int LCAConn::update_dW(int axonId)
{ // compute dW but don't add them to the weights yet.
// That takes place in reduceKernels, so that the output is
// independent of the number of processors.
int nExt = preSynapticLayer()->getNumExtended();
int numKernelIndices = getNumDataPatches();
const pvdata_t * preactbuf = preSynapticLayer()->getLayerData(getDelay(axonId));
const pvdata_t * postactbuf = postSynapticLayer()->getLayerData(getDelay(axonId));
int sya = (post->getLayerLoc()->nf * (post->getLayerLoc()->nx + 2*post->getLayerLoc()->nb));
for(int kExt=0; kExt<nExt;kExt++) {
PVPatch * weights = getWeights(kExt,axonId);
size_t offset = getAPostOffset(kExt, axonId);
pvdata_t preact = preactbuf[kExt];
int ny = weights->ny;
int nk = weights->nx * nfp;
const pvdata_t * postactRef = &postactbuf[offset];
pvdata_t * dwdata = get_dwData(axonId, kExt);
int lineoffsetw = 0;
int lineoffseta = 0;
for( int y=0; y<ny; y++ ) {
for( int k=0; k<nk; k++ ) {
dwdata[lineoffsetw + k] += updateRule_dW(preact, postactRef[lineoffseta+k],lineoffseta+k);
}
lineoffsetw += syp;
lineoffseta += sya;
}
}
// Divide by (numNeurons/numKernels)
int divisor = pre->getNumNeurons()/numKernelIndices;
assert( divisor*numKernelIndices == pre->getNumNeurons() );
for( int kernelindex=0; kernelindex<numKernelIndices; kernelindex++ ) {
int numpatchitems = nxp*nyp*nfp;
pvdata_t * dwpatchdata = get_dwDataHead(axonId,kernelindex);
for( int n=0; n<numpatchitems; n++ ) {
dwpatchdata[n] /= divisor;
}
}
lastUpdateTime = parent->simulationTime();
return PV_SUCCESS;
}
int MomentumConn::applyMomentum(int arbor_ID){
int nExt = preSynapticLayer()->getNumExtended();
const PVLayerLoc * loc = preSynapticLayer()->getLayerLoc();
if(sharedWeights){
int numKernels = getNumDataPatches();
//Shared weights done in parallel, parallel in numkernels
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for(int kernelIdx = 0; kernelIdx < numKernels; kernelIdx++){
pvwdata_t * dwdata_start = get_dwDataHead(arbor_ID, kernelIdx);
pvwdata_t * prev_dw_start = get_prev_dwDataHead(arbor_ID, kernelIdx);
pvwdata_t * wdata_start = get_wDataHead(arbor_ID, kernelIdx);
if(!strcmp(momentumMethod, "simple")){
for(int k = 0; k < nxp*nyp*nfp; k++){
dwdata_start[k] += momentumTau * prev_dw_start[k] - momentumDecay*wdata_start[k];
}
}
else if(!strcmp(momentumMethod, "viscosity")){
for(int k = 0; k < nxp*nyp*nfp; k++){
//dwdata_start[k] = momentumTau * (prev_dw_start[k] + dwdata_start[k]) * (1 - exp(-1.0/ momentumTau)) - momentumDecay*wdata_start[k];
dwdata_start[k] = (prev_dw_start[k] * exp(-1.0/ momentumTau)) + dwdata_start[k] - momentumDecay*wdata_start[k];
}
}
else if(!strcmp(momentumMethod, "alex")){
for(int k = 0; k < nxp*nyp*nfp; k++){
//weight_inc[i] := momW * weight_inc[i-1] - wc * epsW * weights[i-1] + epsW * weight_grads[i]
// weights[i] := weights[i-1] + weight_inc[i]
dwdata_start[k] = momentumTau * prev_dw_start[k] - momentumDecay * getDWMax()* wdata_start[k] + dwdata_start[k];
}
}
}
}
else{
std::cout << "Warning: Momentum not implemented for non-shared weights, not implementing momentum\n";
}
return PV_SUCCESS;
}
int PoolingConn::deliverPostsynapticPerspective(PVLayerCube const * activity, int arborID) {
//Check channel number for noupdate
if(getChannel() == CHANNEL_NOUPDATE) {
return PV_SUCCESS;
}
assert(post->getChannel(getChannel()));
assert(arborID >= 0);
//Get number of neurons restricted target
const int numPostRestricted = post->getNumNeurons();
float dt_factor = getConvertToRateDeltaTimeFactor();
const PVLayerLoc * sourceLoc = preSynapticLayer()->getLayerLoc();
const PVLayerLoc * targetLoc = post->getLayerLoc();
const int sourceNx = sourceLoc->nx;
const int sourceNy = sourceLoc->ny;
const int sourceNf = sourceLoc->nf;
const int targetNx = targetLoc->nx;
const int targetNy = targetLoc->ny;
const int targetNf = targetLoc->nf;
const PVHalo * sourceHalo = &sourceLoc->halo;
const PVHalo * targetHalo = &targetLoc->halo;
//get source layer's extended y stride
int sy = (sourceNx+sourceHalo->lt+sourceHalo->rt)*sourceNf;
//The start of the gsyn buffer
pvdata_t * gSynPatchHead = post->getChannel(this->getChannel());
clearGateIdxBuffer();
int* gatePatchHead = NULL;
if(needPostIndexLayer) {
gatePatchHead = postIndexLayer->getChannel(CHANNEL_EXC);
}
long * startSourceExtBuf = getPostToPreActivity();
if(!startSourceExtBuf) {
std::cout << "HyPerLayer::recvFromPost error getting preToPostActivity from connection. Is shrink_patches on?\n";
exit(EXIT_FAILURE);
}
float resetVal = 0;
if(getPvpatchAccumulateType() == ACCUMULATE_MAXPOOLING) {
resetVal = -INFINITY;
}
for(int b = 0; b < parent->getNBatch(); b++) {
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for (int kTargetRes = 0; kTargetRes < numPostRestricted; kTargetRes++) {
pvdata_t * activityBatch = activity->data + b * (sourceNx + sourceHalo->rt + sourceHalo->lt) * (sourceNy + sourceHalo->up + sourceHalo->dn) * sourceNf;
pvdata_t * gSynPatchHeadBatch = gSynPatchHead + b * targetNx * targetNy * targetNf;
//Change restricted to extended post neuron
int kTargetExt = kIndexExtended(kTargetRes, targetNx, targetNy, targetNf, targetHalo->lt, targetHalo->rt, targetHalo->dn, targetHalo->up);
//Read from buffer
long startSourceExt = startSourceExtBuf[kTargetRes];
//Calculate target's start of gsyn
pvdata_t * gSynPatchPos = gSynPatchHeadBatch + kTargetRes;
//Initialize patch as a huge negative number
*gSynPatchPos = resetVal;
int* gatePatchPos = NULL;
if(needPostIndexLayer) {
gatePatchPos = gatePatchHead + b * postIndexLayer->getNumNeurons() + kTargetRes;
//Initialize gatePatchPos as a negative number
*gatePatchPos = -1;
}
float* activityStartBuf = &(activityBatch[startSourceExt]);
pvwdata_t * weightY = NULL; //No weights in pooling
int sf = postConn->fPatchSize();
int yPatchSize = postConn->yPatchSize();
int numPerStride = postConn->xPatchSize() * postConn->fPatchSize();
const PVLayerLoc * postLoc = post->getLayerLoc();
const int kfPost = featureIndex(kTargetExt, postLoc->nx + postLoc->halo.lt + postLoc->halo.rt, postLoc->ny + postLoc->halo.dn + postLoc->halo.up, postLoc->nf);
int offset = kfPost;
pvwdata_t w = 1.0;
if(getPvpatchAccumulateType() == ACCUMULATE_SUMPOOLING) {
float relative_XScale = pow(2, (post->getXScale() - pre->getXScale()));
float relative_YScale = pow(2, (post->getYScale() - pre->getYScale()));
w = 1.0/(nxp*nyp*relative_XScale*relative_YScale);
}
for (int ky = 0; ky < yPatchSize; ky++) {
int kPreExt = startSourceExt + ky*sy+offset;
const int kxPreExt = kxPos(kPreExt, sourceLoc->nx + sourceLoc->halo.lt + sourceLoc->halo.rt, sourceLoc->ny + sourceLoc->halo.dn + sourceLoc->halo.up, sourceLoc->nf);
const int kyPreExt = kyPos(kPreExt, sourceLoc->nx + sourceLoc->halo.lt + sourceLoc->halo.rt, sourceLoc->ny + sourceLoc->halo.dn + sourceLoc->halo.up, sourceLoc->nf);
const int kfPre = featureIndex(kPreExt, sourceLoc->nx + sourceLoc->halo.lt + sourceLoc->halo.rt, sourceLoc->ny + sourceLoc->halo.dn + sourceLoc->halo.up, sourceLoc->nf);
const int kxPreGlobalExt = kxPreExt + sourceLoc->kx0;
const int kyPreGlobalExt = kyPreExt + sourceLoc->ky0;
const int kPreGlobalExt = kIndex(kxPreGlobalExt, kyPreGlobalExt, kfPre, sourceLoc->nxGlobal + sourceLoc->halo.lt + sourceLoc->halo.rt, sourceLoc->nyGlobal + sourceLoc->halo.up + sourceLoc->halo.dn, sourceLoc->nf);
float * activityY = &(activityStartBuf[ky*sy+offset]);
(accumulateFunctionFromPostPointer)(kPreGlobalExt, numPerStride, gSynPatchPos, activityY, &w, dt_factor, gatePatchPos, sf);
}
}
}
return PV_SUCCESS;
}
int PoolingConn::deliverPresynapticPerspective(PVLayerCube const * activity, int arborID) {
//Check if we need to update based on connection's channel
if(getChannel() == CHANNEL_NOUPDATE) {
return PV_SUCCESS;
}
assert(post->getChannel(getChannel()));
float dt_factor;
if (getPvpatchAccumulateType()==ACCUMULATE_STOCHASTIC) {
dt_factor = getParent()->getDeltaTime();
}
else {
dt_factor = getConvertToRateDeltaTimeFactor();
}
const PVLayerLoc * preLoc = preSynapticLayer()->getLayerLoc();
const PVLayerLoc * postLoc = postSynapticLayer()->getLayerLoc();
assert(arborID >= 0);
const int numExtended = activity->numItems;
float resetVal = 0;
if(getPvpatchAccumulateType() == ACCUMULATE_MAXPOOLING) {
resetVal = -INFINITY;
float* gSyn = post->getChannel(getChannel());
//gSyn is res
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for(int i = 0; i < post->getNumNeuronsAllBatches(); i++) {
gSyn[i] = resetVal;
}
}
clearGateIdxBuffer();
for(int b = 0; b < parent->getNBatch(); b++) {
pvdata_t * activityBatch = activity->data + b * (preLoc->nx + preLoc->halo.rt + preLoc->halo.lt) * (preLoc->ny + preLoc->halo.up + preLoc->halo.dn) * preLoc->nf;
pvdata_t * gSynPatchHeadBatch = post->getChannel(getChannel()) + b * postLoc->nx * postLoc->ny * postLoc->nf;
int* gatePatchHeadBatch = NULL;
if(needPostIndexLayer) {
gatePatchHeadBatch = postIndexLayer->getChannel(CHANNEL_EXC) + b * postIndexLayer->getNumNeurons();
}
unsigned int * activeIndicesBatch = NULL;
if(activity->isSparse) {
activeIndicesBatch = activity->activeIndices + b * (preLoc->nx + preLoc->halo.rt + preLoc->halo.lt) * (preLoc->ny + preLoc->halo.up + preLoc->halo.dn) * preLoc->nf;
}
int numLoop;
if(activity->isSparse) {
numLoop = activity->numActive[b];
}
else {
numLoop = numExtended;
}
if(thread_gateIdxBuffer) {
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for(int i = 0; i < parent->getNumThreads() * post->getNumNeurons(); i++) {
int ti = i/post->getNumNeurons();
int ni = i % post->getNumNeurons();
thread_gateIdxBuffer[ti][ni] = -1;
}
}
#ifdef PV_USE_OPENMP_THREADS
//Clear all gsyn buffers
if(thread_gSyn) {
int numNeurons = post->getNumNeurons();
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for(int i = 0; i < parent->getNumThreads() * numNeurons; i++) {
int ti = i/numNeurons;
int ni = i % numNeurons;
thread_gSyn[ti][ni] = resetVal;
}
}
#endif // PV_USE_OPENMP_THREADS
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for schedule(static)
#endif
for (int loopIndex = 0; loopIndex < numLoop; loopIndex++) {
int kPreExt;
if(activity->isSparse) {
kPreExt = activeIndicesBatch[loopIndex];
}
else {
kPreExt = loopIndex;
}
float a = activityBatch[kPreExt] * dt_factor;
//if (a == 0.0f) continue;
//If we're using thread_gSyn, set this here
pvdata_t * gSynPatchHead;
//float * gatePatchHead = NULL;
int * gatePatchHead = NULL;
#ifdef PV_USE_OPENMP_THREADS
if(thread_gSyn) {
int ti = omp_get_thread_num();
gSynPatchHead = thread_gSyn[ti];
}
else {
gSynPatchHead = gSynPatchHeadBatch;
}
if(needPostIndexLayer) {
if(thread_gateIdxBuffer) {
int ti = omp_get_thread_num();
gatePatchHead = thread_gateIdxBuffer[ti];
}
else {
gatePatchHead = gatePatchHeadBatch;
}
}
#else // PV_USE_OPENMP_THREADS
gSynPatchHead = gSynPatchHeadBatch;
if(needPostIndexLayer) {
gatePatchHead = gatePatchHeadBatch;
}
#endif // PV_USE_OPENMP_THREADS
//deliverOnePreNeuronActivity(kPreExt, arborID, a, gSynPatchHead, gatePatchHead);
PVPatch * weights = getWeights(kPreExt, arborID);
const int nk = weights->nx * fPatchSize();
const int ny = weights->ny;
const int sy = getPostNonextStrides()->sy; // stride in layer
pvwdata_t * weightDataStart = NULL;
pvgsyndata_t * postPatchStart = gSynPatchHead + getGSynPatchStart(kPreExt, arborID);
int* postGatePatchStart = gatePatchHead + getGSynPatchStart(kPreExt, arborID);
//float* postGatePatchStart = gatePatchHead + getGSynPatchStart(kPreExt, arborID);
const int kxPreExt = kxPos(kPreExt, preLoc->nx + preLoc->halo.lt + preLoc->halo.rt, preLoc->ny + preLoc->halo.dn + preLoc->halo.up, preLoc->nf);
const int kyPreExt = kyPos(kPreExt, preLoc->nx + preLoc->halo.lt + preLoc->halo.rt, preLoc->ny + preLoc->halo.dn + preLoc->halo.up, preLoc->nf);
const int kfPre = featureIndex(kPreExt, preLoc->nx + preLoc->halo.lt + preLoc->halo.rt, preLoc->ny + preLoc->halo.dn + preLoc->halo.up, preLoc->nf);
const int kxPreGlobalExt = kxPreExt + preLoc->kx0;
const int kyPreGlobalExt = kyPreExt + preLoc->ky0;
const int kPreGlobalExt = kIndex(kxPreGlobalExt, kyPreGlobalExt, kfPre, preLoc->nxGlobal + preLoc->halo.lt + preLoc->halo.rt, preLoc->nyGlobal + preLoc->halo.up + preLoc->halo.dn, preLoc->nf);
int offset = kfPre;
int sf = fPatchSize();
pvwdata_t w = 1.0;
if(getPvpatchAccumulateType() == ACCUMULATE_SUMPOOLING) {
float relative_XScale = pow(2, (post->getXScale() - pre->getXScale()));
float relative_YScale = pow(2, (post->getYScale() - pre->getYScale()));
w = 1.0/(nxp*nyp*relative_XScale*relative_YScale);
}
void* auxPtr = NULL;
for (int y = 0; y < ny; y++) {
if(needPostIndexLayer) {
auxPtr = (postGatePatchStart+ y*sy + offset);
}
(accumulateFunctionPointer)(kPreGlobalExt, nk, postPatchStart + y*sy + offset, a, &w, auxPtr, sf);
}
}
#ifdef PV_USE_OPENMP_THREADS
//Accumulate back into gSyn // Should this be done in HyPerLayer where it can be done once, as opposed to once per connection?
if(thread_gSyn) {
pvdata_t * gSynPatchHead = gSynPatchHeadBatch;
//float* gateIdxBuffer = postIndexLayer->getChannel(CHANNEL_EXC);
int * gateIdxBuffer = NULL;
if(needPostIndexLayer && thread_gateIdxBuffer) {
gateIdxBuffer = gatePatchHeadBatch;
}
int numNeurons = post->getNumNeurons();
//Looping over neurons first to be thread safe
#pragma omp parallel for
for(int ni = 0; ni < numNeurons; ni++) {
//Different for maxpooling
if(getPvpatchAccumulateType() == ACCUMULATE_MAXPOOLING) {
for(int ti = 0; ti < parent->getNumThreads(); ti++) {
if(gSynPatchHead[ni] < thread_gSyn[ti][ni]) {
gSynPatchHead[ni] = thread_gSyn[ti][ni];
if(needPostIndexLayer && thread_gateIdxBuffer) {
gateIdxBuffer[ni] = thread_gateIdxBuffer[ti][ni];
assert(gateIdxBuffer >= 0);
}
}
}
}
else {
for(int ti = 0; ti < parent->getNumThreads(); ti++) {
gSynPatchHead[ni] += thread_gSyn[ti][ni];
}
}
}
}
#endif
}
if(activity->isSparse) {
pvdata_t * gSyn = post->getChannel(getChannel());
for (int k=0; k<post->getNumNeuronsAllBatches(); k++) {
if (gSyn[k]==-INFINITY) {
gSyn[k] = 0.0f;
}
}
}
return PV_SUCCESS;
}
int LCALIFLateralKernelConn::update_dW(int axonId) {
if (parent->simulationTime() < dWUpdateTime) {
return PV_SUCCESS;
}
dWUpdateTime += dWUpdatePeriod;
int nExt = preSynapticLayer()->getNumExtended();
int numKernelIndices = getNumDataPatches();
updateIntegratedSpikeCount();
float target_rate_sq = getTargetRateKHz()*getTargetRateKHz();
const pvdata_t * preactbuf = integratedSpikeCount;
const pvdata_t * postactbuf = integratedSpikeCount;
int sya = (post->getLayerLoc()->nf * (post->getLayerLoc()->nx + post->getLayerLoc()->halo.lt + post->getLayerLoc()->halo.rt));
const PVLayerLoc * preloc = pre->getLayerLoc();
int nxpre = preloc->nx;
int nypre = preloc->ny;
int nfpre = preloc->nf;
int nxglob = preloc->nxGlobal;
int nyglob = preloc->nyGlobal;
int kx0 = preloc->kx0;
int ky0 = preloc->ky0;
for(int kExt=0; kExt<nExt;kExt++) {
int xglob = kxPos(kExt, nxpre + preloc->halo.lt + preloc->halo.rt, nypre + preloc->halo.dn + preloc->halo.up, nfpre) + kx0 - preloc->halo.lt;
int yglob = kyPos(kExt, nxpre + preloc->halo.lt + preloc->halo.rt, nypre + preloc->halo.dn + preloc->halo.up, nfpre) + ky0 - preloc->halo.dn;
if (xglob < 0 || xglob >= nxglob || yglob < 0 || yglob >= nyglob) {
continue;
}
PVPatch * weights = getWeights(kExt,axonId);
size_t offset = getAPostOffset(kExt, axonId);
pvdata_t preactrate = preactbuf[kExt]/integrationTimeConstant;
int ny = weights->ny;
int nk = weights->nx * nfp;
pvwdata_t * dwdata = get_dwData(axonId, kExt);
int lineoffsetw = 0;
int lineoffseta = 0;
for( int y=0; y<ny; y++ ) {
for( int k=0; k<nk; k++ ) {
int postactindex = offset+lineoffseta+k;
if (postactindex != kExt) { // Neurons don't inhibit themselves
pvdata_t postactrate = postactbuf[postactindex]/integrationTimeConstant;
pvdata_t dw = preactrate*postactrate-target_rate_sq;
dwdata[lineoffsetw + k] += dw;
}
}
lineoffsetw += syp;
lineoffseta += sya;
}
}
// Divide each dw by the number of correlations that contributed to that dw (divisorptr was summed over all MPI processes in initialization).
// Also divide by target_rate_sq to normalize to a dimensionless quantity.
// The nonlinear filter and the multiplication by dt/tauINH takes place in updateWeights, because the filter has to be applied after reduceKernels
// and the multiplication by dt/tauINH needs to take place after the filter.
int patch_size = nxp*nyp*nfp;
for( int kernelindex=0; kernelindex<numKernelIndices; kernelindex++ ) {
pvwdata_t * dwpatchdata = get_dwDataHead(axonId,kernelindex);
float * divisorptr = &interiorCounts[axonId][kernelindex*patch_size];
for( int n=0; n<patch_size; n++ ) {
assert(divisorptr[n]>0 || dwpatchdata[n]==0);
if (divisorptr[n]>0) dwpatchdata[n] /= target_rate_sq * divisorptr[n];
}
}
lastUpdateTime = parent->simulationTime();
return PV_SUCCESS;
}
int TransposePoolingConn::deliverPresynapticPerspective(PVLayerCube const * activity, int arborID) {
//Check if we need to update based on connection's channel
if(getChannel() == CHANNEL_NOUPDATE){
return PV_SUCCESS;
}
assert(post->getChannel(getChannel()));
const PVLayerLoc * preLoc = preSynapticLayer()->getLayerLoc();
const PVLayerLoc * postLoc = postSynapticLayer()->getLayerLoc();
assert(arborID >= 0);
const int numExtended = activity->numItems;
//Grab postIdxLayer's data
int* postIdxData = NULL;
if(pvpatchAccumulateType == ACCUMULATE_MAXPOOLING){
PoolingIndexLayer* postIndexLayer = originalConn->getPostIndexLayer();
assert(postIndexLayer);
//Make sure this layer is an integer layer
assert(postIndexLayer->getDataType() == PV_INT);
DataStore * store = parent->icCommunicator()->publisherStore(postIndexLayer->getLayerId());
int delay = getDelay(arborID);
//TODO this is currently a hack, need to properly implement data types.
postIdxData = (int*) store->buffer(LOCAL, delay);
}
for(int b = 0; b < parent->getNBatch(); b++){
pvdata_t * activityBatch = activity->data + b * (preLoc->nx + preLoc->halo.rt + preLoc->halo.lt) * (preLoc->ny + preLoc->halo.up + preLoc->halo.dn) * preLoc->nf;
pvdata_t * gSynPatchHeadBatch = post->getChannel(getChannel()) + b * postLoc->nx * postLoc->ny * postLoc->nf;
int * postIdxDataBatch = NULL;
if(pvpatchAccumulateType == ACCUMULATE_MAXPOOLING){
postIdxDataBatch = postIdxData + b * originalConn->getPostIndexLayer()->getNumExtended();
}
unsigned int * activeIndicesBatch = NULL;
if(activity->isSparse){
activeIndicesBatch = activity->activeIndices + b * (preLoc->nx + preLoc->halo.rt + preLoc->halo.lt) * (preLoc->ny + preLoc->halo.up + preLoc->halo.dn) * preLoc->nf;
}
int numLoop;
if(activity->isSparse){
numLoop = activity->numActive[b];
}
else{
numLoop = numExtended;
}
#ifdef PV_USE_OPENMP_THREADS
//Clear all thread gsyn buffer
if(thread_gSyn){
int numNeurons = post->getNumNeurons();
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for(int i = 0; i < parent->getNumThreads() * numNeurons; i++){
int ti = i/numNeurons;
int ni = i % numNeurons;
thread_gSyn[ti][ni] = 0;
}
}
#endif // PV_USE_OPENMP_THREADS
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for schedule(static)
#endif
for (int loopIndex = 0; loopIndex < numLoop; loopIndex++) {
int kPreExt;
if(activity->isSparse){
kPreExt = activeIndicesBatch[loopIndex];
}
else{
kPreExt = loopIndex;
}
float a = activityBatch[kPreExt];
if (a == 0.0f) continue;
//If we're using thread_gSyn, set this here
pvdata_t * gSynPatchHead;
#ifdef PV_USE_OPENMP_THREADS
if(thread_gSyn){
int ti = omp_get_thread_num();
gSynPatchHead = thread_gSyn[ti];
}
else{
gSynPatchHead = gSynPatchHeadBatch;
}
#else // PV_USE_OPENMP_THREADS
gSynPatchHead = gSynPatchHeadBatch;
#endif // PV_USE_OPENMP_THREADS
const int kxPreExt = kxPos(kPreExt, preLoc->nx + preLoc->halo.lt + preLoc->halo.rt, preLoc->ny + preLoc->halo.dn + preLoc->halo.up, preLoc->nf);
const int kyPreExt = kyPos(kPreExt, preLoc->nx + preLoc->halo.lt + preLoc->halo.rt, preLoc->ny + preLoc->halo.dn + preLoc->halo.up, preLoc->nf);
const int kfPre = featureIndex(kPreExt, preLoc->nx + preLoc->halo.lt + preLoc->halo.rt, preLoc->ny + preLoc->halo.dn + preLoc->halo.up, preLoc->nf);
if(pvpatchAccumulateType == ACCUMULATE_MAXPOOLING){
const int kxPreGlobalExt = kxPreExt + preLoc->kx0;
const int kyPreGlobalExt = kyPreExt + preLoc->ky0;
if(kxPreGlobalExt < preLoc->halo.lt || kxPreGlobalExt >= preLoc->nxGlobal + preLoc->halo.lt ||
kyPreGlobalExt < preLoc->halo.up || kyPreGlobalExt >= preLoc->nyGlobal + preLoc->halo.up){
continue;
}
//Convert stored global extended index into local extended index
int postGlobalExtIdx = postIdxDataBatch[kPreExt];
// If all inputs are zero and input layer is sparse, postGlobalExtIdx will still be -1.
if(postGlobalExtIdx == -1) { continue; }
//Make sure the index is in bounds
assert(postGlobalExtIdx >= 0 && postGlobalExtIdx <
(postLoc->nxGlobal + postLoc->halo.lt + postLoc->halo.rt) *
(postLoc->nyGlobal + postLoc->halo.up + postLoc->halo.dn) *
postLoc->nf);
const int kxPostGlobalExt = kxPos(postGlobalExtIdx, postLoc->nxGlobal + postLoc->halo.lt + postLoc->halo.rt, postLoc->nyGlobal + postLoc->halo.dn + postLoc->halo.up, postLoc->nf);
const int kyPostGlobalExt = kyPos(postGlobalExtIdx, postLoc->nxGlobal + postLoc->halo.lt + postLoc->halo.rt, postLoc->nyGlobal + postLoc->halo.dn + postLoc->halo.up, postLoc->nf);
const int kfPost = featureIndex(postGlobalExtIdx, postLoc->nxGlobal + postLoc->halo.lt + postLoc->halo.rt, postLoc->nyGlobal + postLoc->halo.dn + postLoc->halo.up, postLoc->nf);
const int kxPostLocalRes = kxPostGlobalExt - postLoc->kx0 - postLoc->halo.lt;
const int kyPostLocalRes = kyPostGlobalExt - postLoc->ky0 - postLoc->halo.up;
if(kxPostLocalRes < 0 || kxPostLocalRes >= postLoc->nx||
kyPostLocalRes < 0 || kyPostLocalRes >= postLoc->ny){
continue;
}
const int kPostLocalRes = kIndex(kxPostLocalRes, kyPostLocalRes, kfPost, postLoc->nx, postLoc->ny, postLoc->nf);
gSynPatchHeadBatch[kPostLocalRes] = a;
}
else{
PVPatch * weights = getWeights(kPreExt, arborID);
const int nk = weights->nx * fPatchSize();
const int ny = weights->ny;
pvgsyndata_t * postPatchStart = gSynPatchHead + getGSynPatchStart(kPreExt, arborID);
const int sy = getPostNonextStrides()->sy; // stride in layer
int offset = kfPre;
int sf = fPatchSize();
pvwdata_t w = 1.0;
if(getPvpatchAccumulateType() == ACCUMULATE_SUMPOOLING){
float relative_XScale = pow(2, (post->getXScale() - pre->getXScale()));
float relative_YScale = pow(2, (post->getYScale() - pre->getYScale()));
w = 1.0/(nxp*nyp*relative_XScale*relative_YScale);
}
void* auxPtr = NULL;
for (int y = 0; y < ny; y++) {
(accumulateFunctionPointer)(0, nk, postPatchStart + y*sy + offset, a, &w, auxPtr, sf);
}
}
}
#ifdef PV_USE_OPENMP_THREADS
//Set back into gSyn
if(thread_gSyn){
pvdata_t * gSynPatchHead = gSynPatchHeadBatch;
int numNeurons = post->getNumNeurons();
//Looping over neurons first to be thread safe
#pragma omp parallel for
for(int ni = 0; ni < numNeurons; ni++){
for(int ti = 0; ti < parent->getNumThreads(); ti++){
if(pvpatchAccumulateType == ACCUMULATE_MAXPOOLING){
if(gSynPatchHead[ni] < fabs(thread_gSyn[ti][ni])){
gSynPatchHead[ni] = thread_gSyn[ti][ni];
}
}
else{
gSynPatchHead[ni] += thread_gSyn[ti][ni];
}
}
}
}
#endif
}
return PV_SUCCESS;
}
