pool.md

Memory Pools

Seep allocates memory when a graph is instantiated for internal nodes, result nodes, input nodes without linked arrays, and input nodes whose linked arrays are not of the correct type. By default Seep allocates memory by constructing new arrays. To reduce memory usage, and to share memory between instances, you may pass a memory pool to the instance function.

x = ANode(4); y = 2*x; z = 2*y;
memory = zeros(Float32, 16)
pool = BuddyPool(memory)

func,dict = instance(pool, z)

Seep will use the provided memory to store the data of this instance graph.

After setting the input data to something other than zeros, you can see the changes reflected in the backing array.

dict[x][:] = 1:4
println(memory')

Seep reclaims memory eagerly within the memory pool. After running the instance function, the only arrays that are guaranteed to be populated are those corresponding to the nodes passed to the instance constructor.

func()
println(dict[z]') # prints [4.0  8.0  12.0  16.0]
println(dict[y]') # error, y wasn't passed to instance
println(dict[x]') # also prints [4.0  8.0  12.0  16.0]!  Since x was not passed to instance, Seep reused the memory for z

Note that you must also be careful not to reference the instance arrays after the memory pool has been deallocated, since they will still point into the invalid block.