[SYCL][CUDA][HIP] Add support for querying device architecture #10573
Conversation
…prove performance
MartinWehking
changed the title
|
Can you also update the extension specification to remove the following limitation:
|
sycl/source/detail/device_info.hpp
Outdated
| @@ -568,64 +568,108 @@ struct get_device_info_impl<range<Dimensions>, | |||
| } | |||
| }; | |||
|
|
|||
| // This macro is only for AMD, NVIDIA and Intel GPU architectures | |||
| #define ARCHES(X) \ | |||
There was a problem hiding this comment.
This can be done in a separate PR later during some refactoring but we can have two std::unordered_maps: one for string keys and another for uint32_t keys: this way in case of Intel platforms we will not convert number to string and compare strings - it will be faster. For AMD and NVIDIA archs with std::unordered_map the access time will be O(1) in a good case and O(N) in a bad case.
There was a problem hiding this comment.
It turns out that the way the uint32_t keys were converted was wrong anyways. The converted strings were not in hexadecimal form anymore and that's also why the CI failed.
I've added another macro for now to avoid the conversion and leave the DeviceIP as an integer.
|
@aelovikov-intel re-triggering of SYCL Pre Commit on Linux / test (AMD/HIP and Intel) didn't help, how it can be fixed? Possibly, some problem with CI machines. |
Looks like a common problem at least with Intel devices: #9845 (comment) |
|
Restarted the checks, it was reported that issue with Intel devices should be fixed now. |
|
Unrelated failure on HIP: |
Implements
sycl_ext_oneapi_device_architecturefor AMD and NVIDIA.The device architecture can be queried through
dev.get_info<sycl::ext::oneapi::experimental::info::device::architecture>()(
devis asycl::device)