use mapreduce from AcceleratedKernels in calc_max_scaled_speed

Author: vchuravy

Project.toml

@@ -4,6 +4,7 @@ version = "0.15.10-DEV"
 authors = ["Michael Schlottke-Lakemper <michael.schlottke-lakemper@uni-a.de>", "Gregor Gassner <ggassner@uni-koeln.de>", "Hendrik Ranocha <mail@ranocha.de>", "Andrew R. Winters <andrew.ross.winters@liu.se>", "Jesse Chan <jesse.chan@rice.edu>", "Andrés Rueda-Ramírez <am.rueda@upm.es>"]
 
 [deps]
+AcceleratedKernels = "6a4ca0a5-0e36-4168-a932-d9be78d558f1"
 Accessors = "7d9f7c33-5ae7-4f3b-8dc6-eff91059b697"
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 CodeTracking = "da1fd8a2-8d9e-5ec2-8556-3022fb5608a2"
@@ -69,6 +70,7 @@ TrixiNLsolveExt = "NLsolve"
 TrixiSparseConnectivityTracerExt = "SparseConnectivityTracer"
 
 [compat]
+AcceleratedKernels = "0.4.3"
 Accessors = "0.1.42"
 Adapt = "4.3"
 CUDA = "5.8.2"

src/Trixi.jl

@@ -62,6 +62,7 @@ using ForwardDiff: ForwardDiff
 using HDF5: HDF5, h5open, attributes, create_dataset, datatype, dataspace
 using KernelAbstractions: KernelAbstractions, @index, @kernel, get_backend, Backend,
                           allocate
+using AcceleratedKernels: AcceleratedKernels
 using LinearMaps: LinearMap
 if _PREFERENCE_LOOPVECTORIZATION
     using LoopVectorization: LoopVectorization, @turbo, indices

src/callbacks_step/stepsize.jl

@@ -207,26 +207,18 @@ function calc_max_scaled_speed(backend::Backend, u, mesh, constant_speed, equati
     @unpack contravariant_vectors, inverse_jacobian = cache.elements
 
     num_elements = nelements(dg, cache)
-    max_scaled_speeds = allocate(backend, eltype(u), num_elements)
-
-    kernel! = max_scaled_speed_KAkernel!(backend)
-    kernel!(max_scaled_speeds, u, typeof(mesh), constant_speed, equations, dg,
-            contravariant_vectors,
-            inverse_jacobian;
-            ndrange = num_elements)
-
-    return maximum(max_scaled_speeds)
-end
+    init = neutral = AcceleratedKernels.neutral_element(Base.max, eltype(u))
+
+    # Provide a custom neutral and init element since we "reduce" over 1:num_elements
+    max_scaled_speed = AcceleratedKernels.mapreduce(Base.max, 1:num_elements, backend; init, neutral) do element
+        max_scaled_speed_per_element(u, mesh, constant_speed,
+                                     equations, dg,
+                                     contravariant_vectors,
+                                     inverse_jacobian,
+                                     element)
+    end
 
-@kernel function max_scaled_speed_KAkernel!(max_scaled_speeds, u, MeshT, constant_speed,
-                                            equations,
-                                            dg, contravariant_vectors, inverse_jacobian)
-    element = @index(Global)
-    max_scaled_speeds[element] = max_scaled_speed_per_element(u, MeshT, constant_speed,
-                                                              equations, dg,
-                                                              contravariant_vectors,
-                                                              inverse_jacobian,
-                                                              element)
+    return max_scaled_speed
 end
 
 include("stepsize_dg1d.jl")