acn-accel

ACN-Accel is a hybrid Python + C + Assembly project with AVX2 acceleration for core neural-network primitives.

The runtime chooses between baseline C, AVX2 intrinsics, and AVX2 assembly implementations based on:

CPU feature availability (AVX2)
input size threshold

Implemented Ops

vector_add(float* a, float* b, float* out, int n)
relu(float* x, float* out, int n)
matmul(float* A, float* B, float* C, int M, int N, int K)

All tensors use row-major layout.

Architecture

Python API (ctypes)
      |
      v
  libacn (shared)
      |
      +--> adaptive.c (dispatcher + AVX2 detection)
             |            |
             |            +--> simd_avx2_asm.S (AVX2 assembly kernel)
             |
             |            +--> simd_avx2.c (AVX2 intrinsics)
             |
             +--> core.c (baseline C)

Adaptive Execution Logic

Runtime AVX2 detection: acn_cpu_has_avx2()
Dispatch rules:
- vector_add: ASM path when n >= 2048, otherwise AVX2 intrinsics for n >= 256 (if AVX2 is available)
- relu: AVX2 path when n >= 256 and AVX2 is available
- matmul: AVX2 path when M*N*K > 100000 and AVX2 is available
- otherwise baseline C path

Build Instructions

1. Configure + build

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

This produces shared library:

Linux: build/libacn.so
macOS: build/libacn.dylib (or equivalent by generator)

2. Python environment

Install dependencies:

python -m pip install numpy pytest

3. Run tests

PYTHONPATH=python ACN_LIB_PATH=build/libacn.so pytest -q

On macOS use .dylib path in ACN_LIB_PATH.

4. Run benchmark

PYTHONPATH=python ACN_LIB_PATH=build/libacn.so python python/benchmark.py

On macOS use .dylib path in ACN_LIB_PATH.

Benchmark Output Example

[ACN Benchmark]

Platform: Linux-...
CPU: AVX2 = True

Matrix 512x512:
NumPy:   120.00 ms
C:        85.00 ms
SIMD:     40.00 ms
AUTO:     41.00 ms
Speedup SIMD vs NumPy: 3.00x

Notes on Memory Handling

SIMD code uses aligned loads/stores when input/output pointers are 32-byte aligned.
Unaligned pointers are handled safely with _mm256_loadu_ps/_mm256_storeu_ps.
Optional helpers are provided:
- acn_aligned_alloc(alignment, size)
- acn_aligned_free(ptr)

Assembly Kernel

File: src/simd_avx2_asm.S
Function: vector_add_avx2_asm(const float* a, const float* b, float* out, int n)
ISA: AVX2 (ymm registers, vmovups, vaddps, scalar tail loop)
Python access: acn.vector_add(a, b, impl="asm")

Project Layout

SVG ↗

Source