George Parrish (Portfolio) - Boned mesh

Boned meshes - draw call optimisation

Problem

On mobile we tend to be quite draw call limited, especially on Android.

Android GLES draw call submission can be an expensive CPU operation in comparison to iOS Metal.

We budgeted for 300 draw calls in total.
We have up to 80 characters on the screen, with multiple materials and realtime shadows.
We also have Environments, UI and VFX particle effects to render.
Draw call counts add up quickly.

The Tower ( gun turret ) assets in the game were made as individual pieces and intentionally not combined.

We wanted to animate the assembly of the pieces into a whole Tower; or explode them apart again into smaller pieces, and bounce them on the environment.

5 Towers are visible in the game at all times. Each Tower is constructed from 15 pieces.
If each piece was rendered as a single draw call, the Towers would cost 75 draw calls in total, plus draw calls for shadow geometry.
This was a good chunk of our draw call budget. We needed to significantly reduce the number of draw calls for the Towers.

Tower assembly (animated)

lswb_tower_assembly.mov

Solution

We developed a system which allowed us to combine the mesh pieces into a single mesh ( and thus a single draw call ) while still being able to move and rotate the individual mesh pieces in the game as if they were separate.

This is very much like a boning system.

Offline process:

We combined the mesh geometry together to produce a single mesh asset.
We tagged each vertex as it is copied over into the combined mesh. The tag indicates which piece the vertex belongs to ( 0..14 )
We stored this tag value in the color alpha channel of each vertex.

Then at game runtime:

We updated an array of shader registers, each element contained the world position and quaternion rotation of each piece in the mesh.
We rendered the combined mesh ( one draw call ).

Our vertex shader

Determined the piece index of each vertex, using the vertex color alpha value.

Indexed into the shader registers using the piece index, to obtain the world position and world rotation of the piece.
Transformed each vertex into the world space of the piece.

Summary

We rendered and animated the assembly of a Tower in one draw call.

The CPU cost of setting 15 position and rotation registers was less than the cost of rendering 14 additional drawcalls.
When we rendered 5 Towers in the game, we reduced draw call count from 75 to 5.