Cycle args for optimizing, gpu syncing improvements, etc

emer · Nov 21, 2024 · 69345bd · 69345bd
1 parent 30f657d
commit 69345bd
Show file tree

Hide file tree

Showing 18 changed files with 711 additions and 759 deletions.
diff --git a/axon/act-layer.go b/axon/act-layer.go
diff --git a/axon/act-layer.goal b/axon/act-layer.goal
@@ -290,7 +290,7 @@ func (ly *LayerParams) GInteg(ctx *Context, pi, ni, di uint32) {
 // drvAct is for Pulvinar layers, activation of driving neuron
 func (ly *LayerParams) SpecialPreGs(ctx *Context, pi, ni, di uint32, drvGe float32, nonDrivePct float32) float32 {
 	saveVal := float32(0)               // sometimes we need to use a value computed here, for the post Gs step
-	pi0 := pi - 1 // 0-n pool index
+	pi0 := pi - ly.PoolSt - 1 // 0-n pool index
 	pnn := uint32(PoolNNeurons(pi))
 	pni := NeuronIxs[NrnNeurIndex, ni] - uint32(PoolsInt[PoolNeurSt, pi, di])
 	nrnCtxtGe := Neurons[CtxtGe, ni, di]
@@ -462,7 +462,7 @@ func (ly *LayerParams) GFromRawSyn(ctx *Context, ni, di uint32) {
 		extraSyn = md * nrnGModSyn * ly.Acts.Dend.ModGain
 	default:
 		if ly.Acts.Dend.HasMod.IsTrue() {
-			md := ly.Acts.Dend.ModBase + ly.Acts.Dend.ModGain*nrnGModSyn
+			md := ly.Acts.Dend.ModBase + ly.Acts.Dend.ModGain * nrnGModSyn
 			if md > 1 {
 				md = 1
 			}
@@ -566,7 +566,7 @@ func (ly *LayerParams) SendSpike(ctx *Context, ni, di uint32) {
 func (ly *LayerParams) PostSpikeSpecial(ctx *Context, lpi, pi, ni, di uint32) {
 	Neurons[Burst, ni, di] = Neurons[CaSpkP, ni, di]
 	li := ly.Index
-	pi0 := pi - 1 // 0-n pool index
+	pi0 := pi - ly.PoolSt - 1 // 0-n pool index
 	pnn := uint32(PoolNNeurons(pi))
 	pni := NeuronIxs[NrnNeurIndex, ni] - uint32(PoolsInt[PoolNeurSt, pi, di])
 	hasRew := GlobalScalars[GvHasRew, di] > 0

diff --git a/axon/act-net.go b/axon/act-net.go
diff --git a/axon/act-net.goal b/axon/act-net.goal
@@ -4,24 +4,36 @@
 
 package axon
 
-// Cycle runs one cycle of activation updating using threading methods.
-func (nt *Network) Cycle() {
-	// todo: chunks of 10 cycles
+// todo: don't even need layer-level ultimately.
+
+// Cycle runs n cycles of activation updating.
+// If getNeurons is true, then neuron state is synced back
+// from the GPU (for cycle-level display etc). Otherwise only
+// layer-level state is synced.
+func (nt *Network) Cycle(ncyc int, getNeurons bool) {
 	nix := nt.NetIxs()
 	ctx := nt.Context()
 	nd := int(nix.NNeurons * ctx.NData)
 	ld := int(nix.NLayers * ctx.NData)
 	pd := int(nix.NPools * ctx.NData)
-	RunGatherSpikes(nd)
-	RunLayerGi(ld)
-	RunBetweenGi(ld)
-	RunPoolGi(pd)
-	RunCycleNeuron(nd)
-	RunSendSpike(nd)
-	RunCyclePost(ld)
-	RunCycleInc(1)
 
-	RunDoneLayers()
+	ToGPUCtxGlobal()
+	for range ncyc {
+		RunGatherSpikes(nd)
+		RunLayerGi(ld)
+		RunBetweenGi(ld)
+		RunPoolGi(pd)
+		RunCycleNeuron(nd)
+		RunSendSpike(nd)
+		RunCyclePost(ld)
+		RunCycleInc(1)
+	}
+
+	if getNeurons {
+		RunDoneLayersNeurons()
+	} else {
+		RunDoneLayers()
+	}
 
 	// todo: fix this:
 	// var ldt, vta *Layer
@@ -64,6 +76,7 @@ func (nt *Network) ApplyExts() {
 	if !UseGPU {
 		return
 	}
+	ToGPU(ExtsVar)
 	nix := nt.NetIxs()
 	ctx := nt.Context()
 	nd := int(nix.NNeurons * ctx.NData)
@@ -79,6 +92,7 @@ func (nt *Network) MinusPhase() {
 	RunMinusPhasePool(pd)
 	RunMinusPhaseNeuron(nd)
 	nt.MinusPhasePost()
+	ToGPULayersNeurons()
 	// todo:
 	// nt.GPU.SyncStateToGPU()
 }
@@ -112,6 +126,7 @@ func (nt *Network) PlusPhase() {
 	RunPlusPhasePool(pd)
 	RunPlusPhaseNeuron(nd)
 	nt.PlusPhasePost()
+	ToGPULayersNeurons()
 	// todo:
 	// nt.GPU.SyncStateToGPU()
 }