Elementwise

benchmarks/GraphBLAS-sharp.Benchmarks/BenchmarksEWiseAdd.fs

@@ -274,7 +274,7 @@ module M =
 type EWiseAddBenchmarks4Float32COOWithoutDataTransfer() =
 
     inherit EWiseAddBenchmarksWithoutDataTransfer<Backend.COOMatrix<float32>,float32>(
-        (fun context wgSize -> Backend.COOMatrix.eWiseAdd context Backend.Common.StandardOperations.float32Sum wgSize),
+        (fun context wgSize -> Backend.COOMatrix.elementwise context Backend.Common.StandardOperations.float32Sum wgSize),
         float32,
         (fun _ -> Utils.nextSingle (System.Random())),
         M.buildCooMatrix
@@ -286,7 +286,7 @@ type EWiseAddBenchmarks4Float32COOWithoutDataTransfer() =
 type EWiseAddBenchmarks4Float32COOWithDataTransfer() =
 
     inherit EWiseAddBenchmarksWithDataTransfer<Backend.COOMatrix<float32>,float32>(
-        (fun context wgSize -> Backend.COOMatrix.eWiseAdd context Backend.Common.StandardOperations.float32Sum wgSize),
+        (fun context wgSize -> Backend.COOMatrix.elementwise context Backend.Common.StandardOperations.float32Sum wgSize),
         float32,
         (fun _ -> Utils.nextSingle (System.Random())),
         M.buildCooMatrix,
@@ -300,7 +300,7 @@ type EWiseAddBenchmarks4Float32COOWithDataTransfer() =
 type EWiseAddBenchmarks4BoolCOOWithoutDataTransfer() =
 
     inherit EWiseAddBenchmarksWithoutDataTransfer<Backend.COOMatrix<bool>,bool>(
-        (fun context wgSize -> Backend.COOMatrix.eWiseAdd context Backend.Common.StandardOperations.boolSum wgSize),
+        (fun context wgSize -> Backend.COOMatrix.elementwise context Backend.Common.StandardOperations.boolSum wgSize),
         (fun _ -> true),
         (fun _ -> true),
         M.buildCooMatrix
@@ -313,7 +313,7 @@ type EWiseAddBenchmarks4BoolCOOWithoutDataTransfer() =
 type EWiseAddBenchmarks4Float32CSRWithoutDataTransfer() =
 
     inherit EWiseAddBenchmarksWithoutDataTransfer<Backend.CSRMatrix<float32>,float32>(
-        (fun context wgSize -> Backend.CSRMatrix.eWiseAdd context Backend.Common.StandardOperations.float32Sum wgSize),
+        (fun context wgSize -> Backend.CSRMatrix.elementwise context Backend.Common.StandardOperations.float32Sum wgSize),
         float32,
         (fun _ -> Utils.nextSingle (System.Random())),
         M.buildCsrMatrix
@@ -326,7 +326,7 @@ type EWiseAddBenchmarks4Float32CSRWithoutDataTransfer() =
 type EWiseAddBenchmarks4BoolCSRWithoutDataTransfer() =
 
     inherit EWiseAddBenchmarksWithoutDataTransfer<Backend.CSRMatrix<bool>,bool>(
-        (fun context wgSize -> Backend.CSRMatrix.eWiseAdd context Backend.Common.StandardOperations.boolSum wgSize),
+        (fun context wgSize -> Backend.CSRMatrix.elementwise context Backend.Common.StandardOperations.boolSum wgSize),
         (fun _ -> true),
         (fun _ -> true),
         M.buildCsrMatrix
@@ -340,7 +340,7 @@ type EWiseAddBenchmarks4BoolCSRWithoutDataTransfer() =
 type EWiseAddAtLeastOneBenchmarks4BoolCOOWithoutDataTransfer() =
 
     inherit EWiseAddBenchmarksWithoutDataTransfer<Backend.COOMatrix<bool>,bool>(
-        (fun context wgSize -> Backend.COOMatrix.eWiseAddAtLeastOne context Backend.Common.StandardOperations.boolSumAtLeastOne wgSize),
+        (fun context wgSize -> Backend.COOMatrix.elementwiseAtLeastOne context Backend.Common.StandardOperations.boolSumAtLeastOne wgSize),
         (fun _ -> true),
         (fun _ -> true),
         M.buildCooMatrix
@@ -352,7 +352,7 @@ type EWiseAddAtLeastOneBenchmarks4BoolCOOWithoutDataTransfer() =
 type EWiseAddAtLeastOneBenchmarks4BoolCSRWithoutDataTransfer() =
 
     inherit EWiseAddBenchmarksWithoutDataTransfer<Backend.CSRMatrix<bool>,bool>(
-        (fun context wgSize -> Backend.CSRMatrix.eWiseAddAtLeastOne context Backend.Common.StandardOperations.boolSumAtLeastOne wgSize),
+        (fun context wgSize -> Backend.CSRMatrix.elementwiseAtLeastOne context Backend.Common.StandardOperations.boolSumAtLeastOne wgSize),
         (fun _ -> true),
         (fun _ -> true),
         M.buildCsrMatrix
@@ -364,7 +364,7 @@ type EWiseAddAtLeastOneBenchmarks4BoolCSRWithoutDataTransfer() =
 type EWiseAddAtLeastOneBenchmarks4Float32COOWithoutDataTransfer() =
 
     inherit EWiseAddBenchmarksWithoutDataTransfer<Backend.COOMatrix<float32>,float32>(
-        (fun context wgSize -> Backend.COOMatrix.eWiseAddAtLeastOne context Backend.Common.StandardOperations.float32SumAtLeastOne wgSize),
+        (fun context wgSize -> Backend.COOMatrix.elementwiseAtLeastOne context Backend.Common.StandardOperations.float32SumAtLeastOne wgSize),
         float32,
         (fun _ -> Utils.nextSingle (System.Random())),
         M.buildCooMatrix
@@ -376,7 +376,7 @@ type EWiseAddAtLeastOneBenchmarks4Float32COOWithoutDataTransfer() =
 type EWiseAddAtLeastOneBenchmarks4Float32CSRWithoutDataTransfer() =
 
     inherit EWiseAddBenchmarksWithoutDataTransfer<Backend.CSRMatrix<float32>,float32>(
-        (fun context wgSize -> Backend.CSRMatrix.eWiseAddAtLeastOne context Backend.Common.StandardOperations.float32SumAtLeastOne wgSize),
+        (fun context wgSize -> Backend.CSRMatrix.elementwiseAtLeastOne context Backend.Common.StandardOperations.float32SumAtLeastOne wgSize),
         float32,
         (fun _ -> Utils.nextSingle (System.Random())),
         M.buildCsrMatrix

benchmarks/GraphBLAS-sharp.Benchmarks/BenchmarksMxv.fs

@@ -1,6 +1,7 @@
 namespace GraphBLAS.FSharp.Benchmarks
 
 open GraphBLAS.FSharp
+open GraphBLAS.FSharp.Backend
 open BenchmarkDotNet.Attributes
 open BenchmarkDotNet.Configs
 open BenchmarkDotNet.Columns
@@ -75,4 +76,4 @@ type MxvBenchmarks() =
                 | ".mtx" -> MtxReader(Utils.getFullPathToMatrix "Common" matrixFilename)
                 | _ -> failwith "Unsupported matrix format"
             )
-*)
+*)

benchmarks/GraphBLAS-sharp.Benchmarks/Helpers.fs

@@ -239,11 +239,11 @@ module Utils =
 
     let rowIndices2rowPointers (rowIndices: int []) rowCount =
         let nnzPerRow = Array.zeroCreate rowCount
-        let rowPointers = Array.zeroCreate rowCount
+        let rowPointers = Array.zeroCreate (rowCount + 1)
 
         Array.iter (fun rowIndex -> nnzPerRow.[rowIndex] <- nnzPerRow.[rowIndex] + 1) rowIndices
 
-        for i in 1 .. rowCount - 1 do
+        for i in 1 .. rowCount do
             rowPointers.[i] <- rowPointers.[i - 1] + nnzPerRow.[i - 1]
 
         rowPointers

src/GraphBLAS-sharp.Backend/Common/StandardOperations.fs

@@ -88,7 +88,7 @@ module StandardOperations =
             | Both _ -> res <- true
             | _ -> ()
 
-            if res then None else (Some true) @>
+            if res then Some true else None @>
 
     let intMulAtLeastOne = mkNumericMulAtLeastOne 0
     let byteMulAtLeastOne = mkNumericMulAtLeastOne 0uy

src/GraphBLAS-sharp.Backend/Common/Utils.fs

@@ -1,6 +1,10 @@
 namespace GraphBLAS.FSharp.Backend.Common
 
+open Brahma.FSharp
+
 module internal Utils =
+    let defaultWorkGroupSize = 32
+
     let floorToPower2 =
         fun x -> x ||| (x >>> 1)
         >> fun x -> x ||| (x >>> 2)

src/GraphBLAS-sharp.Backend/GraphBLAS-sharp.Backend.fsproj

@@ -1,4 +1,4 @@
-<?xml version="1.0" encoding="utf-8"?>
+<?xml version="1.0" encoding="utf-8"?>
 <Project Sdk="Microsoft.NET.Sdk">
 
   <PropertyGroup>
@@ -18,12 +18,17 @@
     <Compile Include="Common/BitonicSort.fs" />
     <Compile Include="Common/StandardOperations.fs" />
     <Compile Include="Predefined/PrefixSum.fs" />
-    <Compile Include="Matrices.fs" />
+    <Compile Include="Objects/Common.fs" />
+    <Compile Include="Objects/ArraysExtentions.fs" />
+    <Compile Include="Objects/Vector.fs" />
+    <Compile Include="Objects/Matrix.fs" />
     <Compile Include="Matrix/COOMatrix/COOMatrix.fs" />
+    <Compile Include="Matrix/CSRMatrix/Elementwise.fs" />
     <Compile Include="Matrix/CSRMatrix/CSRMatrix.fs" />
     <Compile Include="Matrix/CSRMatrix/SpGEMM.fs" />
     <Compile Include="Matrix/CSRMatrix/SpMV.fs" />
     <Compile Include="Matrix/Matrix.fs" />
+    <Folder Include="Vector" />
     <!--Compile Include="Backend/CSRMatrix/GetTuples.fs" /-->
     <!--Compile Include="Backend/CSRMatrix/SpMSpV.fs" /-->
     <!--Compile Include="Backend/CSRMatrix/Transpose.fs" /-->

src/GraphBLAS-sharp.Backend/Matrix/COOMatrix/COOMatrix.fs

@@ -11,16 +11,15 @@ module COOMatrix =
     let private setPositions<'a when 'a: struct> (clContext: ClContext) workGroupSize =
 
         let setPositions =
-            <@ fun (ndRange: Range1D) prefixSumArrayLength (allRowsBuffer: ClArray<int>) (allColumnsBuffer: ClArray<int>) (allValuesBuffer: ClArray<'a>) (prefixSumArrayBuffer: ClArray<int>) (resultRowsBuffer: ClArray<int>) (resultColumnsBuffer: ClArray<int>) (resultValuesBuffer: ClArray<'a>) ->
+            <@ fun (ndRange: Range1D) prefixSumArrayLength resultLength (allRowsBuffer: ClArray<int>) (allColumnsBuffer: ClArray<int>) (allValuesBuffer: ClArray<'a>) (prefixSumArrayBuffer: ClArray<int>) (resultRowsBuffer: ClArray<int>) (resultColumnsBuffer: ClArray<int>) (resultValuesBuffer: ClArray<'a>) ->
 
                 let i = ndRange.GlobalID0
+                let index = prefixSumArrayBuffer.[i]
 
-                if i = prefixSumArrayLength - 1
-                   || i < prefixSumArrayLength
-                      && prefixSumArrayBuffer.[i]
-                         <> prefixSumArrayBuffer.[i + 1] then
-                    let index = prefixSumArrayBuffer.[i]
-
+                if (i < prefixSumArrayLength - 1
+                    && index <> prefixSumArrayBuffer.[i + 1])
+                   || (i = prefixSumArrayLength - 1
+                       && index < resultLength) then
                     resultRowsBuffer.[index] <- allRowsBuffer.[i]
                     resultColumnsBuffer.[index] <- allColumnsBuffer.[i]
                     resultValuesBuffer.[index] <- allValuesBuffer.[i] @>
@@ -82,6 +81,7 @@ module COOMatrix =
                         kernel.KernelFunc
                             ndRange
                             prefixSumArrayLength
+                            resultLength
                             allRows
                             allColumns
                             allValues
@@ -385,7 +385,7 @@ module COOMatrix =
     ///<param name="clContext">.</param>
     ///<param name="opAdd">.</param>
     ///<param name="workGroupSize">Should be a power of 2 and greater than 1.</param>
-    let eWiseAdd<'a, 'b, 'c when 'a: struct and 'b: struct and 'c: struct and 'c: equality>
+    let elementwise<'a, 'b, 'c when 'a: struct and 'b: struct and 'c: struct and 'c: equality>
         (clContext: ClContext)
         (opAdd: Expr<'a option -> 'b option -> 'c option>)
         workGroupSize
@@ -613,7 +613,7 @@ module COOMatrix =
     ///<param name="clContext">.</param>
     ///<param name="opAdd">.</param>
     ///<param name="workGroupSize">Should be a power of 2 and greater than 1.</param>
-    let eWiseAddAtLeastOne<'a, 'b, 'c when 'a: struct and 'b: struct and 'c: struct and 'c: equality>
+    let elementwiseAtLeastOne<'a, 'b, 'c when 'a: struct and 'b: struct and 'c: struct and 'c: equality>
         (clContext: ClContext)
         (opAdd: Expr<AtLeastOne<'a, 'b> -> 'c option>)
         workGroupSize

src/GraphBLAS-sharp.Backend/Matrix/CSRMatrix/CSRMatrix.fs

@@ -3,6 +3,7 @@ namespace GraphBLAS.FSharp.Backend
 open Brahma.FSharp
 open GraphBLAS.FSharp.Backend
 open GraphBLAS.FSharp.Backend.Common
+open GraphBLAS.FSharp.Backend.Elementwise
 open Microsoft.FSharp.Quotations
 
 module CSRMatrix =
@@ -79,12 +80,13 @@ module CSRMatrix =
               Columns = matrix.Columns
               Values = matrix.Values }
 
-    let eWiseAdd (clContext: ClContext) (opAdd: Expr<'a option -> 'b option -> 'c option>) workGroupSize =
+    ///<remarks>Old version</remarks>
+    let elementwiseWithCOO (clContext: ClContext) (opAdd: Expr<'a option -> 'b option -> 'c option>) workGroupSize =
 
         let prepareRows = prepareRows clContext workGroupSize
 
         let eWiseCOO =
-            COOMatrix.eWiseAdd clContext opAdd workGroupSize
+            COOMatrix.elementwise clContext opAdd workGroupSize
 
         let toCSRInplace =
             COOMatrix.toCSRInplace clContext workGroupSize
@@ -113,12 +115,17 @@ module CSRMatrix =
 
             toCSRInplace processor m3COO
 
-    let eWiseAddAtLeastOne (clContext: ClContext) (opAdd: Expr<AtLeastOne<'a, 'b> -> 'c option>) workGroupSize =
+    ///<remarks>Old version</remarks>
+    let elementwiseAtLeastOneWithCOO
+        (clContext: ClContext)
+        (opAdd: Expr<AtLeastOne<'a, 'b> -> 'c option>)
+        workGroupSize
+        =
 
         let prepareRows = prepareRows clContext workGroupSize
 
         let eWiseCOO =
-            COOMatrix.eWiseAddAtLeastOne clContext opAdd workGroupSize
+            COOMatrix.elementwiseAtLeastOne clContext opAdd workGroupSize
 
         let toCSRInplace =
             COOMatrix.toCSRInplace clContext workGroupSize
@@ -177,3 +184,139 @@ module CSRMatrix =
             let coo = toCOO queue matrix
             let transposedCoo = transposeInplace queue coo
             toCSRInplace queue transposedCoo
+
+    let elementwiseToCOO<'a, 'b, 'c when 'a: struct and 'b: struct and 'c: struct and 'c: equality>
+        (clContext: ClContext)
+        (opAdd: Expr<'a option -> 'b option -> 'c option>)
+        workGroupSize
+        =
+
+        let merge = merge clContext workGroupSize
+
+        let preparePositions =
+            preparePositions clContext opAdd Utils.defaultWorkGroupSize
+
+        let setPositions =
+            setPositions<'c> clContext Utils.defaultWorkGroupSize
+
+        fun (queue: MailboxProcessor<_>) (matrixLeft: CSRMatrix<'a>) (matrixRight: CSRMatrix<'b>) ->
+
+            let allRows, allColumns, leftMergedValues, rightMergedValues, isRowEnd, isLeft =
+                merge
+                    queue
+                    matrixLeft.RowPointers
+                    matrixLeft.Columns
+                    matrixLeft.Values
+                    matrixRight.RowPointers
+                    matrixRight.Columns
+                    matrixRight.Values
+
+            let positions, allValues =
+                preparePositions queue allColumns leftMergedValues rightMergedValues isRowEnd isLeft
+
+            queue.Post(Msg.CreateFreeMsg<_>(leftMergedValues))
+            queue.Post(Msg.CreateFreeMsg<_>(rightMergedValues))
+
+            let resultRows, resultColumns, resultValues, positions, positionsSum =
+                setPositions queue allRows allColumns allValues positions
+
+            queue.Post(Msg.CreateFreeMsg<_>(allRows))
+            queue.Post(Msg.CreateFreeMsg<_>(isLeft))
+            queue.Post(Msg.CreateFreeMsg<_>(isRowEnd))
+            queue.Post(Msg.CreateFreeMsg<_>(positions))
+            queue.Post(Msg.CreateFreeMsg<_>(allColumns))
+            queue.Post(Msg.CreateFreeMsg<_>(allValues))
+
+            { Context = clContext
+              RowCount = matrixLeft.RowCount
+              ColumnCount = matrixLeft.ColumnCount
+              Rows = resultRows
+              Columns = resultColumns
+              Values = resultValues }
+
+    let elementwise<'a, 'b, 'c when 'a: struct and 'b: struct and 'c: struct and 'c: equality>
+        (clContext: ClContext)
+        (opAdd: Expr<'a option -> 'b option -> 'c option>)
+        workGroupSize
+        =
+
+        let elementwiseToCOO =
+            elementwiseToCOO clContext opAdd workGroupSize
+
+        let toCSRInplace =
+            COOMatrix.toCSRInplace clContext Utils.defaultWorkGroupSize
+
+        fun (queue: MailboxProcessor<_>) (matrixLeft: CSRMatrix<'a>) (matrixRight: CSRMatrix<'b>) ->
+
+            let cooRes =
+                elementwiseToCOO queue matrixLeft matrixRight
+
+            toCSRInplace queue cooRes
+
+    let elementwiseAtLeastOneToCOO<'a, 'b, 'c when 'a: struct and 'b: struct and 'c: struct and 'c: equality>
+        (clContext: ClContext)
+        (opAdd: Expr<AtLeastOne<'a, 'b> -> 'c option>)
+        workGroupSize
+        =
+
+        let merge = merge clContext workGroupSize
+
+        let preparePositions =
+            preparePositionsAtLeastOne clContext opAdd Utils.defaultWorkGroupSize
+
+        let setPositions =
+            setPositions<'c> clContext Utils.defaultWorkGroupSize
+
+        fun (queue: MailboxProcessor<_>) (matrixLeft: CSRMatrix<'a>) (matrixRight: CSRMatrix<'b>) ->
+
+            let allRows, allColumns, leftMergedValues, rightMergedValues, isRowEnd, isLeft =
+                merge
+                    queue
+                    matrixLeft.RowPointers
+                    matrixLeft.Columns
+                    matrixLeft.Values
+                    matrixRight.RowPointers
+                    matrixRight.Columns
+                    matrixRight.Values
+
+            let positions, allValues =
+                preparePositions queue allColumns leftMergedValues rightMergedValues isRowEnd isLeft
+
+            queue.Post(Msg.CreateFreeMsg<_>(leftMergedValues))
+            queue.Post(Msg.CreateFreeMsg<_>(rightMergedValues))
+
+            let resultRows, resultColumns, resultValues, positions, positionsSum =
+                setPositions queue allRows allColumns allValues positions
+
+            queue.Post(Msg.CreateFreeMsg<_>(allRows))
+            queue.Post(Msg.CreateFreeMsg<_>(isLeft))
+            queue.Post(Msg.CreateFreeMsg<_>(isRowEnd))
+            queue.Post(Msg.CreateFreeMsg<_>(positions))
+            queue.Post(Msg.CreateFreeMsg<_>(allColumns))
+            queue.Post(Msg.CreateFreeMsg<_>(allValues))
+
+            { Context = clContext
+              RowCount = matrixLeft.RowCount
+              ColumnCount = matrixLeft.ColumnCount
+              Rows = resultRows
+              Columns = resultColumns
+              Values = resultValues }
+
+    let elementwiseAtLeastOne<'a, 'b, 'c when 'a: struct and 'b: struct and 'c: struct and 'c: equality>
+        (clContext: ClContext)
+        (opAdd: Expr<AtLeastOne<'a, 'b> -> 'c option>)
+        workGroupSize
+        =
+
+        let elementwiseAtLeastOneToCOO =
+            elementwiseAtLeastOneToCOO clContext opAdd workGroupSize
+
+        let toCSRInplace =
+            COOMatrix.toCSRInplace clContext Utils.defaultWorkGroupSize
+
+        fun (queue: MailboxProcessor<_>) (matrixLeft: CSRMatrix<'a>) (matrixRight: CSRMatrix<'b>) ->
+
+            let cooRes =
+                elementwiseAtLeastOneToCOO queue matrixLeft matrixRight
+
+            toCSRInplace queue cooRes