As seibert says, there is no double-precision texturing support in the hardware. If you want to use texture merely for faster access to double-precision data (due to a not fully coalesced acces pattern), the recommended approach is to map an int2 texture over the double data, and re-interpreting each int2 element as a double on read, using the CUDA device function __hiloint2double(). One can abstract this in any way desired, e.g.
size_t texofs = 0;
cudaBindTexture (&texofs, tex, x, N*sizeof(x));
#define fetch(i) (temp=tex1Dfetch(tex,texofs+(i)),__hiloint2double(temp.y,temp.x))