void Householder<T>::evalHHmatrixData(const NDArray<T>& x, NDArray<T>& tail, T& coeff, T& normX) {
// input validation
if(!x.isVector() && !x.isScalar())
throw "ops::helpers::Householder::evalHHmatrixData method: input array must be vector or scalar!";
if(!x.isScalar() && x.lengthOf() != tail.lengthOf() + 1)
throw "ops::helpers::Householder::evalHHmatrixData method: input tail vector must have length less than unity compared to input x vector!";
normX = x.template reduceNumber<simdOps::Norm2<T>>();
const T min = DataTypeUtils::min<T>();
if(normX*normX - x(0)*x(0) <= min) {
normX = x(0);
coeff = (T)0.;
tail = (T)0.;
}
else {
if(x(0) >= (T)0.)
normX = -normX; // choose opposite sign to lessen roundoff error
T u0 = x(0) - normX;
coeff = -u0 / normX;
if(x.isRowVector())
tail.assign(x({{}, {1, -1}}) / u0);
else
tail.assign(x({{1, -1}, {}}) / u0);
}
}
NDArray<T> Householder<T>::evalHHmatrix(const NDArray<T>& x) {
// input validation
if(!x.isVector() && !x.isScalar())
throw "ops::helpers::Householder::evalHHmatrix method: input array must be vector or scalar!";
NDArray<T> w((int)x.lengthOf(), 1, x.ordering(), x.getWorkspace()); // column-vector
NDArray<T> wT(1, (int)x.lengthOf(), x.ordering(), x.getWorkspace()); // row-vector (transposed w)
T coeff;
T normX = x.template reduceNumber<simdOps::Norm2<T>>();
const T min = DataTypeUtils::min<T>();
if(normX*normX - x(0)*x(0) <= min) {
normX = x(0);
coeff = (T)0.;
w = (T)0.;
}
else {
if(x(0) >= (T)0.)
normX = -normX; // choose opposite sign to lessen roundoff error
T u0 = x(0) - normX;
coeff = -u0 / normX;
w.assign(x / u0);
}
w(0) = (T)1.;
wT.assign(&w);
NDArray<T> identity((int)x.lengthOf(), (int)x.lengthOf(), x.ordering(), x.getWorkspace());
identity.setIdentity(); // identity matrix
return identity - mmul(w, wT) * coeff;
}
void Householder<T>::evalHHmatrixDataI(const NDArray<T>& x, T& coeff, T& normX) {
int rows = (int)x.lengthOf()-1;
int num = 1;
if(rows == 0) {
rows = 1;
num = 0;
}
NDArray<T> tail(rows, 1, x.ordering(), x.getWorkspace());
evalHHmatrixData(x, tail, coeff, normX);
if(x.isRowVector()) {
NDArray<T>* temp = x.subarray({{}, {num, x.sizeAt(1)}});
temp->assign(tail);
delete temp;
}
else {
NDArray<T>* temp = x.subarray({{num, x.sizeAt(0)}, {}});
temp->assign(tail);
delete temp;
}
}
void skipgramBatchExec_(NDArray &s0, NDArray &s1, NDArray &s1n, void *vexpTable, void *vnegTable, void *vinfVector, NDArray &targets, NDArray &negStarters, NDArray &indices, NDArray &codes, NDArray &lr, NDArray &nextRandom, const int nsRounds, const int vocabSize, const int vectorLength, const int expLength, const int negLength, const bool preciseMode, const int numThreads) {
//auto syn0 = reinterpret_cast<T*>(vsyn0);
//auto syn1 = reinterpret_cast<T*>(vsyn1);
//auto syn1Neg = reinterpret_cast<T*>(vsyn1Neg);
const auto expTable = reinterpret_cast<T*>(vexpTable);
const auto negTable = reinterpret_cast<T*>(vnegTable);
const auto infVector = reinterpret_cast<T*>(vinfVector);
T sneu1e[600];
//const auto numThreads = omp_get_max_threads();
const auto idxShift = indices.isEmpty() ? 0 : indices.sizeAt(1);
const auto hsRounds = codes.isEmpty() ? 0 : codes.sizeAt(1);
// regular mode provides 0 guarantees for reproducibility
auto numTargets = targets.lengthOf();
auto bTarget = targets.bufferAsT<int>();
auto bIndices = indices.bufferAsT<int>();
auto bCodes = codes.bufferAsT<int8_t>();
#pragma omp parallel for num_threads(numThreads) private(sneu1e) default(shared) schedule(static)
for (int t = 0; t < numTargets; t++) {
T* neu1e = vectorLength <= 600 ? sneu1e : new T[vectorLength];
memset(neu1e, 0, vectorLength * sizeof(T));
auto target = bTarget[t];
auto alpha = lr.e<double>(t);
unsigned long long randomValue = nextRandom.e<Nd4jLong>(t);
auto syn0row = reinterpret_cast<T*>(s0.bufferWithOffset(target * vectorLength));
if (hsRounds > 0) {
int irow = 0;
auto cShift = t * idxShift;
for (int e = 0; e < hsRounds; e++) {
irow = bIndices[e + cShift];
if (irow < 0 || irow >= vocabSize)
continue;
auto syn1row = s1.bufferWithOffset(irow * vectorLength);
auto code = bCodes[e + cShift];
//nd4j_printf("syn0: [%i]; syn1: [%i]; code: [%i]\n", target, irow, code);
hSoftmax_<T>(syn0row, syn1row, expTable, neu1e, alpha, vectorLength, code, expLength, false);
}
}
if (nsRounds > 0) {
int irow = negStarters.e<int>(t);
int nsStarter = irow;
for (int r = 0; r < nsRounds + 1; r++) {
if (r == 0) {
// target is known in advance
} else {
randomValue = randomValue * (unsigned long long) 25214903917 + 11;
auto idx = nd4j::math::nd4j_abs<Nd4jLong >((randomValue >> 16) % negLength);
irow = idx >= negLength ? -1 : static_cast<int>(negTable[idx]);
if (irow < 0 || irow >= vocabSize)
irow = randomValue % (vocabSize - 1) + 1;
if (irow == nsStarter)
continue;
}
nSampling_<T>(syn0row, s1n.bufferWithOffset(irow * vectorLength), expTable, neu1e, alpha, vectorLength, r == 0 ? 1 : 0, expLength, infVector != nullptr);
}
}
#pragma omp simd
for (int e = 0; e < vectorLength; e++)
syn0row[e] += neu1e[e];
// optionally release temp arrays
if (vectorLength > 600)
delete[] neu1e;
}
