Library Overview

Library Overview

High-level overview of AOCL-DLP architecture, components, and design goals.

Components

• GEMM kernels and drivers
• Post-operations framework (metadata-driven)
• Element-wise utilities
• Threading and parallelization controls

Data Types

• float32, bfloat16, int8/uint8, int32; mixed-precision flows
• Types: Types API Reference

BFloat16 API Behavior

AOCL-DLP automatically handles BF16 operations on hardware lacking native AVX512_BF16 ISA support by transparently rerouting to F32 implementations.

Hardware Support:

• Native BF16: Intel Cooper Lake/Sapphire Rapids+, AMD Zen4+ (uses AVX512_BF16 instructions)
• F32 Fallback: Automatically activated on:
  • AVX2 machines (uses AVX2 F32 kernels)
  • AVX512 without BF16 support: Intel Skylake, Cascade Lake, Ice Lake (uses AVX512 F32 kernels)

Key Points:

• BF16 API calls work unchanged across all hardware
• Library performs runtime detection and automatic rerouting
• When fallback is active: BF16→F32 conversion, F32 computation, F32→BF16 conversion (if needed)
• Performance impact on fallback: conversion overhead, 2x memory bandwidth usage

Call Layers

1. Prepare data (layouts, leading dimensions)
2. Optional reordering for repeated use
3. Configure dlp_metadata_t for fused post-ops
4. Call GEMM or eltwise
5. Optional de/reordering for outputs

Hardware Features

Targets AVX2/FMA3, AVX512, AVX512_VNNI, AVX512_BF16 on supported AMD CPUs.