Create cache for GPU arrays to optimize data transfer (#53)

* Start * Rename API * Add comments * Fix typos * Complete 1D * Complete 2D * Complete 3D
trixi-gpu · Sep 24, 2024 · 80a3b75 · 80a3b75
1 parent d8deac8
commit 80a3b75
Show file tree

Hide file tree

Showing 37 changed files with 387 additions and 178 deletions.
diff --git a/README.md b/README.md
@@ -77,7 +77,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition_convergen
 ###############################################################################
 # ODE solvers, callbacks etc.
 
-ode = semidiscretize_gpu(semi, (0.0, 1.0)) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, (0.0, 1.0)) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/advection_basic_1d.jl b/examples/advection_basic_1d.jl
@@ -28,7 +28,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition_convergen
 # ODE solvers, callbacks etc.
 
 # Create ODE problem with time span from 0.0 to 1.0
-ode = semidiscretize_gpu(semi, (0.0, 1.0)) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, (0.0, 1.0)) # from TrixiCUDA.jl
 
 # At the beginning of the main loop, the SummaryCallback prints a summary of the simulation setup
 # and resets the timers

diff --git a/examples/advection_basic_2d.jl b/examples/advection_basic_2d.jl
@@ -28,7 +28,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition_convergen
 # ODE solvers, callbacks etc.
 
 # Create ODE problem with time span from 0.0 to 1.0
-ode = semidiscretize_gpu(semi, (0.0, 1.0)) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, (0.0, 1.0)) # from TrixiCUDA.jl
 
 # At the beginning of the main loop, the SummaryCallback prints a summary of the simulation setup
 # and resets the timers

diff --git a/examples/advection_basic_3d.jl b/examples/advection_basic_3d.jl
@@ -28,7 +28,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition_convergen
 # ODE solvers, callbacks etc.
 
 # Create ODE problem with time span from 0.0 to 1.0
-ode = semidiscretize_gpu(semi, (0.0, 1.0)) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, (0.0, 1.0)) # from TrixiCUDA.jl
 
 # At the beginning of the main loop, the SummaryCallback prints a summary of the simulation setup
 # and resets the timers

diff --git a/examples/advection_mortar_2d.jl b/examples/advection_mortar_2d.jl
@@ -27,7 +27,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 1.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/advection_mortar_3d.jl b/examples/advection_mortar_3d.jl
@@ -30,7 +30,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 5.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/euler_ec_1d.jl b/examples/euler_ec_1d.jl
@@ -25,7 +25,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 0.4)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/euler_ec_2d.jl b/examples/euler_ec_2d.jl
@@ -27,7 +27,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 0.4)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/euler_ec_3d.jl b/examples/euler_ec_3d.jl
@@ -26,7 +26,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 0.4)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/euler_shockcapturing_1d.jl b/examples/euler_shockcapturing_1d.jl
@@ -35,7 +35,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 1.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/euler_shockcapturing_2d.jl b/examples/euler_shockcapturing_2d.jl
@@ -35,7 +35,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 1.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/euler_shockcapturing_3d.jl b/examples/euler_shockcapturing_3d.jl
@@ -37,7 +37,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 0.4)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/euler_source_terms_1d.jl b/examples/euler_source_terms_1d.jl
@@ -27,7 +27,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 2.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/euler_source_terms_2d.jl b/examples/euler_source_terms_2d.jl
@@ -24,7 +24,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 2.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl   
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl   
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/euler_source_terms_3d.jl b/examples/euler_source_terms_3d.jl
@@ -26,7 +26,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 5.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/eulermulti_ec_1d.jl b/examples/eulermulti_ec_1d.jl
@@ -26,7 +26,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 0.4)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/eulermulti_ec_2d.jl b/examples/eulermulti_ec_2d.jl
@@ -26,7 +26,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 0.4)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/hypdiff_nonperiodic_1d.jl b/examples/hypdiff_nonperiodic_1d.jl
@@ -29,7 +29,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 5.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/hypdiff_nonperiodic_2d.jl b/examples/hypdiff_nonperiodic_2d.jl
@@ -32,7 +32,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 5.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/hypdiff_nonperiodic_3d.jl b/examples/hypdiff_nonperiodic_3d.jl
@@ -33,7 +33,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 5.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/shallowwater_dirichlet_1d.jl b/examples/shallowwater_dirichlet_1d.jl
@@ -39,7 +39,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 1.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/shallowwater_dirichlet_2d.jl b/examples/shallowwater_dirichlet_2d.jl
@@ -39,7 +39,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 1.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/shallowwater_ec_1d.jl b/examples/shallowwater_ec_1d.jl
@@ -65,7 +65,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solver
 
 tspan = (0.0, 2.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 ###############################################################################
 # Callbacks

diff --git a/examples/shallowwater_ec_2d.jl b/examples/shallowwater_ec_2d.jl
@@ -37,7 +37,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver)
 # ODE solver
 
 tspan = (0.0, 2.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 ###############################################################################
 # Workaround to set a discontinuous bottom topography and initial condition for debugging and testing.

diff --git a/examples/shallowwater_source_terms_1d.jl b/examples/shallowwater_source_terms_1d.jl
@@ -36,7 +36,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 1.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/examples/shallowwater_source_terms_2d.jl b/examples/shallowwater_source_terms_2d.jl
@@ -36,7 +36,7 @@ semi = SemidiscretizationHyperbolic(mesh, equations, initial_condition, solver,
 # ODE solvers, callbacks etc.
 
 tspan = (0.0, 1.0)
-ode = semidiscretize_gpu(semi, tspan) # from TrixiCUDA.jl
+ode = semidiscretizeGPU(semi, tspan) # from TrixiCUDA.jl
 
 summary_callback = SummaryCallback()
 

diff --git a/src/TrixiCUDA.jl b/src/TrixiCUDA.jl
@@ -3,16 +3,20 @@ module TrixiCUDA
 # Include other packages that are used in TrixiCUDA.jl
 # using Reexport: @reexport
 
+using CUDA
 using CUDA: @cuda, CuArray, HostKernel,
             threadIdx, blockIdx, blockDim, similar, launch_configuration
 
-using Trixi: AbstractEquations, TreeMesh, DGSEM,
+using Trixi: AbstractEquations, True, False,
+             TreeMesh, DGSEM,
              BoundaryConditionPeriodic, SemidiscretizationHyperbolic,
              VolumeIntegralWeakForm, VolumeIntegralFluxDifferencing, VolumeIntegralShockCapturingHG,
-             flux, ntuple, nvariables,
-             True, False,
+             LobattoLegendreMortarL2,
+             flux, ntuple, nvariables, nnodes, nelements, nmortars,
+             local_leaf_cells, init_elements, init_interfaces, init_boundaries, init_mortars,
              wrap_array, compute_coefficients, have_nonconservative_terms,
              boundary_condition_periodic,
+             digest_boundary_conditions, check_periodicity_mesh_boundary_conditions,
              set_log_type!, set_sqrt_type!
 
 import Trixi: get_node_vars, get_node_coords, get_surface_node_vars
@@ -25,12 +29,15 @@ using StaticArrays: SVector
 
 # Include other source files
 include("auxiliary/auxiliary.jl")
+include("semidiscretization/semidiscretization.jl")
 include("solvers/solvers.jl")
 
+# Change to use the Base.log and Base.sqrt - need to be fixed to avoid outputs
 set_log_type!("log_Base")
 set_sqrt_type!("sqrt_Base")
 
 # Export the public APIs
-export semidiscretize_gpu
+export SemidiscretizationHyperbolicGPU
+export semidiscretizeGPU
 
 end
diff --git a/src/auxiliary/auxiliary.jl b/src/auxiliary/auxiliary.jl
@@ -1,2 +1,2 @@
 include("configurators.jl")
-include("helpers.jl")
+include("stable.jl")
diff --git a/src/auxiliary/helpers.jl → src/auxiliary/stable.jl b/src/auxiliary/helpers.jl → src/auxiliary/stable.jl
@@ -1,18 +1,23 @@
-# Some helper functions and function extensions from Trixi.jl.
+# Some helper functions and function extensions that are invoked by GPU 
+# kernels, mainly to ensure the functions themselves or other functions 
+# remain stable on the GPU.
 
-# Ref: `get_node_vars(u, equations, solver::DG, indices...)` in Trixi.jl
+# See also `get_node_vars(u, equations, solver::DG, indices...)` in Trixi.jl
+# `DG` type is not stable on GPU
 @inline function get_node_vars(u, equations, indices...)
     return SVector(ntuple(@inline(v->u[v, indices...]),
                           Val(nvariables(equations))))
 end
 
-# Ref: `get_node_coords(x, equations, solver::DG, indices...)` in Trixi.jl
+# See also `get_node_coords(x, equations, solver::DG, indices...)` in Trixi.jl
+# `DG` type is not stable on GPU
 @inline function get_node_coords(x, equations, indices...)
     return SVector(ntuple(@inline(idx->x[idx, indices...]),
                           Val(ndims(equations))))
 end
 
-# Ref: `get_surface_node_vars(u, equations, solver::DG, indices...)` in Trixi.jl
+# See also `get_surface_node_vars(u, equations, solver::DG, indices...)` in Trixi.jl
+# `DG` type is not stable on GPU
 @inline function get_surface_node_vars(u, equations, indices...)
     u_ll = SVector(ntuple(@inline(v->u[1, v, indices...]),
                           Val(nvariables(equations))))

diff --git a/src/semidiscretization/semidiscretization.jl b/src/semidiscretization/semidiscretization.jl
@@ -0,0 +1 @@
+include("semidiscretization_hyperbolic.jl")
diff --git a/src/semidiscretization/semidiscretization_hyperbolic.jl b/src/semidiscretization/semidiscretization_hyperbolic.jl
@@ -0,0 +1,27 @@
+# This file is part of the package `Semidiscretizations.jl`.
+
+function SemidiscretizationHyperbolicGPU(mesh, equations, initial_condition, solver;
+                                         source_terms = nothing,
+                                         boundary_conditions = boundary_condition_periodic,
+                                         # `RealT` is used as real type for node locations etc.
+                                         # while `uEltype` is used as element type of solutions etc.
+                                         RealT = real(solver), uEltype = RealT,
+                                         initial_cache = NamedTuple())
+    @assert ndims(mesh) == ndims(equations)
+
+    cache = (; create_cache_gpu(mesh, equations, solver, RealT, uEltype)...,
+             initial_cache...)
+    _boundary_conditions = digest_boundary_conditions(boundary_conditions, mesh, solver,
+                                                      cache)
+
+    check_periodicity_mesh_boundary_conditions(mesh, _boundary_conditions)
+
+    SemidiscretizationHyperbolic{typeof(mesh), typeof(equations),
+                                 typeof(initial_condition),
+                                 typeof(_boundary_conditions), typeof(source_terms),
+                                 typeof(solver), typeof(cache)}(mesh, equations,
+                                                                initial_condition,
+                                                                _boundary_conditions,
+                                                                source_terms, solver,
+                                                                cache)
+end