Remove @linearidx in permutedims kernel

[GiggleLiu]

NEW

➜  scripts git:(master) ✗ JULIA_CUDA_USE_BINARYBUILDER=false julia permutedims.jl       
BenchmarkTools.Trial: 118 samples with 1 evaluation.
 Range (min … max):  42.366 ms … 42.511 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     42.393 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   42.398 ms ± 17.662 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

                   ▃▄▁█▄▆▃ ▁ ▁ ▁ ▃                             
  ▄▁▁▁▄▁▆▄▄▆▆▆▁▇▇▄▇███████▆█▇█▁█▇█▆▇▁▁▆▁▆▄▁▄▄▆▄▇▆▁▁▇▇▁▁▄▁▁▁▁▆ ▄
  42.4 ms         Histogram: frequency by time        42.4 ms <

 Memory estimate: 9.05 KiB, allocs estimate: 95.%

OLD

➜  scripts git:(master) ✗ JULIA_CUDA_USE_BINARYBUILDER=false julia permutedims.jl
BenchmarkTools.Trial: 31 samples with 1 evaluation.
 Range (min … max):  165.607 ms … 165.798 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     165.701 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   165.698 ms ±  50.178 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

                       ▃        █           █                    
  ▇▇▁▁▁▇▁▁▇▁▁▇▁▇▁▁▇▁▁▇▇█▇▁▁▇▁▇▁▇█▇▇▇▁▁▇▇▁▇▁▁█▁▁▁▁▁▁▇▁▁▁▁▁▁▁▁▇▇▇ ▁
  166 ms           Histogram: frequency by time          166 ms <

SCRIPT

➜  scripts git:(master) ✗ cat permutedims.jl 
using CUDA, BenchmarkTools, Random

let D = 28
    a = CUDA.randn(fill(2, D)...)
    b = CUDA.randn(fill(2, D)...)
    pm = (randperm(D)...,)
    display(@benchmark (CUDA.@sync permutedims!($b, $a, $pm)))
end

Please also check the discussion in JuliaGPU/CUDA.jl#1301

[maleadt]

Eh. Let me try storing the array length instead, that follows how Base works.

(but thanks for the PR and finding the issue, of course)