Related issues:

• ULP requirements for fp16 divide #1278
• math bruteforce correctly rounded square root vs. non-correctly rounded test confusion OpenCL-CTS#2387

In short, we would like to revisit the accuracy requirement for an fp16 square root. As a reminder, the current accuracy requirement for the fp16 square root is "correctly rounded", compared to "<= 3 ulp" for the fp32 square root. We would like to propose relaxing the accuracy requirement for the fp16 square root to "<= 1 ulp". This relaxation would be used by our current GPU implementation, which we have been shipping for many years. We would be open to an option to get back to a correctly rounded fp16 square root if desired, say via a new build option, or by applying the fp32 correctly rounded build option to fp16 as well.

We only found this issue now due to the CTS issue mentioned above ☹