DX9 : sending surface back to CPU

Old question I know...lot of people discourage such procedure.

It's an implementation of Occlusion Test in GTA San Andreas (so DX9)...which is used not only for rendering, but for several internal decisions on this game.

It's working already...but it takes 10~12 milliseconds, which causes huge FPS drop (60 --> 40).

Reducing surface resolution has little effect (width / 4, height / 4 ---> still getting 8 mS delay).

I've heard that compressing the surface to h.264 (GPU side) using IMFTransform or SinkWriter, and sending back the compressed frame would be (maybe) faster.

In my first attempt for a solution, I've got stuck due to IMFTransform::ProcessInput is rejecting my surface (error 0x80004001). Maybe driver issues, or my RX-460 is too old).

Should I try via SinkWriter ? (Any direction would be really appreciated)