NEC lambda problem

[raver119]

NCC compiler and/or VE have major performance problems with lambdas.

Please consider this code: https://gist.github.com/raver119/f9aca08f8c3895b840a96d36b37e6bd5

On x86 gcc 7.4 output is:

Time omp: [215 us]; Time lambda: [222 us]

On Aurora output is:

export OMP_NUM_THREADS=8
...
Time omp: [61 us]; Time lambda: [904 us]

export OMP_NUM_THREADS=1
...
Time omp: [202 us]; Time lambda: [4131 us]

[raver119]

cc @mikishin ^^^

[Y-Shibuya]

It's not a complete solution, but it's an improvement from the development team.

One of the reasons for the difference in performance is that the function called by lambda() is not expanded inline.
Specifying -finline as a compile option improves performance to some extent.

$ /opt/nec/ve/bin/nc++ -O3 -fopenmp -finline-functions -finline-max-times=20 -finline-max-depth=20 test_lambda.cpp -o aurora_lamdba
$ ./aurora_lamdba
Time omp: [58 us]; Time lambda: [325 us]