Prism

Replication Package for ASE2026 This repository contains the source code and experimental setup for the paper: "PRISM: Practical and Rapid Floating-Point Error Estimation for Deep Learning Operators via Perturbation Injection." PRISM is designed for researchers to quantify numerical uncertainty and floating-point errors in Deep Learning operators by injecting directed ULP perturbations. 1. Triton Environment Setup (triton) This environment is used for running the Triton-based operator experiments and the PRISM Triton backend. # Create and activate conda environment conda create -n triton python=3.11 -y conda activate triton # Install core dependencies pip install torch==2.4.0 numpy numba tabulate # Build and install Triton from source cd triton pip install ninja cmake wheel pybind11 pip install -e python 2. PyTorch GPU Environment (torch-gpu) This environment is configured to build PyTorch from source with CUDA support enabled, specifically for GPU-based perturbation injection experiments. conda create -n torch-gpu python=3.11 -y conda activate torch-gpu conda install -c conda-forge cmake=3.27 ninja -y pip install pyyaml numpy typing_extensions # Set compilation environment variables cd /PRISM/torch-gpu/pytorch export CUDA_HOME=/usr/local/cuda # Update this to your local path if necessary export CUDACXX=$CUDA_HOME/bin/nvcc export CC=/usr/bin/gcc export CXX=/usr/bin/g++ # Optimization: Disable unnecessary modules to speed up compilation export USE_MKL=0 export USE_MKLDNN=0 export USE_DISTRIBUTED=0 export USE_MPI=0 export BUILD_TEST=0 export USE_XNNPACK=0 export USE_FBGEMM=0 # Compile and install python setup.py install 3. PyTorch CPU Environment (torch-cpu) This environment uses OpenBLAS as the backend for CPU-side operator testing and numerical analysis. Build OpenBLAS # In the OpenBLAS source directory cd /PRISM/OpenBLAS make clean make TARGET=GENERIC USE_OPENMP=1 -j$(nproc) make install PREFIX=$HOME/software/OpenBLAS Build PyTorch conda activate torch-cpu cd /PRISM/torch-cpu/pytorch # Configure backend and paths export BLAS=OpenBLAS export OpenBLAS_HOME=$HOME/software/OpenBLAS export LD_PRELOAD=$OpenBLAS_HOME/lib/libopenblas.so # Disable CUDA and enable OpenBLAS export USE_CUDA=0 export USE_CUDNN=0 export USE_MKL=0 export USE_MKLDNN=0 # Build in development mode python setup.py clean python setup.py develop Run To run the code, you can use the following commands: Triton part: cd /Prism/triton_scripts python run_perturb.py # generate interval for each case python judge_perturb.py # judge the result Pytorch part: For cpu:. cd /Prism/torch_scripts/torch-examples/xx conda activate triton python3 xx.py export LD_PRELOAD=/OpenBLAS/lib/libopenblas.so conda activate torch-cpu USE_MY_CUTLASS=1 python3 xx_perturb.py (perturb version) For gpu: cd /Prism/torch_scripts/torch-examples/xx conda activate triton python3 xx.py conda activate torch-gpu USE_MY_CUTLASS=1 python3 xx_perturb.py (perturb version)