Introduction to Lift

Lift is a library of components useful for parallel programming. It exposes capabilities similar to NVIDIA’s Thrust, but using somewhat different semantics aimed at creating a library that is more robust, transparent and suitable for production code.

Design Goals

The goal of Lift is to provide implementations for common building blocks when writing compute code for GPU and many-core CPU machines:

• Device-specific memory containers
• Cross-device memory transfers
• Well-known parallel primitives
• Interoperability with existing libraries in the same space

The general design philosophy can be summarized:

• Expose slow operations explicitly
• Abstract common operations where this can be done with minimal performance impact
• Provide a code base that is easily understood
• Avoid hidden functionality
• Provide a stable, predictable code base

Quick Example

Sort a set of key-value pairs:

#include <lift/memory.h>
#include <lift/backends.h>
#include <lift/parallel.h>

using namespace lift;

template <target_system system>
void test(void)
{
    scoped_allocation<system, int> buf = { 3, 2, 5, 8, 10, 29, 4, 76, 6 };
    scoped_allocation<system, int> temp_keys(buf.size());
    scoped_allocation<system, uint8> temp_storage;

    parallel<system>::sort(buf, temp_keys, temp_storage);

    printf("%d %d %d\n", buf.peek(0), buf.peek(1), buf.peek(2));
}

int main(int argc, char **argv)
{
    test<host>();
    test<cuda>();
}