Discussion regarding float packing for data sent to GPU

[Qix-]

Hey, I realize this is a vague question. Apologies!

A while back I remember reading some of the rationale for removing non-float datatypes from either uniforms or instance buffers (can't remember which) because of some investigation @bkaradzic had done where he had concluded this data was being sent up in packs of floats at a time under the hood. This lead to some API being changed or removed or something of the sort.

I can't remember what it was about, but it had some interesting technical information that I'd really like to read over again - does anyone know what I'm referring to and might have the link?

Thanks so much!

[bkaradzic]

Try using different type uniforms with D3D compiler, and observe layout in disassembly.

You would get:

float3 x; // 1x 32-bit padding here;
float2 y; // 2x 32-bit padding here;
float4 z; // no padding
float  w; // 3x 32-bit padding here;

[Qix-]

Interesting, thanks!

Is this only D3D? Or does this layout occur in GL, Vulkan, etc?

[Qix-]

Oh I found the issue - #653. Uniform layouts were exactly what I was remembering.

[bkaradzic]

Is this only D3D? Or does this layout occur in GL, Vulkan, etc?

Everywhere is similar issue, esp. with explicit APIs, with exception of old-GL where it's implementation dependent. Modern GL has the same issue (see info about layout(std140)): https://stackoverflow.com/questions/15750788/glsl-layout-std140-padding-dilemma

[Qix-]

Great info, this explains a bit. Thanks as always @bkaradzic :)