HannesDec 2, 2023
@tojiro
Quick question: If a big storage buffer b1 needs to be copied to another buffer b2 at the beginning of a compute shader (where those buffers are parallelized with index "i = global_id.x"), would it be faster to call "b2[i] = b1[i]" in the shader or call "commandEncoder.copyBufferToBuffer(b1, 0, b2)" before the shader call?
#webgpu
Show threadBrandon JonesDec 2, 2023

@hanesu Hm, I think it may depend on the hardware, but if you're doing the compute shader anyway then doing the copy element-by-element at the start of the shader function strikes me as potentially more efficient overall.

If you didn't have other shader work to do that's already touching each buffer element I think I'd always prefer the copy method.

Show threadHannes
@tojiro Makes sense, thanks!
Dec 2, 2023 at 4:17pmWeb