lattice/latticegpu.jl
1
0
Fork 0
mirror of https://igit.ific.uv.es/alramos/latticegpu.jl.git synced 2025-05-15 19:53:41 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Working HMC for SU(2) and SU(3)

Browse source
This commit is contained in:
Alberto Ramos 2021-07-28 13:50:32 +02:00
parent f6be70070e
commit c378648508
6 changed files with 119 additions and 40 deletions
Download patch file Download diff file Expand all files Collapse all files

10
src/YM/YM.jl
View file

@ -43,6 +43,16 @@ struct YMworkspace
rs = zeros(Float64, lp.iL...)
return new(f1, f2, mm, u1, replace_storage(CuArray, cs))
end
if (T == SU3)
f1 = field(SU3alg, lp)
f2 = field(SU3alg, lp)
mm = field(SU3alg, lp)
u1 = field(SU3, lp)
cs = zeros(ComplexF64,lp.iL...)
rs = zeros(Float64, lp.iL...)
return new(f1, f2, mm, u1, replace_storage(CuArray, cs))
end
return nothing
end
end

18
src/YM/YMact.jl
View file

@ -17,7 +17,7 @@ function krnl_plaq!(plx, U, ipl, lp::SpaceParm)
Xu1 = up(X, id1, lp)
Xu2 = up(X, id2, lp)
plx[X] = tr(U[X, id1]*U[Xu1, id2] / (U[X, id2]*U[Xu2, id1]))
@inbounds plx[X] = tr(U[X, id1]*U[Xu1, id2] / (U[X, id2]*U[Xu2, id1]))
return nothing
end
@ -27,8 +27,8 @@ function krnl_plaq!(plx, U, lp::SpaceParm)
X = map2latt((CUDA.threadIdx().x,CUDA.threadIdx().y,CUDA.threadIdx().z),
(CUDA.blockIdx().x,CUDA.blockIdx().y,CUDA.blockIdx().z))
plx[X] = complex(0.0)
for ipl in 1:lp.npls
@inbounds plx[X] = complex(0.0)
@inbounds for ipl in 1:lp.npls
id1, id2 = lp.plidx[ipl]
Xu1 = up(X, id1, lp)
Xu2 = up(X, id2, lp)
@ -55,11 +55,13 @@ function krnl_force_wilson_pln!(frc1, frc2, U, ipl, lp::SpaceParm, gp::GaugeParm
F2 = projalg(a*b)
F3 = projalg(b*a)
frc1[X ,id1] -= F1
frc1[X ,id2] += F1
frc2[Xu1,id2] -= F2
frc2[Xu2,id1] += F3
@inbounds begin
frc1[X ,id1] -= F1
frc1[X ,id2] += F1
frc2[Xu1,id2] -= F2
frc2[Xu2,id1] += F3
end
return nothing
end

52
src/YM/YMfields.jl
View file

@ -21,6 +21,10 @@ function field(::Type{T}, lp::SpaceParm) where {T <: Union{Group, Algebra}}
zeros(Float64, sz)))
elseif (T == SU3)
As = StructArray{SU3}((ones(ComplexF64, sz), zeros(ComplexF64, sz), zeros(ComplexF64, sz), zeros(ComplexF64, sz), ones(ComplexF64, sz), zeros(ComplexF64, sz)))
# As = Array{SU3, lp.ndim+1}(undef, sz)
# CUDA.@sync begin
# CUDA.@cuda threads=kp.threads blocks=kp.blocks krnl_SU3_zero!(As, lp)
# end
elseif (T == SU3alg)
As = StructArray{SU3alg}((zeros(Float64, sz),
zeros(Float64, sz),
@ -30,6 +34,11 @@ function field(::Type{T}, lp::SpaceParm) where {T <: Union{Group, Algebra}}
zeros(Float64, sz),
zeros(Float64, sz),
zeros(Float64, sz)))
# As = Array{SU3alg, lp.ndim+1}(undef, sz)
# CUDA.@sync begin
# CUDA.@cuda threads=kp.threads blocks=kp.blocks krnl_SU3alg_zero!(As, lp)
# end
end
return replace_storage(CuArray, As)
@ -71,6 +80,9 @@ function zero!(X)
fill!(LazyRows(X).t6, 0.0)
fill!(LazyRows(X).t7, 0.0)
fill!(LazyRows(X).t8, 0.0)
# CUDA.@sync begin
# CUDA.@cuda threads=kp.threads blocks=kp.blocks krnl_SU3alg_zero!(X, lp)
# end
end
if (eltype(X) == SU2)
@ -85,6 +97,9 @@ function zero!(X)
fill!(LazyRows(X).u21, complex(0.0))
fill!(LazyRows(X).u22, complex(1.0))
fill!(LazyRows(X).u23, complex(0.0))
# CUDA.@sync begin
# CUDA.@cuda threads=kp.threads blocks=kp.blocks krnl_SU3_zero!(X, lp)
# end
end
return nothing
@ -97,7 +112,7 @@ function norm2(X)
d = CUDA.mapreduce(x->x^2, +, LazyRows(X).t1) +
CUDA.mapreduce(x->x^2, +, LazyRows(X).t2) +
CUDA.mapreduce(x->x^2, +, LazyRows(X).t3)
elseif (eltype(X) == SU2alg)
elseif (eltype(X) == SU3alg)
d = CUDA.mapreduce(x->x^2, +, LazyRows(X).t1) +
CUDA.mapreduce(x->x^2, +, LazyRows(X).t2) +
CUDA.mapreduce(x->x^2, +, LazyRows(X).t3) +
@ -106,7 +121,42 @@ function norm2(X)
CUDA.mapreduce(x->x^2, +, LazyRows(X).t6) +
CUDA.mapreduce(x->x^2, +, LazyRows(X).t7) +
CUDA.mapreduce(x->x^2, +, LazyRows(X).t8)
# d = CUDA.mapreduce(norm2, +, X)
end
return d
end
function krnl_SU3_zero!(G, lp::SpaceParm)
X = map2latt((CUDA.threadIdx().x,CUDA.threadIdx().y,CUDA.threadIdx().z),
(CUDA.blockIdx().x,CUDA.blockIdx().y,CUDA.blockIdx().z))
for id in 1:lp.ndim
G[X,id].u11 = complex(1.0)
G[X,id].u12 = complex(0.0)
G[X,id].u13 = complex(0.0)
G[X,id].u21 = complex(0.0)
G[X,id].u22 = complex(1.0)
G[X,id].u23 = complex(0.0)
end
return nothing
end
function krnl_SU3alg_zero!(G, lp::SpaceParm)
X = map2latt((CUDA.threadIdx().x,CUDA.threadIdx().y,CUDA.threadIdx().z),
(CUDA.blockIdx().x,CUDA.blockIdx().y,CUDA.blockIdx().z))
for id in 1:lp.ndim
G[X,id].t1 = 0.0
G[X,id].t2 = 0.0
G[X,id].t3 = 0.0
G[X,id].t4 = 0.0
G[X,id].t5 = 0.0
G[X,id].t6 = 0.0
G[X,id].t7 = 0.0
G[X,id].t8 = 0.0
end
return nothing
end

14
src/YM/YMhmc.jl
View file

@ -15,7 +15,7 @@ function gauge_action(U, lp::SpaceParm, gp::GaugeParm, kp::KernelParm, ymws::YMw
CUDA.@cuda threads=kp.threads blocks=kp.blocks krnl_plaq!(ymws.cm, U, lp)
end
S = gp.beta*( prod(lp.iL)*lp.npls -
CUDA.mapreduce(real, +, real.(ymws.cm))/gp.ng )
CUDA.mapreduce(real, +, ymws.cm)/gp.ng )
return S
end
@ -29,8 +29,12 @@ function plaquette(U, lp::SpaceParm, gp::GaugeParm, kp::KernelParm, ymws::YMwork
return CUDA.mapreduce(real, +, real.(ymws.cm))/(prod(lp.iL)*lp.npls)
end
hamiltonian(mom, U, lp, gp, kp, ymws) = norm2(mom)/2.0 +
gauge_action(U, lp, gp, kp, ymws)
function hamiltonian(mom, U, lp, gp, kp, ymws)
K = norm2(mom)/2.0
V = gauge_action(U, lp, gp, kp, ymws)
println("K: ", K, " V: ", V)
return K+V
end
function HMC!(U, eps, ns, lp::SpaceParm, gp::GaugeParm, kp::KernelParm, ymws::YMworkspace; noacc=false)
@ -65,9 +69,9 @@ function krnl_updt!(mom, frc1, frc2, eps1, U, eps2, lp::SpaceParm)
X = map2latt((CUDA.threadIdx().x,CUDA.threadIdx().y,CUDA.threadIdx().z),
(CUDA.blockIdx().x,CUDA.blockIdx().y,CUDA.blockIdx().z))
for id in 1:lp.ndim
@inbounds for id in 1:lp.ndim
mom[X,id] = mom[X,id] + eps1 * (frc1[X,id]+frc2[X,id])
U[X,id] = expm(U[X,id], mom[X,id], eps2)
U[X,id] = expm(U[X,id],mom[X,id], eps2)
end
return nothing