Open argilo opened 10 months ago
Related: #686
I looked at this as well and I wrote a new kernel based on the rsqrt() intrinsic. It's faster and more accurate than this kernel.
Should I submit a PR for this?
The invsqrt
kernel needs to stay until we do another major release because we follow semver and can't just remove a kernel. We can remove a specific implementation of a kernel though.
@Ka-zam your rsqrt
implementation would probably fit best in the invsqrt
kernel.
Currently there is no test for volk_32f_invsqrt_32f, and it doesn't look like there's been one in the past.
I tried adding the following line to kernel_tests.h:
But when running the test in a loop (
ctest -R invsqrt --output-on-failure --repeat until-fail:100000
), it soon fails: