void operator()(const communicator& comm, const config& cfg,
T alpha, MatrixA& A, MatrixB& B, T beta, MatrixC& C)
{
using namespace matrix_constants;
len_type m = (Mat == MAT_A ? A.length(0) : Mat == MAT_B ? B.length(0) : C.length(0));
len_type n = (Mat == MAT_A ? A.length(1) : Mat == MAT_B ? B.length(1) : C.length(1));
if (!rscat)
{
if (comm.master())
{
scat_buffer = Pool.allocate<stride_type>(2*m + 2*n);
rscat = scat_buffer.get<stride_type>();
}
comm.broadcast(rscat);
cscat = rscat+m;
rbs = cscat+n;
cbs = rbs+m;
}
matrify_and_run<Mat>(*this, comm, cfg, alpha, A, B, beta, C);
}
void operator()(const communicator& comm, const config& cfg,
T alpha, MatrixA& A, MatrixB& B, T beta, MatrixC& C)
{
using namespace matrix_constants;
const len_type MR = (Mat == MAT_B ? cfg.gemm_kr.def<T>()
: cfg.gemm_mr.def<T>());
const len_type NR = (Mat == MAT_A ? cfg.gemm_kr.def<T>()
: cfg.gemm_nr.def<T>());
len_type m = (Mat == MAT_A ? A.length(0) : Mat == MAT_B ? B.length(0) : C.length(0));
len_type n = (Mat == MAT_A ? A.length(1) : Mat == MAT_B ? B.length(1) : C.length(1));
m = round_up(m, MR);
n = round_up(n, NR);
auto& pack_buffer = child.pack_buffer;
auto& pack_ptr = child.pack_ptr;
if (!pack_ptr)
{
if (comm.master())
{
len_type scatter_size = size_as_type<stride_type,T>(2*m + 2*n);
pack_buffer = Pool.allocate<T>(m*n + std::max(m,n)*TBLIS_MAX_UNROLL + scatter_size);
pack_ptr = pack_buffer.get();
}
comm.broadcast(pack_ptr);
rscat = convert_and_align<T,stride_type>(static_cast<T*>(pack_ptr) + m*n);
cscat = rscat+m;
rbs = cscat+n;
cbs = rbs+m;
}
Sib::operator()(comm, cfg, alpha, A, B, beta, C);
}