL_PROF_HOME}/bin:${PATH}

🔁 You can optionally add these lines to your ~/.bashrc or ~/.zshrc to persist them across sessions.

Quick Test

Once the build completes, you should see multiple *.so files in the lib/ directory. Run a basic profiling task to verify the installation works. Refer to the Usage section for a detailed walkthrough.