Add cache to kernel signature

trixi-gpu · Sep 7, 2024 · 11e42fb · 11e42fb
1 parent 559f84f
commit 11e42fb
Show file tree

Hide file tree

Showing 14 changed files with 89 additions and 44 deletions.
diff --git a/src/solvers/dg_1d.jl b/src/solvers/dg_1d.jl
@@ -393,7 +393,7 @@ end
 
 # Pack kernels for calculating volume integrals
 function cuda_volume_integral!(du, u, mesh::TreeMesh{1}, nonconservative_terms,
-                               equations, volume_integral::VolumeIntegralWeakForm, dg::DGSEM)
+                               equations, volume_integral::VolumeIntegralWeakForm, dg::DGSEM, cache)
     derivative_dhat = CuArray{Float64}(dg.basis.derivative_dhat)
     flux_arr = similar(u)
 
@@ -410,7 +410,7 @@ end
 
 function cuda_volume_integral!(du, u, mesh::TreeMesh{1}, nonconservative_terms::False,
                                equations, volume_integral::VolumeIntegralFluxDifferencing,
-                               dg::DGSEM)
+                               dg::DGSEM, cache)
     volume_flux = volume_integral.volume_flux
 
     derivative_split = dg.basis.derivative_split
@@ -435,7 +435,7 @@ function cuda_volume_integral!(du, u, mesh::TreeMesh{1}, nonconservative_terms::
 end
 
 function cuda_volume_integral!(du, u, mesh::TreeMesh{1}, nonconservative_terms::True, equations,
-                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM)
+                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM, cache)
     symmetric_flux, nonconservative_flux = dg.volume_integral.volume_flux
 
     derivative_split = dg.basis.derivative_split
@@ -470,7 +470,19 @@ function cuda_volume_integral!(du, u, mesh::TreeMesh{1}, nonconservative_terms::
 end
 
 function cuda_volume_integral!(du, u, mesh::TreeMesh{1}, nonconservative_terms::False, equations,
-                               volume_integral::VolumeIntegralShockCapturingHG, dg::DGSEM)
+                               volume_integral::VolumeIntegralShockCapturingHG, dg::DGSEM, cache)
+    element_ids_dg, element_ids_dgfv = cache.element_ids_dg, cache.element_ids_dgfv
+    volume_flux_dg, volume_flux_fv = dg.volume_integral.volume_flux_dg,
+                                     dg.volume_integral.volume_flux_fv
+    indicator = dg.volume_integral.indicator
+
+    # This will cause scalar indexing on the GPU
+    # alpha =  CuArray{Float64}(indicator(u, mesh, equations, dg, cache)) 
+
+    # TODO: Get copies of `u` and `du` on both device and host
+    alpha = indicator(Array(u), mesh, equations, dg, cache)
+    alpha = CuArray{Float64}(alpha)
+
     return nothing
 end
 
@@ -737,7 +749,7 @@ function rhs_gpu!(du_cpu, u_cpu, t, mesh::TreeMesh{1}, equations, boundary_condi
     du, u = copy_to_device!(du_cpu, u_cpu)
 
     cuda_volume_integral!(du, u, mesh, have_nonconservative_terms(equations), equations,
-                          dg.volume_integral, dg)
+                          dg.volume_integral, dg, cache)
 
     cuda_prolong2interfaces!(u, mesh, equations, cache)
 

diff --git a/src/solvers/dg_2d.jl b/src/solvers/dg_2d.jl
@@ -675,7 +675,7 @@ end
 
 # Pack kernels for calculating volume integrals
 function cuda_volume_integral!(du, u, mesh::TreeMesh{2}, nonconservative_terms, equations,
-                               volume_integral::VolumeIntegralWeakForm, dg::DGSEM)
+                               volume_integral::VolumeIntegralWeakForm, dg::DGSEM, cache)
     derivative_dhat = CuArray{Float64}(dg.basis.derivative_dhat)
     flux_arr1 = similar(u)
     flux_arr2 = similar(u)
@@ -696,7 +696,7 @@ function cuda_volume_integral!(du, u, mesh::TreeMesh{2}, nonconservative_terms,
 end
 
 function cuda_volume_integral!(du, u, mesh::TreeMesh{2}, nonconservative_terms::False, equations,
-                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM)
+                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM, cache)
     volume_flux = volume_integral.volume_flux
 
     derivative_split = dg.basis.derivative_split
@@ -727,7 +727,7 @@ function cuda_volume_integral!(du, u, mesh::TreeMesh{2}, nonconservative_terms::
 end
 
 function cuda_volume_integral!(du, u, mesh::TreeMesh{2}, nonconservative_terms::True, equations,
-                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM)
+                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM, cache)
     symmetric_flux, nonconservative_flux = dg.volume_integral.volume_flux
 
     derivative_split = dg.basis.derivative_split
@@ -775,7 +775,7 @@ function cuda_volume_integral!(du, u, mesh::TreeMesh{2}, nonconservative_terms::
 end
 
 function cuda_volume_integral!(du, u, mesh::TreeMesh{2}, nonconservative_terms::False, equations,
-                               volume_integral::VolumeIntegralShockCapturingHG, dg::DGSEM)
+                               volume_integral::VolumeIntegralShockCapturingHG, dg::DGSEM, cache)
     return nothing
 end
 
@@ -1175,7 +1175,7 @@ function rhs_gpu!(du_cpu, u_cpu, t, mesh::TreeMesh{2}, equations, boundary_condi
     du, u = copy_to_device!(du_cpu, u_cpu)
 
     cuda_volume_integral!(du, u, mesh, have_nonconservative_terms(equations), equations,
-                          dg.volume_integral, dg)
+                          dg.volume_integral, dg, cache)
 
     cuda_prolong2interfaces!(u, mesh, equations, cache)
 

diff --git a/src/solvers/dg_3d.jl b/src/solvers/dg_3d.jl
@@ -741,7 +741,7 @@ end
 
 # Pack kernels for calculating volume integrals
 function cuda_volume_integral!(du, u, mesh::TreeMesh{3}, nonconservative_terms, equations,
-                               volume_integral::VolumeIntegralWeakForm, dg::DGSEM)
+                               volume_integral::VolumeIntegralWeakForm, dg::DGSEM, cache)
     derivative_dhat = CuArray{Float64}(dg.basis.derivative_dhat)
     flux_arr1 = similar(u)
     flux_arr2 = similar(u)
@@ -765,7 +765,7 @@ function cuda_volume_integral!(du, u, mesh::TreeMesh{3}, nonconservative_terms,
 end
 
 function cuda_volume_integral!(du, u, mesh::TreeMesh{3}, nonconservative_terms::False, equations,
-                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM)
+                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM, cache)
     volume_flux = volume_integral.volume_flux
 
     derivative_split = CuArray{Float64}(dg.basis.derivative_split)
@@ -799,13 +799,13 @@ function cuda_volume_integral!(du, u, mesh::TreeMesh{3}, nonconservative_terms::
 end
 
 function cuda_volume_integral!(du, u, mesh::TreeMesh{3}, nonconservative_terms::True, equations,
-                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM)
+                               volume_integral::VolumeIntegralFluxDifferencing, dg::DGSEM, cache)
     # Wait for the unmutable MHD implementation in Trixi.jl
     return nothing
 end
 
 function cuda_volume_integral!(du, u, mesh::TreeMesh{3}, nonconservative_terms::False, equations,
-                               volume_integral::VolumeIntegralShockCapturingHG, dg::DGSEM)
+                               volume_integral::VolumeIntegralShockCapturingHG, dg::DGSEM, cache)
     return nothing
 end
 
@@ -1160,7 +1160,7 @@ function rhs_gpu!(du_cpu, u_cpu, t, mesh::TreeMesh{3}, equations, boundary_condi
     du, u = copy_to_device!(du_cpu, u_cpu)
 
     cuda_volume_integral!(du, u, mesh, have_nonconservative_terms(equations), equations,
-                          dg.volume_integral, dg)
+                          dg.volume_integral, dg, cache)
 
     cuda_prolong2interfaces!(u, mesh, equations, cache)
 

diff --git a/test/dgsem_tree_test_suite.jl b/test/dgsem_tree_test_suite.jl
@@ -29,7 +29,8 @@ function run_dgsem_tree_tests_1D(semi, tspan)
     # Test `cuda_volume_integral!`
     TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                    Trixi.have_nonconservative_terms(equations_gpu),
-                                   equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                   equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                   cache_gpu)
     Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                 equations, solver.volume_integral, solver, cache)
     @test CUDA.@allowscalar du_gpu ≈ du
@@ -117,7 +118,8 @@ function run_dgsem_tree_tests_2D(semi, tspan)
     # Test `cuda_volume_integral!`
     TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                    Trixi.have_nonconservative_terms(equations_gpu),
-                                   equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                   equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                   cache_gpu)
     Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                 equations, solver.volume_integral, solver, cache)
     @test CUDA.@allowscalar du_gpu ≈ du
@@ -228,7 +230,8 @@ function run_dgsem_tree_tests_3D(semi, tspan)
     # Test `cuda_volume_integral!`
     TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                    Trixi.have_nonconservative_terms(equations_gpu),
-                                   equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                   equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                   cache_gpu)
     Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                 equations, solver.volume_integral, solver, cache)
     @test CUDA.@allowscalar du_gpu ≈ du

diff --git a/test/test_adevction_mortar.jl b/test/test_adevction_mortar.jl
@@ -57,7 +57,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du
@@ -188,7 +189,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du

diff --git a/test/test_advection_basic.jl b/test/test_advection_basic.jl
@@ -54,7 +54,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du
@@ -157,7 +158,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du
@@ -283,7 +285,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du

diff --git a/test/test_euler_ec.jl b/test/test_euler_ec.jl
@@ -55,7 +55,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du
@@ -160,7 +161,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du
@@ -287,7 +289,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du

diff --git a/test/test_euler_source_terms.jl b/test/test_euler_source_terms.jl
@@ -55,7 +55,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du
@@ -158,7 +159,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du
@@ -286,7 +288,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du

diff --git a/test/test_eulermulti_ec.jl b/test/test_eulermulti_ec.jl
@@ -57,7 +57,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du
@@ -163,7 +164,8 @@ isdir(outdir) && rm(outdir, recursive = true)
         # Test `cuda_volume_integral!`
         TrixiGPU.cuda_volume_integral!(du_gpu, u_gpu, mesh_gpu,
                                        Trixi.have_nonconservative_terms(equations_gpu),
-                                       equations_gpu, solver_gpu.volume_integral, solver_gpu)
+                                       equations_gpu, solver_gpu.volume_integral, solver_gpu,
+                                       cache_gpu)
         Trixi.calc_volume_integral!(du, u, mesh, Trixi.have_nonconservative_terms(equations),
                                     equations, solver.volume_integral, solver, cache)
         @test CUDA.@allowscalar du_gpu ≈ du