prometeo

This is prometeo, an experimental modeling tool for embedded high-performance computing. prometeo provides a domain specific language (DSL) based on a subset of the Python language that allows one to conveniently write scientific computing programs in a high-level language (Python itself) that can be transpiled to high-performance self-contained C code easily deployable on embedded devices.

features

• Python compatible syntax : prometeo is a DSL embedded into the Python language. prometeo programs can be executed from the Python interpreter.
• efficient : prometeo programs transpile to high-performance C code.
• statically typed : prometeo uses Python's native type hints to strictly enforce static typing.
• deterministic memory usage : a specific program structure is required and enforced through static analysis. In this way prometeo transpiled programs have a guaranteed maximum heap usage.
• fast memory management : thanks to its static analysis, prometeo can avoid allocating and garbage-collecting memory, resulting in faster and safer execution.
• self-contained and embeddable : unlike other similar tools and languages, prometeo targets specifically embedded applications and programs written in prometeo transpile to self-contained C code that does not require linking against the Python run-time library.

documentation (preliminary)

prometeo's documentation can be found on Read the Docs at https://prometeo.readthedocs.io/en/latest/index.html.

hello world!

A simple hello world example that shows how to either run a trivial prometeo program from Python or transpile it to C, build it and run it can be found here. The output shows the outcome of the heap usage analysis and the execution time (in this case there is not much to see :p).

performance

Since prometeo programs transpile to pure C code that calls the high performance linear algebra library BLASFEO (publication: https://arxiv.org/abs/1704.02457, code: https://github.com/giaf/blasfeo), execution time can be comparable to hand-written high-performance code. The figure below shows a comparison of the CPU time necessary to carry out a Riccati factorization using highly optimized hand-written C code with calls to BLASFEO and the ones obtained with prometeo transpiled code from this example. The computation times obtained with NumPy and Julia are added too for comparison - notice however that these last two implementations of the Riccati factorization are not as easily embeddable as the C code generated by prometeo and the hand-coded C implementation. All the benchmarks have been run on a Dell XPS-9360 equipped with an i7-7560U CPU running at 2.30 GHz (to avoid frequency fluctuations due to thermal throttling).

Moreover, prometeo can largely outperform state-of-the-art Python compilers such as Nuitka. The table below shows the CPU times obtained on a Fibonacci benchmark.

parser/compiler	CPU time [s]
Python 3.7 (CPython)	11.787
Nuitka	10.039
PyPy	1.78
prometeo	0.657

PyPI installation

prometeo can be installed through PyPI with pip install prometeo-dsl. Notice that, since prometeo makes extensive use of type hints to equip Python code with static typing information, the minimum Python version required is 3.6.

manual installation

If you want to install prometeo building the sources on your local machine you can proceed as follows:

• Run git submodule update --init to clone the submodules.
• Run make install_shared from <prometeo_root>/prometeo/cpmt to compile and install the shared library associated with the C backend. Notice that the default installation path is <prometeo_root>/prometeo/cpmt/install.
• You need Python 3.6. or later.
• Optional: to keep things clean you can setup a virtual environment with virtualenv --python=<path_to_python3.6> <path_to_new_virtualenv>.
• Run pip install -e . from <prometeo_root> to install the Python package.

Finally, you can run the examples in <root>/examples with pmt <example_name>.py --cgen=<True/False>, where the --cgen flag determines whether the code is executed by the Python interpreter or C code is generated compiled and run.

a simple example

The Python code (examples/simple_example/simple_example.py)

from prometeo import *

n : dims = 10

def main() -> int:

    A: pmat = pmat(n, n)
    for i in range(10):
        for j in range(10):
            A[i, j] = 1.0

    B: pmat = pmat(n, n)
    for i in range(10):
        B[0, i] = 2.0

    C: pmat = pmat(n, n)
    C = A * B
    pmat_print(C)
    return 0

can be run by the standard Python interpreter (version >3.6 required) and it will perform the described linear algebra operations using the command pmt simple_example.py --cgen=False. At the same time, the code can be parsed by prometeo and its abstract syntax tree (AST) analyzed in order to generate the following high-performance C code:

#include "stdlib.h"
#include "simple_example.h"
void * ___c_pmt_8_heap;
void * ___c_pmt_64_heap;
void * ___c_pmt_8_heap_head;
void * ___c_pmt_64_heap_head;

#include "prometeo.h"
int main() {
    ___c_pmt_8_heap = malloc(10000); 
    ___c_pmt_8_heap_head = ___c_pmt_8_heap;
    char * pmem_ptr = (char *)___c_pmt_8_heap;
    align_char_to(8, &pmem_ptr);
    ___c_pmt_8_heap = pmem_ptr;
    ___c_pmt_64_heap = malloc(1000000);
    ___c_pmt_64_heap_head = ___c_pmt_64_heap;
    pmem_ptr = (char *)___c_pmt_64_heap;
    align_char_to(64, &pmem_ptr);
    ___c_pmt_64_heap = pmem_ptr;
	void *callee_pmt_8_heap = ___c_pmt_8_heap;
	void *callee_pmt_64_heap = ___c_pmt_64_heap;

    struct pmat * A = c_pmt_create_pmat(n, n);
    for(int i = 0; i < 10; i++) {
        for(int j = 0; j < 10; j++) {
            c_pmt_pmat_set_el(A, i, j, 1.0);
    }

    }

    struct pmat * B = c_pmt_create_pmat(n, n);
    for(int i = 0; i < 10; i++) {
        c_pmt_pmat_set_el(B, 0, i, 2.0);
    }

    struct pmat * C = c_pmt_create_pmat(n, n);
    c_pmt_pmat_fill(C, 0.0);
    c_pmt_gemm_nn(A, B, C, C);
    c_pmt_pmat_print(C);
	___c_pmt_8_heap = callee_pmt_8_heap;
	___c_pmt_64_heap = callee_pmt_64_heap;

	free(___c_pmt_8_heap_head);
	free(___c_pmt_64_heap_head);
	return 0;
}

which relies on the high-performance linear algebra package BLASFEO. The generated code will be readily compiled and run with when running pmt simple_example.py --cgen=True.

how does it work?

Although translating a program written in a language into another with a comparable level of abstraction can be significantly easier than translating to one with a very different level of abstraction (especially if the target language is of much lower level), translating Python programs into C programs still involves a considerable abstraction gap it is not an easy task in general. Loosely speaking, the challenge lies in the necessity to reimplement features that are natively supported by the source language in the target language. In particular, when translating Python to C, the difficulty comes both from the different level of abstraction of the two languages and from the fact that the source and target language are of two very different types: Python is an interpreted, duck-typed and garbage-collected language and C is a compiled and statically typed language.

The task of transpiling Python to C becomes even more challenging if we add the constraint that the generated C code must be efficient (even for small to medium scale computations) and deployable on embedded hardware. In fact these two requirements directly imply that the generated code cannot make use of: i) sophisticated runtime libraries, e.g., the Python runtime library, which are generally not available on embedded hardware ii) dynamic memory allocation that would make the execution slow and unreliable (exception made for memory that is allocated in a setup phase and whose size is known a priori).

Since source-to-source code transformation, or transpilation, and in particular transpilation of Python code into C code is not an unexplored realm, in the following, we mention a few existing projects that address it. In doing so, we highlight where and how they do not satisfy one of the two requirements outlined above, namely (small scale) efficiency and embeddability.

related work

Several software packages exist that address Python-to-C translation in various forms.

In the context of high-performance computing, Numba is a just-in-time compiler for numerical functions written in Python. As such, its aim is to convert properly annotated Python functions, not entire programs, into high-performance LLVM code such that their execution can be sped up. Numba uses an internal representation of the code to be translated and performs a (potentially partial) type inference on the variables involved in order to generate LLVM code that can be called either from Python or from C/C++. In some cases, namely the ones where a complete type inference can be carried out successfully, code that does not rely on the C API can be generated (using the nopython flag). However, the emitted LLVM code would still rely on Numpy for BLAS and LAPACK operations.

Nuitka is a source-to-source compiler that can translate every Python construct into C code that links against thelibpython library and it is therefore able to transpile a large class of Python programs. In order to do so, it relies on the fact that one of the most used implementations of the Python language, namely CPython, is written in C. In fact, Nuitka generates C code that contains calls to CPython that would normally be carried out by the Python parser. Despite its attractive and general transpilation approach, it cannot be easily deployed on embedded hardware due to its intrinsic dependency on libpython. At the same time, since it maps rather closely Python constructs to their CPython implementation, a number of performance issues can be expected when it comes to small to medium scale high-performance computing. This is particularly due to the fact that operations associated with, for example, type checking, memory allocation and garbage collection that can slow down the execution are carried out by the transpiled program too.

Cython is a programming language whose goal is to facilitate writing C extensions for the Python language. In particular, it can translate (optionally) statically typed Python-like code into C code that relies on CPython. Similarly to the considerations made for *Nu