Optimize kernel gradient

Author: efaulhaber

src/general/corrections.jl

@@ -35,15 +35,16 @@ end
 # `rho_mean` is the mean density of the fluid, which is used to determine correction values near the free surface.
 #  Return a tuple `(viscosity_correction, pressure_correction, surface_tension_correction)` representing the correction terms.
 @inline function free_surface_correction(correction::AkinciFreeSurfaceCorrection,
-                                         particle_system, rho_mean)
+                                         particle_system, rho_a, rho_b)
     # Equation 4 in ref
+    rho_mean = (rho_a + rho_b) / 2
     k = correction.rho0 / rho_mean
 
     # Viscosity, pressure, surface_tension
     return k, 1, k
 end
 
-@inline function free_surface_correction(correction, particle_system, rho_mean)
+@inline function free_surface_correction(correction, particle_system, rho_a, rho_b)
     return 1, 1, 1
 end
 

src/schemes/fluid/weakly_compressible_sph/rhs.jl

@@ -32,8 +32,9 @@ function interact!(dv, v_particle_system, u_particle_system,
 
         # In 3D, this function can combine velocity and density load into one wide load,
         # which gives a significant speedup on GPUs.
-        v_a, rho_a = @inbounds velocity_and_density(v_particle_system, particle_system,
-                                                    particle)
+        v_a,
+        rho_a = @inbounds velocity_and_density(v_particle_system, particle_system,
+                                               particle)
 
         # Accumulate the RHS contributions over all neighbors before writing to `dv`,
         # to reduce the number of memory writes.
@@ -58,9 +59,9 @@ function interact!(dv, v_particle_system, u_particle_system,
 
             # `foreach_neighbor` makes sure that `neighbor` is in bounds of `neighbor_system`
             m_b = @inbounds hydrodynamic_mass(neighbor_system, neighbor)
-            v_b, rho_b = @inbounds velocity_and_density(v_neighbor_system, neighbor_system,
-                                                        neighbor)
-            rho_mean = (rho_a + rho_b) / 2
+            v_b,
+            rho_b = @inbounds velocity_and_density(v_neighbor_system, neighbor_system,
+                                                   neighbor)
             vdiff = v_a - v_b
 
             # The following call is equivalent to