PHP-ORT

PHP - Oh Really, Tensors?

This extension provides high performance tensor mathematics. It supports a wide range of mathematical operations on tensors, including element-wise computations, reductions, linear algebra functions, and more.

This extension also includes optional integration with Microsoft's ONNX Runtime for efficient model execution.

Features

• High-Performance Mathematics: Hardware Acceleration
  • CPU (SIMD): WASM, NEON, RISCV64, AVX512, AVX2, SSE4.1, SSE2
  • GPU: CUDA
• Hybrid Parallelism: Automatic work distribution across CPU/GPU
• Comprehensive Type Support: 10 data types with automatic promotion
• Memory Efficient: Reference-counted tensors with zero-copy slicing
• ONNX Integration: Optional and Direct ONNX model loading and inference

Quick Start

Installation

# Build extension
phpize
./configure --enable-ort
make
sudo make install

Installation with ONNX Runtime

# Fetch onnxruntime from github
wget https://github.com/microsoft/onnxruntime/releases/download/v1.22.0/onnxruntime-linux-x64-1.22.0.tgz

# Install onnxruntime
tar -C /usr/local -xvf onnxruntime-linux-x64-1.22.0.tgz --strip-components 1

# Build and install extension
phpize
./configure --with-ort-onnx=/usr/local --enable-ort
make -j$(nproc)
sudo make install

Note: Use microsoft provided releases of ONNX, not brew or ppa or apt provided releases.

Installation with CUDA

See docs/gpu.md for more information about GPU acceleration.

Post Installation

Add to your php.ini:

extension=ort.so

Basic Usage

use ORT\Tensor\Transient;
use ORT\Math;

// Create tensors
$a = new Transient([1000, 1000], $matrix_data, ORT\Tensor::FLOAT);
$b = new Transient([1000, 1000], $matrix_data, ORT\Tensor::FLOAT);

// High-performance matrix multiplication
$result = Math\matmul($a, $b);

// Element-wise operations
$sum = Math\add($a, $b);
$scaled = Math\multiply($a, 2.5);  // Scalar operations

// Reductions
$total = Math\reduce\tensor\sum($a);
$row_sums = Math\reduce\axis\sum($a, 1);

Performance

• SIMD Acceleration: Up to 8x faster with AVX2 vectorization
• Multi-Core Scaling: Automatic scaling to available CPU cores
• Memory Optimization: Aligned memory allocation for optimal performance

Check your system configuration:

echo "Backend: " . (ORT\Math\backend() ?: "scalar") . "\n";
echo "Cores: " . ORT\Math\cores() . "\n";

API Reference (not exhaustive)

Tensor Types

• ORT\Tensor\Transient - Temporary tensors
• ORT\Tensor\Persistent - Named, persistent tensors

Mathematical Operations

• Binary: add, subtract, multiply, divide, pow, mod
• Unary: sin, cos, exp, log, sqrt, abs, neg
• Matrix: matmul, dot, transpose
• Reduction: sum, mean, min, max, argmax, argmin
• Transformation: softmax

ONNX Integration

• ORT\Model - Load and manage ONNX models
• ORT\Runtime - Execute ONNX model inference

See docs/api.md for more extensive API reference.

Configuration

Environment Variables

• ORT_SCALE_CORES - Set thread pool size (default: CPU cores)

Build Options

• --enable-ort-backend=type - Enable SIMD optimizations (default: auto)
• --with-ort-onnx - Link against ONNX Runtime (default: no)

Backend Types

• auto - detect backend type for platform
• wasm - use wasm backend
• neon - use neon backend
• riscv64 - use riscv64 backend
• avx2 - use avx2 backend
• sse41 - use sse41 backend
• sse2 - use sse2 backend
• cuda - use cuda backend
• none - disable backend

Technical Details

See docs for detailed technical documentation.

Requirements

• PHP: 7.4+ or 8.0+
• ONNX Runtime: 1.16+ (optional)
• Compiler: GCC 4.8+ or Clang 3.8+
• CPU: x86_64, aarch64, riscv64, wasm

License

This project is licensed under the PHP License 3.01.

Contributing

Contributions are welcome! Please ensure:

• Code follows existing style conventions
• SIMD implementations include proper fallbacks
• Thread safety is maintained
• Tests pass across supported PHP versions

Support

For issues and feature requests, please use the project's issue tracker.