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

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

  • Basic indexing (indexing by ints, slices, newaxes, and Ellipsis)
  • Most of Advanced indexing (except for some indexing patterns with boolean masks)
  • Data types (dtypes): bool_, int8, int16, int32, int64, uint8, uint16, uint32, uint64, float16, float32, float64, complex64, complex128
  • Most of the array creation routines (empty, ones_like, diag, etc.)
  • Most of the array manipulation routines (reshape, rollaxis, concatenate, etc.)
  • All operators with broadcasting
  • All universal functions for elementwise operations (except those for complex numbers).
  • Linear algebra functions, including product (dot, matmul, etc.) and decomposition (cholesky, svd, etc.), accelerated by cuBLAS.
  • Reduction along axes (sum, max, argmax, etc.)

CuPy also includes the following features for performance:

  • User-defined elementwise CUDA kernels
  • User-defined reduction CUDA kernels
  • Fusing CUDA kernels to optimize user-defined calculation
  • Customizable memory allocator and memory pool
  • 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.