refactor: PrefixSum

Author: IgorErin

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

@@ -8,10 +8,10 @@ module SubSum =
            fun step lid wgSize (localBuffer: 'a []) ->
                 let i = step * (lid + 1) - 1
 
-                let buff =
-                    (%opAdd) localBuffer.[i - (step >>> 1)] localBuffer.[i]
+                let firstValue = localBuffer.[i - (step >>> 1)]
+                let secondValue = localBuffer.[i]
 
-                localBuffer.[i] <- buff
+                localBuffer.[i] <- (%opAdd) firstValue secondValue
         @>
 
     let private sequentialAccess<'a> opAdd =
@@ -23,7 +23,7 @@ module SubSum =
                 localBuffer.[lid] <- (%opAdd) firstValue secondValue
         @>
 
-    let sumGeneral<'a> memoryAccess =
+    let private sumGeneral<'a> memoryAccess =
         <@
             fun wgSize lid (localBuffer: 'a []) ->
                 let mutable step = 2
@@ -39,7 +39,7 @@ module SubSum =
 
     let sequentialSum<'a> opAdd = sumGeneral<'a> <| sequentialAccess<'a> opAdd
 
-    let treeSum<'a> opAdd = sumGeneral<'a> <| treeAccess opAdd
+    let treeSum<'a> opAdd = sumGeneral<'a> <| treeAccess<'a> opAdd
 
 module PreparePositions =
     let both<'c> =

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

@@ -1,6 +1,7 @@
 namespace GraphBLAS.FSharp.Backend.Common
 
 open Brahma.FSharp
+open GraphBLAS.FSharp.Backend
 open Microsoft.FSharp.Quotations
 
 module internal PrefixSum =
@@ -47,6 +48,8 @@ module internal PrefixSum =
         workGroupSize
         =
 
+        let subSum = SubSum.treeSum opAdd
+
         let scan =
             <@ fun (ndRange: Range1D) inputArrayLength verticesLength (resultBuffer: ClArray<'a>) (verticesBuffer: ClArray<'a>) (totalSumBuffer: ClCell<'a>) (zero: ClCell<'a>) (mirror: ClCell<bool>) ->
 
@@ -68,23 +71,10 @@ module internal PrefixSum =
                 else
                     resultLocalBuffer.[localID] <- zero
 
-                let mutable step = 2
-
-                while step <= workGroupSize do
-                    barrierLocal ()
-
-                    if localID < workGroupSize / step then
-                        let i = step * (localID + 1) - 1
-
-                        let buff =
-                            (%opAdd) resultLocalBuffer.[i - (step >>> 1)] resultLocalBuffer.[i]
-
-                        resultLocalBuffer.[i] <- buff
-
-                    step <- step <<< 1
-
                 barrierLocal ()
 
+                (%subSum) workGroupSize localID resultLocalBuffer
+
                 if localID = workGroupSize - 1 then
                     if verticesLength <= 1 && localID = gid then
                         totalSumBuffer.Value <- resultLocalBuffer.[localID]
@@ -93,7 +83,7 @@ module internal PrefixSum =
                     (%beforeLocalSumClear) resultBuffer resultLocalBuffer.[localID] inputArrayLength gid i
                     resultLocalBuffer.[localID] <- zero
 
-                step <- workGroupSize
+                let mutable step = workGroupSize
 
                 while step > 1 do
                     barrierLocal ()

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

@@ -54,8 +54,6 @@ module internal Sum =
 
             processor.Post(Msg.CreateRunMsg<_, _>(kernel))
 
-            ()
-
     let private scanToCell
         (clContext: ClContext)
         (workGroupSize: int)