Clean up and triage unit tests (#3546)

* clean up and triage unit tests

* add back in onnx and closeness fails

* Update unit_test.manifest

src/ngraph/runtime/plaidml/unit_test.manifest

@@ -39,88 +39,79 @@ topk_2d_min_one                         # No plans to implement TopK
 topk_int64                              # No plans to implement TopK
 topk_5d_max_partial                     # No plans to implement TopK
 topk_1d_i32_max_all                     # No plans to implement TopK
+topk_2d_max_one_with_equal_values       # No plans to implement TopK
+model_top_k                             # No plans to implement TopK
 
+# unsupported op: `Erf`
+erf
+gelu_f32
+gelu_f64
+gelu_backprop_factor_f32
+gelu_backprop_factor_f64
+backwards_gelu_f32
+backwards_gelu_f64
+model_erf
+model_erf_int32
 
-# Tests that PlaidML might be able to run at some point.
-backwards_maxpool_n2_c1_hw5_3x3_str2_max_pad1x2_2x3
-backwards_maxpool_n4c1h4w4_kh2kw2_sh1sw1
-backwards_maxpool_n2c1h5w5_kh3kw3_sh2sw2
-backwards_maxpool_n4_c1_hw4_2x2_max
-backwards_maxpool_n2_c1_hw5_3x3_str2_max
-backwards_slice
-batchnorm_fprop_bprop  # To debug
-batchnorm_fprop_bprop_2step  # To debug
-batch_norm_inference_0eps_f64
-batch_norm_inference_f64
-batch_norm_training_0eps_f64
-softmax_axis_3d_double  # To debug
-replace_slice_matrix_inplace
-max_pool_2d_1channel_1image_overpadded
-max_pool_3d
-maxpool_bprop_larger_than_cache
-generate_mask
-generate_mask2
-avg_pool_3d
-avg_pool_3d_uneven_strided_padded_include_in_computation
-quantize_clamp_int32                    # Requires fp64 inputs, which won't work on GPUs
-numeric_float_nan
-numeric_double_nan
-shape_of_scalar
-shape_of_vector
-shape_of_matrix
-shape_of_5d
-sum_stable_acc_double  # To debug: precision errors
+# unsupported ops: `BroadcastDistributed`
+broadcastdistributed
+
+# unsupported ops: 'QuantizedConvolution', 'QuantizedDot', 'EmbeddingLookup'
+model_quant_conv_linear
+model_conv_integer_no_zero_point
+model_matmul_integer_no_zero_point
+model_matmul_integer_4d_no_zero_point
+model_qlinear_matmul
+model_qlinear_matmul_3d
+model_matmul_integer
+model_matmul_integer_zero_point_zero
+model_matmul_integer_scalar
+model_matmul_integer_4d
+model_matmul_integer_4d_zero_point
+model_hardmax
+quantized_convolution
+quantized_conv_int32_output
+quantized_dot_u8u8
+quantized_dot_int32_output
 embedding_lookup_4x5_reverse
 embedding_lookup_10x1_arbitrary
 embedding_lookup_10x1_arbitrary_index_type_int
 embedding_lookup_10x1_arbitrary_index_type_int64
-floor_int32
-gather_nd_scalar_from_2d
-gather_nd_1d_from_2d
-gather_nd_scalar_from_3d
-gather_nd_1d_from_3d
-gather_nd_2d_from_3d
-gather_nd_batch_scalar_from_2d
-gather_nd_batch_1d_from_2d
-gather_nd_batch_scalar_from_3d
-gather_nd_batch_1d_from_3d
-gather_nd_batch_2d_from_3d
-gather_nd_single_indices
-gather_4d_indices_no_axis_uint8
-gather_scalar_indices_axis_1_2d_input
-gather_1d_indices_axis_2_4d_input
-gather_2d_indices_axis_1_2d_input
-gather_scalar_indices_no_axis_2d_input
-gather_1d_indices_no_axis_1d_input
-gather_2d_indices_no_axis_2d_input
-gather_3d_indices_no_axis_2d_input
-gather_4d_indices_no_axis_2d_input
-scatter_add_4d_indices
-scatter_add_3d_indices
-scatter_add_2d_indices
-scatter_add_1d_indices
-scatter_add_scalar_indices
-scatter_nd_add_batch_2d_to_3d
-scatter_nd_add_2d_to_3d
 
-# To be triaged -- bad kernels, numerical accuracy, edge conditions,
-# unimplemented functionality, &c
-cos
-erf
-sin
-tan
-not
-abc_int64
-concat_matrix_int64
-select_double
+# unsupported op: `ReverseSequence`
+model_lstm_bdir_short_input_seq
+model_lstm_mixed_seq_reverse
+model_reverse_sequence_0_batch_1
+model_reverse_sequence_1_batch_0
+
+# result mismatch
+model_dequantize_linear_scalar_zero_scale_int8
+model_softmax
+avg_pool_3d_uneven_strided_padded
+rnn_cell_activation_function
+gru_cell_bias_clip
+gru_cell_linear_before_reset
+softmax_axis_3d
+relu_2Dfprop_i32
+avg_pool_uint8
+avg_pool_int8
+one_hot_vector_many_categories
+conv_bias_1d
+conv_bias_2d
+conv_bias_3d
+conv_bias_add_2d
+normalize_across_hw_4d
+divide_python_rounding_int32
 convert_int32_bool
 convert_float32_bool
-tensor_constant_int64
-constant_equality_bool
-numeric_float_inf
-numeric_double_inf
-computation_reuse
-pad_negative_exterior_1d_check_limits
+batch_norm_inference_0eps_f64
+batch_norm_inference_f64
+batch_norm_training_0eps_f64
+batch_norm_inference_parameters_duplication
+batch_norm_fprop_b1c2h2w2
+batch_norm_fprop_b2c2h2w1
+batch_norm_fprop_b2c2d2h1w1
+batch_norm_fprop_inference_b2c2h2w1
 pad_edge_1d
 pad_edge_1d_top_neg
 pad_edge_1d_top_neg_bigger_than_tensor
@@ -136,29 +127,84 @@ pad_reflect_1d_bottom_neg_bigger_than_tensor
 pad_reflect_1d_multi_reflect
 pad_reflect_2d
 pad_reflect_2d_with_neg
-pad_negative_exterior_2d
-pad_negative_exterior_2d_all_negative
-pad_negative_exterior_4d
 pad_symmetric
-max_trivial_int8
-max_trivial_5d_int32
+
+# No double precision FP support in PlaidML
+sum_trivial_in_double
+sum_stable_acc_double
+sum_stable_simple_double
+softmax_axis_3d_double
+select_double
+numeric_double_nan
+numeric_double_inf
 max_3d_to_scalar_double
-softmax_axis_3d
-logical_and
-logical_or
-batch_norm_inference_parameters_duplication
-batch_norm_fprop_b1c2h2w2
-batch_norm_fprop_b2c2h2w1
-batch_norm_fprop_b2c2d2h1w1
-batch_norm_fprop_inference_b2c2h2w1
+argmin_trivial_in_double
+
+# unsupported op: `ShapeOf`
+shape_of_vector
+shape_of_matrix
+shape_of_5d
+
+# unsupported ops: `ScatterAdd` `ScatterNDAdd`
+scatter_add_3d_indices
+scatter_add_2d_indices
+scatter_add_1d_indices
+scatter_add_scalar_indices
+scatter_nd_add_batch_2d_to_3d
+scatter_nd_add_2d_to_3d
+
+# c++ runtime exception
+replace_slice_matrix_inplace
+quantize_clamp_int32
+pad_negative_exterior_1d_check_limits
+backwards_slice
+backwards_softmax_3d
+argmin_3D_i32
+argmin_3D_i64
+argmax_3D_i32
+argmax_3D_i64
 argmax_3D_axis_0
 argmax_3D_axis_1
 argmax_3D_axis_2
-argmin_trivial_in_double
-topk_2d_max_one_with_equal_values
-sum_trivial_in_double
-sum_stable_simple_double
-one_hot_vector_many_categories
+argmin_4D_i64
+argmax_4D_i64
+
+# PlaidML doesn't support over-padded MaxPool
+max_pool_3d
+backwards_maxpool_n2_c1_hw5_3x3_str2_max_pad1x2_2x3
+
+# c++ compilation failure
+max_trivial_int8
+max_trivial_5d_int32
+floor_int32
+any_trivial
+any_2x2x3_eliminate_dim_0
+
+# unsupported op: `GenerateMask`
+generate_mask
+generate_mask2
+
+# unsupported op: `Gather`
+gather_4d_indices_no_axis_uint8
+gather_4d_indices_no_axis_2d_input
+gather_3d_indices_no_axis_2d_input
+gather_2d_indices_no_axis_2d_input
+gather_1d_indices_no_axis_1d_input
+gather_scalar_indices_no_axis_2d_input
+gather_2d_indices_axis_1_2d_input
+gather_1d_indices_axis_2_4d_input
+gather_scalar_indices_axis_1_2d_input
+gather_nd_single_indices
+gather_nd_scalar_from_2d
+gather_nd_1d_from_2d
+gather_nd_scalar_from_3d
+gather_nd_1d_from_3d
+gather_nd_2d_from_3d
+gather_nd_batch_scalar_from_2d
+gather_nd_batch_1d_from_2d
+gather_nd_batch_scalar_from_3d
+gather_nd_batch_1d_from_3d
+gather_nd_batch_2d_from_3d
 gather_no_axis_int8
 gather_no_axis_int16
 gather_no_axis_int32
@@ -168,93 +214,24 @@ gather_no_axis_uint16
 gather_no_axis_uint32
 gather_no_axis_uint64
 gather_no_axis_bool
-elu
-elu_negative_alpha
-prelu
-hardsigmoid
-prelu_shared_slope
-prelu_negative_slope
-relu_2Dfprop_i32
-conv_bias_1d
-conv_bias_2d
-conv_bias_3d
-conv_bias_bprop_2d
-conv_bias_add_2d
-space_to_depth
-depth_to_space
-normalize_across_chw_4d
-normalize_across_chw_4d_max_bias
-normalize_across_chw_3d
-normalize_across_chw_2d
-normalize_across_hw_4d
-gemm
-fused_clamp
-mvn_mean_normalization
-mvn_mean_normalization_split_channels
-mvn_mean_variance_normalization
-mvn_mean_variance_normalization_split_channels
-grn_4d
-grn_2d_with_bias
-scale_shift_no_broadcast
-scale_shift
-shuffle_channels_simple
-shuffle_channels_negative_axis
-shuffle_channels_float
-squeeze
-squeeze_default_axes
-squared_difference
-squared_difference_broadcast
-fake_quantize
-fake_quantize_with_clip
-fake_quantize_with_clip_across_channels
-dot_0_0
-dot_2x0_0
-equal
-notequal
-greater
-greater_int64
-greatereq
-less
-lesseq
-lesseq_int32
-lesseq_bool
-broadcast_vector_rowwise_int64
-minimum_int64
-maximum_int64
-auto_bcast_binary_elementwise
-any_trivial
-any_2x2x3_eliminate_dim_0
-backwards_acos
-backwards_asin
-backwards_atan
-backwards_softmax_all
-backwards_softmax_axis
-backwards_softmax_underflow
-backwards_softmax_3d
-batch_mat_mul_forward
-dot_matrix_2x0_0x2
-max_pool_uint8
-max_pool_int8
-avg_pool_uint8
-avg_pool_int8
 
-# Need erf
-gelu_f32
-gelu_f64
-gelu_backprop_factor_f32
-gelu_backprop_factor_f64
-backwards_gelu_f32
-backwards_gelu_f64
+# unsupported op: `BatchMatMul`
+batch_mat_mul_forward
+backwards_batchmatmul_tensor2_tensor2
 
-# From onnx tests
+# onnx tests
 model_quant_conv_linear_2d
 model_quant_conv_linear_3d
 model_conv_integer
 model_conv_integer_zero_point_zero
 model_conv_integer_pads
+model_lstm_fwd_hardsigmoid_activation
+model_lstm_fwd_with_clip
+model_lstm_fwd_mixed_seq
 model_lstm_fwd_large_batch_no_clip
 model_global_lp_pool_p3
 model_argmin_no_keepdims
+model_reduce_log_sum_exp
 model_elu
 model_selu
 model_sigmoid
@@ -264,81 +241,18 @@ model_argmin_int32
 model_lp_norm_default
 model_instance_normalization
 
-# failings on plaidml_nGPU
-argmin_4D_i64
-argmax_4D_i64
-
-# dgkutnic ww24.5: these tests are to be triaged by the PlaidML team
-# ww25.2: re-scrubbed this list of tests after fixing check_inputs
-# initial debug points to some of these failing due to precision issues
-sqrt
-batch_norm_inference_0eps_f32
-batch_norm_inference_f32
-batch_norm_training_0eps_f32
-argmin_trivial
-argmax_trivial
-argmin_trivial_in_i32
-argmin_3D_i32
-argmin_3D_i64
-argmax_3D_i32
-argmax_3D_i64
-sum_large_1d_to_scalar
-sum_stable_acc
-one_hot_scalar_2_in_3
-one_hot_scalar_1_in_3
-one_hot_scalar_0_in_3
+# passing locally, fails closeness checks in CI which may be too strict
+elu
+elu_negative_alpha
 lstm_cell_no_bias_no_peepholes
 lstm_cell_bias_peepholes
 lstm_cell_bias_peepholes_clip_input_forget
 lstm_cell_activaction_functions
-divide_python_rounding_int32
-backwards_batchmatmul_tensor2_tensor2
-
-# unsupported ops: `BroadcastDistributed`
-broadcastdistributed
-
-# unsupported ops: 'QuantizedConvolution', 'QuantizedDot', 'TopK', 'Erf', 'EmbeddingLookup'
-model_quant_conv_linear
-model_conv_integer_no_zero_point
-model_matmul_integer_no_zero_point
-model_matmul_integer_4d_no_zero_point
-model_qlinear_matmul
-model_qlinear_matmul_3d
-model_matmul_integer
-model_matmul_integer_zero_point_zero
-model_matmul_integer_scalar
-model_matmul_integer_4d
-model_matmul_integer_4d_zero_point
-model_top_k
-model_erf
-model_erf_int32
-model_hardmax
-quantized_convolution
-quantized_conv_int32_output
-quantized_dot_u8u8
-quantized_dot_int32_output
-
-# unsupported op: `ReverseSequence`
-model_lstm_bdir_short_input_seq
-model_lstm_mixed_seq_reverse
-model_reverse_sequence_0_batch_1
-model_reverse_sequence_1_batch_0
-
-# node validation error: "Argument shapes are inconsistent."
-model_lstm_fwd_with_clip
-model_lstm_fwd_mixed_seq
-model_lstm_fwd_hardsigmoid_activation
-model_reduce_log_sum
-model_reduce_log_sum_exp
-model_reduce_mean
-
-# result mismatch
-model_dequantize_linear_scalar_zero_scale_int8
-model_softmax
-avg_pool_3d_uneven_strided_padded
-rnn_cell_activation_function
-gru_cell_bias_clip
-gru_cell_linear_before_reset
+dot_0_0
+dot_matrix_2x0_0x2
+dot_2x0_0
+auto_bcast_binary_elementwise
+max_pool_2d_1channel_1image_overpadded
 
 # axes input param not supported
 lrn_across_h