mirror of
https://igit.ific.uv.es/alramos/latticegpu.jl.git
synced 2025-05-15 11:43:42 +02:00
Working multi-precision simulations
The pure gauge theory with groups SU(2) and SU(3) is working properly.
This commit is contained in:
Alberto Ramos
parent
1416efdbee
commit
09a09153b9
8 changed files with 188 additions and 283 deletions
|
|
@ -11,196 +11,66 @@
|
|||
|
||||
un(t) = t <: Union{Group, Complex}
|
||||
|
||||
function field(::Type{T}, lp::SpaceParm) where {T <: Union{Group, Algebra}}
|
||||
function field(::Type{G}, ::Type{T}, lp::SpaceParm) where {G <: Union{Group, Algebra}, T <: AbstractFloat}
|
||||
|
||||
sz = lp.bsz, lp.ndim, lp.rsz
|
||||
if (T == SU2)
|
||||
if (G == SU2)
|
||||
# As = StructArray{SU2}(undef, sz, unwrap=un)
|
||||
return CuArray{SU2, 3}(undef, sz)
|
||||
elseif (T == SU2alg)
|
||||
return CuArray{SU2{T}, 3}(undef, sz)
|
||||
elseif (G == SU2alg)
|
||||
# As = StructArray{SU2alg}(undef, sz, unwrap=un)
|
||||
return CuArray{SU2alg, 3}(undef, sz)
|
||||
elseif (T == SU3)
|
||||
return CuArray{SU2alg{T}, 3}(undef, sz)
|
||||
elseif (G == SU3)
|
||||
# As = StructArray{SU3}(undef, sz, unwrap=un)
|
||||
return CuArray{SU3, 3}(undef, sz)
|
||||
return CuArray{SU3{T}, 3}(undef, sz)
|
||||
# As = Array{SU3, lp.ndim+1}(undef, sz)
|
||||
elseif (T == SU3alg)
|
||||
elseif (G == SU3alg)
|
||||
# As = StructArray{SU3alg}(undef, sz, unwrap=un)
|
||||
return CuArray{SU3alg, 3}(undef, sz)
|
||||
return CuArray{SU3alg{T}, 3}(undef, sz)
|
||||
# As = Array{SU3alg, lp.ndim+1}(undef, sz)
|
||||
end
|
||||
|
||||
|
||||
return replace_storage(CuArray, As)
|
||||
|
||||
end
|
||||
|
||||
function krnl_SU3_one!(G, lp::SpaceParm)
|
||||
|
||||
b, r = CUDA.threadIdx().x, CUDA.blockIdx().x
|
||||
|
||||
for id in 1:lp.ndim
|
||||
G[b,id,r] = SU3(1.0,0.0,0.0,0.0,1.0,0.0)
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
function krnl_SU2_one!(G, lp::SpaceParm)
|
||||
|
||||
b, r = CUDA.threadIdx().x, CUDA.blockIdx().x
|
||||
|
||||
for id in 1:lp.ndim
|
||||
G[b,id,r] = SU2(1.0,0.0)
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
function krnl_SU3alg_zero!(G, lp::SpaceParm)
|
||||
|
||||
b, r = CUDA.threadIdx().x, CUDA.blockIdx().x
|
||||
|
||||
for id in 1:lp.ndim
|
||||
G[b,id,r] = SU3alg(0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0)
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
function krnl_SU2alg_zero!(G, lp::SpaceParm)
|
||||
|
||||
b, r = CUDA.threadIdx().x, CUDA.blockIdx().x
|
||||
|
||||
for id in 1:lp.ndim
|
||||
G[b,id,r] = SU2alg(0.0,0.0,0.0)
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
function krnl_SU3alg_assign!(G, M, lp::SpaceParm)
|
||||
|
||||
b, r = CUDA.threadIdx().x, CUDA.blockIdx().x
|
||||
|
||||
for id in 1:lp.ndim
|
||||
G[b,id,r] = SU3alg(M[b,id,r,1], M[b,id,r,2], M[b,id,r,3], M[b,id,r,4],
|
||||
M[b,id,r,5], M[b,id,r,6], M[b,id,r,7], M[b,id,r,8])
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
function krnl_SU2alg_assign!(G, M, lp::SpaceParm)
|
||||
|
||||
b, r = CUDA.threadIdx().x, CUDA.blockIdx().x
|
||||
|
||||
for id in 1:lp.ndim
|
||||
G[b,id,r] = SU2alg(M[b,id,r,1], M[b,id,r,2], M[b,id,r,3])
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
function randomn!(X, lp)
|
||||
|
||||
if (eltype(X) == SU2alg)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t1)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t2)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t3)
|
||||
M = CuArray{Float64}(undef, lp.bsz, lp.ndim, lp.rsz, 3)
|
||||
randn!(CURAND.default_rng(), M)
|
||||
CUDA.@sync begin
|
||||
CUDA.@cuda threads=lp.bsz blocks=lp.rsz krnl_SU2alg_assign!(X, M, lp)
|
||||
end
|
||||
elseif (eltype(X) == SU3alg)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t1)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t2)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t3)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t4)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t5)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t6)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t7)
|
||||
# randn!(CURAND.default_rng(), LazyRows(X).t8)
|
||||
M = CuArray{Float64}(undef, lp.bsz, lp.ndim, lp.rsz, 8)
|
||||
randn!(CURAND.default_rng(), M)
|
||||
CUDA.@sync begin
|
||||
CUDA.@cuda threads=lp.bsz blocks=lp.rsz krnl_SU3alg_assign!(X, M, lp)
|
||||
end
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
function zero!(X, lp)
|
||||
|
||||
if (eltype(X) == SU2alg)
|
||||
# fill!(LazyRows(X).t1, 0.0)
|
||||
# fill!(LazyRows(X).t2, 0.0)
|
||||
# fill!(LazyRows(X).t3, 0.0)
|
||||
CUDA.@sync begin
|
||||
CUDA.@cuda threads=lp.bsz blocks=lp.rsz krnl_SU2alg_zero!(X, lp)
|
||||
end
|
||||
end
|
||||
|
||||
if (eltype(X) == SU3alg)
|
||||
# fill!(LazyRows(X).t1, 0.0)
|
||||
# fill!(LazyRows(X).t2, 0.0)
|
||||
# fill!(LazyRows(X).t3, 0.0)
|
||||
# fill!(LazyRows(X).t4, 0.0)
|
||||
# fill!(LazyRows(X).t5, 0.0)
|
||||
# fill!(LazyRows(X).t6, 0.0)
|
||||
# fill!(LazyRows(X).t7, 0.0)
|
||||
# fill!(LazyRows(X).t8, 0.0)
|
||||
CUDA.@sync begin
|
||||
CUDA.@cuda threads=lp.bsz blocks=lp.rsz krnl_SU3alg_zero!(X, lp)
|
||||
end
|
||||
end
|
||||
|
||||
if (eltype(X) == SU2)
|
||||
# fill!(LazyRows(X).t1.re, 1.0)
|
||||
# fill!(LazyRows(X).t1.im, 0.0)
|
||||
# fill!(LazyRows(X).t2.re, 0.0)
|
||||
# fill!(LazyRows(X).t2.im, 0.0)
|
||||
CUDA.@sync begin
|
||||
CUDA.@cuda threads=lp.bsz blocks=lp.rsz krnl_SU2_one!(X, lp)
|
||||
end
|
||||
end
|
||||
|
||||
if (eltype(X) == SU3)
|
||||
# fill!(LazyRows(X).u11.re, 1.0)
|
||||
# fill!(LazyRows(X).u11.im, 0.0)
|
||||
# fill!(LazyRows(X).u12.re, 0.0)
|
||||
# fill!(LazyRows(X).u12.im, 0.0)
|
||||
# fill!(LazyRows(X).u13.re, 0.0)
|
||||
# fill!(LazyRows(X).u13.im, 0.0)
|
||||
# fill!(LazyRows(X).u21.re, 0.0)
|
||||
# fill!(LazyRows(X).u21.im, 0.0)
|
||||
# fill!(LazyRows(X).u22.re, 1.0)
|
||||
# fill!(LazyRows(X).u22.im, 0.0)
|
||||
# fill!(LazyRows(X).u23.re, 0.0)
|
||||
# fill!(LazyRows(X).u23.im, 0.0)
|
||||
CUDA.@sync begin
|
||||
CUDA.@cuda threads=lp.bsz blocks=lp.rsz krnl_SU3_one!(X, lp)
|
||||
end
|
||||
end
|
||||
|
||||
return nothing
|
||||
end
|
||||
function randomize!(f, lp::SpaceParm, ymws::YMworkspace)
|
||||
|
||||
function norm_field(X)
|
||||
|
||||
return CUDA.mapreduce(norm2, +, X)
|
||||
# d = 0.0
|
||||
if (eltype(X) == SU2alg)
|
||||
# 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) == 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) +
|
||||
# CUDA.mapreduce(x->x^2, +, LazyRows(X).t4) +
|
||||
# CUDA.mapreduce(x->x^2, +, LazyRows(X).t5) +
|
||||
# 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)
|
||||
if ymws.ALG == SU2alg
|
||||
m = CUDA.randn(ymws.PRC, lp.bsz,lp.ndim,3,lp.rsz)
|
||||
CUDA.@sync begin
|
||||
CUDA.@cuda threads=lp.bsz blocks=lp.rsz krnl_assign_SU2!(f,m,lp)
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
return d
|
||||
|
||||
if ymws.ALG == SU3alg
|
||||
m = CUDA.randn(ymws.PRC, lp.bsz,lp.ndim,8,lp.rsz)
|
||||
CUDA.@sync begin
|
||||
CUDA.@cuda threads=lp.bsz blocks=lp.rsz krnl_assign_SU3!(f,m,lp)
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
return nothing
|
||||
end
|
||||
|
||||
|
||||
function krnl_assign_SU3!(frc, m, lp::SpaceParm)
|
||||
|
||||
b, r = CUDA.threadIdx().x, CUDA.blockIdx().x
|
||||
for id in 1:lp.ndim
|
||||
frc[b,id,r] = SU3alg(m[b,id,1,r], m[b,id,2,r], m[b,id,3,r],
|
||||
m[b,id,4,r], m[b,id,5,r], m[b,id,6,r],
|
||||
m[b,id,7,r], m[b,id,8,r])
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
||||
function krnl_assign_SU2!(frc, m, lp::SpaceParm)
|
||||
|
||||
b, r = CUDA.threadIdx().x, CUDA.blockIdx().x
|
||||
for id in 1:lp.ndim
|
||||
frc[b,id,r] = SU2alg(m[b,id,1,r], m[b,id,2,r], m[b,id,3,r])
|
||||
end
|
||||
return nothing
|
||||
end
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue