np.clip(a, mn, mx) on Windows runs slower than np.clip(np.clip(a, None, mx), mn, None))

On Windows platform, np.clip() on floating point inputs with both bounds provided runs significantly slower than calling np.clip() twice, once with the lower bound, and again with the upper bound.

Reproducing code example:

from time import perf_counter
import numpy as np

def myclip(data, vmin, vmax, out=None):
    out = np.clip(data, None, vmax, out=out)
    return np.clip(out, vmin, None, out=out)

data = np.random.random((1000,1000))
out = np.ones_like(data)
vmin, vmax = 0.1, 0.9

assert np.all(np.clip(data, vmin, vmax) == myclip(data, vmin, vmax))

t0 = perf_counter()
for _ in range(10):
    np.clip(data, vmin, vmax, out=out)
t1 = perf_counter()
print(t1 - t0)

t0 = perf_counter()
for _ in range(10):
    myclip(data, vmin, vmax, out=out)
t1 = perf_counter()
print(t1 - t0)

Error message:

The timing outputs where np.clip is significantly slower than myclip:
0.17425939999999995
0.06601459999999992

NumPy/Python version information:

1.20.1 3.8.8 (tags/v3.8.8:024d805, Feb 19 2021, 13:18:16) [MSC v.1928 64 bit (AMD64)]

I can't replicate this behavior, albeit on linux.

In [9]: %timeit np.clip(data, vmin, vmax, out=out)
1.33 ms ± 4.83 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

In [10]: %timeit myclip(data, vmin, vmax, out=out)
3.06 ms ± 30.9 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Maybe I should have specified more clearly that the issue is only for Windows. Yes, on Linux, np.clip is faster, as would be expected.

No you were plenty clear 👍 , I was just surprised by the platform dependence.

how about if we use joblib here ?

I noticed that the use of _NPY_CLIP macro would result in multiple re-evaluations of the input arguments.
By avoiding the use of the _NPY_CLIP macro and using a temporary variable, I was able to get np.clip to run faster than myclip in the reproducing example script above. (Windows 10 x64, MSVC 19.28)

file numpy\core\src\umath\clip.c.src

NPY_NO_EXPORT void
@name@_clip(char **args, npy_intp const *dimensions, npy_intp const *steps, void *NPY_UNUSED(func))
{
    if (steps[1] == 0 && steps[2] == 0) {
        /* min and max are constant throughout the loop, the most common case */
        /* NOTE: it may be possible to optimize these checks for nan */
        @type@ min_val = *(@type@ *)args[1];
        @type@ max_val = *(@type@ *)args[2];

        char *ip1 = args[0], *op1 = args[3];
        npy_intp is1 = steps[0], os1 = steps[3];
        npy_intp n = dimensions[0];

        /* contiguous, branch to let the compiler optimize */
        if (is1 == sizeof(@type@) && os1 == sizeof(@type@)) {
            for(npy_intp i = 0; i < n; i++, ip1 += is1, op1 += os1) {
                // *(@type@ *)op1 = _NPY_CLIP(*(@type@ *)ip1, min_val, max_val);
                @type@ t = *(@type@ *)ip1;
                t = _NPY_@name@_MAX(t, min_val);
                t = _NPY_@name@_MIN(t, max_val);
                *(@type@ *)op1 = t;
            }
        }
        else {
            for(npy_intp i = 0; i < n; i++, ip1 += is1, op1 += os1) {
                // *(@type@ *)op1 = _NPY_CLIP(*(@type@ *)ip1, min_val, max_val);
                @type@ t = *(@type@ *)ip1;
                t = _NPY_@name@_MAX(t, min_val);
                t = _NPY_@name@_MIN(t, max_val);
                *(@type@ *)op1 = t;
            }
        }
    }
    else {
        TERNARY_LOOP {
            *(@type@ *)op1 = _NPY_CLIP(*(@type@ *)ip1, *(@type@ *)ip2, *(@type@ *)ip3);
        }
    }
    npy_clear_floatstatus_barrier((char*)dimensions);
}

BUG: fix win32 np.clip slowness

fixed by #20134

Navigation Menu

Repository navigation

np.clip(a, mn, mx) on Windows runs slower than np.clip(np.clip(a, None, mx), mn, None)) #18673

Reproducing code example:

Error message:

NumPy/Python version information:

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

np.clip(a, mn, mx) on Windows runs slower than np.clip(np.clip(a, None, mx), mn, None)) #18673

Description

Reproducing code example:

Error message:

NumPy/Python version information:

Activity

rossbar commented on Mar 24, 2021

pijyoi commented on Mar 24, 2021

rossbar commented on Mar 24, 2021

rajpratyush commented on Apr 14, 2021

pijyoi commented on Oct 19, 2021

pijyoi commented on Oct 22, 2021

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions