Hyperdiffusion W

Jobs/NHSphere/JobNHHeldSuarezMoistSphere

@@ -1,4 +1,4 @@
-mpirun -n 6 julia --project Examples/testNHSphere.jl \
+mpirun -n 4 julia --project Examples/testNHSphere.jl \
   --Problem="HeldSuarezMoistSphere" \
   --Device="CPU" \
   --GPUType="Metal" \
@@ -14,13 +14,13 @@ mpirun -n 6 julia --project Examples/testNHSphere.jl \
   --Forcing=true \
   --Curl=false \
   --ModelType="VectorInvariant" \
-  --VerticalDiffusion=true \
-  --SurfaceFlux=true \
+  --VerticalDiffusion=false \
+  --SurfaceFlux=false \
   --Coriolis=true \
   --Upwind=true \
-  --HorLimit=true \
+  --HorLimit=false \
   --Equation=CompressibleMoist \
-  --Microphysics=true \
+  --Microphysics=false \
   --TypeMicrophysics="SimpleMicrophysics" \
   --Buoyancy=true \
   --Damping=true \

Project.toml

@@ -23,6 +23,7 @@ NLsolve = "2774e3e8-f4cf-5e23-947b-6d7e65073b56"
 NearestNeighbors = "b8a86587-4115-5ab1-83bc-aa920d37bbce"
 NetCDF = "30363a11-5582-574a-97bb-aa9a979735b9"
 PairedLinkedLists = "7a42b37b-ed3b-477a-9848-3661f53bb718"
+Polynomials = "f27b6e38-b328-58d1-80ce-0feddd5e7a45"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
 ProfileCanvas = "efd6af41-a80b-495e-886c-e51b0c7d77a3"
 Revise = "295af30f-e4ad-537b-8983-00126c2a3abe"

src/FiniteElements/FiniteElements.jl

@@ -0,0 +1,15 @@
+module FiniteElements
+
+using Polynomials
+
+abstract type FiniteElement end
+
+struct NodalElement <: FiniteElement 
+  NumBases::Int
+  NodalBases::Array{Polynomial,1}
+  NodalPoints::Array{Float64,2}
+  Diff1Matrix::Array{Float64,2}
+end
+
+
+end

src/GPU/FcnGPU.jl

@@ -104,13 +104,13 @@ end
 
 function FcnGPU!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
 
-  @show sum(abs.(U))
   backend = get_backend(F)
   FT = eltype(F)
   Glob = FE.Glob
   DS = FE.DS
   DW = FE.DW
   M = FE.M
+  MW = FE.MW
   dXdxI = Metric.dXdxI
   X = Metric.X
   J = Metric.J
@@ -128,7 +128,6 @@ function FcnGPU!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
   lat = Metric.lat  
   dz = Metric.dz  
   zP = Metric.zP  
-  @show size(zP)
   DoF = FE.DoF
   N = size(FE.DS,1)
   Nz = size(F,1)
@@ -157,7 +156,8 @@ function FcnGPU!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
   @views FRhoTr = F[:,:,5]
 # Cache
   @views CacheF = Temp1[:,:,1:6]
-  @views CacheFF = Temp1[:,:,1:6+NumTr]
+  @views CacheFF = Temp1[:,:,1:6+NumTr+1]
+  @views Cachew = Temp1[:,:,6 + 1 + NumTr]  
   @views p = Cache.AuxG[:,:,1]
   KV = Cache.KV
   TSurf = Cache.TSurf
@@ -182,6 +182,10 @@ function FcnGPU!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
   NDoFG = min(div(NumberThreadGPU,Nz),NDoF)
   groupG = (Nz, NDoFG)  
   ndrangeG = (Nz, NDoF)  
+  NzG = min(div(NumberThreadGPU,N*N),Nz-1)
+  groupw = (N, N, NzG, 1)
+  ndrangewB = (Nz-1, NBF)  
+  ndrangewI = (Nz-1, NF-NBF)  
 
   KRhoGradKinKernel! = RhoGradKinKernel!(backend,group)
   KGradKernel! = GradKernel!(backend,group)
@@ -195,6 +199,8 @@ function FcnGPU!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
   end  
   KHyperViscTracerKernel! = HyperViscTracerKernel!(backend, groupTr)
   KHyperViscTracerKoeffKernel! = HyperViscTracerKoeffKernel!(backend, groupTr)
+  KHyperViscWKernel! = HyperViscWKernel!(backend, groupTr)
+  KHyperViscWKoeffKernel! = HyperViscWKoeffKernel!(backend, groupTr)
   KDivRhoTrUpwind3Kernel! = DivRhoTrUpwind3Kernel!(backend, groupTr)
   if Global.Model.SurfaceFlux
     NFG = min(div(NumberThreadGPU,N*N),NF)
@@ -218,6 +224,8 @@ function FcnGPU!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
     @views KHyperViscTracerKernel!(CacheTr,U[:,:,iT+NumV],Rho,DS,DW,dXdxI,J,M,Glob,ndrange=ndrangeB)
     KernelAbstractions.synchronize(backend)
   end  
+  @views KHyperViscWKernel!(Cachew,U[:,:,4],DS,DW,dXdxI,J,MW,Glob,ndrange=ndrangeB)
+  KernelAbstractions.synchronize(backend)
   Parallels.ExchangeData3DSendGPU(CacheFF,Exchange)
 
   KHyperViscKernel!(CacheF,MRho,U,DS,DW,dXdxI_I,J_I,M,Glob_I,ndrange=ndrangeI)
@@ -227,6 +235,8 @@ function FcnGPU!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
     @views KHyperViscTracerKernel!(CacheTr,U[:,:,iT+NumV],Rho,DS,DW,dXdxI_I,J_I,M,Glob_I,ndrange=ndrangeI)
     KernelAbstractions.synchronize(backend)
   end  
+  @views KHyperViscWKernel!(Cachew,U[:,:,4],DS,DW,dXdxI_I,J_I,MW,Glob_I,ndrange=ndrangeB)
+  KernelAbstractions.synchronize(backend)
 
   Parallels.ExchangeData3DRecvGPU!(CacheFF,Exchange)
   KernelAbstractions.synchronize(backend)
@@ -244,6 +254,8 @@ function FcnGPU!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
       KoeffDiv,ndrange=ndrangeB)
     KernelAbstractions.synchronize(backend)
   end  
+  @views KHyperViscWKoeffKernel!(F[:,:,4],Cachew,DS,DW,dXdxI,J,MW,Glob,KoeffDiv,ndrange=ndrangeB)
+  KernelAbstractions.synchronize(backend)
   KGradKernel!(F,U,p,DS,dXdxI,J,M,MRho,Glob,Phys,ndrange=ndrangeB)
   KernelAbstractions.synchronize(backend)
   if Global.Model.Coriolis
@@ -286,6 +298,8 @@ function FcnGPU!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
       KoeffDiv,ndrange=ndrangeI)
     KernelAbstractions.synchronize(backend)
   end  
+  @views KHyperViscWKoeffKernel!(F[:,:,4],Cachew,DS,DW,dXdxI_I,J_I,MW,Glob,KoeffDiv,ndrange=ndrangeI)
+  KernelAbstractions.synchronize(backend)
   KGradKernel!(F,U,p,DS,dXdxI_I,J_I,M,MRho,Glob_I,Phys,ndrange=ndrangeI)
   KernelAbstractions.synchronize(backend)
   if Global.Model.Coriolis
@@ -351,6 +365,7 @@ function FcnGPUAMD!(F,U,FE,Metric,Phys,Cache,Exchange,Global,Param,DiscType)
   DS = FE.DS
   DW = FE.DW
   M = FE.M
+  MW = FE.MW
   dXdxI = Metric.dXdxI
   X = Metric.X
   J = Metric.J

src/GPU/OperatorKernel.jl

@@ -557,6 +557,59 @@ end
   end
 end
 
+@kernel function HyperViscWKernel!(Fw,@Const(w),@Const(D),@Const(DW),@Const(dXdxI),
+  @Const(JJ),@Const(MW),@Const(Glob)) 
+
+  I, J, iz   = @index(Local, NTuple)
+  _,_,Iz,IF = @index(Global, NTuple)
+
+  ColumnTilesDim = @uniform @groupsize()[3]
+  N = @uniform @groupsize()[1]
+  Nz = @uniform @ndrange()[3]
+  NF = @uniform @ndrange()[4]
+
+  ID = I + (J - 1) * N  
+  @inbounds ind = Glob[ID,IF]
+
+  wCol = @localmem eltype(Fw) (N,N, ColumnTilesDim)
+  wCxCol = @localmem eltype(Fw) (N,N, ColumnTilesDim)
+  wCyCol = @localmem eltype(Fw) (N,N, ColumnTilesDim)
+  if Iz <= Nz && IF <= NF
+    @inbounds wCol[I,J,iz] = w[Iz,ind] 
+  end
+  @synchronize
+
+  ID = I + (J - 1) * N  
+  @inbounds ind = Glob[ID,IF]
+
+  if Iz <= Nz && IF <= NF
+    @inbounds Dxc = D[I,1] * wCol[1,J,iz]
+    @inbounds Dyc = D[J,1] * wCol[I,1,iz]
+    for k = 2 : N
+      @inbounds Dxc += D[I,k] * wCol[k,J,iz]
+      @inbounds Dyc += D[J,k] * wCol[I,k,iz] 
+    end
+    @views @inbounds (GradDx, GradDy) = Grad12(Dxc,Dyc,dXdxI[1:2,1:2,:,ID,Iz,IF],JJ[ID,:,Iz,IF])
+    @views @inbounds (tempx, tempy) = Contra12(GradDx,GradDy,dXdxI[1:2,1:2,:,ID,Iz,IF])
+    @inbounds wCxCol[I,J,iz] = tempx
+    @inbounds wCyCol[I,J,iz] = tempy
+  end
+
+  @synchronize 
+
+  ID = I + (J - 1) * N  
+  @inbounds ind = Glob[ID,IF]
+  if Iz <= Nz && IF <= NF
+    @inbounds Divw = DW[I,1] * wCxCol[1,J,iz] + DW[J,1] * wCyCol[I,1,iz]
+    for k = 2 : N
+      @inbounds Divw += DW[I,k] * wCxCol[k,J,iz] + DW[J,k] * wCyCol[I,k,iz]
+    end
+    if Iz < Nz 
+      @inbounds @atomic Fw[Iz+1,ind] += Divw / MW[Iz,ind]
+    end  
+  end
+end
+
 @kernel function HyperViscKoeffKernel!(F,@Const(U),@Const(Cache),@Const(D),@Const(DW),@Const(dXdxI),
   @Const(JJ),@Const(M),@Const(Glob),KoeffCurl,KoeffGrad,KoeffDiv) 
 
@@ -692,6 +745,59 @@ end
   end
 end
 
+@kernel function HyperViscWKoeffKernel!(Fw,@Const(w),@Const(D),@Const(DW),@Const(dXdxI),
+  @Const(JJ),@Const(MW),@Const(Glob),KoeffDivW) 
+
+  I, J, iz   = @index(Local, NTuple)
+  _,_,Iz,IF = @index(Global, NTuple)
+
+  ColumnTilesDim = @uniform @groupsize()[3]
+  N = @uniform @groupsize()[1]
+  Nz = @uniform @ndrange()[3]
+  NF = @uniform @ndrange()[4]
+
+  ID = I + (J - 1) * N  
+  @inbounds ind = Glob[ID,IF]
+
+  wCol = @localmem eltype(Fw) (N,N, ColumnTilesDim)
+  wCxCol = @localmem eltype(Fw) (N,N, ColumnTilesDim)
+  wCyCol = @localmem eltype(Fw) (N,N, ColumnTilesDim)
+  if Iz <= Nz && IF <= NF
+    @inbounds wCol[I,J,iz] = w[Iz,ind] 
+  end
+  @synchronize
+
+  ID = I + (J - 1) * N  
+  @inbounds ind = Glob[ID,IF]
+
+  if Iz <= Nz && IF <= NF
+    @inbounds Dxc = D[I,1] * wCol[1,J,iz]
+    @inbounds Dyc = D[J,1] * wCol[I,1,iz]
+    for k = 2 : N
+      @inbounds Dxc += D[I,k] * wCol[k,J,iz]
+      @inbounds Dyc += D[J,k] * wCol[I,k,iz] 
+    end
+    @views @inbounds (GradDx, GradDy) = Grad12(Dxc,Dyc,dXdxI[1:2,1:2,:,ID,Iz,IF],JJ[ID,:,Iz,IF])
+    @views @inbounds (tempx, tempy) = Contra12(GradDx,GradDy,dXdxI[1:2,1:2,:,ID,Iz,IF])
+    @inbounds wCxCol[I,J,iz] = tempx
+    @inbounds wCyCol[I,J,iz] = tempy
+  end
+
+  @synchronize 
+
+  ID = I + (J - 1) * N  
+  @inbounds ind = Glob[ID,IF]
+  if Iz <= Nz && IF <= NF
+    @inbounds Divw = DW[I,1] * wCxCol[1,J,iz] + DW[J,1] * wCyCol[I,1,iz]
+    for k = 2 : N
+      @inbounds Divw += DW[I,k] * wCxCol[k,J,iz] + DW[J,k] * wCyCol[I,k,iz]
+    end
+    if Iz < Nz 
+      @inbounds @atomic Fw[Iz,ind] += -KoeffDivW * Divw / MW[Iz,ind]
+    end  
+  end
+end
+
 @kernel function DivRhoTrCentralKernel!(F,@Const(c),@Const(uC),@Const(vC),@Const(w),
   @Const(D),@Const(dXdxI),
   @Const(JJ),@Const(M),@Const(Glob))