bare_NN

A minimal neural network framework written in CUDA PTX for educational purposes and low-level GPU programming exploration.

Overview

bare_NN is a bare-metal approach to understanding neural network primitives at the GPU assembly level. This project provides:

• PTX code written for basic operations of neural network
• Corresponding cuda c written to call and operate and synchronize

Features

Core Functionality

• vector operations
• Matrix multiplications
• Activation
• Reduction operation
• Matrix multiplication advanced
• Loss
• Gradient
• Forward pass
• Backward Pass

For PTX conscise tabular doc refer PTX.md

Code structure

• ptx this folder will contain the handwritten ptx
• src this folder will contain linking code to make NN
• reference this folder will contain the cuda kernels for nn components for comparison
• tests this folder will contain test cuda codes to call ptx written kernels individually
• cmake CMake helper modules for building
• include common headers and utilities

Building

Prerequisites

• CUDA Toolkit 11.0+ (tested with 12.x)
• CMake 3.18+
• C++17 compatible compiler
• Ninja (recommended) or Make

Quick Build

# Configure and build (using presets)
cmake --preset default
cmake --build --preset default

# Or manually
mkdir build && cd build
cmake ..
cmake --build .

Build Presets

Preset	Description
default	Release build with Ninja
debug	Debug build with symbols
release	Optimized release build
msvc	Visual Studio 2022
ptx-gen	Build + generate PTX files

# Debug build
cmake --preset debug
cmake --build --preset debug

# Visual Studio (Windows)
cmake --preset msvc
cmake --build --preset msvc-release

Build Options

Option	Default	Description
BUILD_TESTS	ON	Build test executables
BUILD_EXAMPLES	ON	Build example programs
GENERATE_PTX	OFF	Generate PTX from .cu files
ENABLE_FAST_MATH	ON	Enable fast math optimizations
CMAKE_CUDA_ARCHITECTURES	75;86;89	Target GPU architectures

# Custom build example
cmake -B build -DCMAKE_CUDA_ARCHITECTURES=86 -DBUILD_TESTS=ON
cmake --build build --config Release

Running Tests

# Run all tests
cd build
ctest --output-on-failure

# Or use the custom target
cmake --build build --target run_tests

# Run specific test
./build/tests/test_vector_op

Project Structure After Build

bare_NN/
├── build/
│   ├── main.exe              # Main executable
│   ├── ptx/                  # Copied PTX files
│   └── tests/
│       ├── test_vector_op.exe
│       ├── test_matmul.exe
│       └── ptx/              # PTX files for tests
├── src/
├── tests/
└── ptx/

Resources

• CUDA C Programming Guide
• PTX ISA Documentation
• NVIDIA Nsight Tools

---

Happy GPU programming!