Memory & Allocation

Optimizing memory usage in FastInterpolations.jl focuses on three key areas: vector handling, allocation patterns, and grid selection.

1. Vector Queries: Avoid Loops

When querying multiple points (xq is a vector), never loop over scalar queries.

Bad:

results = zeros(length(xq))
for i in eachindex(xq)
   results[i] = linear_interp(x, y, xq[i]) # High overhead per call
end

Good (Vector API):

results = linear_interp(x, y, xq) # Optimized batch processing

The vector API handles grid search and overhead much more efficiently internally.

To achieve true zero-allocation, combine the Vector API with pre-allocated output arrays using the ! (bang) versions of the functions.

Good (Standard Vector API):

# Allocates a new array for 'results' every time
results = linear_interp(x, y, xq)

Best (Zero-Allocation):

# 1. Pre-allocate output array (once, outside loop)
out = similar(xq)

# 2. Use in-place version inside loop
linear_interp!(out, x, y, xq)

Use this pattern in hot loops to eliminate garbage collection overhead.

For uniform grids, always use Range instead of Vector:

# ✅ Good: Range (O(1) lookup, minimal memory)
x = 0.0:0.1:10.0

# ❌ Bad: collect() degrades to Vector (O(log n) lookup)
x = collect(0.0:0.1:10.0)

For non-uniform grids where Vector is required, see Search & Hints for optimizing lookup performance.