Memory Model

The framework’s memory model is explicit: the C engine owns tensor buffers and PHP objects manage lifetime.

Allocation paths

• Tensor::zeros() / Tensor::ones() allocate new buffers.
• Tensor::emptyLike() allocates uninitialized memory.
• tensor_dataset_from_csv() can create direct tensor buffers from CSV.
• arena_create() allocates a bulk memory pool.

Data layout

• Row-major layout is used for all tensors.
• stride metadata tracks contiguous and non-contiguous views.
• Non-contiguous views are valid but may incur copy costs in BLAS.

Zero-copy and views

• tensor_view() and tensor_slice() produce views with shared data.
• Parent references keep the backing memory alive.
• A view becomes invalid if the parent is freed.

Page and cache behavior

• Contiguous tensors reduce page crossing.
• Large tensors should be accessed in sequential order.
• Strided access patterns on slices will reduce throughput.

Best practices

• Keep the training dataset contiguous before matrix multiplication.
• Use arena allocation for repeated temporary tensors.
• Use Tensor::copy() when a non-contiguous view must be repeatedly reused.