void buildProgram(cl::Program &prog, const int num_files,
const char **ker_strs, const int *ker_lens, std::string options)
{
try {
Program::Sources setSrc;
setSrc.emplace_back(USE_DBL_SRC_STR.c_str(), USE_DBL_SRC_STR.length());
setSrc.emplace_back(KParam_hpp, KParam_hpp_len);
for (int i = 0; i < num_files; i++) {
setSrc.emplace_back(ker_strs[i], ker_lens[i]);
}
static std::string defaults =
std::string(" -cl-std=CL1.1") + std::string(" -D dim_type=") +
std::string(dtype_traits<dim_type>::getName());
prog = cl::Program(getContext(), setSrc);
std::vector<cl::Device> targetDevices;
targetDevices.push_back(getDevice());
prog.build(targetDevices, (defaults + options).c_str());
} catch (...) {
SHOW_BUILD_INFO(prog);
throw;
}
}
void random(cl::Buffer out, dim_type elements)
{
try {
static unsigned counter;
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static Program ranProgs[DeviceManager::MAX_DEVICES];
static Kernel ranKernels[DeviceManager::MAX_DEVICES];
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
Program::Sources setSrc;
setSrc.emplace_back(random_cl, random_cl_len);
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D repeat="<< REPEAT
<< " -D " << random_name<T, isRandu>().name();
if (std::is_same<T, double>::value) {
options << " -D USE_DOUBLE";
options << " -D IS_64";
}
if (std::is_same<T, char>::value) {
options << " -D IS_BOOL";
}
buildProgram(ranProgs[device], random_cl, random_cl_len, options.str());
ranKernels[device] = Kernel(ranProgs[device], "random");
});
auto randomOp = make_kernel<cl::Buffer, uint, uint, uint, uint>(ranKernels[device]);
uint groups = divup(elements, THREADS * REPEAT);
counter += divup(elements, THREADS * groups);
NDRange local(THREADS, 1);
NDRange global(THREADS * groups, 1);
randomOp(EnqueueArgs(getQueue(), global, local),
out, elements, counter, random_seed[0], random_seed[1]);
} catch(cl::Error ex) {
CL_TO_AF_ERROR(ex);
}
}
void
set(Buffer &ptr, T val, const size_t &elements)
{
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static Program setProgs[DeviceManager::MAX_DEVICES];
static Kernel setKernels[DeviceManager::MAX_DEVICES];
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
Program::Sources setSrc;
setSrc.emplace_back(set_cl, set_cl_len);
setProgs[device] = Program(getContext(), setSrc);
string opt = string("-D T=") + dtype_traits<T>::getName();
setProgs[device].build(opt.c_str());
setKernels[device] = Kernel(setProgs[device], "set");
});
auto setKern = make_kernel<Buffer, T, const unsigned long>(setKernels[device]);
setKern(EnqueueArgs(getQueue(), NDRange(elements)), ptr, val, elements);
}