Adding rocfft

[psteinb]

even though #124

[tdd11235813]

ok, just checked their readme to see, that the rocfft uses hip under the hood, so your code might run on both platforms by either using nvcc or hcc. The hip calls should be error-handled like in CUDA. rocFFT itself has a hipFFT wrapper. That helps when you come from cuFFT code.

rocFFT specs (copied)

• 1D, 2D & 3D
• power of 2, 3, 5 and mix of them, from size 1 to 2^27. That is, the input sizes that can be written in the form 2^a × 3^b × 5^c. The smaller the prime factor, the better the performance, i.e., powers of two are fastest.
• Single and double precision
• Complex transforms
• Real transforms (internally uses complex)
• In-place and out-of-place transforms
• Interleaved layout, planar layout is not supported currently
• Unit (=1) stride (packed) for 2D and/or 3D. Unit and Non-unit stride (non-packed data) for 1D

[psteinb]

56d402c builds a rocm 1.9.2 based rocfft runner, but it throws a runtime exception which I have to investigate

[psteinb]

ready to rock and roll! I just produced this with rocfft in rocm 1.9.2 on my fiji nano. A bit of an xmas present to @tdd11235813 and gearshifft. ;)

$ cat result.csv 
; "Fiji [Radeon R9 FURY / NANO Series]", "CC", 3.0, "PCI Bus ID", 3, "PCI Device ID", 0, "Multiprocessors", 64, "Memory [MiB]", 4096, "MemoryFree [MiB]", 3840, "HostMemory [MiB]", 64190, "MemClock [MHz]", 500, "GPUClock [MHz]", 1000, "rocfft", 806,"NumberWarmups",2,"NumberWarmRuns",10,"NumberTotalRuns",12,"ErrorBound",1e-05,"CurrentTime",1545302436,"CurrentTimeLocal","Thu Dec 20 11:40:36 2018","Hostname","islay.mpi-cbg.de","gearshifft","0.3.0"
; "Time_ContextCreate [ms]", 0.139483
; "Time_ContextDestroy [ms]", 0.084825
"library","inplace","complex","precision","dim","kind","nx","ny","nz","run","id","success","Time_Allocation [ms]","Time_PlanInitFwd [ms]","Time_PlanInitInv [ms]","Time_Upload [ms]","Time_FFT [ms]","Time_iFFT [ms]","Time_Download [ms]","Time_PlanDestroy [ms]","Time_Total [ms]","Size_DeviceBuffer [bytes]","Size_DevicePlan [bytes]","Size_DeviceTransfer [bytes]","Error_StandardDeviation","Error_Mismatches"
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,0,0,"Warmup",0.00462,1.514999,0.00344,0.32063698769,1.8060619831,1.8831809759,0.26831799746,0.03605,5.922963,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,1,0,"Warmup",0.004873,1.479619,0.005321,0.32303699851,1.8236620426,1.8718210459,0.30271700025,0.033771,5.991953,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,2,0,"Success",0.00452,1.536595,1.57312,0.28335699439,1.8033419847,1.8231819868,0.32591700554,0.034452,7.506252,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,3,0,"Success",0.003991,1.534916,1.323269,0.2974370122,1.8092620373,1.7956620455,0.32255700231,0.036562,7.204425,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,4,0,"Success",0.004337,1.541563,0.003636,0.3150370121,1.7907019854,1.8723009825,0.28927698731,0.032172,5.996714,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,5,0,"Success",0.0044,1.528481,0.005542,0.29535698891,1.8105419874,1.8059020042,0.32575699687,0.035631,5.896988,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,6,0,"Success",0.004275,1.542887,0.005118,0.30975699425,1.8286219835,1.8831809759,0.28495699167,0.031162,6.033764,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,7,0,"Success",0.004367,1.515764,0.003337,0.30735701323,1.826382041,1.866541028,0.34447699785,0.029631,6.047815,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,8,0,"Success",0.003804,1.529423,0.003619,0.27839699388,1.8143819571,1.8556619883,0.32111701369,0.035784,5.93105,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,9,0,"Success",0.004242,1.568875,0.003564,0.3084770143,1.8206219673,1.8628610373,0.34495699406,0.031359,6.089821,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,10,0,"Success",0.004232,1.529613,0.003135,0.27663698792,1.8127820492,1.8612610102,0.31919699907,0.035513,5.930612,1063368,0,1063364,2.1931190051e-07,0
"RocFFT","Inplace","Real","float",1,"oddshape",265841,0,0,11,0,"Success",0.004447,1.681499,0.003732,0.31951698661,1.8051019907,1.8540619612,0.31711700559,0.034131,6.11645,1063368,0,1063364,2.1931190051e-07,0```

[tdd11235813]

YES awesome, thanks for the update :) I would start to integrate this in PR #124 after it is merged as discussed offline, so just let me know, if you want me to merge this. Thanks again and Happy Holidays! :)

[psteinb]

Happy New Year. Yes, please merge this PR and then we'll go for the super builds. I just tested with rocm 2.0 and it works flawlessly.

[psteinb]

thanks for the thorough review @tdd11235813, I hope I adressed all your concerns

[tdd11235813]

awesome, thank you so much for your work!