Simplify logic by adding an extra array slot.

Adds SubgroupSize*4 bytes overhead though

Author: planetis-m

src/computesim.nim

@@ -160,7 +160,7 @@ proc workGroupProc[A, B, C](
     workgroupID: UVec3,
     wg: WorkGroupContext,
     compute: ThreadGenerator[A, B, C],
-    ssbo: A, smem: ptr B, args: C) {.nimcall.} =
+    ssbo: A, smem: ptr B, args: C) =
   # Auxiliary proc for work group management
   var wg = wg # Shadow for modification
   wg.gl_WorkGroupID = workgroupID

src/computesim/core.nim

@@ -132,14 +132,15 @@ type
                              thread: ThreadContext, threadId: uint32): SubgroupCommand
   SubgroupResults* = array[SubgroupSize, SubgroupResult]
   SubgroupCommands* = array[SubgroupSize, SubgroupCommand]
-  SubgroupThreadIDs* = array[SubgroupSize, uint32]
+  SubgroupThreadIDs* = array[SubgroupSize + 1, uint32]
   SubgroupThreads* = array[SubgroupSize, ThreadClosure]
   ThreadContexts* = array[SubgroupSize, ThreadContext]
 
 const
   InvalidId* = high(uint32) # Sentinel value for empty/invalid
 
 iterator threadsInGroup*(group: SubgroupThreadIDs): uint32 =
-  for member in group.items:
-    if member == InvalidId: break
-    yield member
+  var idx: uint32 = 0
+  while group[idx] != InvalidId:
+    yield group[idx]
+    inc idx

src/computesim/lockstep.nim

@@ -37,12 +37,14 @@ proc runThreads*(threads: SubgroupThreads; workGroup: WorkGroupContext,
   var
     anyThreadsActive = true
     allThreadsHalted = false
+    threadGroups {.noinit.}: array[SubgroupSize, SubgroupThreadIDs]
     threadStates {.noinit.}: array[SubgroupSize, ThreadState]
     commands {.noinit.}: SubgroupCommands
     results {.noinit.}: SubgroupResults
     minReconvergeId: uint32 = 0
     barrierId = InvalidId
     barrierThreadCount: uint32 = 0
+    numGroups: uint32 = 0
 
   shouldShowDebugOutput(showDebugOutput)
   threadStates.fill(running)
@@ -105,9 +107,7 @@ proc runThreads*(threads: SubgroupThreads; workGroup: WorkGroupContext,
                          barrierThreadCount, numActiveThreads)
 
     # Group matching operations
-    var
-      threadGroups {.noinit.}: array[SubgroupSize, SubgroupThreadIDs]
-      numGroups: uint32 = 0
+    numGroups = 0
 
     # Group by operation id
     for threadId in 0..<numActiveThreads:
@@ -122,8 +122,7 @@ proc runThreads*(threads: SubgroupThreads; workGroup: WorkGroupContext,
             for slot in 0..<SubgroupSize:
               if threadGroups[groupIdx][slot] == InvalidId:
                 threadGroups[groupIdx][slot] = threadId
-                if slot + 1 < SubgroupSize:
-                  threadGroups[groupIdx][slot + 1] = InvalidId
+                threadGroups[groupIdx][slot + 1] = InvalidId
                 break
             found = true
             break

tests/config.nims

@@ -5,6 +5,7 @@
 --define:"release"
 --define:"useMalloc"
 --define:"ThreadPoolSize=10"
+# --define:"SubgroupSize=32"
 --define:debugSubgroup
 switch("define", "debugSubgroupID:1")