gpu: nvidia: amd: Get native context through device

[hdelan]

Description

Since UR now supports multi device context in HIP adapter oneapi-src/unified-runtime#999 (and the same work is soon to follow in the CUDA adapter) it is no longer sensible to use the get_native method for contexts.

The mapping of native contexts to a sycl::context is now many to one, so when get_native_context is called the plugin no longer knows which native context to return. The multi device context UR PR has made the get native context return the native context of the first device in the context. See https://github.com/oneapi-src/unified-runtime/pull/999/files#diff-259fb15eb14976a3bc1939b9bb8197f51d129a111309343bb84677a655758b54R125 . But this will break for multi GPU systems.

This change instead uses the sycl::device to get the native context, since there is a one to one mapping of sycl::device to native contexts.

This change means that old versions of oneDNN will be compatible with newer plugins only if a one GPU context is being used. In order for a multi GPU system to work with the new plugin, a oneDNN version with this patch included will be necessary.

Test results to follow

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?

No. The following tests fail for the master branch as well as for my PR branch:

On Nvidia A100:

$ ninja test
...
The following tests FAILED:
4 - gpu-cnn-training-bf16-cpp (Failed)
11 - gpu-primitives-batch-normalization-cpp (Failed)
28 - test_batch_normalization_gpu (Failed)
29 - test_batch_normalization_buffer_gpu (Failed)
30 - test_binary_gpu (SEGFAULT)
31 - test_binary_buffer_gpu (SEGFAULT)
40 - test_convolution_eltwise_forward_f32_gpu (Failed)
41 - test_convolution_eltwise_forward_f32_buffer_gpu (Failed)
59 - test_iface_attr_quantization_gpu (Failed)
60 - test_iface_attr_quantization_buffer_gpu (Failed)

HIP tests all passing for ninja test on gfx1031.

[vpirogov]

@hdelan, could you please address questions and feedback on this PR?

[hdelan]

@vpirogov thanks for ping. Changes made

[vpirogov]

Thanks, @hdelan! @densamoilov is currently out, we'll have this promoted after the long weekend in US.

[densamoilov]

@hdelan, please let me know once the PR is ready for review.

[hdelan]

@densamoilov I am running tests but am seeing a lot of failures on AMD one the master branch of oneDNN 443600a30fc3427d6ad1522bceefb52100180a17 with DPC++ build e0f74157c87a2a6eb3438936eaad861e501cc418. Tests are failing with PI_ERROR_INVALID_MEM_OBJECT from piextUSMFree as well as some other things.

As a result it is hard to test the correctness of these changes. Can you recommend a particular setup where tests pass on the master branch of oneDNN, ie a certain DPC++ or release version? So that I can test just the contents of this PR

[hdelan]

@densamoilov, using release version 2024.0.2 I have the same failures with this patch as the failures on the master branch.

AMD MI210 fail list (same for master branch as well as my patch):

The following tests FAILED:
	 17 - test_concat_gpu (Failed)
	 18 - test_concat_buffer_gpu (Failed)
	 35 - test_cross_engine_reorder_buffer (Failed)
	 38 - test_eltwise_gpu (Failed)
	 39 - test_eltwise_buffer_gpu (Failed)
	 44 - test_iface_attr_quantization_gpu (Failed)
	 45 - test_iface_attr_quantization_buffer_gpu (Failed)
	 50 - test_iface_pd_gpu (Failed)
	 51 - test_iface_pd_buffer_gpu (Failed)
	 52 - test_iface_pd_iter_gpu (Failed)
	 53 - test_iface_pd_iter_buffer_gpu (Failed)
	 56 - test_iface_runtime_dims_gpu (Failed)
	 57 - test_iface_runtime_dims_buffer_gpu (Failed)
	 62 - test_inner_product_backward_data_gpu (Failed)
	 63 - test_inner_product_backward_data_buffer_gpu (Failed)
	 64 - test_inner_product_backward_weights_gpu (Failed)
	 65 - test_inner_product_backward_weights_buffer_gpu (Failed)
	 66 - test_inner_product_forward_gpu (Failed)
	 67 - test_inner_product_forward_buffer_gpu (Failed)
	 68 - test_layer_normalization_gpu (Failed)
	 69 - test_layer_normalization_buffer_gpu (Failed)
	 72 - test_matmul_gpu (Failed)
	 73 - test_matmul_buffer_gpu (Failed)
	 80 - test_prelu_gpu (Failed)
	 81 - test_prelu_buffer_gpu (Failed)
	 82 - test_primitive_cache_mt_gpu (Failed)
	 83 - test_primitive_cache_mt_buffer_gpu (Failed)
	 86 - test_reorder_gpu (Failed)
	 87 - test_reorder_buffer_gpu (Failed)
	 88 - test_resampling_gpu (Failed)
	 89 - test_resampling_buffer_gpu (Failed)
	 92 - test_shuffle_gpu (Failed)
	 93 - test_shuffle_buffer_gpu (Failed)
	 96 - test_sum_gpu (Failed)
	 97 - test_sum_buffer_gpu (Failed)

NVIDIA A100 fail list (same for master as well as my patch):

	  4 - gpu-cnn-training-bf16-cpp (Failed)
	 11 - gpu-primitives-batch-normalization-cpp (Failed)
	 28 - test_batch_normalization_gpu (Failed)
	 29 - test_batch_normalization_buffer_gpu (Failed)
	 30 - test_binary_gpu (SEGFAULT)
	 31 - test_binary_buffer_gpu (SEGFAULT)
	 40 - test_convolution_eltwise_forward_f32_gpu (Failed)
	 41 - test_convolution_eltwise_forward_f32_buffer_gpu (Failed)
	 59 - test_iface_attr_quantization_gpu (Failed)
	 60 - test_iface_attr_quantization_buffer_gpu (Failed)

So I suppose this PR is ready for review