MIOpenBatchNormBwdPerAct.cl kernel bug

[joye]

in https://github.com/ROCmSoftwarePlatform/MIOpen/blob/master/src/kernels/MIOpenBatchNormBwdPerAct.cl#L174

the tmp2's calculation is wrong, which should be

dyelem = (_FLOAT_PREC)(dy_in[index]);
tmp2 = mad((_FLOAT_PREC)N, dyelem * pvt_scale, -tmp1);

please check the formula of this derivative.

After this change, the result can match with cuDNN result for the same input.

[ce1adon]

@joye @atamazov
Thank you Mei! You are right. This calculation is wrong.

Based on this information the calculation at
https://github.com/ROCmSoftwarePlatform/MIOpen/blob/master/src/kernels/MIOpenBatchNormBwdPerAct.cl#L265
also need to be corrected.

Considering we don't have error report in ctest or driver, I think the verification in ctest and driver also need to be changed. Can you setup a priority level?

[ce1adon]

I think instead of doing a patch here, maybe put it in the next release. What do you think? @atamazov

[atamazov]

Yes, let's fix during normal development process. Anyone who needs the fix ASAP can ask us explicitly (right here) for the patch.

[daniellowell]

I’ll fix it for the next release.

[daniellowell]

Fixed in MIOpen 2.0.0