Open Table Of Contents

Annotating Numba SourceΒΆ

Numba comes with a script to annotate numba source code. The tool can dump generated code or types to the terminal or generate a webpage with annotated code.

$ numba -h
usage: numba [-h] [--annotate] [--dump-llvm] [--dump-optimized] [--dump-cfg]
             [--dump-ast] [--fancy]

positional arguments:
  filename          Python source filename

optional arguments:
  -h, --help        show this help message and exit
  --annotate        Annotate source
  --dump-llvm       Print generated llvm assembly
  --dump-optimized  Dump the optimized llvm assembly
  --dump-cfg        Dump the control flow graph
  --dump-ast        Dump the AST
  --fancy           Try to output fancy files (.dot or .html)

For instance, it’s easy to dump the generated llvm code to the terminal.

$ numba --dump-llvm

define double @__numba_specialized_0_myfunc(double %y) {
  %return_value = alloca double
  %0 = fmul double %y, 2.000000e+00
  store double %0, double* %return_value
  br label %cleanup_label

cleanup_label:                                    ; preds = %entry, %error_label
  %1 = load double* %return_value
  ret double %1

error_label:                                      ; No predecessors!
  store double 0x7FF8000000000000, double* %return_value
  br label %cleanup_label

$ numba --dump-optimized

define double @__numba_specialized_0_myfunc(double %y) nounwind readnone {
  %0 = fmul double %y, 2.000000e+00
  ret double %0

To get a clear idea of what the types of variables are and where parts of the code may be slow (rely on Python), use the --annotate option. This can also be used from the code itself:

@jit(double(double), annotate=True)
def square(x):
    return x * x

When the –annotate and –fancy flags are used together, the output is saved in html format. The html file name is the same as the Python script name, with the .py extension replaced with a .html extension. When viewed in a web browser, the Python source code of each jitted function is displayed along with the function call and the file name and line number from where the function was called. Moving your mouse cursor over a line of Python code displays a box showing each Python variable and it’s type assigned by numba. Clicking on a line of python displays the LLVM IR that is generated by that line. Any line of python that results in LLVM IR containing a call back into a Python function is indicated by an asterick next to the line number.