Performance Optimization

This page describes the low-level optimizations that make the framework efficient on CPU-bound workloads.

Core optimizations

• native C kernels for tensor math
• BLAS-backed matrix multiplication and vector operations
• OpenMP parallel loops for reduction and elementwise kernels
• fused operations to reduce boundary overhead

Fused kernels

The C backend exposes fused primitives such as:

• tensor_linear() for X @ W^T + bias
• tensor_add_relu() for fused activation
• tensor_fused_adam_step() for optimizer updates

Fused kernels reduce memory traffic and FFI call count.

BLAS and LAPACKE

• tensor_matmul() uses OpenBLAS when available.
• tensor_matmul_ex() supports transposed inputs without extra copies.
• Dense matrix operations are optimized for row-major layout.

Hot path minimization

• Cache the FFI instance in PHP.
• Avoid repeated metadata lookup in inner loops.
• Use native reduction kernels like tensor_sum_axis() instead of PHP loops.

When to use

• Use native fused kernels for training kernels and inference.
• Use BLAS-backed matmul for large matrix multiplication.

When not to use

• Avoid elementwise operations in PHP if the same result can be computed with a bulk kernel.
• Avoid repeated conversions between contiguous and non-contiguous tensors.