# Numba performance doesn't scale as well as NumPy in vectorized max function

**URL:** https://numba.discourse.group/t/numba-performance-doesnt-scale-as-well-as-numpy-in-vectorized-max-function/782
**Category:** Community Support
**Created:** [June 30, 2021, 9:32pm UTC](https://numba.discourse.group/t/numba-performance-doesnt-scale-as-well-as-numpy-in-vectorized-max-function/782 "2021-06-30T21:32:34Z")
**Posts on this page:** 1
**Showing post:** 9

<div class="post-metadata">

### Author: ![stuartarchibald](https://yyz2.discourse-cdn.com/free1/user_avatar/numba.discourse.group/stuartarchibald/32/10_2.png) [@stuartarchibald](https://numba.discourse.group/u/stuartarchibald)
#### Post date: [July 16, 2021, 10:26am UTC](https://numba.discourse.group/t/numba-performance-doesnt-scale-as-well-as-numpy-in-vectorized-max-function/782/9 "2021-07-16T10:26:22Z")

</div>

@brandonwillard, I’m thinking about the way loops are represented in Numba in general and how the `ufunc` loop nests (as referenced) are quite involved. The control flow graphs for the LLVM IR can be seen with:

```python
import numpy as np
import numba

@numba.vectorize(["float64(float64, float64)"], identity="reorderable")
def custom_op_fn(x, y):
    if x > y:
        return x
    else:
        return y

@numba.njit(debug=True)
def max_reduce_axis_1(x):
    res = np.full((x.shape[0],), -np.inf, dtype=x.dtype)
    x_transpose = np.transpose(x)
    for m in range(x.shape[1]):
        custom_op_fn(res, x_transpose[m], res)
    return res

X = np.random.normal(size=(5000, 5000))
res_1 = max_reduce_axis_1(X)

@numba.njit(debug=True, error_model='numpy')
def max_reduce_axis_2(x):
    res = np.empty((x.shape[0],), dtype=x.dtype)
    for i in range(x.shape[0]):
        res[i] = -np.inf
        for j in range(x.shape[1]):
            tmp = x[i, j]
            if res[i] < tmp:
                res[i] = tmp
    return res

res_2 = max_reduce_axis_2(X)
assert np.array_equal(res_1, res_2)
assert np.array_equal(res_1, np.max(X, axis=1))

def show_llvm_cfg(func):
    func.inspect_cfg(func.signatures[0]).pretty_printer(interleave=True, view=True)

show_llvm_cfg(max_reduce_axis_1)
show_llvm_cfg(max_reduce_axis_2)

```

---

_[View the full topic](https://numba.discourse.group/t/numba-performance-doesnt-scale-as-well-as-numpy-in-vectorized-max-function/782)._