GPU| reduce bubbles and kernel overhead (#1132)

Author: abussy

src/densities.jl

@@ -38,7 +38,6 @@ using an optional `occupation_threshold`. By default all occupation numbers are
                                           .* (basis.fft_grid.ifft_normalization)^2
                                           .* abs2.(storage.ψnk_real))
 
-        synchronize_device(basis.architecture)
     end
     ρ = sum(getfield.(storages, :ρ))
 
@@ -93,7 +92,6 @@ end
                                                           .* storage.δψnk_real
               .+ δoccupation[ik][n] .* basis.kweights[ik] .* abs2.(storage.ψnk_real))
 
-        synchronize_device(basis.architecture)
     end
     δρ = sum(getfield.(storages, :δρ))
 

src/terms/Hamiltonian.jl

@@ -148,11 +148,10 @@ end
         to = TimerOutput()  # Thread-local timer output
         ψ_real = storage.ψ_reals
 
-        @timeit to "local+kinetic" begin
+        @timeit to "local" begin
             ifft!(ψ_real, H.basis, H.kpoint, ψ[:, iband]; normalize=false)
             ψ_real .*= potential
             fft!(Hψ[:, iband], H.basis, H.kpoint, ψ_real; normalize=false)  # overwrites ψ_real
-            Hψ[:, iband] .+= H.fourier_op.multiplier .* ψ[:, iband]
         end
 
         if have_divAgrad
@@ -167,10 +166,11 @@ end
         if Threads.threadid() == 1
             merge!(DFTK.timer, to; tree_point=[t.name for t in DFTK.timer.timer_stack])
         end
-
-        synchronize_device(H.basis.architecture)
     end
 
+    # Kinetic term
+    Hψ .+= H.fourier_op.multiplier .* ψ
+
     # Apply the nonlocal operator.
     if !isnothing(H.nonlocal_op)
         @timing "nonlocal" begin