d4rkc0d3r/UnityBlendShapePerformance.md

## UnityBlendShapePerformance.md

      
    Raw
  

              UnityBlendShapePerformance.md
            
          
    Here is what I learned about unity skinned meshes and blendshapes and how they are dealt with on the GPU.

I use Nvidia Nsight Graphics to profile performance and read buffer sizes on GPU. Performance numbers are for a 4090 clocked at 2310 MHz core and stock memory. Unity version is 2019.4.31f1
TLDR

For Unity 2019 Split meshes with blendshapes into two meshes, one for blendshapes and one without.
This does not help much with the memory usage, but it does help a lot with performance. The only VRAM savings is one less copy of POSITION, NORMAL, TANGENT.
For Unity 2021+ merge skinned meshes regardless of if they have blendshapes. No more extra copies and it always uses a fast compute shader now for the bone skinning.
Strip unused blendshapes from the mesh. This can save a lot of VRAM.
Blendshape buffer on GPU


Always has all blendshapes of the mesh, even if they are not used.
Always has delta normals and tangents even if they are all 0.
Buffer stores only deltas for vertices if the delta is non-zero.

struct Blendshape
{
    uint vertexIndex;
    float3 deltaPosition;
    float3 deltaNormal;
    float3 deltaTangent;
};

Blendshape skinning on GPU


Copy POSITION, NORMAL, TANGENT to temporary buffer A
For each active blendshape apply deltas multiplied by weight in place to temporary buffer A with one compute shader call each

This is only done for vertices that have a non-zero delta. So if you move less vertices in a blendshape it will be faster.


Copy temporary buffer A to temporary buffer B
Do regular bone skinning vertex shader with temporary buffer B as input and buffer A as stream output

My test meshes

First variant is single mesh with all blendshapes.

Mesh 155K polygons 115K vertices

4.42 MiB for POSITION, NORMAL, TANGENT + 2 copies of it
1.32 MiB for COLOR, TEXCOORD0
3.53 MiB for BLENDINDICES, BLENDWEIGHTS
8.05 MiB for 24 blendshapes
or 20.15 MiB for 61 blendshapes


Total of 26.16 MiB and 38.26 MiB
Thry VRAM calculator says 49.27 MiB and 73.48 MiB respectively

Second variant is mesh split in two, one for blendshapes and one without.

Mesh 139K polygons 105K vertices

4.02 MiB for POSITION, NORMAL, TANGENT + 1 copy of it
1.20 MiB for COLOR, TEXCOORD0
3.21 MiB for BLENDINDICES, BLENDWEIGHTS


Mesh 16K polygons 10K vertices

410 KiB for POSITION, NORMAL, TANGENT + 2 copies of it
123 KiB for COLOR, TEXCOORD0
164 KiB for BLENDINDICES, BLENDWEIGHTS (only 2 bones per vertex)
8.05 MiB for 24 blendshapes
or 20.15 Mib for 61 blendshapes


Total of 22.01 MiB and 34.11 MiB
Thry VRAM calculator says 48.52 MiB and 72.74 MiB respectively

Mesh import settings have the legacy blendshape normals option enabled.
Performance


Merged mesh no active blendshapes

0.04 ms for vertex shader


Merged mesh 1 active blendshape

0.01 ms compute shader
0.37 ms copy to temporary buffer
0.37 ms for vertex shader
total 0.75 ms


Merged mesh 10 active blendshapes

0.09 ms compute shader (10 calls)
0.36 ms copy to temporary buffer
0.41 ms for vertex shader
total 0.86 ms


Split mesh no active blendshapes

0.005 ms for vertex shader
0.035 ms for vertex shader
total of 0.04 ms


Split mesh 1 active blendshape

0.01 ms compute shader
0.035 ms copy to temporary buffer
0.01 ms for vertex shader
0.035 ms for vertex shader
total of 0.09 ms


Split mesh 10 active blendshapes

0.09 ms compute shader (10 calls)
0.035 ms copy to temporary buffer
0.01 ms for vertex shader
0.035 ms for vertex shader
total of 0.17 ms


Extra test on Unity 2021.3.21f1

Merged mesh no active blendshapes

0.025 ms for compute shader (bones)


Merged mesh 1 active blendshape

0.005 ms for compute shader (blendshape)
0.025 ms for compute shader (bones)
total 0.03 ms


Merged mesh 10 active blendshapes

0.080 ms for compute shaders (blendshapes)
0.025 ms for compute shader (bones)
total 0.105 ms


Notes about the captured performance data

In this profiling run the first copy of POSITION, NORMAL, TANGENT is not showing up in timing data. When I did these tests on a 1080 Ti it was showing up. I am not certain that the timings correspond 1 to 1 to the event they are attached to. For example the final vertex shader is identical for both blendshape and non blendshape skinning, however it is much slower when any blendshapes are active. The time taken for the copies is also very suspiciously large. Copying 4.42 MiB of data with 1 TB/s bandwidth of a 4090 should take 0.0042 ms, so roughly 100x faster. I expect most of the time is just massive stalls in the GPU pipeline.
It is definitely this slow though since creating 20 copies of the merged mesh test avatar with a blendshape active predictable drops the fps while the split mesh doesn't suffer the same massive drop.