void laset(int m, int n,
T offdiag, T diag,
cl_mem dA, size_t dA_offset, magma_int_t ldda)
{
std::string refName = laset_name<uplo>() + std::string("_") +
std::string(dtype_traits<T>::getName()) +
std::to_string(uplo);
int device = getActiveDeviceId();
kc_entry_t entry = kernelCache(device, refName);
if (entry.prog==0 && entry.ker==0) {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D BLK_X=" << BLK_X
<< " -D BLK_Y=" << BLK_Y
<< " -D IS_CPLX=" << af::iscplx<T>();
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value)
options << " -D USE_DOUBLE";
const char* ker_strs[] = {laset_cl};
const int ker_lens[] = {laset_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, laset_name<uplo>());
addKernelToCache(device, refName, entry);
}
int groups_x = (m - 1) / BLK_X + 1;
int groups_y = (n - 1) / BLK_Y + 1;
NDRange local(BLK_X, 1);
NDRange global(groups_x * local[0], groups_y * local[1]);
// retain the cl_mem object during cl::Buffer creation
cl::Buffer dAObj(dA, true);
auto lasetOp = KernelFunctor<int, int, T, T, Buffer, unsigned long long, int>(*entry.ker);
lasetOp(EnqueueArgs(getQueue(), global, local), m, n, offdiag, diag, dAObj, dA_offset, ldda);
}
void swapdblk(int n, int nb, cl_mem dA, size_t dA_offset, int ldda, int inca,
cl_mem dB, size_t dB_offset, int lddb, int incb,
cl_command_queue queue) {
std::string refName =
std::string("swapdblk_") + std::string(dtype_traits<T>::getName());
int device = getActiveDeviceId();
kc_entry_t entry = kernelCache(device, refName);
if (entry.prog == 0 && entry.ker == 0) {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName();
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value)
options << " -D USE_DOUBLE";
const char* ker_strs[] = {swapdblk_cl};
const int ker_lens[] = {swapdblk_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "swapdblk");
addKernelToCache(device, refName, entry);
}
int nblocks = n / nb;
if (nblocks == 0) return;
int info = 0;
if (n < 0) {
info = -1;
} else if (nb < 1 || nb > 1024) {
info = -2;
} else if (ldda < (nblocks - 1) * nb * inca + nb) {
info = -4;
} else if (inca < 0) {
info = -5;
} else if (lddb < (nblocks - 1) * nb * incb + nb) {
info = -7;
} else if (incb < 0) {
info = -8;
}
if (info != 0) {
AF_ERROR("Invalid configuration", AF_ERR_INTERNAL);
return;
}
NDRange local(nb);
NDRange global(nblocks * nb);
cl::Buffer dAObj(dA, true);
cl::Buffer dBObj(dB, true);
auto swapdOp =
KernelFunctor<int, Buffer, unsigned long long, int, int, Buffer,
unsigned long long, int, int>(*entry.ker);
cl::CommandQueue q(queue);
swapdOp(EnqueueArgs(q, global, local), nb, dAObj, dA_offset, ldda, inca,
dBObj, dB_offset, lddb, incb);
}
