CuPy Overview

CuPy is an implementation of NumPy-compatible multi-dimensional array on CUDA. CuPy consists of the core multi-dimensional array class, cupy.ndarray, and many functions on it. It supports a subset of numpy.ndarray interface that is enough for Chainer.

The following is a brief overview of supported subset of NumPy interface:

  • Basic indexing (indexing by ints, slices, newaxes, and Ellipsis)
  • Element types (dtypes): bool_, (u)int{8, 16, 32, 64}, float{16, 32, 64}
  • Most of the array creation routines
  • Reshaping and transposition
  • All operators with broadcasting
  • All Universal functions (a.k.a. ufuncs) for elementwise operations except those for complex numbers
  • Dot product functions (except einsum) using cuBLAS
  • Reduction along axes (sum, max, argmax, etc.)

CuPy also includes following features for performance:

  • Customizable memory allocator, and a simple memory pool as an example
  • User-defined elementwise kernels
  • User-defined reduction kernels
  • cuDNN utilities

CuPy uses on-the-fly kernel synthesis: when a kernel call is required, it compiles a kernel code optimized for the shapes and dtypes of given arguments, sends it to the GPU device, and executes the kernel. The compiled code is cached to $(HOME)/.cupy/kernel_cache directory (this cache path can be overwritten by setting the CUPY_CACHE_DIR environment variable). It may make things slower at the first kernel call, though this slow down will be resolved at the second execution. CuPy also caches the kernel code sent to GPU device within the process, which reduces the kernel transfer time on further calls.

A list of supported attributes, properties, and methods of ndarray

Memory layout

base ctypes itemsize flags nbytes shape size strides

Data type

dtype

Other attributes

T

Item selection and manipulation

take() diagonal()

Arithmetic and comparison operations

__lt__() __le__() __gt__() __ge__() __eq__() __ne__() __nonzero__() __neg__() __pos__() __abs__() __invert__() __add__() __sub__() __mul__() __div__() __truediv__() __floordiv__() __mod__() __divmod__() __pow__() __lshift__() __rshift__() __and__() __or__() __xor__() __iadd__() __isub__() __imul__() __idiv__() __itruediv__() __ifloordiv__() __imod__() __ipow__() __ilshift__() __irshift__() __iand__() __ior__() __ixor__()

Special methods

__copy__() __deepcopy__() __reduce__() __array__() __len__() __getitem__() __setitem__() __int__() __long__() __float__() __oct__() __hex__() __repr__() __str__()

Memory transfer

get() set()

A list of supported routines of cupy module

Indexing routines

take() diagonal()

Padding

pad()

External Functions

scatter_add()

Other

asnumpy()