//
// A replacement for globalIndexFromLocal from conversions.h.
// WARNING - any changes in conversions.h should be reflected here.
static inline int globalIndexFromLocal_nompi(int kl, PVLayerLoc loc)
{
int kxg = loc.kx0 + kxPos(kl, loc.nx, loc.ny, loc.nf);
int kyg = loc.ky0 + kyPos(kl, loc.nx, loc.ny, loc.nf);
int kf = featureIndex(kl, loc.nx, loc.ny, loc.nf);
return kIndex(kxg, kyg, kf, loc.nxGlobal, loc.nyGlobal, loc.nf);
}
int BIDSSensorLayer::updateState(double timef, double dt){
pvdata_t * output = getCLayer()->V;
pvdata_t * input = blayer->getCLayer()->activity->data;
int index;
//Iterate through post layer
for (int i = 0; i < nx * ny; i++){
assert(nf == 1);
//Iterate through features
// std::cout << "Node (" << coords[i].xCoord << ", " << coords[i].yCoord << ")\n";
for (int k = 0; k < nf; k++){
int x = i % nx;
int y = (int) floor(i/nx);
index = kIndex(x, y, k, nx, ny, nf);
data[i][buf_index] = input[index] - (neutral_val / 256);
// std::cout << "\tBuf_index: " << buf_index << ": " << data[i][buf_index] << "\n";
//Next buf index, or reset if at end
float out = matchFilter(i, (int)(timef * dt));
output[index] = out * weight;
}
}
if(buf_index < buf_size - 1){
buf_index++;
}
else{
buf_index = 0;
}
HyPerLayer::setActivity();
return PV_SUCCESS;
}
int ImageTestLayer::updateStateWrapper(double time, double dt)
{
Image::updateStateWrapper(time, dt);
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int nbatch = loc->nbatch;
for(int b = 0; b < nbatch; b++){
pvdata_t * dataBatch = data + b * getNumExtended();
for(int nkRes = 0; nkRes < getNumNeurons(); nkRes++){
//Calculate extended index
int nkExt = kIndexExtended(nkRes, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
//checkVal is the value from batch index 0
pvdata_t checkVal = dataBatch[nkExt] * 255;
int kxGlobal = kxPos(nkRes, nx, ny, nf) + loc->kx0;
int kyGlobal = kyPos(nkRes, nx, ny, nf) + loc->ky0;
int kf = featureIndex(nkRes, nx, ny, nf);
pvdata_t expectedVal = kIndex(kxGlobal, kyGlobal, kf, loc->nxGlobal, loc->nyGlobal, nf);
if(fabs(checkVal - expectedVal) >= 1e-5){
std::cout << "ImageFileIO test Expected: " << expectedVal << " Actual: " << checkVal << "\n";
exit(-1);
}
}
}
return PV_SUCCESS;
}
int KernelProbe::outputState(double timed) {
InterColComm * icComm = parent->icCommunicator();
const int rank = icComm->commRank();
if( rank != 0 ) return PV_SUCCESS;
assert(getTargetConn()!=NULL);
int nxp = getTargetHyPerConn()->xPatchSize();
int nyp = getTargetHyPerConn()->yPatchSize();
int nfp = getTargetHyPerConn()->fPatchSize();
int patchSize = nxp*nyp*nfp;
const pvwdata_t * wdata = getTargetHyPerConn()->get_wDataStart(arborID)+patchSize*kernelIndex;
const pvwdata_t * dwdata = outputPlasticIncr ?
getTargetHyPerConn()->get_dwDataStart(arborID)+patchSize*kernelIndex : NULL;
fprintf(outputstream->fp, "Time %f, Conn \"%s\", nxp=%d, nyp=%d, nfp=%d\n",
timed, getTargetConn()->getName(),nxp, nyp, nfp);
for(int f=0; f<nfp; f++) {
for(int y=0; y<nyp; y++) {
for(int x=0; x<nxp; x++) {
int k = kIndex(x,y,f,nxp,nyp,nfp);
fprintf(outputstream->fp, " x=%d, y=%d, f=%d (index %d):", x, y, f, k);
if(getOutputWeights()) {
fprintf(outputstream->fp, " weight=%f", (float)wdata[k]);
}
if(getOutputPlasticIncr()) {
fprintf(outputstream->fp, " dw=%f", (float)dwdata[k]);
}
fprintf(outputstream->fp,"\n");
}
}
}
return PV_SUCCESS;
}
int StreamReconLayer::updateState(double timef, double dt) {
update_timer->start();
pvdata_t * V = getV();
int nx = getLayerLoc()->nx;
int ny = getLayerLoc()->ny;
int nf = getLayerLoc()->nf;
for (int i = 0; i < nx; i++) {
int vx = i;
int gx = i;
for (int j = 0; j < nf; j ++) {
int vf = j;
int gf;
if (vf + bufferLevel < nf) {
gf = bufferLevel + j;
}
else {
gf = bufferLevel + j - nf;
}
int vindex = kIndex(vx, 0, vf, nx, ny, nf);
int gindex = kIndex(gx, 0, gf, nx, ny, nf);
V[vindex] = GSyn[0][gindex];
}
}
//Copy V to A buffer
PV::HyPerLayer::setActivity();
if (bufferLevel < nf - 1) {
bufferLevel++; }
else {
bufferLevel = 0;
}
update_timer->stop();
return PV_SUCCESS;
} // end update state
int dumponeweight(HyPerConn * conn) {
int status = PV_SUCCESS;
bool errorfound = false;
int rank = conn->getParent()->icCommunicator()->commRank();
int nxp = conn->xPatchSize();
int nyp = conn->yPatchSize();
int nfp = conn->fPatchSize();
int xcenter = (nxp-1)/2;
int ycenter = (nyp-1)/2;
int nxpre = conn->preSynapticLayer()->getLayerLoc()->nxGlobal;
int nypre = conn->preSynapticLayer()->getLayerLoc()->nyGlobal;
bool usingMirrorBCs = conn->preSynapticLayer()->useMirrorBCs();
// If xScaleDiff > 0, it's a many-to-one connection.
int xScaleDiff = conn->postSynapticLayer()->getXScale() - conn->preSynapticLayer()->getXScale();
float xFalloff = powf(2,xScaleDiff);
int yScaleDiff = conn->postSynapticLayer()->getYScale() - conn->preSynapticLayer()->getYScale();
float yFalloff = powf(2,yScaleDiff);
for( int p=0; p<conn->getNumDataPatches(); p++ ) {
pvwdata_t * wgtData = conn->get_wDataHead(0,p); // conn->getKernelPatch(0,p)->data;
for( int f=0; f<nfp; f++ ) {
for( int x=0; x<nxp; x++ ) {
int xoffset = abs((int) floor((x-xcenter)*xFalloff));
for( int y=0; y<nyp; y++ ) {
int yoffset = abs((int) floor((y-ycenter)*yFalloff));
int idx = kIndex(x, y, f, nxp, nyp, nfp);
//TODO-CER-2014.4.4 - weight conversion
pvdata_t wgt = wgtData[idx];
//pvdata_t correct = usingMirrorBCs ? 1 : (nxpre-xoffset)*(nypre-yoffset)/((pvdata_t) (nxpre*nypre));
//New normalization takes into account if pre is not active
//The pixel value from the input is actually 127, where we divide it by 255.
//Not exaclty .5, a little less
//Squared because both pre and post is grabbing it's activity from the image
pvdata_t correct = usingMirrorBCs ? pow(float(127)/float(255),2) : (float(127)/float(255)) * .5;
if( fabs(wgt-correct)>1.0e-5 ) {
pvErrorNoExit(errorMessage);
if( errorfound == false ) {
errorfound = true;
for( int k=0; k<72; k++ ) { pvInfo().printf("="); }
errorMessage.printf("\n");
errorMessage.printf("Rank %d, Connection \"%s\":\n",rank, conn->getName());
}
errorMessage.printf("Rank %d, Patch %d, x=%d, y=%d, f=%d: weight=%f, correct=%f, off by a factor of %f\n", rank, p, x, y, f, wgt, correct, wgt/correct);
status = PV_FAILURE;
}
}
}
}
}
if( status == PV_SUCCESS ) {
pvInfo().printf("Rank %d, connection \"%s\": Weights are correct.\n", rank, conn->getName());
}
return status;
}
//Makes a layer such that the restricted space is the index, but with spinning order be [x, y, f] as opposed to [f, x, y]
int InputLayer::updateState(double timef, double dt){
//Grab layer size
const PVLayerLoc* loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int nxGlobal = loc->nxGlobal;
int nyGlobal = loc->nyGlobal;
int kx0 = loc->kx0;
int ky0 = loc->ky0;
for(int b = 0; b < parent->getNBatch(); b++){
pvdata_t * A = getActivity() + b * getNumExtended();
//looping over ext
for(int iY = 0; iY < ny+loc->halo.up+loc->halo.dn; iY++){
for(int iX = 0; iX < nx+loc->halo.lt+loc->halo.rt; iX++){
//Calculate x and y global extended
int xGlobalExt = iX + loc->kx0;
int yGlobalExt = iY + loc->ky0;
//Calculate x and y in restricted space
int xGlobalRes = xGlobalExt - loc->halo.lt;
int yGlobalRes = yGlobalExt - loc->halo.up;
//Calculate base value
//xGlobal and yGlobalRes can be negative
int baseActivityVal = yGlobalRes * nxGlobal + xGlobalRes;
for(int iFeature = 0; iFeature < nf; iFeature++){
int ext_idx = kIndex(iX, iY, iFeature, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
//Feature gives an offset, since it spins slowest
int activityVal = baseActivityVal + iFeature * nxGlobal * nyGlobal;
A[ext_idx] = activityVal;
}
}
}
}
////Printing for double checking
//printf("\nOutMat\n");
////looping over ext
//for(int iFeature = 0; iFeature < nf; iFeature++){
// for(int iY = 0; iY < ny+loc->halo.up+loc->halo.dn; iY++){
// for(int iX = 0; iX < nx+loc->halo.lt+loc->halo.rt; iX++){
// int ext_idx = kIndex(iX, iY, iFeature, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
// printf("%03d ", (int)A[ext_idx]);
// }
// printf("\n");
// }
// printf("\n\n");
//}
return PV_SUCCESS;
}
int GatePoolTestLayer::updateState(double timef, double dt) {
//Do update state of ANN Layer first
ANNLayer::updateState(timef, dt);
//Grab layer size
const PVLayerLoc* loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nxGlobal = loc->nxGlobal;
int nyGlobal = loc->nyGlobal;
int nf = loc->nf;
int kx0 = loc->kx0;
int ky0 = loc->ky0;
bool isCorrect = true;
//Grab the activity layer of current layer
for(int b = 0; b < loc->nbatch; b++) {
const pvdata_t * A = getActivity() + b * getNumExtended();
//We only care about restricted space, but iY and iX are extended
for(int iY = loc->halo.up; iY < ny + loc->halo.up; iY++) {
for(int iX = loc->halo.lt; iX < nx + loc->halo.lt; iX++) {
for(int iFeature = 0; iFeature < nf; iFeature++) {
int ext_idx = kIndex(iX, iY, iFeature, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
float actualvalue = A[ext_idx];
int xval = (iX + kx0 - loc->halo.lt)/2;
int yval = (iY + ky0 - loc->halo.up)/2;
assert(xval >= 0 && xval < loc->nxGlobal);
assert(yval >= 0 && yval < loc->nxGlobal);
float expectedvalue;
expectedvalue = iFeature * 64 + yval * 16 + xval * 2 + 4.5;
expectedvalue*=4;
if(fabs(actualvalue - expectedvalue) >= 1e-4) {
pvErrorNoExit() << "Connection " << name << " Mismatch at (" << iX << "," << iY << ") : actual value: " << actualvalue << " Expected value: " << expectedvalue << ". Discrepancy is a whopping " << actualvalue - expectedvalue << "! Horrors!" << "\n";
isCorrect = false;
}
}
}
}
}
if(!isCorrect) {
InterColComm * icComm = parent->icCommunicator();
MPI_Barrier(icComm->communicator()); // If there is an error, make sure that MPI doesn't kill the run before process 0 reports the error.
exit(-1);
}
return PV_SUCCESS;
}
int CPTestInputLayer::initializeV() {
assert(parent->parameters()->value(name, "restart", 0.0f, false)==0.0f); // initializeV should only be called if restart is false
const PVLayerLoc * loc = getLayerLoc();
for (int b = 0; b < parent->getNBatch(); b++){
pvdata_t * VBatch = getV() + b * getNumNeurons();
for (int k = 0; k < getNumNeurons(); k++){
int kx = kxPos(k,loc->nx,loc->nx,loc->nf);
int ky = kyPos(k,loc->nx,loc->ny,loc->nf);
int kf = featureIndex(k,loc->nx,loc->ny,loc->nf);
int kGlobal = kIndex(loc->kx0+kx,loc->ky0+ky,kf,loc->nxGlobal,loc->nyGlobal,loc->nf);
VBatch[k] = (pvdata_t) kGlobal;
}
}
return PV_SUCCESS;
}
int BIDSCloneLayer::mapCoords(){
//Copy restricted clone data to current clayer data
for(int i = 0; i < numNodes; i++){
}
const PVLayerLoc origLoc = originalLayer->getCLayer()->loc;
for(int i = 0; i < numNodes; i++){
int index = kIndex(coords[i].xCoord, coords[i].yCoord, 0, clayer->loc.nx, clayer->loc.ny, clayer->loc.nf);
int destIndexEx = kIndexExtended(index, clayer->loc.nx, clayer->loc.ny, clayer->loc.nf, clayer->loc.halo.lt, clayer->loc.halo.rt, clayer->loc.halo.dn, clayer->loc.halo.up);
int srcIndexEx = kIndexExtended(index, origLoc.nx, origLoc.ny, origLoc.nf, origLoc.halo.lt, origLoc.halo.rt, origLoc.halo.dn, origLoc.halo.up);
this->clayer->activity->data[destIndexEx] = originalLayer->getCLayer()->activity->data[srcIndexEx] == 0 ? 0:1;
}
return PV_SUCCESS;
}
int InputLayer::updateState(double timef, double dt){
//Grab the activity layer of current layer
pvdata_t * A = getActivity();
//Grab layer size
const PVLayerLoc* loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int kx0 = loc->kx0;
int ky0 = loc->ky0;
assert(nf == 4);
assert(loc->nxGlobal == 2 && loc->nyGlobal == 2);
//We only care about restricted space
for(int iY = loc->halo.up; iY < ny + loc->halo.up; iY++){
for(int iX = loc->halo.lt; iX < nx + loc->halo.lt; iX++){
for(int iF = 0; iF < loc->nf; iF++){
int idx = kIndex(iX, iY, iF, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
int xval = iX+kx0-loc->halo.lt;
int yval = iY+ky0-loc->halo.up;
if(timef == 10 && xval == 0 && yval == 0 && iF == 0){
A[idx] = 1;
}
else if(timef == 10 && xval == 1 && yval == 0 && iF == 1){
A[idx] = 1;
}
else if(timef == 10 && xval == 0 && yval == 1 && iF == 2){
A[idx] = 1;
}
else if(timef == 10 && xval == 1 && yval == 1 && iF == 3){
A[idx] = 1;
}
else{
A[idx] = 0;
}
}
}
}
return PV_SUCCESS;
}
/**
* @time
* @l
* @k
* @kex
* NOTES:
* - Only the activity buffer covers the extended frame - this is the frame that
* includes boundaries.
* - The other dynamic variables (G_E, G_I, V, Vth) cover the "real" or "restricted"
* frame.
*/
int PointLIFProbe::writeState(double timed)
{
if (parent->columnId()==0 && timed >= writeTime) {
pvAssert(outputStream);
writeTime += writeStep;
PVLayerLoc const * loc = getTargetLayer()->getLayerLoc();
const int k = kIndex(xLoc, yLoc, fLoc, loc->nxGlobal, loc->nyGlobal, loc->nf);
double * valuesBuffer = getValuesBuffer();
outputStream->printf("%s t=%.1f %d"
"G_E=" CONDUCTANCE_PRINT_FORMAT
" G_I=" CONDUCTANCE_PRINT_FORMAT
" G_IB=" CONDUCTANCE_PRINT_FORMAT
" V=" CONDUCTANCE_PRINT_FORMAT
" Vth=" CONDUCTANCE_PRINT_FORMAT
" a=%.1f",
getMessage(), timed, k,
valuesBuffer[0], valuesBuffer[1], valuesBuffer[2],
valuesBuffer[3], valuesBuffer[4], valuesBuffer[5]);
output() << std::endl;
}
return PV_SUCCESS;
}
int PursuitLayer::updateState(double time, double dt) {
if (!updateReady) return PV_SUCCESS;
int nx = getLayerLoc()->nx;
int ny = getLayerLoc()->ny;
int nf = getLayerLoc()->nf;
PVHalo const * halo = &getLayerLoc()->halo;
pvdata_t * activity = getActivity();
memset(activity, 0, getNumExtended()*sizeof(*activity));
int nxy = nx*ny;
for (int kxy=0; kxy<nxy; kxy++) {
int kf = foundFeatures[kxy];
if (kf>=0) {
int kx = kxPos(kxy,nx,ny,1);
int ky = kyPos(kxy,nx,ny,1);
int kex = kIndex(kx+halo->lt, ky+halo->up, kf, nx+halo->lt+halo->rt, ny+halo->dn+halo->up, nf); /* Is this correct? Before splitting x- and y- margin widths, the ny argument was ny*nb, which seems weird. */
activity[kex] = gSynSparse[kxy];
}
}
//resetGSynBuffers_HyPerLayer(getNumNeurons(), getNumChannels(), GSyn[0]);
updateReady = false;
return PV_SUCCESS;
}
int OjaKernelSpikeRateProbe::allocateDataStructures() {
targetOjaKernelConn = dynamic_cast<OjaKernelConn *>(getTargetConn());
if (targetOjaKernelConn == NULL) {
if (getParent()->columnId()==0) {
fprintf(stderr, "LCATraceProbe error: connection \"%s\" must be an LCALIFLateralConn.\n", getTargetConn()->getName());
}
abort();
}
HyPerLayer * targetLayer = NULL;
if (isInputRate) {
targetLayer = targetOjaKernelConn->preSynapticLayer();
}
else {
targetLayer = targetOjaKernelConn->postSynapticLayer();
}
const PVLayerLoc * loc = targetLayer->getLayerLoc();
int x_local = xg - loc->kx0;
int y_local = yg - loc->ky0;
bool inBounds = (x_local >= 0 && x_local < loc->nx && y_local >= 0 && y_local < loc->ny);
if(inBounds ) { // if inBounds
int krestricted = kIndex(x_local, y_local, feature, loc->nx, loc->ny, loc->nf);
if (isInputRate) {
int kextended = kIndexExtended(krestricted, loc->nx, loc->ny, loc->nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
spikeRate = &targetOjaKernelConn->getInputFiringRate(arbor)[kextended];
}
else {
spikeRate = &targetOjaKernelConn->getOutputFiringRate()[krestricted];
}
}
else {
outputstream = NULL;
}
//This is now being done in BaseConnectionProbe
//getTargetConn()->insertProbe(this);
return PV_SUCCESS;
}
int MoviePvpTestLayer::updateStateWrapper(double time, double dt)
{
MoviePvp::updateStateWrapper(time, dt);
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int nbatch = loc->nbatch;
for(int b = 0; b < nbatch; b++){
pvdata_t * dataBatch = data + b * getNumExtended();
int frameIdx;
if(strcmp(getBatchMethod(), "byImage") == 0){
frameIdx = (time-1) * nbatch + b;
}
else if(strcmp(getBatchMethod(), "byMovie") == 0){
frameIdx = b * 2 + (time-1);
}
for(int nkRes = 0; nkRes < getNumNeurons(); nkRes++){
//Calculate extended index
int nkExt = kIndexExtended(nkRes, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
//checkVal is the value from batch index 0
pvdata_t checkVal = dataBatch[nkExt];
int kxGlobal = kxPos(nkRes, nx, ny, nf) + loc->kx0;
int kyGlobal = kyPos(nkRes, nx, ny, nf) + loc->ky0;
int kf = featureIndex(nkRes, nx, ny, nf);
pvdata_t expectedVal = kIndex(kxGlobal, kyGlobal, kf, loc->nxGlobal, loc->nyGlobal, nf) + frameIdx*192;
if(fabs(checkVal - expectedVal) >= 1e-5){
std::cout << "ImageFileIO " << name << " test Expected: " << expectedVal << " Actual: " << checkVal << "\n";
//exit(-1);
}
}
}
return PV_SUCCESS;
}
/**
* @timef
* NOTES:
* - kPost, kxPost, kyPost are indices in the restricted post-synaptic layer.
*
*/
int PostConnProbe::outputState(double timef)
{
int k, kxPre, kyPre;
HyPerConn * c = getTargetHyPerConn();
PVPatch * w;
PVPatch *** wPost = c->convertPreSynapticWeights(timef);
// TODO - WARNING: currently only works if nfPre==0
const PVLayer * lPre = c->preSynapticLayer()->clayer;
const PVLayer * lPost = c->postSynapticLayer()->clayer;
const int nxPre = lPre->loc.nx;
const int nyPre = lPre->loc.ny;
const int nfPre = lPre->loc.nf;
const PVHalo * haloPre = &lPre->loc.halo;
const int nxPost = lPost->loc.nx;
const int nyPost = lPost->loc.ny;
const int nfPost = lPost->loc.nf;
const PVHalo * haloPost = &lPost->loc.halo;
// calc kPost if needed
if (kPost < 0) {
kPost = kIndex(kxPost, kyPost, kfPost, nxPost, nyPost, nfPost);
}
else {
kxPost = kxPos(kPost, nxPost, nyPost, nfPost);
kyPost = kyPos(kPost, nxPost, nyPost, nfPost);
kfPost = featureIndex(kPost, nxPost, nyPost, nfPost);
}
c->preSynapticPatchHead(kxPost, kyPost, kfPost, &kxPre, &kyPre);
const int kxPreEx = kxPre + haloPre->lt;
const int kyPreEx = kyPre + haloPre->up;
const int kxPostEx = kxPost + haloPost->lt;
const int kyPostEx = kyPost + haloPost->up;
const int kPostEx = kIndex(kxPostEx, kyPostEx, kfPost, nxPost+haloPost->lt+haloPost->rt, nyPost+haloPost->dn+haloPost->up, nfPost);
const bool postFired = lPost->activity->data[kPostEx] > 0.0;
w = wPost[getArborID()][kPost];
pvwdata_t * wPostData = c->getWPostData(getArborID(),kPost);
const int nw = w->nx * w->ny * nfPost; //w->nf;
if (wPrev == NULL) {
wPrev = (pvwdata_t *) calloc(nw, sizeof(pvwdata_t));
for (k = 0; k < nw; k++) {
wPrev[k] = wPostData[k]; // This is broken if the patch is shrunken
}
}
if (wActiv == NULL) {
wActiv = (pvwdata_t *) calloc(nw, sizeof(pvwdata_t));
}
k = 0;
for (int ky = 0; ky < w->ny; ky++) {
for (int kx = 0; kx < w->nx; kx++) {
int kPre = kIndex(kx+kxPreEx, ky+kyPreEx, 0, nxPre+haloPre->lt+haloPre->rt, nyPre+haloPre->dn+haloPre->up, nfPre);
wActiv[k++] = lPre->activity->data[kPre];
}
}
bool changed = false;
for (k = 0; k < nw; k++) {
if (wPrev[k] != wPostData[k] || wActiv[k] != 0.0) {
changed = true;
break;
}
}
FILE * fp = getStream()->fp;
if (stdpVars && (postFired || changed)) {
if (postFired) fprintf(fp, "*");
else fprintf(fp, " ");
fprintf(fp, "t=%.1f w%d(%d,%d,%d) prePatchHead(%d,%d): ", timef, kPost, kxPost,
kyPost, kfPost, kxPre, kyPre);
if (image) fprintf(fp, "tag==%d ", image->tag());
fprintf(fp, "\n");
}
if (stdpVars && changed) {
text_write_patch_extra(fp, w, wPostData, wPrev, wActiv, getTargetHyPerConn());
fflush(fp);
}
for (k = 0; k < nw; k++) {
wPrev[k] = wPostData[k];
}
if (outputIndices) {
fprintf(fp, "w%d(%d,%d,%d) prePatchHead(%d,%d): ", kPost, kxPost, kyPost, kfPost, kxPre, kyPre);
if(!stdpVars){
fprintf(fp,"\n");
}
const PVLayer * lPre = c->preSynapticLayer()->clayer;
write_patch_indices(fp, w, &lPre->loc, kxPre, kyPre, 0);
fflush(fp);
}
return 0;
}
int BinningLayer::doUpdateState(double timed, double dt, const PVLayerLoc * origLoc, const PVLayerLoc * currLoc, const pvdata_t * origData, pvdata_t * currA, float binMax, float binMin) {
int status = PV_SUCCESS;
//update_timer->start();
int numBins = currLoc->nf;
int nx = currLoc->nx;
int ny = currLoc->ny;
//Check that both nb are the same
assert(origLoc->halo.lt == currLoc->halo.lt &&
origLoc->halo.rt == currLoc->halo.rt &&
origLoc->halo.dn == currLoc->halo.dn &&
origLoc->halo.up == currLoc->halo.up);
assert(origLoc->nf == 1);
PVHalo const * halo = &origLoc->halo;
float binRange = binMax - binMin;
float stepSize = float(binRange)/numBins;
int nbatch = currLoc->nbatch;
for(int b = 0; b < nbatch; b++){
const pvdata_t * origDataBatch = origData + b * (origLoc->nx + origLoc->halo.lt + origLoc->halo.rt) * (origLoc->ny + origLoc->halo.up + origLoc->halo.dn) * origLoc->nf;
pvdata_t * currABatch = currA + b * (currLoc->nx + currLoc->halo.lt + currLoc->halo.rt) * (currLoc->ny + currLoc->halo.up + currLoc->halo.dn) * currLoc->nf;
// each y value specifies a different target so ok to thread here (sum, sumsq are defined inside loop)
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for
#endif
for (int iY = 0; iY < (ny+halo->dn+halo->up); iY++){
for (int iX = 0; iX < (nx+halo->lt+halo->rt); iX++){
int origIdx = kIndex(iX, iY, 0, nx+halo->lt+halo->rt, ny+halo->dn+halo->up, origLoc->nf);
float inVal = origDataBatch[origIdx];
//If inVal is out of bounds in either binMax or binMin, set the value to be the maximum or minimum val
if(inVal < binMin){
inVal = binMin;
}
if(inVal > binMax){
inVal = binMax;
}
if(zeroDCR && inVal == 0){
for(int iF = 0; iF < numBins; iF++){
int currIdx = kIndex(iX, iY, iF, nx+halo->lt+halo->rt, ny+halo->dn+halo->up, numBins);
currABatch[currIdx] = 0;
}
}
else{
//A sigma of zero means only the centered bin value should get input
int featureIdx = round((inVal-binMin)/stepSize);
for(int iF = 0; iF < numBins; iF++){
if(binSigma == 0){
int currIdx = kIndex(iX, iY, iF, nx+halo->lt+halo->rt, ny+halo->dn+halo->up, numBins);
if(iF == featureIdx){
currABatch[currIdx] = 1;
}
//Resetting value
else{
if(zeroNeg){
currABatch[currIdx] = 0;
}
else{
currABatch[currIdx] = -1;
}
}
}
else{
//Calculate center value for featureIdx (the bin that the value belongs to without a sigma) is binning
float mean;
if(normalDist){
mean = featureIdx * stepSize + (stepSize/2);
}
else{
mean = featureIdx;
}
//Possible bins
int intSigma = ceil(binSigma);
int currIdx = kIndex(iX, iY, iF, nx+halo->lt+halo->rt, ny+halo->dn+halo->up, numBins);
if(iF >= featureIdx-intSigma && iF <= featureIdx+intSigma){
//Get center of that aBin for the x pos of the normal dist
float xVal;
if(normalDist){
xVal = iF * stepSize + (stepSize/2);
}
else{
xVal = iF;
}
//Calculate normal dist
float outVal = calcNormDist(xVal, mean, binSigma);
//Put into activity buffer
currABatch[currIdx] = outVal;
}
//Resetting value
else{
if(zeroNeg){
currABatch[currIdx] = 0;
}
else{
currABatch[currIdx] = -1;
}
}
}
}
}
}
}
}
//update_timer->stop();
return status;
}
int PursuitLayer::recvSynapticInput(HyPerConn * conn, const PVLayerCube * activity, int arborID) {
if (parent->simulationTime()<nextUpdate) return PV_SUCCESS;
nextUpdate += updatePeriod;
recvsyn_timer->start();
assert(arborID >= 0);
if (conn->usingSharedWeights() == false) {
fprintf(stderr, "Error: PursuitLayer can only be the postsynaptic layer of a connection using shared weights (this condition should be removed eventually).\n");
abort();
}
HyPerLayer * pre = conn->preSynapticLayer();
const PVLayerLoc * pre_loc = pre->getLayerLoc();
if (pre_loc->nx != getLayerLoc()->nx || pre_loc->ny != getLayerLoc()->ny) {
fprintf(stderr, "Error: PursuitLayer requires incoming connections to be one-to-one.\n");
abort();
}
#ifdef DEBUG_OUTPUT
int rank;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
//printf("[%d]: HyPerLayr::recvSyn: neighbor=%d num=%d actv=%p this=%p conn=%p\n", rank, neighbor, numExtended, activity, this, conn);
printf("[%d]: HyPerLayr::recvSyn: neighbor=%d num=%d actv=%p this=%p conn=%p\n", rank, 0, numExtended, activity, this, conn);
fflush(stdout);
#endif // DEBUG_OUTPUT
const int numExtended = activity->numItems;
for (int kPre = 0; kPre < numExtended; kPre++) {
float a = activity->data[kPre];
if (a == 0.0f) continue;
PVPatch * weights = conn->getWeights(kPre, arborID);
// WARNING - assumes every value in weights maps into a valid value in GSyn
// - assumes patch stride sf is 1
int nk = conn->fPatchSize() * weights->nx;
int ny = weights->ny;
int sy = conn->getPostNonextStrides()->sy; // stride in layer
int syw = conn->yPatchStride(); // stride in patch
pvdata_t * gSynPatchHead = this->getChannel(conn->getChannel());
pvdata_t * gSynPatchStart = gSynPatchHead + conn->getGSynPatchStart(kPre, arborID);
pvwdata_t * data = conn->get_wData(arborID,kPre);
for (int y = 0; y < ny; y++) {
(conn->accumulateFunctionPointer)(nk, gSynPatchStart + y*sy, a, data + y*syw, NULL);
}
}
// Set |w(:,:,f)|^2. Since this is a one-to-one connection with one presynaptic feature,
// only have to do once for each feature of a single (x,y) site and then copy.
int nxp = conn->xPatchSize();
int nyp = conn->yPatchSize();
int nfp = conn->fPatchSize();
int num_weights = nxp*nyp*nfp;
assert(zUnitCellSize(pre->getXScale(), getXScale())==1);
assert(zUnitCellSize(pre->getYScale(), getYScale())==1);
assert(conn->getNumDataPatches()==1);
for (int kf=0; kf<nfp; kf++) {
pvwdata_t * weight = conn->get_wDataHead(arborID, 0);
pvdata_t sum = 0.0;
for (int k=0; k<num_weights; k+=nfp) {
pvwdata_t w = weight[k + kf]; // Assumes stride in features is 1.
//TODO-CER-2014.4.4 - convert weights
sum += w*w;
}
wnormsq[kf] = sum;
}
pvdata_t * gSynStart = GSyn[conn->getChannel()];
int nx = getLayerLoc()->nx;
int ny = getLayerLoc()->ny;
int nxy = nx*ny;
// TODO: Can I compute energyDropsBestFeature and minLocationsBestFeature without storing all the energyDrops and minimumLocations?
for (int kxy=0; kxy<nxy; kxy++) {
for (int kf=0; kf<nfp; kf++) {
int k=kxy*nfp+kf; // Assumes stride in features is 1.
minimumLocations[k] = gSynStart[k]/wnormsq[kf];
energyDrops[k] = -gSynStart[k]*minimumLocations[k]/2;
}
}
for (int kxy=0; kxy<nxy; kxy++) {
minFeatures[kxy] = -1;
energyDropsBestFeature[kxy] = FLT_MAX;
int index0 = kxy*nfp; // assumes stride in features is 1.
if (foundFeatures[kxy]>=0) {
energyDropsBestFeature[kxy] = energyDrops[kxy*nfp+foundFeatures[kxy]];
minFeatures[kxy] = foundFeatures[kxy];
}
else {
for (int kf=0; kf<nfp; kf++) {
if (energyDrops[index0+kf] < energyDropsBestFeature[kxy]) {
minFeatures[kxy] = kf;
energyDropsBestFeature[kxy] = energyDrops[index0+kf];
}
}
}
}
for (int kxy=0; kxy<nxy; kxy++) {
assert(minFeatures[kxy]>=0 && minFeatures[kxy]<nfp);
int baseindex = kxy*nfp;
minLocationsBestFeature[kxy] = minimumLocations[baseindex+minFeatures[kxy]];
}
bool mask[nxy];
memset(mask, false, nxy*sizeof(*mask));
pvdata_t smallestEnergyDrop;
int minloc;
while (constrainMinima(), minloc = filterMinEnergies(mask, &smallestEnergyDrop), smallestEnergyDrop<FLT_MAX) {
assert(foundFeatures[minloc]<0 || foundFeatures[minloc]==minFeatures[minloc]);
foundFeatures[minloc] = minFeatures[minloc];
gSynSparse[minloc] += minLocationsBestFeature[minloc];
if (gSynSparse[minloc] < 1e-4) {
gSynSparse[minloc]=0;
foundFeatures[minloc] = -1;
}
int minlocx = kxPos(minloc,nx,ny,1);
int maskstartx = minlocx-(nxp-1);
if (maskstartx<0) maskstartx=0;
int maskstopx = minlocx+nxp;
if (maskstopx>nx) maskstopx=nx;
int minlocy = kyPos(minloc,nx,ny,1);
int maskstarty = minlocy-(nyp-1);
if (maskstarty<0) maskstarty=0;
int maskstopy = minlocy+nyp;
if (maskstopy>ny) maskstopy=ny;
for (int ky=maskstarty; ky<maskstopy; ky++) {
for (int kx=maskstartx; kx<maskstopx; kx++) {
mask[kIndex(kx,ky,0,nx,ny,1)]=true;
}
}
}
recvsyn_timer->stop();
updateReady = true;
return 0;
}
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 PointLIFProbe::calcValues(double timevalue) {
// TODO: Reduce duplicated code between PointProbe::calcValues and PointLIFProbe::calcValues.
assert(this->getNumValues()==NUMBER_OF_VALUES);
LIF * LIF_layer = dynamic_cast<LIF *>(getTargetLayer());
assert(LIF_layer != NULL);
pvconductance_t const * G_E = LIF_layer->getConductance(CHANNEL_EXC) + batchLoc * LIF_layer->getNumNeurons();
pvconductance_t const * G_I = LIF_layer->getConductance(CHANNEL_INH) + batchLoc * LIF_layer->getNumNeurons();
pvconductance_t const * G_IB = LIF_layer->getConductance(CHANNEL_INHB) + batchLoc * LIF_layer->getNumNeurons();
pvdata_t const * V = getTargetLayer()->getV();
pvdata_t const * Vth = LIF_layer->getVth();
pvdata_t const * activity = getTargetLayer()->getLayerData();
assert(V && activity && G_E && G_I && G_IB && Vth);
double * valuesBuffer = this->getValuesBuffer();
//We need to calculate which mpi process contains the target point, and send that info to the root process
//Each process calculates local index
const PVLayerLoc * loc = getTargetLayer()->getLayerLoc();
//Calculate local cords from global
const int kx0 = loc->kx0;
const int ky0 = loc->ky0;
const int kb0 = loc->kb0;
const int nx = loc->nx;
const int ny = loc->ny;
const int nf = loc->nf;
const int nbatch = loc->nbatch;
const int xLocLocal = xLoc - kx0;
const int yLocLocal = yLoc - ky0;
const int nbatchLocal = batchLoc - kb0;
//if in bounds
if( xLocLocal >= 0 && xLocLocal < nx &&
yLocLocal >= 0 && yLocLocal < ny &&
nbatchLocal >= 0 && nbatchLocal < nbatch){
const pvdata_t * V = getTargetLayer()->getV();
const pvdata_t * activity = getTargetLayer()->getLayerData();
//Send V and A to root
const int k = kIndex(xLocLocal, yLocLocal, fLoc, nx, ny, nf);
const int kbatch = k + nbatchLocal*getTargetLayer()->getNumNeurons();
valuesBuffer[0] = G_E[kbatch];
valuesBuffer[1] = G_I[kbatch];
valuesBuffer[2] = G_IB[kbatch];
valuesBuffer[3] = V[kbatch];
valuesBuffer[4] = Vth[kbatch];
const int kex = kIndexExtended(k, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
valuesBuffer[5] = activity[kex + nbatchLocal * getTargetLayer()->getNumExtended()];
//If not in root process, send to root process
if(parent->columnId()!=0){
MPI_Send(valuesBuffer, NUMBER_OF_VALUES, MPI_DOUBLE, 0, 0, parent->icCommunicator()->communicator());
}
}
//Root process
if(parent->columnId()==0){
//Calculate which rank target neuron is
//TODO we need to calculate rank from batch as well
int xRank = xLoc/nx;
int yRank = yLoc/ny;
int srcRank = rankFromRowAndColumn(yRank, xRank, parent->icCommunicator()->numCommRows(), parent->icCommunicator()->numCommColumns());
//If srcRank is not root process, MPI_Recv from that rank
if(srcRank != 0){
MPI_Recv(valuesBuffer, NUMBER_OF_VALUES, MPI_DOUBLE, srcRank, 0, parent->icCommunicator()->communicator(), MPI_STATUS_IGNORE);
}
}
return PV_SUCCESS;
}
int BackwardsBatchNorm::updateState(double timef, double dt) {
int status = PV_SUCCESS;
//We are filling this activity buffer
pvdata_t * thisA = clayer->activity->data;
assert(thisA);
//We need the normalized input vals, orig input vals, and the input gradients
const pvdata_t * inputGradA = originalLayer->getCLayer()->activity->data;
const pvdata_t * forwardA = forwardLayer->getCLayer()->activity->data;
const pvdata_t * origInputA = forwardLayer->getOriginalLayer()->getCLayer()->activity->data;
assert(inputGradA && forwardA && origInputA);
//Get locs for all buffers
const PVLayerLoc * thisLoc = getLayerLoc();
const PVLayerLoc * inputGradLoc = originalLayer->getLayerLoc();
const PVLayerLoc * forwardLoc = forwardLayer->getLayerLoc();
const PVLayerLoc * origInputLoc = forwardLayer->getOriginalLayer()->getLayerLoc();
int nbatch = thisLoc->nbatch;
//All nx, ny, and nf should be the same
int nx = thisLoc->nx;
int ny = thisLoc->ny;
int nf = thisLoc->nf;
//Get buffer margins here
int xThisMargin = thisLoc->halo.lt + thisLoc->halo.rt;
int yThisMargin = thisLoc->halo.up + thisLoc->halo.dn;
int xInputGradMargin = inputGradLoc->halo.lt + inputGradLoc->halo.rt;
int yInputGradMargin = inputGradLoc->halo.up + inputGradLoc->halo.dn;
int xForwardMargin = forwardLoc->halo.lt + forwardLoc->halo.rt;
int yForwardMargin = forwardLoc->halo.up + forwardLoc->halo.dn;
int xOrigInputMargin = origInputLoc->halo.lt + origInputLoc->halo.rt;
int yOrigInputMargin = origInputLoc->halo.up + origInputLoc->halo.dn;
//We also need various mean and var buffers from the forward layer
const float* batchMean = forwardLayer->getBatchMean();
const float* batchVar = forwardLayer->getBatchVar();
float* batchMeanShift = forwardLayer->getBatchMeanShift();
float* batchVarShift = forwardLayer->getBatchVarShift();
float epsilon = forwardLayer->getEpsilon();
//Total number of neurons to divide by for each feature
float normVal = parent->getNBatchGlobal() * thisLoc->nyGlobal * thisLoc->nxGlobal;
//We're accumulating into delta buffers, so clear
clearDelta();
//Ioffe et. al. Batch Normalization
//Calculate deltaVar
//TODO parallize over threads
for(int iF = 0; iF < nf; iF++) {
float secondTerm = -.5*(powf(batchVar[iF] + epsilon, -1.5));
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchOrigInputA = origInputA + b * forwardLayer->getOriginalLayer()->getNumExtended();
const pvdata_t* batchInputGradA = inputGradA + b * originalLayer->getNumExtended();
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
int kExtOrigInput = kIndex(iX, iY, iF, nx+xOrigInputMargin, ny+yOrigInputMargin, nf);
int kExtInputGrad = kIndex(iX, iY, iF, nx+xInputGradMargin, ny+yInputGradMargin, nf);
float deltaNorm = batchInputGradA[kExtInputGrad] * batchVarShift[iF];
deltaVar[iF] += deltaNorm * (batchOrigInputA[kExtOrigInput] - batchMean[iF]);
}
}
}
//Multiply deltaVar by secondTerm
deltaVar[iF] = deltaVar[iF] * secondTerm;
}
//Reduce deltaVar
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, deltaVar, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
#endif // PV_USE_MPI
//Calculate deltaMean
//Calculate first term first
//TODO parallize over threads
for(int iF = 0; iF < nf; iF++) {
float multiplier = -1.0/(sqrtf(batchVar[iF]+epsilon));
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchInputGradA = inputGradA + b * originalLayer->getNumExtended();
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
int kExtInputGrad = kIndex(iX, iY, iF, nx+xInputGradMargin, ny+yInputGradMargin, nf);
float deltaNorm = batchInputGradA[kExtInputGrad] * batchVarShift[iF];
deltaMean[iF] += deltaNorm * multiplier;
}
}
}
}
//Reduce deltaMean across mpi
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, deltaMean, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
#endif // PV_USE_MPI
//Calculate second term
//TODO parallize over threads
for(int iF = 0; iF < nf; iF++) {
float tmpMean = 0;
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchOrigInputA = origInputA + b * forwardLayer->getOriginalLayer()->getNumExtended();
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
int kExtOrigInput = kIndex(iX, iY, iF, nx+xOrigInputMargin, ny+yOrigInputMargin, nf);
tmpMean += -2 * (batchOrigInputA[kExtOrigInput] - batchMean[iF]);
}
}
}
//Reduce tmpMean
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, &tmpMean, 1, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
#endif // PV_USE_MPI
tmpMean = tmpMean / normVal;
//Add second term to first term
deltaMean[iF] += deltaVar[iF] * tmpMean;
}
//No more sums, go with efficient loop
//TODO Is the efficient loop better for optimization or do we put
//features on the outer most loop for precalculation of constants over features?
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchOrigInputA = origInputA + b * forwardLayer->getOriginalLayer()->getNumExtended();
const pvdata_t* batchInputGradA = inputGradA + b * originalLayer->getNumExtended();
pvdata_t* batchThisA = thisA + b * getNumExtended();
#ifdef PV_USE_OPENMP_THREADS
#pragma omp parallel for collapse(3)
#endif
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
for(int iF = 0; iF < nf; iF++) {
int kExtOrigInput = kIndex(iX, iY, iF, nx+xOrigInputMargin, ny+yOrigInputMargin, nf);
int kExtInputGrad = kIndex(iX, iY, iF, nx+xInputGradMargin, ny+yInputGradMargin, nf);
int kExtThis = kIndex(iX, iY, iF, nx+xThisMargin, ny+yThisMargin, nf);
float deltaNorm = batchInputGradA[kExtInputGrad] * batchVarShift[iF];
float firstTerm = deltaNorm/sqrtf(batchVar[iF] + epsilon);
float secondTerm = deltaVar[iF] * (2*(batchOrigInputA[kExtOrigInput] - batchMean[iF])/normVal);
float thirdTerm = deltaMean[iF]/normVal;
batchThisA[kExtThis] = firstTerm + secondTerm + thirdTerm;
}
}
}
}
//We calculate delta varShift and deltaMeanShift here
//TODO parallize over threads
//Since we're summing into delta*shift buffers, we have to sequentialize over features
for(int iF = 0; iF < nf; iF++) {
for(int b = 0; b < nbatch; b++) {
const pvdata_t* batchForwardA = forwardA + b * forwardLayer->getNumExtended();
const pvdata_t* batchInputGradA = inputGradA + b * originalLayer->getNumExtended();
for(int iY = 0; iY < ny; iY++) {
for(int iX = 0; iX < nx; iX++) {
int kExtInputGrad = kIndex(iX, iY, iF, nx+xInputGradMargin, ny+yInputGradMargin, nf);
int kExtForwardA = kIndex(iX, iY, iF, nx+xForwardMargin, ny + yForwardMargin, nf);
deltaVarShift[iF] += batchInputGradA[kExtInputGrad] * batchForwardA[kExtForwardA];
deltaMeanShift[iF] += batchInputGradA[kExtInputGrad];
}
}
}
}
//Reduce delta*Shift across all mpi
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, deltaVarShift, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
MPI_Allreduce(MPI_IN_PLACE, deltaMeanShift, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->globalCommunicator());
#endif // PV_USE_MPI
//TODO implement learning rule for meanShift and varShift
return status;
}
void MLPOutputLayer::multiclassNonlocalStats(){
const PVLayerLoc * loc = getLayerLoc();
int nx = loc->nx;
int ny = loc->ny;
int nf = loc->nf;
int numNeurons = getNumNeurons();
pvdata_t * A = getCLayer()->activity->data;
pvdata_t * gtA = gtLayer->getCLayer()->activity->data;
float sumsq = 0;
//Winner take all in the output layer
int currNumRight = 0;
int currNumWrong = 0;
assert(classBuffer);
//Clear classBuffer
for(int i = 0; i < nf; i++){
classBuffer[i] = 0;
}
//Only go through restricted
//Calculate the sum squared error
for(int ni = 0; ni < numNeurons; ni++){
int nExt = kIndexExtended(ni, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
int fi = featureIndex(nExt, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
//Sum over x and y direction
classBuffer[fi] += A[nExt];
sumsq += pow(A[nExt] - gtA[nExt], 2);
}
//Normalize classBuffer to find mean
for(int i = 0; i < nf; i++){
classBuffer[i] /= nx*ny;
}
//Reduce all classBuffers through a mean
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, &sumsq, 1, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, classBuffer, nf, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->communicator());
//Normalize classBuffer across processors
for(int i = 0; i < nf; i++){
classBuffer[i] /= parent->icCommunicator()->commSize();
}
#endif // PV_USE_MPI
//Find max
float estMaxF = -1000;
int estMaxFi = -1;
float actualMaxF = -1000;
int actualMaxFi = -1;
for(int i = 0; i < nf; i++){
if(classBuffer[i] >= estMaxF){
estMaxF = classBuffer[i];
estMaxFi = i;
}
int nExt = kIndex(loc->halo.lt, loc->halo.up, i, nx+loc->halo.lt+loc->halo.rt, ny+loc->halo.dn+loc->halo.up, nf);
if(gtA[nExt] >= actualMaxF){
actualMaxF = gtA[nExt];
actualMaxFi = i;
}
}
//Calculate stats
//Found winning feature, compare to ground truth
if(estMaxFi == actualMaxFi){
currNumRight++;
}
else{
currNumWrong++;
}
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, &currNumRight, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
MPI_Allreduce(MPI_IN_PLACE, &currNumWrong, 1, MPI_INT, MPI_SUM, parent->icCommunicator()->communicator());
#endif // PV_USE_MPI
numRight += currNumRight;
numWrong += currNumWrong;
progressNumRight += currNumRight;
progressNumWrong += currNumWrong;
//Print if need
float timef = parent->simulationTime();
if(timef >= nextStatProgress){
//Update nextStatProgress
nextStatProgress += statProgressPeriod;
if (parent->columnId()==0) {
float totalScore = 100*float(numRight)/float(numRight+numWrong);
float progressScore = 100*float(progressNumRight)/float(progressNumRight+progressNumWrong);
fprintf(stdout, "time:%f layer:\"%s\" total:%f%% progressStep:%f%% energy:%f\n", timef, name, totalScore, progressScore, sumsq/2);
}
//Reset progressStats
progressNumRight = 0;
progressNumWrong = 0;
}
}
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;
}
int LCALIFLateralKernelConn::allocateDataStructures() {
int status = HyPerConn::allocateDataStructures();
// Neurons don't inhibit themselves, only their neighbors; set self-interaction weights to mmzero.
assert(nxp % 2 == 1 && nyp % 2 == 1 && getNumDataPatches()==nfp);
for (int k=0; k<getNumDataPatches(); k++) {
int n = kIndex((nxp-1)/2, (nyp-1)/2, k, nxp, nyp, nfp);
get_wDataHead(0, k)[n] = 0.0f;
}
integratedSpikeCountCube = pvcube_new(pre->getLayerLoc(), pre->getNumExtended());
integratedSpikeCount = integratedSpikeCountCube->data;
for (int k=0; k<pre->getNumExtended(); k++) {
integratedSpikeCount[k] = integrationTimeConstant*getTargetRateKHz(); // Spike counts initialized to equilibrium value
}
mpi_datatype = Communicator::newDatatypes(pre->getLayerLoc());
if (mpi_datatype==NULL) {
fprintf(stderr, "LCALIFLateralKernelConn \"%s\" error creating mpi_datatype\n", name);
abort();
}
// Compute the number of times each patch contributes to dw, for proper averaging.
int num_arbors = numberOfAxonalArborLists();
interiorCounts = (float **) calloc(num_arbors, sizeof(float *));
if (interiorCounts==NULL) {
fprintf(stderr, "LCALIFLateralKernelConn::initialize \"%s\" error: unable to allocate memory for interiorCounts pointer\n", name);
}
interiorCounts[0] = (float *) calloc(getNumDataPatches()*nxp*nyp*nfp, sizeof(float));
if (interiorCounts[0]==NULL) {
fprintf(stderr, "LCALIFLateralKernelConn::initialize \"%s\" error: unable to allocate memory for interiorCounts\n", name);
}
for (int arbor=1; arbor<num_arbors; arbor++) {
interiorCounts[arbor] = interiorCounts[0]+arbor*getNumDataPatches()*nxp*nyp*nfp;
}
const PVLayerLoc * preloc = pre->getLayerLoc();
int nxpre = preloc->nx;
int nypre = preloc->ny;
int nfpre = preloc->nf;
int nExt = pre->getNumExtended();
int sya = getPostExtStrides()->sy;
int nxglob = preloc->nxGlobal;
int nyglob = preloc->nyGlobal;
int kx0 = preloc->kx0;
int ky0 = preloc->ky0;
for (int arbor=0; arbor<numberOfAxonalArborLists(); arbor++) {
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, nypre + preloc->halo.lt + preloc->halo.rt, nypre + preloc->halo.dn + preloc->halo.up, nfpre) + ky0 - preloc->halo.up;
if (xglob < 0 || xglob >= nxglob || yglob < 0 || yglob >= nyglob) {
continue;
}
PVPatch * weights = getWeights(kExt,arbor);
int offset = (int) getAPostOffset(kExt, arbor);
int ny = weights->ny;
int nk = weights->nx * nfp;
int interiorCountOffset = get_wData(arbor, kExt)-get_wDataStart(arbor);
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
interiorCounts[arbor][interiorCountOffset + lineoffsetw + k]++;
}
}
lineoffsetw += syp;
lineoffseta += sya;
}
}
}
int bufsize = numberOfAxonalArborLists() * getNumDataPatches() * nxp * nyp * nfp;
// TODO-CER-2014.3.26 - Ensure that reduction is done when not using MPI
#ifdef PV_USE_MPI
MPI_Allreduce(MPI_IN_PLACE, interiorCounts[0], bufsize, MPI_FLOAT, MPI_SUM, parent->icCommunicator()->communicator());
#endif
return status;
}
int PointProbe::calcValues(double timevalue) {
assert(this->getNumValues()==2);
double * valuesBuffer = this->getValuesBuffer();
//We need to calculate which mpi process contains the target point, and send that info to the root process
//Each process calculates local index
const PVLayerLoc * loc = getTargetLayer()->getLayerLoc();
//Calculate local cords from global
const int kx0 = loc->kx0;
const int ky0 = loc->ky0;
const int kb0 = loc->kb0;
const int nx = loc->nx;
const int ny = loc->ny;
const int nf = loc->nf;
const int nbatch = loc->nbatch;
const int xLocLocal = xLoc - kx0;
const int yLocLocal = yLoc - ky0;
const int nbatchLocal = batchLoc - kb0;
//if in bounds
if( xLocLocal >= 0 && xLocLocal < nx &&
yLocLocal >= 0 && yLocLocal < ny &&
nbatchLocal >= 0 && nbatchLocal < nbatch){
const pvdata_t * V = getTargetLayer()->getV();
const pvdata_t * activity = getTargetLayer()->getLayerData();
//Send V and A to root
const int k = kIndex(xLocLocal, yLocLocal, fLoc, nx, ny, nf);
if(V){
valuesBuffer[0] = V[k + nbatchLocal*getTargetLayer()->getNumNeurons()];
}
else {
valuesBuffer[0] = 0.0;
}
if(activity){
const int kex = kIndexExtended(k, nx, ny, nf, loc->halo.lt, loc->halo.rt, loc->halo.dn, loc->halo.up);
valuesBuffer[1] = activity[kex + nbatchLocal * getTargetLayer()->getNumExtended()];
}
else {
valuesBuffer[1] = 0.0;
}
//If not in root process, send to root process
if(parent->columnId()!=0){
MPI_Send(&valuesBuffer, 2, MPI_DOUBLE, 0, 0, parent->icCommunicator()->communicator());
}
}
//Root process
if(parent->columnId()==0){
//Calculate which rank target neuron is
//TODO we need to calculate rank from batch as well
int xRank = xLoc/nx;
int yRank = yLoc/ny;
int srcRank = rankFromRowAndColumn(yRank, xRank, parent->icCommunicator()->numCommRows(), parent->icCommunicator()->numCommColumns());
//If srcRank is not root process, MPI_Recv from that rank
if(srcRank != 0){
MPI_Recv(&valuesBuffer, 2, MPI_DOUBLE, srcRank, 0, parent->icCommunicator()->communicator(), MPI_STATUS_IGNORE);
}
}
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;
}
