Nightly test regressions in HIP support after the merge of PR 681 (ndtri function)

Submitted by Deven Desai

Assigned to Nobody

Link to original bugzilla bug (#1745)
Version: 3.4 (development)

Description

There are regressions in the nightly tests for Eigen with HIP support as a consequence of the code added in the PR-681 (commit : e38dd48a) that was merged yesterday (September 04, 2019).

(At least one of) the cause(s) seems to be the moving of function cepheas::polevl to internal::ppolevl

internal::ppolevl is defined in the file Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h, which in turn does not get picked up when compiling with GPU support

rocm-user@ixt-rack-04:~$ grep GenericPacketMathFunctions -r . | grep include  

./eigen/Eigen/src/Core/arch/SSE/MathFunctions.h:#include "../Default/GenericPacketMathFunctions.h"  

./eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h:#include "../Default/GenericPacketMathFunctions.h"  

./eigen/Eigen/src/Core/arch/NEON/MathFunctions.h:#include "../Default/GenericPacketMathFunctions.h"  

I suspect the same failure will also occur when compiling with GPU/CUDA support.

What is the correct way to fix this?

Thanks

deven