sycl-thrust

sycl-thrust is an implementation of the Thrust library using SYCL. This allows you to run Thrust programs using standard C++ data structures and algorithms on Intel GPUs.

⚠️ Warning: sycl-thrust is experimental, incomplete, and under active development.

Implemented Features

The current state implements just enough to allocate data on the device, initialize it, and copy back and forth to the host. It also implements some basic execution policies.

Feature	Status
device_vector	✅ Implemented
device_allocator	✅ Implemented
copy	✅ Implemented
fill	✅ Implemented
sort	❌ Missing
reduce	❌ Missing
...others	❌ Missing

Example

#include <thrust/device_vector.h>
#include <thrust/copy.h>
#include <thrust/fill.h>

#include <fmt/ranges.h>

int main(int argc, char** argv) {
  // Create a device vector initialized with 10
  // elements, all equal to 1.
  thrust::device_vector<int> d_v(10, 1);

  // Fill the value "7" into the first 7 elements.
  thrust::fill(d_v.begin(), d_v.begin() + 7, 7);

  // Create a std::vector on the CPU.
  std::vector<int> v(10);

  // Copy our device vector to the CPU.
  thrust::copy(d_v.begin(), d_v.begin() + 10, v.begin());

  std::fill(v.begin(), v.end(), 12);

  thrust::copy(v.begin(), v.end(), d_v.begin());

  fmt::print("{}\n", d_v);

  return 0;
}

Default Device Behavior

By default, sycl-thrust will use the sycl::default_selector_v selector to pick the default device for both device_allocator and the device execution policy. The sycl::default_selector_v will generally pick a GPU device if available, falling back to a CPU device if none is available. This allows code to run successfully even if no GPU is available. If you would like to force sycl-thrust to use a GPU, you can build with CMake flag -DTHRUST_DEFAULT_GPU=ON, or by directly setting the compile directive -DTHRUST_DEFAULT_GPU if you are not using CMake. This may be necessary to run on NVIDIA or AMD GPUs using the Codeplay oneAPI plugin, as the default selector may fail to select these GPUs.