[libclc] erfc accuracy issues

[rjodinchr]

I am not able to pass the OpenCL-CTS for erfc using clvk on the following platform:

• Intel & AMD GPU on ChromeOS
• Nvidia (Tesla V100-SXM2-16GB) on Linux
• Swiftshader on Linux
• llvmpipe on Linux

I'm not sure if it's about clvk or if the implementation is not accurate enough.

Testing has been added to the OpenCL-CTS here:
KhronosGroup/OpenCL-CTS@97cf4c7

@kpet, @alan-baker, @frasercrmck, @karolherbst, is it passing on your side using libclc?

[karolherbst]

Not passing with rusticl either

[frasercrmck]

I don't have a perfect set-up right now (mixing and matching libclc targets with an OCK OpenCL implementation), but I do indeed see failures when using the libclc erfc function:

ERROR: erfc: -213.066727 ulp error at 0x1.38p+3 (0x411c0000): *0x1.aap-142 vs. 0x0p+0
ERROR: erfc: -289.349731 ulp error at 0x1.378p+3 (0x411bc000): *0x1.21p-141 vs. 0x0p+0
ERROR: erfc: -84.825691 ulp error at 0x1.398p+3 (0x411cc000): *0x1.54p-143 vs. 0x0p+0
ERROR: erfc: -392.753082 ulp error at 0x1.37p+3 (0x411b8000): *0x1.89p-141 vs. 0x0p+0
ERROR: erfc: -156.818466 ulp error at 0x1.388p+3 (0x411c4000): *0x1.3ap-142 vs. 0x0p+0
ERROR: erfc: -45.794533 ulp error at 0x1.3a8p+3 (0x411d4000): *0x1.7p-144 vs. 0x0p+0
ERROR: erfc: -18.099249 ulp error at 0x1.3cp+3 (0x411e0000): *0x1.2p-145 vs. 0x0p+0
ERROR: erfc: -24.674929 ulp error at 0x1.3b8p+3 (0x411dc000): *0x1.9p-145 vs. 0x0p+0
ERROR: erfc: -115.363335 ulp error at 0x1.39p+3 (0x411c8000): *0x1.ccp-143 vs. 0x0p+0
ERROR: erfc: -62.341320 ulp error at 0x1.3ap+3 (0x411d0000): *0x1.fp-144 vs. 0x0p+0
ERROR: erfc: -33.623295 ulp error at 0x1.3bp+3 (0x411d8000): *0x1.1p-144 vs. 0x0p+0
ERROR: erfc: -532.850159 ulp error at 0x1.368p+3 (0x411b4000): *0x1.0a8p-140 vs. 0x0p+0

FWIW, only on float - I see double and half passes.

It is passing with OCK's own implementation of the builtin.

[rjodinchr]

I also see it passing for double and half.

I'm going to try to port the OCK implementation to libclc to see if it passes

https://github.com/uxlfoundation/oneapi-construction-kit/blob/f7c4f7ec7595df318f623a4ca77afbca9cfb0d71/modules/compiler/builtins/abacus/source/abacus_math/erfc.cpp#L89

[rjodinchr]

the OCK implementation is passing on Nvidia, but not on any of the other platforms I've been trying (Intel and AMD on ChromeOS, swiftshader, llvmpipe).