diff --git a/BatchScripts/Mac/TestKernels.sh b/BatchScripts/Mac/TestKernels.sh
new file mode 100755
index 0000000..3249103
--- /dev/null
+++ b/BatchScripts/Mac/TestKernels.sh
@@ -0,0 +1,2 @@
+export JuliaDevice="CPU"
+julia --project TestKernels/testKernels.jl
diff --git a/Examples/testFVGrad.jl b/Examples/testFVGrad.jl
index 88daea2..a51b3c2 100644
--- a/Examples/testFVGrad.jl
+++ b/Examples/testFVGrad.jl
@@ -213,10 +213,16 @@ UNew = similar(U)
 @views UNewu = UNew[uPosS:uPosE]
 
 h = zeros(FTB,Grid.NumFaces)
+Div = zeros(FTB,Grid.NumFaces)
+DivEx = zeros(FTB,Grid.NumFaces)
+uN = zeros(FTB,Grid.NumEdges)
+uT = zeros(FTB,Grid.NumEdges)
+uTEx = zeros(FTB,Grid.NumEdges)
 hGrad = zeros(FTB,Grid.NumEdges)
-hGradE = zeros(FTB,Grid.NumEdges)
+hGradEx = zeros(FTB,Grid.NumEdges)
 VelSp = zeros(Grid.NumFaces,2)
 VelSpE = zeros(Grid.NumFaces,2)
+
 # Test gradient
 for iF = 1 : Grid.NumFaces
   x = Grid.Faces[iF].Mid.x  
@@ -232,28 +238,50 @@ for iE = 1 : Grid.NumEdges
   hGradx = 12 * x^3
   hGrady = 12 * y^2
   hGradz = 10 * z
-  hGradE[iE] = hGradx * Grid.Edges[iE].n.x +
+  hGradEx[iE] = hGradx * Grid.Edges[iE].n.x +
                hGrady * Grid.Edges[iE].n.y
                hGradz * Grid.Edges[iE].n.z
 end               
 mul!(hGrad,Grad,h)
 FiniteVolumes.ConvertVelocitySp!(backend,FTB,VelSp,hGrad,Grid)
 Outputs.vtkSkeleton!(vtkSkeletonMesh, GridType*"FVGrad", Proc, ProcNumber, [h VelSp], 0)
-FiniteVolumes.ConvertVelocitySp!(backend,FTB,VelSpE,hGradE,Grid)
+FiniteVolumes.ConvertVelocitySp!(backend,FTB,VelSpE,hGradEx,Grid)
 Outputs.vtkSkeleton!(vtkSkeletonMesh, GridType*"FVGrad", Proc, ProcNumber, [h VelSpE], 1)
 Outputs.vtkSkeleton!(vtkSkeletonMesh, GridType*"FVGrad", Proc, ProcNumber, [h VelSp-VelSpE], 2)
 
-# Test divergence
+# Test tangential
+VelCart=zeros(3)
+VelSphere=zeros(3)
 for iE = 1 : Grid.NumEdges
   x = Grid.Edges[iF].Mid.x
   y = Grid.Edges[iF].Mid.y
   z = Grid.Edges[iF].Mid.z
   (lon,lat,r)= Grids.cart2sphere(x,y,z)
-  uS = 3*lon^4 + 4*lat^3 
-  vS = 4*lon^3 + 5 * lat^2
+  VelSphere[1] 3*lon^4 + 4*lat^3 
+  VelSphere[2] = 4*lon^3 + 5 * lat^2
+# dulondlon = 12*lon^3
+# dulatdlat = 10*lat
+  VelCart = VelSphere2Cart(VelSphere,lon,lat)
+  n1 = Grid.Edges[iE].n.x
+  n2 = Grid.Edges[iE].n.y
+  n3 = Grid.Edges[iE].n.z
+  uN[iE] = n1 * VelCa[1] + n2 * VelCa[2] + n3 * VelCa[3]
+  t1 = Grid.Edges[iE].t.x
+  t2 = Grid.Edges[iE].t.y
+  t3 = Grid.Edges[iE].t.z
+  uTEx[iE] = t1 * VelCa[1] + t2 * VelCa[2] + t3 * VelCa[3]
 end
+mul!(uT,Tang,uN)
+FiniteVolumes.ConvertVelocityTSp!(backend,FTB,VelSpE,uTEx,Grid)
+FiniteVolumes.ConvertVelocityTSp!(backend,FTB,VelSp,uT,Grid)
+Outputs.vtkSkeleton!(vtkSkeletonMesh, GridType*"FVGrad", Proc, ProcNumber, [h VelSp], 3)
+Outputs.vtkSkeleton!(vtkSkeletonMesh, GridType*"FVGrad", Proc, ProcNumber, [h VelSpE], 4)
+Outputs.vtkSkeleton!(vtkSkeletonMesh, GridType*"FVGrad", Proc, ProcNumber, [h VelSp-VelSpE], 5)
 stop
 
+# theta lat 
+# phi   lon
+# Div = 1/sin(theta)*d/dtheta(sin(theta)*u_theta) + 1/sin(theta)*d/dphi(u_phi)
 
 
 
diff --git a/src/Examples/force.jl b/src/Examples/force.jl
index d78738f..81123ef 100644
--- a/src/Examples/force.jl
+++ b/src/Examples/force.jl
@@ -3,15 +3,15 @@ abstract type AbstractForcing end
 Base.@kwdef struct HeldSuarezForcing <: AbstractForcing end
 
 function (Force::HeldSuarezForcing)(Param,Phys)
-  function local_force(U,p,lat)
+  function local_force(F,U,p,lat)
     FT = eltype(U)
     Sigma = p / Phys.p0
     SigmaPowKappa = fast_powGPU(Sigma,Phys.kappa)
     height_factor = max(FT(0), (Sigma - Param.sigma_b) / (FT(1) - Param.sigma_b))
     coslat = cos(lat)
     sinlat = sin(lat)
-    Fu = -(Param.k_f * height_factor) * U[2]
-    Fv = -(Param.k_f * height_factor) * U[3]
+    F[2] += -(Param.k_f * height_factor) * U[2]
+    F[3] += -(Param.k_f * height_factor) * U[3]
     if Sigma < FT(0.7)
       kT = Param.k_a + (Param.k_s - Param.k_a) * height_factor * coslat * coslat * coslat * coslat
     else
@@ -21,8 +21,7 @@ function (Force::HeldSuarezForcing)(Param,Phys)
       Param.DeltaTh_z * log(Sigma) * coslat * coslat) * SigmaPowKappa
     Teq = max(Param.T_min, Teq)
     DeltaT =  kT * (Phys.p0 * Sigma / (U[1] * Phys.Rd) - Teq)
-    FRhoTh  = -U[1] * DeltaT / SigmaPowKappa
-    return FT(0),Fu,Fv,FT(0),FRhoTh  
+    F[5]  += -U[1] * DeltaT / SigmaPowKappa
   end  
   return local_force
 end
diff --git a/src/Examples/initial.jl b/src/Examples/initial.jl
index 35e0bbe..d66e8ec 100755
--- a/src/Examples/initial.jl
+++ b/src/Examples/initial.jl
@@ -456,23 +456,25 @@ function (::HeldSuarezDryExample)(Param,Phys)
     w = FT(0)
     return (Rho,uS,vS,w,Th,qv)
   end
-  function Force(U,p,lat)
+  function Force(F,U,p,lat)
     FT = eltype(U)
     Sigma = p / Phys.p0
+    SigmaPowKappa = fast_powGPU(Sigma,Phys.kappa)
     height_factor = max(FT(0), (Sigma - Param.sigma_b) / (FT(1) - Param.sigma_b))
-    Fu = -(Param.k_f * height_factor) * U[2]
-    Fv = -(Param.k_f * height_factor) * U[3]
+    coslat = cos(lat)
+    sinlat = sin(lat)
+    F[2] += -(Param.k_f * height_factor) * U[2]
+    F[3] += -(Param.k_f * height_factor) * U[3]
     if Sigma < FT(0.7)
-      kT = Param.k_a + (Param.k_s - Param.k_a) * height_factor * cos(lat) * cos(lat) * cos(lat) * cos(lat)
+      kT = Param.k_a + (Param.k_s - Param.k_a) * height_factor * coslat * coslat * coslat * coslat
     else
       kT = FT(0)
     end
-    Teq = (Param.T_equator - Param.DeltaT_y * sin(lat) * sin(lat) -
-      Param.DeltaTh_z * log(Sigma) * cos(lat) * cos(lat)) * Sigma^Phys.kappa
+    Teq = (Param.T_equator - Param.DeltaT_y * sinlat * sinlat -
+      Param.DeltaTh_z * log(Sigma) * coslat * coslat) * SigmaPowKappa
     Teq = max(Param.T_min, Teq)
     DeltaT =  kT * (Phys.p0 * Sigma / (U[1] * Phys.Rd) - Teq)
-    FRhoTh  = -U[1] * DeltaT / Sigma^Phys.kappa
-    return FT(0),Fu,Fv,FT(0),FRhoTh
+    F[5]  += -U[1] * DeltaT / SigmaPowKappa
   end
   return profile,Force
 end
@@ -641,3 +643,7 @@ function z2Eta(x,y,z,Param,Phys)
   return Eta
 
 end
+
+# we may be hitting a slow path:
+# https://stackoverflow.com/questions/14687665/very-slow-stdpow-for-bases-very-close-to-1
+fast_powGPU(x::FT, y::FT) where {FT <: AbstractFloat} = exp(y * log(x))
diff --git a/src/GPU/OperatorKernel.jl b/src/GPU/OperatorKernel.jl
index 240d754..47ceaf2 100644
--- a/src/GPU/OperatorKernel.jl
+++ b/src/GPU/OperatorKernel.jl
@@ -755,12 +755,12 @@ end
   NumG = @uniform @ndrange()[2]
 
   if IC <= NumG
-    FRho,Fu,Fv,Fw,FRhoTh = Force(view(U,Iz,IC,1:5),p[Iz,IC],xS[2,IC])
-    F[Iz,IC,1] += FRho
-    F[Iz,IC,2] += Fu
-    F[Iz,IC,3] += Fv
-    F[Iz,IC,4] += Fw
-    F[Iz,IC,5] += FRhoTh
+    Force(view(F,Iz,IC,:),view(U,Iz,IC,:),p[Iz,IC],xS[2,IC])
+#   F[Iz,IC,1] += FRho
+#   F[Iz,IC,2] += Fu
+#   F[Iz,IC,3] += Fv
+#   F[Iz,IC,4] += Fw
+#   F[Iz,IC,5] += FRhoTh
   end
 end