With the speedup in inversion, cache lookup runs slower than always computing the inversion.

Well, that's interesting ... do you suggest effectively undoing #1216 ?

EDIT: I may attempt to profile on my couple of year old macbook air, it's the slowest PC I have easy access to setup a development environment to. We probably should verify those benchmarks with slower hardware.

I get the following results with the script included below:
Finished cache lookup: 16.0494 ms
Finished direct compute: 7.2663 ms

import fileinput

import numpy as np
import PySide2
from PySide2 import QtGui

from pyqtgraph.util.lru_cache import LRUCache
from pyqtgraph.debug import Profiler

_mapRectFromViewGlobalCache = LRUCache(100, 70)


def invertQTransform(tr):
    try:
        det = tr.determinant()
        detr = 1.0 / det
        inv = tr.adjoint()
        inv *= detr
        return inv
    except ZeroDivisionError:
        raise


def cache_lookup(vt):
    cache = _mapRectFromViewGlobalCache
    k = (
        vt.m11(), vt.m12(), vt.m13(),
        vt.m21(), vt.m22(), vt.m23(),
        vt.m31(), vt.m32(), vt.m33(),
    )

    try:
        inv_vt = cache[k]
    except KeyError:
        inv_vt = invertQTransform(vt)
        cache[k] = inv_vt
    return inv_vt


transforms = [eval(line.strip()) for line in fileinput.input("transforms.txt")]

# use a size that results in always getting a cache hit
transforms = transforms[:50]

# populate cache
for tr in transforms:
    cache_lookup(tr)

profiler = Profiler(disabled=False, delayed=False)

for _ in range(100):
    [cache_lookup(tr) for tr in transforms]
profiler("Finished cache lookup")
for _ in range(100):
    [invertQTransform(tr) for tr in transforms]
profiler("Finished direct compute")

...hard to believe the unpacking of the QTransform takes so long....thanks for the code above, I'll try and benchmark on the macbook air tomorrow.

Ok, that was unexpected;

  Finished cache lookup: 8.2791 ms
  Finished direct compute: 6.0091 ms

I think this PR should strip the cache lookup functionality unless we can identify some circumstance where it is beneficial.

Maybe that should be a separate PR, ideally by the author (@erikmansson) of #1216.

I'll give @erikmansson some time to respond, but I think removing the cache hit can be in the same PR, provided it's in a different commit (and we don't squash on merge); of course alternatively it can just be a different PR.

Please feel free to remove the cache again if the new method is quicker. I have no objections. 😊

@pijyoi I was right about to merge, you pushed the commit when I was ~5 seconds away from pushing the button :P no matter... thanks for updating the PR ... I'll merge when CI finishes up.

Thanks for the PR @pijyoi thanks for chiming in @erikmansson , merging