Skip to content

prisma-ai/stampede

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

38 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

What is this

Graph-level abstraction over heavy computations.

  1. Graph built at compile-time (zero overhead)
  2. Full typesafety
  3. Separate independent computational traits (i.e. caching)

Table of contents

  1. How to use
  2. Building

How to use

Suppose we want to add two number, the calculation of which is a heavy operation.

f(x) = Add( Heavy_0(Id(x)) , Heavy_1(Id(x)) )

  1. Implement Node's with necessary computations

Prototype:

declare_node({UserDefined}, {StateType}, {OutputType}, {InputTypesWithComma})
{OutputType} {UserDefined}::runImpl({InputTypesWithComma} args...) {
    ...
}

that expands to:

template <typename...>
class {UserDefined} : public Node<{UserDefined}, {StateType}, {OutputType}, {InputTypesWithComma}> {
    friend class Node;

public:
    {UserDefined}(): Node<{UserDefined}, {StateType}, {OutputType}, {InputTypesWithComma}>("{user_defined}") {}
private:
    {OutputType} runImpl({InputTypesWithComma} args...) {
      ...
    }
};

Example nodes:

#include <Graph.h>

template <typename...>
class Id : public Node<Sum, int, int> {
    friend class Node;

public:
    Id(): Node<Id, int, int>("id") {}
private:
    int runImpl(int arg) {
        return arg;
    }
};


template <typename...>
class Sum : public Node<Sum, int, int, int> {
    friend class Node;

public:
    Summer(): Node<Summer, int, int, int>("sum") {}
private:
    int runImpl(int arg0, int arg1) {
        return arg0 + arg1;
    }
};

template <typename...>
class Heavy_0 : public Heavy_0<Heavy_0, int, int> {
    friend class Node;

public:
    Heavy_0(): Node<Heavy_0, int, int>("Heavy_0") {}
private:
    int runImpl(int arg) {
        return someHeavyOperation(arg);
    }
};

template <typename...>
class Heavy_1 : public Heavy_1<Heavy_1, int, int> {
    friend class Node;

public:
    Heavy_1(): Node<Heavy_1, int, int>("Heavy_1") {}
private:
    int runImpl(int arg) {
        return anotherHeavyOperation(arg);
    }
};
  1. Extends your Node's with necessary traits

In example we want to parallelize our heavy operations.

#include <Trait.h>

using AsyncHeavy_0 = AsyncTrait<Heavy_0<>>;
using AsyncHeavy_1 = AsyncTrait<Heavy_1<>>;
  1. Enumerate Node's with indicies

In example we assigning:

  • Id<> with id 0
  • AsyncHeavy_0<> with id 1
  • AsyncHeavy_1<> with id 2
  • Sum<> with id 3
#define NODES IndexedNode <0, Id<>> , IndexedNode <1, AsyncHeavy_0<>>, IndexedNode <2, AsyncHeavy_1<>>, IndexedNode <3, Sum<>>
  1. Declare edges that forms node_id <- dependecies_ids

In example we assuming that:

  • inputs for Sum<> (id 3) is AsyncHeavy_0<> (id 1) and AsyncHeavy_1<> (id 2)
  • input for AsyncHeavy_0<> (id 1) is Id<> (id 0)
  • input for AsyncHeavy_0<> (id 2) is Id<> (id 0)
#define EDGES Edge<3, std::tuple< Int<1>, Int<2> >> , Edge<2, std::tuple< Int<0> >>, Edge<1, std::tuple< Int<0> >>
  1. Select garbage-collection policy

By default all nodes instances has same lifecycle as Context (GCPlan == NoPlan).

If you need clear intermediate data (such as a cache) as soon as possible (then there are no other nodes remains thats uses this data ) you can select this behaviour.

#define PLAN BFSLastRecentlyUsedGCPlanImpl
  1. Run computations

In example we're assuming that Id<> (id 0) will be source, Sum<> (id 3) will be target node

std::tuple<int> input = ...;
auto output = withNodes<NODES>::andEdges<EDGES>{}.execute<Inputs<Int<0>>, 3, PLAN>( { input } );

See example of semipractical case of migrating at stampede in src/cv_samples

Building

For building example / test:

mkdir build
cd build
cmake ..
make
./graph_proc

For using in another CMake project:

CMakeLists.txt

...
include_directory({path/to/stampede/src})
...

Special thanks

@evsluzh for early-stage review

@opedge for early-stage review

@ashagraev for help in YaTalks report

Used in

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published