-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathdiff_3D_nonlin_multixpu_perf.jl
165 lines (155 loc) · 10.1 KB
/
diff_3D_nonlin_multixpu_perf.jl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
const use_return = haskey(ENV, "USE_RETURN" ) ? parse(Bool, ENV["USE_RETURN"] ) : false
const USE_GPU = haskey(ENV, "USE_GPU" ) ? parse(Bool, ENV["USE_GPU"] ) : false
const do_viz = haskey(ENV, "DO_VIZ" ) ? parse(Bool, ENV["DO_VIZ"] ) : false
const do_save = haskey(ENV, "DO_SAVE" ) ? parse(Bool, ENV["DO_SAVE"] ) : false
const do_save_viz = haskey(ENV, "DO_SAVE_VIZ") ? parse(Bool, ENV["DO_SAVE_VIZ"]) : false
const nx = haskey(ENV, "NX" ) ? parse(Int , ENV["NX"] ) : 64
const ny = haskey(ENV, "NY" ) ? parse(Int , ENV["NY"] ) : 64
const nz = haskey(ENV, "NZ" ) ? parse(Int , ENV["NZ"] ) : 64
###
using ParallelStencil
using ParallelStencil.FiniteDifferences3D
@static if USE_GPU
@init_parallel_stencil(CUDA, Float64, 3)
else
@init_parallel_stencil(Threads, Float64, 3)
end
using ImplicitGlobalGrid, Plots, Printf, LinearAlgebra, MAT
import MPI
norm_g(A) = (sum2_l = sum(A.^2); sqrt(MPI.Allreduce(sum2_l, MPI.SUM, MPI.COMM_WORLD)))
@views inn(A) = A[2:end-1,2:end-1,2:end-1]
macro innH3(ix,iy,iz) esc(:( H[$ix+1,$iy+1,$iz+1]*H[$ix+1,$iy+1,$iz+1]*H[$ix+1,$iy+1,$iz+1] )) end
macro av_xi_H3(ix,iy,iz) esc(:( 0.5*(H[$ix,$iy+1,$iz+1]+H[$ix+1,$iy+1,$iz+1]) * 0.5*(H[$ix,$iy+1,$iz+1]+H[$ix+1,$iy+1,$iz+1]) * 0.5*(H[$ix,$iy+1,$iz+1]+H[$ix+1,$iy+1,$iz+1]) )) end
macro av_yi_H3(ix,iy,iz) esc(:( 0.5*(H[$ix+1,$iy,$iz+1]+H[$ix+1,$iy+1,$iz+1]) * 0.5*(H[$ix+1,$iy,$iz+1]+H[$ix+1,$iy+1,$iz+1]) * 0.5*(H[$ix+1,$iy,$iz+1]+H[$ix+1,$iy+1,$iz+1]) )) end
macro av_zi_H3(ix,iy,iz) esc(:( 0.5*(H[$ix+1,$iy+1,$iz]+H[$ix+1,$iy+1,$iz+1]) * 0.5*(H[$ix+1,$iy+1,$iz]+H[$ix+1,$iy+1,$iz+1]) * 0.5*(H[$ix+1,$iy+1,$iz]+H[$ix+1,$iy+1,$iz+1]) )) end
macro av_xi_Re(ix,iy,iz) esc(:( π + sqrt(π*π + max_lxyz2 / @av_xi_H3($ix,$iy,$iz) * _dt) )) end
macro av_yi_Re(ix,iy,iz) esc(:( π + sqrt(π*π + max_lxyz2 / @av_yi_H3($ix,$iy,$iz) * _dt) )) end
macro av_zi_Re(ix,iy,iz) esc(:( π + sqrt(π*π + max_lxyz2 / @av_zi_H3($ix,$iy,$iz) * _dt) )) end
macro Re(ix,iy,iz) esc(:( π + sqrt(π*π + max_lxyz2 / @innH3($ix,$iy,$iz) * _dt) )) end
macro av_xi_θr_dτ(ix,iy,iz) esc(:( max_lxyz / Vpdτ / @av_xi_Re($ix,$iy,$iz) * Resc )) end
macro av_yi_θr_dτ(ix,iy,iz) esc(:( max_lxyz / Vpdτ / @av_yi_Re($ix,$iy,$iz) * Resc )) end
macro av_zi_θr_dτ(ix,iy,iz) esc(:( max_lxyz / Vpdτ / @av_zi_Re($ix,$iy,$iz) * Resc )) end
macro dτ_ρ(ix,iy,iz) esc(:( Vpdτ * max_lxyz / @innH3($ix,$iy,$iz) / @Re($ix,$iy,$iz) * Resc )) end
@parallel_indices (ix,iy,iz) function compute_flux!(qHx, qHy, qHz, H, Vpdτ, Resc, _dt, max_lxyz, max_lxyz2, _dx, _dy, _dz)
if (ix<=size(qHx,1) && iy<=size(qHx,2) && iz<=size(qHx,3)) qHx[ix,iy,iz] = (qHx[ix,iy,iz] * @av_xi_θr_dτ(ix,iy,iz) - @av_xi_H3(ix,iy,iz) * _dx * (H[ix+1,iy+1,iz+1] - H[ix,iy+1,iz+1]) ) / (1.0 + @av_xi_θr_dτ(ix,iy,iz)) end
if (ix<=size(qHy,1) && iy<=size(qHy,2) && iz<=size(qHy,3)) qHy[ix,iy,iz] = (qHy[ix,iy,iz] * @av_yi_θr_dτ(ix,iy,iz) - @av_yi_H3(ix,iy,iz) * _dy * (H[ix+1,iy+1,iz+1] - H[ix+1,iy,iz+1]) ) / (1.0 + @av_yi_θr_dτ(ix,iy,iz)) end
if (ix<=size(qHz,1) && iy<=size(qHz,2) && iz<=size(qHz,3)) qHz[ix,iy,iz] = (qHz[ix,iy,iz] * @av_zi_θr_dτ(ix,iy,iz) - @av_zi_H3(ix,iy,iz) * _dz * (H[ix+1,iy+1,iz+1] - H[ix+1,iy+1,iz]) ) / (1.0 + @av_zi_θr_dτ(ix,iy,iz)) end
return
end
@parallel_indices (ix,iy,iz) function compute_update!(H, Hold, qHx, qHy, qHz, Vpdτ, Resc, _dt, max_lxyz, max_lxyz2, _dx, _dy, _dz, size_innH_1, size_innH_2, size_innH_3)
if (ix<=size_innH_1 && iy<=size_innH_2 && iz<=size_innH_3) H[ix+1,iy+1,iz+1] = (H[ix+1,iy+1,iz+1] + @dτ_ρ(ix,iy,iz) * (_dt * Hold[ix+1,iy+1,iz+1] - (_dx * (qHx[ix+1,iy,iz] - qHx[ix,iy,iz]) + _dy * (qHy[ix,iy+1,iz] - qHy[ix,iy,iz]) + _dz * (qHz[ix,iy,iz+1] - qHz[ix,iy,iz])) )) / (1.0 + _dt * @dτ_ρ(ix,iy,iz)) end
return
end
@parallel_indices (ix,iy,iz) function compute_flux_res!(qHx2, qHy2, qHz2, H, _dx, _dy, _dz)
if (ix<=size(qHx2,1) && iy<=size(qHx2,2) && iz<=size(qHx2,3)) qHx2[ix,iy,iz] = -@av_xi_H3(ix,iy,iz) * _dx * (H[ix+1,iy+1,iz+1] - H[ix,iy+1,iz+1]) end
if (ix<=size(qHy2,1) && iy<=size(qHy2,2) && iz<=size(qHy2,3)) qHy2[ix,iy,iz] = -@av_yi_H3(ix,iy,iz) * _dy * (H[ix+1,iy+1,iz+1] - H[ix+1,iy,iz+1]) end
if (ix<=size(qHz2,1) && iy<=size(qHz2,2) && iz<=size(qHz2,3)) qHz2[ix,iy,iz] = -@av_zi_H3(ix,iy,iz) * _dz * (H[ix+1,iy+1,iz+1] - H[ix+1,iy+1,iz]) end
return
end
@parallel_indices (ix,iy,iz) function check_res!(ResH, H, Hold, qHx2, qHy2, qHz2, _dt, _dx, _dy, _dz)
if (ix<=size(ResH,1) && iy<=size(ResH,2) && iz<=size(ResH,3)) ResH[ix,iy,iz] = -_dt * (H[ix+1,iy+1,iz+1] - Hold[ix+1,iy+1,iz+1]) - (_dx * (qHx[ix+1,iy,iz] - qHx[ix,iy,iz]) + _dy * (qHy[ix,iy+1,iz] - qHy[ix,iy,iz]) + _dz * (qHz[ix,iy,iz+1] - qHz[ix,iy,iz])) end
return
end
@parallel_indices (ix,iy,iz) function assign!(Hold, H)
if (ix<=size(H,1) && iy<=size(H,2) && iz<=size(H,3)) Hold[ix,iy,iz] = H[ix,iy,iz] end
return
end
@views function diffusion_3D_()
# Physics
lx, ly, lz = 10.0, 10.0, 10.0 # domain size
ttot = 0.4 # total simulation time
dt = 0.2 # physical time step
# Numerics
tol = 1e-8 # tolerance
itMax = 1e3 # max number of iterations
nout = 2000 # tol check
CFL = 1 / sqrt(3) # CFL number
Resc = 1 / 1.2 # iteration parameter scaling
me, dims, nprocs = init_global_grid(nx, ny, nz) # MPI initialisation
b_width = (16, 4, 4) # boundary width for comm/comp overlap
# Derived numerics
dx, dy, dz = lx/nx_g(), ly/ny_g(), lz/nz_g() # cell sizes
Vpdτ = CFL * min(dx, dy, dz)
max_lxyz = max(lx, ly, lz)
max_lxyz2 = max_lxyz^2
xc, yc, zc = LinRange(dx/2, lx-dx/2, nx), LinRange(dy/2, ly-dy/2, ny), LinRange(dz/2, lz-dz/2, nz)
_dx, _dy, _dz, _dt = 1.0/dx, 1.0/dy, 1.0/dz, 1.0/dt
# Array allocation
qHx = @zeros(nx-1,ny-2,nz-2)
qHy = @zeros(nx-2,ny-1,nz-2)
qHz = @zeros(nx-2,ny-2,nz-1)
qHx2 = @zeros(nx-1,ny-2,nz-2)
qHy2 = @zeros(nx-2,ny-1,nz-2)
qHz2 = @zeros(nx-2,ny-2,nz-1)
ResH = @zeros(nx-2,ny-2,nz-2)
# Initial condition
H0 = zeros(nx,ny,nz)
H0 = Data.Array([exp(-(x_g(ix,dx,H0)-0.5*lx+dx/2)*(x_g(ix,dx,H0)-0.5*lx+dx/2) - (y_g(iy,dy,H0)-0.5*ly+dy/2)*(y_g(iy,dy,H0)-0.5*ly+dy/2) - (z_g(iz,dz,H0)-0.5*lz+dz/2)*(z_g(iz,dz,H0)-0.5*lz+dz/2)) for ix=1:size(H0,1), iy=1:size(H0,2), iz=1:size(H0,3)])
Hold = @ones(nx,ny,nz) .* H0
H = @ones(nx,ny,nz) .* H0
size_innH_1, size_innH_2, size_innH_3 = size(H,1)-2, size(H,2)-2, size(H,3)-2
len_ResH_g = ((nx-2-2)*dims[1]+2)*((ny-2-2)*dims[2]+2)*((nz-2-2)*dims[3]+2)
if do_viz || do_save_viz
if (me==0) ENV["GKSwstype"]="nul"; if do_viz !ispath("../../figures") && mkdir("../../figures") end; end
nx_v, ny_v, nz_v = (nx-2)*dims[1], (ny-2)*dims[2], (nz-2)*dims[3]
if (nx_v*ny_v*nz_v*sizeof(Data.Number) > 0.8*Sys.free_memory()) error("Not enough memory for visualization.") end
H_v = zeros(nx_v, ny_v, nz_v) # global array for visu
H_inn = zeros(nx-2, ny-2, nz-2) # no halo local array for visu
z_sl = Int(ceil(nz_g()/2)) # Central z-slice
Xi_g, Yi_g = dx+dx/2:dx:lx-dx-dx/2, dy+dy/2:dy:ly-dy-dy/2 # inner points only
end
t = 0.0; it = 0; ittot = 0; nt = Int(ceil(ttot/dt)); niter = 0
# Physical time loop
while it<nt
iter = 0; err = 2*tol
# Pseudo-transient iteration
while err>tol && iter<itMax
if (it==1 && iter==0) tic(); niter = 0 end
@parallel compute_flux!(qHx, qHy, qHz, H, Vpdτ, Resc, _dt, max_lxyz, max_lxyz2, _dx, _dy, _dz)
@hide_communication b_width begin # communication/computation overlap
@parallel compute_update!(H, Hold, qHx, qHy, qHz, Vpdτ, Resc, _dt, max_lxyz, max_lxyz2, _dx, _dy, _dz, size_innH_1, size_innH_2, size_innH_3)
update_halo!(H)
end
iter += 1; niter += 1
if iter % nout == 0
@parallel compute_flux_res!(qHx2, qHy2, qHz2, H, _dx, _dy, _dz)
@parallel check_res!(ResH, H, Hold, qHx2, qHy2, qHz2, _dt, _dx, _dy, _dz)
err = norm_g(ResH) / sqrt(len_ResH_g)
end
end
ittot += iter; it += 1; t += dt
@parallel assign!(Hold, H)
if isnan(err) error("NaN") end
end
t_toc = toc()
A_eff = (2*4+2)/1e9*nx*ny*nz*sizeof(Data.Number) # Effective main memory access per iteration [GB]
t_it = t_toc/niter # Execution time per iteration [s]
T_eff = A_eff/t_it # Effective memory throughput [GB/s]
if (me==0) @printf("PERF: Time = %1.3f sec, T_eff = %1.2f GB/s (niter = %d)\n", t_toc, round(T_eff, sigdigits=3), niter) end
if (me==0) @printf("Total time = %1.2f, time steps = %d, nx = %d, iterations tot = %d \n", round(ttot, sigdigits=2), it, nx_g(), ittot) end
# Visualise
if do_viz || do_save_viz
H_inn .= Array(inn(H)); gather!(H_inn, H_v)
if me==0 && do_viz
heatmap(Xi_g, Yi_g, H_v[:,:,z_sl]', dpi=150, aspect_ratio=1, framestyle=:box, xlims=(Xi_g[1],Xi_g[end]), ylims=(Yi_g[1],Yi_g[end]), xlabel="lx", ylabel="ly", c=:viridis, clims=(0,1), title="nonlinear diffusion (nt=$it, iters=$ittot)")
savefig("../../figures/diff_3D_nonlin_multixpu_perf_$(nx_g()).png")
end
end
if me==0 && do_save
!ispath("../../output") && mkdir("../../output")
open("../../output/out_diff_3D_nonlin_multixpu_perf.txt","a") do io
println(io, "$(nprocs) $(nx_g()) $(ny_g()) $(nz_g()) $(ittot) $(t_toc) $(A_eff) $(t_it) $(T_eff)")
end
end
if me==0 && do_save_viz
!ispath("../../out_visu") && mkdir("../../out_visu")
matwrite("../../out_visu/diff_3D_nonlin_multixpu_perf.mat", Dict("H_3D"=> Array(H_v), "xc_3D"=> Array(xc), "yc_3D"=> Array(yc), "zc_3D"=> Array(zc)); compress = true)
end
finalize_global_grid()
return xc, yc, zc, H
end
if use_return
xc, yc, zc, H = diffusion_3D_();
else
diffusion_3D = begin diffusion_3D_(); return; end
end