Fix Numpy 2.0 related test failures

[ksunden]

PR summary

Closes #27645 (technically combination of this and #27624, but the comments I've left on there have all been about this part)

Numpy made tweaks to dtype promotions that affected some computation (but only at the limits of floating point precision)

This PR counter acts these:

• pie image tests have a tolerance introduced
  • could perhaps tweak some of the exact values down a bit, but none were more than (0.01 RMS)
  • All differences were imperceptable as a human, and some tolerance would be required regardless, so I did not replace the images
  • Technically could be changed by setting the env var NPY_PROMOTION_state=legacy, but that is changing intendid numpy 2.0 behavior
• pylab updated to ensure builtins are used for two more functions now included in numpy's namespace that occlude builtints
  • pow and bool
• Assertions in test_scalarmappable_to_rgba changed to np.testing.assert_almost_equal
  • Since some values are close to 0, some of the assert_almost_equal_nulp and similar options proved insufficient.
  • Could use assert_allclose instead, but would have to specify tolerances.

PR checklist

• "closes #0000" is in the body of the PR description to link the related issue
• new and changed code is tested
• Plotting related features are demonstrated in an example
• New Features and API Changes are noted with a directive and release note
• Documentation complies with general and docstring guidelines

[tacaswell]

The second xref is to the pytest8 PR which I do not understand how it is related.

[tacaswell]

oh, I understand, this + #27624 close #27645

[QuLogic]

Running with 1.24.1 and NPY_PROMOTION_STATE=weak_and_warn, I think we may be introducing an unintentional copy here?

__________ test_norm_update_figs[png] __________

ext = 'png', request = <FixtureRequest for <Function test_norm_update_figs[png]>>, args = (), kwargs = {}, file_name = 'test_norm_update_figs[png]'
fig_test = <Figure size 640x480 with 1 Axes>, fig_ref = <Figure size 640x480 with 1 Axes>, figs = []

    @pytest.mark.parametrize("ext", extensions)
    def wrapper(*args, ext, request, **kwargs):
        if 'ext' in old_sig.parameters:
            kwargs['ext'] = ext
        if 'request' in old_sig.parameters:
            kwargs['request'] = request

        file_name = "".join(c for c in request.node.name
                            if c in ALLOWED_CHARS)
        try:
            fig_test = plt.figure("test")
            fig_ref = plt.figure("reference")
            with _collect_new_figures() as figs:
>               func(*args, fig_test=fig_test, fig_ref=fig_ref, **kwargs)

lib/matplotlib/testing/decorators.py:411:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
lib/matplotlib/tests/test_colors.py:1661: in test_norm_update_figs
    fig_test.canvas.draw()
lib/matplotlib/backends/backend_agg.py:387: in draw
    self.figure.draw(self.renderer)
lib/matplotlib/artist.py:95: in draw_wrapper
    result = draw(artist, renderer, *args, **kwargs)
lib/matplotlib/artist.py:72: in draw_wrapper
    return draw(artist, renderer)
lib/matplotlib/figure.py:3117: in draw
    mimage._draw_list_compositing_images(
lib/matplotlib/image.py:132: in _draw_list_compositing_images
    a.draw(renderer)
lib/matplotlib/artist.py:72: in draw_wrapper
    return draw(artist, renderer)
lib/matplotlib/axes/_base.py:3095: in draw
    mimage._draw_list_compositing_images(
lib/matplotlib/image.py:132: in _draw_list_compositing_images
    a.draw(renderer)
lib/matplotlib/artist.py:72: in draw_wrapper
    return draw(artist, renderer)
lib/matplotlib/image.py:653: in draw
    im, l, b, trans = self.make_image(
lib/matplotlib/image.py:945: in make_image
    return self._make_image(self._A, bbox, transformed_bbox, clip,
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = <matplotlib.image.AxesImage object at 0x7f02f6900c40>
A = masked_array(
  data=[[ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9],
        [10, 11, 12, 13, 14, 15, 16, 17, 18, 19],
    ... 82, 83, 84, 85, 86, 87, 88, 89],
        [90, 91, 92, 93, 94, 95, 96, 97, 98, 99]],
  mask=False,
  fill_value=999999)
in_bbox = Bbox([[-0.5, 9.5], [9.5, -0.5]]), out_bbox = <matplotlib.transforms.TransformedBbox object at 0x7f02f68f6370>
clip_bbox = <matplotlib.transforms.TransformedBbox object at 0x7f02f6900e80>, magnification = 1.0, unsampled = False, round_to_pixel_border = True

    def _make_image(self, A, in_bbox, out_bbox, clip_bbox, magnification=1.0,
                    unsampled=False, round_to_pixel_border=True):
        """
        Normalize, rescale, and colormap the image *A* from the given *in_bbox*
        (in data space), to the given *out_bbox* (in pixel space) clipped to
        the given *clip_bbox* (also in pixel space), and magnified by the
        *magnification* factor.

        *A* may be a greyscale image (M, N) with a dtype of `~numpy.float32`,
        `~numpy.float64`, `~numpy.float128`, `~numpy.uint16` or `~numpy.uint8`,
        or an (M, N, 4) RGBA image with a dtype of `~numpy.float32`,
        `~numpy.float64`, `~numpy.float128`, or `~numpy.uint8`.

        If *unsampled* is True, the image will not be scaled, but an
        appropriate affine transformation will be returned instead.

        If *round_to_pixel_border* is True, the output image size will be
        rounded to the nearest pixel boundary.  This makes the images align
        correctly with the Axes.  It should not be used if exact scaling is
        needed, such as for `FigureImage`.

        Returns
        -------
        image : (M, N, 4) `numpy.uint8` array
            The RGBA image, resampled unless *unsampled* is True.
        x, y : float
            The upper left corner where the image should be drawn, in pixel
            space.
        trans : `~matplotlib.transforms.Affine2D`
            The affine transformation from image to pixel space.
        """
        if A is None:
            raise RuntimeError('You must first set the image '
                               'array or the image attribute')
        if A.size == 0:
            raise RuntimeError("_make_image must get a non-empty image. "
                               "Your Artist's draw method must filter before "
                               "this method is called.")

        clipped_bbox = Bbox.intersection(out_bbox, clip_bbox)

        if clipped_bbox is None:
            return None, 0, 0, None

        out_width_base = clipped_bbox.width * magnification
        out_height_base = clipped_bbox.height * magnification

        if out_width_base == 0 or out_height_base == 0:
            return None, 0, 0, None

        if self.origin == 'upper':
            # Flip the input image using a transform.  This avoids the
            # problem with flipping the array, which results in a copy
            # when it is converted to contiguous in the C wrapper
            t0 = Affine2D().translate(0, -A.shape[0]).scale(1, -1)
        else:
            t0 = IdentityTransform()

        t0 += (
            Affine2D()
            .scale(
                in_bbox.width / A.shape[1],
                in_bbox.height / A.shape[0])
            .translate(in_bbox.x0, in_bbox.y0)
            + self.get_transform())

        t = (t0
             + (Affine2D()
                .translate(-clipped_bbox.x0, -clipped_bbox.y0)
                .scale(magnification)))

        # So that the image is aligned with the edge of the Axes, we want to
        # round up the output width to the next integer.  This also means
        # scaling the transform slightly to account for the extra subpixel.
        if ((not unsampled) and t.is_affine and round_to_pixel_border and
                (out_width_base % 1.0 != 0.0 or out_height_base % 1.0 != 0.0)):
            out_width = math.ceil(out_width_base)
            out_height = math.ceil(out_height_base)
            extra_width = (out_width - out_width_base) / out_width_base
            extra_height = (out_height - out_height_base) / out_height_base
            t += Affine2D().scale(1.0 + extra_width, 1.0 + extra_height)
        else:
            out_width = int(out_width_base)
            out_height = int(out_height_base)
        out_shape = (out_height, out_width)

        if not unsampled:
            if not (A.ndim == 2 or A.ndim == 3 and A.shape[-1] in (3, 4)):
                raise ValueError(f"Invalid shape {A.shape} for image data")
            if A.ndim == 2 and self._interpolation_stage != 'rgba':
                # if we are a 2D array, then we are running through the
                # norm + colormap transformation.  However, in general the
                # input data is not going to match the size on the screen so we
                # have to resample to the correct number of pixels

                # TODO slice input array first
                a_min = A.min()
                a_max = A.max()
                if a_min is np.ma.masked:  # All masked; values don't matter.
                    a_min, a_max = np.int32(0), np.int32(1)
                if A.dtype.kind == 'f':  # Float dtype: scale to same dtype.
                    scaled_dtype = np.dtype(
                        np.float64 if A.dtype.itemsize > 4 else np.float32)
                    if scaled_dtype.itemsize < A.dtype.itemsize:
                        _api.warn_external(f"Casting input data from {A.dtype}"
                                           f" to {scaled_dtype} for imshow.")
                else:  # Int dtype, likely.
                    # Scale to appropriately sized float: use float32 if the
                    # dynamic range is small, to limit the memory footprint.
                    da = a_max.astype(np.float64) - a_min.astype(np.float64)
                    scaled_dtype = np.float64 if da > 1e8 else np.float32

                # Scale the input data to [.1, .9].  The Agg interpolators clip
                # to [0, 1] internally, and we use a smaller input scale to
                # identify the interpolated points that need to be flagged as
                # over/under.  This may introduce numeric instabilities in very
                # broadly scaled data.

                # Always copy, and don't allow array subtypes.
                A_scaled = np.array(A, dtype=scaled_dtype)
                # Clip scaled data around norm if necessary.  This is necessary
                # for big numbers at the edge of float64's ability to represent
                # changes.  Applying a norm first would be good, but ruins the
                # interpolation of over numbers.
                self.norm.autoscale_None(A)
                dv = np.float64(self.norm.vmax) - np.float64(self.norm.vmin)
                vmid = np.float64(self.norm.vmin) + dv / 2
                fact = 1e7 if scaled_dtype == np.float64 else 1e4
                newmin = vmid - dv * fact
                if newmin < a_min:
                    newmin = None
                else:
                    a_min = np.float64(newmin)
                newmax = vmid + dv * fact
                if newmax > a_max:
                    newmax = None
                else:
                    a_max = np.float64(newmax)
                if newmax is not None or newmin is not None:
                    np.clip(A_scaled, newmin, newmax, out=A_scaled)

                # Rescale the raw data to [offset, 1-offset] so that the
                # resampling code will run cleanly.  Using dyadic numbers here
                # could reduce the error, but would not fully eliminate it and
                # breaks a number of tests (due to the slightly different
                # error bouncing some pixels across a boundary in the (very
                # quantized) colormapping step).
                offset = .1
                frac = .8
                # Run vmin/vmax through the same rescaling as the raw data;
                # otherwise, data values close or equal to the boundaries can
                # end up on the wrong side due to floating point error.
                vmin, vmax = self.norm.vmin, self.norm.vmax
                if vmin is np.ma.masked:
                    vmin, vmax = a_min, a_max
                vrange = np.array([vmin, vmax], dtype=scaled_dtype)

>               A_scaled -= a_min
E               UserWarning: result dtype changed due to the removal of value-based promotion from NumPy. Changed from float32 to float64.

lib/matplotlib/image.py:492: UserWarning

[QuLogic]

It looks like this is caused by:

matplotlib/lib/matplotlib/axes/_axes.py

Lines 3238 to 3240 in 690aaf3

	# The use of float32 is "historical", but can't be changed without
	# regenerating the test baselines.
	x = np.asarray(x, np.float32)

(which ironically is to avoid changing test images), so it could be fixed by explicitly upcasting again:

diff --git a/lib/matplotlib/axes/_axes.py b/lib/matplotlib/axes/_axes.py
index b1343b5c65..d035e9b042 100644
--- a/lib/matplotlib/axes/_axes.py
+++ b/lib/matplotlib/axes/_axes.py
@@ -3284,7 +3284,7 @@ class Axes(_AxesBase):
         slices = []
         autotexts = []

-        for frac, label, expl in zip(x, labels, explode):
+        for frac, label, expl in zip(x.astype(np.float64), labels, explode):
             x, y = center
             theta2 = (theta1 + frac) if counterclock else (theta1 - frac)
             thetam = 2 * np.pi * 0.5 * (theta1 + theta2)