Compiling With Intel® SDC

What if I get a compilation error

There are a few reasons why Intel SDC cannot compile your code out-of-the-box.

  1. Intel SDC does support only a subset of Pandas* APIs.

  2. Intel SDC and Numba* can compile only a subset of Python data types.

  3. Intel SDC cannot infer the type of a variable at compile time.

Unsupported APIs

Intel® SDC is able to compile variety of the most typical workflows that involve Pandas* operations but not all. Sometimes it means that your code cannot be compiled out-of-the-box:

import numba
import pandas

def read_df(filename):
    return pandas.read_excel(filename)



Traceback (most recent call last):
numba.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
Unknown attribute 'read_excel' of type Module(<module 'pandas' from ...)

In such case you have the following options:

import numba
import pandas

def read_df():
    return pandas.read_csv("data.csv")

import numba
import pandas

def cummax():
    s = pandas.Series([0, 1, 0, 2, 0, 3, 0, 4])

    with numba.objmode(r='intp[:]'):
        r = s.cummax().values

    return pandas.Series(r)

Please note, that an array is returned from objmode. Returning Series or DataFrame from objmode is not a trivial task.

  • Exclude such calls from jit region:

import numba
import pandas

def cummax():
    def create_series():
        return pandas.Series([0, 1, 0, 2, 0, 3, 0, 4])

    s = create_series()

    return s.cummax()

Please note that last two options would result in performing boxing/unboxing which could significantly affect performance.

For more details on performance see Getting Performance With Intel® SDC

For list of supported functions see API Reference

Unsupported Data Types

The other common reason why Intel® SDC or Numba* cannot compile the code is because it does not support a certain data type. e.g. Numba* doesn’t support heterogeneous lists and dicts:

a = [0, 2, 5, "a", "b"]

Literal heterogeneous lists usually could be replaced with tuples:

a = (0, 2, 5, "a", "b")

While heterogeneous dicts are not supported, it could be passed as parameter to pandas.DataFrame or pandas.read_csv:

data = {'A': np.ranf(10), 'B': np.ones(10)}
df = pandas.DataFrame(data=data)

Intel® SDC supports pandas.Series only of boolean, integer, float and string types. Other types like Series of datetime or categorical are not supported.

Type Inference And Type Stability

The last but certainly not the least why Intel® SDC cannot compile your code is because it cannot infer the type at the time of compilation. The most frequent cause for that is the type instability.

The static compilation is a powerful technology to obtain high efficiency of a code but the flip side is the compiler should be able to infer all variable types at the time of compilation and these types remain stable within the region being compiled.

The following is an example of the type-unstable variable a, and hence this code cannot be compiled by Numba*

if flag:
   a = 1.0
   a = np.ones(10)

The use of isinstance() function often means type instability and is not supported. Similarly, function calls should also be deterministic. The below example is not supported since the function f() is not known in advance:

if flag:
    f = np.zeros
    f = np.random.ranf
a = f(10)

Dealing With Integer NaN Values

The pandas.Series are built upon numpy.ndarray, which does not support NaN values for integers and booleans. For that reason Pandas* dynamically converts integer columns to floating point ones when NaN values are needed. Intel SDC doesn’t perform such conversion and it is user responsibility to manually convert from integer data to floating point data.

Type Inference In I/O Operations

If the filename is constant, the Intel SDC may be able to determine file schema at compilation time. It will allow to perform type inference of columns in respective Pandas* dataframe.

df = pandas.read_csv("data.csv")

If Intel SDC fails to infer types from the file, the schema must be manually specified.

names = ['A', 'B']
usecols = ['A']
dtypes={'A': np.float64}
pd.read_csv(file_name, names=names, usecols=usecols, dtype=dtypes)

Alternatively you can take file reading out of the compiled region.

Note: if data file contains integer data with empty positions (Nans) it is highly recommended to manually specify column type to float.