inline void test_fill(T v1, T v2, T v3, bc::command_queue queue) {
if(boost::is_same<typename bc::scalar_type<T>::type, bc::double_>::value &&
!queue.get_device().supports_extension("cl_khr_fp64")) {
std::cerr << "Skipping test_fill<" << bc::type_name<T>() << ">() "
"on device which doesn't support cl_khr_fp64" << std::endl;
return;
}
bc::vector<T> vector(4, queue.get_context());
bc::fill(vector.begin(), vector.end(), v1, queue);
queue.finish();
CHECK_RANGE_EQUAL(T, 4, vector, (v1, v1, v1, v1));
vector.resize(1000, queue);
bc::fill(vector.begin(), vector.end(), v2, queue);
queue.finish();
BOOST_CHECK_EQUAL(vector.front(), v2);
BOOST_CHECK_EQUAL(vector.back(), v2);
bc::fill(vector.begin() + 500, vector.end(), v3, queue);
queue.finish();
BOOST_CHECK_EQUAL(vector.front(), v2);
BOOST_CHECK_EQUAL(vector[499], v2);
BOOST_CHECK_EQUAL(vector[500], v3);
BOOST_CHECK_EQUAL(vector.back(), v3);
}
inline void test_fill_n(T v1, T v2, T v3, bc::command_queue queue) {
if(boost::is_same<typename bc::scalar_type<T>::type, bc::double_>::value &&
!queue.get_device().supports_extension("cl_khr_fp64")) {
std::cerr << "Skipping test_fill_n<" << bc::type_name<T>() << ">() "
"on device which doesn't support cl_khr_fp64" << std::endl;
return;
}
bc::vector<T> vector(4, queue.get_context());
bc::fill_n(vector.begin(), 4, v1, queue);
queue.finish();
CHECK_RANGE_EQUAL(T, 4, vector, (v1, v1, v1, v1));
bc::fill_n(vector.begin(), 3, v2, queue);
queue.finish();
CHECK_RANGE_EQUAL(T, 4, vector, (v2, v2, v2, v1));
bc::fill_n(vector.begin() + 1, 2, v3, queue);
queue.finish();
CHECK_RANGE_EQUAL(T, 4, vector, (v2, v3, v3, v1));
bc::fill_n(vector.begin(), 4, v2, queue);
queue.finish();
CHECK_RANGE_EQUAL(T, 4, vector, (v2, v2, v2, v2));
// fill last element
bc::fill_n(vector.end() - 1, 1, v3, queue);
queue.finish();
CHECK_RANGE_EQUAL(T, 4, vector, (v2, v2, v2, v3));
// fill first element
bc::fill_n(vector.begin(), 1, v1, queue);
queue.finish();
CHECK_RANGE_EQUAL(T, 4, vector, (v1, v2, v2, v3));
}
/// Select best launch configuration for the given shared memory requirements.
void config(const boost::compute::command_queue &queue, std::function<size_t(size_t)> smem) {
boost::compute::device dev = queue.get_device();
size_t ws;
if ( is_cpu(queue) ) {
ws = 1;
} else {
// Select workgroup size that would fit into the device.
ws = dev.get_info<std::vector<size_t>>(CL_DEVICE_MAX_WORK_ITEM_SIZES)[0] / 2;
size_t max_ws = max_threads_per_block(queue);
size_t max_smem = max_shared_memory_per_block(queue);
// Reduce workgroup size until it satisfies resource requirements:
while( (ws > max_ws) || (smem(ws) > max_smem) )
ws /= 2;
}
config(num_workgroups(queue), ws);
}
void tune_accumulate(const compute::vector<T>& data,
const size_t trials,
compute::command_queue& queue)
{
boost::shared_ptr<compute::detail::parameter_cache>
params = compute::detail::parameter_cache::get_global_cache(queue.get_device());
const std::string cache_key =
std::string("__boost_reduce_on_gpu_") + compute::type_name<T>();
const compute::uint_ tpbs[] = { 4, 8, 16, 32, 64, 128, 256, 512, 1024 };
const compute::uint_ vpts[] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16 };
double min_time = std::numeric_limits<double>::max();
compute::uint_ best_tpb = 0;
compute::uint_ best_vpt = 0;
for(size_t i = 0; i < sizeof(tpbs) / sizeof(*tpbs); i++){
params->set(cache_key, "tpb", tpbs[i]);
for(size_t j = 0; j < sizeof(vpts) / sizeof(*vpts); j++){
params->set(cache_key, "vpt", vpts[j]);
try {
const double t = perf_accumulate(data, trials, queue);
if(t < min_time){
best_tpb = tpbs[i];
best_vpt = vpts[j];
min_time = t;
}
}
catch(compute::opencl_error&){
// invalid parameters for this device, skip
}
}
}
// store optimal parameters
params->set(cache_key, "tpb", best_tpb);
params->set(cache_key, "vpt", best_vpt);
}
size_t preferred_work_group_size_multiple(const boost::compute::command_queue &q) const {
return K.get_work_group_info<size_t>(q.get_device(), CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE);
}
/// The size in bytes of shared memory per block available for this kernel.
size_t max_shared_memory_per_block(const boost::compute::command_queue &q) const {
boost::compute::device d = q.get_device();
return d.local_memory_size() - K.get_work_group_info<cl_ulong>(d, CL_KERNEL_LOCAL_MEM_SIZE);
}
/// The maximum number of threads per block, beyond which a launch of the kernel would fail.
size_t max_threads_per_block(const boost::compute::command_queue &q) const {
return K.get_work_group_info<size_t>(q.get_device(), CL_KERNEL_WORK_GROUP_SIZE);
}
/// Standard number of workgroups to launch on a device.
static inline size_t num_workgroups(const boost::compute::command_queue &q) {
// This is a simple heuristic-based estimate. More advanced technique may
// be employed later.
return 8 * q.get_device().compute_units();
}
