GPU weird bug in NewState

[rcoreilly]

There is a very strange bug in GPU NewState:

layerparams.go:

// NewStateNeuron handles all initialization at start of new input pattern.
// Should already have presented the external input to the network at this point.
func (ly *LayerParams) NewStateNeuron(ctx *Context, ni, di uint32, vals *LayerVals) {
    SetNrnV(ctx, ni, di, BurstPrv, NrnV(ctx, ni, di, Burst))
    SetNrnV(ctx, ni, di, SpkPrv, NrnV(ctx, ni, di, CaSpkD))

This works fine in CPU mode, and is clearly free of an actual coding error, but the GPU version returns 0 for those NrnV access calls of the Burst and CaSpkD variables. It sets the new values just fine, including setting Burst etc, but reading these values just plain fails for reasons that are entirely unclear.

Originally this code was being called from a Pool level dispatch, but changing it to neuron-level did not fix the problem (but is more performant).

There is some kind of read cache breakage here that we need to figure out. Meanwhile, using a temporary solution of just doing the NewState on CPU -- not a big deal performance-wise.

[rcoreilly]

Important: when switching back to NewState on GPU, need to update to use proper logic for NewStateLayerActAvg