Merge pull request #1905 from CliMA/ck/benchmark_script

Add index-swapping benchmark script

Author: charleskawczynski

benchmarks/scripts/index_swapping.jl

@@ -0,0 +1,213 @@
+#=
+julia --project=.buildkite
+julia -g2 --check-bounds=yes --project=.buildkite
+using Revise; include(joinpath("benchmarks", "scripts", "index_swapping.jl"))
+
+# Info:
+This script compares the performance
+of our universal index support (for
+mixed DataLayout operations) against
+specialized index support for uniform
+DataLayout operations.
+
+In particular,
+ - `at_dot_call!` is a reference for the speed of light
+                  we could achieve on the hardware, as
+                  memory coalescence comes for free on
+                  vectors (as opposed to arrays).
+ - `custom_kernel_bc!(; swap = 0)` mimics our specialized operations
+ - `custom_kernel_bc!(; swap = 1)` mimics our generalized pointwise operations
+ - `custom_kernel_bc!(; swap = 2)` mimics our generalized stencil operations
+
+# Benchmark results:
+
+Clima A100
+```
+at_dot_call!($X_vector, $Y_vector):
+     6 milliseconds, 19 microseconds
+custom_kernel_bc!($X_array, $Y_array, $uss, swap = 0):
+     6 milliseconds, 329 microseconds
+custom_kernel_bc!($X_array, $Y_array, $uss, swap = 1):
+     14 milliseconds, 232 microseconds
+custom_kernel_bc!($X_array, $Y_array, $uss, swap = 2):
+     15 milliseconds, 960 microseconds
+```
+=#
+
+#! format: off
+import CUDA
+using BenchmarkTools, Dates
+using LazyBroadcast: @lazy
+ArrayType = CUDA.CuArray;
+# ArrayType = identity;
+
+if ArrayType === identity
+    macro pretty_belapsed(expr)
+        return quote
+            println($(string(expr)), ":")
+            print("     ")
+            print_time_and_units(BenchmarkTools.@belapsed(esc($expr)))
+        end
+    end
+else
+    macro pretty_belapsed(expr)
+        return quote
+            println($(string(expr)), ":")
+            print("     ")
+            print_time_and_units(
+                BenchmarkTools.@belapsed(CUDA.@sync((esc($expr))))
+            )
+        end
+    end
+    macro pretty_elapsed(expr)
+        return quote
+            println($(string(expr)), ":")
+            print("     ")
+            print_time_and_units(
+                BenchmarkTools.@elapsed(CUDA.@sync((esc($expr))))
+            )
+        end
+    end
+end
+print_time_and_units(x) = println(time_and_units_str(x))
+time_and_units_str(x::Real) =
+    trunc_time(string(compound_period(x, Dates.Second)))
+function compound_period(x::Real, ::Type{T}) where {T <: Dates.Period}
+    nf = Dates.value(convert(Dates.Nanosecond, T(1)))
+    ns = Dates.Nanosecond(ceil(x * nf))
+    return Dates.canonicalize(Dates.CompoundPeriod(ns))
+end
+trunc_time(s::String) = count(',', s) > 1 ? join(split(s, ",")[1:2], ",") : s
+foo(x1, x2, x3) = x1
+function at_dot_call!(X, Y)
+    (; x1, x2, x3) = X
+    (; y1) = Y
+    for i in 1:100 # reduce variance / impact of launch latency
+        @. y1 = foo(x1, x2, x3) # 3 reads, 1 write
+    end
+    return nothing
+end;
+
+struct UniversalSizesStatic{Nv, Nij, Nh} end
+
+get_Nv(::UniversalSizesStatic{Nv}) where {Nv} = Nv
+get_Nij(::UniversalSizesStatic{Nv, Nij}) where {Nv, Nij} = Nij
+get_Nh(::UniversalSizesStatic{Nv, Nij, Nh}) where {Nv, Nij, Nh} = Nh
+get_N(us::UniversalSizesStatic{Nv, Nij}) where {Nv, Nij} = prod((Nv,Nij,Nij,1,get_Nh(us)))
+UniversalSizesStatic(Nv, Nij, Nh) = UniversalSizesStatic{Nv, Nij, Nh}()
+using Test
+
+function custom_kernel_bc!(X, Y, us::UniversalSizesStatic; swap=0, printtb=false)
+    (; x1, x2, x3) = X
+    (; y1) = Y
+    bc = @lazy @. y1 = foo(x1, x2, x3)
+    @assert !(y1 isa Array)
+    f = if swap==0
+        custom_kernel_knl_bc_no_swap!
+    elseif swap == 1
+        custom_kernel_knl_bc_swap!
+    elseif swap == 2
+        custom_kernel_knl_bc_2swap!
+    else
+        error("oops")
+    end
+    kernel =
+        CUDA.@cuda always_inline = true launch = false f(
+            y1,
+            bc,
+            us,
+        )
+    N = get_N(us)
+    config = CUDA.launch_configuration(kernel.fun)
+    threads = min(N, config.threads)
+    blocks = cld(N, threads)
+    printtb && @show blocks, threads
+    for i in 1:100 # reduce variance / impact of launch latency
+        kernel(y1, bc,us; threads, blocks)
+    end
+    return nothing
+end;
+
+# Mimics how indexing works in generalized pointwise kernels
+function custom_kernel_knl_bc_swap!(y1, bc, us)
+    @inbounds begin
+        tidx = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x
+        if tidx ≤ get_N(us)
+            n = (get_Nij(us), get_Nij(us), 1, get_Nv(us), get_Nh(us))
+            GCI = CartesianIndices(map(x -> Base.OneTo(x), n))[tidx]
+            # Perform index swap (as in `getindex(::AbstractData, ::CartesianIndex)`)
+            i, j, _, v, h = GCI.I
+            CI = CartesianIndex(v, i, j, 1, h)
+            y1[CI] = bc[CI]
+        end
+    end
+    return nothing
+end
+
+# Mimics how indexing works in specialized kernels
+function custom_kernel_knl_bc_no_swap!(y1, bc, us)
+    @inbounds begin
+        tidx = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x
+        if tidx ≤ get_N(us)
+            n = (get_Nv(us), get_Nij(us), get_Nij(us), 1, get_Nh(us))
+            CI = CartesianIndices(map(x -> Base.OneTo(x), n))[tidx]
+            y1[CI] = bc[CI] # requires special broadcasted index support
+        end
+    end
+    return nothing
+end
+# Mimics how indexing works in generalized stencil kernels
+function custom_kernel_knl_bc_2swap!(y1, bc, us)
+    @inbounds begin
+        tidx = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x
+        if tidx ≤ get_N(us)
+            # We start with a VIJFH-specific CartesianIndex
+            n = (get_Nv(us), get_Nij(us), get_Nij(us), 1, get_Nh(us))
+            CIK = CartesianIndices(map(x -> Base.OneTo(x), n))[tidx] # data-specific in kernel
+
+            # Swap in `getidx`
+            (v, i, j, _, h) = CIK.I
+            GCI = CartesianIndex(i, j, 1, v, h)
+
+            # Swap again (in `getindex(::AbstractData, ::CartesianIndex)`)
+            (i, j, _, v, h) = GCI.I
+            CI = CartesianIndex(v, i, j, 1, h)
+            y1[CI] = bc[CI]
+        end
+    end
+    return nothing
+end
+
+import Random
+function test_custom_kernel_bc!(X_array, Y_array, uss; swap)
+    Random.seed!(1234)
+    X_array.x1 .= typeof(X_array.x1)(rand(eltype(X_array.x1), size(X_array.x1)))
+    Y_array_cp = deepcopy(Y_array)
+    custom_kernel_bc!(X_array, Y_array_cp, uss; swap=0)
+    custom_kernel_bc!(X_array, Y_array, uss; swap)
+    @test all(Y_array_cp.y1 .== Y_array.y1)
+end
+
+FT = Float32;
+arr(T) = T(zeros(63,4,4,1,5400))
+X_array = (;x1 = arr(ArrayType),x2 = arr(ArrayType),x3 = arr(ArrayType));
+Y_array = (;y1 = arr(ArrayType),);
+to_vec(ξ) = (;zip(propertynames(ξ), map(θ -> vec(θ), values(ξ)))...);
+X_vector = to_vec(X_array);
+Y_vector = to_vec(Y_array);
+N = length(X_vector.x1)
+(Nv, Nij, _, _, Nh) = size(Y_array.y1);
+uss = UniversalSizesStatic(Nv, Nij, Nh);
+at_dot_call!(X_vector, Y_vector)
+custom_kernel_bc!(X_array, Y_array, uss; swap=0)
+custom_kernel_bc!(X_array, Y_array, uss; swap=1)
+custom_kernel_bc!(X_array, Y_array, uss; swap=2)
+test_custom_kernel_bc!(X_array, Y_array, uss; swap=1)
+test_custom_kernel_bc!(X_array, Y_array, uss; swap=2)
+
+@pretty_belapsed at_dot_call!($X_vector, $Y_vector)
+@pretty_belapsed custom_kernel_bc!($X_array, $Y_array, $uss, swap=0)
+@pretty_belapsed custom_kernel_bc!($X_array, $Y_array, $uss, swap=1)
+@pretty_belapsed custom_kernel_bc!($X_array, $Y_array, $uss, swap=2)
+
+#! format: on