NVIDIA GPU Code Generation Guide

Overview

GPU code comes in two forms:

• SASS: Machine code specific to GPU architecture.
• PTX: Intermediate code JIT-compiled to SASS at runtime.

These are packaged in containers (CUBIN/FATBIN) during compilation.

Build System Syntax

Usage	Contents	Syntax	Example
PyTorch	SASS only	Space separated with -real suffix	80-real 86-real 87-real
PyTorch	SASS+PTX	Space separated	80 86 87
PyTorch	PTX only	Space separated with +PTX suffix	80+PTX 86+PTX 87+PTX
CMake	SASS only	Semicolon with -real suffix	80-real;86-real;87-real
CMake	SASS+PTX	Semicolon separated	80;86;87
CMake	PTX only	Semicolon with -virtual suffix	80-virtual;86-virtual;87-virtual
NVCC	SASS only	sm_ prefix	sm_80 sm_86 sm_87
NVCC	SASS+PTX	gencode specification	arch=compute_80,code=sm_80
NVCC	PTX only	compute_ prefix	compute_80 compute_86 compute_87

Environment Variables

PyTorch style:

• SASS only: TORCH_CUDA_ARCH_LIST="80-real 86-real 87-real"
• SASS+PTX: TORCH_CUDA_ARCH_LIST="80 86 87"
• PTX only: TORCH_CUDA_ARCH_LIST="80+PTX 86+PTX 87+PTX"

CMake style:

• SASS only: CMAKE_CUDA_ARCHITECTURES="80-real;86-real;87-real"
• SASS+PTX: CMAKE_CUDA_ARCHITECTURES="80;86;87"
• PTX only: CMAKE_CUDA_ARCHITECTURES="80-virtual;86-virtual;87-virtual"

NVCC flags:

• SASS only: CUDA_NVCC_FLAGS="-arch=sm_80 -arch=sm_86 -arch=sm_87"
• SASS+PTX: CUDA_NVCC_FLAGS="-gencode arch=compute_80,code=sm_80 -gencodearch=compute_86,code=sm_86 -gencode arch=compute_87,code=sm_87"`
• PTX only: CUDA_NVCC_FLAGS="-arch=compute_80 -arch=compute_86 -arch=compute_87"