Use multithreading in row_group_slots refarray method

docs/src/lib/internals.md

@@ -15,4 +15,5 @@ gennames
 getmaxwidths
 ourshow
 ourstrwidth
+tforeach
 ```

src/dataframerow/utils.jl

@@ -338,46 +338,53 @@ function row_group_slots(cols::NTuple{N, AbstractVector},
             end
             refmap
         end
-        @inbounds for i in eachindex(groups)
-            local refs_i
-            let i=i # Workaround for julia#15276
-                refs_i = map(c -> c[i], refarrays)
-            end
-            vals = map((m, r, s, fi) -> m[r-fi+1] * s, refmaps, refs_i, strides, firstinds)
-            j = sum(vals) + 1
-            # x < 0 happens with -1 in refmap, which corresponds to missing
-            if skipmissing && any(x -> x < 0, vals)
-                j = 0
-            else
-                seen[j] = true
+        tforeach(eachindex(groups), basesize=1_000_000) do i
+            @inbounds begin
+                local refs_i
+                let i=i # Workaround for julia#15276
+                    refs_i = map(c -> c[i], refarrays)
+                end
+                vals = map((m, r, s, fi) -> m[r-fi+1] * s, refmaps, refs_i, strides, firstinds)
+                j = sum(vals) + 1
+                # x < 0 happens with -1 in refmap, which corresponds to missing
+                if skipmissing && any(x -> x < 0, vals)
+                    j = 0
+                else
+                    seen[j] = true
+                end
+                groups[i] = j
             end
-            groups[i] = j
         end
     else
-        @inbounds for i in eachindex(groups)
-            local refs_i
-            let i=i # Workaround for julia#15276
-                refs_i = map(refarrays, missinginds) do ref, missingind
-                    r = Int(ref[i])
-                    if skipmissing
-                        return r == missingind ? -1 : (r > missingind ? r-1 : r)
-                    else
-                        return r
+        tforeach(eachindex(groups), basesize=1_000_000) do i
+            @inbounds begin
+                local refs_i
+                let i=i # Workaround for julia#15276
+                    refs_i = map(refarrays, missinginds) do ref, missingind
+                        r = Int(ref[i])
+                        if skipmissing
+                            return r == missingind ? -1 : (r > missingind ? r-1 : r)
+                        else
+                            return r
+                        end
                     end
                 end
+                vals = map((r, s, fi) -> (r-fi) * s, refs_i, strides, firstinds)
+                j = sum(vals) + 1
+                # x < 0 happens with -1, which corresponds to missing
+                if skipmissing && any(x -> x < 0, vals)
+                    j = 0
+                else
+                    seen[j] = true
+                end
+                groups[i] = j
             end
-            vals = map((r, s, fi) -> (r-fi) * s, refs_i, strides, firstinds)
-            j = sum(vals) + 1
-            # x < 0 happens with -1, which corresponds to missing
-            if skipmissing && any(x -> x < 0, vals)
-                j = 0
-            else
-                seen[j] = true
-            end
-            groups[i] = j
         end
     end
-    if !all(seen) # Compress group indices to remove unused ones
+    # If some groups are unused, compress group indices to drop them
+    # sum(seen) is faster than all(seen) when not short-circuiting,
+    # and short-circuit would only happen in the slower case anyway
+    if sum(seen) < length(seen)
         oldngroups = ngroups
         remap = zeros(Int, ngroups)
         ngroups = 0

src/other/utils.jl

@@ -83,3 +83,48 @@ else
 end
 
 funname(c::ComposedFunction) = Symbol(funname(c.outer), :_, funname(c.inner))
+
+# Compute chunks of indices, each with at least `basesize` entries
+# This method ensures balanced sizes by avoiding a small last chunk
+function split_indices(len::Integer, basesize::Integer)
+    len′ = Int64(len) # Avoid overflow on 32-bit machines
+    np = max(1, div(len, basesize))
+    return (Int(1 + ((i - 1) * len′) ÷ np):Int((i * len′) ÷ np) for i in 1:np)
+end
+
+"""
+    tforeach(f, x::AbstractArray; basesize::Integer)
+
+Apply function `f` to each entry in `x` in parallel, spawning
+one separate task for each block of at least `basesize` entries.
+
+A number of task higher than `Threads.nthreads()` may be spawned,
+since that can allow for a more efficient load balancing in case
+some threads are busy (nested parallelism).
+"""
+function tforeach(f, x::AbstractArray; basesize::Integer)
+    @assert firstindex(x) == 1
+
+    @static if VERSION >= v"1.4"
+        nt = Threads.nthreads()
+        len = length(x)
+        if nt > 1 && len > basesize
+            @sync for p in split_indices(len, basesize)
+                Threads.@spawn begin
+                    for i in p
+                        f(@inbounds x[i])
+                    end
+                end
+            end
+        else
+            for i in eachindex(x)
+                f(@inbounds x[i])
+            end
+        end
+    else
+        for i in eachindex(x)
+            f(@inbounds x[i])
+        end
+    end
+    return
+end

test/grouping.jl

@@ -3831,4 +3831,22 @@ end
                         ((x, y, z) -> x[1] <= 5 ? unwrap(y[1]) : unwrap(z[1])) => :res)
 end
 
+@testset "groupby multithreading" begin
+    for x in (PooledArray(rand(1:10, 1_100_000)),
+              PooledArray(rand([1:9; missing], 1_100_000))),
+        y in (PooledArray(rand(["a", "b", "c", "d"], 1_100_000)),
+              PooledArray(rand(["a"; "b"; "c"; missing], 1_100_000)))
+        df = DataFrame(x=x, y=y)
+
+        # Checks are done by groupby_checked
+        @test length(groupby_checked(df, :x)) == 10
+        @test length(groupby_checked(df, :x, skipmissing=true)) ==
+            length(unique(skipmissing(x)))
+
+        @test length(groupby_checked(df, [:x, :y])) == 40
+        @test length(groupby_checked(df, [:x, :y], skipmissing=true)) ==
+            length(unique(skipmissing(x))) * length(unique(skipmissing(y)))
+    end
+end
+
 end # module

test/utils.jl

@@ -101,4 +101,28 @@ end
     @test fetch(t) === true
 end
 
+@testset "split_indices" begin
+    for len in 0:12
+        basesize = 10
+        x = DataFrames.split_indices(len, basesize)
+
+        @test length(x) == max(1, div(len, basesize))
+        @test reduce(vcat, x) === 1:len
+        vmin, vmax = extrema(length, x)
+        @test vmin + 1 == vmax || vmin == vmax
+        @test len < basesize || vmin >= basesize
+    end
+
+    # Check overflow on 32-bit
+    len = typemax(Int32)
+    basesize = 100_000_000
+    x = collect(DataFrames.split_indices(len, basesize))
+    @test length(x) == div(len, basesize)
+    @test x[1][1] === 1
+    @test x[end][end] === Int(len)
+    vmin, vmax = extrema(length, x)
+    @test vmin + 1 == vmax || vmin == vmax
+    @test len < basesize || vmin >= basesize
+end
+
 end # module