Merge remote-tracking branch 'YaccConstructor/dev' into dev

.github/workflows/build-and-benchmark.yml

@@ -36,7 +36,7 @@ jobs:
         with:
           name: BFS
           tool: 'benchmarkdotnet'
-          output-file-path: BenchmarkDotNet.Artifacts/results/GraphBLAS.FSharp.Benchmarks.BFSWithoutTransferBenchmarkInt32-report-brief.json
+          output-file-path: BenchmarkDotNet.Artifacts/results/GraphBLAS.FSharp.Benchmarks.Algorithms.BFS.BFSWithoutTransferBenchmarkInt32-report-brief.json
           # Access token to deploy GitHub Pages branch
           github-token: ${{ secrets._GITHUB_TOKEN }}
           # Push and deploy GitHub pages branch automatically

benchmarks/GraphBLAS-sharp.Benchmarks/Algorithms/BFS.fs

@@ -1,4 +1,4 @@
-namespace GraphBLAS.FSharp.Benchmarks.Algorithms.BFS
+namespace GraphBLAS.FSharp.Benchmarks.Algorithms.BFS
 
 open System.IO
 open BenchmarkDotNet.Attributes
@@ -12,8 +12,8 @@ open GraphBLAS.FSharp.Objects.ArraysExtensions
 open GraphBLAS.FSharp.Backend.Quotes
 
 [<AbstractClass>]
-[<IterationCount(10)>]
-[<WarmupCount(3)>]
+[<IterationCount(100)>]
+[<WarmupCount(10)>]
 [<Config(typeof<Configs.Matrix>)>]
 type Benchmarks<'elem when 'elem : struct>(
     buildFunToBenchmark,
@@ -211,4 +211,3 @@ type BFSWithTransferBenchmarkBool() =
 
     static member InputMatrixProvider =
         Benchmarks<_>.InputMatrixProviderBuilder "BFSBenchmarks.txt"
-

benchmarks/GraphBLAS-sharp.Benchmarks/Configs/Context.txt

@@ -1,3 +1,3 @@
-NVIDIA*
+AMD*
 Gpu
-32
+64

benchmarks/GraphBLAS-sharp.Benchmarks/Configs/WorkflowTargets.txt

@@ -1 +1 @@
-BFSBenchmark
+BFSWithoutTransfer

benchmarks/GraphBLAS-sharp.Benchmarks/GraphBLAS-sharp.Benchmarks.fsproj

@@ -25,7 +25,6 @@
     <Compile Include="Matrix/Map2/MathNET.fs" />
     <Compile Include="Vector/Map2.fs" />
     <Compile Include="Algorithms/BFS.fs" />
-    <Compile Include="Algorithms/PageRank.fs" />
     <Compile Include="Program.fs" />
     <Folder Include="Datasets" />
   </ItemGroup>

benchmarks/GraphBLAS-sharp.Benchmarks/Program.fs

@@ -5,8 +5,6 @@ open BenchmarkDotNet.Running
 let main argv =
     let benchmarks =
         BenchmarkSwitcher [| typeof<Algorithms.BFS.BFSWithoutTransferBenchmarkBool>
-                             typeof<Algorithms.BFS.BFSPushPullWithoutTransferBenchmarkBool>
-                             typeof<Algorithms.PageRank.PageRankWithoutTransferBenchmarkFloat32> |]
 
     benchmarks.Run argv |> ignore
     0

benchmarks/GraphBLAS-sharp.Benchmarks/Scripts/Benchmark.py

@@ -7,8 +7,8 @@
 
 from dataclasses import dataclass
 
-ROOT = pathlib.Path(__file__).parent.parent.parent.parent
-BENCHMARKS = pathlib.Path(__file__).parent.parent
+ROOT = pathlib.Path(__file__).resolve().parent.parent.parent.parent
+BENCHMARKS = pathlib.Path(__file__).resolve().parent.parent
 CONFIGS = BENCHMARKS / "Configs"
 BINARIES = BENCHMARKS / "bin" / "Release" / "net7.0"
 RESULTS = ROOT / "BenchmarkDotNet.Artifacts" / "results"

src/GraphBLAS-sharp.Backend/Algorithms/Algorithms.fs

@@ -12,10 +12,10 @@ module Algorithms =
 
         let singleSourcePushPull = BFS.singleSourcePushPull
 
-    module SSSP =
-        let run = SSSP.run
+    module MSBFS =
+        let runLevels = MSBFS.Levels.run
 
-    module PageRank =
-        let run = PageRank.run
+        let runParents = MSBFS.Parents.run
 
-        let prepareMatrix = PageRank.prepareMatrix
+    module SSSP =
+        let run = SSSP.run

src/GraphBLAS-sharp.Backend/Algorithms/BFS.fs

@@ -87,7 +87,7 @@ module internal BFS =
         let fillSubVectorTo =
             Vector.assignByMaskInPlace Mask.assign clContext workGroupSize
 
-        fun (queue: MailboxProcessor<Msg>) (matrix: ClMatrix<'a>) (source: int) ->
+        fun (queue: MailboxProcessor<Msg>) (matrix: ClMatrix<bool>) (source: int) ->
             let vertexCount = matrix.RowCount
 
             let levels =
@@ -149,7 +149,7 @@ module internal BFS =
             Vector.map2Sparse Mask.complementedOp clContext workGroupSize
 
         let fillSubVectorInPlace =
-            Vector.assignByMaskInPlace Mask.assign clContext workGroupSize
+            Vector.assignByMaskInPlace (Mask.assign) clContext workGroupSize
 
         let toSparse = Vector.toSparse clContext workGroupSize
 

src/GraphBLAS-sharp.Backend/Algorithms/MSBFS.fs

@@ -0,0 +1,265 @@
+namespace GraphBLAS.FSharp.Backend.Algorithms
+
+open Brahma.FSharp
+open FSharp.Quotations
+open GraphBLAS.FSharp
+open GraphBLAS.FSharp.Objects
+open GraphBLAS.FSharp.Common
+open GraphBLAS.FSharp.Objects.ClMatrix
+open GraphBLAS.FSharp.Objects.ArraysExtensions
+open GraphBLAS.FSharp.Objects.ClContextExtensions
+open GraphBLAS.FSharp.Objects.ClCellExtensions
+open GraphBLAS.FSharp.Backend.Quotes
+open GraphBLAS.FSharp.Backend.Matrix.LIL
+open GraphBLAS.FSharp.Backend.Matrix.COO
+
+module internal MSBFS =
+    let private frontExclude (clContext: ClContext) workGroupSize =
+
+        let invert =
+            ClArray.mapInPlace ArithmeticOperations.intNotQ clContext workGroupSize
+
+        let prefixSum =
+            PrefixSum.standardExcludeInPlace clContext workGroupSize
+
+        let scatterIndices =
+            Scatter.lastOccurrence clContext workGroupSize
+
+        let scatterValues =
+            Scatter.lastOccurrence clContext workGroupSize
+
+        fun (queue: MailboxProcessor<_>) allocationMode (front: ClMatrix.COO<_>) (intersection: ClArray<int>) ->
+
+            invert queue intersection
+
+            let length =
+                (prefixSum queue intersection).ToHostAndFree queue
+
+            if length = 0 then
+                None
+            else
+                let rows =
+                    clContext.CreateClArrayWithSpecificAllocationMode(allocationMode, length)
+
+                let columns =
+                    clContext.CreateClArrayWithSpecificAllocationMode(allocationMode, length)
+
+                let values =
+                    clContext.CreateClArrayWithSpecificAllocationMode(allocationMode, length)
+
+                scatterIndices queue intersection front.Rows rows
+                scatterIndices queue intersection front.Columns columns
+                scatterValues queue intersection front.Values values
+
+                { Context = clContext
+                  Rows = rows
+                  Columns = columns
+                  Values = values
+                  RowCount = front.RowCount
+                  ColumnCount = front.ColumnCount }
+                |> Some
+
+    module Levels =
+        let private updateFrontAndLevels (clContext: ClContext) workGroupSize =
+
+            let updateFront = frontExclude clContext workGroupSize
+
+            let mergeDisjoint =
+                Matrix.mergeDisjoint clContext workGroupSize
+
+            let setLevel = ClArray.fill clContext workGroupSize
+
+            let findIntersection =
+                Intersect.findKeysIntersection clContext workGroupSize
+
+            fun (queue: MailboxProcessor<_>) allocationMode (level: int) (front: ClMatrix.COO<_>) (levels: ClMatrix.COO<_>) ->
+
+                // Find intersection of levels and front indices.
+                let intersection =
+                    findIntersection queue DeviceOnly front levels
+
+                // Remove mutual elements
+                let newFront =
+                    updateFront queue allocationMode front intersection
+
+                intersection.Free queue
+
+                match newFront with
+                | Some f ->
+                    let levelClCell = clContext.CreateClCell level
+
+                    // Set current level value to all remaining front positions
+                    setLevel queue levelClCell 0 f.Values.Length f.Values
+
+                    levelClCell.Free queue
+
+                    // Update levels
+                    let newLevels = mergeDisjoint queue levels f
+
+                    newLevels, newFront
+                | _ -> levels, None
+
+        let run<'a when 'a: struct>
+            (add: Expr<int -> int -> int option>)
+            (mul: Expr<int -> 'a -> int option>)
+            (clContext: ClContext)
+            workGroupSize
+            =
+
+            let spGeMM =
+                Operations.SpGeMM.COO.expand add mul clContext workGroupSize
+
+            let copy = Matrix.copy clContext workGroupSize
+
+            let updateFrontAndLevels =
+                updateFrontAndLevels clContext workGroupSize
+
+            fun (queue: MailboxProcessor<Msg>) (matrix: ClMatrix<'a>) (source: int list) ->
+                let vertexCount = matrix.RowCount
+                let sourceVertexCount = source.Length
+
+                let source = source |> List.sort
+
+                let startMatrix =
+                    source |> List.mapi (fun i vertex -> i, vertex, 1)
+
+                let mutable levels =
+                    startMatrix
+                    |> Matrix.ofList clContext DeviceOnly sourceVertexCount vertexCount
+
+                let mutable front = copy queue DeviceOnly levels
+
+                let mutable level = 1
+                let mutable stop = false
+
+                while not stop do
+                    level <- level + 1
+
+                    //Getting new frontier
+                    match spGeMM queue DeviceOnly (ClMatrix.COO front) matrix with
+                    | None ->
+                        front.Dispose queue
+                        stop <- true
+
+                    | Some newFrontier ->
+                        front.Dispose queue
+
+                        //Filtering visited vertices
+                        match updateFrontAndLevels queue DeviceOnly level newFrontier levels with
+                        | l, Some f ->
+                            front <- f
+
+                            levels.Dispose queue
+
+                            levels <- l
+
+                            newFrontier.Dispose queue
+
+                        | _, None ->
+                            stop <- true
+                            newFrontier.Dispose queue
+
+                ClMatrix.COO levels
+
+    module Parents =
+        let private updateFrontAndParents (clContext: ClContext) workGroupSize =
+            let frontExclude = frontExclude clContext workGroupSize
+
+            let mergeDisjoint =
+                Matrix.mergeDisjoint clContext workGroupSize
+
+            let findIntersection =
+                Intersect.findKeysIntersection clContext workGroupSize
+
+            let copyIndices = ClArray.copyTo clContext workGroupSize
+
+            fun (queue: MailboxProcessor<Msg>) allocationMode (front: ClMatrix.COO<_>) (parents: ClMatrix.COO<_>) ->
+
+                // Find intersection of levels and front indices.
+                let intersection =
+                    findIntersection queue DeviceOnly front parents
+
+                // Remove mutual elements
+                let newFront =
+                    frontExclude queue allocationMode front intersection
+
+                intersection.Free queue
+
+                match newFront with
+                | Some f ->
+                    // Update parents
+                    let newParents = mergeDisjoint queue parents f
+
+                    copyIndices queue f.Columns f.Values
+
+                    newParents, Some f
+
+                | _ -> parents, None
+
+        let run<'a when 'a: struct> (clContext: ClContext) workGroupSize =
+
+            let spGeMM =
+                Operations.SpGeMM.COO.expand
+                    (ArithmeticOperations.min)
+                    (ArithmeticOperations.fst)
+                    clContext
+                    workGroupSize
+
+            let updateFrontAndParents =
+                updateFrontAndParents clContext workGroupSize
+
+            fun (queue: MailboxProcessor<Msg>) (inputMatrix: ClMatrix<'a>) (source: int list) ->
+                let vertexCount = inputMatrix.RowCount
+                let sourceVertexCount = source.Length
+
+                let source = source |> List.sort
+
+                let matrix =
+                    match inputMatrix with
+                    | ClMatrix.CSR m ->
+                        { Context = clContext
+                          RowPointers = m.RowPointers
+                          Columns = m.Columns
+                          Values = m.Columns
+                          RowCount = m.RowCount
+                          ColumnCount = m.ColumnCount }
+                        |> ClMatrix.CSR
+                    | _ -> failwith "Incorrect format"
+
+                let mutable parents =
+                    source
+                    |> List.mapi (fun i vertex -> i, vertex, -1)
+                    |> Matrix.ofList clContext DeviceOnly sourceVertexCount vertexCount
+
+                let mutable front =
+                    source
+                    |> List.mapi (fun i vertex -> i, vertex, vertex)
+                    |> Matrix.ofList clContext DeviceOnly sourceVertexCount vertexCount
+
+                let mutable stop = false
+
+                while not stop do
+                    //Getting new frontier
+                    match spGeMM queue DeviceOnly (ClMatrix.COO front) matrix with
+                    | None ->
+                        front.Dispose queue
+                        stop <- true
+
+                    | Some newFrontier ->
+                        front.Dispose queue
+
+                        //Filtering visited vertices
+                        match updateFrontAndParents queue DeviceOnly newFrontier parents with
+                        | p, Some f ->
+                            front <- f
+
+                            parents.Dispose queue
+                            parents <- p
+
+                            newFrontier.Dispose queue
+
+                        | _, None ->
+                            stop <- true
+                            newFrontier.Dispose queue
+
+                ClMatrix.COO parents

src/GraphBLAS-sharp.Backend/Algorithms/SSSP.fs

@@ -7,11 +7,11 @@ open GraphBLAS.FSharp.Backend.Quotes
 open GraphBLAS.FSharp.Objects.ClContextExtensions
 open GraphBLAS.FSharp.Objects.ClCellExtensions
 
-module internal SSSP =
+module SSSP =
     let run (clContext: ClContext) workGroupSize =
 
         let less = ArithmeticOperations.less<int>
-        let min = ArithmeticOperations.min<int>
+        let min = ArithmeticOperations.minOption<int>
         let plus = ArithmeticOperations.intSumAsMul
 
         let spMVInPlace =
@@ -76,4 +76,6 @@ module internal SSSP =
             front1.Dispose queue
             front2.Dispose queue
 
-            distance
+            match distance with
+            | ClVector.Dense dist -> dist
+            | _ -> failwith "not implemented"