toSparse unsorted and tests

Author: kirillgarbar

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

@@ -148,7 +148,8 @@ module internal BFS =
         let fillSubVectorInPlace =
             Vector.assignByMaskInPlace (Mask.assign) clContext workGroupSize
 
-        let toSparse = Vector.toSparse clContext workGroupSize
+        let toSparse =
+            Vector.toSparseUnsorted clContext workGroupSize
 
         let toDense = Vector.toDense clContext workGroupSize
 

src/GraphBLAS-sharp.Backend/Vector/Dense/Vector.fs

@@ -140,7 +140,63 @@ module Vector =
 
             valueCell.Free()
 
+    // TODO: toSparseUnsorted + bitonic probably would work faster
     let toSparse<'a when 'a: struct> (clContext: ClContext) workGroupSize =
+        let scatterValues =
+            Common.Scatter.lastOccurrence clContext workGroupSize
+
+        let scatterIndices =
+            Common.Scatter.lastOccurrence clContext workGroupSize
+
+        let getBitmap =
+            Map.map (Map.option 1 0) clContext workGroupSize
+
+        let prefixSum =
+            Common.PrefixSum.standardExcludeInPlace clContext workGroupSize
+
+        let allIndices =
+            ClArray.init Map.id clContext workGroupSize
+
+        let allValues =
+            Map.map (Map.optionToValueOrZero Unchecked.defaultof<'a>) clContext workGroupSize
+
+        fun (processor: RawCommandQueue) allocationMode (vector: ClArray<'a option>) ->
+
+            let positions = getBitmap processor DeviceOnly vector
+
+            let resultLength =
+                (prefixSum processor positions)
+                    .ToHostAndFree(processor)
+
+            // compute result indices
+            let resultIndices =
+                clContext.CreateClArrayWithSpecificAllocationMode<int>(allocationMode, resultLength)
+
+            let allIndices =
+                allIndices processor DeviceOnly vector.Length
+
+            scatterIndices processor positions allIndices resultIndices
+
+            allIndices.Free()
+
+            // compute result values
+            let resultValues =
+                clContext.CreateClArrayWithSpecificAllocationMode<'a>(allocationMode, resultLength)
+
+            let allValues = allValues processor DeviceOnly vector
+
+            scatterValues processor positions allValues resultValues
+
+            allValues.Free()
+
+            positions.Free()
+
+            { Context = clContext
+              Indices = resultIndices
+              Values = resultValues
+              Size = vector.Length }
+
+    let toSparseUnsorted<'a when 'a: struct> (clContext: ClContext) workGroupSize =
 
         let kernel =
             <@ fun (ndRange: Range1D) (inputLength: int) (inputValues: ClArray<'a option>) (resultSize: ClCell<int>) (resultIndices: ClArray<int>) (resultValues: ClArray<'a>) ->

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

@@ -129,6 +129,26 @@ module Vector =
                 <| toSparse processor allocationMode vector
             | ClVector.Sparse _ -> copy processor allocationMode vector
 
+    /// <summary>
+    /// Sparsifies the given vector if it is in a dense format.
+    /// If the given vector is already sparse, copies it.
+    /// Works faster than regular version, but indices of the sparse vector are unsorted.
+    /// </summary>
+    /// <param name="clContext">OpenCL context.</param>
+    /// <param name="workGroupSize">Should be a power of 2 and greater than 1.</param>
+    let toSparseUnsorted (clContext: ClContext) workGroupSize =
+        let toSparse =
+            Dense.Vector.toSparseUnsorted clContext workGroupSize
+
+        let copy = copy clContext workGroupSize
+
+        fun (processor: RawCommandQueue) allocationMode (vector: ClVector<'a>) ->
+            match vector with
+            | ClVector.Dense vector ->
+                ClVector.Sparse
+                <| toSparse processor allocationMode vector
+            | ClVector.Sparse _ -> copy processor allocationMode vector
+
     /// <summary>
     /// Densifies the given vector if it is in a sparse format.
     /// If the given vector is already dense, copies it.

tests/GraphBLAS-sharp.Tests/Backend/Vector/Convert.fs

@@ -22,6 +22,7 @@ let wgSize = Constants.Common.defaultWorkGroupSize
 let makeTest
     formatFrom
     (convertFun: RawCommandQueue -> AllocationFlag -> ClVector<'a> -> ClVector<'a>)
+    (convertFunUnsorted: option<RawCommandQueue -> AllocationFlag -> ClVector<'a> -> ClVector<'a>>)
     isZero
     case
     (array: 'a [])
@@ -37,7 +38,7 @@ let makeTest
 
         let actual =
             let clVector = vector.ToDevice context
-            let convertedVector = convertFun q HostInterop clVector
+            let convertedVector = convertFun q DeviceOnly clVector
 
             let res = convertedVector.ToHost q
 
@@ -56,6 +57,27 @@ let makeTest
 
         Expect.equal actual expected "Vectors must be the same"
 
+        match convertFunUnsorted with
+        | None -> ()
+        | Some convertFunUnsorted ->
+            let clVector = vector.ToDevice context
+            let convertedVector = convertFunUnsorted q DeviceOnly clVector
+
+            let res = convertedVector.ToHost q
+
+            match res, expected with
+            | Vector.Sparse res, Vector.Sparse expected ->
+                let iv = Array.zip res.Indices res.Values
+                let resSorted = Array.sortBy (fun (i, v) -> i) iv
+                let indices, values = Array.unzip resSorted
+                Expect.equal indices expected.Indices "Indices must be the same"
+                Expect.equal values expected.Values "Values must be the same"
+                Expect.equal res.Size expected.Size "Size must be the same"
+            | _ -> ()
+
+            clVector.Dispose()
+            convertedVector.Dispose()
+
 let testFixtures case =
     let getCorrectnessTestName datatype formatFrom =
         sprintf $"Correctness on %s{datatype}, %A{formatFrom} -> %A{case.Format}"
@@ -68,19 +90,21 @@ let testFixtures case =
     match case.Format with
     | Sparse ->
         [ let convertFun = Vector.toSparse context wgSize
+          let convertFunUnsorted = Vector.toSparseUnsorted context wgSize
 
           Utils.listOfUnionCases<VectorFormat>
           |> List.map
               (fun formatFrom ->
-                  makeTest formatFrom convertFun ((=) 0) case
+                  makeTest formatFrom convertFun (Some convertFunUnsorted) ((=) 0) case
                   |> testPropertyWithConfig config (getCorrectnessTestName "int" formatFrom))
 
           let convertFun = Vector.toSparse context wgSize
+          let convertFunUnsorted = Vector.toSparseUnsorted context wgSize
 
           Utils.listOfUnionCases<VectorFormat>
           |> List.map
               (fun formatFrom ->
-                  makeTest formatFrom convertFun ((=) false) case
+                  makeTest formatFrom convertFun (Some convertFunUnsorted) ((=) false) case
                   |> testPropertyWithConfig config (getCorrectnessTestName "bool" formatFrom)) ]
         |> List.concat
     | Dense ->
@@ -89,15 +113,15 @@ let testFixtures case =
           Utils.listOfUnionCases<VectorFormat>
           |> List.map
               (fun formatFrom ->
-                  makeTest formatFrom convertFun ((=) 0) case
+                  makeTest formatFrom convertFun None ((=) 0) case
                   |> testPropertyWithConfig config (getCorrectnessTestName "int" formatFrom))
 
           let convertFun = Vector.toDense context wgSize
 
           Utils.listOfUnionCases<VectorFormat>
           |> List.map
               (fun formatFrom ->
-                  makeTest formatFrom convertFun ((=) false) case
+                  makeTest formatFrom convertFun None ((=) false) case
                   |> testPropertyWithConfig config (getCorrectnessTestName "bool" formatFrom)) ]
         |> List.concat