/** Validate every local entry of a matrix has a given value. */
void validate_mat(DistMat& mat, DataType expected) {
for (int i = 0; i < mat.LocalHeight(); ++i) {
for (int j = 0; j < mat.LocalWidth(); ++j) {
ASSERT_EQ(mat.GetLocal(i, j), expected);
}
}
}
static bool readDist(int fd, const char* filename, DistMat& M, uint64_t* bytes)
{
struct layer_header header;
ssize_t read_rc = read(fd, &header, sizeof(header));
if (read_rc != sizeof(header)) {
// error!
}
*bytes += read_rc;
// check that header values match up
Int height = header.height;
Int width = header.width;
M.Resize(height, width);
if(M.ColStride() == 1 && M.RowStride() == 1) {
if(M.Height() == M.LDim()) {
void* buf = (void*) M.Buffer();
size_t bufsize = height * width * sizeof(DataType);
read_rc = read(fd, buf, bufsize);
if (read_rc != bufsize) {
// error!
}
*bytes += read_rc;
} else {
for(Int j = 0; j < width; ++j) {
void* buf = (void*) M.Buffer(0, j);
size_t bufsize = height * sizeof(DataType);
read_rc = read(fd, buf, bufsize);
if (read_rc != bufsize) {
// error!
}
*bytes += read_rc;
}
}
} else {
const Int localHeight = M.LocalHeight();
const Int localWidth = M.LocalWidth();
const Int lDim = M.LDim();
if(localHeight == lDim) {
void* buf = (void*) M.Buffer();
size_t bufsize = localHeight * localWidth * sizeof(DataType);
read_rc = read(fd, buf, bufsize);
if (read_rc != bufsize) {
// error!
}
*bytes += read_rc;
} else {
for(Int jLoc = 0; jLoc < localWidth; ++jLoc) {
void* buf = (void*) M.Buffer(0, jLoc);
size_t bufsize = localHeight * sizeof(DataType);
read_rc = read(fd, buf, bufsize);
if (read_rc != bufsize) {
// error!
}
*bytes += read_rc;
}
}
}
return true;
}
static bool writeDist(int fd, const char* filename, const DistMat& M, uint64_t* bytes)
{
struct layer_header header;
header.rank = (uint64_t) M.Grid().Rank();
header.width = (uint64_t) M.Width();
header.height = (uint64_t) M.Height();
header.localwidth = (uint64_t) M.LocalWidth();
header.localheight = (uint64_t) M.LocalHeight();
header.ldim = (uint64_t) M.LDim();
ssize_t write_rc = write(fd, &header, sizeof(header));
if (write_rc != sizeof(header)) {
// error!
}
*bytes += write_rc;
const Int localHeight = M.LocalHeight();
const Int localWidth = M.LocalWidth();
const Int lDim = M.LDim();
if(localHeight == lDim) {
void* buf = (void*) M.LockedBuffer();
size_t bufsize = localHeight * localWidth * sizeof(DataType);
write_rc = write(fd, buf, bufsize);
if (write_rc != bufsize) {
// error!
}
*bytes += write_rc;
} else {
for(Int j = 0; j < localWidth; ++j) {
void* buf = (void*) M.LockedBuffer(0, j);
size_t bufsize = localHeight * sizeof(DataType);
write_rc = write(fd, buf, bufsize);
if (write_rc != bufsize) {
// error!
}
*bytes += write_rc;
}
}
return true;
}
/** Create a new matrix. */
void create_mat(DistMat& mat, DataType def = 1.0) {
mat.Resize(LBANN_COMM_TEST_NROWS, LBANN_COMM_TEST_NCOLS);
El::Fill(mat, def);
}
double cg(gl_types::core * _glcore, std::vector<double> & means, double &diff, advanced_config & config){
glcore = _glcore;
int tmp=ps.n; ps.n=ps.m; ps.m = tmp;
diff = NAN;
init_row_cols();
DistMat A;
DistVec b(0,true), prec(1,true), r(2, true), p(3,true), x(4,true), Ap(5), t(6,true);
//initialize startng guess
if (!config.square)
x = ones(0.5,ps.m);
else
x = ones(0.5,ps.n);
DistDouble rsold, rnew, alpha;
/* r = -A*x+b;
if (~sqaure)
r=A'*r;
end
p = r;
rsold = r'*r;
*/
r=-A*x+b;
if (!config.square)
r=A._transpose()*r;
p = r;
rsold = r._transpose()*r;
/*
for i=1:size(A,1)
Ap=A*p;
if (~config.square)
Ap=A'*Ap;
end
alpha=rsold/(p'*Ap);
x=x+alpha*p;
r=r-alpha*Ap;
rsnew=r'*r;
if sqrt(rsnew)<1e-10
break;
end
p=r+rsnew/rsold*p;
rsold=rsnew;
end
*/
for (int i=1; i <= std::min(config.iter,size(A,1)); i++){
Ap=A*p;
if (!config.square)
Ap= A._transpose()*Ap;
alpha = rsold/(p._transpose()*Ap);
x=x+alpha*p;
if (config.cg_resid){
t=A*x-b;
logstream(LOG_INFO)<<"Iteration " << i << " approximated solution redidual is " << norm(t).toDouble() << std::endl;
}
r=r-alpha*Ap;
rnew=r._transpose()*r;
if (sqrt(rnew)<1e-10){
diff = sqrt(rnew).toDouble();
logstream(LOG_INFO)<<" Conjugate gradient converged in iteration "<<i<<" to an accuracy of " << diff << std::endl;
break;
}
p=r+(rnew/rsold)*p;
rsold=rnew;
}
x.to_vec(means);
return runtime;
}
