Introduction to CuPy

Features of CuPy

CuPy is now used very frequently. Some of its features include

• With CuPy, we can write custom Python code that leverages CUDA and GPU speedups if you have something specific that still needs to be supported. 
• CuPy’s interface mirrors the Numpy. 
• It can be used as a direct replacement.
• CuPy supports the array operations of Numpy, like indexing, broadcasting, math on arrays, and matrix transformations.

Functions Implementation in CuPy

CuPy implements many functions on cupy.ndarray objects. 

To use CuPy, you will first need to install it and ensure that you have a compatible GPU and the necessary CUDA drivers and tools. You can then import CuPy and use it to define and run functions on the GPU.

Here's an example of a simple CuPy function:

import cupy as cp
def add(a, b):
    return a + b
# Create two arrays
a = cp.array([1, 2, 3, 4])
b = cp.array([5, 6, 7, 8])
# Run the function on the GPU
res = add(a, b)
print(res) # prints [6 8 10 12]

You can also try this code with Online Python Compiler

Run Code

There are various other functions, such as cp.zeros() or cp.ones(), to create arrays on the GPU and use its math functions, such as cp.exp() or cp.sin(), to perform element-wise operations on GPU arrays.

It's important to note that CuPy functions can only operate on GPU arrays, so if you want to use a NumPy array as input to a CuPy function, you will need to convert it to a CuPy array using the cp.asarray() function. Similarly, if you want to convert a CuPy array to a NumPy array, you can use the cp.asnumpy() function.

Memory Management with CuPy

CuPy has a feature memory pool which makes the performance better by lowering the overhead for memory allocations and CPU/GPU synchronization.
 

There are 2 different memory pools:

• Device memory pool is GPU device memory used for GPU memory allocations.
• Pinned memory pool is non-swappable CPU memory used during CPU-to-GPU data transfer.

Memory Pool Operations

The Memory pool instance provides memory allocation statistics. To access the default memory pool instance, we use 

• cupy.get_default_memory_pool() 
• cupy.get_default_pinned_memory_pool

Limiting GPU Memory Usage

We can put a limit on the amount of GPU memory that can be allocated using the CUPY_GPU_MEMORY_LIMIT environment variable.

The limit can be set by using cupy.cuda.MemoryPool.set_limit(). This way, we can use a different limit for each GPU device. The limit counts only the memory allocated only inside the memory pool and not the memory allocated outside of it. 

Changing Memory Pool

We can use the memory allocator rather than the default memory pool. It can be done by passing the memory allocation function to cupy.cuda.set_allocator() or cupy.cuda.set_pinned_memory_allocator() which takes 1 argument and returns  cupy.cuda.MemoryPointer. We can also disable the default memory pool accordingly.

Frequently Asked Questions

Is CuPy compatible with NumPy? 

CuPy is designed to be fully compatible with NumPy so that you can use it as a drop-in replacement for NumPy in most cases. You can use CuPy arrays instead of NumPy arrays, and many NumPy functions will also work on CuPy arrays.

Is CuPy faster than NumPy?

CuPy executes faster than NumPy as it operates on GPU rather than CPU.

Are all NumPy functions compatible with CuPy?

Not all NumPy functions are compatible with CuPy, so you may need to use CuPy-specific functions for some operations.

Conclusions

We hope this blog helped you understand the difference between NumPy and CuPy. Firstly, We discussed the basic features of both, followed by their comparison under various fields.

Refer to the Basics of C++ with Data Structure, DBMS, and Operating System by Coding Ninjas, and keep practicing on our platform Coding Ninjas Studio. You can check out the mock test series on code studio.

You can also refer to our Guided Path on Coding Ninjas Studio to upskill yourself in domains like Data Structures and Algorithms, Competitive Programming, Aptitude, and many more! Refer to the interview bundle if you want to prepare for placement interviews. Check out interview experiences to understand various companies' interview questions.

Give your career an edge over others by considering our premium courses!

Happy Learning!