Slow Text creation in manim 0.17.3

[jesi-rgb]

Description of bug / unexpected behavior

I am working on a project involving moderately large groups of mobjects (100s), which has actually never been a problem. But upgrading to 17.3 has made this process incredibly slow.

My code is a simple VGroup of VGroups. These inner VGroups are composed of a Rectangle and a Text object. That is ~100s of these VGroups which add up to double the quantity in total VMobjects.

In 0.17.2 this took absolutely no time, in 0.17.3 it takes tenths of seconds.

Setup

Inspecting a bit, turns out is about the multiple text objects. Let me provide some examples. All these are running in manim 0.17.3

Very slow

The following piece of code is a minimal working example of the setup I have. I actually have several classes and inheritance in my code, so this is what it'd look stripped down. We have a random 2D array and we iterate through it, creating a VGroup of a Square and some Text. Importantly, we are setting text properties such as the font attribute.

This piece of code takes a very long time to run. Unfortunately, I don't know how to measure manim run time. But hopefully the details help to those that can.

class Test(MovingCameraScene):
    def construct(self):
        big_kernel = np.random.randint(0, 255, (10, 10))

        big_kernel_mob = (
            VGroup(
                *[
                    VGroup(
                        Square(),
                        Text(
                            f"{c}",
                            font="Arial",
                            weight=BOLD,
                        ),
                    ).arrange(ORIGIN)
                    for row in big_kernel
                    for c in row
                ]
            )
            .arrange_in_grid(rows=10, cols=10, buff=0)
            .scale(0.3)
        )

        self.add(big_kernel_mob)

A bit faster

On the other hand, this very similar one runs normally fast. It's not immediate, but takes just under a second. The only difference is that the font is not being set, so it defaults to Times.

class Test(MovingCameraScene):
    def construct(self):
        big_kernel = np.random.randint(0, 255, (10, 10))

        big_kernel_mob = (
            VGroup(
                *[
                    VGroup(
                        Square(),
                        Text(
                            f"{c}",
                            # font="Arial",
                            weight=BOLD,
                        ),
                    ).arrange(ORIGIN)
                    for row in big_kernel
                    for c in row
                ]
            )
            .arrange_in_grid(rows=10, cols=10, buff=0)
            .scale(0.3)
        )

        self.add(big_kernel_mob)

The fastest

Even more so, if we ignore the matrix values and always use the same number, it runs even faster. One can assume that some text caching is going on with Pango, but I am not completely sure. Here, setting the font does not matter.

class Test(MovingCameraScene):
    def construct(self):
        big_kernel = np.random.randint(0, 255, (10, 10))

        big_kernel_mob = (
            VGroup(
                *[
                    VGroup(
                        Square(),
                        Text(
                            f"{20}",
                            font="Arial",
                            weight=BOLD,
                        ),
                    ).arrange(ORIGIN)
                    for row in big_kernel
                    for c in row
                ]
            )
            .arrange_in_grid(rows=10, cols=10, buff=0)
            .scale(0.3)
        )

        self.add(big_kernel_mob)

Changing to 0.17.2

However, if I reinstall manim at 17.2, the slowest piece of code now runs just as fast as usual. So, I guess some regression has happened in the new release.

System specifications

• OS: MacOS 13.2.1
• RAM: 16GB

Hopefully I didn't miss anything. Please let me know if any other details are needed! Thanks so much beforehand.

[behackl]

Thanks for the report! This is a bit unfortunate given that we intended to improve performance of the creation of large texts.

@kjlubick, could you perhaps take a closer look?

[kjlubick]

My guess was the removal of the cache in https://github.com/ManimCommunity/manim/pull/3181/files

@jesi-rgb can you try adding use_svg_cache=True to your Text() calls?

[kjlubick]

(btw, https://jiffyclub.github.io/snakeviz/ is a neat way to visualize profiles that you can generate from Python or the command line)

[jesi-rgb]

@kjlubick So, like this?

Text("a", use_svg_cache=True, font="Arial")

[jesi-rgb]

I get this error:

TypeError: manim.mobject.svg.svg_mobject.SVGMobject.__init__() got multiple values for keyword argument 'use_svg_cache'

[jesi-rgb]

Apparently, this is already set in SVGMobject, if I'm not mistaken

[kjlubick]

I guess that means I forgot to pop a kwarg.

[naveen521kk]

I think it's because of the PR which checks whether the font is installed or not, #3046. Calling manimpango.list_fonts() multiple times without caching would makes things slow.

[kjlubick]

Good insight Naveen. Maybe try adding warn_missing_font=False to the Text calls?

[naveen521kk]

Looks like it's the main cause, it's from the first slow example provided

[MrDiver]

Measuring manim is as easy as executing a profiler on the python script, there are multiple options with multiple viewers, i personally prefer vprof for ease of use.

Then you would just add

if __name__ == "__main__":
    with tempconfig({"renderer": "cairo"}):
        Test().render()

to your python script and run vprof test.py

https://github.com/nvdv/vprof

Other:

• https://github.com/plasma-umass/scalene
• https://jiffyclub.github.io/snakeviz/ + CProfile builtin

[MrDiver]

I can confirm @naveen521kk 's observation for the first example.

It improves with caching

Now the slowest part is parsing the svg.