Support multithreading in groupreduce

Keep the default to a single thread until we find a reliable way of
predicting a reasonably optimal number of threads.

Author: nalimilan

.github/workflows/ci.yml

@@ -35,6 +35,7 @@ jobs:
       - uses: actions/cache@v1
         env:
           cache-name: cache-artifacts
+          JULIA_NUM_THREADS: 2
         with:
           path: ~/.julia/artifacts
           key: ${{ runner.os }}-test-${{ env.cache-name }}-${{ hashFiles('**/Project.toml') }}

NEWS.md

@@ -1,3 +1,12 @@
+# DataFrames v1.0 Release Notes
+
+## New functionalities
+
+* `combine`, `select` and `transform` with `GroupedDataFrame` now accept
+  a `nthreads` argument which enables multithreading for some optimized
+  grouped reductions ([#2491](https://github.com/JuliaData/DataFrames.jl/pull/2491)).
+
+
 # DataFrames v0.22 Release Notes
 
 ## Breaking changes

src/DataFrames.jl

@@ -3,7 +3,7 @@ module DataFrames
 using Statistics, Printf, REPL
 using Reexport, SortingAlgorithms, Compat, Unicode, PooledArrays, CategoricalArrays
 @reexport using Missings, InvertedIndices
-using Base.Sort, Base.Order, Base.Iterators
+using Base.Sort, Base.Order, Base.Iterators, Base.Threads
 using TableTraits, IteratorInterfaceExtensions
 import LinearAlgebra: norm
 using Markdown

src/abstractdataframe/selection.jl

@@ -644,9 +644,10 @@ end
     select(df::AbstractDataFrame, args...; copycols::Bool=true, renamecols::Bool=true)
     select(args::Callable, df::DataFrame; renamecols::Bool=true)
     select(gd::GroupedDataFrame, args...; copycols::Bool=true, keepkeys::Bool=true,
-           ungroup::Bool=true, renamecols::Bool=true)
+           ungroup::Bool=true, renamecols::Bool=true, nthreads::Integer=1)
     select(f::Base.Callable, gd::GroupedDataFrame; copycols::Bool=true,
-           keepkeys::Bool=true, ungroup::Bool=true, renamecols::Bool=true)
+           keepkeys::Bool=true, ungroup::Bool=true,
+           renamecols::Bool=true, nthreads::Integer=1)
 
 Create a new data frame that contains columns from `df` or `gd` specified by
 `args` and return it. The result is guaranteed to have the same number of rows
@@ -664,6 +665,9 @@ $TRANSFORMATION_COMMON_RULES
   data frame.
 - `ungroup::Bool=true` : whether the return value of the operation on `gd` should be a data
   frame or a `GroupedDataFrame`.
+- `nthreads::Integer=1` : the number of CPU threads to use. Passing a value higher than 1
+  currently has an effect only for some optimized grouped reductions. Values higher than
+  `Threads.nthreads()` will be replaced with that value.
 
 # Examples
 ```jldoctest
@@ -858,9 +862,11 @@ end
     transform(df::AbstractDataFrame, args...; copycols::Bool=true, renamecols::Bool=true)
     transform(f::Callable, df::DataFrame; renamecols::Bool=true)
     transform(gd::GroupedDataFrame, args...; copycols::Bool=true,
-              keepkeys::Bool=true, ungroup::Bool=true, renamecols::Bool=true)
+              keepkeys::Bool=true, ungroup::Bool=true,
+              renamecols::Bool=true, nthreads::Integer=1)
     transform(f::Base.Callable, gd::GroupedDataFrame; copycols::Bool=true,
-              keepkeys::Bool=true, ungroup::Bool=true, renamecols::Bool=true)
+              keepkeys::Bool=true, ungroup::Bool=true,
+              renamecols::Bool=true, nthreads::Integer=1)
 
 Create a new data frame that contains columns from `df` or `gd` plus columns
 specified by `args` and return it. The result is guaranteed to have the same
@@ -877,6 +883,9 @@ $TRANSFORMATION_COMMON_RULES
   data frame.
 - `ungroup::Bool=true` : whether the return value of the operation on `gd` should be a data
   frame or a `GroupedDataFrame`.
+- `nthreads::Integer=1` : the number of CPU threads to use. Passing a value higher than 1
+  currently has an effect only for some optimized grouped reductions. Values higher than
+  `Threads.nthreads()` will be replaced with that value.
 
 Note that when the first argument is a `GroupedDataFrame`, `keepkeys=false`
 is needed to be able to return a different value for the grouping column:
@@ -924,9 +933,11 @@ end
     combine(df::AbstractDataFrame, args...; renamecols::Bool=true)
     combine(f::Callable, df::AbstractDataFrame; renamecols::Bool=true)
     combine(gd::GroupedDataFrame, args...;
-            keepkeys::Bool=true, ungroup::Bool=true, renamecols::Bool=true)
+            keepkeys::Bool=true, ungroup::Bool=true,
+            renamecols::Bool=true, nthreads::Integer=1)
     combine(f::Base.Callable, gd::GroupedDataFrame;
-            keepkeys::Bool=true, ungroup::Bool=true, renamecols::Bool=true)
+            keepkeys::Bool=true, ungroup::Bool=true,
+            renamecols::Bool=true, nthreads::Integer=1)
 
 Create a new data frame that contains columns from `df` or `gd` specified by
 `args` and return it. The result can have any number of rows that is determined
@@ -941,6 +952,9 @@ $TRANSFORMATION_COMMON_RULES
   data frame.
 - `ungroup::Bool=true` : whether the return value of the operation on `gd` should be a data
   frame or a `GroupedDataFrame`.
+- `nthreads::Integer=1` : the number of CPU threads to use. Passing a value higher than 1
+  currently has an effect only for some optimized grouped reductions. Values higher than
+  `Threads.nthreads()` will be replaced with that value.
 
 # Examples
 ```jldoctest

src/groupeddataframe/fastaggregates.jl

@@ -157,24 +157,84 @@ function copyto_widen!(res::AbstractVector{T}, x::AbstractVector) where T
 end
 
 function groupreduce!(res::AbstractVector, f, op, condf, adjust, checkempty::Bool,
-                      incol::AbstractVector, gd::GroupedDataFrame)
+                      incol::AbstractVector, gd::GroupedDataFrame, nthreads::Integer)
     n = length(gd)
+    groups = gd.groups
     if adjust !== nothing || checkempty
         counts = zeros(Int, n)
     end
-    groups = gd.groups
-    @inbounds for i in eachindex(incol, groups)
-        gix = groups[i]
-        x = incol[i]
-        if gix > 0 && (condf === nothing || condf(x))
-            # this check should be optimized out if U is not Any
-            if eltype(res) === Any && !isassigned(res, gix)
-                res[gix] = f(x, gix)
-            else
-                res[gix] = op(res[gix], f(x, gix))
+    nt = min(nthreads, Threads.nthreads())
+    if nt <= 1 || axes(incol) != axes(groups)
+        @inbounds for i in eachindex(incol, groups)
+            gix = groups[i]
+            x = incol[i]
+            if gix > 0 && (condf === nothing || condf(x))
+                # this check should be optimized out if eltype is not Any
+                if eltype(res) === Any && !isassigned(res, gix)
+                    res[gix] = f(x, gix)
+                else
+                    res[gix] = op(res[gix], f(x, gix))
+                end
+                if adjust !== nothing || checkempty
+                    counts[gix] += 1
+                end
+            end
+        end
+    else
+        res_vec = Vector{typeof(res)}(undef, nt)
+        # needs to be always allocated to fix type instability with @threads
+        counts_vec = Vector{Vector{Int}}(undef, nt)
+        res_vec[1] = res
+        if adjust !== nothing || checkempty
+            counts_vec[1] = counts
+        end
+        for i in 2:nt
+            res_vec[i] = copy(res)
+            if adjust !== nothing || checkempty
+                counts_vec[i] = zeros(Int, n)
             end
+        end
+        Threads.@threads for tid in 1:nt
+            res′ = res_vec[tid]
             if adjust !== nothing || checkempty
-                counts[gix] += 1
+                counts′ = counts_vec[tid]
+            end
+            start = 1 + ((tid - 1) * length(groups)) ÷ nt
+            stop = (tid * length(groups)) ÷ nt
+            @inbounds for i in start:stop
+                gix = groups[i]
+                x = incol[i]
+                if gix > 0 && (condf === nothing || condf(x))
+                    # this check should be optimized out if eltype is not Any
+                    if eltype(res′) === Any && !isassigned(res′, gix)
+                        res′[gix] = f(x, gix)
+                    else
+                        res′[gix] = op(res′[gix], f(x, gix))
+                    end
+                    if adjust !== nothing || checkempty
+                        counts′[gix] += 1
+                    end
+                end
+            end
+        end
+        for i in 2:length(res_vec)
+            resi = res_vec[i]
+            @inbounds @simd for j in eachindex(res)
+                # this check should be optimized out if eltype is not Any
+                if eltype(res) === Any
+                    if isassigned(resi, j) && isassigned(res, j)
+                        res[j] = op(res[j], resi[j])
+                    elseif isassigned(resi, j)
+                        res[j] = resi[j]
+                    end
+                else
+                    res[j] = op(res[j], resi[j])
+                end
+            end
+        end
+        if adjust !== nothing || checkempty
+            for i in 2:length(counts_vec)
+                counts .+= counts_vec[i]
             end
         end
     end
@@ -218,26 +278,31 @@ end
 
 # function barrier works around type instability of groupreduce_init due to applicable
 groupreduce(f, op, condf, adjust, checkempty::Bool,
-            incol::AbstractVector, gd::GroupedDataFrame) =
+            incol::AbstractVector, gd::GroupedDataFrame,
+            nthreads::Integer) =
     groupreduce!(groupreduce_init(op, condf, adjust, incol, gd),
-                 f, op, condf, adjust, checkempty, incol, gd)
+                 f, op, condf, adjust, checkempty, incol, gd, nthreads)
 # Avoids the overhead due to Missing when computing reduction
 groupreduce(f, op, condf::typeof(!ismissing), adjust, checkempty::Bool,
-            incol::AbstractVector, gd::GroupedDataFrame) =
+            incol::AbstractVector, gd::GroupedDataFrame,
+            nthreads::Integer) =
     groupreduce!(disallowmissing(groupreduce_init(op, condf, adjust, incol, gd)),
-                 f, op, condf, adjust, checkempty, incol, gd)
+                 f, op, condf, adjust, checkempty, incol, gd, nthreads)
 
-(r::Reduce)(incol::AbstractVector, gd::GroupedDataFrame) =
-    groupreduce((x, i) -> x, r.op, r.condf, r.adjust, r.checkempty, incol, gd)
+(r::Reduce)(incol::AbstractVector, gd::GroupedDataFrame;
+            nthreads::Integer=1) =
+    groupreduce((x, i) -> x, r.op, r.condf, r.adjust, r.checkempty, incol, gd, nthreads)
 
 # this definition is missing in Julia 1.0 LTS and is required by aggregation for var
 # TODO: remove this when we drop 1.0 support
 if VERSION < v"1.1"
     Base.zero(::Type{Missing}) = missing
 end
 
-function (agg::Aggregate{typeof(var)})(incol::AbstractVector, gd::GroupedDataFrame)
-    means = groupreduce((x, i) -> x, Base.add_sum, agg.condf, /, false, incol, gd)
+function (agg::Aggregate{typeof(var)})(incol::AbstractVector, gd::GroupedDataFrame;
+                                       nthreads::Integer=1)
+    means = groupreduce((x, i) -> x, Base.add_sum, agg.condf, /, false,
+                        incol, gd, nthreads)
     # !ismissing check is purely an optimization to avoid a copy later
     if eltype(means) >: Missing && agg.condf !== !ismissing
         T = Union{Missing, real(eltype(means))}
@@ -247,32 +312,38 @@ function (agg::Aggregate{typeof(var)})(incol::AbstractVector, gd::GroupedDataFra
     res = zeros(T, length(gd))
     return groupreduce!(res, (x, i) -> @inbounds(abs2(x - means[i])), +, agg.condf,
                         (x, l) -> l <= 1 ? oftype(x / (l-1), NaN) : x / (l-1),
-                        false, incol, gd)
+                        false, incol, gd, nthreads)
 end
 
-function (agg::Aggregate{typeof(std)})(incol::AbstractVector, gd::GroupedDataFrame)
-    outcol = Aggregate(var, agg.condf)(incol, gd)
+function (agg::Aggregate{typeof(std)})(incol::AbstractVector, gd::GroupedDataFrame;
+                                       nthreads::Integer=1)
+    outcol = Aggregate(var, agg.condf)(incol, gd; nthreads=nthreads)
     if eltype(outcol) <: Union{Missing, Rational}
         return sqrt.(outcol)
     else
         return map!(sqrt, outcol, outcol)
     end
 end
 
-for f in (first, last)
-    function (agg::Aggregate{typeof(f)})(incol::AbstractVector, gd::GroupedDataFrame)
-        n = length(gd)
-        outcol = similar(incol, n)
-        fillfirst!(agg.condf, outcol, incol, gd, rev=agg.f === last)
-        if isconcretetype(eltype(outcol))
-            return outcol
-        else
-            return copyto_widen!(Tables.allocatecolumn(typeof(first(outcol)), n), outcol)
+for f in (:first, :last)
+    # Without using @eval the presence of a keyword argument triggers a Julia bug
+    @eval begin
+        function (agg::Aggregate{typeof($f)})(incol::AbstractVector, gd::GroupedDataFrame;
+                                              nthreads::Integer=1)
+            n = length(gd)
+            outcol = similar(incol, n)
+            fillfirst!(agg.condf, outcol, incol, gd, rev=agg.f === last)
+            if isconcretetype(eltype(outcol))
+                return outcol
+            else
+                return copyto_widen!(Tables.allocatecolumn(typeof(first(outcol)), n), outcol)
+            end
         end
     end
 end
 
-function (agg::Aggregate{typeof(length)})(incol::AbstractVector, gd::GroupedDataFrame)
+function (agg::Aggregate{typeof(length)})(incol::AbstractVector, gd::GroupedDataFrame;
+                                          nthreads::Integer=1)
     if getfield(gd, :idx) === nothing
         lens = zeros(Int, length(gd))
         @inbounds for gix in gd.groups