Function size vs speed

test_gpu_bigFN_compilation.jl

using BenchmarkTools
using CUDA
BenchmarkTools.DEFAULT_PARAMETERS.seconds = 1.00


test_sum(a, b, c) = begin
  I = (blockIdx().x - 1) * blockDim().x + threadIdx().x
  if I > 1000
    return
  end
  Base.Cartesian.@nexprs 10 i -> begin
    c[I, i] += b[I] + a[I] + i
  end
  return
end

N = 100
a = CuArray(fill(1.0f0, 1000))
b = CuArray(fill(1.0f0, 1000))
c = CuArray(fill(1.0f0, 1000, 200));
#%%
_test_sum = cufunction(test_sum, Tuple{CuDeviceArray{Float32,1,1},CuDeviceArray{Float32,1,1},CuDeviceArray{Float32,2,1},})
@time CUDA.@sync _test_sum(a, b, c, threads=512, blocks=2)
@time CUDA.@sync _test_sum(a, b, c, threads=512, blocks=2)
@time CUDA.@sync _test_sum(a, b, c, threads=512, blocks=2)
@time CUDA.@sync _test_sum(a, b, c, threads=512, blocks=2)
@time CUDA.@sync _test_sum(a, b, c, threads=512, blocks=2)
@time CUDA.@sync @cuda threads = 512 blocks = 2 test_sum(a, b, c)
@time CUDA.@sync @cuda threads = 512 blocks = 2 test_sum(a, b, c)

@show c[1:1], (N * N / 2 + 200) * 4


# (@benchmark CUDA.@sync @cuda threads = 512 blocks = 2 test_sum($a, $b, $c)) |> display
(@benchmark CUDA.@sync _test_sum($a, $b, $c, threads=512, blocks=2)) |> display
;