diff --git a/paddle/fluid/eager/auto_code_generator/CMakeLists.txt b/paddle/fluid/eager/auto_code_generator/CMakeLists.txt index 03cec80b682b115f0279885eda3efd99325bd4c1..187c3db44522296c797f9076f17b986a7b183f77 100644 --- a/paddle/fluid/eager/auto_code_generator/CMakeLists.txt +++ b/paddle/fluid/eager/auto_code_generator/CMakeLists.txt @@ -47,12 +47,12 @@ if(WIN32) endif() add_custom_target(eager_codegen - COMMAND "${eager_generator_path}/eager_generator.exe" "${PADDLE_SOURCE_DIR}/paddle/fluid/eager/api/generated/fluid_generated" + COMMAND "${eager_generator_path}/eager_generator.exe" "${PADDLE_SOURCE_DIR}/paddle/fluid/eager/api/generated/fluid_generated" "${PADDLE_SOURCE_DIR}/paddle/fluid/eager/auto_code_generator/op_list.txt" DEPENDS ${EAGER_CODEGEN_DEPS} VERBATIM) else() add_custom_target(eager_codegen - COMMAND "${CMAKE_CURRENT_BINARY_DIR}/eager_generator" "${PADDLE_SOURCE_DIR}/paddle/fluid/eager/api/generated/fluid_generated" + COMMAND "${CMAKE_CURRENT_BINARY_DIR}/eager_generator" "${PADDLE_SOURCE_DIR}/paddle/fluid/eager/api/generated/fluid_generated" "${PADDLE_SOURCE_DIR}/paddle/fluid/eager/auto_code_generator/op_list.txt" DEPENDS eager_generator VERBATIM) endif() diff --git a/paddle/fluid/eager/auto_code_generator/eager_generator.cc b/paddle/fluid/eager/auto_code_generator/eager_generator.cc index 136eaebe2cc4bfe25c507878240efdefb4341f19..283153585866ab89db3116142ea4b67d77c44a18 100644 --- a/paddle/fluid/eager/auto_code_generator/eager_generator.cc +++ b/paddle/fluid/eager/auto_code_generator/eager_generator.cc @@ -12,6 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include #include #include #include @@ -26,6 +27,9 @@ #include "paddle/fluid/pybind/pybind.h" #include "paddle/fluid/string/string_helper.h" +DEFINE_bool(generate_all, false, + "Generate all operators currently registered in Paddle"); + static std::unordered_set operators_to_skip = { "fused_elemwise_add_activation", // No Default Attr "fused_elemwise_activation", // No Default Attr @@ -40,12 +44,10 @@ static std::unordered_set operators_to_skip = { "pull_box_sparse", "fused_attention", "diag_v2", -}; - -static std::unordered_set operators_to_codegen = { - "sigmoid", "matmul_v2", "reduce_sum", "elementwise_add", - "share_buffer", "var_conv_2d", "split"}; + "transfer_dtype", + "c_split"}; +static std::unordered_set operators_to_codegen = {}; static std::unordered_set skipped_operators = {}; namespace paddle { @@ -353,7 +355,10 @@ static bool CheckOpProto(proto::OpProto* op_proto) { // Only handle matmul_v2 for now VLOG(1) << "------ Analyzing Op ------: " << op_type; - if (!operators_to_codegen.count(op_type)) return false; + if (!FLAGS_generate_all) { + if (!operators_to_codegen.count(op_type)) return false; + } + if (operators_to_skip.count(op_type)) return false; return true; @@ -976,7 +981,7 @@ static std::pair GenerateForwardFunctionContents( paddle::string::Sprintf(FWD_NUM_ARG_TEMPLATE, outnum); dygraph_function_args_str += arg_str; const char* FWD_OUTS_CONTENT_TEMPLATE = - "{ \"%s\", egr::ConstructDuplicableOutput(%s) },"; + "{ \"%s\", egr::EagerUtils::ConstructDuplicableOutput(%s) },"; outs_contents_str += paddle::string::Sprintf(FWD_OUTS_CONTENT_TEMPLATE, output_name, outnum); } else { @@ -1253,7 +1258,7 @@ static std::string GenerateGradNodeCCContents( if (duplicable_input_name_set.count(fwd_input_name)) { const char* GRAD_OUTS_CONTENT_TEMPLATE = - "{ \"%s\", egr::ConstructDuplicableOutput( " + "{ \"%s\", egr::EagerUtils::ConstructDuplicableOutput( " "this->OutputMeta()[%d].Size() ) },"; outs_contents_str += paddle::string::Sprintf( GRAD_OUTS_CONTENT_TEMPLATE, grad_output_name, fwd_input_position); @@ -1639,13 +1644,30 @@ static void DygraphCodeGeneration(const std::string& output_dir) { } // namespace framework } // namespace paddle +static void CollectOperatorsToCodeGen(const std::string& op_list_path) { + std::string line; + std::ifstream op_list_file(op_list_path); + if (op_list_file.is_open()) { + while (getline(op_list_file, line)) { + operators_to_codegen.insert(line); + } + op_list_file.close(); + } else { + PADDLE_THROW( + paddle::platform::errors::Fatal("Unable to open op_list.txt file")); + } +} + int main(int argc, char* argv[]) { - if (argc != 2) { + if (argc != 3) { std::cerr << "argc must be 2" << std::endl; return -1; } std::string eager_root = argv[1]; + std::string op_list_path = argv[2]; + + CollectOperatorsToCodeGen(op_list_path); paddle::framework::DygraphCodeGeneration(eager_root); return 0; diff --git a/paddle/fluid/eager/auto_code_generator/op_list.txt b/paddle/fluid/eager/auto_code_generator/op_list.txt index 00a9abde156fbaef5be4cd6ae95a8578ce0a7f83..6bfba753633f3381679e06111235481deadee20c 100644 --- a/paddle/fluid/eager/auto_code_generator/op_list.txt +++ b/paddle/fluid/eager/auto_code_generator/op_list.txt @@ -2,3 +2,504 @@ sigmoid matmul_v2 reduce_sum elementwise_add +rsqrt +multihead_matmul +addmm +gru +round +rank_attention +fused_embedding_fc_lstm +where_index +bicubic_interp +arg_min +tile +bilinear_tensor_product +ctc_align +pow2_decay_with_linear_warmup +split +fc +clear_float_status +load +elementwise_max +adadelta +chunk_eval +check_finite_and_unscale +sparse_momentum +tan +adam +fsp +where +logical_xor +multiclass_nms3 +one_hot_v2 +sequence_softmax +affine_channel +triangular_solve +sequence_topk_avg_pooling +space_to_depth +reverse +fused_embedding_eltwise_layernorm +expand_v2 +lgamma +solve +deformable_psroi_pooling +instance_norm +decode_jpeg +gather_nd +reduce_prod +matrix_rank +asin +lstmp +iou_similarity +huber_loss +one_hot +sequence_slice +lookup_table +softplus +depthwise_conv2d +fused_fc_elementwise_layernorm +sigmoid_cross_entropy_with_logits +exp +scatter +equal_all +searchsorted +fusion_squared_mat_sub +unique +log +conv_shift +smooth_l1_loss +linear_interp_v2 +momentum +temporal_shift +nce +mv +proximal_gd +memcpy_h2d +add_position_encoding +cosh +hash +grad_add +sign +prelu +linspace +fill_diagonal +logsigmoid +load_combine +fetch_v2 +randperm +sequence_scatter +partial_sum +relu6 +conv3d +lstm_unit +not_equal +transpose2 +uniform_random_batch_size_like +unfold +lrn +softmax_with_cross_entropy +isfinite_v2 +bernoulli +max_pool3d_with_index +gaussian_random +flatten2 +matmul +cvm +adamax +masked_select +range +bitwise_not +trace +multinomial +modified_huber_loss +roll +squared_l2_distance +conv3d_transpose +share_data +fake_quantize_abs_max +unique_with_counts +fill +concat +fill_zeros_like +hierarchical_sigmoid +isinf_v2 +squeeze +multiclass_nms2 +bpr_loss +fft_c2c +bicubic_interp_v2 +reshape +coalesce_tensor +roi_align +reshape2 +reduce_any +unstack +scatter_nd_add +sequence_reshape +bilateral_slice +fill_any_like +empty +pad_constant_like +pool2d +size +imag +eigh +stack +dgc_momentum +lamb +generate_proposals_v2 +bitwise_or +gru_unit +fake_channel_wise_quantize_dequantize_abs_max +sampling_id +unsqueeze2 +average_accumulates +sequence_enumerate +fusion_seqconv_eltadd_relu +bce_loss +generate_proposal_labels +im2sequence +isinf +adagrad +linear_chain_crf +retinanet_target_assign +fusion_group +teacher_student_sigmoid_loss +random_crop +lookup_table_v2 +detection_map +l1_norm +sqrt +fused_elemwise_activation +slogdeterminant +share_buffer +bitwise_and +diag_embed +unbind +dropout +moving_average_abs_max_scale +beam_search +log_loss +greater_than +kron +sigmoid_focal_loss +rmsprop +conv2d +uniform_random_inplace +maxout +linear_interp +auc +logical_or +batch_norm +acos +unpool +cumprod +sample_logits +pull_box_extended_sparse +crop_tensor +fill_constant +deformable_conv +generate_mask_labels +locality_aware_nms +expand_as +matrix_power +greater_equal +generate_proposals +bilinear_interp +inplace_abn +softshrink +mul +data_norm +get_tensor_from_selected_rows +spp +floor +gelu +retinanet_detection_output +minus +push_dense +silu +sequence_erase +real +nearest_interp_v2 +dgc_clip_by_norm +squeeze2 +strided_slice +conj +precision_recall +save +fusion_seqexpand_concat_fc +fake_quantize_range_abs_max +depthwise_conv2d_transpose +positive_negative_pair +square +var_conv_2d +log1p +fused_softmax_mask_upper_triangle +clip_by_norm +atan2 +box_decoder_and_assign +fft_r2c +roi_pool +overlap_add +fill_constant_batch_size_like +fill_any +dequantize_log +max_pool2d_with_index +pad3d +norm +viterbi_decode +mish +box_coder +flatten +elementwise_mod +margin_cross_entropy +pull_sparse +logical_and +pow +stanh +label_smooth +merged_momentum +ascend_trigger +fused_feedforward +rpn_target_assign +roi_perspective_transform +expand +prroi_pool +pool3d +memcpy +distribute_fpn_proposals +frame +bincount +shape +group_norm +resnet_unit +sequence_expand_as +cos_sim +eigvals +save_combine +class_center_sample +read_file +isfinite +arg_max +equal +fake_dequantize_max_abs +qr +anchor_generator +layer_norm +merge_selected_rows +less_equal +rnn +fusion_lstm +lars_momentum +hard_sigmoid +isnan +elementwise_floordiv +correlation +histogram +gather_tree +segment_pool +sync_batch_norm +fusion_repeated_fc_relu +nop +fused_attention +expand_as_v2 +filter_by_instag +diag_v2 +pull_box_sparse +nll_loss +dot +scale +ncclBcast +shuffle_batch +ncclReduce +diag +multiplex +leaky_relu +allclose +adamw +elementwise_pow +prior_box +p_norm +unique_consecutive +lod_reset +pad +sequence_conv +log10 +set_value +bitwise_xor +center_loss +randint +attention_lstm +uniform_random +slice +meshgrid +hard_swish +sin +mean_iou +pad2d +inverse +spectral_norm +shuffle_channel +psroi_pool +seed +ceil +eig +reduce_min +cos +ncclAllReduce +cudnn_lstm +digamma +assign_value +increment +tdm_sampler +fused_softmax_mask +sequence_reverse +eigvalsh +diagonal +trunc +log2 +marker +tanh +yolov3_loss +graph_send_recv +accuracy +atan +less_than +unsqueeze +crf_decoding +log_softmax +ftrl +matrix_nms +top_k_v2 +cast +tanh_shrink +hard_shrink +multiclass_nms +fusion_transpose_flatten_concat +sequence_unpad +fused_elemwise_add_activation +pull_sparse_v2 +frobenius_norm +crop +cross_entropy2 +skip_layernorm +tdm_child +fused_embedding_seq_pool +erf +conv2d_inception_fusion +trilinear_interp +logsumexp +fusion_seqpool_concat +alloc_float_status +sequence_concat +fusion_seqpool_cvm_concat +similarity_focus +argsort +sequence_expand +sgd +fused_bn_add_activation +bilinear_interp_v2 +clip +deformable_conv_v1 +hinge_loss +determinant +conv2d_transpose +memcpy_d2h +softsign +fake_quantize_dequantize_abs_max +broadcast_tensors +grid_sampler +fft_c2r +pyramid_hash +fake_quantize_dequantize_moving_average_abs_max +multi_dot +sequence_pool +transpose +top_k +dist +affine_grid +gaussian_random_batch_size_like +fake_channel_wise_dequantize_max_abs +reciprocal +sequence_mask +fill_diagonal_tensor +abs +partial_concat +elu +index_select +row_conv +cross +elementwise_mul +decayed_adagrad +bipartite_match +run_program +fake_quantize_moving_average_abs_max +mine_hard_examples +target_assign +lstm +truncated_gaussian_random +match_matrix_tensor +elementwise_div +kldiv_loss +cumsum +sum +proximal_adagrad +update_loss_scaling +shard_index +selu +mean +gumbel_softmax +sequence_pad +tree_conv +assign +flatten_contiguous_range +tril_triu +brelu +celu +reduce_mean +sinh +rank_loss +reduce_max +fusion_gru +fill_zeros_like2 +expm1 +squared_l2_norm +elementwise_sub +margin_rank_loss +faster_tokenizer +relu +is_empty +reduce_all +edit_distance +bmm +yolo_box +soft_relu +density_prior_box +eye +swish +cross_entropy +dpsgd +cholesky +batch_fc +nearest_interp +gather +trilinear_interp_v2 +box_clip +isnan_v2 +softmax +conv2d_fusion +fused_batch_norm_act +get_float_status +index_sample +elementwise_min +logical_not +collect_fpn_proposals +pixel_shuffle +thresholded_relu +polygon_box_transform +lookup_table_dequant +warpctc +fake_channel_wise_quantize_abs_max +dequantize_abs_max +svd +flip diff --git a/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc b/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc index c7c27dcc1d1508945295ca80f9c3631eacc23325..ea9aae83ff1891341471c9c6316d83b209a7c714 100644 --- a/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc +++ b/paddle/fluid/eager/tests/task_tests/eager_utils_test.cc @@ -60,7 +60,7 @@ TEST(EagerUtils, AutoGradMeta) { std::vector autograd_metas = EagerUtils::multi_autograd_meta(&ets); std::vector unsafe_autograd_metas = - EagerUtils::unsafe_autograd_meta(&ets); + EagerUtils::unsafe_autograd_meta(ets); CHECK_NOTNULL(unsafe_autograd_metas[0]); CHECK_NOTNULL(unsafe_autograd_metas[1]); diff --git a/paddle/fluid/eager/utils.cc b/paddle/fluid/eager/utils.cc index 28eefd62c5aa0acea8bfd09171fc8990989c3d6a..be06bf9eb344ba3c15a5fc4bdf21099772562cb0 100644 --- a/paddle/fluid/eager/utils.cc +++ b/paddle/fluid/eager/utils.cc @@ -48,9 +48,9 @@ AutogradMeta* EagerUtils::unsafe_autograd_meta(const egr::EagerTensor& target) { } std::vector EagerUtils::unsafe_autograd_meta( - std::vector* targets) { + const std::vector& targets) { std::vector metas; - for (const egr::EagerTensor& t : *targets) { + for (const egr::EagerTensor& t : targets) { metas.push_back(unsafe_autograd_meta(t)); } return metas; diff --git a/paddle/fluid/eager/utils.h b/paddle/fluid/eager/utils.h index f7e226a2aba36c0540ceb2f06d3890d0dfbe6097..03f922e5bf9ba9759762e14ede2c998918016d9a 100644 --- a/paddle/fluid/eager/utils.h +++ b/paddle/fluid/eager/utils.h @@ -114,7 +114,7 @@ class EagerUtils { // This method will return an AutogradMeta pointer unsafely. static AutogradMeta* unsafe_autograd_meta(const egr::EagerTensor& target); static std::vector unsafe_autograd_meta( - std::vector* targets); + const std::vector& targets); template static bool ComputeRequireGrad(T trace_backward, Args&&... args) {