I remember this game accessing the depth buffer too often per frame. As many light sources on the screen as many lfb-lock to read back the corresponding depth value to detect if the lightsource is in coverage or not (for the light aura rendering).
From my understanding, that's what causes the game to slow down dramatically.

Do you play this game with Glide\ForceEmulatingTruePCIAccess enabled? By that way,

1) The lfb readbacks should be much faster for those lock-read a pixel or a small segment-unlock API driving pattern
2) As for QEmu, the client OS and the game never reaches the physical D3D buffer directly by the CPU (only normal host system memory) , so the problems of mapping different type of memories like GPU into the client space should not matter if I get it right (I read Kjliew's concerns about that)