16struct KernelExecutionInfo {
17 double elapsed_summary = 0.0;
18 int enqueue_count = 0;
23OneapiDeviceQueue::OneapiDeviceQueue(OneapiDevice *device)
35 const int max_num_threads = oneapi_device_->get_num_multiprocessors() *
36 oneapi_device_->get_max_num_threads_per_multiprocessor();
39 params.
baseline = 16 *
max(8 * max_num_threads, 65536);
45 params.
baseline = 32768 * oneapi_device_->get_num_multiprocessors();
51 if (oneapi_device_->get_max_num_threads_per_multiprocessor() < 128) {
52 const int max_num_threads = oneapi_device_->get_num_multiprocessors() *
53 oneapi_device_->get_max_num_threads_per_multiprocessor();
54 params.
baseline = 16 *
max(8 * max_num_threads, 65536);
68void OneapiDeviceQueue::get_memory_info(
size_t &total,
size_t &free)
const
70 oneapi_device_->get_device_memory_info(total, free);
73int OneapiDeviceQueue::num_sort_partitions(
int max_num_paths, uint )
const
75 int sort_partition_elements = (oneapi_device_->get_max_num_threads_per_multiprocessor() >= 128) ?
80 return max(max_num_paths / sort_partition_elements, 1);
83void OneapiDeviceQueue::init_execution()
85 oneapi_device_->load_image_info(
nullptr);
87 SyclQueue *device_queue = oneapi_device_->sycl_queue();
88 void *kg_dptr = oneapi_device_->kernel_globals_device_pointer();
91 kernel_context_ = make_unique<KernelContext>();
92 kernel_context_->queue = device_queue;
93 kernel_context_->kernel_globals = kg_dptr;
95 debug_init_execution();
98void OneapiDeviceQueue::load_image_info()
100 oneapi_device_->load_image_info(
this);
104 const int signed_kernel_work_size,
107 if (oneapi_device_->have_error()) {
112 if (oneapi_device_->load_image_info(
nullptr)) {
113 if (!synchronize()) {
118 void **args =
const_cast<void **
>(_args.
values);
120 debug_enqueue_begin(kernel, signed_kernel_work_size);
121 assert(signed_kernel_work_size >= 0);
122 size_t kernel_global_size = (size_t)signed_kernel_work_size;
123 size_t kernel_local_size;
126 kernel_context_->scene_max_shaders = oneapi_device_->scene_max_shaders();
128 oneapi_device_->get_adjusted_global_and_local_sizes(
129 kernel_context_->queue, kernel, kernel_global_size, kernel_local_size);
132 bool is_finished_ok = oneapi_device_->enqueue_kernel(
133 kernel_context_.get(), kernel, kernel_global_size, kernel_local_size, args);
135 if (is_finished_ok ==
false) {
137 "\" execution error: got runtime exception \"" +
138 oneapi_device_->oneapi_error_message() +
"\"");
143 return is_finished_ok;
146bool OneapiDeviceQueue::synchronize()
148 if (oneapi_device_->have_error()) {
152 bool is_finished_ok = oneapi_device_->queue_synchronize(oneapi_device_->sycl_queue());
153 if (is_finished_ok ==
false) {
154 oneapi_device_->set_error(
"oneAPI unknown kernel execution error: got runtime exception \"" +
155 oneapi_device_->oneapi_error_message() +
"\"");
160 return !(oneapi_device_->have_error());
165 oneapi_device_->mem_zero(mem);
170 oneapi_device_->mem_copy_to(mem);
175 oneapi_device_->mem_copy_from(mem);
178void *OneapiDeviceQueue::copy_from_device_synchronized(
device_memory &mem,
186 oneapi_device_->mem_copy_from(mem, 0, 0, 0, 0,
storage.data());
191# ifdef SYCL_LINEAR_MEMORY_INTEROP_AVAILABLE
194 return make_unique<OneapiDeviceGraphicsInterop>(
this);
#define CCL_NAMESPACE_END
const char * device_kernel_as_string(DeviceKernel kernel)
#define assert(assertion)
#define storage(slot, qualifiers)