do it by 1/sqrt. need to be refined for performance
877ab2a backend: add double version of rsqrt
backend/src/libocl/tmpl/ocl_math_common.tmpl.cl | 5 +++++
backend/src/libocl/tmpl/ocl_math_common.tmpl.h | 1 +
2 files changed, 6 insertions(+)
Upstream: cgit.freedesktop.org