Maximising Python Speed¶

Algorithms¶

The most important aspect of designing any routine for performance is ensuring that the best algorithm is employed. This is a topic for textbooks rather than for a MicroPython guide but spectacular performance gains can sometimes be achieved by adopting algorithms known for their efficiency.

RAM Allocation¶

To design efficient MicroPython code it is necessary to have an understanding of the way the interpreter allocates RAM. When an object is created or grows in size (for example where an item is appended to a list) the necessary RAM is allocated from a block known as the heap. This takes a significant amount of time; further it will on occasion trigger a process known as garbage collection which can take several milliseconds.

Consequently the performance of a function or method can be improved if an object is created once only and not permitted to grow in size. This implies that the object persists for the duration of its use: typically it will be instantiated in a class constructor and used in various methods.

This is covered in further detail Controlling garbage collection below.

Buffers¶

An example of the above is the common case where a buffer is required, such as one used for communication with a device. A typical driver will create the buffer in the constructor and use it in its I/O methods which will be called repeatedly.

The MicroPython libraries typically provide support for pre-allocated buffers. For example, objects which support stream interface (e.g., file or UART) provide read() method which allocate new buffer for read data, but also a readinto() method to read data into an existing buffer.

Floating Point¶

Some MicroPython ports allocate floating point numbers on heap. Some other ports may lack dedicated floating-point coprocessor, and perform arithmetic operations on them in “software” at considerably lower speed than on integers. Where performance is important, use integer operations and restrict the use of floating point to sections of the code where performance is not paramount. For example, capture ADC readings as integers values to an array in one quick go, and only then convert them to floating-point numbers for signal processing.

Arrays¶

Consider the use of the various types of array classes as an alternative to lists. The array module supports various element types with 8-bit elements supported by Python’s built in bytes and bytearray classes. These data structures all store elements in contiguous memory locations. Once again to avoid memory allocation in critical code these should be pre-allocated and passed as arguments or as bound objects.

When passing slices of objects such as bytearray instances, Python creates a copy which involves allocation of the size proportional to the size of slice. This can be alleviated using a memoryview object. memoryview itself is allocated on heap, but is a small, fixed-size object, regardless of the size of slice it points too.

ba = bytearray(10000)  # big array
func(ba[30:2000])      # a copy is passed, ~2K new allocation
mv = memoryview(ba)    # small object is allocated
func(mv[30:2000])      # a pointer to memory is passed

A memoryview can only be applied to objects supporting the buffer protocol - this includes arrays but not lists. Small caveat is that while memoryview object is live, it also keeps alive the original buffer object. So, a memoryview isn’t a universal panacea. For instance, in the example above, if you are done with 10K buffer and just need those bytes 30:2000 from it, it may be better to make a slice, and let the 10K buffer go (be ready for garbage collection), instead of making a long-living memoryview and keeping 10K blocked for GC.

Nonetheless, memoryview is indispensable for advanced preallocated buffer management. .readinto() method discussed above puts data at the beginning of buffer and fills in entire buffer. What if you need to put data in the middle of existing buffer? Just create a memoryview into the needed section of buffer and pass it to .readinto().

The const() declaration¶

MicroPython provides a const() declaration. This works in a similar way to #define in C in that when the code is compiled to bytecode the compiler substitutes the numeric value for the identifier. This avoids a dictionary lookup at runtime. The argument to const() may be anything which, at compile time, evaluates to an integer e.g. 0x100 or 1 << 8.

Caching object references¶

Where a function or method repeatedly accesses objects performance is improved by caching the object in a local variable:

class foo(object):
    def __init__(self):
        ba = bytearray(100)
    def bar(self, obj_display):
        ba_ref = self.ba
        fb = obj_display.framebuffer
        # iterative code using these two objects

This avoids the need repeatedly to look up self.ba and obj_display.framebuffer in the body of the method bar().

Controlling garbage collection¶

When memory allocation is required, MicroPython attempts to locate an adequately sized block on the heap. This may fail, usually because the heap is cluttered with objects which are no longer referenced by code. If a failure occurs, the process known as garbage collection reclaims the memory used by these redundant objects and the allocation is then tried again - a process which can take several milliseconds.

There are benefits in pre-empting this by periodically issuing gc.collect(). Firstly doing a collection before it is actually required is quicker - typically on the order of 1ms if done frequently. Secondly you can determine the point in code where this time is used rather than have a longer delay occur at random points, possibly in a speed critical section. Finally performing collections regularly can reduce fragmentation in the heap. Severe fragmentation can lead to non-recoverable allocation failures.

Accessing hardware directly¶

This comes into the category of more advanced programming and involves some knowledge of the target MCU. Consider the example of toggling an output pin on the Pyboard. The standard approach would be to write

mypin.value(mypin.value() ^ 1) # mypin was instantiated as an output pin

This involves the overhead of two calls to the Pin instance’s value() method. This overhead can be eliminated by performing a read/write to the relevant bit of the chip’s GPIO port output data register (odr). To facilitate this the stm module provides a set of constants providing the addresses of the relevant registers. A fast toggle of pin P4 (CPU pin A14) - corresponding to the green LED - can be performed as follows:

BIT14 = const(1 << 14)
stm.mem16[stm.GPIOA + stm.GPIO_ODR] ^= BIT14