Relative errors for fp16 is not correctly set in the CTS

KhronosGroup / OpenCL-CTS

The OpenCL Conformance Tests

Apache License 2.0

184 stars 196 forks source link

Relative errors for fp16 is not correctly set in the CTS #1998

Closed rjodinchr closed 3 months ago

rjodinchr commented 3 months ago

This is true at least for divide: https://github.com/KhronosGroup/OpenCL-CTS/blob/main/test_conformance/math_brute_force/function_list.cpp#L425

ERROR: divide: 0.500000 ulp error at {0x1.428p-13 (0x090a), -0x1.ep+6 (0xd780)}
Expected: -0x1.6p-20  (half 0x8016)
Actual: -0x1.5p-20 (half 0x8015) at index: 61714

rjodinchr commented 3 months ago

My mistake I was looking at the embedded profile in the spec