Example

In the following example, a fixed_resource_policy is used when the code is compiled so that it selects a specific device. When USE_CPU is defined at compile-time, this example always uses the CPU queue. When USE_GPU is defined at compile-time, it always uses the GPU queue. Otherwise, it uses an auto_tune_policy to dynamically select between these two queues. Such a scenario could be used for debugging or simply to maintain the dynamic selection code even if the best device to use is known for some subset of platforms.

#include <oneapi/dpl/dynamic_selection>
#include <sycl/sycl.hpp>
#include <iostream>

namespace ex = oneapi::dpl::experimental;

int main() {
  std::vector<sycl::queue> r { sycl::queue{sycl::cpu_selector_v},
                               sycl::queue{sycl::gpu_selector_v} };

  const std::size_t N = 10000;
  std::vector<float> av(N, 0.0);
  std::vector<float> bv(N, 0.0);
  std::vector<float> cv(N, 0.0);
  for (int i = 0; i < N; ++i) {
    av[i] = bv[i] = i;
  }

#if USE_CPU
  ex::fixed_resource_policy p{r};    // (1) uses index 0 of r, the cpu
#elif USE_GPU
  ex::fixed_resource_policy p{r, 1}; // (2) uses index 1 of r, the gpu
#else
  ex::auto_tune_policy p{r};
#endif

  {
    sycl::buffer<float> a_b(av);
    sycl::buffer<float> b_b(bv);
    sycl::buffer<float> c_b(cv);


    for (int i = 0; i < 6; ++i) {
      ex::submit_and_wait(p, [&](sycl::queue q) { // (3)
        // (4)
        std::cout << (q.get_device().is_cpu() ? "using cpu\n" : "using gpu\n");
        return q.submit([&](sycl::handler &h) {   // (5)
          sycl::accessor a_a(a_b, h, sycl::read_only);
          sycl::accessor b_a(b_b, h, sycl::read_only);
          sycl::accessor c_a(c_b, h, sycl::read_write);
          h.parallel_for(N, [=](auto i) { c_a[i] = a_a[i] + b_a[i]; });
        });
      });
    };
  }

  for (int i = 0; i < N; ++i) {
    if (cv[i] != 2*i) {
       std::cout << "ERROR!\n";
    }
  }
  std::cout << "Done.\n";
}

The key points in this example are:

1. If USE_CPU is defined, a fixed_resouce_policy is constructed that targets the CPU.

2. If USE_GPU is defined, a fixed_resouce_policy is constructed that targets the GPU.

3. submit_and_wait is invoked with the policy as the first argument. The selected queue will be passed to the user-provided function.

4. For clarity when run, the type of device is displayed.

5. The queue is used in a function to perform an asynchronous offload. The SYCL event returned from the call to submit is returned. Returning an event is required for functions passed to submit and submit_and_wait.

Selection Algorithm

The selection algorithm for fixed_resource_policy always returns the same specific resource from its set of resources. The index of the resource is set during construction or deferred initialization.

Simplified, expository implementation of the selection algorithm:

template<typename... Args>
selection_type fixed_resource_policy::select(Args&& ...) {
  if (initialized_) {
    return selection_type{*this, resources_[fixed_offset_]};
  } else {
    throw std::logic_error(“select called before initialization”);
  }
}

where resources_ is a container of resources, such as std::vector of sycl::queue, and fixed_offset_ stores a fixed integer offset. Both resources_ and fixed_offset are set during construction or deferred initialization of the policy and then remain constant.

Constructors

fixed_resource_policy provides three constructors.

Deferred Initialization

A fixed_resource_policy that was constructed with deferred initialization must be initialized by calling one its initialize member functions before it can be used to select or submit.

Queries

A fixed_resource_policy has get_resources and get_submission_group member functions.

Reporting Requirements

If a resource returned by select is used directly without calling submit or submit_and_wait, it may be necessary to call report to provide feedback to the policy. However, the fixed_resource_policy does not require any feedback about the system state or the behavior of the workload. Therefore, no explicit reporting of execution information is needed, as is summarized in the table below.

In generic code, it is possible to perform compile-time checks to avoid reporting overheads when reporting is not needed, while still writing code that will work with any policy, as demonstrated below:

auto s = select(my_policy);
if constexpr (report_info_v<decltype(s), execution_info::task_submission_t>)
{
  s.report(execution_info::task_submission);
}