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_programn"
    "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
Support for CUDA versions 7.0, 7.5, 8.0, 9.x, 10.x
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: -ldl -lcuda -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.