libdevice

Type: Technology Tags: NVIDIA, CUDA, libdevice, device math, compiler SDK, LLVM bitcode Related: NVIDIA-CUDA, CUDA-Math-API, libNVVM, NVVM-IR, NVCC, PTX-ISA Sources: https://docs.nvidia.com/cuda/libdevice-users-guide/index.html Last Updated: 2026-04-29

Summary

libdevice is NVIDIA’s device-side bitcode library used by CUDA compiler paths to implement common mathematical and utility functions for GPU kernels. It sits in the CUDA compiler SDK layer alongside libNVVM and NVVM-IR.

Detail

Purpose

CUDA device code often calls math functions and helper operations that are not directly written into every kernel source file. libdevice provides compiler-consumable implementations so front ends can lower those operations consistently into GPU code.

Key capabilities

• Supplies device-function implementations in LLVM bitcode form.
• Supports CUDA math and utility functions used by GPU kernels.
• Integrates with libNVVM and NVVM IR compilation flows.
• Helps custom GPU language front ends target NVIDIA GPUs without reimplementing core device math support.

NVIDIA context

For most CUDA developers, libdevice is invisible. For compiler engineers and advanced CUDA library authors, it is part of the supported path from high-level code to PTX and GPU execution.

Connections

• CUDA-Math-API - many device-side math functions relate to libdevice-provided implementations.
• libNVVM - compiler library that works with NVVM IR and device bitcode.
• NVVM-IR - compiler IR layer used with libdevice.
• PTX-ISA - final virtual ISA target after NVVM/libdevice lowering.
• NVCC - standard CUDA compiler flow uses these components internally.

Source Excerpts

• NVIDIA documents libdevice as a user guide under the CUDA compiler SDK documentation set.