[SYCL] `platform()` return a platform with no device

[TApplencourt]

Describe the bug

The sycl::plarform() constructor return a platform with no device. This is forbidden by the spec. But maybe I do something wronrg...

To Reproduce

$cat foo.cpp
#include <CL/sycl.hpp>
namespace sycl = cl::sycl;
int main() {
#ifdef DEFAULT
  sycl::platform P = sycl::platform(sycl::default_selector{});
#else
  sycl::platform P = sycl::platform();
#endif
  assert(!P.get_devices().empty());
}
$dpcpp foo.cpp
$./a.out
a.out: foo.cpp:10: int main(): Assertion `!P.get_devices().empty()' failed.
Aborted

Environment (please complete the following information):

• OS: Linux
• Target device and vendor: Intel GPU / OpenCL / Level 0
• DPC++ version: 2021.x.0.20210819

Additional context

https://www.khronos.org/registry/SYCL/specs/sycl-2020/html/sycl-2020.html#sec:platform-class said for the platform() constructor:

Constructs a SYCL platform instance that is a copy of the platform which contains the device returned by default_selector_v.

This doesn't seem to be the case here. As the spec said for default_selector_v:

Since all implementations must support at least one device, this selector must always return a device.

But our platform is empty. Also, Compiling the code with -DDEFAULT behaves as expected and returns a platform with at least one device.

Don't hesitate if you have any questions,

[romanovvlad]

One part of the problem is that SYCL1.2.1 says that the platform default constructor should construct a host platform. We have not updated it to SYCL2020 behavior.
Another part is that the host device is disabled in dpcpp builds which are distributed as part of oneAPI, that's why it has no devices.

[TApplencourt]

Thanks for the explanation! So it's more a "feature request" to implement SYCL2020 behavior than a bug :).

This is not a super high priority feature, just a little verbose to always pass a selector to the platform constructor in my teaching example. But maybe It's a good practice so worth the verbosity.

Thanks again,
Thomas

[alexbatashev]

Related PR: #5388