Home

AOCL-DLP Documentation Hub

Welcome to the AOCL-DLP (AMD Optimizing CPU Libraries - Deep Learning Primitives) documentation hub! This wiki serves as your central navigation point to all AOCL-DLP documentation resources.

About AOCL-DLP

AOCL-DLP is a high-performance library designed to provide optimized deep learning primitives for AMD processors. It implements Low Precision GEMM (LPGEMM) operations for machine learning applications, supporting multiple data types, pre-operations, and post-operations. The library is specifically optimized to leverage AMD hardware capabilities including AVX2, AVX512, AVX512_VNNI, and AVX512_BF16 instruction sets.

Key Features

• Highly Optimized GEMM Operations - High-performance matrix multiplication targeting AMD CPUs
• Multiple Data Type Support - Various precision formats (float, bfloat16, int8, uint8, int32)
• Comprehensive Post-Operations - SUM, ELTWISE, BIAS, SCALE, MATRIX_ADD, MATRIX_MUL
• Batch GEMM Support - Optimized for handling multiple GEMM operations
• Symmetric Quantization - Specialized routines for quantized operations
• Thread Optimization - Parallel execution via OpenMP

📚 Documentation Resources

Project Information

• 🏠 Project README - Overview and feature summary
• 🔧 Build Instructions - Compilation guide
• 📦 Installation Guide - Installation steps
• 📄 License - BSD 3-Clause License
• 📚 Examples - Examples of how to use the library

Community & Support

• 🤝 Contributing Guide
• 📋 Code of Conduct
• 🔒 Security Policy
• 💬 Support Resources

Development Tools & Coverage

• 📊 Code Coverage Tools - Comprehensive coverage analysis and HTML report generation

🔧 API Documentation

Core APIs (C Interface)

GEMM API

High-performance General Matrix Multiplication operations supporting various data type combinations:

• Float Operations: f32f32f32of32
• Bfloat16 Operations: bf16bf16f32of32, bf16bf16f32obf16, bf16s4f32of32, bf16s4f32obf16
• Integer Operations: u8s8s32os32, u8s8s32os8, u8s8s32ou8, u8s8s32of32, u8s8s32obf16
• Signed Integer: s8s8s32os32, s8s8s32os8, s8s8s32ou8, s8s8s32of32, s8s8s32obf16

📖 Detailed API Reference: Doxygen API Documentation

Batch GEMM API

Optimized batch processing for multiple GEMM operations with support for all data type combinations.

📖 Detailed API Reference: Doxygen Batch GEMM Documentation

Post-Operations API

Comprehensive post-processing operations that can be chained with GEMM:

• SUM: Element-wise addition with scaling and zero point
• ELTWISE: Activation functions (RELU, PRELU, GELU_TANH, GELU_ERF, CLIP, SWISH, TANH, SIGMOID)
• BIAS: Bias addition to results
• SCALE: Scaling operations
• MATRIX_ADD/MUL: Matrix operations with scaling

📖 Detailed API Reference: Doxygen Post-Operations Documentation

Quantization API

Specialized symmetric quantization routines:

• s8s8s32of32_sym_quant
• s8s8s32obf16_sym_quant

📖 Detailed API Reference: Doxygen Quantization Documentation

Utility Functions API

Standalone mathematical operations:

• gelu_tanh_f32 - GELU activation with tanh approximation
• gelu_erf_f32 - GELU activation with erf approximation
• softmax_f32 - Softmax function for float

📖 Detailed API Reference: Doxygen Utility Functions Documentation

Eltwise Operations API

Specialized element-wise operations supporting various input/output type combinations:

• bf16of32, bf16obf16, f32of32, f32obf16, f32os32, f32os8

📖 Detailed API Reference: Doxygen Eltwise Operations Documentation

Library Management API

Essential functions for library initialization, cleanup, and configuration.

📖 Detailed API Reference: Doxygen Library Interface Documentation

📋 Complete API Reference

Comprehensive Documentation

• 🔗 Complete Doxygen API Documentation - Full API reference with detailed function descriptions
• 🔗 Sphinx Documentation - User guide and tutorials

💡 Examples and Tutorials

Code Examples

Explore practical implementations and usage patterns:

• 🎯 Simple GEMM Operations

  • Float32 GEMM
  • Bfloat16 GEMM
  • Integer GEMM

• 🔄 Batch GEMM Examples

• ⚡ Post-Operations Examples

  • GEMM with Bias
  • GEMM with ReLU
  • Combined Post-Operations

• 🎨 Eltwise Operations

• 📊 Quantization Examples

• 🔄 Matrix Reordering

Tutorial Documentation

• 🔗 Getting Started Tutorial
• 🔗 Performance Optimization Guide
• 🔗 Advanced Usage Patterns

🧪 Testing and Validation

Testing Framework

Comprehensive testing suite for validation and performance benchmarking:

• Integration Tests - End-to-end workflow testing
• Performance Benchmarks - Speed and accuracy measurements
• Correctness Validation - Numerical accuracy verification

📖 Testing Documentation: Testing Framework Guide

🏗️ Development Resources

Build System

• CMake Configuration - Modern CMake-based build system
• Cross-platform Support - Linux, Windows compatibility
• Dependency Management - Automated dependency resolution
• Build Options - Customizable build configurations

Development Tools

• Code Formatting - Consistent code style with clang-format
• Pre-commit Hooks - Automated code quality checks
• Continuous Integration - Automated testing and validation

⚡ Performance and Hardware Support

Optimized for AMD Hardware

• Zen1+ Architecture - AVX2/FMA3 support
• Zen4+ Architecture - AVX512, AVX512_VNNI, AVX512_BF16 support
• x86_64 Compatibility - Runs on any compatible x86_64 CPU

Performance Features

• NUMA Awareness - Optimized memory access patterns
• Thread Scaling - Efficient parallel execution
• Memory Management - Optimized data layouts and caching

---

This documentation hub is continuously updated. For the latest information, please refer to the linked resources above.