Test the performance of permutation instances with different parameters

[CongMa13]

permutation_tuning_3 is able to test all rank 3 permutation instances. It accepts command line parameters: lens of rank3 tensor and the output dims.

For example,

root@x1000c7s5b1n0:/home/congma13/work/forks/build/hiptensor_da972cadcf9b_ck_223a2abe62f# ./bin/permutation_tuning_3 F32 124 1580 64  2 0 1
There are 70 instances.                                                                                                                                                                          DeviceElementwiseImpl<3, 256, 64, 64, 4, 4, 0, 1, 1, 1>:        Perf: 0.155879 ms, 0.16088 TFlops, 643.519 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 4, 4, 0, 1, 2, 2>:        Perf: 0.122778 ms, 0.204253 TFlops, 817.012 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 4, 4, 0, 1, 4, 4>:        Perf: 0.103642 ms, 0.241966 TFlops, 967.862 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 4, 4, 1, 0, 1, 1>:        Perf: 0.168602 ms, 0.148739 TFlops, 594.957 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 4, 4, 1, 0, 2, 2>:        Perf: 0.119111 ms, 0.210542 TFlops, 842.166 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 4, 4, 1, 0, 4, 4>:        Perf: 0.107594 ms, 0.233078 TFlops, 932.312 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 0, 1, 1, 1>:      Perf: 0.555214 ms, 0.0451677 TFlops, 180.671 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 0, 1, 2, 2>:      Perf: 0.282154 ms, 0.0888796 TFlops, 355.518 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 0, 1, 4, 4>:      Perf: 0.259383 ms, 0.0966823 TFlops, 386.729 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 0, 1, 8, 8> does not support this input tensor:
DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 0, 1, 16, 16> does not support this input tensor:
DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 1, 0, 1, 1>:      Perf: 0.54868 ms, 0.0457056 TFlops, 182.823 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 1, 0, 2, 2>:      Perf: 0.286791 ms, 0.0874426 TFlops, 349.771 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 1, 0, 4, 4>:      Perf: 0.258455 ms, 0.0970295 TFlops, 388.118 GB/s
DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 1, 0, 8, 8> does not support this input tensor:                                                                                                    DeviceElementwiseImpl<3, 256, 64, 64, 16, 16, 1, 0, 16, 16> does not support this input tensor:                                                                                                  DeviceElementwiseImpl<3, 256, 128, 128, 8, 8, 0, 1, 1, 1>:      Perf: 0.223434 ms, 0.112238 TFlops, 448.951 GB/s
DeviceElementwiseImpl<3, 256, 128, 128, 8, 8, 0, 1, 2, 2>:      Perf: 0.163738 ms, 0.153158 TFlops, 612.631 GB/s
DeviceElementwiseImpl<3, 256, 128, 128, 8, 8, 0, 1, 4, 4>:      Perf: 0.140621 ms, 0.178336 TFlops, 713.343 GB/s
DeviceElementwiseImpl<3, 256, 128, 128, 8, 8, 0, 1, 8, 8> does not support this input tensor:                                                                                                    DeviceElementwiseImpl<3, 256, 128, 128, 8, 8, 1, 0, 1, 1>:      Perf: 0.224637 ms, 0.111637 TFlops, 446.547 GB/s
DeviceElementwiseImpl<3, 256, 128, 128, 8, 8, 1, 0, 2, 2>:      Perf: 0.174285 ms, 0.143889 TFlops, 575.557 GB/s
DeviceElementwiseImpl<3, 256, 128, 128, 8, 8, 1, 0, 4, 4>:      Perf: 0.153869 ms, 0.162981 TFlops, 651.924 GB/s
DeviceElementwiseImpl<3, 256, 128, 128, 8, 8, 1, 0, 8, 8> does not support this input tensor:                   

[cgmillette]

Do we need to change the sample?

[cgmillette]

Would it make sense here to have rank determined from argc?
RANK = (argc - 2) >>1; // as long as (argc -2) % 2 == 0

Then we can have 1 executable and combine all the cases?