GPU map2

MathNet.Numerics.sln

@@ -28,27 +28,27 @@ EndProject
 Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Build", "Build", "{B54A0B40-DE22-49FB-B1C0-6E5BDA3B0B2F}"
 	ProjectSection(SolutionItems) = preProject
 		.editorconfig = .editorconfig
+		.gitattributes = .gitattributes
+		.gitignore = .gitignore
 		appveyor.yml = appveyor.yml
+		build.cmd = build.cmd
+		build.sh = build.sh
 		Directory.Build.props = Directory.Build.props
 		global.json = global.json
-		build\MathNet.Numerics.nuspec = build\MathNet.Numerics.nuspec
-		build\MathNet.Numerics.FSharp.nuspec = build\MathNet.Numerics.FSharp.nuspec
 		build\MathNet.Numerics.CUDA.Win.nuspec = build\MathNet.Numerics.CUDA.Win.nuspec
+		build\MathNet.Numerics.FSharp.nuspec = build\MathNet.Numerics.FSharp.nuspec
 		build\MathNet.Numerics.MKL.Linux-x64.nuspec = build\MathNet.Numerics.MKL.Linux-x64.nuspec
 		build\MathNet.Numerics.MKL.Linux-x86.nuspec = build\MathNet.Numerics.MKL.Linux-x86.nuspec
 		build\MathNet.Numerics.MKL.Linux.nuspec = build\MathNet.Numerics.MKL.Linux.nuspec
 		build\MathNet.Numerics.MKL.Win-x64.nuspec = build\MathNet.Numerics.MKL.Win-x64.nuspec
 		build\MathNet.Numerics.MKL.Win-x86.nuspec = build\MathNet.Numerics.MKL.Win-x86.nuspec
 		build\MathNet.Numerics.MKL.Win.nuspec = build\MathNet.Numerics.MKL.Win.nuspec
+		build\MathNet.Numerics.nuspec = build\MathNet.Numerics.nuspec
 		build\MathNet.Numerics.OpenBLAS.Win.nuspec = build\MathNet.Numerics.OpenBLAS.Win.nuspec
 		build\NativeProvider.targets = build\NativeProvider.targets
+		NuGet.config = NuGet.config
 		paket.dependencies = paket.dependencies
 		paket.lock = paket.lock
-		build.cmd = build.cmd
-		build.sh = build.sh
-		.gitignore = .gitignore
-		.gitattributes = .gitattributes
-		NuGet.config = NuGet.config
 		restore.cmd = restore.cmd
 		restore.sh = restore.sh
 	EndProjectSection
@@ -69,6 +69,11 @@ Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "Data.Text", "src\Data.Text\
 EndProject
 Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "Data.Tests", "src\Data.Tests\Data.Tests.csproj", "{1FBB8988-3F40-4FCE-BEA8-2C705888804A}"
 EndProject
+Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Solution Items", "Solution Items", "{8EC462FD-D22E-90A8-E5CE-7E832BA40C5D}"
+	ProjectSection(SolutionItems) = preProject
+		kernels.fs = kernels.fs
+	EndProjectSection
+EndProject
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
 		Debug|Any CPU = Debug|Any CPU

src/FSharp.Tests/DenseMatrixGPUTests.fs

@@ -25,8 +25,8 @@ module DenseMatrixGPUTests =
 
     [<Test>]
     let ``DenseMatrixGPU.map2`` () =
-        GPU.Matrix.map2 id largeM largeM |> shouldEqual largeM
+        GPU.Matrix.map2 <@(+)@> largeM largeM |> shouldEqual (Matrix.map (fun x -> x + x) largeM)
 
     [<Test>]
     let ``DenseMatrixGPU.mXm`` () =
-        GPU.Matrix.mXm largeM largeM |> shouldEqual largeM
+        GPU.Matrix.mXm largeM largeM |> shouldEqual largeM

src/FSharp.Tests/FSharp.Tests.fsproj

@@ -3,7 +3,7 @@
     <PropertyGroup>
         <OutputType>Exe</OutputType>
         <GenerateProgramFile>false</GenerateProgramFile>
-        <TargetFrameworks>net8.0;net48</TargetFrameworks>
+        <TargetFrameworks>net8.0;</TargetFrameworks>
         <AssemblyName>MathNet.Numerics.FSharp.Tests</AssemblyName>
         <RootNamespace>MathNet.Numerics.Tests</RootNamespace>
         <IsPackable>false</IsPackable>

src/FSharp/FSharp.fsproj

@@ -2,7 +2,7 @@
 <Project Sdk="Microsoft.NET.Sdk">
     <PropertyGroup>
         <OutputType>Library</OutputType>
-        <TargetFrameworks>net8.0;net6.0;net48;netstandard2.0</TargetFrameworks>
+        <TargetFrameworks>net8.0</TargetFrameworks>
         <AssemblyName>MathNet.Numerics.FSharp</AssemblyName>
         <RootNamespace>MathNet.Numerics</RootNamespace>
         <IsPackable>true</IsPackable>
@@ -17,11 +17,16 @@ many contributions, proper release notes with attributions will follow. thank yo
         <PackageTags>fsharp F# math numeric statistics probability integration interpolation regression solve fit linear algebra matrix fft</PackageTags>
         <GeneratePackageOnBuild>false</GeneratePackageOnBuild>
         <EnableDefaultCompileItems>false</EnableDefaultCompileItems>
-        <NoWarn>2003;NU1604</NoWarn>
+        <NoWarn>2003;NU1604;NU1608</NoWarn>
     </PropertyGroup>
     <ItemGroup>
         <ProjectReference Include="..\Numerics\Numerics.csproj" />
     </ItemGroup>
+    <ItemGroup Condition="'$(TargetFramework)' == 'net8.0'">
+        <PackageReference Include="Brahma.FSharp">
+        <Version>2.0.6</Version>
+      </PackageReference>
+    </ItemGroup>
     <ItemGroup>
         <Compile Include="AssemblyInfo.fs" />
         <Compile Include="Statistics.fs" />

src/FSharp/LinearAlgebra.Matrix.GPU.fs

@@ -32,7 +32,7 @@ namespace MathNet.Numerics.LinearAlgebra.GPU
 open System
 open MathNet.Numerics
 open MathNet.Numerics.LinearAlgebra
-
+open Brahma.FSharp
 
 /// A module which implements functional matrix operations.
 [<CompilationRepresentation(CompilationRepresentationFlags.ModuleSuffix)>]
@@ -42,7 +42,58 @@ module Matrix =
     let iter f (v: #Matrix<_>) = v
 
     /// !!!
-    let map2 f (v: #Matrix<_>) (w: #Matrix<_>) = v
+    let map2 (f: Quotations.Expr<'a -> 'b -> 'c>) (v: #Matrix<'a>) (w: #Matrix<'b>) =
+        let device = ClDevice.GetFirstAppropriateDevice() // ?
+        let context = RuntimeContext(device).ClContext
+
+        let n = v.RowCount
+        let m = v.ColumnCount
+
+        let m1_gpu = context.CreateClArray<_>(v.ToColumnMajorArray(), HostAccessMode.NotAccessible)
+        let m2_gpu = context.CreateClArray<_>(w.ToColumnMajorArray(), HostAccessMode.NotAccessible)
+        let m3_gpu =
+            context.CreateClArray(
+                n * m,
+                HostAccessMode.NotAccessible,
+                deviceAccessMode=DeviceAccessMode.WriteOnly,
+                allocationMode = AllocationMode.Default
+            )
+
+        let localWorkSize = 4
+
+        let kernel =
+            <@
+                fun (r: Range2D) (m1: ClArray<_>) (m2: ClArray<_>) (m3: ClArray<_>)  ->
+                    let i = r.GlobalID0
+                    let j = r.GlobalID1
+
+                    m3.[i * m + j] <- (%f) m1[i * m + j] m2[i * m + j]
+            @>
+
+        let kernel = context.Compile kernel
+
+        let ndRange =
+            Range2D(
+                n,
+                m,
+                localWorkSize,
+                localWorkSize
+            )
+
+        let commandQueue = context.QueueProvider.CreateQueue()
+        let kernel = kernel.GetKernel()
+
+        commandQueue.Post(Msg.MsgSetArguments(fun () -> kernel.KernelFunc ndRange m1_gpu m2_gpu m3_gpu))
+        commandQueue.Post(Msg.CreateRunMsg<_, _> kernel)
+        let result : 'c[] = Array.zeroCreate(n * m)
+        let result = commandQueue.PostAndReply(fun ch -> Msg.CreateToHostMsg(m3_gpu, result, ch))
+        commandQueue.Post(Msg.CreateFreeMsg m1_gpu)
+
+        commandQueue.Post(Msg.CreateFreeMsg m2_gpu)
+
+        commandQueue.Post(Msg.CreateFreeMsg m3_gpu)
+
+        result |> DenseMatrix.raw n m
 
     /// !!!
     let mXm (v: #Matrix<_>) (w: #Matrix<_>) = v