Reduce use of DataLayouts internals in ClimaCore

ext/cuda/cuda_utils.jl

@@ -37,15 +37,15 @@ to benchmark compare against auto-determined threads/blocks (if `auto=false`).
 function auto_launch!(
     f!::F!,
     args,
-    data;
+    nitems::Union{Integer, Nothing} = nothing;
     auto = false,
     threads_s = nothing,
     blocks_s = nothing,
     always_inline = true,
     caller = :unknown,
 ) where {F!}
     if auto
-        nitems = get_n_items(data)
+        @assert !isnothing(nitems)
         if nitems ≥ 0
             kernel = CUDA.@cuda always_inline = true launch = false f!(args...)
             config = CUDA.launch_configuration(kernel.fun)
@@ -64,7 +64,7 @@ function auto_launch!(
         # CUDA.registers(kernel) > 50 || return nothing # for debugging
         # occursin("single_field_solve_kernel", string(nameof(F!))) || return nothing
         if !haskey(reported_stats, key)
-            nitems = get_n_items(data)
+            @assert !isnothing(nitems)
             kernel = CUDA.@cuda always_inline = true launch = false f!(args...)
             config = CUDA.launch_configuration(kernel.fun)
             threads = min(nitems, config.threads)

ext/cuda/data_layouts_copyto.jl

@@ -23,8 +23,7 @@ function Base.copyto!(
     if Nh > 0
         auto_launch!(
             knl_copyto!,
-            (dest, bc),
-            dest;
+            (dest, bc);
             threads_s = (Nij, Nij),
             blocks_s = (Nh, 1),
         )
@@ -42,8 +41,7 @@ function Base.copyto!(
         Nv_blocks = cld(Nv, Nv_per_block)
         auto_launch!(
             knl_copyto!,
-            (dest, bc),
-            dest;
+            (dest, bc);
             threads_s = (Nij, Nij, Nv_per_block),
             blocks_s = (Nh, Nv_blocks),
         )
@@ -59,8 +57,7 @@ function Base.copyto!(
     if Nv > 0
         auto_launch!(
             knl_copyto!,
-            (dest, bc),
-            dest;
+            (dest, bc);
             threads_s = (1, 1),
             blocks_s = (1, Nv),
         )
@@ -73,13 +70,7 @@ function Base.copyto!(
     bc::DataLayouts.BroadcastedUnionDataF{S},
     ::ToCUDA,
 ) where {S}
-    auto_launch!(
-        knl_copyto!,
-        (dest, bc),
-        dest;
-        threads_s = (1, 1),
-        blocks_s = (1, 1),
-    )
+    auto_launch!(knl_copyto!, (dest, bc); threads_s = (1, 1), blocks_s = (1, 1))
     return dest
 end
 
@@ -100,7 +91,8 @@ function cuda_copyto!(dest::AbstractData, bc)
     (_, _, Nv, _, Nh) = DataLayouts.universal_size(dest)
     us = DataLayouts.UniversalSize(dest)
     if Nv > 0 && Nh > 0
-        auto_launch!(knl_copyto_flat!, (dest, bc, us), dest; auto = true)
+        nitems = prod(DataLayouts.universal_size(dest))
+        auto_launch!(knl_copyto_flat!, (dest, bc, us), nitems; auto = true)
     end
     return dest
 end

ext/cuda/data_layouts_fill.jl

@@ -14,7 +14,8 @@ function cuda_fill!(dest::AbstractData, val)
     (_, _, Nv, _, Nh) = DataLayouts.universal_size(dest)
     us = DataLayouts.UniversalSize(dest)
     if Nv > 0 && Nh > 0
-        auto_launch!(knl_fill_flat!, (dest, val, us), dest; auto = true)
+        nitems = prod(DataLayouts.universal_size(dest))
+        auto_launch!(knl_fill_flat!, (dest, val, us), nitems; auto = true)
     end
     return dest
 end

ext/cuda/data_layouts_fused_copyto.jl

@@ -50,8 +50,7 @@ function fused_copyto!(
         Nv_blocks = cld(Nv, Nv_per_block)
         auto_launch!(
             knl_fused_copyto!,
-            (fmbc,),
-            dest1;
+            (fmbc,);
             threads_s = (Nij, Nij, Nv_per_block),
             blocks_s = (Nh, Nv_blocks),
         )
@@ -68,8 +67,7 @@ function fused_copyto!(
     if Nh > 0
         auto_launch!(
             knl_fused_copyto!,
-            (fmbc,),
-            dest1;
+            (fmbc,);
             threads_s = (Nij, Nij),
             blocks_s = (Nh, 1),
         )
@@ -85,8 +83,7 @@ function fused_copyto!(
     if Nv > 0 && Nh > 0
         auto_launch!(
             knl_fused_copyto!,
-            (fmbc,),
-            dest1;
+            (fmbc,);
             threads_s = (1, 1),
             blocks_s = (Nh, Nv),
         )
@@ -101,8 +98,7 @@ function fused_copyto!(
 ) where {S}
     auto_launch!(
         knl_fused_copyto!,
-        (fmbc,),
-        dest1;
+        (fmbc,);
         threads_s = (1, 1),
         blocks_s = (1, 1),
     )

ext/cuda/data_layouts_mapreduce.jl

@@ -28,7 +28,7 @@ function mapreduce_cuda(
     pdata = parent(data)
     T = eltype(pdata)
     (Ni, Nj, Nk, Nv, Nh) = size(data)
-    Nf = div(length(pdata), prod(size(data))) # length of field dimension
+    Nf = DataLayouts.ncomponents(data) # length of field dimension
     pwt = parent(weighted_jacobian)
 
     nitems = Nv * Ni * Nj * Nk * Nh

ext/cuda/limiters.jl

@@ -21,39 +21,26 @@ function compute_element_bounds!(
     ρ,
     ::ClimaComms.CUDADevice,
 )
-    S = size(Fields.field_values(ρ))
-    (Ni, Nj, _, Nv, Nh) = S
+    (_, _, Nv, _, Nh) = DataLayouts.universal_size(ρ)
     nthreads, nblocks = config_threadblock(Nv, Nh)
 
     args = (
         limiter,
         Fields.field_values(Operators.strip_space(ρq, axes(ρq))),
         Fields.field_values(Operators.strip_space(ρ, axes(ρ))),
-        Nv,
-        Nh,
-        Val(Ni),
-        Val(Nj),
     )
     auto_launch!(
         compute_element_bounds_kernel!,
-        args,
-        ρ;
+        args;
         threads_s = nthreads,
         blocks_s = nblocks,
     )
     return nothing
 end
 
 
-function compute_element_bounds_kernel!(
-    limiter,
-    ρq,
-    ρ,
-    Nv,
-    Nh,
-    ::Val{Ni},
-    ::Val{Nj},
-) where {Ni, Nj}
+function compute_element_bounds_kernel!(limiter, ρq, ρ)
+    (Ni, Nj, Nv, _, Nh) = DataLayouts.universal_size(ρ)
     n = (Nv, Nh)
     tidx = thread_index()
     @inbounds if valid_range(tidx, prod(n))
@@ -88,21 +75,18 @@ function compute_neighbor_bounds_local!(
     ::ClimaComms.CUDADevice,
 )
     topology = Spaces.topology(axes(ρ))
-    Ni, Nj, _, Nv, Nh = size(Fields.field_values(ρ))
+    us = DataLayouts.UniversalSize(ρ)
+    (Ni, Nj, Nv, _, Nh) = DataLayouts.universal_size(us)
     nthreads, nblocks = config_threadblock(Nv, Nh)
     args = (
         limiter,
         topology.local_neighbor_elem,
         topology.local_neighbor_elem_offset,
-        Nv,
-        Nh,
-        Val(Ni),
-        Val(Nj),
+        us,
     )
     auto_launch!(
         compute_neighbor_bounds_local_kernel!,
-        args,
-        ρ;
+        args;
         threads_s = nthreads,
         blocks_s = nblocks,
     )
@@ -112,12 +96,9 @@ function compute_neighbor_bounds_local_kernel!(
     limiter,
     local_neighbor_elem,
     local_neighbor_elem_offset,
-    Nv,
-    Nh,
-    ::Val{Ni},
-    ::Val{Nj},
-) where {Ni, Nj}
-
+    us::DataLayouts.UniversalSize,
+)
+    (Ni, Nj, Nv, _, Nh) = DataLayouts.universal_size(us)
     n = (Nv, Nh)
     tidx = thread_index()
     @inbounds if valid_range(tidx, prod(n))
@@ -147,27 +128,24 @@ function apply_limiter!(
     ::ClimaComms.CUDADevice,
 )
     ρq_data = Fields.field_values(ρq)
-    (Ni, Nj, _, Nv, Nh) = size(ρq_data)
-    Nf = DataLayouts.ncomponents(ρq_data)
+    us = DataLayouts.UniversalSize(ρq_data)
+    (Ni, Nj, _, Nv, Nh) = DataLayouts.universal_size(us)
     maxiter = Ni * Nj
+    Nf = DataLayouts.ncomponents(ρq_data)
     WJ = Spaces.local_geometry_data(axes(ρq)).WJ
     nthreads, nblocks = config_threadblock(Nv, Nh)
     args = (
         limiter,
         Fields.field_values(Operators.strip_space(ρq, axes(ρq))),
         Fields.field_values(Operators.strip_space(ρ, axes(ρ))),
         WJ,
-        Nv,
-        Nh,
+        us,
         Val(Nf),
-        Val(Ni),
-        Val(Nj),
         Val(maxiter),
     )
     auto_launch!(
         apply_limiter_kernel!,
-        args,
-        ρ;
+        args;
         threads_s = nthreads,
         blocks_s = nblocks,
     )
@@ -179,15 +157,13 @@ function apply_limiter_kernel!(
     ρq_data,
     ρ_data,
     WJ_data,
-    Nv,
-    Nh,
+    us::DataLayouts.UniversalSize,
     ::Val{Nf},
-    ::Val{Ni},
-    ::Val{Nj},
     ::Val{maxiter},
-) where {Nf, Ni, Nj, maxiter}
+) where {Nf, maxiter}
     (; q_bounds_nbr, rtol) = limiter
     converged = true
+    (Ni, Nj, Nv, _, Nh) = DataLayouts.universal_size(us)
     n = (Nv, Nh)
     tidx = thread_index()
     @inbounds if valid_range(tidx, prod(n))