refined align_supported_dtypes and test_ops.py skip list (#779)

1. updated align_supported_dyptes()
a) if forward does not support bfloat16, do not add it for backward
b) for _refs.xxx, if op xxx is supported by torch-xpu-ops, also align
the backward dtypes with CUDA.
2. added more ops in _xpu_computation_op_list to align the dtypes with
cuda
3. updated skip list: 
a) Added some cases in skip list because with 1 and 2 more cases are
added, while some cases are XFAIL with CUDA but passed with XPU.
b) Removed cases from skip list:
        # No this case anymore
        # "test_python_ref__refs_linspace_tensor_overload_xpu_int16",
        # "test_python_ref__refs_linspace_tensor_overload_xpu_int32",
        # "test_python_ref__refs_linspace_tensor_overload_xpu_int64",
        # "test_python_ref__refs_linspace_tensor_overload_xpu_int8",
        # "test_python_ref__refs_linspace_tensor_overload_xpu_uint8",
        # "test_python_ref__refs_linspace_xpu_int16",
        # "test_python_ref__refs_linspace_xpu_int32",
        # "test_python_ref__refs_linspace_xpu_int64",
        # "test_python_ref__refs_linspace_xpu_int8",
        # "test_python_ref__refs_linspace_xpu_uint8",
        # "test_python_ref__refs_logaddexp_xpu_complex128",
        # "test_python_ref__refs_logaddexp_xpu_complex64",
        # "test_python_ref__refs_native_layer_norm_xpu_bfloat16",
        # "test_python_ref__refs_native_layer_norm_xpu_float16",
        # "test_python_ref__refs_native_layer_norm_xpu_float32",
#
"test_python_ref__refs_nn_functional_hinge_embedding_loss_xpu_bfloat16",
#
"test_python_ref__refs_nn_functional_hinge_embedding_loss_xpu_float16",
#
"test_python_ref__refs_nn_functional_margin_ranking_loss_xpu_bfloat16",
# "test_python_ref__refs_nn_functional_margin_ranking_loss_xpu_float16",
# "test_python_ref__refs_nn_functional_triplet_margin_loss_xpu_uint8",
        # "test_python_ref__refs_square_xpu_bool",
        # "test_python_ref__refs_trunc_xpu_float64",
        
        # skipped 
# "test_python_ref_executor__refs_geometric_executor_aten_xpu_bfloat16",
# "test_python_ref_executor__refs_geometric_executor_aten_xpu_float16",
# "test_python_ref_executor__refs_geometric_executor_aten_xpu_float32",
# "test_python_ref_executor__refs_geometric_executor_aten_xpu_float64",
# "test_python_ref_executor__refs_geometric_executor_aten_xpu_int16",
# "test_python_ref_executor__refs_geometric_executor_aten_xpu_int32",
# "test_python_ref_executor__refs_geometric_executor_aten_xpu_int64",
# "test_python_ref_executor__refs_geometric_executor_aten_xpu_int8",
# "test_python_ref_executor__refs_geometric_executor_aten_xpu_uint8",
# "test_python_ref_executor__refs_linspace_executor_aten_xpu_int16",
# "test_python_ref_executor__refs_linspace_executor_aten_xpu_int32",
# "test_python_ref_executor__refs_linspace_executor_aten_xpu_int64",
# "test_python_ref_executor__refs_linspace_executor_aten_xpu_int8",
# "test_python_ref_executor__refs_linspace_executor_aten_xpu_uint8",
#
"test_python_ref_executor__refs_linspace_tensor_overload_executor_aten_xpu_int16",
#
"test_python_ref_executor__refs_linspace_tensor_overload_executor_aten_xpu_int32",
#
"test_python_ref_executor__refs_linspace_tensor_overload_executor_aten_xpu_int64",
#
"test_python_ref_executor__refs_linspace_tensor_overload_executor_aten_xpu_int8",
#
"test_python_ref_executor__refs_linspace_tensor_overload_executor_aten_xpu_uint8",
#
"test_python_ref_executor__refs_log_normal_executor_aten_xpu_bfloat16",
# "test_python_ref_executor__refs_log_normal_executor_aten_xpu_float16",
# "test_python_ref_executor__refs_log_normal_executor_aten_xpu_float32",
# "test_python_ref_executor__refs_log_normal_executor_aten_xpu_float64",

        # skipped
#
"test_python_ref_executor__refs_native_layer_norm_executor_aten_xpu_bfloat16",
#
"test_python_ref_executor__refs_native_layer_norm_executor_aten_xpu_float16",

        # need to add native_layer_norm to list

#"test_python_ref_executor__refs_native_layer_norm_executor_aten_xpu_float32",

        # skipped
#
"test_python_ref_executor__refs_nn_functional_alpha_dropout_executor_aten_xpu_bfloat16",
#
"test_python_ref_executor__refs_nn_functional_alpha_dropout_executor_aten_xpu_float16",
#
"test_python_ref_executor__refs_nn_functional_alpha_dropout_executor_aten_xpu_float32",
#
"test_python_ref_executor__refs_nn_functional_alpha_dropout_executor_aten_xpu_float64",

        # skipped
#
"test_python_ref_executor__refs_nn_functional_hinge_embedding_loss_executor_aten_xpu_bfloat16",
#
"test_python_ref_executor__refs_nn_functional_hinge_embedding_loss_executor_aten_xpu_float16",
#
"test_python_ref_executor__refs_nn_functional_margin_ranking_loss_executor_aten_xpu_bfloat16",
#
"test_python_ref_executor__refs_nn_functional_margin_ranking_loss_executor_aten_xpu_float16",
#
"test_python_ref_executor__refs_nn_functional_triplet_margin_loss_executor_aten_xpu_uint8",

 # skipped
# "test_python_ref_executor__refs_vdot_executor_aten_xpu_complex128",
# "test_python_ref_executor__refs_vdot_executor_aten_xpu_complex64",

        # skipped
#
"test_python_ref_torch_fallback__refs_linspace_tensor_overload_xpu_int16",
#
"test_python_ref_torch_fallback__refs_linspace_tensor_overload_xpu_int32",
#
"test_python_ref_torch_fallback__refs_linspace_tensor_overload_xpu_int64",
#
"test_python_ref_torch_fallback__refs_linspace_tensor_overload_xpu_int8",
#
"test_python_ref_torch_fallback__refs_linspace_tensor_overload_xpu_uint8",
        # "test_python_ref_torch_fallback__refs_linspace_xpu_int16",
        # "test_python_ref_torch_fallback__refs_linspace_xpu_int32",
        # "test_python_ref_torch_fallback__refs_linspace_xpu_int64",
        # "test_python_ref_torch_fallback__refs_linspace_xpu_int8",
# "test_python_ref_torch_fallback__refs_linspace_xpu_uint8",
# "test_python_ref_torch_fallback__refs_logaddexp_xpu_complex128",
# "test_python_ref_torch_fallback__refs_logaddexp_xpu_complex64",
# "test_python_ref_torch_fallback__refs_native_layer_norm_xpu_bfloat16",
# "test_python_ref_torch_fallback__refs_native_layer_norm_xpu_float16",

        # skipped? added native_layer_norm in op list
#"test_python_ref_torch_fallback__refs_native_layer_norm_xpu_float32",

        # skipped
#
"test_python_ref_torch_fallback__refs_nn_functional_hinge_embedding_loss_xpu_bfloat16",
#
"test_python_ref_torch_fallback__refs_nn_functional_hinge_embedding_loss_xpu_float16",
#
"test_python_ref_torch_fallback__refs_nn_functional_margin_ranking_loss_xpu_bfloat16",
#
"test_python_ref_torch_fallback__refs_nn_functional_margin_ranking_loss_xpu_float16",
# "test_python_ref_torch_fallback__refs_sinh_xpu_complex128",
#
"test_python_ref_torch_fallback__refs_special_multigammaln_mvlgamma_p_5_xpu_int32",

        # skipped?
        #"test_python_ref_torch_fallback__refs_square_xpu_bool",

        # skipped
        # "test_python_ref_torch_fallback__refs_vdot_xpu_complex128",
        # "test_python_ref_torch_fallback__refs_vdot_xpu_complex64",

        # skipped 
        #"test_compare_cpu__refs_special_zeta_xpu_float32",

        # skipped ?
        #"test_python_ref__refs_heaviside_xpu_int64",

        # skipped
        # "test_python_ref__refs_special_bessel_j0_xpu_int64",
        # "test_python_ref_errors__refs_dstack_xpu",
        # "test_python_ref_errors__refs_hstack_xpu",
        # "test_python_ref_errors__refs_linalg_cross_xpu",
        # "test_python_ref_errors__refs_vstack_xpu",
# "test_python_ref_executor__refs_mul_executor_aten_xpu_complex32",
# "test_python_ref__refs_special_multigammaln_mvlgamma_p_5_xpu_float64",
#
"test_python_ref_executor__refs_special_multigammaln_mvlgamma_p_3_executor_aten_xpu_float64",
        # "test_python_ref__refs_square_xpu_complex128",
        # "test_python_ref__refs_square_xpu_complex64",
# "test_python_ref_executor__refs_istft_executor_aten_xpu_complex128",
# "test_python_ref_executor__refs_square_executor_aten_xpu_complex128",
        # "test_python_ref_torch_fallback__refs_square_xpu_complex128",
        # "test_python_ref_torch_fallback__refs_square_xpu_complex64",


#Fixed with xpu_test_utils.py and not add bfloat16 in backward if
bloat16 is not enabled in forward
#"test_dtypes_view_as_complex_xpu", # Didn't align with CUDA, The
following dtypes did not work in backward but are listed by the OpInfo:
{torch.bfloat16}
#"test_dtypes_view_as_real_xpu", # Didn't align with CUDA, The following
dtypes did not work in backward but are listed by the OpInfo:
{torch.bfloat16}

        # no this test now
#"test_noncontiguous_samples_native_dropout_backward_xpu_int64", # The
implementation aligns with CUDA, RuntimeError: "masked_scale" not
implemented for 'Long'.
#"test_non_standard_bool_values_native_dropout_backward_xpu_bool", # The
implementation aligns with CUDA, RuntimeError: "masked_scale" not
implemented for 'Bool'.
#"test_non_standard_bool_values_scatter_reduce_amax_xpu_bool", # Align
with CUDA dtypes - "scatter_gather_base_kernel_func" not implemented for
'Bool'
#"test_non_standard_bool_values_scatter_reduce_amin_xpu_bool", # Align
with CUDA dtypes - "scatter_gather_base_kernel_func" not implemented for
'Bool'
#"test_non_standard_bool_values_scatter_reduce_prod_xpu_bool", # Align
with CUDA dtypes - "scatter_gather_base_kernel_func" not implemented for
'Bool'

        # skipped by addeding argsort and sort to list
#"test_non_standard_bool_values_argsort_xpu_bool", # The implementation
aligns with CUDA, RuntimeError: "argsort" not implemented for 'Bool'.

        # skipped
        #"test_out_triangular_solve_xpu_float32",

# add square in op list, this op is defined in aten, no need backend.
        #"test_python_ref_executor__refs_square_executor_aten_xpu_bool",

       # need to add native_layer_norm to list

#"test_python_ref_executor__refs_native_layer_norm_executor_aten_xpu_float32",

---------

Co-authored-by: Huaiyu, Zheng <[email protected]>

test/xpu/extended/skip_list_common.py

@@ -146,5 +146,21 @@
     "test_compare_cpu_nanmedian_xpu_int64",
     "test_compare_cpu_nanmedian_xpu_int8",
     "test_compare_cpu_nanmedian_xpu_uint8",
+
+    # sort algorithm is different to cpu
+    "test_compare_cpu_argsort_xpu_bfloat16",
+    "test_compare_cpu_argsort_xpu_float16",
+    "test_compare_cpu_argsort_xpu_int8",
+    "test_compare_cpu_argsort_xpu_uint8",
+
+    # AssertionError: The values for attribute 'dtype' do not match: torch.float32 != torch.bfloat16
+    # https://github.com/intel/torch-xpu-ops/issues/780
+    "test_compare_cpu_native_layer_norm_xpu_bfloat16",
+    "test_compare_cpu_native_layer_norm_xpu_float16",
+
+    # AssertionError: Tensor-likes are not close!
+    # https://github.com/intel/torch-xpu-ops/issues/781
+    "test_compare_cpu_square_xpu_complex64",
+
     ),
 }