Maximizing Python Speed

This tutorial describes ways of improving the performance of MicroPython code. Optimizations involving other languages are covered elsewhere, namely the use of modules written in C and the MicroPython inline ARM Thumb-2 assembler.

The process of developing high performance code comprises the following stages which should be performed in the order listed.

  • Design for speed.
  • Code and debug.

Optimization steps:

  • Identify the slowest section of code.
  • Improve the efficiency of the Python code.
  • Use the native code emitter.
  • Use the viper code emitter.

Designing for speed

Performance issues should be considered at the outset. This involves taking a view on the sections of code which are most performance critical and devoting particular attention to their design. The process of optimization begins when the code has been tested: if the design is correct at the outset optimization will be straightforward and may actually be unnecessary.

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 the stream interface (e.g., file or UART) provide a read() method which allocates a 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 a dedicated floating-point coprocessor (the OpenMV Cam has a 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. A memoryview object 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 object can only be applied to objects supporting the buffer protocol - this includes arrays but not lists. A small caveat is that while the memoryview object is live, it also keeps alive the original buffer object. So, a memoryview object 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 object and keeping 10K blocked for GC.

Nonetheless, the memoryview class is indispensable for advanced preallocated buffer management. The .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().

Identifying the slowest section of code

This is a process known as profiling and is covered in textbooks and (for standard Python) supported by various software tools. For the type of smaller embedded application likely to be running on MicroPython platforms the slowest function or method can usually be established by judicious use of the timing ticks group of functions documented here. Code execution time can be measured in ms, us, or CPU cycles.

The following enables any function or method to be timed by adding an @timed_function decorator:

def timed_function(f, *args, **kwargs):
    myname = str(f).split(' ')[1]
    def new_func(*args, **kwargs):
        t = time.ticks_us()
        result = f(*args, **kwargs)
        delta = time.ticks_diff(t, time.ticks_us())
        print('Function {} Time = {:6.3f}ms'.format(myname, delta/1000))
        return result
    return new_func

MicroPython code improvements

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

The Native code emitter

This causes the MicroPython compiler to emit ARM native opcodes rather than bytecode. It covers the bulk of the Python language so most functions will require no adaptation (but see below). It is invoked by means of a function decorator:

@micropython.native
def foo(self, arg):
    buf = self.linebuf # Cached object
    # code

There are certain limitations in the current implementation of the native code emitter.

  • Context managers are not supported (the with statement).
  • Generators are not supported.
  • If raise is used an argument must be supplied.

The trade-off for the improved performance (roughly twice as fast as bytecode) is an increase in compiled code size.

The Viper code emitter

The optimizations discussed above involve standards-compliant Python code. The Viper code emitter is not fully compliant. It supports special Viper native data types in pursuit of performance. Integer processing is non-compliant because it uses machine words: arithmetic on 32 bit hardware is performed modulo 2**32.

Like the Native emitter Viper produces machine instructions but further optimizations are performed, substantially increasing performance especially for integer arithmetic and bit manipulations. It is invoked using a decorator:

@micropython.viper
def foo(self, arg: int) -> int:
    # code

As the above fragment illustrates it is beneficial to use Python type hints to assist the Viper optimizer. Type hints provide information on the data types of arguments and of the return value; these are a standard Python language feature formally defined here PEP0484. Viper supports its own set of types namely int, uint (unsigned integer), ptr, ptr8, ptr16 and ptr32. The ptrX types are discussed below. Currently the uint type serves a single purpose: as a type hint for a function return value. If such a function returns 0xffffffff Python will interpret the result as 2**32 -1 rather than as -1.

In addition to the restrictions imposed by the native emitter the following constraints apply:

  • Functions may have up to four arguments.
  • Default argument values are not permitted.
  • Floating point may be used but is not optimized.

Viper provides pointer types to assist the optimizer. These comprise

  • ptr Pointer to an object.
  • ptr8 Points to a byte.
  • ptr16 Points to a 16 bit half-word.
  • ptr32 Points to a 32 bit machine word.

The concept of a pointer may be unfamiliar to Python programmers. It has similarities to a Python memoryview object in that it provides direct access to data stored in memory. Items are accessed using subscript notation, but slices are not supported: a pointer can return a single item only. Its purpose is to provide fast random access to data stored in contiguous memory locations - such as data stored in objects which support the buffer protocol, and memory-mapped peripheral registers in a microcontroller. It should be noted that programming using pointers is hazardous: bounds checking is not performed and the compiler does nothing to prevent buffer overrun errors.

Typical usage is to cache variables:

@micropython.viper
def foo(self, arg: int) -> int:
    buf = ptr8(self.linebuf) # self.linebuf is a bytearray or bytes object
    for x in range(20, 30):
        bar = buf[x] # Access a data item through the pointer
        # code omitted

In this instance the compiler “knows” that buf is the address of an array of bytes; it can emit code to rapidly compute the address of buf[x] at runtime. Where casts are used to convert objects to Viper native types these should be performed at the start of the function rather than in critical timing loops as the cast operation can take several microseconds. The rules for casting are as follows:

  • Casting operators are currently: int, bool, uint, ptr, ptr8, ptr16 and ptr32.
  • The result of a cast will be a native Viper variable.
  • Arguments to a cast can be a Python object or a native Viper variable.
  • If argument is a native Viper variable, then cast is a no-op (i.e. costs nothing at runtime) that just changes the type (e.g. from uint to ptr8) so that you can then store/load using this pointer.
  • If the argument is a Python object and the cast is int or uint, then the Python object must be of integral type and the value of that integral object is returned.
  • The argument to a bool cast must be integral type (boolean or integer); when used as a return type the viper function will return True or False objects.
  • If the argument is a Python object and the cast is ptr, ptr, ptr16 or ptr32, then the Python object must either have the buffer protocol with read-write capabilities (in which case a pointer to the start of the buffer is returned) or it must be of integral type (in which case the value of that integral object is returned).

The following example illustrates the use of a ptr16 cast to toggle pin X1 on the Pyboard n times:

BIT0 = const(1)
@micropython.viper
def toggle_n(n: int):
    odr = ptr16(stm.GPIOA + stm.GPIO_ODR)
    for _ in range(n):
        odr[0] ^= BIT0

A detailed technical description of the three code emitters may be found on Kickstarter here Note 1 and here Note 2