Inner Product Comparison with AVX-512, and AMX using BF16

This repository demonstrates the use of various hardware optimization techniques for computing the inner product of BF16 vectors using:

1. Standard CPU instructions
2. AVX-512 (Advanced Vector Extensions)
3. AMX (Advanced Matrix Extensions)

The code benchmarks the performance of these techniques and provides insights into their speed-ups for computational tasks.

---

Requirements

Hardware

• Processor: Must support AVX-512 and AMX instruction sets.
• Operating System: Linux-based systems with kernel support for AMX extensions.

Software

• Compiler: icpx

---

Features

• Data Generation: Randomly generate BF16 vectors for testing.
• Inner Product Implementations:
  • Standard Scalar Implementation: Performs inner product computation in a straightforward way.
  • AVX-512 Optimized: Leverages SIMD intrinsics for efficient computation.
  • AMX Optimized: Uses tile-based operations for high-performance dot products.

---

Build And Run

make
./out/inner-product

Output

Runnging test with dimesion:32, data size: 1600000, test size: 1600
Data generated.
Start test.

Inner product calculations completed in 366.002 seconds.

AVX-512 calculations completed in 20.9063 seconds.

AMX calculations completed in 0.482679 seconds.


Using AVX-512 is 17.5068 times faster than standard.
Using AMX is 758.272 times faster than standard and 43.313 times faster than AVX-512.