Merge develop

Author: sneaxiy

CMakeLists.txt

@@ -24,6 +24,8 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
         "${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
 message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
         "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
+message(STATUS "AR tools: ${CMAKE_AR}")
+
 if(WIN32)
     set(CMAKE_SUPPRESS_REGENERATION ON)
     set(CMAKE_STATIC_LIBRARY_PREFIX lib)

cmake/operators.cmake

@@ -110,7 +110,7 @@ function(op_library TARGET)
     # Define operators that don't need pybind here.
     foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
 "tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op"
-"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op")
+"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op" "sync_batch_norm_op")
         if ("${TARGET}" STREQUAL "${manual_pybind_op}")
             set(pybind_flag 1)
         endif()

paddle/fluid/API.spec

@@ -77,7 +77,7 @@ paddle.fluid.initializer.MSRAInitializer.__init__ (ArgSpec(args=['self', 'unifor
 paddle.fluid.initializer.force_init_on_cpu (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', '6d0f3e22c90d9d500d36ff57daf056ee'))
 paddle.fluid.initializer.init_on_cpu (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', 'a6d7011ca3d8c0d454dac3a56eae0c29'))
 paddle.fluid.initializer.NumpyArrayInitializer.__init__ (ArgSpec(args=['self', 'value'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.layers.fc (ArgSpec(args=['input', 'size', 'num_flatten_dims', 'param_attr', 'bias_attr', 'act', 'is_test', 'name'], varargs=None, keywords=None, defaults=(1, None, None, None, False, None)), ('document', '1929058262994f212620599c63aea6bd'))
+paddle.fluid.layers.fc (ArgSpec(args=['input', 'size', 'num_flatten_dims', 'param_attr', 'bias_attr', 'act', 'is_test', 'name'], varargs=None, keywords=None, defaults=(1, None, None, None, False, None)), ('document', '424e898365195e3ccbc2e7dc8b63605e'))
 paddle.fluid.layers.embedding (ArgSpec(args=['input', 'size', 'is_sparse', 'is_distributed', 'padding_idx', 'param_attr', 'dtype'], varargs=None, keywords=None, defaults=(False, False, None, None, 'float32')), ('document', '89c2c55a0b0656b106064048e068e77a'))
 paddle.fluid.layers.dynamic_lstm (ArgSpec(args=['input', 'size', 'h_0', 'c_0', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'float32', None)), ('document', 'dfbb624f85015df29e994ca6999e8ff6'))
 paddle.fluid.layers.dynamic_lstmp (ArgSpec(args=['input', 'size', 'proj_size', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'proj_activation', 'dtype', 'name', 'h_0', 'c_0', 'cell_clip', 'proj_clip'], varargs=None, keywords=None, defaults=(None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'tanh', 'float32', None, None, None, None, None)), ('document', 'b4b608b986eb9617aa0525e1be21d32d'))
@@ -100,7 +100,7 @@ paddle.fluid.layers.pool2d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'po
 paddle.fluid.layers.pool3d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)), ('document', '043de7333b79ee0ac55053c14ed81625'))
 paddle.fluid.layers.adaptive_pool2d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None)), ('document', '859b887174d06f361658f69cb7c06d95'))
 paddle.fluid.layers.adaptive_pool3d (ArgSpec(args=['input', 'pool_size', 'pool_type', 'require_index', 'name'], varargs=None, keywords=None, defaults=('max', False, None)), ('document', '120f4323a3d7ed9c0916f15a59f0e497'))
-paddle.fluid.layers.batch_norm (ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False)), ('document', 'c527b71b8a4c60dca8df8a745c2b598d'))
+paddle.fluid.layers.batch_norm (ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False)), ('document', '320c6973b02ea179fa89fecc80796464'))
 paddle.fluid.layers.data_norm (ArgSpec(args=['input', 'act', 'epsilon', 'param_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var'], varargs=None, keywords=None, defaults=(None, 1e-05, None, 'NCHW', False, None, None, None, False)), ('document', 'e45e09e65a2658e07cad987222f0d9ab'))
 paddle.fluid.layers.beam_search_decode (ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'b0b8d53821716cd50c42e09b593f3feb'))
 paddle.fluid.layers.conv2d_transpose (ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)), ('document', '03993955ab1e6d3044c44e6f17fc85e9'))
@@ -339,7 +339,8 @@ paddle.fluid.layers.generate_mask_labels (ArgSpec(args=['im_info', 'gt_classes',
 paddle.fluid.layers.iou_similarity (ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '587845f60c5d97ffdf2dfd21da52eca1'))
 paddle.fluid.layers.box_coder (ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name', 'axis'], varargs=None, keywords=None, defaults=('encode_center_size', True, None, 0)), ('document', '032d0f4b7d8f6235ee5d91e473344f0e'))
 paddle.fluid.layers.polygon_box_transform (ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '0e5ac2507723a0b5adec473f9556799b'))
-paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '991e934c3e09abf0edec7c9c978b4691'))
+paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'gtscore', 'use_label_smooth', 'name'], varargs=None, keywords=None, defaults=(None, True, None)), ('document', '57fa96922e42db8f064c3fb77f2255e8'))
+paddle.fluid.layers.yolo_box (ArgSpec(args=['x', 'img_size', 'anchors', 'class_num', 'conf_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '5566169a5ab993d177792c023c7fb340'))
 paddle.fluid.layers.box_clip (ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '397e9e02b451d99c56e20f268fa03f2e'))
 paddle.fluid.layers.multiclass_nms (ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None)), ('document', 'ca7d1107b6c5d2d6d8221039a220fde0'))
 paddle.fluid.layers.distribute_fpn_proposals (ArgSpec(args=['fpn_rois', 'min_level', 'max_level', 'refer_level', 'refer_scale', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '7bb011ec26bace2bc23235aa4a17647d'))
@@ -376,7 +377,7 @@ paddle.fluid.contrib.BeamSearchDecoder.read_array (ArgSpec(args=['self', 'init',
 paddle.fluid.contrib.BeamSearchDecoder.update_array (ArgSpec(args=['self', 'array', 'value'], varargs=None, keywords=None, defaults=None), ('document', '5754e9b3212b7c09497151516a0de5a7'))
 paddle.fluid.contrib.memory_usage (ArgSpec(args=['program', 'batch_size'], varargs=None, keywords=None, defaults=None), ('document', '8fcb2f93bb743693baa8d4860a5ccc47'))
 paddle.fluid.contrib.op_freq_statistic (ArgSpec(args=['program'], varargs=None, keywords=None, defaults=None), ('document', '4d43687113c4bf5b29d15aee2f4e4afa'))
-paddle.fluid.contrib.QuantizeTranspiler.__init__ (ArgSpec(args=['self', 'weight_bits', 'activation_bits', 'activation_quantize_type', 'weight_quantize_type', 'window_size'], varargs=None, keywords=None, defaults=(8, 8, 'abs_max', 'abs_max', 10000)), ('document', '14b39f1fcd5667ff556b1aad94357d1d'))
+paddle.fluid.contrib.QuantizeTranspiler.__init__ (ArgSpec(args=['self', 'weight_bits', 'activation_bits', 'activation_quantize_type', 'weight_quantize_type', 'window_size', 'moving_rate'], varargs=None, keywords=None, defaults=(8, 8, 'abs_max', 'abs_max', 10000, 0.9)), ('document', '14b39f1fcd5667ff556b1aad94357d1d'))
 paddle.fluid.contrib.QuantizeTranspiler.convert_to_int8 (ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.contrib.QuantizeTranspiler.freeze_program (ArgSpec(args=['self', 'program', 'place', 'fuse_bn', 'scope'], varargs=None, keywords=None, defaults=(False, None)), ('document', '909675a1ab055c69b436a7893fcae4fd'))
 paddle.fluid.contrib.QuantizeTranspiler.training_transpile (ArgSpec(args=['self', 'program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None)), ('document', '6dd9909f10b283ba2892a99058a72884'))
@@ -401,9 +402,9 @@ paddle.fluid.contrib.MagnitudePruner.__init__ (ArgSpec(args=['self', 'threshold'
 paddle.fluid.contrib.MagnitudePruner.prune (ArgSpec(args=['self', 'param', 'threshold'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.contrib.RatioPruner.__init__ (ArgSpec(args=['self', 'ratios'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e7a81a325b296a9ca502ee5adb4fc85d'))
 paddle.fluid.contrib.RatioPruner.prune (ArgSpec(args=['self', 'param', 'ratio'], varargs=None, keywords=None, defaults=(None,)), ('document', '358cbf2978c91028fb96a195a9884645'))
-paddle.fluid.contrib.load_persistables_for_increment (ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var', 'lookup_table_var_path'], varargs=None, keywords=None, defaults=None), ('document', '11fbf7e8dd2289805de291b453a33ee7'))
-paddle.fluid.contrib.load_persistables_for_inference (ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var_name'], varargs=None, keywords=None, defaults=None), ('document', '5b5577bb3d24070da819674255d16196'))
-paddle.fluid.contrib.convert_dist_to_sparse_program (ArgSpec(args=['program'], varargs=None, keywords=None, defaults=None), ('document', '4efbd93876832d4d35497cdbc7a1e6d8'))
+paddle.fluid.contrib.load_persistables_for_increment (ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var', 'lookup_table_var_path'], varargs=None, keywords=None, defaults=None), ('document', '2ab36d4f7a564f5f65e455807ad06c67'))
+paddle.fluid.contrib.load_persistables_for_inference (ArgSpec(args=['dirname', 'executor', 'program', 'lookup_table_var_name'], varargs=None, keywords=None, defaults=None), ('document', '59066bac9db0ac6ce414d05780b7333f'))
+paddle.fluid.contrib.convert_dist_to_sparse_program (ArgSpec(args=['program'], varargs=None, keywords=None, defaults=None), ('document', '74c39c595dc70d6be2f16d8e462d282b'))
 paddle.fluid.contrib.HDFSClient.__init__ (ArgSpec(args=['self', 'hadoop_home', 'configs'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.contrib.HDFSClient.delete (ArgSpec(args=['self', 'hdfs_path'], varargs=None, keywords=None, defaults=None), ('document', 'c3721aa2d4d9ef5a857dd47b2681c03e'))
 paddle.fluid.contrib.HDFSClient.download (ArgSpec(args=['self', 'hdfs_path', 'local_path', 'overwrite', 'unzip'], varargs=None, keywords=None, defaults=(False, False)), ('document', 'ca55bde92184d3fd0f9f5c963b25e634'))

paddle/fluid/framework/details/build_strategy.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <glog/logging.h>
 #include <memory>
+#include <utility>
 
 #include "paddle/fluid/framework/details/memory_optimize_helper.h"
 #include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
@@ -49,6 +50,11 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
       AppendPass("sequential_execution_pass");
     }
 
+    // Add op fusion.
+    if (strategy.sync_batch_norm_) {
+      AppendPass("sync_batch_norm_pass");
+    }
+
     // Add op fusion.
     if (strategy.fuse_relu_depthwise_conv_) {
       AppendPass("fuse_relu_depthwise_conv_pass");
@@ -227,6 +233,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
 }  // namespace framework
 }  // namespace paddle
 
+USE_PASS(sync_batch_norm_pass);
 USE_PASS(fuse_relu_depthwise_conv_pass);
 USE_PASS(fuse_elewise_add_act_pass);
 USE_PASS(graph_viz_pass);

paddle/fluid/framework/details/build_strategy.h

@@ -77,6 +77,8 @@ struct BuildStrategy {
 
   bool fuse_relu_depthwise_conv_{false};
 
+  bool sync_batch_norm_{false};
+
   bool memory_optimize_{true};
   // TODO(dzhwinter):
   // make enable_inplace, memory_optimize_

paddle/fluid/framework/details/memory_optimize_helper.cc

@@ -337,7 +337,6 @@ bool NodeCanReused(const VarDesc& node) {
   auto type = node.GetType();
   // only these types holds bulk of gpu memory
   if (!(type == proto::VarType::LOD_TENSOR ||
-        type == proto::VarType::SELECTED_ROWS ||
         type == proto::VarType::LOD_TENSOR_ARRAY)) {
     return false;
   }

paddle/fluid/framework/grad_op_desc_maker.h

@@ -14,7 +14,9 @@ limitations under the License. */
 
 #pragma once
 #include <algorithm>
+#include <memory>
 #include <string>
+#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 #include "paddle/fluid/framework/op_desc.h"
@@ -55,11 +57,11 @@ class GradOpDescMakerBase {
                    std::back_inserter(ret_val),
                    [this](const std::string& fwd_var_name) -> std::string {
                      auto g_name = GradVarName(fwd_var_name);
-                     if (no_grad_set_.count(g_name)) {
-                       return kEmptyVarName;
-                     } else {
+                     if (no_grad_set_.empty() || !no_grad_set_.count(g_name)) {
                        (*this->grad_to_var_)[g_name] = fwd_var_name;
                        return g_name;
+                     } else {
+                       return kEmptyVarName;
                      }
                    });
     if (!drop_empty_grad) {

paddle/fluid/framework/ir/CMakeLists.txt

@@ -46,6 +46,8 @@ cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS pass)
 pass_library(graph_to_program_pass base)
 pass_library(graph_viz_pass base)
 pass_library(lock_free_optimize_pass base)
+pass_library(cpu_quantize_pass inference)
+pass_library(cpu_quantize_squash_pass inference)
 pass_library(fc_fuse_pass inference)
 pass_library(attention_lstm_fuse_pass inference)
 pass_library(infer_clean_graph_pass inference)
@@ -66,6 +68,7 @@ pass_library(conv_elementwise_add_fuse_pass inference)
 pass_library(conv_affine_channel_fuse_pass inference)
 pass_library(transpose_flatten_concat_fuse_pass inference)
 pass_library(identity_scale_op_clean_pass base)
+pass_library(sync_batch_norm_pass base)
 
 # There may be many transpose-flatten structures in a model, and the output of
 # these structures will be used as inputs to the concat Op. This pattern will
@@ -100,6 +103,11 @@ cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS g
 cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto)
 cc_test(test_seqpool_concat_fuse_pass SRCS seqpool_concat_fuse_pass_tester.cc DEPS seqpool_concat_fuse_pass framework_proto)
 cc_test(test_is_test_pass SRCS is_test_pass_tester.cc DEPS is_test_pass)
+cc_test(test_cpu_quantize_pass SRCS cpu_quantize_pass_tester.cc DEPS cpu_quantize_pass naive_executor)
+cc_test(test_cpu_quantize_squash_pass SRCS cpu_quantize_squash_pass_tester.cc DEPS cpu_quantize_squash_pass naive_executor)
+if(NOT WIN32)
+    cc_test(test_sync_batch_norm_pass SRCS sync_batch_norm_pass_tester.cc DEPS sync_batch_norm_pass)
+endif()
 if (WITH_MKLDNN)
     cc_test(test_depthwise_conv_mkldnn_pass SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
     cc_test(test_conv_bias_mkldnn_fuse_pass SRCS mkldnn/conv_bias_mkldnn_fuse_pass_tester.cc DEPS conv_bias_mkldnn_fuse_pass naive_executor)