introduce muladd, update AMR tests

trixi-framework · Sep 20, 2023 · d9e6af9 · d9e6af9
1 parent 73384ac
commit d9e6af9
Show file tree

Hide file tree

Showing 4 changed files with 28 additions and 4 deletions.
diff --git a/src/equations/compressible_navier_stokes_1d.jl b/src/equations/compressible_navier_stokes_1d.jl
@@ -1,3 +1,10 @@
+# By default, Julia/LLVM does not use fused multiply-add operations (FMAs).
+# Since these FMAs can increase the performance of many numerical algorithms,
+# we need to opt-in explicitly.
+# See https://ranocha.de/blog/Optimizing_EC_Trixi for further details.
+@muladd begin
+#! format: noindent
+
 @doc raw"""
     CompressibleNavierStokesDiffusion1D(equations; mu, Pr,
                                         gradient_variables=GradientVariablesPrimitive())
@@ -401,3 +408,4 @@ end
                                                                                                                                       })
     return SVector(flux_inner[1], flux_inner[2], flux_inner[3])
 end
+end # @muladd
diff --git a/src/equations/compressible_navier_stokes_2d.jl b/src/equations/compressible_navier_stokes_2d.jl
@@ -1,3 +1,10 @@
+# By default, Julia/LLVM does not use fused multiply-add operations (FMAs).
+# Since these FMAs can increase the performance of many numerical algorithms,
+# we need to opt-in explicitly.
+# See https://ranocha.de/blog/Optimizing_EC_Trixi for further details.
+@muladd begin
+#! format: noindent
+
 @doc raw"""
     CompressibleNavierStokesDiffusion2D(equations; mu, Pr,
                                         gradient_variables=GradientVariablesPrimitive())
@@ -433,3 +440,4 @@ end
                                                                                                                                       })
     return SVector(flux_inner[1], flux_inner[2], flux_inner[3], flux_inner[4])
 end
+end # @muladd
diff --git a/src/equations/compressible_navier_stokes_3d.jl b/src/equations/compressible_navier_stokes_3d.jl
@@ -1,3 +1,10 @@
+# By default, Julia/LLVM does not use fused multiply-add operations (FMAs).
+# Since these FMAs can increase the performance of many numerical algorithms,
+# we need to opt-in explicitly.
+# See https://ranocha.de/blog/Optimizing_EC_Trixi for further details.
+@muladd begin
+#! format: noindent
+
 @doc raw"""
     CompressibleNavierStokesDiffusion3D(equations; mu, Pr,
                                         gradient_variables=GradientVariablesPrimitive())
@@ -471,3 +478,4 @@ end
     return SVector(flux_inner[1], flux_inner[2], flux_inner[3], flux_inner[4],
                    flux_inner[5])
 end
+end # @muladd
diff --git a/test/test_parabolic_1d.jl b/test/test_parabolic_1d.jl
@@ -80,8 +80,8 @@ isdir(outdir) && rm(outdir, recursive=true)
     @test_trixi_include(joinpath(examples_dir(), "tree_1d_dgsem", "elixir_navierstokes_convergence_walls_amr.jl"),
       equations_parabolic = CompressibleNavierStokesDiffusion1D(equations, mu=mu(),
                                                                 Prandtl=prandtl_number()),
-      l2 = [2.527877257772131e-5, 2.5539911566937718e-5, 0.0001211860451244785],
-      linf = [0.00014663867588948776, 0.00019422448348348196, 0.0009556439394007299]
+      l2 = [2.5278824700860636e-5, 2.5540078777006958e-5, 0.00012118655083858043],
+      linf = [0.0001466387075579334, 0.00019422427462629705, 0.0009556446847707178]
     )
   end
 
@@ -90,8 +90,8 @@ isdir(outdir) && rm(outdir, recursive=true)
       equations_parabolic = CompressibleNavierStokesDiffusion1D(equations, mu=mu(),
                                                                 Prandtl=prandtl_number(), 
                                                                 gradient_variables = GradientVariablesEntropy()),
-      l2 = [2.4593699163175966e-5, 2.392863645712634e-5, 0.00011252526651714956],
-      linf = [0.00011850555445525046, 0.0001898777490968537, 0.0009597561467877824]
+      l2 = [2.459359632523962e-5, 2.3928390718460263e-5, 0.00011252414117082376],
+      linf = [0.0001185052018830568, 0.00018987717854305393, 0.0009597503607920999]
     )
   end
 end