Automatically add padding to make aligned loads possible.

Author: efaulhaber

src/general/semidiscretization.jl

@@ -94,10 +94,21 @@ function Semidiscretization(systems::Union{AbstractSystem, Nothing}...;
 
     sizes_u = [u_nvariables(system) * n_integrated_particles(system)
                for system in systems]
-    ranges_u = Tuple((sum(sizes_u[1:(i - 1)]) + 1):sum(sizes_u[1:i])
-                     for i in eachindex(sizes_u))
     sizes_v = [v_nvariables(system) * n_integrated_particles(system)
                for system in systems]
+
+    # Align sizes to 64 bytes by adding padding if necessary.
+    # This ensures that aligned loads can be used on the integration arrays, which can
+    # significantly improve performance on GPUs. Performance benefits on CPUs remain
+    # to be investigated.
+    for i in eachindex(systems)
+        block_size = div(64, sizeof(eltype(systems[i])))
+        sizes_u[i] = div(sizes_u[i], block_size, RoundUp) * block_size
+        sizes_v[i] = div(sizes_v[i], block_size, RoundUp) * block_size
+    end
+
+    ranges_u = Tuple((sum(sizes_u[1:(i - 1)]) + 1):sum(sizes_u[1:i])
+                     for i in eachindex(sizes_u))
     ranges_v = Tuple((sum(sizes_v[1:(i - 1)]) + 1):sum(sizes_v[1:i])
                      for i in eachindex(sizes_v))
 

src/schemes/fluid/weakly_compressible_sph/rhs.jl

@@ -186,14 +186,10 @@ use_aligned_vrho_load(v, system, density_calculator) = false
 #   The unaligned version `vload` does not produce wide load instructions on GPUs.
 function use_aligned_vrho_load(v::AbstractGPUArray, system, ::ContinuityDensity)
     if !can_use_aligned_load(v, 4)
-        # If aligned loads are possible for the problem, but not allowed due to alignment,
-        # we don't fall back to the non-SIMD version and throw an error instead.
-        # This is likely a configuration error (see the error message below), and notifying
-        # the user is better than silently falling back to slower loads and thus
-        # non-deterministic performance.
-        error("on GPUs in 3D, all WCSPH systems with `ContinuityDensity` must be the " *
-              "first systems in the semidiscretization to ensure that their integration " *
-              "arrays are aligned for SIMD loads.")
+        # Aligned loads should always be possible on GPUs because the slices of `v_ode`
+        # are aligned to 64 bytes in `Semidiscretization` and arrays on GPUs are always
+        # aligned to full pages.
+        error("illegal alignment of `v` integration array. Please report this issue.")
     end
 
     return true