Port stress/response calculations to the GPU

ext/DFTKCUDAExt.jl

@@ -5,6 +5,7 @@ import DFTK: CPU, GPU, DispatchFunctional, precompilation_workflow
 using DftFunctionals
 using DFTK
 using Libxc
+import ForwardDiff: Dual
 
 DFTK.synchronize_device(::GPU{<:CUDA.CuArray}) = CUDA.synchronize()
 
@@ -20,6 +21,13 @@ for fun in (:potential_terms, :kernel_terms)
     end
 end
 
+# Ensure DFTK's custom ForwardDiff rule for FFTs is used.
+# See: https://github.com/JuliaGPU/CUDA.jl/issues/3018
+function Base.:*(p::CUFFT.CuFFTPlan{T,S,K,false},
+                 x::CuArray{<:Complex{<:Dual{Tg}}}) where {T,S,K,Tg}
+    DFTK.dual_fft_mul(p, x)
+end
+
 # Insure pre-compilation can proceed without error (old Julia/packages versions)
 if Libxc.has_cuda() && !isnothing(Base.get_extension(Libxc, :LibxcCudaExt))
 

src/DispatchFunctional.jl

@@ -26,7 +26,7 @@ function DftFunctionals.has_energy(func::LibxcFunctional)
     0 in Libxc.supported_derivatives(Libxc.Functional(func.identifier))
 end
 
-function libxc_unfold_spin(data::Matrix, n_spin::Int)
+function libxc_unfold_spin(data::AbstractMatrix, n_spin::Int)
     n_p = size(data, 2)
     if n_spin == 1
         data  # Only one spin component

src/gpu/gpu_arrays.jl

@@ -13,13 +13,13 @@ function LinearAlgebra.norm(A::Hermitian{T, <:AbstractGPUArray}) where {T}
     sqrt(2upper_triangle - diago)
 end
 
-
+# Make sure that there is a CPU fallback for AbstractGPUArrays (e.g. for Duals)
 for fun in (:potential_terms, :kernel_terms)
     @eval function DftFunctionals.$fun(fun::DispatchFunctional, ρ::AT,
-                                       args...) where {AT <: AbstractGPUArray{Float64}}
+                                       args...) where {AT <: AbstractGPUArray}
         # Fallback implementation for the GPU: Transfer to the CPU and run computation there
         cpuify(::Nothing) = nothing
         cpuify(x::AbstractArray) = Array(x)
         $fun(fun, Array(ρ), cpuify.(args)...)
     end
-end
+end

src/response/chi0.jl

@@ -409,6 +409,7 @@ to the Hamiltonian change `δH` represented by the matrix-vector products `δHψ
     # We then use the extra information we have from these additional bands,
     # non-necessarily converged, to split the Sternheimer_solver with a Schur
     # complement.
+    occupation = [to_cpu(oc) for oc in occupation]
     (mask_occ, mask_extra) = occupied_empty_masks(occupation, occupation_threshold)
 
     ψ_occ   = [ψ[ik][:, maskk] for (ik, maskk) in enumerate(mask_occ)]
@@ -561,6 +562,7 @@ function construct_bandtol(Bandtol::Type, basis::PlaneWaveBasis, ψ, occupation:
     Ω  = basis.model.unit_cell_volume
     Ng = prod(basis.fft_size)
     Nk = length(basis.kpoints)
+    occupation = [to_cpu(oc) for oc in occupation]
     mask_occ = occupied_empty_masks(occupation, occupation_threshold).mask_occ
 
     # Including k-points the expression (3.11) in 2505.02319 becomes

src/terms/local_nonlinearity.jl

@@ -11,9 +11,12 @@ end
 
 function ene_ops(term::TermLocalNonlinearity, basis::PlaneWaveBasis{T}, ψ, occupation;
                  ρ, kwargs...) where {T}
+
+    # FD on the GPU, when T<:Dual causes all sorts of troubles, at least on AMD.
+    # Because this is not often use, simply treat it on the CPU, and move result to GPU.
     fp(ρ) = ForwardDiff.derivative(term.f, ρ)
     E = sum(fρ -> convert_dual(T, fρ), term.f.(ρ)) * basis.dvol
-    potential = convert_dual.(T, fp.(ρ))
+    potential = to_device(basis.architecture, convert_dual.(T, fp.(to_cpu(ρ))))
 
     # In the case of collinear spin, the potential is spin-dependent
     ops = [RealSpaceMultiplication(basis, kpt, potential[:, :, :, kpt.spin])
@@ -22,16 +25,16 @@ function ene_ops(term::TermLocalNonlinearity, basis::PlaneWaveBasis{T}, ψ, occu
 end
 
 
-function compute_kernel(term::TermLocalNonlinearity, ::AbstractBasis{T}; ρ, kwargs...) where {T}
+function compute_kernel(term::TermLocalNonlinearity, basis::AbstractBasis{T}; ρ, kwargs...) where {T}
     fp(ρ) = ForwardDiff.derivative(term.f, ρ)
     fpp(ρ) = ForwardDiff.derivative(fp, ρ)
-    Diagonal(vec(convert_dual.(T, fpp.(ρ))))
+    Diagonal(to_device(basis.architecture, vec(convert_dual.(T, fpp.(to_cpu(ρ))))))
 end
 
-function apply_kernel(term::TermLocalNonlinearity, ::AbstractBasis{T},
+function apply_kernel(term::TermLocalNonlinearity, basis::AbstractBasis{T},
                       δρ::AbstractArray{Tδρ}; ρ, kwargs...) where {T, Tδρ}
     S = promote_type(T, Tδρ)
     fp(ρ) = ForwardDiff.derivative(term.f, ρ)
     fpp(ρ) = ForwardDiff.derivative(fp, ρ)
-    convert_dual.(S, fpp.(ρ) .* δρ)
+    to_device(basis.architecture, convert_dual.(S, fpp.(to_cpu(ρ)) .* to_cpu(δρ)))
 end

src/terms/xc.jl

@@ -42,7 +42,7 @@ function (xc::Xc)(basis::PlaneWaveBasis{T}) where {T}
         # Strip duals from functional parameters if needed
         params = parameters(fun)
         if !isempty(params)
-            newparams = convert_dual.(T, params)
+            newparams = map(p -> convert_dual(T, p), params)
             fun = change_parameters(fun, newparams; keep_identifier=true)
         end
         fun
@@ -427,7 +427,7 @@ function apply_kernel(term::TermXc, basis::PlaneWaveBasis{T}, δρ::AbstractArra
 
     # If the XC functional is not supported for an architecture, terms is on the CPU
     terms = kernel_terms(term.functionals, density)
-    δV = zeros(Tδρ, size(ρ)...)  # [ix, iy, iz, iσ]
+    δV = zeros_like(δρ, Tδρ, size(ρ)...)  # [ix, iy, iz, iσ]
 
     Vρρ = to_device(basis.architecture, reshape(terms.Vρρ, n_spin, n_spin, basis.fft_size...))
     @views for s = 1:n_spin, t = 1:n_spin  # LDA term
@@ -529,11 +529,20 @@ _matify(data::AbstractArray) = reshape(data, size(data, 1), :)
 
 for fun in (:potential_terms, :kernel_terms)
     @eval begin
-        function DftFunctionals.$fun(xc::Functional, density::LibxcDensities)
+        function DftFunctionals.$fun(xc::DispatchFunctional, density::LibxcDensities)
             $fun(xc, _matify(density.ρ_real), _matify(density.σ_real),
                      _matify(density.τ_real), _matify(density.Δρ_real))
         end
 
+        # Ensure functionals from DftFunctionals are sent to the CPU
+        # TODO: Allow GPUArrys once DftFunctionals is refactored to support GPU. 
+        function DftFunctionals.$fun(fun::DftFunctionals.Functional, density::LibxcDensities)
+            maticpuify(::Nothing) = nothing
+            maticpuify(x::AbstractArray) = reshape(Array(x), size(x, 1), :)
+            DftFunctionals.$fun(fun, maticpuify(density.ρ_real), maticpuify(density.σ_real),
+                                 maticpuify(density.τ_real), maticpuify(density.Δρ_real))
+        end
+
         function DftFunctionals.$fun(xcs::Vector{Functional}, density::LibxcDensities)
             isempty(xcs) && return NamedTuple()
             result = $fun(xcs[1], density)

src/workarounds/forwarddiff_rules.jl

@@ -33,7 +33,9 @@ function LinearAlgebra.mul!(y::AbstractArray{<:Union{Complex{<:Dual}}},
                             x::AbstractArray{<:Union{Complex{<:Dual}}})
     copyto!(y, p*x)
 end
-function Base.:*(p::AbstractFFTs.Plan, x::AbstractArray{<:Complex{<:Dual{Tg}}}) where {Tg}
+# Custom rule for FFTs used to overload the Base.:* operator. This is a function
+# of it's own because DFTKCUDAExt.jl uses it to implement a specific overload.
+function dual_fft_mul(p::AbstractFFTs.Plan, x::AbstractArray{<:Complex{<:Dual{Tg}}}) where {Tg}
     # TODO do we want x::AbstractArray{<:Dual{T}} too?
     xtil = p * ForwardDiff.value.(x)
     dxtils = ntuple(ForwardDiff.npartials(eltype(x))) do n
@@ -46,6 +48,9 @@ function Base.:*(p::AbstractFFTs.Plan, x::AbstractArray{<:Complex{<:Dual{Tg}}})
         )
     end
 end
+function Base.:*(p::AbstractFFTs.Plan, x::AbstractArray{<:Complex{<:Dual{Tg}}}) where {Tg}
+    dual_fft_mul(p, x)
+end
 
 function build_fft_plans!(tmp::AbstractArray{Complex{T}}) where {T<:Dual}
     opFFT  = AbstractFFTs.plan_fft(tmp)
@@ -230,9 +235,9 @@ end
 
 
 @timing "self_consistent_field ForwardDiff" function self_consistent_field(
-        basis_dual::PlaneWaveBasis{T};
+        basis_dual::PlaneWaveBasis{<:Dual{Tg,V,N}};
         response=ResponseOptions(),
-        kwargs...) where {T <: Dual}
+        kwargs...) where {Tg,V,N}
     # Note: No guarantees on this interface yet.
 
     # Primal pass
@@ -251,28 +256,29 @@ end
     end
     # Implicit differentiation
     response.verbose && println("Solving response problem")
-    δresults = ntuple(ForwardDiff.npartials(T)) do α
+    δresults = ntuple(N) do α
         δHextψ = [ForwardDiff.partials.(δHextψk, α) for δHextψk in Hψ_dual]
         δtemperature = ForwardDiff.partials(basis_dual.model.temperature, α)
         solve_ΩplusK_split(scfres, δHextψ; δtemperature,
                            tol=last(scfres.history_Δρ), response.verbose)
     end
 
     # Convert and combine
-    DT = Dual{ForwardDiff.tagtype(T)}
     ψ = map(scfres.ψ, getfield.(δresults, :δψ)...) do ψk, δψk...
         map(ψk, δψk...) do ψnk, δψnk...
-            Complex(DT(real(ψnk), real.(δψnk)),
-                    DT(imag(ψnk), imag.(δψnk)))
+            Complex(Dual{Tg}(real(ψnk), real.(δψnk)),
+                    Dual{Tg}(imag(ψnk), imag.(δψnk)))
         end
     end
     eigenvalues = map(scfres.eigenvalues, getfield.(δresults, :δeigenvalues)...) do εk, δεk...
-        map((εnk, δεnk...) -> DT(εnk, δεnk), εk, δεk...)
+        map((εnk, δεnk...) -> Dual{Tg}(εnk, δεnk), εk, δεk...)
     end
     occupation = map(scfres.occupation, getfield.(δresults, :δoccupation)...) do occk, δocck...
-        map((occnk, δoccnk...) -> DT(occnk, δoccnk), occk, δocck...)
+        occk_cpu = to_cpu(occk)
+        to_device(basis_dual.architecture,
+                  map((occk_cpu, δocck...) -> Dual{Tg}(occk_cpu, δocck), occk_cpu, δocck...))
     end
-    εF = DT(scfres.εF, getfield.(δresults, :δεF)...)
+    εF = Dual{Tg}(scfres.εF, getfield.(δresults, :δεF)...)
 
     # For strain, basis_dual contributes an explicit lattice contribution which
     # is not contained in δresults, so we need to recompute ρ here