Does hardcoding the the coordinates to (0.0f, 0.0f)-(1.0f, 1.0f) make any difference? Might be a quick way to rule out texture coordinates as a cause.
I tried that, doesn't make any difference. I'll do some tests outside wine with glDrawPixels, because I'm not really sure if it's really that slow on my card. Maybe the d3d initialisation drives it into a bad state.