Skip to content

Numpy ring buffer at a fixed memory address to allow for significantly sped up numpy, sigpy, numba & pyFFTW calculations.

License

Notifications You must be signed in to change notification settings

Dennis-van-Gils/python-dvg-ringbuffer

Repository files navigation

https://static.pepy.tech/personalized-badge/dvg-ringbuffer?period=month&units=international_system&left_color=gray&right_color=blue&left_text=%E2%86%93%20per%20month https://img.shields.io/pypi/v/dvg-ringbuffer https://img.shields.io/pypi/pyversions/dvg-ringbuffer

DvG_RingBuffer

Provides a numpy ring buffer at a fixed memory address to allow for significantly sped up numpy, sigpy, numba & pyFFTW calculations.

Installation:

pip install dvg-ringbuffer

Based on:

https://pypi.org/project/numpy_ringbuffer/ by Eric Wieser.

DvG_RingBuffer can be used as a drop-in replacement for numpy_ringbuffer and provides several optimizations and extra features, but requires Python 3.

If and only if the ring buffer is completely full, will it return its array data as a contiguous C-style numpy array at a single fixed memory address per ring buffer instance. It does so by unwrapping the discontiguous ring buffer array into a second extra unwrap buffer that is a private member of the ring buffer class. This is advantegeous for other accelerated computations by, e.g., numpy, sigpy, numba & pyFFTW, that benefit from being fed with contiguous arrays at the same memory address each time again, such that compiler optimizations and data planning are made possible.

When the ring buffer is not completely full, it will return its data as a contiguous C-style numpy array, but at different memory addresses. This is how the original numpy-buffer always operates.

Commonly, collections.deque() is used to act as a ring buffer. The benefits of a deque is that it is thread safe and fast (enough) for most situations. However, there is an overhead whenever the deque -- a list-like container -- needs to be transformed into a numpy array. Because DvG_RingBuffer already returns numpy arrays it will outperform a collections.deque() easily, tested to be a factor of ~60.

Warning

  • This ring buffer is not thread safe. You'll have to implement your own mutex locks when using this ring buffer in multithreaded operations.
  • The data array that is returned by a full ring buffer is a pass by reference of the unwrap buffer. It is not a copy! Hence, changing values in the returned data array is identical to changing values in the unwrap buffer.

API

class RingBuffer(capacity, dtype=np.float64, allow_overwrite=True)

Create a new ring buffer with the given capacity and element type.

Args:
capacity (int):
The maximum capacity of the ring buffer
dtype (data-type, optional):

Desired type of buffer elements. Use a type like (float, 2) to produce a buffer with shape (capacity, 2).

Default: np.float64

allow_overwrite (bool, optional):

If False, throw an IndexError when trying to append to an already full buffer.

Default: True

Methods

  • clear()

  • append(value)

    Append a single value to the ring buffer.

    rb = RingBuffer(3, dtype=np.int)  #  []
    rb.append(1)                      #  [1]
    rb.append(2)                      #  [1, 2]
    rb.append(3)                      #  [1, 2, 3]
    rb.append(4)                      #  [2, 3, 4]
  • appendleft(value)

    Append a single value to the ring buffer from the left side.

    rb = RingBuffer(3, dtype=np.int)  #  []
    rb.appendleft(1)                  #  [1]
    rb.appendleft(2)                  #  [2, 1]
    rb.appendleft(3)                  #  [3, 2, 1]
    rb.appendleft(4)                  #  [4, 3, 2]
  • extend(values)

    Extend the ring buffer with a list of values.

    rb = RingBuffer(3, dtype=np.int)  #  []
    rb.extend([1])                    #  [1]
    rb.extend([2, 3])                 #  [1, 2, 3]
    rb.extend([4, 5, 6, 7])           #  [5, 6, 7]
  • extendleft(values)

    Extend the ring buffer with a list of values from the left side.

    rb = RingBuffer(3, dtype=np.int)  #  []
    rb.extendleft([1])                #  [1]
    rb.extendleft([3, 2])             #  [3, 2, 1]
    rb.extendleft([7, 6, 5, 4])       #  [7, 6, 5]
  • pop()

    Remove the right-most item from the ring buffer and return it.

  • popleft()

    Remove the left-most item from the ring buffer and return it.

Properties

  • is_full
  • unwrap_address
  • current_address
  • dtype
  • shape
  • maxlen

Indexing & slicing

  • [] including negative indices and slicing

    from dvg_ringbuffer import RingBuffer
    
    rb = RingBuffer(4, dtype=np.int)  # --> rb[:] = array([], dtype=int32)
    rb.extend([1, 2, 3, 4, 5])        # --> rb[:] = array([2, 3, 4, 5])
    x = rb[0]                         # --> x = 2
    x = rb[-1]                        # --> x = 5
    x = rb[:3]                        # --> x = array([2, 3, 4])
    x = rb[np.array([0, 2, -1])]      # --> x = array([2, 4, 5])
    
    rb = RingBuffer(5, dtype=(np.int, 2))  # --> rb[:] = array([], shape=(0, 2), dtype=int32)
    rb.append([1, 2])                      # --> rb[:] = array([[1, 2]])
    rb.append([3, 4])                      # --> rb[:] = array([[1, 2], [3, 4]])
    rb.append([5, 6])                      # --> rb[:] = array([[1, 2], [3, 4], [5, 6]])
    x = rb[0]                              # --> x = array([1, 2])
    x = rb[0, :]                           # --> x = array([1, 2])
    x = rb[:, 0]                           # --> x = array([1, 3, 5])

About

Numpy ring buffer at a fixed memory address to allow for significantly sped up numpy, sigpy, numba & pyFFTW calculations.

Topics

Resources

License

Stars

Watchers

Forks

Sponsor this project