When using timestamp query to measure the execution time of compute shader on D3D12 backend, we find it return different results on Intel and Nvidia devices. On Intel, the timestamp is written into queries when the Dispatch command is completed (the compute shader is finished by GPU), but on Nvidia, it's written when the command is dispatched (the computer shader is not finished yet), so we need to add resource barrier before writing timestamp to ensure that the result is valid.

The question is what the execution time of Dispatch command should be, and I'm not sure if we need to insert the resource barrier implicitly?