Unity body/shape creation times faster than standalone?

[TBbadmofo]

I was first trying something in Unity and thought I could shave off some overhead by using box2d directly, but I'm currently seeing the opposite.
I'm creating 10,500 bodies and shapes all 1 unit apart. All being created in one loop before the first step. I'm reusing 1 definition for the body and 1 definition for the shape.
These are just square static shapes.

In Unity I'm doing the same thing except using Unity's methods. New GameObject, Add Component of BoxCollider2D, which accepts a previously defined BoxCollider2D.

In Unity this executes in about 50 ms.

In the standalone box2d using the latest code, this takes about 1.2 seconds. where 1 second of it is the b2CreatePolygonShape() and the other 0.2 is b2CreateBody()

Why is Unity so much faster with the creation times?

[erincatto]

It might be that the Unity API uses width whereas the Box2D API uses half-widths. This could lead to many shapes overlapping, creating and n-squared cost.

[TBbadmofo]

I multiplied the positions by 5 so they would be well separated, and see the same times. I also put a -1 filter on them.
I forgot to add they are 1 unit of size. So they would now have 4 units of separation in both X and Y.
If I understood you correctly this is what you were referring to with width and half width?

I also tried adding them all once and deleting them all to have the arrays allocated to that size and then adding them again with the same performance. Although I noticed another problem. When I destroyed all the bodies, my steps were taking 41x longer(41ms vs <1 ms). I added the bodies back after they were destroyed, and performance seemed to be back to before they were destroyed.

[erincatto]

Yeah, if they are spread out then there should be no performance problem. I have a benchmark for body creation that indicates vastly faster performance. This scene of 5K bodies is created in 25ms, which corresponds closely to the Unity timing. Are you able to do any more detailed profiling? Can you hit pause in the debugger to see the source lines hit?

[TBbadmofo]

I will try running the examples and seeing what times they get. I did build this from source with MinGW, not sure if that could be the issue. I will try to learn gprof and get some profiling information. I was just commenting the body creation and shape creation and seeing the times drop.
test.txt

I uploaded some example code of what I'm doing. Also included removing the bodies with the time step taking a lot of time.

[erincatto]

Okay, I tested your code. It does take quite a while in the debug build. This is due to heavy validation. In the release build it takes ~7ms. See BOX2D_VALIDATE and B2_VALIDATE.

[TBbadmofo]

Dang, I was digging into that last night, but I was not making the connection. Thank you so much. Blazing fast now.