Basics of CuPy¶
In this section, you will learn about the following things:
The concept of current device
host-device and device-device array transfer
Basics of cupy.ndarray¶
CuPy is a GPU array backend that implements a subset of NumPy interface.
In the following code,
cp is an abbreviation of
cupy, following the standard convention of abbreviating
>>> import numpy as np >>> import cupy as cp
>>> x_gpu = cp.array([1, 2, 3])
x_gpu above is an instance of
As one can see, CuPy’s syntax here is identical to that of NumPy.
The main difference between
numpy.ndarray is that
the CuPy arrays are allocated on the current device, which we will talk about later.
Most of the array manipulations are also done in the way similar to NumPy.
Take the Euclidean norm (a.k.a L2 norm), for example.
numpy.linalg.norm() function that calculates it on CPU.
>>> x_cpu = np.array([1, 2, 3]) >>> l2_cpu = np.linalg.norm(x_cpu)
Using CuPy, we can perform the same calculations on GPU in a similar way:
>>> x_gpu = cp.array([1, 2, 3]) >>> l2_gpu = cp.linalg.norm(x_gpu)
CuPy implements many functions on
See the reference for the supported subset of NumPy API.
Knowledge of NumPy will help you utilize most of the CuPy features.
We, therefore, recommend you familiarize yourself with the NumPy documentation.
CuPy has a concept of a current device, which is the default GPU device on which
the allocation, manipulation, calculation, etc., of arrays take place.
Suppose ID of the current device is 0.
In such a case, the following code would create an array
x_on_gpu0 on GPU 0.
>>> x_on_gpu0 = cp.array([1, 2, 3, 4, 5])
The current device can be changed using
cupy.cuda.Device.use() as follows:
>>> x_on_gpu0 = cp.array([1, 2, 3, 4, 5]) >>> cp.cuda.Device(1).use() >>> x_on_gpu1 = cp.array([1, 2, 3, 4, 5])
To temporarily switch to another GPU device, use
with context manager:
>>> with cp.cuda.Device(1): ... x_on_gpu1 = cp.array([1, 2, 3, 4, 5]) >>> x_on_gpu0 = cp.array([1, 2, 3, 4, 5])
Most CuPy operations are performed on the currently active device and attempts to process an array stored on a different device will result in an error:
>>> with cp.cuda.Device(0): ... x_on_gpu0 = cp.array([1, 2, 3, 4, 5]) >>> with cp.cuda.Device(1): ... x_on_gpu0 * 2 # raises error Traceback (most recent call last): ... ValueError: Array device must be same as the current device: array device = 0 while current = 1
cupy.ndarray.device attribute indicates the device on which the array is allocated.
>>> with cp.cuda.Device(1): ... x = cp.array([1, 2, 3, 4, 5]) >>> x.device <CUDA Device 1>
When only one device is available, explicit device switching is not needed.
Associated with the concept of current devices are current streams, which help avoid explicitly passing streams in every single operation so as to keep the APIs pythonic and user-friendly. In CuPy, all CUDA operations such as data transfer (see the Data Transfer section) and kernel launches are enqueued onto the current stream, and the queued tasks on the same stream will be executed in serial (but asynchronously with respect to the host).
The default current stream in CuPy is CUDA’s null stream (i.e., stream 0). It is also known as the legacy
default stream, which is unique per device. However, it is possible to change the current stream using the
cupy.cuda.Stream API, please see Accessing CUDA Functionalities for example. The current stream in CuPy can be
It is worth noting that CuPy’s current stream is managed on a per thread basis, meaning that on different Python threads the current stream (if not the null stream) can be different.
Move arrays to a device¶
>>> x_cpu = np.array([1, 2, 3]) >>> x_gpu = cp.asarray(x_cpu) # move the data to the current device.
>>> with cp.cuda.Device(0): ... x_gpu_0 = cp.ndarray([1, 2, 3]) # create an array in GPU 0 >>> with cp.cuda.Device(1): ... x_gpu_1 = cp.asarray(x_gpu_0) # move the array to GPU 1
cupy.asarray() does not copy the input array if possible.
So, if you put an array of the current device, it returns the input object itself.
Move array from a device to the host¶
Moving a device array to the host can be done by
cupy.asnumpy() as follows:
>>> x_gpu = cp.array([1, 2, 3]) # create an array in the current device >>> x_cpu = cp.asnumpy(x_gpu) # move the array to the host.
We can also use
>>> x_cpu = x_gpu.get()
Check Memory Management for a detailed description of how memory is managed in CuPy using memory pools.
How to write CPU/GPU agnostic code¶
CuPy’s compatibility with NumPy makes it possible to write CPU/GPU agnostic code.
For this purpose, CuPy implements the
cupy.get_array_module() function that
returns a reference to
cupy if any of its arguments resides on a GPU
Here is an example of a CPU/GPU agnostic function that computes
>>> # Stable implementation of log(1 + exp(x)) >>> def softplus(x): ... xp = cp.get_array_module(x) # 'xp' is a standard usage in the community ... print("Using:", xp.__name__) ... return xp.maximum(0, x) + xp.log1p(xp.exp(-abs(x)))
When you need to manipulate CPU and GPU arrays, an explicit data
transfer may be required to move them to the same location – either CPU or GPU.
For this purpose, CuPy implements two sister methods called
cupy.asarray(). Here is an example that demonstrates the use of both methods:
>>> x_cpu = np.array([1, 2, 3]) >>> y_cpu = np.array([4, 5, 6]) >>> x_cpu + y_cpu array([5, 7, 9]) >>> x_gpu = cp.asarray(x_cpu) >>> x_gpu + y_cpu Traceback (most recent call last): ... TypeError: Unsupported type <class 'numpy.ndarray'> >>> cp.asnumpy(x_gpu) + y_cpu array([5, 7, 9]) >>> cp.asnumpy(x_gpu) + cp.asnumpy(y_cpu) array([5, 7, 9]) >>> x_gpu + cp.asarray(y_cpu) array([5, 7, 9]) >>> cp.asarray(x_gpu) + cp.asarray(y_cpu) array([5, 7, 9])
cupy.asnumpy() method returns a NumPy array (array on the host),
cupy.asarray() method returns a CuPy array (array on the current device).
Both methods can accept arbitrary input, meaning that they can be applied to any data that
is located on either the host or device and can be converted to an array.