Skip to content
/ jitify Public

A single-header C++ library for simplifying the use of CUDA Runtime Compilation (NVRTC).

License

Notifications You must be signed in to change notification settings

NVIDIA/jitify

Repository files navigation

Jitify

A single-header C++ library for simplifying the use of CUDA Runtime Compilation (NVRTC).

Rationale

Integrating NVRTC into existing and/or templated CUDA code can be tricky. Jitify aims to simplify this process by hiding the complexities behind a simple, high-level interface.

Quick example

const char* program_source = "my_program\n"
    "template<int N, typename T>\n"
    "__global__\n"
    "void my_kernel(T* data) {\n"
    "    T data0 = data[0];\n"
    "    for( int i=0; i<N-1; ++i ) {\n"
    "        data[0] *= data0;\n"
    "    }\n"
    "}\n";
static jitify::JitCache kernel_cache;
jitify::Program program = kernel_cache.program(program_source);
// ...set up data etc.
dim3 grid(1);
dim3 block(1);
using jitify::reflection::type_of;
program.kernel("my_kernel")
       .instantiate(3, type_of(*data))
       .configure(grid, block)
       .launch(data);

Features

Jitify provides/takes care of the following things:

  • All NVRTC and CUDA Driver API calls
  • Simple kernel instantiation and launch syntax
  • Caching compiled kernels
  • Loading source code from strings, files, or embedded in an executable
  • Ignoring host code in runtime-compiled sources
  • Skipping unneeded headers
  • Support for JIT-safe standard library headers (e.g., float.h, stdint.h etc.)
  • Dealing with kernel name mangling
  • Reflecting kernel template parameters into strings
  • Compiling specifically for the current device's compute capability
  • Linking to pre-compiled PTX/CUBIN/FATBIN/object/library files
  • Support for CUDA versions 7.0, 7.5, 8.0, 9.x, 10.x, on both Linux and Windows
  • Convenient parallel_for function and lambda support
  • *New* jitify::experimental API provides serialization capabilities to enable user-managed hashing and caching

Things you can do with Jitify and NVRTC:

  • Rapidly port existing code to use CUDA Runtime Compilation
  • Dramatically reduce code volume and offline-compilation times
  • Increase kernel performance by baking in runtime constants and autotuning

How to build

Jitify is just a single header file:

#include <jitify.hpp>

Compile with: -pthread (not needed if JITIFY_THREAD_SAFE is defined to 0)

Link with: -lcuda -lcudart -lnvrtc

A small utility called stringify is included for converting text files into C string literals, which provides a convenient way to integrate JIT-compiled sources into a build.

Running tests

Tests can be run with the following command:

$ make test

This will automatically download and build the GoogleTest library, which requires CMake to be available on the system.

Documentation

Examples

See jitify_example.cpp for some examples of how to use the library. The Makefile also demonstrates how to use the provided stringify utility.

GTC 2017 Talk by Ben Barsdell and Kate Clark

API documentation

Doxygen documentation can be generated by running:

$ make doc

The HTML and LaTeX results are placed into the doc/ subdirectory.

License

BSD-3-Clause

Authors

Ben Barsdell (NVIDIA, bbarsdell at nvidia dot com)

Kate Clark (NVIDIA, mclark at nvidia dot com)

About

A single-header C++ library for simplifying the use of CUDA Runtime Compilation (NVRTC).

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published