Getting started¶

Requirements¶

• C++17 (the floor set by CCCL, which refuses to build lower).
• A host compiler — tested on g++ 13 and clang++ 18 — and/or nvcc.
• CCCL (libcudacxx), vendored under external/cccl.

teeny is header-only: there is nothing to build or link. Add two include paths:

-I path/to/teeny/include
-I path/to/teeny/external/cccl/libcudacxx/include

With CMake¶

teeny ships a CMakeLists.txt that exposes it as an INTERFACE (header-only) target. Linking it inherits the include directories, the C++17 requirement, and the CCCL dependency, so there is nothing to compile.

# In-tree (vendored, or as a submodule):
add_subdirectory(teeny)
target_link_libraries(app PRIVATE teeny::teeny)

# Or, after `find_package(teeny CONFIG)` on an installed copy:
target_link_libraries(app PRIVATE teeny::teeny)

CCCL (libcudacxx) is resolved in this order: an explicit include dir you point at with -DTEENY_CCCL_INCLUDE=/path/to/cccl/include (the knob for a per-CUDA choice), then an installed CCCL via find_package(CCCL), then the vendored submodule under external/cccl. See CUDA & CCCL compatibility.

Hello, tensor¶

#include <teeny/teeny.h>
#include <cstdio>
using namespace tny;

int main() {
    double buf[6] = {1, 2, 3, 4, 5, 6};
    auto m = wrap(buf, shape<2, 3>{});     // a 2×3 view over `buf`
    m(1, 2) = 60;                          // write through the view
    std::printf("%g  sum=%g\n", m(1, -1), m.sum());  // -1 = last column
    return 0;
}

g++ -std=c++17 -I include -I external/cccl/libcudacxx/include hello.cpp -o hello

Building the tests and examples¶

make run-test        # build + run every tests/test_*.cpp, printing PASS/FAIL
make run-examples    # build + run the standalone example algorithms
make CXX=clang++ run-test    # pick a compiler

One include, one namespace¶

#include <teeny/teeny.h>  // the whole library
using namespace tny;      // the public namespace

teeny/teeny.h pulls in everything. The CUDA compiler and runtime are detected automatically, in which case tensors can allocate and work with CUDA memory (device or pinned host). When CUDA is not available, these variants are not included. You do not need your own __CUDACC__ guards. Define TNY_NO_CUDA to force the CUDA support off.