You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by zh...@apache.org on 2021/09/13 13:30:44 UTC
[incubator-mxnet] branch master updated (bcdfa61 -> e359bcd)
This is an automated email from the ASF dual-hosted git repository.
zhasheng pushed a change to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git.
from bcdfa61 fix macos cmake with TVM_OP ON (#20570)
add e359bcd [Master] Clang-formatter: only src/ directory (#20571)
No new revisions were added by this update.
Summary of changes:
src/api/_api_internal/_api_internal.cc | 110 +-
src/api/cached_op_api.cc | 201 +-
src/api/operator/numpy/linalg/np_det.cc | 17 +-
src/api/operator/numpy/linalg/np_eig.cc | 34 +-
src/api/operator/numpy/linalg/np_eigvals.cc | 56 +-
src/api/operator/numpy/linalg/np_gesvd.cc | 18 +-
src/api/operator/numpy/linalg/np_inv.cc | 17 +-
src/api/operator/numpy/linalg/np_lstsq.cc | 30 +-
src/api/operator/numpy/linalg/np_matrix_rank.cc | 50 +-
src/api/operator/numpy/linalg/np_norm.cc | 17 +-
src/api/operator/numpy/linalg/np_pinv.cc | 33 +-
src/api/operator/numpy/linalg/np_potrf.cc | 34 +-
src/api/operator/numpy/linalg/np_qr.cc | 18 +-
src/api/operator/numpy/linalg/np_slogdet.cc | 27 +-
src/api/operator/numpy/linalg/np_solve.cc | 17 +-
src/api/operator/numpy/linalg/np_tensorinv.cc | 34 +-
src/api/operator/numpy/linalg/np_tensorsolve.cc | 50 +-
src/api/operator/numpy/np_bincount_op.cc | 57 +-
.../numpy/np_broadcast_reduce_op_boolean.cc | 37 +-
.../operator/numpy/np_broadcast_reduce_op_index.cc | 124 +-
.../operator/numpy/np_broadcast_reduce_op_value.cc | 137 +-
src/api/operator/numpy/np_cross.cc | 23 +-
src/api/operator/numpy/np_cumsum.cc | 68 +-
src/api/operator/numpy/np_delete_op.cc | 135 +-
src/api/operator/numpy/np_diff_op.cc | 19 +-
src/api/operator/numpy/np_dot_op.cc | 16 +-
src/api/operator/numpy/np_ediff1d_op.cc | 84 +-
src/api/operator/numpy/np_einsum_op.cc | 70 +-
.../numpy/np_elemwise_broadcast_logic_op.cc | 84 +-
src/api/operator/numpy/np_elemwise_broadcast_op.cc | 189 +-
.../numpy/np_elemwise_broadcast_op_extended_sec.cc | 24 +-
.../operator/numpy/np_elemwise_unary_op_basic.cc | 87 +-
src/api/operator/numpy/np_fill_diagonal_op.cc | 68 +-
src/api/operator/numpy/np_histogram_op.cc | 82 +-
src/api/operator/numpy/np_init_op.cc | 545 ++--
src/api/operator/numpy/np_insert_op.cc | 245 +-
src/api/operator/numpy/np_interp_op.cc | 96 +-
src/api/operator/numpy/np_kron.cc | 19 +-
src/api/operator/numpy/np_matmul_op.cc | 39 +-
src/api/operator/numpy/np_matrix_op.cc | 1015 ++++----
src/api/operator/numpy/np_memory_op.cc | 26 +-
src/api/operator/numpy/np_moments_op.cc | 140 +-
src/api/operator/numpy/np_nan_to_num_op.cc | 70 +-
src/api/operator/numpy/np_nonzero_op.cc | 26 +-
src/api/operator/numpy/np_ordering_op.cc | 63 +-
src/api/operator/numpy/np_pad_op.cc | 31 +-
src/api/operator/numpy/np_percentile_op.cc | 97 +-
src/api/operator/numpy/np_polynomial_op.cc | 26 +-
src/api/operator/numpy/np_repeat_op.cc | 44 +-
src/api/operator/numpy/np_tensordot_op.cc | 40 +-
src/api/operator/numpy/np_trace_op.cc | 17 +-
src/api/operator/numpy/np_tri_op.cc | 20 +-
src/api/operator/numpy/np_tril_op.cc | 17 +-
src/api/operator/numpy/np_triu_op.cc | 15 +-
src/api/operator/numpy/np_unique_op.cc | 64 +-
src/api/operator/numpy/np_where_op.cc | 55 +-
src/api/operator/numpy/np_window_op.cc | 40 +-
src/api/operator/numpy/random/np_choice_op.cc | 104 +-
src/api/operator/numpy/random/np_exponential_op.cc | 81 +-
src/api/operator/numpy/random/np_laplace_op.cc | 116 +-
.../operator/numpy/random/np_location_scale_op.cc | 226 +-
src/api/operator/numpy/random/np_multinomial_op.cc | 78 +-
src/api/operator/numpy/random/np_pareto_op.cc | 81 +-
src/api/operator/numpy/random/np_power_op.cc | 81 +-
src/api/operator/numpy/random/np_rayleigh_op.cc | 81 +-
src/api/operator/numpy/random/np_weibull_op.cc | 81 +-
.../operator/numpy_extension/npx_activation_op.cc | 37 +-
.../operator/numpy_extension/npx_arange_like_op.cc | 95 +-
.../operator/numpy_extension/npx_batch_dot_op.cc | 74 +-
.../operator/numpy_extension/npx_batch_norm_op.cc | 111 +-
.../numpy_extension/npx_broadcast_like_op.cc | 79 +-
.../numpy_extension/npx_control_flow_op.cc | 259 +-
.../operator/numpy_extension/npx_convolution_op.cc | 238 +-
.../numpy_extension/npx_deconvolution_op.cc | 289 ++-
src/api/operator/numpy_extension/npx_dropout_op.cc | 66 +-
.../operator/numpy_extension/npx_embedding_op.cc | 62 +-
.../numpy_extension/npx_fully_connected_op.cc | 65 +-
.../operator/numpy_extension/npx_group_norm_op.cc | 71 +-
.../operator/numpy_extension/npx_layer_norm_op.cc | 95 +-
.../operator/numpy_extension/npx_leaky_relu_op.cc | 101 +-
src/api/operator/numpy_extension/npx_one_hot_op.cc | 66 +-
src/api/operator/numpy_extension/npx_pick_op.cc | 9 +-
src/api/operator/numpy_extension/npx_pooling_op.cc | 183 +-
src/api/operator/numpy_extension/npx_rnn_op.cc | 18 +-
src/api/operator/numpy_extension/npx_softmax_op.cc | 378 +--
src/api/operator/numpy_extension/npx_topk_op.cc | 13 +-
src/api/operator/op_utils.cc | 4 +-
src/api/operator/op_utils.h | 4 +-
src/api/operator/random/np_gamma_op.cc | 22 +-
src/api/operator/random/np_normal_op.cc | 130 +-
src/api/operator/random/np_randint_op.cc | 72 +-
src/api/operator/random/np_uniform_op.cc | 129 +-
src/api/operator/random/shuffle_op.cc | 52 +-
.../elemwise_binary_broadcast_op_extended.cc | 31 +-
src/api/operator/tensor/indexing_op.cc | 17 +-
src/api/operator/tensor/matrix_op.cc | 26 +-
src/api/operator/ufunc_helper.cc | 109 +-
src/api/operator/ufunc_helper.h | 4 +-
src/api/operator/utils.cc | 28 +-
src/api/operator/utils.h | 10 +-
src/base.cc | 4 +-
src/c_api/c_api.cc | 2385 ++++++++++--------
src/c_api/c_api_common.h | 35 +-
src/c_api/c_api_function.cc | 154 +-
src/c_api/c_api_ndarray.cc | 182 +-
src/c_api/c_api_profile.cc | 487 ++--
src/c_api/c_api_symbolic.cc | 947 ++++---
src/c_api/c_api_test.cc | 33 +-
src/common/cuda/rtc.cc | 155 +-
src/common/cuda/rtc.h | 10 +-
src/common/cuda/rtc/reducer-inl.h | 2 -
src/common/cuda/rtc/util-inl.h | 6 +-
src/common/cuda/rtc/vectorization-inl.h | 103 +-
src/common/cuda/utils.cc | 10 +-
src/common/cuda/utils.h | 389 ++-
src/common/exec_utils.cc | 32 +-
src/common/exec_utils.h | 192 +-
src/common/lazy_alloc_array.h | 34 +-
src/common/object_pool.h | 8 +-
src/common/random_generator.cu | 39 +-
src/common/rtc.cc | 137 +-
src/common/static_array.h | 6 +-
src/common/tensor_inspector.h | 306 ++-
src/common/utils.cc | 58 +-
src/common/utils.cu | 18 +-
src/common/utils.h | 338 +--
src/engine/engine.cc | 15 +-
src/engine/engine_impl.h | 12 +-
src/engine/naive_engine.cc | 154 +-
src/engine/openmp.cc | 14 +-
src/engine/openmp.h | 22 +-
src/engine/stream_manager.h | 25 +-
src/engine/thread_pool.h | 8 +-
src/engine/threaded_engine.cc | 239 +-
src/engine/threaded_engine.h | 145 +-
src/engine/threaded_engine_perdevice.cc | 136 +-
src/engine/threaded_engine_pooled.cc | 43 +-
src/imperative/attach_op_execs_pass.cc | 131 +-
src/imperative/attach_op_resource_pass.cc | 39 +-
src/imperative/cached_op.cc | 831 +++---
src/imperative/cached_op.h | 405 ++-
src/imperative/cached_op_threadsafe.cc | 205 +-
src/imperative/cached_op_threadsafe.h | 52 +-
src/imperative/eliminate_common_expr_pass.cc | 56 +-
src/imperative/exec_pass.h | 53 +-
src/imperative/imperative.cc | 428 ++--
src/imperative/imperative_utils.cc | 133 +-
src/imperative/imperative_utils.h | 775 +++---
src/imperative/infer_graph_attr_pass.cc | 283 ++-
src/imperative/inplace_addto_detect_pass.cc | 45 +-
src/imperative/naive_cached_op.cc | 61 +-
src/imperative/naive_cached_op.h | 49 +-
src/imperative/pointwise_fusion_pass.cc | 165 +-
src/imperative/simple_partition_pass.cc | 78 +-
src/imperative/simple_partition_pass.h | 69 +-
src/initialize.cc | 148 +-
src/initialize.h | 10 +-
src/io/batchify.cc | 406 +--
src/io/dataloader.cc | 65 +-
src/io/dataset.cc | 281 ++-
src/io/image_aug_default.cc | 534 ++--
src/io/image_augmenter.h | 19 +-
src/io/image_det_aug_default.cc | 421 ++--
src/io/image_io.cc | 286 +--
src/io/image_iter_common.h | 322 +--
src/io/image_recordio.h | 23 +-
src/io/inst_vector.h | 51 +-
src/io/iter_batchloader.h | 57 +-
src/io/iter_csv.cc | 60 +-
src/io/iter_image_det_recordio.cc | 320 +--
src/io/iter_image_recordio.cc | 208 +-
src/io/iter_image_recordio_2.cc | 718 +++---
src/io/iter_libsvm.cc | 81 +-
src/io/iter_mnist.cc | 96 +-
src/io/iter_normalize.h | 128 +-
src/io/iter_prefetcher.h | 90 +-
src/io/iter_sampler.cc | 60 +-
src/io/iter_sparse.h | 2 +-
src/io/iter_sparse_batchloader.h | 55 +-
src/io/iter_sparse_prefetcher.h | 109 +-
src/io/opencv_compatibility.h | 6 +-
src/ir/expr.cc | 21 +-
src/kvstore/comm.h | 377 +--
src/kvstore/comm_tree.h | 194 +-
src/kvstore/gpu_topology.h | 283 ++-
src/kvstore/gradient_compression-inl.h | 155 +-
src/kvstore/gradient_compression.cc | 180 +-
src/kvstore/gradient_compression.cu | 12 +-
src/kvstore/gradient_compression.h | 50 +-
src/kvstore/kvstore.cc | 8 +-
src/kvstore/kvstore_dist.h | 432 ++--
src/kvstore/kvstore_dist_server.h | 282 ++-
src/kvstore/kvstore_local.h | 191 +-
src/kvstore/kvstore_nccl.h | 285 ++-
src/kvstore/kvstore_utils.cc | 12 +-
src/kvstore/kvstore_utils.cu | 63 +-
src/kvstore/kvstore_utils.h | 5 +-
src/kvstore/p3store_dist.h | 238 +-
src/lib_api.cc | 1166 +++++----
src/libinfo.cc | 77 +-
src/ndarray/ndarray_function-inl.h | 440 ++--
src/ndarray/ndarray_function.cc | 150 +-
src/ndarray/ndarray_function.cu | 204 +-
src/ndarray/ndarray_function.h | 84 +-
src/nnvm/amp_infer_unknown.cc | 58 +-
src/nnvm/error.h | 3 +-
src/nnvm/gradient.cc | 255 +-
src/nnvm/graph_algorithm.h | 28 +-
src/nnvm/graph_editor.cc | 22 +-
src/nnvm/legacy_json_util.cc | 218 +-
src/nnvm/legacy_op_util.cc | 241 +-
src/nnvm/low_precision_pass.cc | 198 +-
src/nnvm/node_op_util.h | 80 +-
src/nnvm/plan_memory.cc | 167 +-
src/nnvm/tvm_bridge.cc | 37 +-
src/operator/all_finite-inl.h | 26 +-
src/operator/all_finite.cc | 167 +-
src/operator/all_finite.cu | 40 +-
src/operator/amp_graph_pass.cc | 8 +-
src/operator/bilinear_sampler-inl.h | 151 +-
src/operator/bilinear_sampler.cc | 214 +-
src/operator/bilinear_sampler.cu | 218 +-
src/operator/c_lapack_api.cc | 244 +-
src/operator/c_lapack_api.h | 1561 +++++++-----
src/operator/channel_op_common.h | 53 +-
src/operator/contrib/adabelief-inl.h | 395 +--
src/operator/contrib/adabelief.cc | 219 +-
src/operator/contrib/adabelief.cu | 19 +-
src/operator/contrib/adamw-inl.h | 397 +--
src/operator/contrib/adamw.cc | 217 +-
src/operator/contrib/adamw.cu | 32 +-
src/operator/contrib/adaptive_avg_pooling-inl.h | 80 +-
src/operator/contrib/adaptive_avg_pooling.cc | 171 +-
src/operator/contrib/adaptive_avg_pooling.cu | 125 +-
src/operator/contrib/allclose_op-inl.h | 109 +-
src/operator/contrib/allclose_op.cc | 54 +-
src/operator/contrib/allclose_op.cu | 30 +-
src/operator/contrib/batch_norm_relu.cc | 257 +-
src/operator/contrib/bilinear_resize-inl.h | 251 +-
src/operator/contrib/bilinear_resize.cc | 177 +-
src/operator/contrib/bilinear_resize.cu | 286 +--
src/operator/contrib/boolean_mask-inl.h | 53 +-
src/operator/contrib/boolean_mask.cc | 149 +-
src/operator/contrib/boolean_mask.cu | 147 +-
src/operator/contrib/bounding_box-common.h | 120 +-
src/operator/contrib/bounding_box-inl.h | 1055 ++++----
src/operator/contrib/bounding_box.cc | 226 +-
src/operator/contrib/bounding_box.cu | 701 +++---
src/operator/contrib/count_sketch-inl.h | 219 +-
src/operator/contrib/count_sketch.cc | 35 +-
src/operator/contrib/count_sketch.cu | 165 +-
.../contrib/deformable_psroi_pooling-inl.h | 186 +-
src/operator/contrib/deformable_psroi_pooling.cc | 711 +++---
src/operator/contrib/deformable_psroi_pooling.cu | 785 +++---
src/operator/contrib/dgl_graph-inl.h | 8 +-
src/operator/contrib/dgl_graph.cc | 807 +++---
src/operator/contrib/dgl_graph.cu | 2 +-
src/operator/contrib/dynamic_shape_ops-inl.h | 41 +-
src/operator/contrib/dynamic_shape_ops.cc | 52 +-
src/operator/contrib/erfinv-inl.h | 137 +-
src/operator/contrib/fft-inl.h | 200 +-
src/operator/contrib/fft.cc | 17 +-
src/operator/contrib/fft.cu | 8 +-
src/operator/contrib/gradient_multiplier_op.cc | 53 +-
src/operator/contrib/gradient_multiplier_op.cu | 10 +-
src/operator/contrib/hawkes_ll-inl.h | 375 ++-
src/operator/contrib/hawkes_ll.cc | 76 +-
src/operator/contrib/hawkes_ll.cu | 5 +-
src/operator/contrib/index_array-inl.h | 33 +-
src/operator/contrib/index_array.cc | 135 +-
src/operator/contrib/index_array.cu | 40 +-
src/operator/contrib/index_copy-inl.h | 8 +-
src/operator/contrib/index_copy.cc | 114 +-
src/operator/contrib/index_copy.cu | 72 +-
.../contrib/intgemm/intgemm_fully_connected_op.cc | 217 +-
src/operator/contrib/intgemm/max_absolute_op.cc | 55 +-
src/operator/contrib/intgemm/prepare_data_op.cc | 72 +-
src/operator/contrib/intgemm/prepare_weight_op.cc | 136 +-
src/operator/contrib/intgemm/take_weight_op.cc | 68 +-
src/operator/contrib/krprod.cc | 68 +-
src/operator/contrib/krprod.h | 150 +-
src/operator/contrib/mrcnn_mask_target-inl.h | 53 +-
src/operator/contrib/mrcnn_mask_target.cu | 203 +-
src/operator/contrib/multi_lamb-inl.h | 241 +-
src/operator/contrib/multi_lamb.cc | 262 +-
src/operator/contrib/multi_lamb.cu | 207 +-
src/operator/contrib/multi_lans-inl.h | 261 +-
src/operator/contrib/multi_lans.cc | 282 ++-
src/operator/contrib/multi_lans.cu | 241 +-
src/operator/contrib/multi_lars-inl.h | 71 +-
src/operator/contrib/multi_lars.cc | 38 +-
src/operator/contrib/multi_lars.cu | 3 +-
src/operator/contrib/multi_proposal-inl.h | 115 +-
src/operator/contrib/multi_proposal.cc | 299 ++-
src/operator/contrib/multi_proposal.cu | 309 +--
src/operator/contrib/multi_sum_sq-inl.h | 39 +-
src/operator/contrib/multi_sum_sq.cc | 85 +-
src/operator/contrib/multi_sum_sq.cu | 107 +-
src/operator/contrib/multibox_detection-inl.h | 148 +-
src/operator/contrib/multibox_detection.cc | 114 +-
src/operator/contrib/multibox_detection.cu | 174 +-
src/operator/contrib/multibox_prior-inl.h | 138 +-
src/operator/contrib/multibox_prior.cc | 61 +-
src/operator/contrib/multibox_prior.cu | 108 +-
src/operator/contrib/multibox_target-inl.h | 195 +-
src/operator/contrib/multibox_target.cc | 191 +-
src/operator/contrib/multibox_target.cu | 292 ++-
src/operator/contrib/nn/deformable_im2col.h | 276 +-
.../contrib/nn/modulated_deformable_im2col.h | 174 +-
src/operator/contrib/nnz.cc | 67 +-
src/operator/contrib/optimizer_op-inl.h | 146 +-
src/operator/contrib/optimizer_op.cc | 41 +-
src/operator/contrib/optimizer_op.cu | 2 +-
src/operator/contrib/preloaded_multi_sgd-inl.h | 221 +-
src/operator/contrib/preloaded_multi_sgd.cc | 326 +--
src/operator/contrib/preloaded_multi_sgd.cu | 12 +-
src/operator/contrib/proposal-inl.h | 115 +-
src/operator/contrib/proposal.cc | 235 +-
src/operator/contrib/proposal.cu | 286 +--
src/operator/contrib/psroi_pooling-inl.h | 105 +-
src/operator/contrib/psroi_pooling.cc | 293 ++-
src/operator/contrib/psroi_pooling.cu | 304 +--
src/operator/contrib/quadratic_op-inl.h | 123 +-
src/operator/contrib/quadratic_op.cc | 50 +-
src/operator/contrib/quadratic_op.cu | 6 +-
src/operator/contrib/reset_arrays-inl.h | 34 +-
src/operator/contrib/reset_arrays.cc | 65 +-
src/operator/contrib/reset_arrays.cu | 7 +-
src/operator/contrib/roi_align-inl.h | 44 +-
src/operator/contrib/roi_align.cc | 538 ++--
src/operator/contrib/roi_align.cu | 358 ++-
src/operator/contrib/rroi_align-inl.h | 26 +-
src/operator/contrib/rroi_align.cc | 291 ++-
src/operator/contrib/stes_op.cc | 61 +-
src/operator/contrib/stes_op.cu | 54 +-
src/operator/contrib/stes_op.h | 46 +-
src/operator/contrib/sync_batch_norm-inl.h | 338 +--
src/operator/contrib/sync_batch_norm.cc | 57 +-
src/operator/contrib/sync_batch_norm.cu | 7 +-
src/operator/contrib/transformer-inl.h | 236 +-
src/operator/contrib/transformer.cc | 892 +++----
src/operator/contrib/transformer.cu | 497 ++--
src/operator/contrib/tvmop/dot.cc | 52 +-
src/operator/contrib/tvmop/ufunc.cc | 104 +-
src/operator/control_flow.cc | 852 +++----
src/operator/correlation-inl.h | 258 +-
src/operator/correlation.cc | 243 +-
src/operator/correlation.cu | 1043 +++++---
src/operator/crop-inl.h | 172 +-
src/operator/crop.cc | 16 +-
src/operator/crop.cu | 4 +-
src/operator/cross_device_copy.cc | 38 +-
src/operator/cudnn_bilinear_sampler-inl.h | 108 +-
src/operator/cudnn_lrn-inl.h | 76 +-
src/operator/cudnn_spatial_transformer-inl.h | 130 +-
src/operator/custom/custom-inl.h | 50 +-
src/operator/custom/custom.cc | 342 +--
src/operator/custom/native_op-inl.h | 145 +-
src/operator/custom/native_op.cc | 12 +-
src/operator/custom/native_op.cu | 4 +-
src/operator/custom/ndarray_op-inl.h | 96 +-
src/operator/custom/ndarray_op.cc | 71 +-
src/operator/deformable_convolution-inl.h | 384 +--
src/operator/deformable_convolution.cc | 35 +-
src/operator/deformable_convolution.cu | 24 +-
src/operator/elemwise_op_common.h | 166 +-
src/operator/fusion/fused_op-inl.h | 355 ++-
src/operator/fusion/fused_op.cc | 195 +-
src/operator/fusion/fused_op.cu | 452 ++--
src/operator/fusion/fused_op.h | 119 +-
src/operator/grid_generator-inl.h | 246 +-
src/operator/grid_generator.cc | 17 +-
src/operator/grid_generator.cu | 10 +-
src/operator/identity_attach_KL_sparse_reg-inl.h | 116 +-
src/operator/identity_attach_KL_sparse_reg.cc | 31 +-
src/operator/identity_attach_KL_sparse_reg.cu | 6 +-
src/operator/image/crop-inl.h | 466 ++--
src/operator/image/crop.cc | 158 +-
src/operator/image/crop.cu | 49 +-
src/operator/image/image_random-inl.h | 931 ++++---
src/operator/image/image_random.cc | 224 +-
src/operator/image/image_random.cu | 354 ++-
src/operator/image/image_utils.h | 44 +-
src/operator/image/resize-inl.h | 161 +-
src/operator/image/resize.cc | 58 +-
src/operator/image/resize.cu | 32 +-
src/operator/instance_norm-inl.h | 148 +-
src/operator/instance_norm.cc | 83 +-
src/operator/instance_norm.cu | 7 +-
src/operator/l2_normalization-inl.h | 256 +-
src/operator/l2_normalization.cc | 83 +-
src/operator/l2_normalization.cu | 8 +-
src/operator/leaky_relu-inl.h | 293 ++-
src/operator/leaky_relu.cc | 195 +-
src/operator/leaky_relu.cu | 8 +-
src/operator/linalg.h | 191 +-
src/operator/linalg_impl.h | 2637 +++++++++++---------
src/operator/loss_binary_op-inl.h | 51 +-
src/operator/loss_binary_op.cc | 54 +-
src/operator/loss_binary_op.cu | 6 +-
src/operator/make_loss-inl.h | 130 +-
src/operator/make_loss.cc | 23 +-
src/operator/make_loss.cu | 13 +-
src/operator/math_functions-inl.h | 59 +-
src/operator/mkl_functions-inl.h | 74 +-
.../modulated_deformable_convolution-inl.h | 494 ++--
src/operator/modulated_deformable_convolution.cc | 48 +-
src/operator/modulated_deformable_convolution.cu | 24 +-
src/operator/mshadow_op.h | 1013 ++++----
src/operator/mxnet_op.h | 1176 ++++-----
src/operator/nn/activation-inl.h | 87 +-
src/operator/nn/activation.cc | 124 +-
src/operator/nn/activation.cu | 67 +-
src/operator/nn/batch_norm.cu | 803 +++---
src/operator/nn/concat-inl.h | 178 +-
src/operator/nn/concat.cc | 256 +-
src/operator/nn/concat.cu | 17 +-
src/operator/nn/convolution-inl.h | 328 +--
src/operator/nn/convolution.cc | 310 +--
src/operator/nn/convolution.cu | 96 +-
src/operator/nn/ctc_loss-inl.h | 250 +-
src/operator/nn/ctc_loss.cc | 99 +-
src/operator/nn/ctc_loss.cu | 38 +-
src/operator/nn/cudnn/cudnn_activation-inl.h | 71 +-
src/operator/nn/cudnn/cudnn_algoreg-inl.h | 100 +-
src/operator/nn/cudnn/cudnn_algoreg.cc | 10 +-
src/operator/nn/cudnn/cudnn_batch_norm-inl.h | 170 +-
src/operator/nn/cudnn/cudnn_batch_norm.cc | 118 +-
src/operator/nn/cudnn/cudnn_convolution-inl.h | 580 ++---
src/operator/nn/cudnn/cudnn_deconvolution-inl.h | 429 ++--
src/operator/nn/cudnn/cudnn_pooling-inl.h | 189 +-
.../nn/cudnn/cudnn_softmax_activation-inl.h | 66 +-
src/operator/nn/deconvolution-inl.h | 399 ++-
src/operator/nn/deconvolution.cc | 348 +--
src/operator/nn/deconvolution.cu | 84 +-
src/operator/nn/depthwise_convolution-inl.h | 201 +-
src/operator/nn/dropout-inl.h | 377 ++-
src/operator/nn/dropout.cc | 186 +-
src/operator/nn/dropout.cu | 9 +-
src/operator/nn/fully_connected-inl.h | 408 +--
src/operator/nn/fully_connected.cc | 221 +-
src/operator/nn/fully_connected.cu | 25 +-
src/operator/nn/group_norm-inl.h | 367 +--
src/operator/nn/group_norm.cc | 117 +-
src/operator/nn/group_norm.cu | 7 +-
src/operator/nn/im2col-inl.h | 189 +-
src/operator/nn/im2col.cc | 329 +--
src/operator/nn/im2col.cu | 14 +-
src/operator/nn/im2col.h | 153 +-
src/operator/nn/layer_norm-inl.h | 232 +-
src/operator/nn/layer_norm.cc | 339 +--
src/operator/nn/layer_norm.cu | 543 ++--
src/operator/nn/log_softmax.cc | 210 +-
src/operator/nn/log_softmax.cu | 12 +-
src/operator/nn/lrn-inl.h | 115 +-
src/operator/nn/lrn.cc | 122 +-
src/operator/nn/lrn.cu | 10 +-
src/operator/nn/mkldnn/mkldnn_base-inl.h | 2 +-
src/operator/nn/mkldnn/mkldnn_layer_norm.cc | 2 +-
src/operator/nn/mkldnn/mkldnn_ops-inl.h | 18 +-
src/operator/nn/moments-inl.h | 100 +-
src/operator/nn/moments.cc | 58 +-
src/operator/nn/moments.cu | 8 +-
src/operator/nn/pool.h | 1171 +++++----
src/operator/nn/pool_utils.h | 36 +-
src/operator/nn/pooling-inl.h | 389 +--
src/operator/nn/pooling.cc | 392 ++-
src/operator/nn/pooling.cu | 47 +-
src/operator/nn/sequence_mask-inl.h | 23 +-
src/operator/nn/softmax-inl.h | 1121 +++++----
src/operator/nn/softmax.cc | 253 +-
src/operator/nn/softmax.cu | 222 +-
src/operator/nn/softmax_activation-inl.h | 79 +-
src/operator/nn/softmax_activation.cc | 46 +-
src/operator/nn/softmax_activation.cu | 13 +-
src/operator/nn/softmin.cc | 56 +-
src/operator/nn/softmin.cu | 6 +-
src/operator/nn/upsampling-inl.h | 159 +-
src/operator/nn/upsampling.cc | 178 +-
src/operator/nn/upsampling.cu | 7 +-
src/operator/npx_control_flow.cc | 854 ++++---
src/operator/npx_control_flow.h | 80 +-
.../numpy/linalg/broadcast_reduce_customized-inl.h | 193 +-
.../numpy/linalg/broadcast_reduce_op_customized.h | 89 +-
src/operator/numpy/linalg/np_eig-inl.h | 192 +-
src/operator/numpy/linalg/np_eig.cc | 87 +-
src/operator/numpy/linalg/np_eig.cu | 6 +-
src/operator/numpy/linalg/np_eigvals-inl.h | 298 ++-
src/operator/numpy/linalg/np_eigvals.cc | 67 +-
src/operator/numpy/linalg/np_eigvals.cu | 6 +-
src/operator/numpy/linalg/np_gesvd-inl.h | 141 +-
src/operator/numpy/linalg/np_gesvd.cc | 85 +-
src/operator/numpy/linalg/np_gesvd.cu | 5 +-
src/operator/numpy/linalg/np_lstsq-inl.h | 528 ++--
src/operator/numpy/linalg/np_lstsq.cc | 55 +-
src/operator/numpy/linalg/np_lstsq.cu | 3 +-
src/operator/numpy/linalg/np_matrix_rank-inl.h | 258 +-
src/operator/numpy/linalg/np_matrix_rank.cc | 113 +-
src/operator/numpy/linalg/np_matrix_rank.cu | 5 +-
src/operator/numpy/linalg/np_norm-inl.h | 515 ++--
src/operator/numpy/linalg/np_norm.cc | 76 +-
src/operator/numpy/linalg/np_norm_backward.cc | 18 +-
src/operator/numpy/linalg/np_norm_backward.cu | 44 +-
src/operator/numpy/linalg/np_norm_forward.cc | 30 +-
src/operator/numpy/linalg/np_norm_forward.cu | 45 +-
src/operator/numpy/linalg/np_pinv-inl.h | 606 +++--
src/operator/numpy/linalg/np_pinv.cc | 110 +-
src/operator/numpy/linalg/np_pinv.cu | 5 +-
src/operator/numpy/linalg/np_potrf.cc | 35 +-
src/operator/numpy/linalg/np_potrf.cu | 2 +-
src/operator/numpy/linalg/np_qr-inl.h | 576 +++--
src/operator/numpy/linalg/np_qr.cc | 49 +-
src/operator/numpy/linalg/np_qr.cu | 6 +-
src/operator/numpy/linalg/np_solve-inl.h | 514 ++--
src/operator/numpy/linalg/np_solve.cc | 74 +-
src/operator/numpy/linalg/np_solve.cu | 4 +-
src/operator/numpy/linalg/np_tensorinv-inl.h | 64 +-
src/operator/numpy/linalg/np_tensorinv.cc | 80 +-
src/operator/numpy/linalg/np_tensorinv.cu | 5 +-
src/operator/numpy/linalg/np_tensorsolve-inl.h | 303 ++-
src/operator/numpy/linalg/np_tensorsolve.cc | 82 +-
src/operator/numpy/linalg/np_tensorsolve.cu | 4 +-
src/operator/numpy/np_bincount_op-inl.h | 81 +-
src/operator/numpy/np_bincount_op.cc | 120 +-
src/operator/numpy/np_bincount_op.cu | 130 +-
src/operator/numpy/np_boolean_mask_assign.cc | 219 +-
src/operator/numpy/np_boolean_mask_assign.cu | 145 +-
src/operator/numpy/np_broadcast_reduce_op.cc | 41 +-
src/operator/numpy/np_broadcast_reduce_op.h | 705 +++---
.../numpy/np_broadcast_reduce_op_boolean.cc | 83 +-
.../numpy/np_broadcast_reduce_op_boolean.cu | 12 +-
src/operator/numpy/np_broadcast_reduce_op_index.cc | 63 +-
src/operator/numpy/np_broadcast_reduce_op_index.cu | 4 +-
src/operator/numpy/np_broadcast_reduce_op_value.cc | 352 +--
src/operator/numpy/np_broadcast_reduce_op_value.cu | 55 +-
src/operator/numpy/np_constraint_check.cc | 40 +-
src/operator/numpy/np_constraint_check.cu | 13 +-
src/operator/numpy/np_constraint_check.h | 22 +-
src/operator/numpy/np_cross-inl.h | 1488 ++++++-----
src/operator/numpy/np_cross.cc | 76 +-
src/operator/numpy/np_cross.cu | 6 +-
src/operator/numpy/np_cumsum-inl.h | 86 +-
src/operator/numpy/np_cumsum.cc | 61 +-
src/operator/numpy/np_cumsum.cu | 6 +-
src/operator/numpy/np_delete_op-inl.h | 227 +-
src/operator/numpy/np_delete_op.cc | 69 +-
src/operator/numpy/np_delete_op.cu | 5 +-
src/operator/numpy/np_diff-inl.h | 97 +-
src/operator/numpy/np_diff.cc | 68 +-
src/operator/numpy/np_diff.cu | 6 +-
src/operator/numpy/np_dot-inl.h | 42 +-
src/operator/numpy/np_dot.cc | 62 +-
src/operator/numpy/np_dot.cu | 6 +-
src/operator/numpy/np_ediff1d_op-inl.h | 106 +-
src/operator/numpy/np_ediff1d_op.cc | 151 +-
src/operator/numpy/np_ediff1d_op.cu | 6 +-
src/operator/numpy/np_einsum_op-inl.h | 619 ++---
src/operator/numpy/np_einsum_op.cc | 202 +-
src/operator/numpy/np_einsum_op.cu | 521 ++--
src/operator/numpy/np_einsum_path_op-inl.h | 280 +--
.../numpy/np_elemwise_broadcast_logic_op.cc | 260 +-
.../numpy/np_elemwise_broadcast_logic_op.cu | 13 +-
src/operator/numpy/np_elemwise_broadcast_op.cc | 286 +--
src/operator/numpy/np_elemwise_broadcast_op.cu | 39 +-
src/operator/numpy/np_elemwise_broadcast_op.h | 243 +-
.../numpy/np_elemwise_broadcast_op_extended.cc | 545 ++--
.../numpy/np_elemwise_broadcast_op_extended.cu | 68 +-
.../numpy/np_elemwise_broadcast_op_extended_sec.cc | 161 +-
.../numpy/np_elemwise_broadcast_op_extended_sec.cu | 36 +-
src/operator/numpy/np_elemwise_unary_op_basic.cc | 543 ++--
src/operator/numpy/np_elemwise_unary_op_basic.cu | 60 +-
src/operator/numpy/np_fill_diagonal_op-inl.h | 72 +-
src/operator/numpy/np_fill_diagonal_op.cc | 44 +-
src/operator/numpy/np_fill_diagonal_op.cu | 2 +-
src/operator/numpy/np_indexing_op.cc | 458 ++--
src/operator/numpy/np_indexing_op.cu | 522 ++--
src/operator/numpy/np_indexing_op.h | 79 +-
src/operator/numpy/np_init_op.cc | 266 +-
src/operator/numpy/np_init_op.cu | 39 +-
src/operator/numpy/np_init_op.h | 202 +-
src/operator/numpy/np_insert_op-inl.h | 233 +-
src/operator/numpy/np_insert_op_scalar-inl.h | 64 +-
src/operator/numpy/np_insert_op_scalar.cc | 89 +-
src/operator/numpy/np_insert_op_scalar.cu | 4 +-
src/operator/numpy/np_insert_op_slice-inl.h | 111 +-
src/operator/numpy/np_insert_op_slice.cc | 83 +-
src/operator/numpy/np_insert_op_slice.cu | 4 +-
src/operator/numpy/np_insert_op_tensor-inl.h | 134 +-
src/operator/numpy/np_insert_op_tensor.cc | 104 +-
src/operator/numpy/np_insert_op_tensor.cu | 4 +-
src/operator/numpy/np_interp_op-inl.h | 199 +-
src/operator/numpy/np_interp_op.cc | 73 +-
src/operator/numpy/np_interp_op.cu | 4 +-
src/operator/numpy/np_kron-inl.h | 165 +-
src/operator/numpy/np_kron.cc | 52 +-
src/operator/numpy/np_kron.cu | 6 +-
src/operator/numpy/np_matmul_op-inl.h | 343 +--
src/operator/numpy/np_matmul_op.cc | 86 +-
src/operator/numpy/np_matmul_op.cu | 6 +-
src/operator/numpy/np_matrix_op-inl.h | 931 +++----
src/operator/numpy/np_matrix_op.cc | 1409 ++++++-----
src/operator/numpy/np_matrix_op.cu | 122 +-
src/operator/numpy/np_memory_op.cc | 30 +-
src/operator/numpy/np_memory_op.cu | 2 +-
src/operator/numpy/np_memory_op.h | 12 +-
src/operator/numpy/np_moments_op.cc | 211 +-
src/operator/numpy/np_nonzero_op-inl.h | 6 +-
src/operator/numpy/np_nonzero_op.cc | 64 +-
src/operator/numpy/np_nonzero_op.cu | 80 +-
src/operator/numpy/np_pad_op-inl.h | 612 ++---
src/operator/numpy/np_pad_op.cc | 56 +-
src/operator/numpy/np_pad_op.cu | 6 +-
src/operator/numpy/np_percentile_op-inl.h | 199 +-
src/operator/numpy/np_percentile_op.cc | 78 +-
src/operator/numpy/np_percentile_op.cu | 25 +-
src/operator/numpy/np_polynomial_op-inl.h | 32 +-
src/operator/numpy/np_polynomial_op.cc | 113 +-
src/operator/numpy/np_polynomial_op.cu | 66 +-
src/operator/numpy/np_repeat_op-inl.h | 122 +-
src/operator/numpy/np_repeat_op.cc | 34 +-
src/operator/numpy/np_repeat_op.cu | 45 +-
src/operator/numpy/np_tensordot_op-inl.h | 339 +--
src/operator/numpy/np_tensordot_op.cc | 152 +-
src/operator/numpy/np_tensordot_op.cu | 9 +-
src/operator/numpy/np_trace_op-inl.h | 93 +-
src/operator/numpy/np_trace_op.cc | 44 +-
src/operator/numpy/np_trace_op.cu | 5 +-
src/operator/numpy/np_tri_op-inl.h | 59 +-
src/operator/numpy/np_tri_op.cc | 34 +-
src/operator/numpy/np_tri_op.cu | 3 +-
src/operator/numpy/np_tril_op-inl.h | 113 +-
src/operator/numpy/np_tril_op.cc | 53 +-
src/operator/numpy/np_tril_op.cu | 6 +-
src/operator/numpy/np_triu_op-inl.h | 109 +-
src/operator/numpy/np_triu_op.cc | 53 +-
src/operator/numpy/np_triu_op.cu | 6 +-
src/operator/numpy/np_true_divide-inl.h | 152 +-
src/operator/numpy/np_true_divide.cc | 130 +-
src/operator/numpy/np_true_divide.cu | 8 +-
src/operator/numpy/np_unique_op.cc | 234 +-
src/operator/numpy/np_unique_op.cu | 236 +-
src/operator/numpy/np_unique_op.h | 28 +-
src/operator/numpy/np_where_op-inl.h | 320 ++-
src/operator/numpy/np_where_op.cc | 356 +--
src/operator/numpy/np_where_op.cu | 16 +-
src/operator/numpy/np_window_op.cc | 57 +-
src/operator/numpy/np_window_op.cu | 9 +-
src/operator/numpy/np_window_op.h | 67 +-
src/operator/numpy/random/dist_common.cc | 4 +-
src/operator/numpy/random/dist_common.cu | 10 +-
src/operator/numpy/random/dist_common.h | 196 +-
src/operator/numpy/random/np_bernoulli_op.cc | 65 +-
src/operator/numpy/random/np_bernoulli_op.cu | 3 +-
src/operator/numpy/random/np_bernoulli_op.h | 93 +-
src/operator/numpy/random/np_choice_op.cc | 82 +-
src/operator/numpy/random/np_choice_op.cu | 6 +-
src/operator/numpy/random/np_choice_op.h | 150 +-
src/operator/numpy/random/np_exponential_op.cc | 130 +-
src/operator/numpy/random/np_exponential_op.cu | 4 +-
src/operator/numpy/random/np_exponential_op.h | 123 +-
src/operator/numpy/random/np_gamma_op.cc | 129 +-
src/operator/numpy/random/np_gamma_op.cu | 6 +-
src/operator/numpy/random/np_gamma_op.h | 326 +--
src/operator/numpy/random/np_laplace_op.cc | 73 +-
src/operator/numpy/random/np_laplace_op.cu | 3 +-
src/operator/numpy/random/np_laplace_op.h | 128 +-
src/operator/numpy/random/np_location_scale_op.cc | 20 +-
src/operator/numpy/random/np_location_scale_op.cu | 20 +-
src/operator/numpy/random/np_location_scale_op.h | 360 +--
src/operator/numpy/random/np_multinomial_op.cc | 38 +-
src/operator/numpy/random/np_multinomial_op.cu | 11 +-
src/operator/numpy/random/np_multinomial_op.h | 128 +-
src/operator/numpy/random/np_normal_op.cc | 211 +-
src/operator/numpy/random/np_normal_op.cu | 8 +-
src/operator/numpy/random/np_normal_op.h | 160 +-
src/operator/numpy/random/np_pareto_op.cc | 130 +-
src/operator/numpy/random/np_pareto_op.cu | 5 +-
src/operator/numpy/random/np_pareto_op.h | 138 +-
src/operator/numpy/random/np_power_op.cc | 74 +-
src/operator/numpy/random/np_power_op.cu | 3 +-
src/operator/numpy/random/np_power_op.h | 85 +-
src/operator/numpy/random/np_rayleigh_op.cc | 130 +-
src/operator/numpy/random/np_rayleigh_op.cu | 5 +-
src/operator/numpy/random/np_rayleigh_op.h | 104 +-
src/operator/numpy/random/np_uniform_op.cc | 146 +-
src/operator/numpy/random/np_uniform_op.cu | 6 +-
src/operator/numpy/random/np_uniform_op.h | 132 +-
src/operator/numpy/random/np_weibull_op.cc | 130 +-
src/operator/numpy/random/np_weibull_op.cu | 5 +-
src/operator/numpy/random/np_weibull_op.h | 130 +-
src/operator/operator.cc | 4 +-
src/operator/operator_common.h | 289 ++-
src/operator/operator_tune-inl.h | 169 +-
src/operator/operator_tune.cc | 640 ++---
src/operator/operator_tune.h | 58 +-
src/operator/operator_util.cc | 542 ++--
src/operator/optimizer_op-inl.h | 2622 ++++++++++---------
src/operator/optimizer_op.cc | 1065 ++++----
src/operator/optimizer_op.cu | 257 +-
src/operator/pad-inl.h | 182 +-
src/operator/pad.cc | 293 +--
src/operator/pad.cu | 392 ++-
src/operator/quantization/calibrate-inl.h | 5 +-
src/operator/quantization/calibrate.cc | 110 +-
src/operator/quantization/dequantize-inl.h | 82 +-
src/operator/quantization/dequantize.cc | 64 +-
src/operator/quantization/dequantize.cu | 2 +-
src/operator/quantization/quantization_utils.h | 108 +-
src/operator/quantization/quantize-inl.h | 98 +-
src/operator/quantization/quantize.cc | 57 +-
src/operator/quantization/quantize.cu | 3 +-
src/operator/quantization/quantize_graph_pass.cc | 181 +-
src/operator/quantization/quantize_v2-inl.h | 211 +-
src/operator/quantization/quantize_v2.cc | 89 +-
src/operator/quantization/quantize_v2.cu | 2 +-
src/operator/quantization/quantized_activation.cc | 94 +-
src/operator/quantization/quantized_batch_norm.cc | 103 +-
src/operator/quantization/quantized_concat.cc | 143 +-
src/operator/quantization/quantized_conv.cc | 170 +-
src/operator/quantization/quantized_conv.cu | 178 +-
.../quantization/quantized_elemwise_add-inl.h | 22 +-
.../quantization/quantized_elemwise_add.cc | 77 +-
.../quantization/quantized_elemwise_mul-inl.h | 31 +-
.../quantization/quantized_elemwise_mul.cc | 139 +-
src/operator/quantization/quantized_flatten-inl.h | 56 +-
src/operator/quantization/quantized_flatten.cc | 80 +-
src/operator/quantization/quantized_flatten.cu | 2 +-
.../quantization/quantized_fully_connected.cc | 296 +--
.../quantization/quantized_fully_connected.cu | 85 +-
src/operator/quantization/quantized_indexing_op.cc | 171 +-
src/operator/quantization/quantized_pooling.cc | 194 +-
src/operator/quantization/quantized_pooling.cu | 42 +-
src/operator/quantization/requantize-inl.h | 198 +-
src/operator/quantization/requantize.cc | 85 +-
src/operator/quantization/requantize.cu | 3 +-
src/operator/random/multisample_op.cc | 141 +-
src/operator/random/multisample_op.cu | 17 +-
src/operator/random/multisample_op.h | 101 +-
src/operator/random/pdf_op.cc | 181 +-
src/operator/random/pdf_op.cu | 17 +-
src/operator/random/pdf_op.h | 668 +++--
src/operator/random/sample_multinomial_op.cc | 86 +-
src/operator/random/sample_multinomial_op.cu | 20 +-
src/operator/random/sample_multinomial_op.h | 132 +-
src/operator/random/sample_op.cc | 121 +-
src/operator/random/sample_op.cu | 8 +-
src/operator/random/sample_op.h | 540 ++--
src/operator/random/sampler.h | 327 +--
src/operator/random/shuffle_op.cc | 93 +-
src/operator/random/shuffle_op.cu | 50 +-
src/operator/random/unique_sample_op.cc | 25 +-
src/operator/random/unique_sample_op.h | 88 +-
src/operator/regression_output-inl.h | 173 +-
src/operator/regression_output.cc | 101 +-
src/operator/regression_output.cu | 23 +-
src/operator/rnn-inl.h | 778 +++---
src/operator/rnn.cc | 249 +-
src/operator/rnn.cu | 8 +-
src/operator/rnn_impl.h | 1365 +++++-----
src/operator/roi_pooling-inl.h | 120 +-
src/operator/roi_pooling.cc | 158 +-
src/operator/roi_pooling.cu | 174 +-
src/operator/sequence_last-inl.h | 211 +-
src/operator/sequence_last.cc | 43 +-
src/operator/sequence_last.cu | 13 +-
src/operator/sequence_mask-inl.h | 181 +-
src/operator/sequence_mask.cc | 70 +-
src/operator/sequence_mask.cu | 64 +-
src/operator/sequence_op_common.h | 29 +-
src/operator/sequence_reverse-inl.h | 187 +-
src/operator/sequence_reverse.cc | 28 +-
src/operator/sequence_reverse.cu | 13 +-
src/operator/slice_channel-inl.h | 142 +-
src/operator/slice_channel.cc | 20 +-
src/operator/slice_channel.cu | 9 +-
src/operator/softmax_output-inl.h | 263 +-
src/operator/softmax_output.cc | 151 +-
src/operator/softmax_output.cu | 7 +-
src/operator/spatial_transformer-inl.h | 209 +-
src/operator/spatial_transformer.cc | 225 +-
src/operator/spatial_transformer.cu | 198 +-
src/operator/special_functions-inl.h | 121 +-
src/operator/subgraph/build_subgraph.cc | 199 +-
src/operator/subgraph/common.h | 62 +-
src/operator/subgraph/default_subgraph_property.cc | 27 +-
.../subgraph/default_subgraph_property_v2.cc | 36 +-
.../partitioner/custom_subgraph_property.h | 238 +-
.../subgraph/static_shape_subgraph_property.cc | 61 +-
src/operator/subgraph/subgraph_property.h | 64 +-
src/operator/subgraph/tensorrt/nnvm_to_onnx-inl.h | 260 +-
src/operator/subgraph/tensorrt/nnvm_to_onnx.cc | 292 ++-
src/operator/subgraph/tensorrt/onnx_to_tensorrt.cc | 76 +-
src/operator/subgraph/tensorrt/onnx_to_tensorrt.h | 44 +-
src/operator/subgraph/tensorrt/tensorrt-inl.h | 101 +-
src/operator/subgraph/tensorrt/tensorrt.cc | 124 +-
src/operator/subgraph/tensorrt/tensorrt.cu | 33 +-
src/operator/subgraph_op_common.cc | 100 +-
src/operator/subgraph_op_common.h | 73 +-
src/operator/svm_output-inl.h | 112 +-
src/operator/svm_output.cc | 86 +-
src/operator/svm_output.cu | 85 +-
src/operator/swapaxis-inl.h | 120 +-
src/operator/swapaxis.cc | 22 +-
src/operator/swapaxis.cu | 13 +-
src/operator/tensor/amp_cast.cc | 285 ++-
src/operator/tensor/amp_cast.cu | 11 +-
src/operator/tensor/amp_cast.h | 54 +-
src/operator/tensor/broadcast_reduce-inl.h | 669 +++--
.../tensor/broadcast_reduce_minmax_value.cc | 44 +-
.../tensor/broadcast_reduce_minmax_value.cu | 18 +-
src/operator/tensor/broadcast_reduce_norm_value.cc | 63 +-
src/operator/tensor/broadcast_reduce_norm_value.cu | 15 +-
src/operator/tensor/broadcast_reduce_op.cc | 51 +-
src/operator/tensor/broadcast_reduce_op.h | 1192 +++++----
src/operator/tensor/broadcast_reduce_op_index.cc | 101 +-
src/operator/tensor/broadcast_reduce_op_index.cu | 17 +-
src/operator/tensor/broadcast_reduce_op_value.cc | 121 +-
src/operator/tensor/broadcast_reduce_op_value.cu | 13 +-
src/operator/tensor/broadcast_reduce_prod_value.cc | 41 +-
src/operator/tensor/broadcast_reduce_prod_value.cu | 17 +-
src/operator/tensor/broadcast_reduce_sum_value.cc | 71 +-
src/operator/tensor/broadcast_reduce_sum_value.cu | 27 +-
src/operator/tensor/cast_storage-inl.h | 200 +-
src/operator/tensor/cast_storage.cc | 34 +-
src/operator/tensor/cast_storage.cu | 4 +-
src/operator/tensor/control_flow_op.cc | 98 +-
src/operator/tensor/control_flow_op.cu | 8 +-
src/operator/tensor/control_flow_op.h | 319 ++-
src/operator/tensor/diag_op-inl.h | 183 +-
src/operator/tensor/diag_op.cc | 50 +-
src/operator/tensor/diag_op.cu | 16 +-
src/operator/tensor/dot-inl.h | 877 ++++---
src/operator/tensor/dot.cc | 245 +-
src/operator/tensor/dot.cu | 11 +-
.../tensor/elemwise_binary_broadcast_op.cc | 177 +-
src/operator/tensor/elemwise_binary_broadcast_op.h | 475 ++--
.../tensor/elemwise_binary_broadcast_op_basic.cc | 189 +-
.../tensor/elemwise_binary_broadcast_op_basic.cu | 32 +-
.../elemwise_binary_broadcast_op_extended.cc | 136 +-
.../elemwise_binary_broadcast_op_extended.cu | 19 +-
.../tensor/elemwise_binary_broadcast_op_logic.cc | 63 +-
.../tensor/elemwise_binary_broadcast_op_logic.cu | 18 +-
src/operator/tensor/elemwise_binary_op-inl.h | 398 +--
src/operator/tensor/elemwise_binary_op.cc | 178 +-
src/operator/tensor/elemwise_binary_op.h | 743 +++---
src/operator/tensor/elemwise_binary_op_basic.cc | 248 +-
src/operator/tensor/elemwise_binary_op_basic.cu | 202 +-
src/operator/tensor/elemwise_binary_op_extended.cc | 93 +-
src/operator/tensor/elemwise_binary_op_extended.cu | 22 +-
src/operator/tensor/elemwise_binary_op_logic.cc | 54 +-
src/operator/tensor/elemwise_binary_op_logic.cu | 21 +-
src/operator/tensor/elemwise_binary_scalar_op.cc | 81 +-
src/operator/tensor/elemwise_binary_scalar_op.h | 345 ++-
.../tensor/elemwise_binary_scalar_op_basic.cc | 181 +-
.../tensor/elemwise_binary_scalar_op_basic.cu | 39 +-
.../tensor/elemwise_binary_scalar_op_extended.cc | 108 +-
.../tensor/elemwise_binary_scalar_op_extended.cu | 26 +-
.../tensor/elemwise_binary_scalar_op_logic.cc | 81 +-
.../tensor/elemwise_binary_scalar_op_logic.cu | 30 +-
src/operator/tensor/elemwise_sum.cc | 136 +-
src/operator/tensor/elemwise_sum.cu | 52 +-
src/operator/tensor/elemwise_sum.h | 38 +-
src/operator/tensor/elemwise_unary_op.cc | 44 +-
src/operator/tensor/elemwise_unary_op.h | 542 ++--
src/operator/tensor/elemwise_unary_op_basic.cc | 953 +++----
src/operator/tensor/elemwise_unary_op_basic.cu | 149 +-
src/operator/tensor/elemwise_unary_op_logexp.cc | 289 +--
src/operator/tensor/elemwise_unary_op_logexp.cu | 30 +-
src/operator/tensor/elemwise_unary_op_pow.cc | 448 ++--
src/operator/tensor/elemwise_unary_op_pow.cu | 36 +-
src/operator/tensor/elemwise_unary_op_trig.cc | 784 +++---
src/operator/tensor/elemwise_unary_op_trig.cu | 80 +-
src/operator/tensor/histogram-inl.h | 59 +-
src/operator/tensor/histogram.cc | 114 +-
src/operator/tensor/histogram.cu | 60 +-
src/operator/tensor/index_add-inl.h | 119 +-
src/operator/tensor/index_add_backward.cc | 61 +-
src/operator/tensor/index_add_backward.cu | 50 +-
src/operator/tensor/index_add_forward.cc | 147 +-
src/operator/tensor/index_add_forward.cu | 49 +-
src/operator/tensor/index_update-inl.h | 143 +-
src/operator/tensor/index_update.cc | 278 ++-
src/operator/tensor/index_update.cu | 178 +-
src/operator/tensor/indexing_op.cc | 848 ++++---
src/operator/tensor/indexing_op.cu | 596 +++--
src/operator/tensor/indexing_op.h | 888 ++++---
src/operator/tensor/init_op.cc | 199 +-
src/operator/tensor/init_op.cu | 36 +-
src/operator/tensor/init_op.h | 509 ++--
src/operator/tensor/la_op-inl.h | 815 +++---
src/operator/tensor/la_op.cc | 872 ++++---
src/operator/tensor/la_op.cu | 70 +-
src/operator/tensor/la_op.h | 614 ++---
src/operator/tensor/matrix_op-inl.h | 1971 ++++++++-------
src/operator/tensor/matrix_op.cc | 1009 ++++----
src/operator/tensor/matrix_op.cu | 288 ++-
src/operator/tensor/ordering_op-inl.h | 957 +++----
src/operator/tensor/ordering_op.cc | 177 +-
src/operator/tensor/ordering_op.cu | 13 +-
src/operator/tensor/ravel.cc | 74 +-
src/operator/tensor/ravel.cu | 6 +-
src/operator/tensor/ravel.h | 97 +-
src/operator/tensor/reduce_rtc.cc | 237 +-
src/operator/tensor/slice-inl.h | 26 +-
src/operator/tensor/sort_op.h | 76 +-
src/operator/tensor/sparse_retain-inl.h | 243 +-
src/operator/tensor/sparse_retain.cc | 51 +-
src/operator/tensor/sparse_retain.cu | 6 +-
src/operator/tensor/square_sum-inl.h | 352 +--
src/operator/tensor/square_sum.cc | 63 +-
src/operator/tensor/square_sum.cu | 60 +-
src/operator/tensor/util/tensor_util-inl.h | 15 +-
src/operator/tvmop/op_module.cc | 39 +-
src/operator/tvmop/op_module.h | 19 +-
src/optimizer/sgd-inl.h | 187 +-
src/profiler/aggregate_stats.cc | 137 +-
src/profiler/aggregate_stats.h | 20 +-
src/profiler/custom_op_profiler.h | 59 +-
src/profiler/nvtx.cc | 1 -
src/profiler/nvtx.h | 8 +-
src/profiler/profiler.cc | 67 +-
src/profiler/profiler.h | 413 +--
src/profiler/storage_profiler.cc | 44 +-
src/profiler/storage_profiler.h | 56 +-
src/profiler/vtune.h | 161 +-
src/resource.cc | 334 +--
src/runtime/c_runtime_api.cc | 58 +-
src/runtime/container.cc | 66 +-
src/runtime/ndarray_handle.cc | 10 +-
src/runtime/object.cc | 42 +-
src/runtime/object_internal.h | 2 +-
src/runtime/registry.cc | 33 +-
src/serialization/cnpy.cc | 479 ++--
src/serialization/cnpy.h | 6 +-
src/storage/cpu_device_storage.h | 3 +-
src/storage/cpu_shared_storage_manager.h | 91 +-
src/storage/pooled_storage_manager.h | 175 +-
src/storage/storage.cc | 108 +-
src/storage/storage_manager.h | 12 +-
src/storage/storage_manager_helpers.h | 66 +-
940 files changed, 84073 insertions(+), 75359 deletions(-)
mode change 100755 => 100644 src/operator/contrib/hawkes_ll.cu
mode change 100755 => 100644 src/operator/contrib/preloaded_multi_sgd.cc