void trainOneEpochDenseCPU(int itask, float *data, float *numerator,
float *denominator, float *codebook,
unsigned int nSomX, unsigned int nSomY,
unsigned int nDimensions, unsigned int nVectors,
unsigned int nVectorsPerRank, float radius,
float scale, string mapType, int *globalBmus)
{
unsigned int p1[2] = {0, 0};
unsigned int *bmus = new unsigned int[nVectorsPerRank*2];
#pragma omp parallel default(shared) private(p1)
{
#pragma omp for
for (unsigned int n = 0; n < nVectorsPerRank; n++) {
if (itask*nVectorsPerRank+n<nVectors) {
/// get the best matching unit
get_bmu_coord(codebook, data, nSomY, nSomX,
nDimensions, p1, n);
bmus[2*n] = p1[0]; bmus[2*n+1] = p1[1];
}
}
}
float *localNumerator = new float[nSomY*nSomX*nDimensions];
float *localDenominator = new float[nSomY*nSomX];
#pragma omp parallel default(shared)
{
#pragma omp for
for (unsigned int som_y = 0; som_y < nSomY; som_y++) {
for (unsigned int som_x = 0; som_x < nSomX; som_x++) {
localDenominator[som_y*nSomX + som_x] = 0.0;
for (unsigned int d = 0; d < nDimensions; d++)
localNumerator[som_y*nSomX*nDimensions + som_x*nDimensions + d] = 0.0;
}
}
/// Accumulate denoms and numers
#pragma omp for
for (unsigned int som_y = 0; som_y < nSomY; som_y++) {
for (unsigned int som_x = 0; som_x < nSomX; som_x++) {
for (unsigned int n = 0; n < nVectorsPerRank; n++) {
if (itask*nVectorsPerRank+n<nVectors) {
float dist = 0.0f;
if (mapType == "planar") {
dist = euclideanDistanceOnPlanarMap(som_x, som_y, bmus[2*n], bmus[2*n+1]);
} else if (mapType == "toroid") {
dist = euclideanDistanceOnToroidMap(som_x, som_y, bmus[2*n], bmus[2*n+1], nSomX, nSomY);
}
float neighbor_fuct = getWeight(dist, radius, scale);
for (unsigned int d = 0; d < nDimensions; d++) {
localNumerator[som_y*nSomX*nDimensions + som_x*nDimensions + d] +=
1.0f * neighbor_fuct
* (*(data + n*nDimensions + d));
}
localDenominator[som_y*nSomX + som_x] += neighbor_fuct;
}
}
}
}
}
#ifdef HAVE_MPI
MPI_Reduce(localNumerator, numerator,
nSomY*nSomX*nDimensions, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(localDenominator, denominator,
nSomY*nSomX, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Gather(bmus, nVectorsPerRank*2, MPI_INT, globalBmus, nVectorsPerRank*2, MPI_INT, 0, MPI_COMM_WORLD);
#else
for (unsigned int i=0; i < nSomY*nSomX*nDimensions; ++i) {
numerator[i] = localNumerator[i];
}
for (unsigned int i=0; i < nSomY*nSomX; ++i) {
denominator[i] = localDenominator[i];
}
for (unsigned int i=0; i < 2*nVectorsPerRank; ++i) {
globalBmus[i]=bmus[i];
}
#endif
delete [] bmus;
delete [] localNumerator;
delete [] localDenominator;
}
void trainOneEpochSparseCPU(int itask, svm_node **sparseData, float *numerator,
float *denominator, float *codebook,
unsigned int nSomX, unsigned int nSomY,
unsigned int nDimensions, unsigned int nVectors,
unsigned int nVectorsPerRank, float radius,
float scale, string mapType, string gridType,
bool compact_support, bool gaussian, int *globalBmus) {
int p1[2] = {0, 0};
int *bmus = new int[nVectorsPerRank * 2];
#ifdef _OPENMP
#pragma omp parallel default(shared) private(p1)
#endif
{
#ifdef _OPENMP
#pragma omp for
#endif
#ifdef _WIN32
for (int n = 0; n < nVectorsPerRank; n++) {
#else
for (unsigned int n = 0; n < nVectorsPerRank; n++) {
#endif
if (itask * nVectorsPerRank + n < nVectors) {
/// get the best matching unit
get_bmu_coord(codebook, sparseData, nSomY, nSomX,
nDimensions, p1, n);
bmus[2 * n] = p1[0];
bmus[2 * n + 1] = p1[1];
}
}
}
float *localNumerator = new float[nSomY * nSomX * nDimensions];
float *localDenominator = new float[nSomY * nSomX];
#ifdef _OPENMP
#pragma omp parallel default(shared)
#endif
{
#ifdef _OPENMP
#pragma omp for
#endif
#ifdef _WIN32
for (int som_y = 0; som_y < nSomY; som_y++) {
#else
for (unsigned int som_y = 0; som_y < nSomY; som_y++) {
#endif
for (unsigned int som_x = 0; som_x < nSomX; som_x++) {
localDenominator[som_y * nSomX + som_x] = 0.0;
for (unsigned int d = 0; d < nDimensions; d++)
localNumerator[som_y * nSomX * nDimensions + som_x * nDimensions + d] = 0.0;
}
}
/// Accumulate denoms and numers
#ifdef _OPENMP
#pragma omp for
#endif
#ifdef _WIN32
for (int som_y = 0; som_y < nSomY; som_y++) {
#else
for (unsigned int som_y = 0; som_y < nSomY; som_y++) {
#endif
for (unsigned int som_x = 0; som_x < nSomX; som_x++) {
for (unsigned int n = 0; n < nVectorsPerRank; n++) {
if (itask * nVectorsPerRank + n < nVectors) {
float dist = 0.0f;
if (gridType == "rectangular") {
if (mapType == "planar") {
dist = euclideanDistanceOnPlanarMap(som_x, som_y, bmus[2 * n], bmus[2 * n + 1]);
}
else if (mapType == "toroid") {
dist = euclideanDistanceOnToroidMap(som_x, som_y, bmus[2 * n], bmus[2 * n + 1], nSomX, nSomY);
}
}
else {
if (mapType == "planar") {
dist = euclideanDistanceOnHexagonalPlanarMap(som_x, som_y, bmus[2 * n], bmus[2 * n + 1]);
}
else if (mapType == "toroid") {
dist = euclideanDistanceOnHexagonalToroidMap(som_x, som_y, bmus[2 * n], bmus[2 * n + 1], nSomX, nSomY);
}
}
float neighbor_fuct = getWeight(dist, radius, scale, compact_support, gaussian);
unsigned int j = 0;
while ( sparseData[n][j].index != -1 ) {
localNumerator[som_y * nSomX * nDimensions +
som_x * nDimensions +
sparseData[n][j].index] +=
1.0f * neighbor_fuct * sparseData[n][j].value;
j++;
}
localDenominator[som_y * nSomX + som_x] += neighbor_fuct;
}
}
}
}
}
#ifdef HAVE_MPI
MPI_Reduce(localNumerator, numerator,
nSomY * nSomX * nDimensions, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Reduce(localDenominator, denominator,
nSomY * nSomX, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD);
MPI_Gather(bmus, nVectorsPerRank * 2, MPI_INT, globalBmus, nVectorsPerRank * 2, MPI_INT, 0, MPI_COMM_WORLD);
#else
for (unsigned int i = 0; i < nSomY * nSomX * nDimensions; ++i) {
numerator[i] = localNumerator[i];
}
for (unsigned int i = 0; i < nSomY * nSomX; ++i) {
denominator[i] = localDenominator[i];
}
for (unsigned int i = 0; i < 2 * nVectorsPerRank; ++i) {
globalBmus[i] = bmus[i];
}
#endif
delete [] bmus;
delete [] localNumerator;
delete [] localDenominator;
}