[Openmp-dev] Unpacking Variadic Templates inside target region is extremely slow with clang compiler

[Rahulkumar Gayatri via Openmp-dev]

Hello All,
I have observed that unpacking variadic templates inside the target region
is extremely slow compared to manually writing member functions for each
case.

I have written a test-case that is attached to this email.
In the test case there are 2 classes for multi-dimensional arrays :
1. Class specialized for 2D arrays (Array2D)
2. Class that accepts a parameter pack and can hence be used to create
generic MD arrays of any reasonable number of dimensions (ArrayMD).

I observe that using a 2D array created with the ArrayMD class is ~50x
slower when it accesses it's elements inside the target compared to
creating the same array using Array2D class.

Clang-version = LLVM/10

There is an associated makefile for the clang compiler that accepts the
following parameters :
make (sequential)
make OPENMP=y (OpenMP3.0)
make OPENMP_TARGET=y (OpenMP4.5)
make OPENMP_TARGET=y VD=y (Use the ArrayMD class to create the 2D array).

Can someone take a look at this test-case.


Regards,
Rahul.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/openmp-dev/attachments/20200702/a0992aad/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: testCase_variadicTemplates.tar
Type: application/x-tar
Size: 20480 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/openmp-dev/attachments/20200702/a0992aad/attachment-0001.tar>