libocca/occa.py

OCCA Python API: JIT Compilation for Multiple Architectures

occa.py's Issues

Redirect std::cout and std::cerr

Jupyter doesn't redirect IO messages coming directly from the C++ API. We need to somehow wrap std::cout and std::err usage in occa to allow for redirects.

We would take wanted outputs and wrap them in some C API methods

• std::cout would use PySys_WriteStdout

• occa::warn and occa::error could use PyErr_WarnEx for that Jupyter pretty print

Pretty print errors given an ast node

Use the node's col_offset and lineno attributes to print a nice error message for the user

myfile.py:10:30:Error Expected a 1 and not a 0
1234567890123456789012345678901234567890
                             ^

or if in Jupyter, get the cell number and print

In [31]:10:30:Error Expected a 1 and not a 0
1234567890123456789012345678901234567890
                             ^

We might be able to get the cell number with inspect

import inspect
inspect.getfile(foo)
> '<ipython-input-31-5643a48472c0>'

@okl.kernel

Python OKL

Create ast parser to convert a restricted Python function to an OKL kernel

@okl.kernel
def add_vectors(a: [float],
                b: [float],
                ab: [float]):
    for i in okl.range(len(a)).tile(16):
      ab[i] = a[i] + b[i]

shorthand for

@okl.kernel
def add_vectors(a: [float],
                b: [float],
                ab: [float]):
    for block in okl.range(0, len(a), 16).outer:
        for i in okl.range(block, block + 16).inner:
            ab[i] = a[i] + b[i]

API

Have props be a restricted kwarg

add_vectors[device](a, b, ab, props={})
# or
add_vectors(a, b, ab, props={})

Docstrings

• Add docstrings to everything!
• Add to README that XCode needs to be installed