libshit

A small CUDA-enabled C++ library with tensor operations, autograd-style gradient tape, and a lightweight layers API.

Requirements

• Windows 10/11
• Visual Studio 2022 or newer with Desktop development for C++
• NVIDIA CUDA Toolkit 11.x or newer (tested with 13.2)
• CMake 3.18 or newer
• Ninja (recommended) or another supported CMake generator

Build Instructions

1. Open a Visual Studio Developer PowerShell or x64 Native Tools command prompt.

2. Change into the repository root:

cd "libshit"

3. Create the build directory if needed and configure the project:

cmake -S . -B build -G Ninja

This will locate the CUDA toolkit via find_package(CUDAToolkit REQUIRED) and generate the build files.

4. Build the library and tests:

cmake --build build --config Debug --target all

Alternatively, build a specific test target:

cmake --build build --config Debug --target grad_tests -- -j1

Run Tests

Run all CTest tests from the build directory:

cd build
ctest -C Debug --output-on-failure

Or run a single test executable directly:

.\layers_model_tests.exe

now personally i ignore all of that and just use my IDE tooling.

Workspace Layout

• include/: public headers
• src/: library and CUDA kernel sources
• tests/: test executables
• build/: out-of-tree CMake build directory

Notes

• The library target libshit is built as a shared DLL.
• CUDA headers are exposed to consumers so headers like shit_errors.hpp can compile in test targets.
• The project includes several tests:
  • basic_tests
  • grad_tests
  • grad_loss_scaling_tests
  • layers_model_tests
  • mse_tests