What to do if I want to use Hip toolchain?

[wangzy0327]

When I compile oneDNN source files using make -j 20,I get error as following:
error: no matching function for call to 'get_native'
error: no matching function for call to 'get_native_mem'

I ask about the problem details on oneDNN

So, what can I get some help?

[alexbatashev]

HIP backend is experimental and currently does not support interoperability APIs.

[TejaX-Alaghari]

By creating the "backend_traits_hip.hpp" with appropriate forward declarations, I was able to compile the code with "get_native" calls when selecting a hip backend.

I'm assuming there's more to the done for running the code on an AMD platform with this experimental support.
@alexbatashev, can you please expand on this further?

I can submit the "backend_traits_hip.hpp", for reference if necessary.

[alexbatashev]

@TejaX-Alaghari you basically need two things to make it work:

1. A specification. Something similar to OpenCL interop: https://www.khronos.org/registry/SYCL/specs/sycl-2020/html/sycl-2020.html#chapter:opencl-backend or Level Zero https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/supported/sycl_ext_oneapi_backend_level_zero.md
   Once it is reviewed and approved, you can proceed with implementation.

2. You need to update HIP plugin to support interop APIs.
   Function definitions can be found in pi.h, you are looking for functions with Native in their names.
   Implementation should be put here: https://github.com/intel/llvm/tree/sycl/sycl/plugins/hip

Since writing a good spec without implementation is hard, you may want to proceed with both at the same time. You can contact @intel/dpcpp-specification-reviewers and @AerialMantis for help with reviewing the spec and implementation.

[wangzy0327]

By creating the "backend_traits_hip.hpp" with appropriate forward declarations, I was able to compile the code with "get_native" calls when selecting a hip backend.

I'm assuming there's more to the done for running the code on an AMD platform with this experimental support. @alexbatashev, can you please expand on this further?

I can submit the "backend_traits_hip.hpp", for reference if necessary.

By creating the backend_traits_hip.hpp with appropriate forward declarations and hip.hpp, I was not able to compile the code with "get_native" calls when selecting a hip backend.
when I When I compile oneDNN source files using make -j 20,I get error as following:
error: no matching function for call to 'get_native'
error: no matching function for call to 'get_native_mem'
Can you help me to find the problem?

[alexbatashev]

@wangzy0327 you should look at how level zero is implemented here, HIP or CUDA would look similarly.

[AerialMantis]

Since writing a good spec without implementation is hard, you may want to proceed with both at the same time. You can contact https://github.com/orgs/intel/teams/dpcpp-specification-reviewers and @AerialMantis for help with reviewing the spec and implementation.

I'd be happy to help review. We're in the process of defining and implementing a backend specification for CUDA so you can also reference this as I suspect there will be a lot of similarities between the CUDA and HIP backends.

• Backend spec: Introduce CUDA backend specification KhronosGroup/SYCL-Docs#197
• SYCL-CTS test plan: Introduce test plan for CUDA backend interop KhronosGroup/SYCL-CTS#207
• Draft implementation of updated interop APIs: [SYCL][CUDA] Add updated CUDA backend interop extension #5658

By creating the backend_traits_hip.hpp with appropriate forward declarations and hip.hpp, I was not able to compile the code with "get_native" calls when selecting a hip backend.

From a glance it looks as though it's not finding the HIP backend specializations, is backend_traits_hip.hpp being included in https://github.com/intel/llvm/blob/sycl/sycl/include/CL/sycl/backend.hpp?

[wangzy0327]

Since writing a good spec without implementation is hard, you may want to proceed with both at the same time. You can contact https://github.com/orgs/intel/teams/dpcpp-specification-reviewers and @AerialMantis for help with reviewing the spec and implementation.

I'd be happy to help review. We're in the process of defining and implementing a backend specification for CUDA so you can also reference this as I suspect there will be a lot of similarities between the CUDA and HIP backends.

• Backend spec: Introduce CUDA backend specification KhronosGroup/SYCL-Docs#197
• SYCL-CTS test plan: Introduce test plan for CUDA backend interop KhronosGroup/SYCL-CTS#207
• Draft implementation of updated interop APIs: [SYCL][CUDA] Add updated CUDA backend interop extension #5658

By creating the backend_traits_hip.hpp with appropriate forward declarations and hip.hpp, I was not able to compile the code with "get_native" calls when selecting a hip backend.

From a glance it looks as though it's not finding the HIP backend specializations, is backend_traits_hip.hpp being included in https://github.com/intel/llvm/blob/sycl/sycl/include/CL/sycl/backend.hpp?

OK，Thank you very much. I will refer these

• Backend spec: [Introduce CUDA backend specification Introduce CUDA backend specification KhronosGroup/SYCL-Docs#197]
• SYCL-CTS test plan: Introduce test plan for CUDA backend interop KhronosGroup/SYCL-CTS#207
• Draft implementation of updated interop APIs: [SYCL][CUDA] Add updated CUDA backend interop extension #5658