I'm trying to find the most efficient way to extract all non zero elements stored in a texture using D3D10 hardware. The resulting array need not be accessed by the CPU. I can think of 3 ways but I'm not sure which one is the most efficient or if there are better ways.

1. The naive way - copy texture to a staging resource and go through the buffer on the CPU.
2. Run a geometry shader with an n x m vertex buffer (1 vertex per pixel) as input and stream out only non zero vertices. Will require a texture lookup and conditional statement in the geometry shader. Sounds slower than (1) and also more wasteful in most circumstances.
3. Use the compute shader to do a scattered write. This seems like the best out of the 3 but I can't find a way to do this using the cs_4_0 profile as it doesn't have atomic operations required to implement a buffer index.