Sped up covariance estimate via numba

[wolph]

Benchmarked using: py.test arch/tests/unitroot/test_fmols_ccr.py

Before:

============================= fixture duration top ============================== 
total          name              num   avg            min
0:00:00.080000             trend  1728        0:00:00 0:00:00
0:00:00.062000              diff  1728        0:00:00 0:00:00
0:00:00.032000            kernel  1728        0:00:00 0:00:00
0:00:00.031000         force_int  1728        0:00:00 0:00:00
0:00:00.016000   trivariate_data     2 0:00:00.008000 0:00:00
0:00:00.016000           x_trend  1728        0:00:00 0:00:00
0:00:00.016000         bandwidth  1728        0:00:00 0:00:00
0:00:00.253000       grand total 10426        0:00:00 0:00:00
============================ test call duration top ============================= 
total          name              num   avg            min
0:00:06.062000  test_fmols_smoke   864        0:00:00 0:00:00
0:00:06.001000    test_ccr_smoke   864        0:00:00 0:00:00
0:00:00.093000   test_ccr_eviews    26        0:00:00 0:00:00
0:00:00.078000 test_fmols_eviews    26        0:00:00 0:00:00
0:00:12.234000       grand total  1784        0:00:00 0:00:00
============================ test setup duration top ============================ 
total          name              num   avg            min
0:00:00.377000  test_fmols_smoke   864        0:00:00 0:00:00
0:00:00.295000    test_ccr_smoke   864        0:00:00 0:00:00
0:00:00.672000       grand total  1784        0:00:00 0:00:00
========================== test teardown duration top =========================== 
total          name              num   avg            min
0:00:00.142000  test_fmols_smoke   864        0:00:00 0:00:00
0:00:00.124000    test_ccr_smoke   864        0:00:00 0:00:00
0:00:00.266000       grand total  1784        0:00:00 0:00:00
============================= 1784 passed in 14.37s ============================= 

After:

============================= fixture duration top ============================== 
total          name              num   avg            min
0:00:00.063000             trend  1728        0:00:00 0:00:00
0:00:00.048000           x_trend  1728        0:00:00 0:00:00
0:00:00.032000         force_int  1728        0:00:00 0:00:00
0:00:00.016000   trivariate_data     2 0:00:00.008000 0:00:00
0:00:00.016000            kernel  1728        0:00:00 0:00:00
0:00:00.015000         bandwidth  1728        0:00:00 0:00:00
0:00:00.190000       grand total 10426        0:00:00 0:00:00
============================ test call duration top ============================= 
total          name              num   avg            min
0:00:04.528000  test_fmols_smoke   864        0:00:00 0:00:00
0:00:03.748000    test_ccr_smoke   864        0:00:00 0:00:00
0:00:00.094000 test_fmols_eviews    26        0:00:00 0:00:00
0:00:00.094000   test_ccr_eviews    26        0:00:00 0:00:00
0:00:08.464000       grand total  1784        0:00:00 0:00:00
============================ test setup duration top ============================ 
total          name              num   avg            min
0:00:00.392000    test_ccr_smoke   864        0:00:00 0:00:00
0:00:00.329000  test_fmols_smoke   864        0:00:00 0:00:00
0:00:00.721000       grand total  1784        0:00:00 0:00:00
========================== test teardown duration top =========================== 
total          name              num   avg            min
0:00:00.109000  test_fmols_smoke   864        0:00:00 0:00:00
0:00:00.064000    test_ccr_smoke   864        0:00:00 0:00:00
0:00:00.016000   test_ccr_eviews    26        0:00:00 0:00:00
0:00:00.189000       grand total  1784        0:00:00 0:00:00
============================= 1784 passed in 10.34s ============================= 

Nothing too dramatic but a useful performance improvement :)

I'm still looking at other bottlenecks right now so depending on your preference I can create separate pull requests or a single large one :)

[codecov]

Codecov Report

All modified lines are covered by tests ✅

Files	Coverage Δ
arch/covariance/kernel.py	100.00% <100.00%> (ø)

📢 Thoughts on this report? Let us know!.

[bashtage]

Looks good. Can you run isort for fix imports. ?

[wolph]

I actually ran isort, black and flake8 and none of those show issues with my code.

I'm guessing there's been a black update this is the diff that black is producing now:

$ git diff
diff --git a/arch/univariate/base.py b/arch/univariate/base.py
index 4847000f8..f59fd7c1c 100644
--- a/arch/univariate/base.py
+++ b/arch/univariate/base.py
@@ -1189,10 +1189,7 @@ class ARCHModelFixedResult(_SummaryRepr):

         stubs = list(self._names)
         header = ["coef"]
-        param_table_data = [
-            [format_float_fixed(param, 10, 4)]
-            for param in self.params
-        ]
+        param_table_data = [[format_float_fixed(param, 10, 4)] for param in self.params]

         mc = self.model.num_params
         vc = self.model.volatility.num_params

I can apply it if you want to, but it's not really part of this PR :)

The other issue is (was, I just pushed an update) that the signature for the jit function is incorrect. It doesn't allow usage as a decorator with arguments

[wolph]

Disregard part of that last comment... I was looking at the wrong branch ;)

The diff above still applies though

[bashtage]

To get the decorator to work, I thnk you can use

try:
    if DISABLE_NUMBA:
        raise ImportError

    from numba import jit

    jit = functools.partial(jit, nopython=True)

except ImportError:

    def jit(
        *args: Any,
        **kwargs: Any,
    ) -> Any:
        if args and callable(args[0]):
            func = args[0]

        def wrap(func):
            def wrapper(*args: Any, **kwargs: Any) -> Callable[..., Any]:
                import warnings

                warnings.warn(performance_warning, PerformanceWarning)
                return func(*args, **kwargs)

            return wrapper

        return wrap

[bashtage]

I pushed an improved version of the fix.

[bashtage]

Thanks. Merged these in after some surgery on the compact layer.

[wolph]

Thanks for the quick merge and the awesome project!