Merge branch 'fast_install_1.3' of https://github.com/JiabinYang/Paddle into...

Merge branch 'fast_install_1.3' of https://github.com/JiabinYang/Paddle into JiabinYang-fast_install_1.3

CMakeLists.txt

@@ -212,7 +212,7 @@ endif()
 if (WITH_JEMALLOC)
     find_package(JeMalloc REQUIRED)
     include_directories(${JEMALLOC_INCLUDE_DIR})
-    add_definitions(-DWITH_JEMALLOC)
+    add_definitions(-DPADDLE_WITH_JEMALLOC)
 endif()
 
 include(generic)            # simplify cmake module
@@ -276,9 +276,3 @@ add_subdirectory(paddle)
 if(WITH_PYTHON)
     add_subdirectory(python)
 endif()
-
-if(WITH_DOC)
-    find_package(Sphinx REQUIRED)
-    find_python_module(recommonmark REQUIRED)
-    add_subdirectory(doc)
-endif()

Dockerfile

@@ -11,12 +11,10 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
 # ENV variables
 ARG WITH_GPU
 ARG WITH_AVX
-ARG WITH_DOC
 
 ENV WOBOQ OFF
 ENV WITH_GPU=${WITH_GPU:-ON}
 ENV WITH_AVX=${WITH_AVX:-ON}
-ENV WITH_DOC=${WITH_DOC:-OFF}
 
 ENV HOME /root
 # Add bash enhancements

cmake/FindSphinx.cmake

-# - This module looks for Sphinx
-# Find the Sphinx documentation generator
-#
-# This modules defines
-#  SPHINX_EXECUTABLE
-#  SPHINX_FOUND
-
-find_program(SPHINX_EXECUTABLE
-  NAMES sphinx-build
-  PATHS
-    /usr/bin
-    /usr/local/bin
-    /opt/local/bin
-  DOC "Sphinx documentation generator"
-)
-
-if( NOT SPHINX_EXECUTABLE )
-  set(_Python_VERSIONS
-    2.7 2.6 2.5 2.4 2.3 2.2 2.1 2.0 1.6 1.5
-  )
-
-  foreach( _version ${_Python_VERSIONS} )
-    set( _sphinx_NAMES sphinx-build-${_version} )
-
-    find_program( SPHINX_EXECUTABLE
-      NAMES ${_sphinx_NAMES}
-      PATHS
-        /usr/bin
-        /usr/local/bin
-        /opt/loca/bin
-      DOC "Sphinx documentation generator"
-    )
-  endforeach()
-endif()
-
-include(FindPackageHandleStandardArgs)
-
-find_package_handle_standard_args(Sphinx DEFAULT_MSG
-  SPHINX_EXECUTABLE
-)
-
-
-option( SPHINX_HTML_OUTPUT "Build a single HTML with the whole content." ON )
-option( SPHINX_DIRHTML_OUTPUT "Build HTML pages, but with a single directory per document." OFF )
-option( SPHINX_HTMLHELP_OUTPUT "Build HTML pages with additional information for building a documentation collection in htmlhelp." OFF )
-option( SPHINX_QTHELP_OUTPUT "Build HTML pages with additional information for building a documentation collection in qthelp." OFF )
-option( SPHINX_DEVHELP_OUTPUT "Build HTML pages with additional information for building a documentation collection in devhelp." OFF )
-option( SPHINX_EPUB_OUTPUT "Build HTML pages with additional information for building a documentation collection in epub." OFF )
-option( SPHINX_LATEX_OUTPUT "Build LaTeX sources that can be compiled to a PDF document using pdflatex." OFF )
-option( SPHINX_MAN_OUTPUT "Build manual pages in groff format for UNIX systems." OFF )
-option( SPHINX_TEXT_OUTPUT "Build plain text files." OFF )
-
-
-mark_as_advanced(
-  SPHINX_EXECUTABLE
-  SPHINX_HTML_OUTPUT
-  SPHINX_DIRHTML_OUTPUT
-  SPHINX_HTMLHELP_OUTPUT
-  SPHINX_QTHELP_OUTPUT
-  SPHINX_DEVHELP_OUTPUT
-  SPHINX_EPUB_OUTPUT
-  SPHINX_LATEX_OUTPUT
-  SPHINX_MAN_OUTPUT
-  SPHINX_TEXT_OUTPUT
-)
-
-function( Sphinx_add_target target_name builder conf cache source destination )
-  add_custom_target( ${target_name} ALL
-    COMMAND ${SPHINX_EXECUTABLE} -b ${builder}
-    -d ${cache}
-    -c ${conf}
-    ${source}
-    ${destination}
-    COMMENT "Generating sphinx documentation: ${builder}"
-    COMMAND cd ${destination} && ln -sf ./index_*.html index.html
-    )
-
-  set_property(
-    DIRECTORY APPEND PROPERTY
-    ADDITIONAL_MAKE_CLEAN_FILES
-    ${destination}
-    )
-endfunction()
-
-# Target dependencies can be optionally listed at the end.
-function( Sphinx_add_targets target_base_name conf source base_destination )
-
-  set( _dependencies )
-
-  foreach( arg IN LISTS ARGN )
-    set( _dependencies ${_dependencies} ${arg} )
-  endforeach()
-
-  if( ${SPHINX_HTML_OUTPUT} )
-    Sphinx_add_target( ${target_base_name}_html html ${conf} ${source} ${base_destination}/html )
-
-    add_dependencies( ${target_base_name}_html ${_dependencies} )
-  endif()
-
-  if( ${SPHINX_DIRHTML_OUTPUT} )
-    Sphinx_add_target( ${target_base_name}_dirhtml dirhtml ${conf} ${source} ${base_destination}/dirhtml )
-
-    add_dependencies( ${target_base_name}_dirhtml ${_dependencies} )
-  endif()
-
-  if( ${SPHINX_QTHELP_OUTPUT} )
-    Sphinx_add_target( ${target_base_name}_qthelp qthelp ${conf} ${source} ${base_destination}/qthelp )
-
-    add_dependencies( ${target_base_name}_qthelp ${_dependencies} )
-  endif()
-
-  if( ${SPHINX_DEVHELP_OUTPUT} )
-    Sphinx_add_target( ${target_base_name}_devhelp devhelp ${conf} ${source} ${base_destination}/devhelp )
-
-    add_dependencies( ${target_base_name}_devhelp ${_dependencies} )
-  endif()
-
-  if( ${SPHINX_EPUB_OUTPUT} )
-    Sphinx_add_target( ${target_base_name}_epub epub ${conf} ${source} ${base_destination}/epub )
-
-    add_dependencies( ${target_base_name}_epub ${_dependencies} )
-  endif()
-
-  if( ${SPHINX_LATEX_OUTPUT} )
-    Sphinx_add_target( ${target_base_name}_latex latex ${conf} ${source} ${base_destination}/latex )
-
-    add_dependencies( ${target_base_name}_latex ${_dependencies} )
-  endif()
-
-  if( ${SPHINX_MAN_OUTPUT} )
-    Sphinx_add_target( ${target_base_name}_man man ${conf} ${source} ${base_destination}/man )
-
-    add_dependencies( ${target_base_name}_man ${_dependencies} )
-  endif()
-
-  if( ${SPHINX_TEXT_OUTPUT} )
-    Sphinx_add_target( ${target_base_name}_text text ${conf} ${source} ${base_destination}/text )
-
-    add_dependencies( ${target_base_name}_text ${_dependencies} )
-  endif()
-
-  if( ${BUILD_TESTING} )
-    sphinx_add_target( ${target_base_name}_linkcheck linkcheck ${conf} ${source} ${base_destination}/linkcheck )
-
-    add_dependencies( ${target_base_name}_linkcheck ${_dependencies} )
-  endif()
-endfunction()

cmake/operators.cmake

@@ -52,8 +52,8 @@ function(op_library TARGET)
         endif()
         if(WITH_MKLDNN)
             string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}")
-            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc)
-                list(APPEND mkldnn_cc_srcs ${MKLDNN_FILE}.cc)
+            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/mkldnn/${MKLDNN_FILE}.cc)
+                list(APPEND mkldnn_cc_srcs mkldnn/${MKLDNN_FILE}.cc)
             endif()
         endif()
     else()

paddle/fluid/API.spec

@@ -122,7 +122,7 @@ paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None,
 paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
 paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0, False))
 paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name', 'path_table', 'path_code', 'is_custom', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, False, False))
-paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'is_accumulated', 'name'], varargs=None, keywords=None, defaults=(0, True, None))
+paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'is_accumulated', 'name', 'return_parent_idx'], varargs=None, keywords=None, defaults=(0, True, None, False))
 paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
 paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
@@ -142,10 +142,10 @@ paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon',
 paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0))
 paddle.fluid.layers.roi_align ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None))
 paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,))
-paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None))
+paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None, True, 1))
 paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',))
-paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None))
-paddle.fluid.layers.resize_nearest ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None))
+paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners', 'align_mode'], varargs=None, keywords=None, defaults=(None, None, None, None, True, 1))
+paddle.fluid.layers.resize_nearest ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape', 'align_corners'], varargs=None, keywords=None, defaults=(None, None, None, None, True))
 paddle.fluid.layers.gather ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,))
@@ -322,9 +322,10 @@ paddle.fluid.layers.generate_proposal_labels ArgSpec(args=['rpn_rois', 'gt_class
 paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'im_info', 'anchors', 'variances', 'pre_nms_top_n', 'post_nms_top_n', 'nms_thresh', 'min_size', 'eta', 'name'], varargs=None, keywords=None, defaults=(6000, 1000, 0.5, 0.1, 1.0, None))
 paddle.fluid.layers.generate_mask_labels ArgSpec(args=['im_info', 'gt_classes', 'is_crowd', 'gt_segms', 'rois', 'labels_int32', 'num_classes', 'resolution'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name'], varargs=None, keywords=None, defaults=('encode_center_size', True, None))
+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))
 paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'class_num', 'ignore_thresh', 'loss_weight_xy', 'loss_weight_wh', 'loss_weight_conf_target', 'loss_weight_conf_notarget', 'loss_weight_class', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None))
+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))
 paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
 paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1))
 paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))

paddle/fluid/framework/CMakeLists.txt

-
 #windows treat symbolic file as a real file, which is different with unix
 #We create a hidden file and compile it instead of origin source file.
 function(windows_symbolic TARGET)
@@ -207,3 +206,24 @@ endif (NOT WIN32)
 
 cc_library(dlpack_tensor SRCS dlpack_tensor.cc DEPS tensor dlpack)
 cc_test(dlpack_tensor_test SRCS dlpack_tensor_test.cc DEPS dlpack_tensor glog)
+
+# Get the current working branch
+execute_process(
+  COMMAND git rev-parse --abbrev-ref HEAD
+  WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
+  OUTPUT_VARIABLE PADDLE_BRANCH
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+
+# Get the latest abbreviated commit hash of the working branch
+execute_process(
+  COMMAND git log -1 --format=%h
+  WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
+  OUTPUT_VARIABLE PADDLE_COMMIT
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+
+message(STATUS "commit: ${PADDLE_COMMIT}")
+message(STATUS "branch: ${PADDLE_BRANCH}")
+
+configure_file(commit.h.in commit.h)

paddle/fluid/framework/commit.h.in

+#pragma once
+
+#include <string>
+
+namespace paddle {
+namespace framework {
+
+static std::string paddle_commit() {
+  return "@PADDLE_COMMIT@";
+}
+
+static std::string paddle_compile_branch() {
+  return "@PADDLE_BRANCH@";
+}
+
+static std::string paddle_version() {
+  return "@PADDLE_VERSION@";
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/ir/CMakeLists.txt

@@ -10,8 +10,22 @@ function(pass_library TARGET DEST)
     set(options "")
     set(oneValueArgs "")
     set(multiValueArgs SRCS DEPS)
+    set(targetPrefix "")
+
+    # Get optional argument
+    set(extraMacroArgs ${ARGN})
+    list(LENGTH extraMacroArgs numExtraMacroArgs)
+    if(numExtraMacroArgs GREATER 0)
+        list(GET extraMacroArgs 0 targetPrefix)
+    endif()
+
     cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-    cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS})
+    if(targetPrefix)
+        cc_library(${TARGET} SRCS ${targetPrefix}/${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS})
+    else()
+        cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS})
+    endif()
+
     # add more DEST here, such as train, dist and collect USE_PASS into a file automatically.
     if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference")
         message(STATUS "add pass ${TARGET} ${DEST}")
@@ -62,11 +76,11 @@ foreach (index RANGE 3 6)
 endforeach()
 
 if(WITH_MKLDNN)
-    pass_library(mkldnn_placement_pass base)
-    pass_library(depthwise_conv_mkldnn_pass base)
-    pass_library(conv_bias_mkldnn_fuse_pass inference)
-    pass_library(conv_relu_mkldnn_fuse_pass inference)
-    pass_library(conv_elementwise_add_mkldnn_fuse_pass inference)
+    pass_library(mkldnn_placement_pass base mkldnn)
+    pass_library(depthwise_conv_mkldnn_pass base mkldnn)
+    pass_library(conv_bias_mkldnn_fuse_pass inference mkldnn)
+    pass_library(conv_relu_mkldnn_fuse_pass inference mkldnn)
+    pass_library(conv_elementwise_add_mkldnn_fuse_pass inference mkldnn)
 endif()
 
 cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector )
@@ -86,7 +100,7 @@ cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framewor
 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)
 if (WITH_MKLDNN)
-    cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
-    cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
-    cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
+    cc_test(test_depthwise_conv_mkldnn_pass SRCS mkldnn/depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass)
+    cc_test(test_conv_relu_mkldnn_fuse_pass SRCS mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass)
+    cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass)
 endif ()

paddle/fluid/framework/ir/graph_traits.cc

@@ -14,6 +14,7 @@
 
 #include "paddle/fluid/framework/ir/graph_traits.h"
 
+#include <set>
 #include <vector>
 
 namespace paddle {
@@ -79,7 +80,7 @@ NodesTSIterator::NodesTSIterator(const std::vector<Node *> &source) {
   }
 
   std::unordered_set<Node *> visited;
-  std::unordered_set<Node *> to_visit{source.begin(), source.end()};
+  std::set<Node *> to_visit{source.begin(), source.end()};
 
   std::vector<Node *> inlink_visited;
   while (!to_visit.empty()) {

paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc → paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h"
+#include "paddle/fluid/framework/ir/mkldnn/conv_bias_mkldnn_fuse_pass.h"
 #include <functional>
 #include <string>
 #include <vector>

paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc → paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
+#include "paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.h"
 #include <functional>
 #include <list>
 #include <map>

paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc → paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass_tester.cc

@@ -15,8 +15,8 @@
 #include <gtest/gtest.h>
 #include <string>
 
-#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h"
 #include "paddle/fluid/framework/ir/graph_traits.h"
+#include "paddle/fluid/framework/ir/mkldnn/conv_elementwise_add_mkldnn_fuse_pass.h"
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc → paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h"
+#include "paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.h"
 #include <string>
 #include <vector>
 #include "paddle/fluid/platform/enforce.h"

paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc → paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass_tester.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h"
+#include "paddle/fluid/framework/ir/mkldnn/conv_relu_mkldnn_fuse_pass.h"
 
 #include <gtest/gtest.h>
 #include "paddle/fluid/framework/op_proto_maker.h"

paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc → paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.cc

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h"
+#include "paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.h"
 #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 
 namespace paddle {

paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc → paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass_tester.cc

@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h"
+#include "paddle/fluid/framework/ir/mkldnn/depthwise_conv_mkldnn_pass.h"
 
 #include <gtest/gtest.h>
 

paddle/fluid/framework/ir/mkldnn_placement_pass.cc → paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.cc

@@ -12,7 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "paddle/fluid/framework/ir/mkldnn_placement_pass.h"
+#include "paddle/fluid/framework/ir/mkldnn/mkldnn_placement_pass.h"
 #include <string>
 
 namespace paddle {

paddle/fluid/framework/ngraph_operator.cc

-/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include <glog/logging.h>
-
-#include <algorithm>
-#include <map>
-
-#include "paddle/fluid/framework/feed_fetch_type.h"
-#include "paddle/fluid/framework/framework.pb.h"
-#include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/ngraph_bridge.h"
-#include "paddle/fluid/framework/ngraph_operator.h"
-#include "paddle/fluid/framework/tensor.h"
-#include "paddle/fluid/framework/var_desc.h"
-#include "paddle/fluid/framework/var_type.h"
-
-#include "ngraph/ngraph.hpp"
-
-namespace paddle {
-namespace framework {
-
-static ngraph::Shape Ddim2Shape(const DDim& dims) {
-  ngraph::Shape sp;
-  for (int i = 0; i < dims.size(); ++i) {
-    int k = dims[i];
-    k = k == 0 ? 1 : k;
-    sp.push_back(k);
-  }
-  return sp;
-}
-
-static std::map<proto::VarType::Type, ngraph::element::Type> pd2ng_type_map = {
-    {proto::VarType::FP32, ngraph::element::f32},
-    {proto::VarType::FP64, ngraph::element::f64},
-    {proto::VarType::INT32, ngraph::element::i32},
-    {proto::VarType::INT64, ngraph::element::i64},
-    {proto::VarType::BOOL, ngraph::element::boolean},
-};
-
-typedef enum {                /* nGraph support state on ops          */
-               FULL_TRAIN,    /* Support full ops for train           */
-               PARTIAL_TRAIN, /* Support partial ops for train        */
-               FULL_TEST,     /* Support full list of ops for test    */
-               PARTIAL_TEST   /* Support partial list of ops for test */
-} op_state;
-
-// perform graph build through bridge and execute computation
-class NgraphEngine {
- public:
-  explicit NgraphEngine(const Scope& scope, const platform::Place& place,
-                        const std::vector<std::shared_ptr<OperatorBase>>& ops,
-                        const std::unordered_map<
-                            std::string, ngraph::element::Type>& var_type_map,
-                        const std::unordered_set<std::string>& persist,
-                        const std::unordered_set<std::string>& fetches,
-                        const std::unordered_set<std::string>& post_op_inputs,
-                        op_state ng_op_state)
-      : scope_(scope),
-        place_(place),
-        fused_ops_(ops),
-        var_type_map_(var_type_map),
-        persistables_(persist),
-        fetches_(fetches),
-        post_op_inputs_(post_op_inputs),
-        ng_op_state_(ng_op_state) {
-    var_in_node_map_ = std::make_shared<
-        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>();
-
-    var_node_map_ = std::make_shared<
-        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>();
-
-    BuildNgIO();
-
-    GetNgFunction();
-  }
-
-  void Run(const Scope& scope, const platform::Place& place) const;
-
- private:
-  static std::unordered_map<std::string, std::shared_ptr<ngraph::Function>>
-      func_cache_;
-  const Scope& scope_;
-  const platform::Place& place_;
-  std::vector<std::shared_ptr<OperatorBase>> fused_ops_;
-  std::unordered_map<std::string, ngraph::element::Type> var_type_map_;
-  std::unordered_set<std::string> persistables_;
-  std::unordered_set<std::string> fetches_;
-  std::unordered_set<std::string> post_op_inputs_;
-  op_state ng_op_state_;
-
-  // ngraph backend eg. CPU
-  static std::shared_ptr<ngraph::runtime::Backend> backend_;
-  // ngraph function to call and execute
-  std::shared_ptr<ngraph::Function> ngraph_function_;
-  // var_name of inputs
-  std::vector<std::string> var_in_;
-  // var_name of outputs from  fetch in order
-  std::vector<std::string> var_out_;
-  // map input vars to nodes
-  std::shared_ptr<
-      std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
-      var_in_node_map_;
-  // map each var name with a ngraph node
-  std::shared_ptr<
-      std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
-      var_node_map_;
-  // cache key to check if function is cached
-  std::shared_ptr<std::string> GetCacheKey();
-  // get ngraph input and define ngraph input parameters
-  void GetNgInputShape(std::shared_ptr<OperatorBase> op);
-  // Call ngraph bridge to map ops
-  void BuildNgNodes();
-  // get the ngraph input and output var list
-  void BuildNgIO();
-  // build ngraph function call
-  void BuildNgFunction();
-  // Check cache for ngraph function or otherwise build the function
-  void GetNgFunction();
-};
-
-std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
-NgraphOperator::NgraphOpIntervals(
-    std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops) {
-  std::vector<std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
-      intervals;
-  if (ops->empty()) {
-    return intervals;
-  }
-  size_t size = ops->size();
-  size_t left = 0;
-  while (left < size && ops->at(left)->Type() != kFeedOpType) {
-    ++left;
-  }
-  if (left == size) {
-    return intervals;
-  }
-  while (left < size && ops->at(left)->Type() == kFeedOpType) {
-    ++left;
-  }
-
-  size_t right = left;
-  while (right < size && ops->at(right)->Type() != kFetchOpType) {
-    ++right;
-  }
-  if (right == size) {
-    return intervals;
-  }
-  if (left >= right) return intervals;
-
-  // (left, right - 1) represents indices between feed and fetch
-  size_t pivot = left;
-  while (pivot < right) {
-    auto op_type = ops->at(pivot)->Type();
-    if (paddle::framework::NgraphBridge::NG_NODE_MAP.find(op_type) ==
-        paddle::framework::NgraphBridge::NG_NODE_MAP.end()) {
-      ++pivot;
-    } else {
-      size_t start = pivot, end = start;
-      while (pivot < right &&
-             (paddle::framework::NgraphBridge::NG_NODE_MAP.find(
-                  ops->at(pivot)->Type()) !=
-              paddle::framework::NgraphBridge::NG_NODE_MAP.end())) {
-        ++pivot;
-        ++end;
-      }
-      std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>
-          interval = {ops->begin() + start, ops->begin() + end};
-      intervals.push_back(interval);
-    }
-  }  // end while
-
-  return intervals;
-}
-
-NgraphOperator::NgraphOperator(
-    const ProgramDesc& prog, size_t block_id,
-    std::vector<std::unique_ptr<OperatorBase>>::iterator start,
-    std::vector<std::unique_ptr<OperatorBase>>::iterator end,
-    const std::string& type, const VariableNameMap& inputs,
-    const VariableNameMap& outputs, const AttributeMap& attrs)
-    : OperatorBase(type, inputs, outputs, attrs),
-      pdesc_(prog),
-      block_(block_id) {
-  for (std::vector<std::unique_ptr<OperatorBase>>::iterator it = start;
-       it != end; ++it) {
-    fused_ops_.push_back(std::move(*it));
-  }
-
-  for (std::vector<std::unique_ptr<OperatorBase>>::iterator it = end;
-       (*it)->Type() != kFetchOpType; ++it) {
-    for (auto& var_name_item : (*it)->Inputs()) {
-      for (auto& var_name : var_name_item.second) {
-        post_op_inputs_.insert(var_name);
-      }
-    }
-  }
-
-  if ((*(start - 1))->Type() == kFeedOpType && (*end)->Type() == kFetchOpType) {
-    is_full_ = true;
-  }
-
-  Process();
-}
-
-void NgraphOperator::Process() {
-  auto& bdesc = pdesc_.Block(block_);
-  for (auto& var : bdesc.AllVars()) {
-    if (!(var->GetType() == proto::VarType::SELECTED_ROWS ||
-          var->GetType() == proto::VarType::LOD_TENSOR ||
-          var->GetType() == proto::VarType::LOD_TENSOR_ARRAY)) {
-      continue;
-    }
-
-    auto var_name = var->Name();
-    if (var->Name() == framework::kEmptyVarName) {
-      continue;
-    }
-
-    if (var_name != "fetch" && var_name != "feed") {
-      auto pd_type = var->GetDataType();
-      if (pd2ng_type_map.find(pd_type) == pd2ng_type_map.end()) {
-        PADDLE_THROW("Data type of var %s not found in pd2ng_type_map",
-                     var_name);
-      }
-      var_type_map_[var_name] = pd2ng_type_map[pd_type];
-    }
-
-    if (var->Persistable()) {
-      persistables_.insert(var->Name());
-    }
-  }
-
-  for (auto* op : bdesc.AllOps()) {
-    if (op->Type() == kFetchOpType) {
-      std::string fetch_target_name = op->Input("X")[0];
-      fetches_.insert(fetch_target_name);
-    }
-  }
-}
-
-void NgraphOperator::RunImpl(const Scope& scope,
-                             const platform::Place& place) const {
-  op_state ng_op_state = PARTIAL_TEST;
-  auto& bdesc = pdesc_.Block(block_);
-  for (auto* op : bdesc.AllOps()) {
-    if (op->Type().find("_grad") != std::string::npos) {
-      ng_op_state = PARTIAL_TRAIN;
-      break;
-    }
-  }
-
-  if (is_full_) {
-    ng_op_state = ng_op_state == PARTIAL_TEST ? FULL_TEST : FULL_TRAIN;
-  }
-
-  NgraphEngine ngraph_engine(scope, place, fused_ops_, var_type_map_,
-                             persistables_, fetches_, post_op_inputs_,
-                             ng_op_state);
-  ngraph_engine.Run(scope, place);
-}
-
-std::unordered_map<std::string, std::shared_ptr<ngraph::Function>>
-    NgraphEngine::func_cache_ = {};
-
-std::shared_ptr<ngraph::runtime::Backend> NgraphEngine::backend_ =
-    ngraph::runtime::Backend::create("CPU");
-
-void NgraphEngine::GetNgInputShape(std::shared_ptr<OperatorBase> op) {
-  RuntimeContext ctx(op->Inputs(), op->Outputs(), scope_);
-  op->RuntimeInferShape(scope_, place_, ctx);
-  for (auto& var_name_item : op->Inputs()) {
-    for (auto& var_name : var_name_item.second) {
-      auto* var = scope_.FindVar(var_name);
-      if (var && var->IsType<LoDTensor>()) {
-        auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var);
-        auto sp = Ddim2Shape(tensor_pd->dims());
-        if (std::find(var_in_.begin(), var_in_.end(), var_name) !=
-            var_in_.end()) {
-          if (var_node_map_->find(var_name) == var_node_map_->end()) {
-            auto ng_type = var_type_map_.at(var_name);
-            auto prm =
-                std::make_shared<ngraph::op::Parameter>(ng_type, sp, true);
-            (*var_node_map_)[var_name] = prm;
-            (*var_in_node_map_)[var_name] = prm;
-          }
-        }
-      }
-    }
-  }
-}
-
-void NgraphEngine::BuildNgNodes() {
-  for (auto& var_name : var_out_) {
-    if (var_node_map_->find(var_name) == var_node_map_->end()) {
-      auto* var = scope_.FindVar(var_name);
-      if (var && var->IsType<LoDTensor>()) {
-        auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var);
-        auto& ddim = tensor_pd->dims();
-        auto ng_shape = Ddim2Shape(ddim);
-        auto ng_type = var_type_map_.at(var_name);
-        auto prm =
-            std::make_shared<ngraph::op::Parameter>(ng_type, ng_shape, true);
-        (*var_node_map_)[var_name] = prm;
-      }
-    }
-  }
-
-  paddle::framework::NgraphBridge ngb(var_node_map_);
-  for (auto& op : fused_ops_) {
-    ngb.BuildNgNode(op);
-  }
-}
-
-void NgraphEngine::BuildNgIO() {
-  std::unordered_set<std::string> inputs;
-  std::unordered_set<std::string> outputs;
-
-  for (auto& op : fused_ops_) {
-    for (auto& var_name_item : op->Inputs()) {
-      for (auto& var_name : var_name_item.second) {
-        inputs.insert(var_name);
-        const bool is_output = outputs.find(var_name) != outputs.end();
-        if (!is_output &&
-            std::find(var_in_.begin(), var_in_.end(), var_name) ==
-                var_in_.end()) {
-          // fill var_in here to keep lhs and rhs order
-          var_in_.push_back(var_name);
-        }
-      }
-    }
-
-    if (op->Type() != "fill_constant") {
-      GetNgInputShape(op);
-    }
-
-    for (auto& var_name_item : op->Outputs()) {
-      PADDLE_ENFORCE_LE(var_name_item.second.size(), 1,
-                        "op %s has more than 1 output - Not handling yet",
-                        op->Type());
-      for (auto& var_name : var_name_item.second) {
-        outputs.insert(var_name);
-      }
-    }
-  }
-
-  // var_out.clear();
-  for (auto& op : fused_ops_) {
-    for (auto& var_name_item : op->Outputs()) {
-      PADDLE_ENFORCE_LE(var_name_item.second.size(), 1,
-                        "op %s has more than 1 output - Not handling yet",
-                        op->Type());
-      for (auto& var_name : var_name_item.second) {
-        switch (ng_op_state_) {
-          case PARTIAL_TEST:
-            if (post_op_inputs_.find(var_name) != post_op_inputs_.end() ||
-                fetches_.find(var_name) != fetches_.end()) {
-              var_out_.push_back(var_name);
-            }
-            break;
-          case FULL_TEST:
-            if (fetches_.find(var_name) != fetches_.end()) {
-              var_out_.push_back(var_name);
-            }
-            break;
-          case PARTIAL_TRAIN:
-            if (fetches_.find(var_name) != fetches_.end() ||
-                post_op_inputs_.find(var_name) != post_op_inputs_.end() ||
-                persistables_.find(var_name) != persistables_.end()) {
-              var_out_.push_back(var_name);
-            }
-            break;
-          case FULL_TRAIN:
-            if (fetches_.find(var_name) != fetches_.end() ||
-                persistables_.find(var_name) != persistables_.end()) {
-              var_out_.push_back(var_name);
-            }
-            break;
-          default:
-            var_out_.push_back(var_name);
-        }
-      }
-    }
-  }
-}
-
-void NgraphEngine::BuildNgFunction() {
-  BuildNgNodes();
-  ngraph_function_ = nullptr;
-  ngraph::NodeVector func_outputs;
-  ngraph::ParameterVector func_inputs;
-
-  for (auto& vo : var_out_) {
-    func_outputs.push_back(var_node_map_->at(vo));
-  }
-
-  for (auto& vi : var_in_) {
-    std::shared_ptr<ngraph::op::Parameter> prm =
-        std::dynamic_pointer_cast<ngraph::op::Parameter>(
-            var_in_node_map_->at(vi));
-    func_inputs.push_back(prm);
-  }
-
-  ngraph_function_ =
-      std::make_shared<ngraph::Function>(func_outputs, func_inputs);
-}
-
-std::shared_ptr<std::string> NgraphEngine::GetCacheKey() {
-  auto cache_key = std::make_shared<std::string>("");
-  *cache_key += std::to_string(fused_ops_.size());
-  for (auto& op : fused_ops_) {
-    *cache_key += op->Type();
-  }
-  for (auto& var_name : var_in_) {
-    auto shape = var_node_map_->at(var_name)->get_shape();
-    *cache_key += var_name;
-    *cache_key += var_type_map_.at(var_name).c_type_string();
-    for (size_t i = 0; i < shape.size(); ++i) {
-      *cache_key += std::to_string(shape.at(i));
-    }
-  }
-
-  for (auto& var_name : var_out_) {
-    auto* var = scope_.FindVar(var_name);
-    if (var && var->IsType<LoDTensor>()) {
-      auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var);
-      auto& ddim = tensor_pd->dims();
-      for (int i = 0; i < ddim.size(); ++i) {
-        *cache_key += std::to_string(ddim[i]);
-      }
-    }
-  }
-  return cache_key;
-}
-
-void NgraphEngine::GetNgFunction() {
-  bool cache_on = true;
-  if (cache_on) {
-    std::string cache_key_val = *GetCacheKey();
-    if (func_cache_.find(cache_key_val) != func_cache_.end()) {
-      ngraph_function_ = func_cache_.at(cache_key_val);
-    } else {
-      BuildNgFunction();
-      func_cache_[cache_key_val] = ngraph_function_;
-    }
-  } else {
-    BuildNgFunction();
-  }
-}
-
-void NgraphEngine::Run(const Scope& scope, const platform::Place& place) const {
-  std::vector<std::shared_ptr<ngraph::runtime::Tensor>> t_in;
-  std::vector<std::shared_ptr<ngraph::runtime::Tensor>> t_out;
-
-  for (size_t i = 0; i < var_in_.size(); ++i) {
-    auto vi = var_in_.at(i);
-    auto sp = var_node_map_->at(vi)->get_shape();
-    std::shared_ptr<ngraph::runtime::Tensor> ti;
-    auto* var = scope.FindVar(vi);
-    if (var && var->IsType<LoDTensor>()) {
-      auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var);
-      PADDLE_ENFORCE(sp == Ddim2Shape(tensor_pd->dims()),
-                     "Ensure ngraph tensor layout align with paddle tensor");
-      if (tensor_pd->type() == proto::VarType::FP32) {
-        const float* arr = tensor_pd->data<float>();
-        ti = backend_->create_tensor(ngraph::element::f32, sp,
-                                     const_cast<float*>(arr));
-      } else if (tensor_pd->type() == proto::VarType::INT32) {
-        const int* arr = tensor_pd->data<int>();
-        ti = backend_->create_tensor(ngraph::element::i32, sp,
-                                     const_cast<int*>(arr));
-      } else if (tensor_pd->type() == proto::VarType::INT64) {
-        const int64_t* arr = tensor_pd->data<int64_t>();
-        ti = backend_->create_tensor(ngraph::element::i64, sp,
-                                     const_cast<int64_t*>(arr));
-      } else if (tensor_pd->type() == proto::VarType::FP64) {
-        const double* arr = tensor_pd->data<double>();
-        ti = backend_->create_tensor(ngraph::element::f64, sp,
-                                     const_cast<double*>(arr));
-      } else if (tensor_pd->type() == proto::VarType::BOOL) {
-        const bool* arr = tensor_pd->data<bool>();
-        ti = backend_->create_tensor(ngraph::element::boolean, sp,
-                                     const_cast<bool*>(arr));
-      } else {
-        PADDLE_THROW("Data type not handling for var %s", vi);
-      }
-    } else {
-      PADDLE_THROW("Cannot find var or tensor with var name %s", vi);
-    }
-    bool is_test = (ng_op_state_ == PARTIAL_TEST || ng_op_state_ == FULL_TEST)
-                       ? true
-                       : false;
-    bool is_persistable =
-        (persistables_.find(vi) != persistables_.end()) ? true : false;
-    if (is_test && is_persistable) {
-      ti->set_stale(false);
-    }
-    t_in.push_back(ti);
-  }
-
-  for (size_t i = 0; i < var_out_.size(); ++i) {
-    auto var_name = var_out_[i];
-    auto* var = scope.FindVar(var_name);
-    std::shared_ptr<ngraph::runtime::Tensor> to;
-    if (var && var->IsType<LoDTensor>()) {
-      auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var);
-      auto dd = tensor_pd->dims();
-      ngraph::Shape sp = Ddim2Shape(dd);
-      auto ng_type = var_type_map_.at(var_name);
-      if (ng_type == ngraph::element::f32) {
-        auto pd_arr = tensor_pd->mutable_data<float>(place);
-        to = backend_->create_tensor(ngraph::element::f32, sp, pd_arr);
-      } else if (ng_type == ngraph::element::i64) {
-        auto pd_arr = tensor_pd->mutable_data<int64_t>(place);
-        to = backend_->create_tensor(ngraph::element::i64, sp, pd_arr);
-      } else if (ng_type == ngraph::element::f64) {
-        auto pd_arr = tensor_pd->mutable_data<double>(place);
-        to = backend_->create_tensor(ngraph::element::f64, sp, pd_arr);
-      } else if (ng_type == ngraph::element::boolean) {
-        auto pd_arr = tensor_pd->mutable_data<bool>(place);
-        to = backend_->create_tensor(ngraph::element::boolean, sp, pd_arr);
-      } else {
-        PADDLE_THROW("Data type not handled in for var %s", var_name);
-      }
-      t_out.push_back(to);
-    } else {
-      PADDLE_THROW("Cannot find var or tensor with var name %s", var_name);
-    }
-  }
-
-  backend_->call(backend_->compile(ngraph_function_), t_out, t_in);
-}  // NgraphEngine::RunImpl
-}  // namespace framework
-}  // namespace paddle

paddle/fluid/framework/operator.cc

@@ -555,18 +555,17 @@ Tensor* ExecutionContext::LegacyOutput<Tensor>(const std::string& name) const {
 template <>
 std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
     const std::string& name) const {
-  auto names = op().Outputs(name);
+  auto it = ctx_.outputs.find(name);
+  if (it == ctx_.outputs.end()) {
+    return {};
+  }
+  const std::vector<Variable*>& vars = it->second;
   std::vector<Tensor*> res;
-  res.reserve(names.size());
-  std::transform(names.begin(), names.end(), std::back_inserter(res),
-                 [&](const std::string& sub_name) -> Tensor* {
-                   auto var = scope_.FindVar(sub_name);
-                   if (var == nullptr) return nullptr;
-                   PADDLE_ENFORCE(
-                       var->IsType<LoDTensor>(),
-                       "%s should be LoDTensor, but the received type is %s",
-                       sub_name, ToTypeName(var->Type()));
-                   return var->GetMutable<LoDTensor>();
+  res.reserve(vars.size());
+  std::transform(vars.begin(), vars.end(), std::back_inserter(res),
+                 [&](Variable* var) -> Tensor* {
+                   return var == nullptr ? nullptr
+                                         : var->GetMutable<LoDTensor>();
                  });
   return res;
 }

paddle/fluid/imperative/layer.cc

@@ -156,6 +156,8 @@ class Autograd {
       for (auto it : candidate->pre_ops_) {
         for (OpBase* pre_op : it.second) {
           if (!pre_op) continue;
+          VLOG(5) << "op dep " << candidate->op_desc_->Type() << " <---- "
+                  << it.first << " <---- " << pre_op->op_desc_->Type();
           if (visited.find(pre_op) == visited.end()) {
             visited.insert(pre_op);
             queue.push_back(pre_op);
@@ -204,59 +206,68 @@ framework::LoDTensor& VarBase::GradValue() {
 }
 
 std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
-  if (!grad_op_desc_ && backward_id_ <= 0) {
+  if (grad_op_descs_.empty() && backward_id_ <= 0) {
     LOG(WARNING) << "op with no grad: " << op_desc_->Type();
     return {};
   }
 
-  std::map<std::string, std::vector<framework::Variable*>> grad_outputs;
+  std::vector<framework::VariableValueMap> grad_outputs;
   if (backward_id_ > 0) {
     VLOG(3) << "py_layer_grad";
-    grad_outputs[framework::GradVarName(PyLayer::kFwdOut)] = PyLayer::ApplyGrad(
-        backward_id_,
-        grad_input_vars_[framework::GradVarName(PyLayer::kFwdInp)]);
+    grad_outputs.resize(1);
+    grad_outputs[0][framework::GradVarName(PyLayer::kFwdOut)] =
+        PyLayer::ApplyGrad(
+            backward_id_,
+            grad_input_vars_[0][framework::GradVarName(PyLayer::kFwdInp)]);
   } else {
-    VLOG(3) << "op grad " << grad_op_desc_->Type();
-    for (auto it : grad_output_vars_) {
-      auto& outputs = grad_outputs[it.first];
-      for (size_t i = 0; i < it.second.size(); ++i) {
-        // Allocate a new variable
-        Variable* tmp_var = new framework::Variable();
-        tmp_var->GetMutable<framework::LoDTensor>();
-        outputs.push_back(tmp_var);
+    grad_outputs.resize(grad_op_descs_.size());
+    for (size_t k = 0; k < grad_op_descs_.size(); ++k) {
+      framework::OpDesc* grad_op_desc = grad_op_descs_[k];
+      VLOG(3) << "op grad " << grad_op_desc->Type();
+      for (auto it : grad_output_vars_[k]) {
+        auto& outputs = grad_outputs[k][it.first];
+        for (size_t i = 0; i < it.second.size(); ++i) {
+          // Allocate a new variable
+          Variable* tmp_var = new framework::Variable();
+          tmp_var->GetMutable<framework::LoDTensor>();
+          outputs.push_back(tmp_var);
+        }
       }
-    }
 
-    framework::RuntimeContext ctx(grad_input_vars_, grad_outputs);
+      framework::RuntimeContext ctx(grad_input_vars_[k], grad_outputs[k]);
 
-    // No need to do compile time infer shape here.
-    // grad_op_desc_->InferShape(*block_);
-    grad_op_desc_->InferVarType(block_);
+      // No need to do compile time infer shape here.
+      // grad_op_desc_->InferShape(*block_);
+      grad_op_desc->InferVarType(block_);
 
-    std::unique_ptr<framework::OperatorBase> opbase =
-        framework::OpRegistry::CreateOp(*grad_op_desc_);
-    framework::OperatorWithKernel* op_kernel =
-        dynamic_cast<framework::OperatorWithKernel*>(opbase.get());
-    PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
+      std::unique_ptr<framework::OperatorBase> opbase =
+          framework::OpRegistry::CreateOp(*grad_op_desc);
+      framework::OperatorWithKernel* op_kernel =
+          dynamic_cast<framework::OperatorWithKernel*>(opbase.get());
+      PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
 
-    framework::Scope scope;
-    PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place_);
-    p.op.RuntimeInferShape(scope, place_, ctx);
-    p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
+      framework::Scope scope;
+      PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place_);
+      p.op.RuntimeInferShape(scope, place_, ctx);
+      p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
+    }
   }
 
-  for (auto it : grad_output_vars_) {
-    auto& outputs = grad_outputs[it.first];
-    auto& origin_outputs = it.second;
-    PADDLE_ENFORCE_EQ(outputs.size(), origin_outputs.size());
-
-    for (size_t i = 0; i < outputs.size(); ++i) {
-      framework::Variable* grad = outputs[i];
-      framework::Variable* orig_grad = origin_outputs[i];
-      AddTo(grad, orig_grad, place_);
-      delete grad;
+  for (size_t k = 0; k < grad_output_vars_.size(); ++k) {
+    for (auto it : grad_output_vars_[k]) {
+      auto& outputs = grad_outputs[k][it.first];
+      auto& origin_outputs = it.second;
+      PADDLE_ENFORCE_EQ(outputs.size(), origin_outputs.size());
+
+      for (size_t i = 0; i < outputs.size(); ++i) {
+        framework::Variable* grad = outputs[i];
+        framework::Variable* orig_grad = origin_outputs[i];
+        AddTo(grad, orig_grad, place_);
+        delete grad;
+      }
     }
   }
+
   return input_vars_;
 }
 

paddle/fluid/imperative/layer.h

@@ -28,6 +28,7 @@
 #include "paddle/fluid/framework/var_desc.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/operators/math/math_function.h"
 
 #include "paddle/fluid/imperative/type_defs.h"
 
@@ -140,16 +141,24 @@ class VarBase {
   void RunBackward();
 
   void TrackPreOp(OpBase* pre_op, const std::string& pre_op_out_name,
-                  int pre_op_out_idx, bool stop_gradient) {
+                  int pre_op_out_idx, bool pre_op_stop_gradient) {
     pre_op_ = pre_op;
     pre_op_out_name_ = pre_op_out_name;
     pre_op_out_idx_ = pre_op_out_idx;
-    stop_gradient_ = stop_gradient;
+    if (pre_op_stop_gradient) {
+      stop_gradient_ = pre_op_stop_gradient;
+    }
   }
 
   void ClearGradient() {
-    delete grads_;
-    grads_ = new VarBase(true);
+    VLOG(1) << "clear gradient of " << var_desc_->Name();
+    if (grads_ && grads_->var_ && grads_->var_->IsInitialized()) {
+      auto grads_t = grads_->var_->GetMutable<framework::LoDTensor>();
+      operators::math::set_constant(
+          *(platform::DeviceContextPool::Instance().Get(
+              grads_->var_->Get<framework::LoDTensor>().place())),
+          grads_t, 0.0);
+    }
   }
 
   framework::LoDTensor& GradValue();
@@ -184,12 +193,13 @@ class OpBase {
   OpBase()
       : op_desc_(nullptr),
         forward_id_(-1),
-        grad_op_desc_(nullptr),
         backward_id_(-1),
         place_(platform::CPUPlace()) {}
 
   virtual ~OpBase() {
-    if (grad_op_desc_) delete grad_op_desc_;
+    for (framework::OpDesc* desc : grad_op_descs_) {
+      delete desc;
+    }
   }
 
   std::map<std::string, std::vector<VarBase*>> ApplyGrad();
@@ -198,9 +208,11 @@ class OpBase {
   // For pure python PyLayer, use `forward_id_`, otherwise, use op_desc_.
   framework::OpDesc* op_desc_;
   int forward_id_;
-  // When has backward, one of `grad_op_desc_` or `backward_id_` is set,
+
+  // When has backward, one of `grad_op_descs_` or `backward_id_` is set,
   // not both.
-  framework::OpDesc* grad_op_desc_;
+  // Note: each fwd op corresponds to a vector of bwd ops.
+  std::vector<framework::OpDesc*> grad_op_descs_;
   int backward_id_;
 
   platform::Place place_;
@@ -210,8 +222,11 @@ class OpBase {
   OpBasePtrMap pre_ops_;
   std::map<std::string, std::vector<int>> pre_ops_out_idx_;
 
-  framework::VariableValueMap grad_input_vars_;
-  framework::VariableValueMap grad_output_vars_;
+  // Inputs to a vector of bwd ops.
+  std::vector<framework::VariableValueMap> grad_input_vars_;
+  // Outputs to a vector of bwd ops.
+  std::vector<framework::VariableValueMap> grad_output_vars_;
+
   framework::BlockDesc* block_;
 };
 

paddle/fluid/imperative/tracer.cc

@@ -24,15 +24,17 @@ namespace imperative {
 void CreateGradOp(const framework::OpDesc& op_desc,
                   const std::unordered_set<std::string>& no_grad_set,
                   const std::vector<framework::BlockDesc*>& grad_sub_block,
-                  framework::OpDesc** grad_op_desc,
+                  std::vector<framework::OpDesc*>* grad_op_descs,
                   std::unordered_map<std::string, std::string>* grad_to_var) {
-  std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
+  PADDLE_ENFORCE(grad_op_descs->empty());
+  std::vector<std::unique_ptr<framework::OpDesc>> descs =
       framework::OpInfoMap::Instance()
           .Get(op_desc.Type())
           .GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
-  PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
-  // TODO(panyx0718): Leak?
-  *grad_op_desc = grad_op_descs[0].release();
+
+  for (auto& desc : descs) {
+    grad_op_descs->emplace_back(desc.release());
+  }
 }
 
 void InitVar(framework::Variable* var, framework::Variable* grad_var,
@@ -83,11 +85,12 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
   op->input_vars_ = inputs;
   for (auto it : op->input_vars_) {
     auto& invars = invars_map[it.first];
+    invars.reserve(it.second.size());
     for (VarBase* inp : it.second) {
       PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr",
                               op->op_desc_->Type(), inp->var_desc_->Name());
 
-      invars.push_back(inp->var_);
+      invars.emplace_back(inp->var_);
       vars[inp->var_desc_->Name()] = inp;
       if (inp->PreOp()) {
         op->pre_ops_[it.first].push_back(inp->PreOp());
@@ -104,9 +107,10 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
   for (auto it : op->output_vars_) {
     auto& outvars = outvars_map[it.first];
     const std::vector<VarBase*>& outputs = it.second;
+    outvars.reserve(outputs.size());
     for (size_t i = 0; i < outputs.size(); ++i) {
       VarBase* out = outputs[i];
-      outvars.push_back(out->var_);
+      outvars.emplace_back(out->var_);
       vars[out->var_desc_->Name()] = out;
 
       framework::VarDesc* var_desc = block->FindVar(out->var_desc_->Name());
@@ -138,49 +142,52 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
       prepared_op.op, scope, *prepared_op.dev_ctx, prepared_op.ctx));
 
   if (!stop_gradient) {
-    framework::OpDesc* grad_op_desc;
-    // TODO(panyx): Is this leaked?
     std::unique_ptr<std::unordered_map<std::string, std::string>> grad_to_var(
         new std::unordered_map<std::string, std::string>());
-    CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var.get());
-    op->grad_op_desc_ = grad_op_desc;
-
-    for (auto it : grad_op_desc->Inputs()) {
-      auto& grad_in_vars = op->grad_input_vars_[it.first];
-      for (const std::string& grad_invar : it.second) {
-        block->FindRecursiveOrCreateVar(grad_invar);
-        auto var_it = grad_to_var->find(grad_invar);
-        if (var_it == grad_to_var->end()) {
-          auto fwd_var_it = vars.find(grad_invar);
-          PADDLE_ENFORCE(fwd_var_it != vars.end());
-          // Forward inputs or outputs.
-          grad_in_vars.push_back(fwd_var_it->second->var_);
-        } else {
+    CreateGradOp(*op_desc, {}, {block}, &op->grad_op_descs_, grad_to_var.get());
+
+    op->grad_input_vars_.resize(op->grad_op_descs_.size());
+    op->grad_output_vars_.resize(op->grad_op_descs_.size());
+    for (size_t i = 0; i < op->grad_op_descs_.size(); ++i) {
+      framework::OpDesc* grad_op_desc = op->grad_op_descs_[i];
+      for (auto it : grad_op_desc->Inputs()) {
+        auto& grad_in_vars = op->grad_input_vars_[i][it.first];
+        for (const std::string& grad_invar : it.second) {
+          block->FindRecursiveOrCreateVar(grad_invar);
+          auto var_it = grad_to_var->find(grad_invar);
+          if (var_it == grad_to_var->end()) {
+            auto fwd_var_it = vars.find(grad_invar);
+            PADDLE_ENFORCE(fwd_var_it != vars.end());
+            // Forward inputs or outputs.
+            grad_in_vars.push_back(fwd_var_it->second->var_);
+          } else {
+            VarBase* var = vars[var_it->second];
+            if (!var->grads_->var_->IsInitialized()) {
+              InitVar(var->var_, var->grads_->var_,
+                      prepared_op.GetDeviceContext());
+            }
+            // Douts.
+            grad_in_vars.push_back(var->grads_->var_);
+          }
+        }
+      }
+
+      for (auto it : grad_op_desc->Outputs()) {
+        auto& grad_out_vars = op->grad_output_vars_[i][it.first];
+        for (const std::string& grad_outvar : it.second) {
+          block->FindRecursiveOrCreateVar(grad_outvar);
+          auto var_it = grad_to_var->find(grad_outvar);
+          PADDLE_ENFORCE(var_it != grad_to_var->end(),
+                         "Could not found the grad op output var, should this "
+                         "operator %s's stop gradient be True",
+                         op_desc->Type());
           VarBase* var = vars[var_it->second];
           if (!var->grads_->var_->IsInitialized()) {
             InitVar(var->var_, var->grads_->var_,
                     prepared_op.GetDeviceContext());
           }
-          // Douts.
-          grad_in_vars.push_back(var->grads_->var_);
-        }
-      }
-    }
-
-    for (auto it : grad_op_desc->Outputs()) {
-      auto& grad_out_vars = op->grad_output_vars_[it.first];
-      for (const std::string& grad_outvar : it.second) {
-        block->FindRecursiveOrCreateVar(grad_outvar);
-        auto var_it = grad_to_var->find(grad_outvar);
-        PADDLE_ENFORCE(var_it != grad_to_var->end(),
-                       "Could not found the grad op output var, should this "
-                       "operator %s's stop gradient be True",
-                       op_desc->Type());
-        VarBase* var = vars[var_it->second];
-        if (!var->grads_->var_->IsInitialized()) {
-          InitVar(var->var_, var->grads_->var_, prepared_op.GetDeviceContext());
+          grad_out_vars.push_back(var->grads_->var_);
         }
-        grad_out_vars.push_back(var->grads_->var_);
       }
     }
   }
@@ -209,10 +216,12 @@ std::vector<VarBase*> Tracer::PyTrace(OpBase* op,
     out->TrackPreOp(op, PyLayer::kFwdOut, i, stop_gradient);
   }
   if (!stop_gradient) {
+    op->grad_input_vars_.resize(1);
+    op->grad_output_vars_.resize(1);
     auto& grad_input_vars =
-        op->grad_input_vars_[framework::GradVarName(PyLayer::kFwdInp)];
+        op->grad_input_vars_[0][framework::GradVarName(PyLayer::kFwdInp)];
     auto& grad_output_vars =
-        op->grad_output_vars_[framework::GradVarName(PyLayer::kFwdOut)];
+        op->grad_output_vars_[0][framework::GradVarName(PyLayer::kFwdOut)];
 
     for (const VarBase* inp : inputs) {
       grad_input_vars.push_back(inp->var_);

paddle/fluid/inference/analysis/argument.h

@@ -28,6 +28,7 @@
 #include "paddle/fluid/framework/ir/graph.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/inference/api/paddle_analysis_config.h"
 #include "paddle/fluid/platform/variant.h"
 
 namespace paddle {
@@ -130,6 +131,8 @@ struct Argument {
   DECL_ARGUMENT_FIELD(tensorrt_max_batch_size, TensorRtMaxBatchSize, int);
   DECL_ARGUMENT_FIELD(tensorrt_workspace_size, TensorRtWorkspaceSize, int);
   DECL_ARGUMENT_FIELD(tensorrt_min_subgraph_size, TensorRtMinSubgraphSize, int);
+  DECL_ARGUMENT_FIELD(tensorrt_precision_mode, TensorRtPrecisionMode,
+                      AnalysisConfig::Precision);
 
   // Memory optimized related.
   DECL_ARGUMENT_FIELD(enable_memory_optim, EnableMemoryOptim, bool);

paddle/fluid/inference/analysis/helper.cc

@@ -36,6 +36,14 @@ void SetAttr<int>(framework::proto::OpDesc *op, const std::string &name,
   attr->set_i(data);
 }
 template <>
+void SetAttr<bool>(framework::proto::OpDesc *op, const std::string &name,
+                   const bool &data) {
+  auto *attr = op->add_attrs();
+  attr->set_name(name);
+  attr->set_type(paddle::framework::proto::AttrType::BOOLEAN);
+  attr->set_b(data);
+}
+template <>
 void SetAttr<int64_t>(framework::proto::OpDesc *op, const std::string &name,
                       const int64_t &data) {
   auto *attr = op->add_attrs();

paddle/fluid/inference/analysis/helper.h

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <sys/stat.h>
 #include <cstdio>
 #include <fstream>
+#include <set>
 #include <string>
 #include <typeindex>
 #include <unordered_map>
@@ -29,9 +30,14 @@ limitations under the License. */
 #include "paddle/fluid/platform/port.h"
 
 #ifdef _WIN32
-#define GCC_ATTRIBUTE(attr__) ;
+#include <direct.h>
+#include <io.h>
+#define GCC_ATTRIBUTE(attr__)
+#define MKDIR(path) _mkdir(path)
 #else
+#include <unistd.h>
 #define GCC_ATTRIBUTE(attr__) __attribute__((attr__));
+#define MKDIR(path) mkdir(path, S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH)
 #endif
 #define __SHOULD_USE_RESULT__ GCC_ATTRIBUTE(warn_unused_result)
 
@@ -163,6 +169,54 @@ static bool PathExists(const std::string &path) {
   return false;
 }
 
+static std::string GetDirRoot(const std::string &path) {
+  char sep = '/';
+
+#ifdef _WIN32
+  sep = '\\';
+#endif
+
+  size_t i = path.rfind(sep, path.length());
+  if (i != std::string::npos) {
+    return (path.substr(0, i));
+  }
+  return path;
+}
+
+static std::string GetOrCreateModelOptCacheDir(const std::string &model_root) {
+  std::string opt_cache_dir = model_root + "/_opt_cache/";
+  if (!PathExists(opt_cache_dir)) {
+    PADDLE_ENFORCE(MKDIR(opt_cache_dir.c_str()) != -1,
+                   "Can not create optimize cache directory: %s, Make sure you "
+                   "have permission to write",
+                   opt_cache_dir);
+  }
+  return opt_cache_dir;
+}
+
+static std::string GetTrtCalibPath(const std::string &model_root,
+                                   const std::string &engine_key) {
+  return model_root + "/trt_calib_" + engine_key;
+}
+
+// If there is no calib table data file in model_opt_cache_dir, return "".
+static std::string GetTrtCalibTableData(const std::string &model_opt_cache_dir,
+                                        const std::string &engine_key,
+                                        bool enable_int8) {
+  std::string trt_calib_table_path =
+      GetTrtCalibPath(model_opt_cache_dir, engine_key);
+  if (enable_int8 && FileExists(trt_calib_table_path)) {
+    VLOG(3) << "Calibration table file: " << trt_calib_table_path
+            << "is found here";
+    std::ifstream infile(trt_calib_table_path, std::ios::in);
+    std::stringstream buffer;
+    buffer << infile.rdbuf();
+    std::string calibration_data(buffer.str());
+    return calibration_data;
+  }
+  return "";
+}
+
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -67,6 +67,20 @@ void IRPassManager::CreatePasses(Argument *argument,
       pass->Set("max_batch_size", new int(argument->tensorrt_max_batch_size()));
       pass->Set("min_subgraph_size",
                 new int(argument->tensorrt_min_subgraph_size()));
+      pass->Set("program",
+                new framework::ProgramDesc *(&argument->main_program()));
+
+      bool enable_int8 = argument->tensorrt_precision_mode() ==
+                         AnalysisConfig::Precision::kInt8;
+
+      pass->Set("enable_int8", new bool(enable_int8));
+      std::string model_opt_cache_dir =
+          argument->Has("model_dir")
+              ? argument->model_dir()
+              : GetDirRoot(argument->model_program_path());
+      pass->Set(
+          "model_opt_cache_dir",
+          new std::string(GetOrCreateModelOptCacheDir(model_opt_cache_dir)));
     }
 
     // graph_ = pass->Apply(std::move(graph_));
@@ -91,11 +105,14 @@ std::unique_ptr<Graph> IRPassManager::Apply(std::unique_ptr<Graph> graph) {
 }
 
 framework::proto::ProgramDesc IRPassManager::AcquireProgram(
-    std::unique_ptr<Graph> *graph, const ProgramDesc &program) const {
+    std::unique_ptr<Graph> *graph, ProgramDesc *program) const {
   auto pass =
       framework::ir::PassRegistry::Instance().Get("graph_to_program_pass");
 
-  ProgramDesc desc(program);
+  // Direct using ProgramDesc desc(argument->main_program()) may cause
+  // incomplete copies of information.
+  ProgramDesc desc;
+  desc.CopyFrom(*program->Proto());
   pass->SetNotOwned("program", &desc);
   auto *the_graph = graph->release();
   *graph = pass->Apply(std::unique_ptr<Graph>(the_graph));

paddle/fluid/inference/analysis/ir_pass_manager.h

@@ -29,6 +29,7 @@
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/analysis/argument.h"
+#include "paddle/fluid/inference/analysis/helper.h"
 
 namespace paddle {
 namespace inference {
@@ -42,8 +43,8 @@ class IRPassManager final {
 
   std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph);
 
-  framework::proto::ProgramDesc AcquireProgram(
-      std::unique_ptr<Graph> *graph, const ProgramDesc &program) const;
+  framework::proto::ProgramDesc AcquireProgram(std::unique_ptr<Graph> *graph,
+                                               ProgramDesc *program) const;
 
   framework::ir::Graph &graph() const { return *graph_; }
 

paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include <algorithm>
+#include <set>
 #include <string>
 #include <vector>
 
@@ -67,12 +68,33 @@ std::unique_ptr<framework::ir::Graph> analysis::TensorRtSubgraphPass::ApplyImpl(
   return graph;
 }
 
+std::string GenerateEngineKey(const std::set<std::string> &engine_inputs,
+                              const std::set<std::string> &engine_outputs) {
+  std::string engine_hash_key = "";
+  for (auto name : engine_inputs) {
+    engine_hash_key += name;
+  }
+  for (auto name : engine_outputs) {
+    engine_hash_key += name;
+  }
+  auto engine_key = std::to_string(std::hash<std::string>()(engine_hash_key));
+  return engine_key;
+}
+
 void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
                                             Graph *graph) const {
   auto *op_desc = node->Op();
   auto &subgraph = *Agent(node).subgraph();
   PADDLE_ENFORCE(!subgraph.empty());
 
+  framework::ProgramDesc *program_desc =
+      Get<framework::ProgramDesc *>("program");
+  // Add new block for TensorRTEngineOP
+  const framework::BlockDesc &main_block =
+      program_desc->Block(framework::kRootBlockIndex);
+  // const framework::BlockDesc& main_block = program_desc->Block(0);
+  framework::BlockDesc *new_block = program_desc->AppendBlock(main_block);
+
   // An fake block desc.
   framework::proto::BlockDesc block_proto;
   framework::BlockDesc block_desc(nullptr, &block_proto);
@@ -82,13 +104,18 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
                           subgraph.size());
 
   for (auto *node : subgraph) {
+    auto *new_block_op = new_block->AppendOp();
     auto *op = block_desc.AppendOp();
+    *new_block_op->Proto() = *node->Op()->Proto();
     *op->Proto() = *node->Op()->Proto();
   }
 
-  // collect inputs
-  std::unordered_set<std::string> input_names;
-  std::unordered_set<std::string> input_names_with_id;
+  // Then, we will use the input_names_with_id and output_names_with_id to
+  // generate the eigine key.
+  // So, We use set instead of unordered_set here to ensure that the engine key
+  // is unique.
+  std::set<std::string> input_names;
+  std::set<std::string> input_names_with_id;
   for (auto *x : node->inputs) {
     input_names.insert(x->Name());
     input_names_with_id.insert(x->Name() + std::to_string(x->id()));
@@ -96,8 +123,8 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
   op_desc->SetInput(
       "Xs", std::vector<std::string>(input_names.begin(), input_names.end()));
 
-  std::unordered_set<std::string> output_names;
-  std::unordered_set<std::string> output_names_with_id;
+  std::set<std::string> output_names;
+  std::set<std::string> output_names_with_id;
   for (auto *x : node->outputs) {
     output_names.insert(x->Name());
     output_names_with_id.insert(x->Name() + std::to_string(x->id()));
@@ -182,7 +209,6 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
   // to Tensor.
   std::vector<std::string> output_mapping;
   for (auto name : output_names) {
-    // LOG(INFO) << name << " " << output_name_map.size();
     PADDLE_ENFORCE(output_name_map.count(name) != 0);
     output_mapping.push_back(output_name_map[name]);
   }
@@ -193,16 +219,29 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
       *vars->Add() = *node->Var()->Proto();
     }
   }
+
   PADDLE_ENFORCE(!block_desc.Proto()->vars().empty(),
                  "the block has no var-desc");
   PADDLE_ENFORCE(!output_mapping.empty());
-  // Set attrs
+  op_desc->SetBlockAttr("sub_block", new_block);
   SetAttr(op_desc->Proto(), "subgraph",
           block_desc.Proto()->SerializeAsString());
+  // Set attrs
   SetAttr(op_desc->Proto(), "max_batch_size", Get<int>("max_batch_size"));
   SetAttr(op_desc->Proto(), "workspace_size", Get<int>("workspace_size"));
   SetAttr(op_desc->Proto(), "parameters", ExtractParameters(graph->Nodes()));
   SetAttr(op_desc->Proto(), "output_name_mapping", output_mapping);
+
+  auto enable_int8 = Get<bool>("enable_int8");
+  auto engine_key =
+      GenerateEngineKey(input_names_with_id, output_names_with_id);
+
+  std::string calibration_data = GetTrtCalibTableData(
+      Get<std::string>("model_opt_cache_dir"), engine_key, enable_int8);
+  SetAttr(op_desc->Proto(), "calibration_data", calibration_data);
+
+  SetAttr(op_desc->Proto(), "enable_int8", enable_int8);
+  SetAttr(op_desc->Proto(), "engine_key", engine_key);
 }
 
 std::vector<std::string> ExtractParameters(

paddle/fluid/inference/analysis/passes/ir_graph_to_program_pass.cc

@@ -31,7 +31,11 @@ void IrGraphToProgramPass::RunImpl(Argument *argument) {
   }
 
   std::unique_ptr<Graph> graph(argument->main_graph_ptr());
-  framework::ProgramDesc desc(argument->main_program());
+
+  // Direct using ProgramDesc desc(argument->main_program()) may cause
+  // incomplete copies of information.
+  framework::ProgramDesc desc;
+  desc.CopyFrom(*argument->main_program().Proto());
   pass->SetNotOwned("program", &desc);
   auto thegraph = pass->Apply(std::move(graph));
   thegraph.release();  // the argument still own the graph.

paddle/fluid/inference/analysis/passes/memory_optimize_pass.h

@@ -13,7 +13,9 @@
 // limitations under the License.
 
 #pragma once
-
+#include <string>
+#include <utility>
+#include <vector>
 #include "paddle/fluid/inference/analysis/analysis_pass.h"
 #include "paddle/fluid/platform/port.h"
 

paddle/fluid/inference/api/analysis_config.cc

@@ -22,7 +22,7 @@
 
 namespace paddle {
 
-PassStrategy *contrib::AnalysisConfig::pass_builder() const {
+PassStrategy *AnalysisConfig::pass_builder() const {
   if (!pass_builder_.get()) {
     if (use_gpu_) {
       LOG(INFO) << "Create GPU IR passes";
@@ -42,27 +42,27 @@ PassStrategy *contrib::AnalysisConfig::pass_builder() const {
   return pass_builder_.get();
 }
 
-contrib::AnalysisConfig::AnalysisConfig(const std::string &model_dir) {
+AnalysisConfig::AnalysisConfig(const std::string &model_dir) {
   model_dir_ = model_dir;
 
   Update();
 }
-contrib::AnalysisConfig::AnalysisConfig(const std::string &prog_file,
-                                        const std::string &params_file) {
+AnalysisConfig::AnalysisConfig(const std::string &prog_file,
+                               const std::string &params_file) {
   prog_file_ = prog_file;
   params_file_ = params_file;
 
   Update();
 }
-void contrib::AnalysisConfig::SetModel(const std::string &prog_file_path,
-                                       const std::string &params_file_path) {
+void AnalysisConfig::SetModel(const std::string &prog_file_path,
+                              const std::string &params_file_path) {
   prog_file_ = prog_file_path;
   params_file_ = params_file_path;
 
   Update();
 }
-void contrib::AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
-                                           int device_id) {
+void AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
+                                  int device_id) {
 #ifdef PADDLE_WITH_CUDA
   use_gpu_ = true;
   memory_pool_init_size_mb_ = memory_pool_init_size_mb;
@@ -74,13 +74,13 @@ void contrib::AnalysisConfig::EnableUseGpu(uint64_t memory_pool_init_size_mb,
 
   Update();
 }
-void contrib::AnalysisConfig::DisableGpu() {
+void AnalysisConfig::DisableGpu() {
   use_gpu_ = false;
 
   Update();
 }
 
-contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) {
+AnalysisConfig::AnalysisConfig(const AnalysisConfig &other) {
 #define CP_MEMBER(member__) member__ = other.member__;
 
   // Model related.
@@ -102,6 +102,7 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) {
   CP_MEMBER(tensorrt_workspace_size_);
   CP_MEMBER(tensorrt_max_batchsize_);
   CP_MEMBER(tensorrt_min_subgraph_size_);
+  CP_MEMBER(tensorrt_precision_mode_);
   // MKLDNN releated.
   CP_MEMBER(use_mkldnn_);
   CP_MEMBER(mkldnn_enabled_op_types_);
@@ -129,7 +130,7 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) {
   Update();
 }
 
-void contrib::AnalysisConfig::EnableMKLDNN() {
+void AnalysisConfig::EnableMKLDNN() {
 #ifdef PADDLE_WITH_MKLDNN
   pass_builder()->EnableMKLDNN();
   use_mkldnn_ = true;
@@ -141,9 +142,9 @@ void contrib::AnalysisConfig::EnableMKLDNN() {
   Update();
 }
 
-void contrib::AnalysisConfig::EnableTensorRtEngine(int workspace_size,
-                                                   int max_batch_size,
-                                                   int min_subgraph_size) {
+void AnalysisConfig::EnableTensorRtEngine(
+    int workspace_size, int max_batch_size, int min_subgraph_size,
+    AnalysisConfig::Precision precision_mode) {
 #ifdef PADDLE_WITH_CUDA
   if (!use_gpu()) {
     LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first";
@@ -154,6 +155,7 @@ void contrib::AnalysisConfig::EnableTensorRtEngine(int workspace_size,
   tensorrt_workspace_size_ = workspace_size;
   tensorrt_max_batchsize_ = max_batch_size;
   tensorrt_min_subgraph_size_ = min_subgraph_size;
+  tensorrt_precision_mode_ = precision_mode;
 
   Update();
 #else
@@ -163,7 +165,7 @@ void contrib::AnalysisConfig::EnableTensorRtEngine(int workspace_size,
 }
 
 // TODO(Superjomn) refactor this, buggy.
-void contrib::AnalysisConfig::Update() {
+void AnalysisConfig::Update() {
   auto info = SerializeInfoCache();
   if (info == serialized_info_cache_) return;
 
@@ -223,7 +225,7 @@ void contrib::AnalysisConfig::Update() {
   }
 }
 
-std::string contrib::AnalysisConfig::SerializeInfoCache() {
+std::string AnalysisConfig::SerializeInfoCache() {
   std::stringstream ss;
   ss << model_dir_;
   ss << prog_file_;
@@ -258,14 +260,14 @@ std::string contrib::AnalysisConfig::SerializeInfoCache() {
   return ss.str();
 }
 
-void contrib::AnalysisConfig::SetCpuMathLibraryNumThreads(
+void AnalysisConfig::SetCpuMathLibraryNumThreads(
     int cpu_math_library_num_threads) {
   cpu_math_library_num_threads_ = cpu_math_library_num_threads;
 
   Update();
 }
 
-float contrib::AnalysisConfig::fraction_of_gpu_memory_for_pool() const {
+float AnalysisConfig::fraction_of_gpu_memory_for_pool() const {
 #ifdef PADDLE_WITH_CUDA
   // Get the GPU memory details and calculate the fraction of memory for the
   // GPU memory pool.
@@ -280,8 +282,8 @@ float contrib::AnalysisConfig::fraction_of_gpu_memory_for_pool() const {
 #endif
 }
 
-void contrib::AnalysisConfig::EnableMemoryOptim(
-    bool static_optim, bool force_update_static_cache) {
+void AnalysisConfig::EnableMemoryOptim(bool static_optim,
+                                       bool force_update_static_cache) {
   enable_memory_optim_ = true;
   static_memory_optim_ = static_optim;
   static_memory_optim_force_update_ = force_update_static_cache;
@@ -289,14 +291,14 @@ void contrib::AnalysisConfig::EnableMemoryOptim(
   Update();
 }
 
-bool contrib::AnalysisConfig::enable_memory_optim() const {
+bool AnalysisConfig::enable_memory_optim() const {
   return enable_memory_optim_;
 }
 
-void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer,
-                                             size_t prog_buffer_size,
-                                             const char *param_buffer,
-                                             size_t param_buffer_size) {
+void AnalysisConfig::SetModelBuffer(const char *prog_buffer,
+                                    size_t prog_buffer_size,
+                                    const char *param_buffer,
+                                    size_t param_buffer_size) {
   prog_file_ = std::string(prog_buffer, prog_buffer + prog_buffer_size);
   params_file_ = std::string(param_buffer, param_buffer + param_buffer_size);
   model_from_memory_ = true;
@@ -304,7 +306,7 @@ void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer,
   Update();
 }
 
-NativeConfig contrib::AnalysisConfig::ToNativeConfig() const {
+NativeConfig AnalysisConfig::ToNativeConfig() const {
   NativeConfig config;
   config.model_dir = model_dir_;
   config.prog_file = prog_file_;

paddle/fluid/inference/api/analysis_predictor.cc

@@ -15,6 +15,7 @@
 #include "paddle/fluid/inference/api/analysis_predictor.h"
 #include <glog/logging.h>
 #include <algorithm>
+#include <fstream>
 #include <memory>
 #include <string>
 #include <vector>
@@ -25,6 +26,7 @@
 #include "paddle/fluid/framework/naive_executor.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/var_type_traits.h"
+#include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/analysis/passes/memory_optimize_pass.h"
 #include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
@@ -37,13 +39,20 @@
 
 #if PADDLE_WITH_TENSORRT
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
+#include "paddle/fluid/inference/tensorrt/trt_int8_calibrator.h"
+
 #endif
 
 DECLARE_bool(profile);
 
 namespace paddle {
 
-using contrib::AnalysisConfig;
+using inference::Singleton;
+#if PADDLE_WITH_TENSORRT
+using inference::tensorrt::TRTInt8Calibrator;
+using inference::tensorrt::TRTCalibratorEngine;
+using inference::tensorrt::TRTCalibratorEngineManager;
+#endif
 
 namespace {
 bool IsPersistable(const framework::VarDesc *var) {
@@ -113,6 +122,15 @@ bool AnalysisPredictor::PrepareProgram(
   if (!program) {
     if (!LoadProgramDesc()) return false;
 
+    // If not cloned, the parameters should be loaded.
+    // If config_.ir_optim() is True, parameters is loaded in
+    // OptimizeInferenceProgram(), but other persistable variables
+    // (like RAW type var) are not created in scope.
+    // If config_.ir_optim() is False, parameters is loaded in LoadParameters(),
+    // still need to create other persistable variables.
+    // So in both case, create persistable variables at first.
+    executor_->CreateVariables(*inference_program_, 0, true, sub_scope_);
+
     // Optimize the program, and load parameters and modify them in the
     // scope_.
     // This will change the scope_ address.
@@ -120,15 +138,6 @@ bool AnalysisPredictor::PrepareProgram(
       status_ir_optim_enabled_ = true;
       OptimizeInferenceProgram();
     } else {
-      // If the parent_scope is passed, we assert that the persistable variables
-      // are already created, so just create the no persistable variables.
-
-      // If not cloned, the parameters should be loaded
-      // OptimizeInferenceProgram.
-      // So in both cases, just the local variables are needed to load, not the
-      // parematers.
-      executor_->CreateVariables(*inference_program_, 0, true, sub_scope_);
-
       // Load parameters
       LOG(INFO) << "load parameters ";
       LoadParameters();
@@ -339,6 +348,8 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
         !config_.params_file().empty(),
         "Either model_dir or (param_file, prog_file) should be set.");
     PADDLE_ENFORCE(!config_.prog_file().empty());
+    std::string dir = inference::analysis::GetDirRoot(config_.prog_file());
+
     argument_.SetModelProgramPath(config_.prog_file());
     argument_.SetModelParamsPath(config_.params_file());
   }
@@ -349,6 +360,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
     argument_.SetTensorRtWorkspaceSize(config_.tensorrt_workspace_size_);
     argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_);
     argument_.SetTensorRtMinSubgraphSize(config_.tensorrt_min_subgraph_size_);
+    argument_.SetTensorRtPrecisionMode(config_.tensorrt_precision_mode_);
   }
 
   if (config_.use_mkldnn_) {
@@ -363,7 +375,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
   }
   argument_.SetIrAnalysisPasses(passes);
   argument_.SetAnalysisPasses(config_.pass_builder()->AnalysisPasses());
-  argument_.SetScopeNotOwned(const_cast<framework::Scope *>(scope_.get()));
+  argument_.SetScopeNotOwned(scope_.get());
   Analyzer().Run(&argument_);
 
   PADDLE_ENFORCE(argument_.scope_valid());
@@ -569,7 +581,67 @@ bool AnalysisPredictor::LoadParameters() {
   return true;
 }
 
+#if PADDLE_WITH_TENSORRT
+bool AnalysisPredictor::SaveTrtCalibToDisk() {
+  PADDLE_ENFORCE(config_.tensorrt_engine_enabled(),
+                 "This func can be invoked only in trt mode");
+  auto &block = inference_program_->Block(0);
+  for (auto &op_desc : block.AllOps()) {
+    if (op_desc->Type() == "tensorrt_engine") {
+      std::string engine_name =
+          boost::get<std::string>(op_desc->GetAttr("engine_key"));
+      if (!Singleton<TRTCalibratorEngineManager>::Global().Has(engine_name)) {
+        LOG(ERROR) << "You should run the predictor(with trt) on the real data "
+                      "to generate calibration info";
+        return false;
+      }
+      TRTCalibratorEngine *calib_engine =
+          Singleton<TRTCalibratorEngineManager>::Global().Get(engine_name);
+      LOG(INFO) << "Wait for calib threads done.";
+      calib_engine->calib_->waitAndSetDone();
+      LOG(INFO) << "Generating TRT Calibration table data, this may cost a lot "
+                   "of time...";
+      calib_engine->thr_->join();
+      std::string calibration_table_data =
+          calib_engine->calib_->getCalibrationTableAsString();
+
+      if (calibration_table_data.empty()) {
+        LOG(ERROR) << "the calibration table is empty.";
+        return false;
+      }
+
+      std::string model_opt_cache_dir =
+          argument_.Has("model_dir")
+              ? argument_.model_dir()
+              : inference::analysis::GetDirRoot(argument_.model_program_path());
+
+      std::string calibration_table_data_path =
+          inference::analysis::GetTrtCalibPath(
+              inference::analysis::GetOrCreateModelOptCacheDir(
+                  model_opt_cache_dir),
+              engine_name);
+
+      std::ofstream ofile(calibration_table_data_path, std::ios::out);
+      LOG(INFO) << "Write Paddle-TRT INT8 calibration table data to file "
+                << calibration_table_data_path;
+      ofile << calibration_table_data;
+      ofile.close();
+    }
+  }
+  // Free all calibrator resources.
+  Singleton<TRTCalibratorEngineManager>::Global().DeleteALL();
+  return true;
+}
+#endif
+
 AnalysisPredictor::~AnalysisPredictor() {
+#if PADDLE_WITH_TENSORRT
+  if (config_.tensorrt_engine_enabled() &&
+      config_.tensorrt_precision_mode_ == AnalysisConfig::Precision::kInt8 &&
+      Singleton<TRTCalibratorEngineManager>::Global().Has()) {
+    SaveTrtCalibToDisk();
+  }
+#endif
   if (FLAGS_profile) {
     platform::DisableProfiler(platform::EventSortingKey::kTotal,
                               "./profile.log");
@@ -653,11 +725,15 @@ bool AnalysisPredictor::need_collect_var_shapes_for_memory_optim() {
   return need;
 }
 
+std::string AnalysisPredictor::GetSeriazlizedProgram() const {
+  return inference_program_->Proto()->SerializeAsString();
+}
+
 template <>
-std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<contrib::AnalysisConfig>(
-    const contrib::AnalysisConfig &config) {
-  return CreatePaddlePredictor<contrib::AnalysisConfig,
-                               PaddleEngineKind::kAnalysis>(config);
+std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<AnalysisConfig>(
+    const AnalysisConfig &config) {
+  return CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
+      config);
 }
 
 }  // namespace paddle

paddle/fluid/inference/api/analysis_predictor.h

@@ -33,7 +33,6 @@ using inference::analysis::Argument;
 using inference::analysis::Analyzer;
 using framework::proto::ProgramDesc;
 using framework::NaiveExecutor;
-using contrib::AnalysisConfig;
 
 /** \brief This predictor is based on the original native predictor with IR and
  * Analysis support.
@@ -75,6 +74,8 @@ class AnalysisPredictor : public PaddlePredictor {
 
   void SetMkldnnThreadID(int tid);
 
+  std::string GetSeriazlizedProgram() const override;
+
  protected:
   // For memory optimization.
   bool need_collect_var_shapes_for_memory_optim();
@@ -97,6 +98,21 @@ class AnalysisPredictor : public PaddlePredictor {
   void GetFetchOne(const framework::LoDTensor &fetchs,
                    PaddleTensor *output_data);
 
+#if PADDLE_WITH_TENSORRT
+  // When we use Paddle-TRT INT8 engine, we need to generate calibration table
+  // data first,
+  // the calibration table contains the range for each op's input and output,
+  // this whole process can be divided into several steps:
+  //
+  // 1. Builds a 32-bit engine, runs it on the calibration set, and records a
+  // histogram for each
+  // tensor of the distribution of activation values.
+  // 2. Builds a calibration table from the histograms.
+  //
+  // After step 2, we need to store the calibration table on disk
+  bool SaveTrtCalibToDisk();
+#endif
+
 // Some more detailed tests, they are made the friends of the predictor, so that
 // the all the details can be tested.
 #if PADDLE_WITH_TESTING
@@ -106,7 +122,7 @@ class AnalysisPredictor : public PaddlePredictor {
 #endif
 
  private:
-  contrib::AnalysisConfig config_;
+  AnalysisConfig config_;
   Argument argument_;
   std::unique_ptr<NaiveExecutor> executor_;
   platform::Place place_;

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -24,7 +24,6 @@
 DEFINE_string(dirname, "", "dirname to tests.");
 
 namespace paddle {
-using contrib::AnalysisConfig;
 
 TEST(AnalysisPredictor, analysis_off) {
   AnalysisConfig config;
@@ -215,6 +214,8 @@ TEST(AnalysisPredictor, memory_optim) {
   {
     // The first predictor help to cache the memory optimize strategy.
     auto predictor = CreatePaddlePredictor<AnalysisConfig>(config);
+    LOG(INFO) << "serialized program: " << predictor->GetSeriazlizedProgram();
+    ASSERT_FALSE(predictor->GetSeriazlizedProgram().empty());
 
     // Run several times to check the parameters are not reused by mistake.
     for (int i = 0; i < 5; i++) {

paddle/fluid/inference/api/api.cc

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <sstream>
+#include "paddle/fluid/framework/commit.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
@@ -97,4 +99,12 @@ void PaddleBuf::Free() {
   }
 }
 
+std::string get_version() {
+  std::stringstream ss;
+  ss << "version: " << framework::paddle_version() << "\n";
+  ss << "commit: " << framework::paddle_commit() << "\n";
+  ss << "branch: " << framework::paddle_compile_branch() << "\n";
+  return ss.str();
+}
+
 }  // namespace paddle

paddle/fluid/inference/api/api_impl_tester.cc

@@ -295,7 +295,7 @@ TEST(inference_api_native, image_classification_gpu) {
 #endif
 
 TEST(PassBuilder, Delete) {
-  contrib::AnalysisConfig config;
+  AnalysisConfig config;
   config.DisableGpu();
   config.pass_builder()->DeletePass("attention_lstm_fuse_pass");
   const auto& passes = config.pass_builder()->AllPasses();

paddle/fluid/inference/api/api_tester.cc

@@ -61,4 +61,10 @@ TEST(paddle_inference_api, demo) {
   predictor->Run({}, &outputs);
 }
 
+TEST(paddle_inference_api, get_version) {
+  LOG(INFO) << "paddle version:\n" << get_version();
+  auto version = get_version();
+  ASSERT_FALSE(version.empty());
+}
+
 }  // namespace paddle

paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc

@@ -36,7 +36,7 @@ namespace demo {
  */
 void Main() {
   std::unique_ptr<PaddlePredictor> predictor;
-  paddle::contrib::AnalysisConfig config;
+  paddle::AnalysisConfig config;
   config.EnableUseGpu(100, 0);
   config.SetModel(FLAGS_modeldir + "/__model__",
                   FLAGS_modeldir + "/__params__");

paddle/fluid/inference/api/demo_ci/vis_demo.cc

@@ -34,7 +34,6 @@ DEFINE_bool(use_gpu, false, "Whether use gpu.");
 namespace paddle {
 namespace demo {
 
-using contrib::AnalysisConfig;
 /*
  * Use the native and analysis fluid engine to inference the demo.
  */

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -29,11 +29,6 @@
 namespace paddle {
 
 class AnalysisPredictor;
-// ==
-//
-// -----------------------------------------------------------------------------------
-// NOTE: The following APIs are not mature yet, we are still working on them.
-namespace contrib {
 
 // NOTE WIP, not stable yet.
 struct AnalysisConfig {
@@ -42,6 +37,10 @@ struct AnalysisConfig {
   explicit AnalysisConfig(const std::string& model_dir);
   explicit AnalysisConfig(const std::string& prog_file,
                           const std::string& params_file);
+  enum class Precision {
+    kFloat32 = 0,
+    kInt8,
+  };
 
   /** Set model with a directory.
    */
@@ -135,7 +134,8 @@ struct AnalysisConfig {
    * subgraph is less than this, it will not transfer to TensorRT engine.
    */
   void EnableTensorRtEngine(int workspace_size = 1 << 20,
-                            int max_batch_size = 1, int min_subgraph_size = 3);
+                            int max_batch_size = 1, int min_subgraph_size = 3,
+                            Precision precision = Precision::kFloat32);
   /** A boolean state telling whether the TensorRT engine is used.
    */
   bool tensorrt_engine_enabled() const { return use_tensorrt_; }
@@ -229,6 +229,7 @@ struct AnalysisConfig {
   //  We set this variable to control the minimum number of nodes in the
   //  subgraph, 3 as default value.
   int tensorrt_min_subgraph_size_{3};
+  Precision tensorrt_precision_mode_;
 
   // memory reuse related.
   bool enable_memory_optim_{false};
@@ -254,5 +255,4 @@ struct AnalysisConfig {
   mutable std::unique_ptr<PassStrategy> pass_builder_;
 };
 
-}  // namespace contrib
 }  // namespace paddle

paddle/fluid/inference/api/paddle_api.h

@@ -215,6 +215,14 @@ class PaddlePredictor {
    */
   virtual ~PaddlePredictor() = default;
 
+  /** \brief Get the serialized model program that executes in inference phase.
+   * Its data type is ProgramDesc, which is a protobuf message.
+   */
+  virtual std::string GetSeriazlizedProgram() const {
+    assert(false);  // Force raise error.
+    return "NotImplemented";
+  }
+
   /** The common configs for all the predictors.
    */
   struct Config {
@@ -288,4 +296,6 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT& config);
 
 int PaddleDtypeSize(PaddleDType dtype);
 
+std::string get_version();
+
 }  // namespace paddle

paddle/fluid/inference/tensorrt/CMakeLists.txt

-nv_library(tensorrt_engine SRCS engine.cc DEPS ${GLOB_OPERATOR_DEPS} framework_proto device_context)
+nv_library(tensorrt_engine SRCS engine.cc trt_int8_calibrator.cc DEPS ${GLOB_OPERATOR_DEPS} framework_proto device_context)
 nv_library(tensorrt_op_teller SRCS op_teller.cc DEPS framework_proto)
 nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
 nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine)

paddle/fluid/inference/tensorrt/engine.cc

@@ -69,6 +69,13 @@ void TensorRTEngine::FreezeNetwork() {
   // build engine.
   infer_builder_->setMaxBatchSize(max_batch_);
   infer_builder_->setMaxWorkspaceSize(max_workspace_);
+  if (enable_int8_) {
+    infer_builder_->setInt8Mode(true);
+    PADDLE_ENFORCE(
+        calibrator_ != nullptr,
+        "The precision mode is 'INT8', the calibrator should not be nullptr");
+    infer_builder_->setInt8Calibrator(calibrator_);
+  }
 
   infer_engine_.reset(infer_builder_->buildCudaEngine(*infer_network_));
   PADDLE_ENFORCE(infer_engine_ != nullptr, "build cuda engine failed!");

paddle/fluid/inference/tensorrt/engine.h

@@ -23,12 +23,14 @@ limitations under the License. */
 #include "paddle/fluid/inference/engine.h"
 #include "paddle/fluid/inference/tensorrt/helper.h"
 #include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
+#include "paddle/fluid/inference/tensorrt/trt_int8_calibrator.h"
 #include "paddle/fluid/inference/utils/singleton.h"
 
 namespace paddle {
 namespace inference {
 namespace tensorrt {
 
+class TRTInt8Calibrator;
 /*
  * TensorRT Engine.
  *
@@ -55,13 +57,16 @@ class TensorRTEngine : public EngineBase {
   };
 
   TensorRTEngine(int max_batch, int max_workspace, cudaStream_t stream,
-                 int device = 0,
+                 int device = 0, bool enable_int8 = false,
+                 TRTInt8Calibrator* calibrator = nullptr,
                  nvinfer1::ILogger& logger = NaiveLogger::Global())
       : max_batch_(max_batch),
         max_workspace_(max_workspace),
         stream_(stream),
-        logger_(logger),
-        device_(device) {}
+        device_(device),
+        enable_int8_(enable_int8),
+        calibrator_(calibrator),
+        logger_(logger) {}
 
   virtual ~TensorRTEngine();
 
@@ -139,8 +144,8 @@ class TensorRTEngine : public EngineBase {
   // In the normal case, the paddle-trt exists bug when runing the googlenet.
   // When there are more than two convolutions of 1 * 1 with the same input, the
   // paddle-tensorrt will do the merging optimization, which fuse those conv
-  // into
-  // one conv, and then trigger bug. So,  We should use strategy to avoid this
+  // into one conv, and then trigger bug. So,  We should use strategy to avoid
+  // this
   // optimization for the time being. This bug will be fixed in the future.
   std::unordered_map<std::string /*name*/, int /*ITensor_quote_num*/>
       itensor_quote_num;
@@ -153,9 +158,14 @@ class TensorRTEngine : public EngineBase {
   // the max memory size the engine uses
   int max_workspace_;
 
+  cudaStream_t stream_;
+  // The specific GPU id that the TensorRTEngine bounded to.
+  int device_;
+
+  bool enable_int8_;
+  TRTInt8Calibrator* calibrator_;
   // batch size of the current data, will be updated each Executation.
   int batch_size_{-1};
-  cudaStream_t stream_;
 
   nvinfer1::ILogger& logger_;
 
@@ -165,8 +175,6 @@ class TensorRTEngine : public EngineBase {
   std::unordered_map<std::string /*name*/, nvinfer1::ITensor* /*ITensor*/>
       itensor_map_;
 
-  // The specific GPU id that the TensorRTEngine bounded to.
-  int device_;
   std::vector<std::unique_ptr<plugin::PluginTensorRT>> owned_plugin_;
 
   // TensorRT related internal members

paddle/fluid/inference/tensorrt/trt_int8_calibrator.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/inference/tensorrt/trt_int8_calibrator.h"
+#include "glog/logging.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+// set the batch size before constructing the thread to execute engine
+int TRTInt8Calibrator::getBatchSize() const { return batch_size_; }
+
+TRTInt8Calibrator::TRTInt8Calibrator(
+    const std::unordered_map<std::string, size_t>& buffers, int batch_size,
+    std::string engine_name, const platform::Place place)
+    : batch_size_(batch_size), engine_name_(engine_name) {
+  int i = 0;
+  VLOG(4) << "Init a new calibrator: " << engine_name_;
+  for (const auto it : buffers) {
+    framework::Tensor temp_tensor;
+    std::string input_name = it.first;
+    int data_size = it.second;
+    int num_ele = data_size / sizeof(int16_t);
+    framework::DDim data_shape = framework::make_ddim({num_ele});
+    temp_tensor.Resize(data_shape);
+    data_tensors_.push_back(temp_tensor);
+    data_buffers_[input_name] = std::pair<void*, size_t>(
+        static_cast<void*>(temp_tensor.mutable_data<int16_t>(place)), num_ele);
+    i += 1;
+  }
+}
+
+TRTInt8Calibrator::TRTInt8Calibrator(const std::string& calib_data)
+    : batch_size_(0),
+      calib_running_(false),
+      data_is_set_(false),
+      done_(true),
+      calibration_table_(calib_data) {}
+
+void TRTInt8Calibrator::waitAndSetDone() {
+  std::unique_lock<std::mutex> lk(mut_);
+  while ((calib_running_ || data_is_set_) && !done_) cond_.wait(lk);
+  if (!done_) {
+    done_ = true;
+    cond_.notify_all();
+  }
+}
+
+// There might be more than one input for trt subgraph,
+// So, we use a map to store input information.
+bool TRTInt8Calibrator::setBatch(
+    const std::unordered_map<std::string, void*>& data) {
+  VLOG(3) << "set batch: " << engine_name_;
+  std::unique_lock<std::mutex> lk(mut_);
+  //  There is a producer and a consumer. The producer set the batch data and
+  //  the consumer get the batch data. The size of the data pool is one.
+  //  So, the producer has to wait for the consumer to finish processing before
+  //  they can set the data.
+  while ((calib_running_ || data_is_set_) && (!done_)) cond_.wait(lk);
+  // The done_ is set to true using waitAndSetDone, When all calibration data
+  // are processed.
+  if (done_) return false;
+
+  // Sets the batch.
+  for (const auto& it : data) {
+    auto dataptr = data_buffers_.find(it.first);
+    if (dataptr == data_buffers_.end()) {
+      LOG(FATAL) << "FATAL " << engine_name_ << " input name '" << it.first
+                 << "' does not match with the buffer names";
+    }
+    const auto& d = dataptr->second;
+    PADDLE_ENFORCE(
+        cudaMemcpy(d.first, it.second, d.second, cudaMemcpyDeviceToDevice),
+        "Fail to cudaMemcpy %s for %s", engine_name_, it.first);
+  }
+
+  data_is_set_ = true;
+  cond_.notify_all();
+  return true;
+}
+
+bool TRTInt8Calibrator::getBatch(void** bindings, const char** names,
+                                 int num_bindings) {
+  VLOG(4) << "get batch: " << engine_name_;
+  std::unique_lock<std::mutex> lk(mut_);
+  // The consumer has just finished processing a data.
+  // The producer can set the data again.
+  calib_running_ = false;
+  cond_.notify_all();
+
+  // As long as there is data in the pool, the consumer can get it.
+  while (!data_is_set_ && !done_) cond_.wait(lk);
+  if (done_) return false;
+
+  // Gets the batch
+  for (int i = 0; i < num_bindings; i++) {
+    auto it = data_buffers_.find(names[i]);
+    if (it == data_buffers_.end()) {
+      LOG(FATAL) << "Calibration engine asked for unknown tensor name '"
+                 << names[i] << "' at position " << i;
+    }
+    bindings[i] = it->second.first;
+  }
+
+  data_is_set_ = false;
+  calib_running_ = true;
+  VLOG(4) << "get batch done: " << engine_name_;
+  return true;
+}
+
+void TRTInt8Calibrator::setDone() {
+  std::unique_lock<std::mutex> lk(mut_);
+  done_ = true;
+  cond_.notify_all();
+}
+
+const void* TRTInt8Calibrator::readCalibrationCache(size_t& length) {
+  if (calibration_table_.empty()) return nullptr;
+  length = calibration_table_.size();
+  return calibration_table_.data();
+}
+
+void TRTInt8Calibrator::writeCalibrationCache(const void* ptr,
+                                              std::size_t length) {
+  calibration_table_ = std::string((const char*)ptr, length);
+  VLOG(4) << "Got calibration data for " << engine_name_ << " " << ptr
+          << " length=" << length;
+}
+TRTInt8Calibrator::~TRTInt8Calibrator() {
+  VLOG(4) << "Destroying calibrator for " << engine_name_;
+}
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tensorrt/trt_int8_calibrator.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <NvInfer.h>
+#include <cuda_runtime_api.h>
+#include <atomic>
+#include <memory>
+#include <mutex>  // NOLINT
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/inference/tensorrt/engine.h"
+#include "paddle/fluid/platform/place.h"
+
+namespace paddle {
+namespace inference {
+namespace tensorrt {
+
+class TensorRTEngine;
+
+struct TRTInt8Calibrator : public nvinfer1::IInt8EntropyCalibrator {
+ public:
+  TRTInt8Calibrator(const std::unordered_map<std::string, size_t>& buffers,
+                    int batch_size, std::string engine_name,
+                    const platform::Place place);
+
+  explicit TRTInt8Calibrator(const std::string& calibration_data);
+  ~TRTInt8Calibrator();
+
+  int getBatchSize() const override;
+
+  bool getBatch(void* bindings[], const char* names[],
+                int num_bindings) override;
+
+  bool setBatch(const std::unordered_map<std::string, void*>& data);
+  void setDone();
+  void waitAndSetDone();
+
+  const void* readCalibrationCache(std::size_t& length) override;
+  void writeCalibrationCache(const void* ptr, std::size_t length) override;
+  const std::string& getCalibrationTableAsString() {
+    return calibration_table_;
+  }
+
+ private:
+  const int batch_size_;
+
+  bool calib_running_{true};
+  bool data_is_set_{false};
+  bool done_{false};
+
+  std::mutex mut_;
+  std::condition_variable cond_;
+
+  std::unordered_map<std::string, std::pair<void*, size_t>> data_buffers_;
+  std::vector<framework::Tensor> data_tensors_;
+
+  std::string engine_name_;
+  std::string calibration_table_;
+};
+
+class TRTCalibratorEngine {
+ public:
+  TRTCalibratorEngine() {}
+  std::unique_ptr<TRTInt8Calibrator> calib_;
+  std::unique_ptr<std::thread> thr_;
+  std::unique_ptr<TensorRTEngine> engine_;
+};
+/*
+ * Manager to control the TensorRT Int8 calibration creation and deltetion.
+ */
+class TRTCalibratorEngineManager {
+ public:
+  bool Has() const { return res_.size() > 0; }
+  bool Has(const std::string& name) const {
+    if (res_.count(name) == 0) return false;
+    return res_.at(name).get() != nullptr;
+  }
+
+  // Get Int8Calibrator via name
+  TRTCalibratorEngine* Get(const std::string& name) const {
+    return res_.at(name).get();
+  }
+
+  // Look up or create a calibrator.
+  TRTCalibratorEngine* LookupOrCreate(const std::string& engine_name) {
+    if (res_.count(engine_name) == 0) {
+      auto* p = new TRTCalibratorEngine;
+      res_[engine_name].reset(p);
+    }
+    return res_.at(engine_name).get();
+  }
+
+  // Create an Int8Calibrator
+  TRTCalibratorEngine* Create(const std::string& engine_name) {
+    auto* p = new TRTCalibratorEngine;
+    res_[engine_name].reset(p);
+    return p;
+  }
+
+  void DeleteALL() {
+    for (auto& item : res_) {
+      item.second.reset(nullptr);
+    }
+  }
+
+ private:
+  std::unordered_map<std::string, std::unique_ptr<TRTCalibratorEngine>> res_;
+};
+
+}  // namespace tensorrt
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -54,6 +54,7 @@ else()
     message(WARNING "These tests has been disabled in OSX or WITH_MKL=OFF before being fixed: \n test_analyzer_seq_pool1")
 endif()
 
+
 # RNN2
 set(RNN2_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn2")
 download_model_and_data(${RNN2_INSTALL_DIR} "rnn2_model.tar.gz" "rnn2_data.txt.tar.gz")
@@ -115,6 +116,10 @@ if (NOT EXISTS ${MOBILENET_INSTALL_DIR})
 endif()
 inference_analysis_api_test_with_refer_result(test_analyzer_mobilenet_transpose ${MOBILENET_INSTALL_DIR} analyzer_vis_tester.cc SERIAL)
 
+# googlenet
+inference_analysis_api_test_with_fake_data(test_analyzer_googlenet
+  "${INFERENCE_DEMO_INSTALL_DIR}/googlenet" analyzer_resnet50_tester.cc "googlenet.tar.gz" SERIAL)
+
 # resnet50
 inference_analysis_api_test_with_fake_data(test_analyzer_resnet50
   "${INFERENCE_DEMO_INSTALL_DIR}/resnet50" analyzer_resnet50_tester.cc "resnet50_model.tar.gz" SERIAL)
@@ -123,6 +128,11 @@ inference_analysis_api_test_with_fake_data(test_analyzer_resnet50
 inference_analysis_api_test_with_fake_data(test_analyzer_mobilenet_depthwise_conv
   "${INFERENCE_DEMO_INSTALL_DIR}/mobilenet_depthwise_conv" analyzer_resnet50_tester.cc "mobilenet_model.tar.gz" SERIAL)
 
+# bert, max_len=20
+set(BERT_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/bert20")
+download_model_and_data(${BERT_INSTALL_DIR} "bert_model.tar.gz" "bert_data_len20.txt.tar.gz")
+inference_analysis_api_test(test_analyzer_bert ${BERT_INSTALL_DIR} analyzer_bert_tester.cc SERIAL)
+
 # anakin
 if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
     # anakin rnn1

paddle/fluid/inference/tests/api/analyzer_bert_tester.cc

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/fluid/inference/tests/api/tester_helper.h"
+
+namespace paddle {
+namespace inference {
+
+using paddle::PaddleTensor;
+
+template <typename T>
+void GetValueFromStream(std::stringstream *ss, T *t) {
+  (*ss) >> (*t);
+}
+
+template <>
+void GetValueFromStream<std::string>(std::stringstream *ss, std::string *t) {
+  *t = ss->str();
+}
+
+// Split string to vector
+template <typename T>
+void Split(const std::string &line, char sep, std::vector<T> *v) {
+  std::stringstream ss;
+  T t;
+  for (auto c : line) {
+    if (c != sep) {
+      ss << c;
+    } else {
+      GetValueFromStream<T>(&ss, &t);
+      v->push_back(std::move(t));
+      ss.str({});
+      ss.clear();
+    }
+  }
+
+  if (!ss.str().empty()) {
+    GetValueFromStream<T>(&ss, &t);
+    v->push_back(std::move(t));
+    ss.str({});
+    ss.clear();
+  }
+}
+
+template <typename T>
+constexpr paddle::PaddleDType GetPaddleDType();
+
+template <>
+constexpr paddle::PaddleDType GetPaddleDType<int64_t>() {
+  return paddle::PaddleDType::INT64;
+}
+
+template <>
+constexpr paddle::PaddleDType GetPaddleDType<float>() {
+  return paddle::PaddleDType::FLOAT32;
+}
+
+// Parse tensor from string
+template <typename T>
+bool ParseTensor(const std::string &field, paddle::PaddleTensor *tensor) {
+  std::vector<std::string> data;
+  Split(field, ':', &data);
+  if (data.size() < 2) return false;
+
+  std::string shape_str = data[0];
+
+  std::vector<int> shape;
+  Split(shape_str, ' ', &shape);
+
+  std::string mat_str = data[1];
+
+  std::vector<T> mat;
+  Split(mat_str, ' ', &mat);
+
+  tensor->shape = shape;
+  auto size =
+      std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>()) *
+      sizeof(T);
+  tensor->data.Resize(size);
+  std::copy(mat.begin(), mat.end(), static_cast<T *>(tensor->data.data()));
+  tensor->dtype = GetPaddleDType<T>();
+
+  return true;
+}
+
+// Parse input tensors from string
+bool ParseLine(const std::string &line,
+               std::vector<paddle::PaddleTensor> *tensors) {
+  std::vector<std::string> fields;
+  Split(line, ';', &fields);
+
+  if (fields.size() < 5) return false;
+
+  tensors->clear();
+  tensors->reserve(5);
+
+  int i = 0;
+  // src_id
+  paddle::PaddleTensor src_id;
+  ParseTensor<int64_t>(fields[i++], &src_id);
+  tensors->push_back(src_id);
+
+  // pos_id
+  paddle::PaddleTensor pos_id;
+  ParseTensor<int64_t>(fields[i++], &pos_id);
+  tensors->push_back(pos_id);
+
+  // segment_id
+  paddle::PaddleTensor segment_id;
+  ParseTensor<int64_t>(fields[i++], &segment_id);
+  tensors->push_back(segment_id);
+
+  // self_attention_bias
+  paddle::PaddleTensor self_attention_bias;
+  ParseTensor<float>(fields[i++], &self_attention_bias);
+  tensors->push_back(self_attention_bias);
+
+  // next_segment_index
+  paddle::PaddleTensor next_segment_index;
+  ParseTensor<int64_t>(fields[i++], &next_segment_index);
+  tensors->push_back(next_segment_index);
+
+  return true;
+}
+
+bool LoadInputData(std::vector<std::vector<paddle::PaddleTensor>> *inputs) {
+  if (FLAGS_infer_data.empty()) {
+    LOG(ERROR) << "please set input data path";
+    return false;
+  }
+
+  std::ifstream fin(FLAGS_infer_data);
+  std::string line;
+  int sample = 0;
+
+  // The unit-test dataset only have 10 samples, each sample have 5 feeds.
+  while (std::getline(fin, line)) {
+    std::vector<paddle::PaddleTensor> feed_data;
+    ParseLine(line, &feed_data);
+    inputs->push_back(std::move(feed_data));
+    sample++;
+    if (!FLAGS_test_all_data && sample == FLAGS_batch_size) break;
+  }
+  LOG(INFO) << "number of samples: " << sample;
+
+  return true;
+}
+
+void SetConfig(AnalysisConfig *config) { config->SetModel(FLAGS_infer_model); }
+
+void profile(bool use_mkldnn = false) {
+  AnalysisConfig config;
+  SetConfig(&config);
+
+  if (use_mkldnn) {
+    config.EnableMKLDNN();
+  }
+
+  std::vector<PaddleTensor> outputs;
+  std::vector<std::vector<PaddleTensor>> inputs;
+  LoadInputData(&inputs);
+  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&config),
+                 inputs, &outputs, FLAGS_num_threads);
+}
+
+TEST(Analyzer_bert, profile) { profile(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_bert, profile_mkldnn) { profile(true); }
+#endif
+
+// Check the fuse status
+TEST(Analyzer_bert, fuse_statis) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  int num_ops;
+  auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
+  auto fuse_statis = GetFuseStatis(
+      static_cast<AnalysisPredictor *>(predictor.get()), &num_ops);
+  LOG(INFO) << "num_ops: " << num_ops;
+}
+
+// Compare result of NativeConfig and AnalysisConfig
+void compare(bool use_mkldnn = false) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  if (use_mkldnn) {
+    cfg.EnableMKLDNN();
+  }
+
+  std::vector<std::vector<PaddleTensor>> inputs;
+  LoadInputData(&inputs);
+  CompareNativeAndAnalysis(
+      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), inputs);
+}
+
+TEST(Analyzer_bert, compare) { compare(); }
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_bert, compare_mkldnn) { compare(true /* use_mkldnn */); }
+#endif
+
+// Compare Deterministic result
+TEST(Analyzer_bert, compare_determine) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+
+  std::vector<std::vector<PaddleTensor>> inputs;
+  LoadInputData(&inputs);
+  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
+                       inputs);
+}
+}  // namespace inference
+}  // namespace paddle

paddle/fluid/inference/tests/api/analyzer_dam_tester.cc

@@ -19,7 +19,6 @@ DEFINE_int32(max_turn_num, 9,
 
 namespace paddle {
 namespace inference {
-using contrib::AnalysisConfig;
 
 constexpr int32_t kMaxTurnLen = 50;
 
@@ -165,7 +164,7 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
   input_slots->push_back(std::move(response_mask_tensor));
 }
 
-void SetConfig(contrib::AnalysisConfig *cfg) {
+void SetConfig(AnalysisConfig *cfg) {
   cfg->SetModel(FLAGS_infer_model + "/__model__", FLAGS_infer_model + "/param");
   cfg->SwitchSpecifyInputNames();
   cfg->SwitchIrOptim(true);
@@ -187,7 +186,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 
 // Easy for profiling independently.
 void profile(bool use_mkldnn = false) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   if (use_mkldnn) {
@@ -223,7 +222,7 @@ TEST(Analyzer_dam, profile_mkldnn) { profile(true /* use_mkldnn */); }
 
 // Check the fuse status
 TEST(Analyzer_dam, fuse_statis) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   int num_ops;
@@ -256,7 +255,7 @@ void compare(bool use_mkldnn = false) {
 TEST(Analyzer_dam, compare_with_static_memory_optim) {
   // The small dam will core in CI, but works in local.
   if (FLAGS_max_turn_num == 9) {
-    contrib::AnalysisConfig cfg, cfg1;
+    AnalysisConfig cfg, cfg1;
     DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
 
     std::vector<std::vector<PaddleTensor>> input_slots_all;
@@ -282,7 +281,7 @@ TEST(Analyzer_dam, compare_with_static_memory_optim) {
 TEST(Analyzer_dam, compare_with_dynamic_memory_optim) {
   // The small dam will core in CI, but works in local.
   if (FLAGS_max_turn_num == 9) {
-    contrib::AnalysisConfig cfg, cfg1;
+    AnalysisConfig cfg, cfg1;
     DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
 
     std::vector<std::vector<PaddleTensor>> input_slots_all;

paddle/fluid/inference/tests/api/analyzer_lac_tester.cc

@@ -18,8 +18,6 @@ namespace paddle {
 namespace inference {
 namespace analysis {
 
-using contrib::AnalysisConfig;
-
 struct DataRecord {
   std::vector<int64_t> data;
   std::vector<size_t> lod;

paddle/fluid/inference/tests/api/analyzer_mm_dnn_tester.cc

@@ -16,7 +16,6 @@
 
 namespace paddle {
 namespace inference {
-using contrib::AnalysisConfig;
 
 struct DataRecord {
   std::vector<std::vector<int64_t>> query, title;
@@ -75,7 +74,7 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
   }
 }
 
-void SetConfig(contrib::AnalysisConfig *cfg) {
+void SetConfig(AnalysisConfig *cfg) {
   cfg->SetModel(FLAGS_infer_model);
   cfg->DisableGpu();
   cfg->SwitchSpecifyInputNames();
@@ -95,7 +94,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 
 // Easy for profiling independently.
 void profile(bool use_mkldnn = false) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
   std::vector<PaddleTensor> outputs;
 
@@ -130,7 +129,7 @@ TEST(Analyzer_MM_DNN, profile_mkldnn) { profile(true /* use_mkldnn */); }
 
 // Check the fuse status
 TEST(Analyzer_MM_DNN, fuse_statis) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   int num_ops;
@@ -141,7 +140,7 @@ TEST(Analyzer_MM_DNN, fuse_statis) {
 
 // Compare result of NativeConfig and AnalysisConfig
 void compare(bool use_mkldnn = false) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   if (use_mkldnn) {

paddle/fluid/inference/tests/api/analyzer_ner_tester.cc

@@ -16,7 +16,6 @@
 
 namespace paddle {
 namespace inference {
-using contrib::AnalysisConfig;
 
 struct DataRecord {
   std::vector<std::vector<int64_t>> word, mention;
@@ -76,7 +75,7 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data) {
   }
 }
 
-void SetConfig(contrib::AnalysisConfig *cfg, bool memory_load = false) {
+void SetConfig(AnalysisConfig *cfg, bool memory_load = false) {
   if (memory_load) {
     std::string buffer_prog, buffer_param;
     ReadBinaryFile(FLAGS_infer_model + "/__model__", &buffer_prog);
@@ -105,7 +104,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 
 // Easy for profiling independently.
 void profile(bool memory_load = false) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg, memory_load);
   std::vector<PaddleTensor> outputs;
 
@@ -136,7 +135,7 @@ TEST(Analyzer_Chinese_ner, profile_memory_load) {
 
 // Check the fuse status
 TEST(Analyzer_Chinese_ner, fuse_statis) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   int num_ops;
@@ -152,7 +151,7 @@ TEST(Analyzer_Chinese_ner, fuse_statis) {
 
 // Compare result of NativeConfig and AnalysisConfig
 TEST(Analyzer_Chinese_ner, compare) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;

paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc

@@ -16,7 +16,6 @@
 
 namespace paddle {
 namespace inference {
-using contrib::AnalysisConfig;
 
 struct DataRecord {
   std::vector<std::vector<int64_t>> query_basic, query_phrase, title_basic,
@@ -103,7 +102,7 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
   }
 }
 
-void SetConfig(contrib::AnalysisConfig *cfg) {
+void SetConfig(AnalysisConfig *cfg) {
   cfg->SetModel(FLAGS_infer_model);
   cfg->DisableGpu();
   cfg->SwitchSpecifyInputNames();
@@ -123,7 +122,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 
 // Easy for profiling independently.
 TEST(Analyzer_Pyramid_DNN, profile) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
   std::vector<PaddleTensor> outputs;
 
@@ -147,7 +146,7 @@ TEST(Analyzer_Pyramid_DNN, profile) {
 
 // Check the fuse status
 TEST(Analyzer_Pyramid_DNN, fuse_statis) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   int num_ops;
@@ -158,7 +157,7 @@ TEST(Analyzer_Pyramid_DNN, fuse_statis) {
 
 // Compare result of NativeConfig and AnalysisConfig
 TEST(Analyzer_Pyramid_DNN, compare) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;

paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc

@@ -20,7 +20,6 @@ namespace paddle {
 namespace inference {
 
 using namespace framework;  // NOLINT
-using namespace contrib;    // NOLINT
 
 struct DataRecord {
   std::vector<std::vector<std::vector<float>>> link_step_data_all;
@@ -223,7 +222,7 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
 
 // Easy for profiling independently.
 TEST(Analyzer_rnn1, profile) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
   cfg.DisableGpu();
   cfg.SwitchIrDebug();
@@ -237,7 +236,7 @@ TEST(Analyzer_rnn1, profile) {
 
 // Check the fuse status
 TEST(Analyzer_rnn1, fuse_statis) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   int num_ops;
@@ -254,7 +253,7 @@ TEST(Analyzer_rnn1, fuse_statis) {
 
 // Compare result of NativeConfig and AnalysisConfig
 TEST(Analyzer_rnn1, compare) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
 
   std::vector<std::vector<PaddleTensor>> input_slots_all;
@@ -276,7 +275,7 @@ TEST(Analyzer_rnn1, compare_determine) {
 
 // Test Multi-Thread.
 TEST(Analyzer_rnn1, multi_thread) {
-  contrib::AnalysisConfig cfg;
+  AnalysisConfig cfg;
   SetConfig(&cfg);
   std::vector<PaddleTensor> outputs;
 

paddle/fluid/inference/tests/api/analyzer_vis_tester.cc

@@ -20,7 +20,6 @@ limitations under the License. */
 namespace paddle {
 namespace inference {
 namespace analysis {
-using contrib::AnalysisConfig;
 
 struct Record {
   std::vector<float> data;

paddle/fluid/inference/tests/api/config_printer.h

@@ -58,9 +58,8 @@ std::ostream &operator<<(std::ostream &os, const NativeConfig &config) {
   return os;
 }
 
-std::ostream &operator<<(std::ostream &os,
-                         const contrib::AnalysisConfig &config) {
-  os << GenSpaces(num_spaces) << "contrib::AnalysisConfig {\n";
+std::ostream &operator<<(std::ostream &os, const AnalysisConfig &config) {
+  os << GenSpaces(num_spaces) << "AnalysisConfig {\n";
   num_spaces++;
   os << config.ToNativeConfig();
   if (!config.model_from_memory()) {

paddle/fluid/inference/tests/api/tester_helper.h

@@ -56,16 +56,9 @@ DECLARE_int32(paddle_num_threads);
 namespace paddle {
 namespace inference {
 
-float Random(float low, float high) {
-  static std::random_device rd;
-  static std::mt19937 mt(rd());
-  std::uniform_real_distribution<double> dist(low, high);
-  return dist(mt);
-}
-
 void PrintConfig(const PaddlePredictor::Config *config, bool use_analysis) {
   const auto *analysis_config =
-      reinterpret_cast<const contrib::AnalysisConfig *>(config);
+      reinterpret_cast<const AnalysisConfig *>(config);
   if (use_analysis) {
     LOG(INFO) << *analysis_config;
     return;
@@ -109,9 +102,9 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
 std::unique_ptr<PaddlePredictor> CreateTestPredictor(
     const PaddlePredictor::Config *config, bool use_analysis = true) {
   const auto *analysis_config =
-      reinterpret_cast<const contrib::AnalysisConfig *>(config);
+      reinterpret_cast<const AnalysisConfig *>(config);
   if (use_analysis) {
-    return CreatePaddlePredictor<contrib::AnalysisConfig>(*analysis_config);
+    return CreatePaddlePredictor<AnalysisConfig>(*analysis_config);
   }
   auto native_config = analysis_config->ToNativeConfig();
   return CreatePaddlePredictor<NativeConfig>(native_config);
@@ -146,7 +139,8 @@ void SetFakeImageInput(std::vector<std::vector<PaddleTensor>> *inputs,
                        const std::string &dirname, bool is_combined = true,
                        std::string model_filename = "model",
                        std::string params_filename = "params",
-                       const std::vector<std::string> *feed_names = nullptr) {
+                       const std::vector<std::string> *feed_names = nullptr,
+                       const int continuous_inuput_index = 0) {
   // Set fake_image_data
   PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data.");
   std::vector<std::vector<int64_t>> feed_target_shapes = GetFeedTargetShapes(
@@ -183,7 +177,8 @@ void SetFakeImageInput(std::vector<std::vector<PaddleTensor>> *inputs,
     float *input_data = static_cast<float *>(input.data.data());
     // fill input data, for profile easily, do not use random data here.
     for (size_t j = 0; j < len; ++j) {
-      *(input_data + j) = Random(0.0, 1.0) / 10.;
+      *(input_data + j) =
+          static_cast<float>((j + continuous_inuput_index) % len) / len;
     }
   }
   (*inputs).emplace_back(input_slots);

paddle/fluid/inference/tests/api/trt_models_tester.cc

@@ -42,9 +42,9 @@ void SetConfig(ConfigType* config, std::string model_dir, bool use_gpu,
 }
 
 template <>
-void SetConfig<contrib::AnalysisConfig>(contrib::AnalysisConfig* config,
-                                        std::string model_dir, bool use_gpu,
-                                        bool use_tensorrt, int batch_size) {
+void SetConfig<AnalysisConfig>(AnalysisConfig* config, std::string model_dir,
+                               bool use_gpu, bool use_tensorrt,
+                               int batch_size) {
   if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) {
     config->SetModel(model_dir + "/" + FLAGS_prog_filename,
                      model_dir + "/" + FLAGS_param_filename);
@@ -75,11 +75,11 @@ void profile(std::string model_dir, bool use_analysis, bool use_tensorrt) {
 
   std::vector<PaddleTensor> outputs;
   if (use_analysis || use_tensorrt) {
-    contrib::AnalysisConfig config;
+    AnalysisConfig config;
     config.EnableUseGpu(100, 0);
     config.pass_builder()->TurnOnDebug();
-    SetConfig<contrib::AnalysisConfig>(&config, model_dir, true, use_tensorrt,
-                                       FLAGS_batch_size);
+    SetConfig<AnalysisConfig>(&config, model_dir, true, use_tensorrt,
+                              FLAGS_batch_size);
     TestPrediction(reinterpret_cast<PaddlePredictor::Config*>(&config),
                    inputs_all, &outputs, FLAGS_num_threads, true);
   } else {
@@ -99,18 +99,18 @@ void compare(std::string model_dir, bool use_tensorrt) {
     SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "");
   }
 
-  contrib::AnalysisConfig analysis_config;
-  SetConfig<contrib::AnalysisConfig>(&analysis_config, model_dir, true,
-                                     use_tensorrt, FLAGS_batch_size);
+  AnalysisConfig analysis_config;
+  SetConfig<AnalysisConfig>(&analysis_config, model_dir, true, use_tensorrt,
+                            FLAGS_batch_size);
   CompareNativeAndAnalysis(
       reinterpret_cast<const PaddlePredictor::Config*>(&analysis_config),
       inputs_all);
 }
 
 void compare_continuous_input(std::string model_dir, bool use_tensorrt) {
-  contrib::AnalysisConfig analysis_config;
-  SetConfig<contrib::AnalysisConfig>(&analysis_config, model_dir, true,
-                                     use_tensorrt, FLAGS_batch_size);
+  AnalysisConfig analysis_config;
+  SetConfig<AnalysisConfig>(&analysis_config, model_dir, true, use_tensorrt,
+                            FLAGS_batch_size);
   auto config =
       reinterpret_cast<const PaddlePredictor::Config*>(&analysis_config);
   auto native_pred = CreateTestPredictor(config, false);
@@ -119,9 +119,10 @@ void compare_continuous_input(std::string model_dir, bool use_tensorrt) {
     std::vector<std::vector<PaddleTensor>> inputs_all;
     if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) {
       SetFakeImageInput(&inputs_all, model_dir, true, FLAGS_prog_filename,
-                        FLAGS_param_filename);
+                        FLAGS_param_filename, nullptr, i);
     } else {
-      SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "");
+      SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "", nullptr,
+                        i);
     }
     CompareNativeAndAnalysis(native_pred.get(), analysis_pred.get(),
                              inputs_all);

paddle/fluid/memory/allocation/legacy_allocator.cc

@@ -13,9 +13,15 @@
 // limitations under the License.
 
 #include "paddle/fluid/memory/allocation/legacy_allocator.h"
+
 #include <string>
 #include <utility>
 #include <vector>
+
+#ifdef PADDLE_WITH_JEMALLOC
+#include <jemalloc/jemalloc.h>
+#endif
+
 #include "glog/logging.h"
 #include "paddle/fluid/memory/detail/buddy_allocator.h"
 #include "paddle/fluid/memory/detail/system_allocator.h"
@@ -95,7 +101,11 @@ struct NaiveAllocator {
 template <>
 void *Alloc<platform::CPUPlace>(const platform::CPUPlace &place, size_t size) {
   VLOG(10) << "Allocate " << size << " bytes on " << platform::Place(place);
+#ifdef PADDLE_WITH_JEMALLOC
+  void *p = malloc(size);
+#else
   void *p = GetCPUBuddyAllocator()->Alloc(size);
+#endif
   if (FLAGS_init_allocated_mem) {
     memset(p, 0xEF, size);
   }
@@ -107,12 +117,21 @@ template <>
 void Free<platform::CPUPlace>(const platform::CPUPlace &place, void *p,
                               size_t size) {
   VLOG(10) << "Free pointer=" << p << " on " << platform::Place(place);
+#ifdef PADDLE_WITH_JEMALLOC
+  free(p);
+#else
   GetCPUBuddyAllocator()->Free(p);
+#endif
 }
 
 template <>
 size_t Used<platform::CPUPlace>(const platform::CPUPlace &place) {
+#ifdef PADDLE_WITH_JEMALLOC
+  // fake the result of used memory when PADDLE_WITH_JEMALLOC is ON
+  return 0U;
+#else
   return GetCPUBuddyAllocator()->Used();
+#endif
 }
 
 #ifdef PADDLE_WITH_CUDA

paddle/fluid/operators/activation_op.cc

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/activation_op.h"
 #include <string>
-#include "paddle/fluid/operators/mkldnn_activation_op.h"
+#include "paddle/fluid/operators/mkldnn/mkldnn_activation_op.h"
 #include "paddle/fluid/platform/port.h"
 
 namespace paddle {

paddle/fluid/operators/beam_search_op.cc

@@ -51,6 +51,9 @@ class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker {
     AddOutput("selected_scores",
               "A LoDTensor containing the accumulated scores corresponding to "
               "Output(selected_ids).");
+    AddOutput(
+        "parent_idx",
+        "A Tensor preserving the selected_ids' parent indice in pre_ids.");
 
     // Attributes stored in AttributeMap
     AddAttr<int>("level", "the level of LoDTensor");

paddle/fluid/operators/beam_search_op.h

@@ -41,13 +41,15 @@ class BeamSearchOpKernel : public framework::OpKernel<T> {
     auto selected_ids = context.Output<framework::LoDTensor>("selected_ids");
     auto selected_scores =
         context.Output<framework::LoDTensor>("selected_scores");
+    auto* parent_idx = context.Output<framework::Tensor>("parent_idx");
     PADDLE_ENFORCE_NOT_NULL(selected_ids);
     PADDLE_ENFORCE_NOT_NULL(selected_scores);
+    PADDLE_ENFORCE_NOT_NULL(parent_idx);
 
     math::BeamSearchFunctor<DeviceContext, T> alg;
     alg(context.template device_context<DeviceContext>(), pre_ids, pre_scores,
-        ids, scores, selected_ids, selected_scores, level, beam_size, end_id,
-        is_accumulated);
+        ids, scores, selected_ids, selected_scores, parent_idx, level,
+        beam_size, end_id, is_accumulated);
   }
 };
 

paddle/fluid/operators/detection/box_coder_op.cc

@@ -10,6 +10,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/detection/box_coder_op.h"
+#include <vector>
 
 namespace paddle {
 namespace operators {
@@ -32,32 +33,57 @@ class BoxCoderOp : public framework::OperatorWithKernel {
 
     if (ctx->IsRuntime()) {
       PADDLE_ENFORCE_EQ(prior_box_dims.size(), 2,
-                        "The rank of Input of PriorBoxVar must be 2");
+                        "The rank of Input PriorBox must be 2");
       PADDLE_ENFORCE_EQ(prior_box_dims[1], 4,
                         "The shape of PriorBox is [N, 4]");
       if (ctx->HasInput("PriorBoxVar")) {
         auto prior_box_var_dims = ctx->GetInputDim("PriorBoxVar");
-        PADDLE_ENFORCE_EQ(prior_box_dims, prior_box_var_dims);
+        PADDLE_ENFORCE(
+            prior_box_var_dims.size() == 1 || prior_box_var_dims.size() == 2,
+            "Input(PriorBoxVar) of BoxCoderOp should be 1 or 2.");
+        if (prior_box_var_dims.size() == 1) {
+          PADDLE_ENFORCE_EQ(
+              prior_box_var_dims[0], 4,
+              "The 1st dimension of Input(PriorBoxVar) should be 4"
+              "when the rank is 1.");
+        } else {
+          PADDLE_ENFORCE_EQ(
+              prior_box_dims, prior_box_var_dims,
+              "The dimension of Input(PriorBoxVar) should be equal to"
+              "the dimension of Input(PriorBox when the rank is 2.)");
+        }
       }
+    }
 
-      auto code_type =
-          GetBoxCodeType(ctx->Attrs().Get<std::string>("code_type"));
-      if (code_type == BoxCodeType::kEncodeCenterSize) {
-        PADDLE_ENFORCE_EQ(target_box_dims.size(), 2,
-                          "The rank of Input of TargetBox must be 2");
-        PADDLE_ENFORCE_EQ(target_box_dims[1], 4,
-                          "The shape of TargetBox is [M, 4]");
-      } else if (code_type == BoxCodeType::kDecodeCenterSize) {
-        PADDLE_ENFORCE_EQ(target_box_dims.size(), 3,
-                          "The rank of Input of TargetBox must be 3");
+    auto code_type = GetBoxCodeType(ctx->Attrs().Get<std::string>("code_type"));
+    int axis = ctx->Attrs().Get<int>("axis");
+    if (code_type == BoxCodeType::kEncodeCenterSize) {
+      PADDLE_ENFORCE_EQ(target_box_dims.size(), 2,
+                        "The rank of Input TargetBox must be 2");
+      PADDLE_ENFORCE_EQ(target_box_dims[1], 4,
+                        "The shape of TargetBox is [M, 4]");
+      ctx->SetOutputDim(
+          "OutputBox",
+          framework::make_ddim({target_box_dims[0], prior_box_dims[0], 4}));
+    } else if (code_type == BoxCodeType::kDecodeCenterSize) {
+      PADDLE_ENFORCE_EQ(target_box_dims.size(), 3,
+                        "The rank of Input TargetBox must be 3");
+      if (axis == 0) {
         PADDLE_ENFORCE_EQ(target_box_dims[1], prior_box_dims[0]);
-        PADDLE_ENFORCE_EQ(target_box_dims[2], prior_box_dims[1]);
+      } else if (axis == 1) {
+        PADDLE_ENFORCE_EQ(target_box_dims[0], prior_box_dims[0]);
+      } else {
+        PADDLE_THROW("axis must be 0 or 1.");
       }
+      PADDLE_ENFORCE_EQ(target_box_dims[2], prior_box_dims[1]);
+      ctx->ShareDim("TargetBox", /*->*/ "OutputBox");
+    }
+
+    if (code_type == BoxCodeType::kDecodeCenterSize && axis == 1) {
+      ctx->ShareLoD("PriorBox", /*->*/ "OutputBox");
+    } else {
+      ctx->ShareLoD("TargetBox", /*->*/ "OutputBox");
     }
-    ctx->SetOutputDim(
-        "OutputBox",
-        framework::make_ddim({target_box_dims[0], prior_box_dims[0], 4}));
-    ctx->ShareLoD("TargetBox", /*->*/ "OutputBox");
   }
 };
 
@@ -100,6 +126,21 @@ class BoxCoderOpMaker : public framework::OpProtoAndCheckerMaker {
                   "(bool, default true) "
                   "whether treat the priorbox as a noramlized box")
         .SetDefault(true);
+    AddAttr<int>("axis",
+                 "(int, default 0)"
+                 "which axis in PriorBox to broadcast for box decode,"
+                 "for example, if axis is 0 and TargetBox has shape"
+                 "[N, M, 4] and PriorBox has shape [M, 4], then PriorBox "
+                 "will broadcast to [N, M, 4] for decoding. It is only valid"
+                 "when code type is decode_center_size")
+        .SetDefault(0)
+        .InEnum({0, 1});
+    AddAttr<std::vector<float>>(
+        "variance",
+        "(vector<float>, default {}),"
+        "variance of prior box with shape [4]. PriorBoxVar and variance can"
+        "not be provided at the same time.")
+        .SetDefault(std::vector<float>{});
     AddOutput("OutputBox",
               "(LoDTensor or Tensor) "
               "When code_type is 'encode_center_size', the output tensor of "
@@ -138,7 +179,11 @@ where `tx`, `ty`, `tw`, `th` denote the target box's center coordinates, width
 and height respectively. Similarly, `px`, `py`, `pw`, `ph` denote the
 priorbox's (anchor) center coordinates, width and height. `pxv`, `pyv`, `pwv`,
 `phv` denote the variance of the priorbox and `ox`, `oy`, `ow`, `oh` denote the
-encoded/decoded coordinates, width and height.
+encoded/decoded coordinates, width and height. 
+
+During Box Decoding, two modes for broadcast are supported. Say target box has 
+shape [N, M, 4], and the shape of prior box can be [N, 4] or [M, 4]. Then prior
+box will broadcast to target box along the assigned axis. 
 )DOC");
   }
 };

paddle/fluid/operators/detection/box_coder_op.cu

@@ -9,6 +9,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#include <thrust/device_vector.h>
+#include <thrust/host_vector.h>
+#include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/operators/detection/box_coder_op.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
 
@@ -16,11 +19,11 @@ namespace paddle {
 namespace operators {
 
 template <typename T>
-__global__ void EncodeCenterSizeKernel(const T* prior_box_data,
-                                       const T* prior_box_var_data,
-                                       const T* target_box_data, const int row,
-                                       const int col, const int len,
-                                       const bool normalized, T* output) {
+__global__ void EncodeCenterSizeKernel(
+    const T* prior_box_data, const T* prior_box_var_data,
+    const T* target_box_data, const int row, const int col, const int len,
+    const bool normalized, const T prior_box_var_size, const float* variance,
+    const int var_size, T* output) {
   const int idx = threadIdx.x + blockIdx.x * blockDim.x;
   if (idx < row * col) {
     const int row_idx = idx / col;
@@ -30,11 +33,9 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data,
     T prior_box_height = prior_box_data[col_idx * len + 3] -
                          prior_box_data[col_idx * len + 1] +
                          (normalized == false);
-    T prior_box_center_x =
-        (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2;
-    T prior_box_center_y = (prior_box_data[col_idx * len + 3] +
-                            prior_box_data[col_idx * len + 1]) /
-                           2;
+    T prior_box_center_x = prior_box_data[col_idx * len] + prior_box_width / 2;
+    T prior_box_center_y =
+        prior_box_data[col_idx * len + 1] + prior_box_height / 2;
 
     T target_box_center_x =
         (target_box_data[row_idx * len + 2] + target_box_data[row_idx * len]) /
@@ -55,58 +56,73 @@ __global__ void EncodeCenterSizeKernel(const T* prior_box_data,
     output[idx * len + 2] = log(fabs(target_box_width / prior_box_width));
     output[idx * len + 3] = log(fabs(target_box_height / prior_box_height));
     if (prior_box_var_data) {
-      output[idx * len] /= prior_box_var_data[col_idx * len];
-      output[idx * len + 1] /= prior_box_var_data[col_idx * len + 1];
-      output[idx * len + 2] /= prior_box_var_data[col_idx * len + 2];
-      output[idx * len + 3] /= prior_box_var_data[col_idx * len + 3];
+      int prior_var_offset = 0;
+      if (prior_box_var_size == 2) {
+        prior_var_offset = col_idx * len;
+      }
+      output[idx * len] /= prior_box_var_data[prior_var_offset];
+      output[idx * len + 1] /= prior_box_var_data[prior_var_offset + 1];
+      output[idx * len + 2] /= prior_box_var_data[prior_var_offset + 2];
+      output[idx * len + 3] /= prior_box_var_data[prior_var_offset + 3];
+    } else if (var_size == 4) {
+      for (int k = 0; k < 4; ++k) {
+        output[idx * len + k] /= static_cast<T>(variance[k]);
+      }
     }
   }
 }
 
 template <typename T>
-__global__ void DecodeCenterSizeKernel(const T* prior_box_data,
-                                       const T* prior_box_var_data,
-                                       const T* target_box_data, const int row,
-                                       const int col, const int len,
-                                       const bool normalized, T* output) {
+__global__ void DecodeCenterSizeKernel(
+    const T* prior_box_data, const T* prior_box_var_data,
+    const T* target_box_data, const int row, const int col, const int len,
+    const bool normalized, const T prior_box_var_size, const float* variance,
+    const int var_size, const int axis, T* output) {
   const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  int prior_box_offset = 0;
   if (idx < row * col) {
     const int col_idx = idx % col;
-    T prior_box_width = prior_box_data[col_idx * len + 2] -
-                        prior_box_data[col_idx * len] + (normalized == false);
-    T prior_box_height = prior_box_data[col_idx * len + 3] -
-                         prior_box_data[col_idx * len + 1] +
+    const int row_idx = idx / col;
+    prior_box_offset = axis == 0 ? col_idx * len : row_idx * len;
+    T prior_box_width = prior_box_data[prior_box_offset + 2] -
+                        prior_box_data[prior_box_offset] +
+                        (normalized == false);
+    T prior_box_height = prior_box_data[prior_box_offset + 3] -
+                         prior_box_data[prior_box_offset + 1] +
                          (normalized == false);
     T prior_box_center_x =
-        (prior_box_data[col_idx * len + 2] + prior_box_data[col_idx * len]) / 2;
-    T prior_box_center_y = (prior_box_data[col_idx * len + 3] +
-                            prior_box_data[col_idx * len + 1]) /
-                           2;
+        prior_box_data[prior_box_offset] + prior_box_width / 2;
+    T prior_box_center_y =
+        prior_box_data[prior_box_offset + 1] + prior_box_height / 2;
     T target_box_width, target_box_height;
     T target_box_center_x, target_box_center_y;
+    T box_var_x = T(1), box_var_y = T(1);
+    T box_var_w = T(1), box_var_h = T(1);
     if (prior_box_var_data) {
-      target_box_width = exp(prior_box_var_data[col_idx * len + 2] *
-                             target_box_data[idx * len + 2]) *
-                         prior_box_width;
-      target_box_height = exp(prior_box_var_data[col_idx * len + 3] *
-                              target_box_data[idx * len + 3]) *
-                          prior_box_height;
-      target_box_center_x = prior_box_var_data[col_idx * len] *
-                                target_box_data[idx * len] * prior_box_width +
-                            prior_box_center_x;
-      target_box_center_y = prior_box_var_data[col_idx * len + 1] *
-                                target_box_data[idx * len + 1] *
-                                prior_box_height +
-                            prior_box_center_y;
-    } else {
-      target_box_width = exp(target_box_data[idx * len + 2]) * prior_box_width;
-      target_box_height =
-          exp(target_box_data[idx * len + 3]) * prior_box_height;
-      target_box_center_x =
-          target_box_data[idx * len] * prior_box_width + prior_box_center_x;
-      target_box_center_y = target_box_data[idx * len + 1] * prior_box_height +
-                            prior_box_center_y;
+      int prior_var_offset = 0;
+      if (prior_box_var_size == 2) {
+        prior_var_offset = axis == 0 ? col_idx * len : row_idx * len;
+      }
+      box_var_x = prior_box_var_data[prior_var_offset];
+      box_var_y = prior_box_var_data[prior_var_offset + 1];
+      box_var_w = prior_box_var_data[prior_var_offset + 2];
+      box_var_h = prior_box_var_data[prior_var_offset + 3];
+    } else if (var_size == 4) {
+      box_var_x = static_cast<T>(variance[0]);
+      box_var_y = static_cast<T>(variance[1]);
+      box_var_w = static_cast<T>(variance[2]);
+      box_var_h = static_cast<T>(variance[3]);
     }
+    target_box_width =
+        exp(box_var_w * target_box_data[idx * len + 2]) * prior_box_width;
+    target_box_height =
+        exp(box_var_h * target_box_data[idx * len + 3]) * prior_box_height;
+    target_box_center_x =
+        box_var_x * target_box_data[idx * len] * prior_box_width +
+        prior_box_center_x;
+    target_box_center_y =
+        box_var_y * target_box_data[idx * len + 1] * prior_box_height +
+        prior_box_center_y;
 
     output[idx * len] = target_box_center_x - target_box_width / 2;
     output[idx * len + 1] = target_box_center_y - target_box_height / 2;
@@ -127,36 +143,64 @@ class BoxCoderCUDAKernel : public framework::OpKernel<T> {
     auto* prior_box_var = context.Input<framework::Tensor>("PriorBoxVar");
     auto* target_box = context.Input<framework::LoDTensor>("TargetBox");
     auto* output_box = context.Output<framework::Tensor>("OutputBox");
-
+    std::vector<float> variance = context.Attr<std::vector<float>>("variance");
     const T* prior_box_data = prior_box->data<T>();
     const T* target_box_data = target_box->data<T>();
     const T* prior_box_var_data = nullptr;
-    if (prior_box_var) prior_box_var_data = prior_box_var->data<T>();
+    auto prior_box_var_size = 0;
+    if (prior_box_var) {
+      PADDLE_ENFORCE(variance.empty(),
+                     "Input 'PriorBoxVar' and attribute 'variance' should not"
+                     "be used at the same time.");
+      prior_box_var_data = prior_box_var->data<T>();
+      prior_box_var_size = prior_box_var->dims().size();
+    }
+    if (!(variance.empty())) {
+      PADDLE_ENFORCE(static_cast<int>(variance.size()) == 4,
+                     "Size of attribute 'variance' should be 4");
+    }
 
     if (target_box->lod().size()) {
       PADDLE_ENFORCE_EQ(target_box->lod().size(), 1,
                         "Only support 1 level of LoD.");
     }
+    const int var_size = static_cast<int>(variance.size());
+
+    auto code_type = GetBoxCodeType(context.Attr<std::string>("code_type"));
+    bool normalized = context.Attr<bool>("box_normalized");
+    int axis = context.Attr<int>("axis");
+
     auto row = target_box->dims()[0];
     auto col = prior_box->dims()[0];
+    if (code_type == BoxCodeType::kDecodeCenterSize) {
+      col = target_box->dims()[1];
+    }
     auto len = prior_box->dims()[1];
     int block = 512;
     int grid = (row * col + block - 1) / block;
     auto& device_ctx = context.cuda_device_context();
 
+    auto& allocator =
+        platform::DeviceTemporaryAllocator::Instance().Get(device_ctx);
+    int bytes = var_size * sizeof(float);
+    auto dev_var = allocator.Allocate(bytes);
+    float* dev_var_data = reinterpret_cast<float*>(dev_var->ptr());
+    auto cplace = platform::CPUPlace();
+    const auto gplace = boost::get<platform::CUDAPlace>(context.GetPlace());
+    memory::Copy(gplace, dev_var_data, cplace, &variance[0], bytes,
+                 device_ctx.stream());
+
     output_box->mutable_data<T>({row, col, len}, context.GetPlace());
     T* output = output_box->data<T>();
 
-    auto code_type = GetBoxCodeType(context.Attr<std::string>("code_type"));
-    bool normalized = context.Attr<bool>("box_normalized");
     if (code_type == BoxCodeType::kEncodeCenterSize) {
       EncodeCenterSizeKernel<T><<<grid, block, 0, device_ctx.stream()>>>(
           prior_box_data, prior_box_var_data, target_box_data, row, col, len,
-          normalized, output);
+          normalized, prior_box_var_size, dev_var_data, var_size, output);
     } else if (code_type == BoxCodeType::kDecodeCenterSize) {
       DecodeCenterSizeKernel<T><<<grid, block, 0, device_ctx.stream()>>>(
           prior_box_data, prior_box_var_data, target_box_data, row, col, len,
-          normalized, output);
+          normalized, prior_box_var_size, dev_var_data, var_size, axis, output);
     }
   }
 };

paddle/fluid/operators/detection/box_coder_op.h

@@ -11,6 +11,7 @@ limitations under the License. */
 
 #pragma once
 #include <string>
+#include <vector>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/math/math_function.h"
 
@@ -34,7 +35,8 @@ class BoxCoderKernel : public framework::OpKernel<T> {
   void EncodeCenterSize(const framework::Tensor* target_box,
                         const framework::Tensor* prior_box,
                         const framework::Tensor* prior_box_var,
-                        const bool normalized, T* output) const {
+                        const bool normalized,
+                        const std::vector<float> variance, T* output) const {
     int64_t row = target_box->dims()[0];
     int64_t col = prior_box->dims()[0];
     int64_t len = prior_box->dims()[1];
@@ -53,10 +55,9 @@ class BoxCoderKernel : public framework::OpKernel<T> {
         T prior_box_height = prior_box_data[j * len + 3] -
                              prior_box_data[j * len + 1] +
                              (normalized == false);
-        T prior_box_center_x =
-            (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
+        T prior_box_center_x = prior_box_data[j * len] + prior_box_width / 2;
         T prior_box_center_y =
-            (prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2;
+            prior_box_data[j * len + 1] + prior_box_height / 2;
 
         T target_box_center_x =
             (target_box_data[i * len + 2] + target_box_data[i * len]) / 2;
@@ -78,10 +79,18 @@ class BoxCoderKernel : public framework::OpKernel<T> {
         output[offset + 3] =
             std::log(std::fabs(target_box_height / prior_box_height));
         if (prior_box_var) {
-          output[offset] /= prior_box_var_data[j * len];
-          output[offset + 1] /= prior_box_var_data[j * len + 1];
-          output[offset + 2] /= prior_box_var_data[j * len + 2];
-          output[offset + 3] /= prior_box_var_data[j * len + 3];
+          int prior_var_offset = 0;
+          if (prior_box_var->dims().size() == 2) {
+            prior_var_offset = j * len;
+          }
+          output[offset] /= prior_box_var_data[prior_var_offset];
+          output[offset + 1] /= prior_box_var_data[prior_var_offset + 1];
+          output[offset + 2] /= prior_box_var_data[prior_var_offset + 2];
+          output[offset + 3] /= prior_box_var_data[prior_var_offset + 3];
+        } else if (!(variance.empty())) {
+          for (int k = 0; k < 4; ++k) {
+            output[offset + k] /= static_cast<T>(variance[k]);
+          }
         }
       }
     }
@@ -89,58 +98,71 @@ class BoxCoderKernel : public framework::OpKernel<T> {
   void DecodeCenterSize(const framework::Tensor* target_box,
                         const framework::Tensor* prior_box,
                         const framework::Tensor* prior_box_var,
-                        const bool normalized, T* output) const {
+                        const bool normalized, const int axis,
+                        const std::vector<float> variance, T* output) const {
     int64_t row = target_box->dims()[0];
-    int64_t col = prior_box->dims()[0];
-    int64_t len = prior_box->dims()[1];
+    int64_t col = target_box->dims()[1];
+    int64_t len = target_box->dims()[2];
 
     auto* target_box_data = target_box->data<T>();
     auto* prior_box_data = prior_box->data<T>();
     const T* prior_box_var_data = nullptr;
     if (prior_box_var) prior_box_var_data = prior_box_var->data<T>();
-
+    int prior_box_offset = 0;
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for collapse(2)
 #endif
     for (int64_t i = 0; i < row; ++i) {
       for (int64_t j = 0; j < col; ++j) {
         size_t offset = i * col * len + j * len;
-        T prior_box_width = prior_box_data[j * len + 2] -
-                            prior_box_data[j * len] + (normalized == false);
-        T prior_box_height = prior_box_data[j * len + 3] -
-                             prior_box_data[j * len + 1] +
+        if (axis == 0) {
+          prior_box_offset = j * len;
+        } else if (axis == 1) {
+          prior_box_offset = i * len;
+        }
+        T prior_box_width = prior_box_data[prior_box_offset + 2] -
+                            prior_box_data[prior_box_offset] +
+                            (normalized == false);
+        T prior_box_height = prior_box_data[prior_box_offset + 3] -
+                             prior_box_data[prior_box_offset + 1] +
                              (normalized == false);
         T prior_box_center_x =
-            (prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
+            prior_box_data[prior_box_offset] + prior_box_width / 2;
         T prior_box_center_y =
-            (prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2;
+            prior_box_data[prior_box_offset + 1] + prior_box_height / 2;
 
         T target_box_center_x = 0, target_box_center_y = 0;
         T target_box_width = 0, target_box_height = 0;
+        T box_var_x = T(1), box_var_y = T(1);
+        T box_var_w = T(1), box_var_h = T(1);
         if (prior_box_var) {
-          target_box_center_x = prior_box_var_data[j * len] *
-                                    target_box_data[offset] * prior_box_width +
-                                prior_box_center_x;
-          target_box_center_y = prior_box_var_data[j * len + 1] *
-                                    target_box_data[offset + 1] *
-                                    prior_box_height +
-                                prior_box_center_y;
-          target_box_width = std::exp(prior_box_var_data[j * len + 2] *
-                                      target_box_data[offset + 2]) *
-                             prior_box_width;
-          target_box_height = std::exp(prior_box_var_data[j * len + 3] *
-                                       target_box_data[offset + 3]) *
-                              prior_box_height;
-        } else {
-          target_box_center_x =
-              target_box_data[offset] * prior_box_width + prior_box_center_x;
-          target_box_center_y = target_box_data[offset + 1] * prior_box_height +
-                                prior_box_center_y;
-          target_box_width =
-              std::exp(target_box_data[offset + 2]) * prior_box_width;
-          target_box_height =
-              std::exp(target_box_data[offset + 3]) * prior_box_height;
+          int prior_var_offset = 0;
+          if (prior_box_var->dims().size() == 2) {
+            if (axis == 0)
+              prior_var_offset = j * len;
+            else if (axis == 1)
+              prior_var_offset = i * len;
+          }
+          box_var_x = prior_box_var_data[prior_var_offset];
+          box_var_y = prior_box_var_data[prior_var_offset + 1];
+          box_var_w = prior_box_var_data[prior_var_offset + 2];
+          box_var_h = prior_box_var_data[prior_var_offset + 3];
+        } else if (!(variance.empty())) {
+          box_var_x = static_cast<T>(variance[0]);
+          box_var_y = static_cast<T>(variance[1]);
+          box_var_w = static_cast<T>(variance[2]);
+          box_var_h = static_cast<T>(variance[3]);
         }
+        target_box_center_x =
+            box_var_x * target_box_data[offset] * prior_box_width +
+            prior_box_center_x;
+        target_box_center_y =
+            box_var_y * target_box_data[offset + 1] * prior_box_height +
+            prior_box_center_y;
+        target_box_width =
+            std::exp(box_var_w * target_box_data[offset + 2]) * prior_box_width;
+        target_box_height = std::exp(box_var_h * target_box_data[offset + 3]) *
+                            prior_box_height;
 
         output[offset] = target_box_center_x - target_box_width / 2;
         output[offset + 1] = target_box_center_y - target_box_height / 2;
@@ -157,26 +179,40 @@ class BoxCoderKernel : public framework::OpKernel<T> {
     auto* prior_box_var = context.Input<framework::Tensor>("PriorBoxVar");
     auto* target_box = context.Input<framework::LoDTensor>("TargetBox");
     auto* output_box = context.Output<framework::Tensor>("OutputBox");
-
+    std::vector<float> variance = context.Attr<std::vector<float>>("variance");
+    const int axis = context.Attr<int>("axis");
     if (target_box->lod().size()) {
       PADDLE_ENFORCE_EQ(target_box->lod().size(), 1UL,
                         "Only support 1 level of LoD.");
     }
+    if (prior_box_var) {
+      PADDLE_ENFORCE(variance.empty(),
+                     "Input 'PriorBoxVar' and attribute 'variance' should not"
+                     "be used at the same time.");
+    }
+    if (!(variance.empty())) {
+      PADDLE_ENFORCE(static_cast<int>(variance.size()) == 4,
+                     "Size of attribute 'variance' should be 4");
+    }
+    auto code_type = GetBoxCodeType(context.Attr<std::string>("code_type"));
+    bool normalized = context.Attr<bool>("box_normalized");
+
     auto row = target_box->dims()[0];
     auto col = prior_box->dims()[0];
+    if (code_type == BoxCodeType::kDecodeCenterSize) {
+      col = target_box->dims()[1];
+    }
     auto len = prior_box->dims()[1];
 
     output_box->mutable_data<T>({row, col, len}, context.GetPlace());
 
-    auto code_type = GetBoxCodeType(context.Attr<std::string>("code_type"));
-    bool normalized = context.Attr<bool>("box_normalized");
     T* output = output_box->data<T>();
     if (code_type == BoxCodeType::kEncodeCenterSize) {
       EncodeCenterSize(target_box, prior_box, prior_box_var, normalized,
-                       output);
+                       variance, output);
     } else if (code_type == BoxCodeType::kDecodeCenterSize) {
-      DecodeCenterSize(target_box, prior_box, prior_box_var, normalized,
-                       output);
+      DecodeCenterSize(target_box, prior_box, prior_box_var, normalized, axis,
+                       variance, output);
     }
   }
 };

paddle/fluid/operators/detection/multiclass_nms_op.cc

@@ -9,9 +9,9 @@ http://www.apache.org/licenses/LICENSE-2.0
 Unless required by applicable law or agreed to in writing, software
 distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-
 limitations under the License. */
 
+#include <glog/logging.h>
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detection/poly_util.h"
 
@@ -35,30 +35,45 @@ class MultiClassNMSOp : public framework::OperatorWithKernel {
 
     auto box_dims = ctx->GetInputDim("BBoxes");
     auto score_dims = ctx->GetInputDim("Scores");
+    auto score_size = score_dims.size();
 
     if (ctx->IsRuntime()) {
+      PADDLE_ENFORCE(score_size == 2 || score_size == 3,
+                     "The rank of Input(Scores) must be 2 or 3");
       PADDLE_ENFORCE_EQ(box_dims.size(), 3,
-                        "The rank of Input(BBoxes) must be 3.");
-      PADDLE_ENFORCE_EQ(score_dims.size(), 3,
-                        "The rank of Input(Scores) must be 3.");
-      PADDLE_ENFORCE(box_dims[2] == 4 || box_dims[2] == 8 ||
-                         box_dims[2] == 16 || box_dims[2] == 24 ||
-                         box_dims[2] == 32,
-                     "The 2nd dimension of Input(BBoxes) must be 4 or 8, "
-                     "represents the layout of coordinate "
-                     "[xmin, ymin, xmax, ymax] or "
-                     "4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or "
-                     "8 points: [xi, yi] i= 1,2,...,8 or "
-                     "12 points: [xi, yi] i= 1,2,...,12 or "
-                     "16 points: [xi, yi] i= 1,2,...,16");
-      PADDLE_ENFORCE_EQ(box_dims[1], score_dims[2],
-                        "The 1st dimensiong of Input(BBoxes) must be equal to "
-                        "3rd dimension of Input(Scores), which represents the "
-                        "predicted bboxes.");
+                        "The rank of Input(BBoxes) must be 3");
+      if (score_size == 3) {
+        PADDLE_ENFORCE(box_dims[2] == 4 || box_dims[2] == 8 ||
+                           box_dims[2] == 16 || box_dims[2] == 24 ||
+                           box_dims[2] == 32,
+                       "The last dimension of Input(BBoxes) must be 4 or 8, "
+                       "represents the layout of coordinate "
+                       "[xmin, ymin, xmax, ymax] or "
+                       "4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or "
+                       "8 points: [xi, yi] i= 1,2,...,8 or "
+                       "12 points: [xi, yi] i= 1,2,...,12 or "
+                       "16 points: [xi, yi] i= 1,2,...,16");
+        PADDLE_ENFORCE_EQ(
+            box_dims[1], score_dims[2],
+            "The 2nd dimension of Input(BBoxes) must be equal to "
+            "last dimension of Input(Scores), which represents the "
+            "predicted bboxes.");
+      } else {
+        PADDLE_ENFORCE(box_dims[2] == 4,
+                       "The last dimension of Input(BBoxes) must be 4");
+        PADDLE_ENFORCE_EQ(box_dims[1], score_dims[1],
+                          "The 2nd dimension of Input(BBoxes)"
+                          "must be equal to the 2nd dimension"
+                          " of Input(Scores)");
+      }
     }
     // Here the box_dims[0] is not the real dimension of output.
     // It will be rewritten in the computing kernel.
-    ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2});
+    if (score_size == 3) {
+      ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2});
+    } else {
+      ctx->SetOutputDim("Out", {-1, box_dims[2] + 2});
+    }
   }
 
  protected:
@@ -123,8 +138,9 @@ static inline T JaccardOverlap(const T* box1, const T* box2,
     const T inter_ymin = std::max(box1[1], box2[1]);
     const T inter_xmax = std::min(box1[2], box2[2]);
     const T inter_ymax = std::min(box1[3], box2[3]);
-    const T inter_w = inter_xmax - inter_xmin;
-    const T inter_h = inter_ymax - inter_ymin;
+    T norm = normalized ? static_cast<T>(0.) : static_cast<T>(1.);
+    T inter_w = inter_xmax - inter_xmin + norm;
+    T inter_h = inter_ymax - inter_ymin + norm;
     const T inter_area = inter_w * inter_h;
     const T bbox1_area = BBoxArea<T>(box1, normalized);
     const T bbox2_area = BBoxArea<T>(box2, normalized);
@@ -139,7 +155,7 @@ T PolyIoU(const T* box1, const T* box2, const size_t box_size,
   T bbox2_area = PolyArea<T>(box2, box_size, normalized);
   T inter_area = PolyOverlapArea<T>(box1, box2, box_size, normalized);
   if (bbox1_area == 0 || bbox2_area == 0 || inter_area == 0) {
-    // If coordinate values are is invalid
+    // If coordinate values are invalid
     // if area size <= 0,  return 0.
     return T(0.);
   } else {
@@ -147,12 +163,35 @@ T PolyIoU(const T* box1, const T* box2, const size_t box_size,
   }
 }
 
+template <class T>
+void SliceOneClass(const platform::DeviceContext& ctx,
+                   const framework::Tensor& items, const int class_id,
+                   framework::Tensor* one_class_item) {
+  T* item_data = one_class_item->mutable_data<T>(ctx.GetPlace());
+  const T* items_data = items.data<T>();
+  const int64_t num_item = items.dims()[0];
+  const int class_num = items.dims()[1];
+  if (items.dims().size() == 3) {
+    int item_size = items.dims()[2];
+    for (int i = 0; i < num_item; ++i) {
+      std::memcpy(item_data + i * item_size,
+                  items_data + i * class_num * item_size + class_id * item_size,
+                  sizeof(T) * item_size);
+    }
+  } else {
+    for (int i = 0; i < num_item; ++i) {
+      item_data[i] = items_data[i * class_num + class_id];
+    }
+  }
+}
+
 template <typename T>
 class MultiClassNMSKernel : public framework::OpKernel<T> {
  public:
   void NMSFast(const Tensor& bbox, const Tensor& scores,
                const T score_threshold, const T nms_threshold, const T eta,
-               const int64_t top_k, std::vector<int>* selected_indices) const {
+               const int64_t top_k, std::vector<int>* selected_indices,
+               const bool normalized) const {
     // The total boxes for each instance.
     int64_t num_boxes = bbox.dims()[0];
     // 4: [xmin ymin xmax ymax]
@@ -178,15 +217,16 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
           T overlap = T(0.);
           // 4: [xmin ymin xmax ymax]
           if (box_size == 4) {
-            overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
-                                        bbox_data + kept_idx * box_size, true);
+            overlap =
+                JaccardOverlap<T>(bbox_data + idx * box_size,
+                                  bbox_data + kept_idx * box_size, normalized);
           }
           // 8: [x1 y1 x2 y2 x3 y3 x4 y4] or 16, 24, 32
           if (box_size == 8 || box_size == 16 || box_size == 24 ||
               box_size == 32) {
-            overlap =
-                PolyIoU<T>(bbox_data + idx * box_size,
-                           bbox_data + kept_idx * box_size, box_size, true);
+            overlap = PolyIoU<T>(bbox_data + idx * box_size,
+                                 bbox_data + kept_idx * box_size, box_size,
+                                 normalized);
           }
           keep = overlap <= adaptive_threshold;
         } else {
@@ -205,37 +245,58 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
 
   void MultiClassNMS(const framework::ExecutionContext& ctx,
                      const Tensor& scores, const Tensor& bboxes,
+                     const int scores_size,
                      std::map<int, std::vector<int>>* indices,
                      int* num_nmsed_out) const {
     int64_t background_label = ctx.Attr<int>("background_label");
     int64_t nms_top_k = ctx.Attr<int>("nms_top_k");
     int64_t keep_top_k = ctx.Attr<int>("keep_top_k");
+    bool normalized = ctx.Attr<bool>("normalized");
     T nms_threshold = static_cast<T>(ctx.Attr<float>("nms_threshold"));
     T nms_eta = static_cast<T>(ctx.Attr<float>("nms_eta"));
     T score_threshold = static_cast<T>(ctx.Attr<float>("score_threshold"));
+    auto& dev_ctx = ctx.template device_context<platform::CPUDeviceContext>();
 
-    int64_t class_num = scores.dims()[0];
-    int64_t predict_dim = scores.dims()[1];
     int num_det = 0;
+
+    int64_t class_num = scores_size == 3 ? scores.dims()[0] : scores.dims()[1];
+    Tensor bbox_slice, score_slice;
     for (int64_t c = 0; c < class_num; ++c) {
       if (c == background_label) continue;
-      Tensor score = scores.Slice(c, c + 1);
-      NMSFast(bboxes, score, score_threshold, nms_threshold, nms_eta, nms_top_k,
-              &((*indices)[c]));
+      if (scores_size == 3) {
+        score_slice = scores.Slice(c, c + 1);
+        bbox_slice = bboxes;
+      } else {
+        score_slice.Resize({scores.dims()[0], 1});
+        bbox_slice.Resize({scores.dims()[0], 4});
+        SliceOneClass<T>(dev_ctx, scores, c, &score_slice);
+        SliceOneClass<T>(dev_ctx, bboxes, c, &bbox_slice);
+      }
+      NMSFast(bbox_slice, score_slice, score_threshold, nms_threshold, nms_eta,
+              nms_top_k, &((*indices)[c]), normalized);
+      if (scores_size == 2) {
+        std::stable_sort((*indices)[c].begin(), (*indices)[c].end());
+      }
       num_det += (*indices)[c].size();
     }
 
     *num_nmsed_out = num_det;
     const T* scores_data = scores.data<T>();
     if (keep_top_k > -1 && num_det > keep_top_k) {
+      const T* sdata;
       std::vector<std::pair<float, std::pair<int, int>>> score_index_pairs;
       for (const auto& it : *indices) {
         int label = it.first;
-        const T* sdata = scores_data + label * predict_dim;
+        if (scores_size == 3) {
+          sdata = scores_data + label * scores.dims()[1];
+        } else {
+          score_slice.Resize({scores.dims()[0], 1});
+          SliceOneClass<T>(dev_ctx, scores, label, &score_slice);
+          sdata = score_slice.data<T>();
+        }
         const std::vector<int>& label_indices = it.second;
         for (size_t j = 0; j < label_indices.size(); ++j) {
           int idx = label_indices[j];
-          PADDLE_ENFORCE_LT(idx, predict_dim);
           score_index_pairs.push_back(
               std::make_pair(sdata[idx], std::make_pair(label, idx)));
         }
@@ -252,31 +313,55 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
         int idx = score_index_pairs[j].second.second;
         new_indices[label].push_back(idx);
       }
+      if (scores_size == 2) {
+        for (const auto& it : new_indices) {
+          int label = it.first;
+          std::stable_sort(new_indices[label].begin(),
+                           new_indices[label].end());
+        }
+      }
       new_indices.swap(*indices);
       *num_nmsed_out = keep_top_k;
     }
   }
 
-  void MultiClassOutput(const Tensor& scores, const Tensor& bboxes,
+  void MultiClassOutput(const platform::DeviceContext& ctx,
+                        const Tensor& scores, const Tensor& bboxes,
                         const std::map<int, std::vector<int>>& selected_indices,
-                        Tensor* outs) const {
+                        const int scores_size, Tensor* outs) const {
+    int64_t class_num = scores.dims()[1];
     int64_t predict_dim = scores.dims()[1];
     int64_t box_size = bboxes.dims()[1];
-    int64_t out_dim = bboxes.dims()[1] + 2;
+    if (scores_size == 2) {
+      box_size = bboxes.dims()[2];
+    }
+    int64_t out_dim = box_size + 2;
     auto* scores_data = scores.data<T>();
     auto* bboxes_data = bboxes.data<T>();
     auto* odata = outs->data<T>();
-
+    const T* sdata;
+    Tensor bbox;
+    bbox.Resize({scores.dims()[0], box_size});
     int count = 0;
     for (const auto& it : selected_indices) {
       int label = it.first;
-      const T* sdata = scores_data + label * predict_dim;
       const std::vector<int>& indices = it.second;
+      if (scores_size == 2) {
+        SliceOneClass<T>(ctx, bboxes, label, &bbox);
+      } else {
+        sdata = scores_data + label * predict_dim;
+      }
       for (size_t j = 0; j < indices.size(); ++j) {
         int idx = indices[j];
-        const T* bdata = bboxes_data + idx * box_size;
-        odata[count * out_dim] = label;           // label
-        odata[count * out_dim + 1] = sdata[idx];  // score
+        odata[count * out_dim] = label;  // label
+        const T* bdata;
+        if (scores_size == 3) {
+          bdata = bboxes_data + idx * box_size;
+          odata[count * out_dim + 1] = sdata[idx];  // score
+        } else {
+          bdata = bbox.data<T>() + idx * box_size;
+          odata[count * out_dim + 1] = *(scores_data + idx * class_num + label);
+        }
         // xmin, ymin, xmax, ymax or multi-points coordinates
         std::memcpy(odata + count * out_dim + 2, bdata, box_size * sizeof(T));
         count++;
@@ -285,52 +370,64 @@ class MultiClassNMSKernel : public framework::OpKernel<T> {
   }
 
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto* boxes = ctx.Input<Tensor>("BBoxes");
-    auto* scores = ctx.Input<Tensor>("Scores");
+    auto* boxes = ctx.Input<LoDTensor>("BBoxes");
+    auto* scores = ctx.Input<LoDTensor>("Scores");
     auto* outs = ctx.Output<LoDTensor>("Out");
 
     auto score_dims = scores->dims();
-
-    int64_t batch_size = score_dims[0];
-    int64_t class_num = score_dims[1];
-    int64_t predict_dim = score_dims[2];
-    int64_t box_dim = boxes->dims()[2];
-    int64_t out_dim = boxes->dims()[2] + 2;
+    auto score_size = score_dims.size();
+    auto& dev_ctx = ctx.template device_context<platform::CPUDeviceContext>();
 
     std::vector<std::map<int, std::vector<int>>> all_indices;
     std::vector<size_t> batch_starts = {0};
-    for (int64_t i = 0; i < batch_size; ++i) {
-      Tensor ins_score = scores->Slice(i, i + 1);
-      ins_score.Resize({class_num, predict_dim});
-
-      Tensor ins_boxes = boxes->Slice(i, i + 1);
-      ins_boxes.Resize({predict_dim, box_dim});
-
+    int64_t batch_size = score_dims[0];
+    int64_t box_dim = boxes->dims()[2];
+    int64_t out_dim = box_dim + 2;
+    int num_nmsed_out = 0;
+    Tensor boxes_slice, scores_slice;
+    int n = score_size == 3 ? batch_size : boxes->lod().back().size() - 1;
+    for (int i = 0; i < n; ++i) {
+      if (score_size == 3) {
+        scores_slice = scores->Slice(i, i + 1);
+        scores_slice.Resize({score_dims[1], score_dims[2]});
+        boxes_slice = boxes->Slice(i, i + 1);
+        boxes_slice.Resize({score_dims[2], box_dim});
+      } else {
+        auto boxes_lod = boxes->lod().back();
+        scores_slice = scores->Slice(boxes_lod[i], boxes_lod[i + 1]);
+        boxes_slice = boxes->Slice(boxes_lod[i], boxes_lod[i + 1]);
+      }
       std::map<int, std::vector<int>> indices;
-      int num_nmsed_out = 0;
-      MultiClassNMS(ctx, ins_score, ins_boxes, &indices, &num_nmsed_out);
+      MultiClassNMS(ctx, scores_slice, boxes_slice, score_size, &indices,
+                    &num_nmsed_out);
       all_indices.push_back(indices);
       batch_starts.push_back(batch_starts.back() + num_nmsed_out);
     }
 
     int num_kept = batch_starts.back();
     if (num_kept == 0) {
-      T* od = outs->mutable_data<T>({1}, ctx.GetPlace());
+      T* od = outs->mutable_data<T>({1, 1}, ctx.GetPlace());
       od[0] = -1;
+      batch_starts = {0, 1};
     } else {
       outs->mutable_data<T>({num_kept, out_dim}, ctx.GetPlace());
-      for (int64_t i = 0; i < batch_size; ++i) {
-        Tensor ins_score = scores->Slice(i, i + 1);
-        ins_score.Resize({class_num, predict_dim});
-
-        Tensor ins_boxes = boxes->Slice(i, i + 1);
-        ins_boxes.Resize({predict_dim, box_dim});
-
+      for (int i = 0; i < n; ++i) {
+        if (score_size == 3) {
+          scores_slice = scores->Slice(i, i + 1);
+          boxes_slice = boxes->Slice(i, i + 1);
+          scores_slice.Resize({score_dims[1], score_dims[2]});
+          boxes_slice.Resize({score_dims[2], box_dim});
+        } else {
+          auto boxes_lod = boxes->lod().back();
+          scores_slice = scores->Slice(boxes_lod[i], boxes_lod[i + 1]);
+          boxes_slice = boxes->Slice(boxes_lod[i], boxes_lod[i + 1]);
+        }
         int64_t s = batch_starts[i];
         int64_t e = batch_starts[i + 1];
         if (e > s) {
           Tensor out = outs->Slice(s, e);
-          MultiClassOutput(ins_score, ins_boxes, all_indices[i], &out);
+          MultiClassOutput(dev_ctx, scores_slice, boxes_slice, all_indices[i],
+                           score_dims.size(), &out);
         }
       }
     }
@@ -346,17 +443,24 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() override {
     AddInput("BBoxes",
-             "(Tensor) A 3-D Tensor with shape "
+             "Two types of bboxes are supported:"
+             "1. (Tensor) A 3-D Tensor with shape "
              "[N, M, 4 or 8 16 24 32] represents the "
              "predicted locations of M bounding bboxes, N is the batch size. "
              "Each bounding box has four coordinate values and the layout is "
-             "[xmin, ymin, xmax, ymax], when box size equals to 4.");
+             "[xmin, ymin, xmax, ymax], when box size equals to 4."
+             "2. (LoDTensor) A 3-D Tensor with shape [M, C, 4]"
+             "M is the number of bounding boxes, C is the class number");
     AddInput("Scores",
-             "(Tensor) A 3-D Tensor with shape [N, C, M] represents the "
+             "Two types of scores are supported:"
+             "1. (Tensor) A 3-D Tensor with shape [N, C, M] represents the "
              "predicted confidence predictions. N is the batch size, C is the "
              "class number, M is number of bounding boxes. For each category "
              "there are total M scores which corresponding M bounding boxes. "
-             " Please note, M is equal to the 1st dimension of BBoxes. ");
+             " Please note, M is equal to the 2nd dimension of BBoxes. "
+             "2. (LoDTensor) A 2-D LoDTensor with shape [M, C]. "
+             "M is the number of bbox, C is the class number. In this case, "
+             "Input BBoxes should be the second case with shape [M, C, 4].");
     AddAttr<int>(
         "background_label",
         "(int, defalut: 0) "
@@ -384,6 +488,10 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
                  "(int64_t) "
                  "Number of total bboxes to be kept per image after NMS "
                  "step. -1 means keeping all bboxes after NMS step.");
+    AddAttr<bool>("normalized",
+                  "(bool, default true) "
+                  "Whether detections are normalized.")
+        .SetDefault(true);
     AddOutput("Out",
               "(LoDTensor) A 2-D LoDTensor with shape [No, 6] represents the "
               "detections. Each row has 6 values: "
@@ -399,24 +507,21 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 This operator is to do multi-class non maximum suppression (NMS) on a batched
 of boxes and scores.
-
 In the NMS step, this operator greedily selects a subset of detection bounding
 boxes that have high scores larger than score_threshold, if providing this
 threshold, then selects the largest nms_top_k confidences scores if nms_top_k
 is larger than -1. Then this operator pruns away boxes that have high IOU
 (intersection over union) overlap with already selected boxes by adaptive
 threshold NMS based on parameters of nms_threshold and nms_eta.
-
 Aftern NMS step, at most keep_top_k number of total bboxes are to be kept
 per image if keep_top_k is larger than -1.
-
 This operator support multi-class and batched inputs. It applying NMS
 independently for each class. The outputs is a 2-D LoDTenosr, for each
 image, the offsets in first dimension of LoDTensor are called LoD, the number
 of offset is N + 1, where N is the batch size. If LoD[i + 1] - LoD[i] == 0,
 means there is no detected bbox for this image. If there is no detected boxes
-for all images, all the elements in LoD are 0, and the Out only contains one
-value which is -1.
+for all images, all the elements in LoD are set to {1}, and the Out only 
+contains one value which is -1.
 )DOC");
   }
 };

paddle/fluid/operators/distributed/proto_encoder_helper.h

@@ -85,7 +85,7 @@ class ProtoEncodeHelper {
 #define REPLACE_ENFORCE_GLOG 1
     // Make sure callers didn't do operations that went over max_size promised
     if (paddle::platform::is_error(p_ <= limit_)) {
-      paddle::platform::throw_on_error(p_ <= limit_);
+      paddle::platform::throw_on_error(p_ <= limit_, "");
     }
 #undef REPLACE_ENFORCE_GLOG
   }

paddle/fluid/operators/dropout_op.cu

@@ -114,4 +114,5 @@ REGISTER_OP_CUDA_KERNEL(
     ops::GPUDropoutKernel<plat::CUDADeviceContext, double>);
 REGISTER_OP_CUDA_KERNEL(
     dropout_grad, ops::DropoutGradKernel<plat::CUDADeviceContext, float>,
+    ops::DropoutGradKernel<plat::CUDADeviceContext, plat::float16>,
     ops::DropoutGradKernel<plat::CUDADeviceContext, double>);

paddle/fluid/operators/gather_op.cu

@@ -31,7 +31,7 @@ class GatherOpCUDAKernel : public framework::OpKernel<T> {
     auto *output = ctx.Output<Tensor>("Out");
 
     output->mutable_data<T>(ctx.GetPlace());
-
+    if (x->numel() == 0) return;
     GPUGather<T>(ctx.device_context(), *x, *index, output);
   }
 };
@@ -45,14 +45,13 @@ class GatherGradOpCUDAKernel : public framework::OpKernel<T> {
     auto *Index = ctx.Input<Tensor>("Index");
     auto *dX = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto *dO = ctx.Input<Tensor>(framework::GradVarName("Out"));
-    auto *x = ctx.Input<Tensor>("X");
 
     dX->mutable_data<T>(ctx.GetPlace());
     auto dxt = framework::EigenVector<T>::Flatten(*dX);
     auto &place = *ctx.template device_context<platform::CUDADeviceContext>()
                        .eigen_device();
     dxt.device(place) = dxt.constant(static_cast<T>(0));
-
+    if (dO->numel() == 0) return;
     GPUScatterAssign<T>(ctx.device_context(), *dO, *Index, dX);
   }
 };
@@ -61,11 +60,14 @@ class GatherGradOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel<float>,
                         ops::GatherOpCUDAKernel<double>,
                         ops::GatherOpCUDAKernel<int64_t>,
-                        ops::GatherOpCUDAKernel<int>);
+                        ops::GatherOpCUDAKernel<int>,
+                        ops::GatherOpCUDAKernel<plat::float16>);
 REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel<float>,
                         ops::GatherGradOpCUDAKernel<double>,
                         ops::GatherGradOpCUDAKernel<int64_t>,
-                        ops::GatherGradOpCUDAKernel<int>);
+                        ops::GatherGradOpCUDAKernel<int>,
+                        ops::GatherGradOpCUDAKernel<plat::float16>);

paddle/fluid/operators/gather_op.h

@@ -35,7 +35,7 @@ class GatherOpKernel : public framework::OpKernel<T> {
     auto *output = ctx.Output<Tensor>("Out");
 
     output->mutable_data<T>(ctx.GetPlace());
-
+    if (x->numel() == 0) return;
     CPUGather<T>(ctx.device_context(), *x, *index, output);
   }
 };
@@ -56,7 +56,7 @@ class GatherGradientOpKernel : public framework::OpKernel<T> {
     auto &place = *ctx.template device_context<platform::CPUDeviceContext>()
                        .eigen_device();
     dxt.device(place) = dxt.constant(static_cast<T>(0));
-
+    if (dO->numel() == 0) return;
     ScatterAssign<T>(ctx.device_context(), *dO, *Index, dX);
   }
 };

paddle/fluid/operators/hierarchical_sigmoid_op.h

@@ -136,7 +136,7 @@ class HierarchicalSigmoidOpKernel : public framework::OpKernel<T> {
     sum.mutable_data<T>(framework::make_ddim(sum_dims), ctx.GetPlace());
     auto sum_mat = EigenMatrix<T>::From(sum);
     out->mutable_data<T>(ctx.GetPlace());
-    auto out_mat = framework::EigenVector<T>::Flatten(*out);
+    auto out_mat = framework::EigenMatrix<T>::From(*out);
     if (bias) {
       bit_code->Add(*bias, pre_out);
     }

paddle/fluid/operators/interpolate_op.cc

@@ -82,6 +82,18 @@ class InterpolateOpMaker : public framework::OpProtoAndCheckerMaker {
                          "bilinear interpolation and \"nearest\" for nearest "
                          "neighbor interpolation.")
         .SetDefault("bilinear");
+    AddAttr<bool>(
+        "align_corners",
+        "an optinal bool. Defaults to True. "
+        "If True, the centers of 4 corner pixels of the input and output "
+        "tensors are aligned, preserving the values at the corner pixels, "
+        "if Flase, are not aligned")
+        .SetDefault(true);
+    AddAttr<int>("align_mode",
+                 "(int, default \'1\'), optional for bilinear interpolation"
+                 "can be \'0\' for src_idx = scale*(dst_indx+0.5)-0.5 , "
+                 "can be \'1\' for src_idx = scale*dst_index .")
+        .SetDefault(1);
     AddComment(R"DOC(
           This operator samples input X to given output shape by using specified
           interpolation method, the interpolation methods can be \"nearest\"
@@ -98,6 +110,64 @@ class InterpolateOpMaker : public framework::OpProtoAndCheckerMaker {
           to perform linear interpolation first in one direction, and then 
           again in the other direction.
 
+          Align_corners and align_mode are optinal parameters,the calculation method 
+          of interpolation can be selected by them.
+          
+          Example:
+
+          For scale:
+          
+            if align_corners = True and out_{size}>1 :
+
+              scale_{factor} = (in_{size}-1.0)/(out_{size}-1.0)
+            
+            else:
+              
+              scale_{factor} = float(in_{size}/out_{size})
+            
+          
+          Nearest neighbor interpolation:
+          
+          if:
+              align_corners = False
+
+              input : (N,C,H_in,W_in)
+              output: (N,C,H_out,W_out) where:
+
+              H_out = \left \lfloor {H_{in} * scale_{}factor}} \right \rfloor
+              W_out = \left \lfloor {W_{in} * scale_{}factor}} \right \rfloor
+
+          else:
+              align_corners = True
+
+              input : (N,C,H_in,W_in)
+              output: (N,C,H_out,W_out) where:
+
+              H_out = round(H_{in} * scale_{factor})
+              W_out = round(W_{in} * scale_{factor})
+
+          Bilinear interpolation:
+
+          if:
+              align_corners = False , align_mode = 0
+              
+              input : (N,C,H_in,W_in)
+              output: (N,C,H_out,W_out) where:
+              
+              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
+              W_out = (W_{in}+0.5) * scale_{factor} - 0.5
+
+
+          else:
+           
+              input : (N,C,H_in,W_in)
+              output: (N,C,H_out,W_out) where:
+
+              H_out = H_{in} * scale_{factor}
+              W_out = W_{in} * scale_{factor}
+
+          
+
           For details of nearest neighbor interpolation, please refer to Wikipedia: 
           https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation
 

paddle/fluid/operators/interpolate_op.cu

@@ -23,7 +23,8 @@ __global__ void KeNearestNeighborInterpFw(
     const T* in, const size_t in_img_h, const size_t in_img_w,
     const size_t input_h, const size_t input_w, T* out, const size_t out_img_h,
     const size_t out_img_w, const size_t output_h, const size_t output_w,
-    const size_t num_channels, const float ratio_h, const float ratio_w) {
+    const size_t num_channels, const float ratio_h, const float ratio_w,
+    const bool align_corners) {
   int nthreads = output_h * output_w;
   int tid = blockIdx.x * blockDim.x + threadIdx.x;
   int stride = blockDim.x * gridDim.x;
@@ -35,10 +36,14 @@ __global__ void KeNearestNeighborInterpFw(
     int channel_id = out_id_w / out_img_size;
 
     int out_img_idy = (out_id_w % out_img_size) / out_img_w;
-    int in_img_idy = static_cast<int>(ratio_h * out_img_idy + 0.5);
+    int in_img_idy = (align_corners)
+                         ? static_cast<int>(ratio_h * out_img_idy + 0.5)
+                         : static_cast<int>(ratio_h * out_img_idy);
 
     int out_img_idx = tid % out_img_w;
-    int in_img_idx = static_cast<int>(ratio_w * out_img_idx + 0.5);
+    int in_img_idx = (align_corners)
+                         ? static_cast<int>(ratio_w * out_img_idx + 0.5)
+                         : static_cast<int>(ratio_w * out_img_idx);
 
     out[tid] = in[out_id_h * input_w + channel_id * in_img_size +
                   in_img_idy * in_img_w + in_img_idx];
@@ -50,7 +55,8 @@ __global__ void KeNearestNeighborInterpBw(
     T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h,
     const size_t input_w, const T* out, const size_t out_img_h,
     const size_t out_img_w, const size_t output_h, const size_t output_w,
-    const size_t num_channels, const float ratio_h, const float ratio_w) {
+    const size_t num_channels, const float ratio_h, const float ratio_w,
+    const bool align_corners) {
   int nthreads = output_h * output_w;
   int tid = blockIdx.x * blockDim.x + threadIdx.x;
   int stride = blockDim.x * gridDim.x;
@@ -62,10 +68,14 @@ __global__ void KeNearestNeighborInterpBw(
     int channel_id = out_id_w / out_img_size;
 
     int out_img_idy = (out_id_w % out_img_size) / out_img_w;
-    int in_img_idy = static_cast<int>(ratio_h * out_img_idy + 0.5);
+    int in_img_idy = (align_corners)
+                         ? static_cast<int>(ratio_h * out_img_idy + 0.5)
+                         : static_cast<int>(ratio_h * out_img_idy);
 
     int out_img_idx = tid % out_img_w;
-    int in_img_idx = static_cast<int>(ratio_w * out_img_idx + 0.5);
+    int in_img_idx = (align_corners)
+                         ? static_cast<int>(ratio_w * out_img_idx + 0.5)
+                         : static_cast<int>(ratio_w * out_img_idx);
 
     T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size +
                     in_img_idy * in_img_w + in_img_idx];
@@ -79,10 +89,12 @@ __global__ void KeBilinearInterpFw(
     const T* in, const size_t in_img_h, const size_t in_img_w,
     const size_t input_h, const size_t input_w, T* out, const size_t out_img_h,
     const size_t out_img_w, const size_t output_h, const size_t output_w,
-    const size_t num_channels, const float ratio_h, const float ratio_w) {
+    const size_t num_channels, const float ratio_h, const float ratio_w,
+    const bool align_corners, const int align_mode) {
   int nthreads = output_h * output_w;
   int tid = blockIdx.x * blockDim.x + threadIdx.x;
   int stride = blockDim.x * gridDim.x;
+  bool align_flag = (align_mode == 0 && !align_corners);
   for (; tid < nthreads; tid += stride) {
     int out_id_h = tid / output_w;
     int out_id_w = tid % output_w;
@@ -91,15 +103,23 @@ __global__ void KeBilinearInterpFw(
     int channel_id = out_id_w / out_img_size;
 
     int out_img_idy = (out_id_w % out_img_size) / out_img_w;
-    int in_img_idy = ratio_h * out_img_idy;
+    int in_img_idy = align_flag
+                         ? static_cast<int>(ratio_h * (out_img_idy + 0.5) - 0.5)
+                         : static_cast<int>(ratio_h * out_img_idy);
+    in_img_idy = (in_img_idy > 0) ? in_img_idy : 0;
     int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0;
-    T h1lambda = ratio_h * out_img_idy - in_img_idy;
+    T h1lambda = align_flag ? ratio_h * (out_img_idy + 0.5) - 0.5 - in_img_idy
+                            : ratio_h * out_img_idy - in_img_idy;
     T h2lambda = 1.f - h1lambda;
 
     int out_img_idx = tid % out_img_w;
-    int in_img_idx = ratio_w * out_img_idx;
+    int in_img_idx = align_flag
+                         ? static_cast<int>(ratio_w * (out_img_idx + 0.5) - 0.5)
+                         : static_cast<int>(ratio_w * out_img_idx);
+    in_img_idx = (in_img_idx > 0) ? in_img_idx : 0;
     int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0;
-    T w1lambda = ratio_w * out_img_idx - in_img_idx;
+    T w1lambda = align_flag ? ratio_w * (out_img_idx + 0.5) - 0.5 - in_img_idx
+                            : ratio_w * out_img_idx - in_img_idx;
     T w2lambda = 1.f - w1lambda;
 
     const T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size +
@@ -118,10 +138,12 @@ __global__ void KeBilinearInterpBw(
     T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h,
     const size_t input_w, const T* out, const size_t out_img_h,
     const size_t out_img_w, const size_t output_h, const size_t output_w,
-    const size_t num_channels, const T ratio_h, const T ratio_w) {
+    const size_t num_channels, const T ratio_h, const T ratio_w,
+    const bool align_corners, const int align_mode) {
   int nthreads = output_h * output_w;
   int tid = blockIdx.x * blockDim.x + threadIdx.x;
   int stride = blockDim.x * gridDim.x;
+  bool align_flag = (align_mode == 0 && !align_corners);
   for (; tid < nthreads; tid += stride) {
     int out_id_h = tid / output_w;
     int out_id_w = tid % output_w;
@@ -130,15 +152,22 @@ __global__ void KeBilinearInterpBw(
     int channel_id = out_id_w / out_img_size;
 
     int out_img_idy = (out_id_w % out_img_size) / out_img_w;
-    int in_img_idy = ratio_h * out_img_idy;
+    int in_img_idy = align_flag ? ratio_h * (out_img_idy + 0.5) - 0.5
+                                : ratio_h * out_img_idy;
+    in_img_idy = (in_img_idy > 0) ? in_img_idy : 0;
     int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0;
-    T h1lambda = ratio_h * out_img_idy - in_img_idy;
+    T h1lambda = align_flag ? ratio_h * (out_img_idy + 0.5) - 0.5 - in_img_idy
+                            : ratio_h * out_img_idy - in_img_idy;
+
     T h2lambda = 1.f - h1lambda;
 
     int out_img_idx = tid % out_img_w;
-    int in_img_idx = ratio_w * out_img_idx;
+    int in_img_idx = align_flag ? ratio_w * (out_img_idx + 0.5) - 0.5
+                                : ratio_w * out_img_idx;
+    in_img_idx = (in_img_idx > 0) ? in_img_idx : 0;
     int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0;
-    T w1lambda = ratio_w * out_img_idx - in_img_idx;
+    T w1lambda = align_flag ? ratio_w * (out_img_idx + 0.5) - 0.5 - in_img_idx
+                            : ratio_w * out_img_idx - in_img_idx;
     T w2lambda = 1.f - w1lambda;
 
     T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size +
@@ -175,6 +204,9 @@ class InterpolateOpCUDAKernel : public framework::OpKernel<T> {
       out_w = size_data[1];
     }
 
+    bool align_corners = ctx.Attr<bool>("align_corners");
+    int align_mode = ctx.Attr<int>("align_mode");
+
     int n = input->dims()[0];
     int c = input->dims()[1];
     int in_h = input->dims()[2];
@@ -188,10 +220,16 @@ class InterpolateOpCUDAKernel : public framework::OpKernel<T> {
     int in_chw = c * in_hw;
     int out_chw = c * out_hw;
 
-    float ratio_h =
-        (out_h > 1) ? static_cast<float>(in_h - 1) / (out_h - 1) : 0.f;
-    float ratio_w =
-        (out_w > 1) ? static_cast<float>(in_w - 1) / (out_w - 1) : 0.f;
+    float ratio_h = 0.f;
+    float ratio_w = 0.f;
+    if (out_h > 1) {
+      ratio_h = (align_corners) ? static_cast<float>(in_h - 1) / (out_h - 1)
+                                : static_cast<float>(in_h) / out_h;
+    }
+    if (out_w > 1) {
+      ratio_w = (align_corners) ? static_cast<float>(in_w - 1) / (out_w - 1)
+                                : static_cast<float>(in_w) / out_w;
+    }
 
     if (in_h == out_h && in_w == out_w) {
       framework::TensorCopy(*input, ctx.GetPlace(), output);
@@ -206,12 +244,12 @@ class InterpolateOpCUDAKernel : public framework::OpKernel<T> {
       KeNearestNeighborInterpFw<
           T><<<grid_dim, 512, 0, ctx.cuda_device_context().stream()>>>(
           input_data, in_h, in_w, n, in_chw, output_data, out_h, out_w, n,
-          out_chw, c, ratio_h, ratio_w);
+          out_chw, c, ratio_h, ratio_w, align_corners);
     } else if ("bilinear" == interp_method) {
       KeBilinearInterpFw<
           T><<<grid_dim, 512, 0, ctx.cuda_device_context().stream()>>>(
           input_data, in_h, in_w, n, in_chw, output_data, out_h, out_w, n,
-          out_chw, c, ratio_h, ratio_w);
+          out_chw, c, ratio_h, ratio_w, align_corners, align_mode);
     }
   }
 };
@@ -234,6 +272,10 @@ class InterpolateGradOpCUDAKernel : public framework::OpKernel<T> {
     int out_h = ctx.Attr<int>("out_h");
     int out_w = ctx.Attr<int>("out_w");
     auto out_size = ctx.Input<Tensor>("OutSize");
+
+    bool align_corners = ctx.Attr<bool>("align_corners");
+    int align_mode = ctx.Attr<int>("align_mode");
+
     if (out_size != nullptr) {
       Tensor sizes;
       framework::TensorCopy(*out_size, platform::CPUPlace(), &sizes);
@@ -252,10 +294,16 @@ class InterpolateGradOpCUDAKernel : public framework::OpKernel<T> {
     int in_chw = c * in_hw;
     int out_chw = c * out_hw;
 
-    float ratio_h =
-        (out_h > 1) ? static_cast<float>(in_h - 1) / (out_h - 1) : 0.f;
-    float ratio_w =
-        (out_w > 1) ? static_cast<float>(in_w - 1) / (out_w - 1) : 0.f;
+    float ratio_h = 0.f;
+    float ratio_w = 0.f;
+    if (out_h > 1) {
+      ratio_h = (align_corners) ? static_cast<float>(in_h - 1) / (out_h - 1)
+                                : static_cast<float>(in_h) / out_h;
+    }
+    if (out_w > 1) {
+      ratio_w = (align_corners) ? static_cast<float>(in_w - 1) / (out_w - 1)
+                                : static_cast<float>(in_w) / out_w;
+    }
 
     if (in_h == out_h && in_w == out_w) {
       framework::TensorCopy(*output_grad, ctx.GetPlace(), input_grad);
@@ -270,12 +318,12 @@ class InterpolateGradOpCUDAKernel : public framework::OpKernel<T> {
       KeNearestNeighborInterpBw<
           T><<<grid_dim, 512, 0, ctx.cuda_device_context().stream()>>>(
           input_grad_data, in_h, in_w, n, in_chw, output_grad_data, out_h,
-          out_w, n, out_chw, c, ratio_h, ratio_w);
+          out_w, n, out_chw, c, ratio_h, ratio_w, align_corners);
     } else if ("bilinear" == interp_method) {
       KeBilinearInterpBw<
           T><<<grid_dim, 512, 0, ctx.cuda_device_context().stream()>>>(
           input_grad_data, in_h, in_w, n, in_chw, output_grad_data, out_h,
-          out_w, n, out_chw, c, ratio_h, ratio_w);
+          out_w, n, out_chw, c, ratio_h, ratio_w, align_corners, align_mode);
     }
   }
 };

paddle/fluid/operators/interpolate_op.h

@@ -26,14 +26,17 @@ template <typename T>
 static void NearestNeighborInterpolate(const Tensor& input, Tensor* output,
                                        const float ratio_h, const float ratio_w,
                                        const int n, const int c,
-                                       const int out_h, const int out_w) {
+                                       const int out_h, const int out_w,
+                                       const bool align_corners) {
   auto input_t = EigenTensor<T, 4>::From(input);
   auto output_t = EigenTensor<T, 4>::From(*output);
   for (int k = 0; k < out_h; k++) {  // loop for images
-    int in_k = static_cast<int>(ratio_h * k + 0.5);
+    int in_k = (align_corners) ? static_cast<int>(ratio_h * k + 0.5)
+                               : static_cast<int>(ratio_h * k);
 
     for (int l = 0; l < out_w; l++) {
-      int in_l = static_cast<int>(ratio_w * l + 0.5);
+      int in_l = (align_corners) ? static_cast<int>(ratio_w * l + 0.5)
+                                 : static_cast<int>(ratio_w * l);
 
       for (int i = 0; i < n; i++) {    // loop for batches
         for (int j = 0; j < c; j++) {  // loop for channels
@@ -48,20 +51,29 @@ template <typename T>
 static void BilinearInterpolation(const Tensor& input, Tensor* output,
                                   const float ratio_h, const float ratio_w,
                                   const int in_h, const int in_w, const int n,
-                                  const int c, const int out_h,
-                                  const int out_w) {
+                                  const int c, const int out_h, const int out_w,
+                                  const bool align_corners,
+                                  const bool align_mode) {
   auto input_t = EigenTensor<T, 4>::From(input);
   auto output_t = EigenTensor<T, 4>::From(*output);
+  bool align_flag = (align_mode == 0 && !align_corners);
   for (int k = 0; k < out_h; k++) {  // loop for images
-    int y_n = static_cast<int>(ratio_h * k);
+    int y_n = align_flag ? static_cast<int>(ratio_h * (k + 0.5) - 0.5)
+                         : static_cast<int>(ratio_h * k);
+    y_n = (y_n > 0) ? y_n : 0;
     int y_s = (y_n + 1) < (in_h - 1) ? (y_n + 1) : (in_h - 1);
-    float d_n = ratio_h * k - y_n;
+    float d_n =
+        align_flag ? ratio_h * (k + 0.5) - 0.5 - y_n : ratio_h * k - y_n;
     float d_s = 1.f - d_n;
 
     for (int l = 0; l < out_w; l++) {
-      int x_w = static_cast<int>(ratio_w * l);
+      int x_w = (align_mode == 0 && !align_corners)
+                    ? static_cast<int>(ratio_w * (l + 0.5) - 0.5)
+                    : static_cast<int>(ratio_w * l);
+      x_w = (x_w > 0) ? x_w : 0;
       int x_e = (x_w + 1) < (in_w - 1) ? (x_w + 1) : (in_w - 1);
-      float d_w = ratio_w * l - x_w;
+      float d_w =
+          align_flag ? ratio_w * (l + 0.5) - 0.5 - x_w : ratio_w * l - x_w;
       float d_e = 1.f - d_w;
 
       for (int i = 0; i < n; i++) {    // loop for batches
@@ -78,19 +90,20 @@ static void BilinearInterpolation(const Tensor& input, Tensor* output,
 }
 
 template <typename T>
-static void NearestNeighborInterpolateGrad(const Tensor& output_grad,
-                                           Tensor* input_grad,
-                                           const float ratio_h,
-                                           const float ratio_w, const int n,
-                                           const int c, const int out_h,
-                                           const int out_w) {
+static void NearestNeighborInterpolateGrad(
+    const Tensor& output_grad, Tensor* input_grad, const float ratio_h,
+    const float ratio_w, const int n, const int c, const int out_h,
+    const int out_w, const bool align_corners) {
   auto input_grad_t = EigenTensor<T, 4>::From(*input_grad);
   auto output_grad_t = EigenTensor<T, 4>::From(output_grad);
+
   for (int k = 0; k < out_h; k++) {  // loop for images
-    int in_k = static_cast<int>(ratio_h * k + 0.5);
+    int in_k = (align_corners) ? static_cast<int>(ratio_h * k + 0.5)
+                               : static_cast<int>(ratio_h * k);
 
     for (int l = 0; l < out_w; l++) {
-      int in_l = static_cast<int>(ratio_w * l + 0.5);
+      int in_l = (align_corners) ? static_cast<int>(ratio_w * l + 0.5)
+                                 : static_cast<int>(ratio_w * l);
 
       for (int i = 0; i < n; i++) {    // loop for batches
         for (int j = 0; j < c; j++) {  // loop for channels
@@ -106,19 +119,28 @@ static void BilinearInterpolationGrad(const Tensor& output_grad,
                                       Tensor* input_grad, const float ratio_h,
                                       const float ratio_w, const int in_h,
                                       const int in_w, const int n, const int c,
-                                      const int out_h, const int out_w) {
+                                      const int out_h, const int out_w,
+                                      const bool align_corners,
+                                      const int align_mode) {
   auto input_grad_t = EigenTensor<T, 4>::From(*input_grad);
   auto output_grad_t = EigenTensor<T, 4>::From(output_grad);
+  bool align_flag = (align_mode == 0 && !align_corners);
   for (int k = 0; k < out_h; k++) {  // loop for images
-    int y_n = static_cast<int>(ratio_h * k);
+    int y_n = align_flag ? static_cast<int>(ratio_h * (k + 0.5) - 0.5)
+                         : static_cast<int>(ratio_h * k);
+    y_n = (y_n > 0) ? y_n : 0;
     int y_s = (y_n + 1) < (in_h - 1) ? (y_n + 1) : (in_h - 1);
-    float d_n = ratio_h * k - y_n;
+    float d_n =
+        align_flag ? ratio_h * (k + 0.5) - 0.5 - y_n : ratio_h * k - y_n;
     float d_s = 1.f - d_n;
 
     for (int l = 0; l < out_w; l++) {
-      int x_w = static_cast<int>(ratio_w * l);
+      int x_w = align_flag ? static_cast<int>(ratio_w * (l + 0.5) - 0.5)
+                           : static_cast<int>(ratio_w * l);
+      x_w = (x_w > 0) ? x_w : 0;
       int x_e = (x_w + 1) < (in_w - 1) ? (x_w + 1) : (in_w - 1);
-      float d_w = ratio_w * l - x_w;
+      float d_w =
+          align_flag ? ratio_w * (l + 0.5) - 0.5 - x_w : ratio_w * l - x_w;
       float d_e = 1.f - d_w;
 
       for (int i = 0; i < n; i++) {    // loop for batches
@@ -134,7 +156,6 @@ static void BilinearInterpolationGrad(const Tensor& output_grad,
     }
   }
 }
-
 template <typename T>
 class InterpolateKernel : public framework::OpKernel<T> {
  public:
@@ -151,6 +172,8 @@ class InterpolateKernel : public framework::OpKernel<T> {
       out_h = out_size_data[0];
       out_w = out_size_data[1];
     }
+    bool align_corners = ctx.Attr<bool>("align_corners");
+    int align_mode = ctx.Attr<int>("align_mode");
 
     const int n = input->dims()[0];
     const int c = input->dims()[1];
@@ -168,17 +191,24 @@ class InterpolateKernel : public framework::OpKernel<T> {
       return;
     }
 
-    float ratio_h =
-        (out_h > 1) ? static_cast<float>(in_h - 1) / (out_h - 1) : 0.f;
-    float ratio_w =
-        (out_w > 1) ? static_cast<float>(in_w - 1) / (out_w - 1) : 0.f;
+    float ratio_h = 0.f;
+    float ratio_w = 0.f;
+
+    if (out_h > 1) {
+      ratio_h = (align_corners) ? static_cast<float>(in_h - 1) / (out_h - 1)
+                                : static_cast<float>(in_h) / out_h;
+    }
+    if (out_w > 1) {
+      ratio_w = (align_corners) ? static_cast<float>(in_w - 1) / (out_w - 1)
+                                : static_cast<float>(in_w) / out_w;
+    }
 
     if ("bilinear" == interp_method) {
       BilinearInterpolation<T>(*input, output, ratio_h, ratio_w, in_h, in_w, n,
-                               c, out_h, out_w);
+                               c, out_h, out_w, align_corners, align_mode);
     } else if ("nearest" == interp_method) {
       NearestNeighborInterpolate<T>(*input, output, ratio_h, ratio_w, n, c,
-                                    out_h, out_w);
+                                    out_h, out_w, align_corners);
     }
   }
 };
@@ -200,6 +230,8 @@ class InterpolateGradKernel : public framework::OpKernel<T> {
       out_h = out_size_data[0];
       out_w = out_size_data[1];
     }
+    bool align_corners = ctx.Attr<bool>("align_corners");
+    int align_mode = ctx.Attr<int>("align_mode");
 
     const int n = input->dims()[0];
     const int c = input->dims()[1];
@@ -217,17 +249,26 @@ class InterpolateGradKernel : public framework::OpKernel<T> {
       return;
     }
 
-    float ratio_h =
-        (out_h > 1) ? static_cast<float>(in_h - 1) / (out_h - 1) : 0.f;
-    float ratio_w =
-        (out_w > 1) ? static_cast<float>(in_w - 1) / (out_w - 1) : 0.f;
+    float ratio_h = 0.f;
+    float ratio_w = 0.f;
+
+    if (out_h > 1) {
+      ratio_h = (align_corners) ? static_cast<float>(in_h - 1) / (out_h - 1)
+                                : static_cast<float>(in_h) / out_h;
+    }
+    if (out_w > 1) {
+      ratio_w = (align_corners) ? static_cast<float>(in_w - 1) / (out_w - 1)
+                                : static_cast<float>(in_w) / out_w;
+    }
 
     if ("bilinear" == interp_method) {
       BilinearInterpolationGrad<T>(*output_grad, input_grad, ratio_h, ratio_w,
-                                   in_h, in_w, n, c, out_h, out_w);
+                                   in_h, in_w, n, c, out_h, out_w,
+                                   align_corners, align_mode);
     } else if ("nearest" == interp_method) {
       NearestNeighborInterpolateGrad<T>(*output_grad, input_grad, ratio_h,
-                                        ratio_w, n, c, out_h, out_w);
+                                        ratio_w, n, c, out_h, out_w,
+                                        align_corners);
     }
   }
 };

paddle/fluid/operators/jit/benchmark.cc

@@ -158,7 +158,7 @@ void BenchAllImpls(const typename KernelTuples::attr_type& attr, Args... args) {
 
 using Tensor = paddle::framework::Tensor;
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void BenchXYZNKernel() {
   for (int d : TestSizes()) {
     Tensor x, y, z;
@@ -175,7 +175,7 @@ void BenchXYZNKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void BenchAXYNKernel() {
   for (int d : TestSizes()) {
     const T a = static_cast<T>(3);
@@ -187,10 +187,23 @@ void BenchAXYNKernel() {
     RandomVec<T>(d, x_data);
     BenchAllImpls<KT, jit::AXYNTuples<T>, PlaceType>(d, &a, x.data<T>(), y_data,
                                                      d);
+    // test inplace
+    BenchAllImpls<KT, jit::AXYNTuples<T>, PlaceType>(d, &a, x.data<T>(), x_data,
+                                                     d);
+  }
+}
+
+template <jit::KernelType KT, typename T, typename PlaceType>
+void BenchXRNKernel() {
+  for (int d : TestSizes()) {
+    Tensor x;
+    RandomVec<T>(d, x.mutable_data<T>({d}, PlaceType()));
+    T res;
+    BenchAllImpls<KT, jit::XRNTuples<T>, PlaceType>(d, x.data<T>(), &res, d);
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void BenchXYNKernel() {
   for (int d : TestSizes()) {
     Tensor x, y;
@@ -203,7 +216,7 @@ void BenchXYNKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void BenchLSTMKernel() {
   for (bool use_peephole : {true, false}) {
     for (int d : TestSizes()) {
@@ -240,7 +253,7 @@ void BenchLSTMKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void BenchGRUKernel() {
   for (int d : TestSizes()) {
     const jit::gru_attr_t attr(d, jit::kVSigmoid, jit::kVTanh);
@@ -262,7 +275,7 @@ void BenchGRUKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void BenchSeqPoolKernel() {
   std::vector<jit::SeqPoolType> pool_types = {
       jit::SeqPoolType::kSum, jit::SeqPoolType::kAvg, jit::SeqPoolType::kSqrt};
@@ -284,7 +297,7 @@ void BenchSeqPoolKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void BenchMatMulKernel() {
   for (int m : {1, 2, 3, 4}) {
     for (int n : TestSizes()) {
@@ -305,57 +318,64 @@ void BenchMatMulKernel() {
   }
 }
 
+template <jit::KernelType KT, typename T, typename PlaceType>
+void BenchSoftmaxKernel() {
+  for (int bs : {1, 2, 10}) {
+    for (int n : TestSizes()) {
+      Tensor x, y;
+      x.Resize({bs, n});
+      y.Resize({bs, n});
+      RandomVec<T>(bs * n, x.mutable_data<T>(PlaceType()), -2.f, 2.f);
+      const T* x_data = x.data<T>();
+      T* y_data = y.mutable_data<T>(PlaceType());
+      BenchAllImpls<KT, jit::SoftmaxTuples<T>, PlaceType>(n, x_data, y_data, n,
+                                                          bs);
+    }
+  }
+}
+
 using T = float;
-using PlaceType = paddle::platform::CPUPlace;
+using CPUPlace = paddle::platform::CPUPlace;
 
 // xyzn
-BENCH_FP32_CPU(kVMul) { BenchXYZNKernel<jit::kVMul, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kVAdd) { BenchXYZNKernel<jit::kVAdd, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kVAddRelu) { BenchXYZNKernel<jit::kVAddRelu, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kVSub) { BenchXYZNKernel<jit::kVSub, T, PlaceType>(); }
+BENCH_FP32_CPU(kVMul) { BenchXYZNKernel<jit::kVMul, T, CPUPlace>(); }
+BENCH_FP32_CPU(kVAdd) { BenchXYZNKernel<jit::kVAdd, T, CPUPlace>(); }
+BENCH_FP32_CPU(kVAddRelu) { BenchXYZNKernel<jit::kVAddRelu, T, CPUPlace>(); }
+BENCH_FP32_CPU(kVSub) { BenchXYZNKernel<jit::kVSub, T, CPUPlace>(); }
 
 // axyn
-BENCH_FP32_CPU(kVScal) { BenchAXYNKernel<jit::kVScal, T, PlaceType>(); }
+BENCH_FP32_CPU(kVScal) { BenchAXYNKernel<jit::kVScal, T, CPUPlace>(); }
+BENCH_FP32_CPU(kVAddBias) { BenchAXYNKernel<jit::kVAddBias, T, CPUPlace>(); }
 
-BENCH_FP32_CPU(kVAddBias) { BenchAXYNKernel<jit::kVAddBias, T, PlaceType>(); }
+// xrn
+BENCH_FP32_CPU(kHSum) { BenchXRNKernel<jit::kHSum, T, CPUPlace>(); }
+BENCH_FP32_CPU(kHMax) { BenchXRNKernel<jit::kHMax, T, CPUPlace>(); }
 
 // xyn
-BENCH_FP32_CPU(kVRelu) { BenchXYNKernel<jit::kVRelu, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kVIdentity) { BenchXYNKernel<jit::kVIdentity, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kVSquare) { BenchXYNKernel<jit::kVSquare, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kVExp) { BenchXYNKernel<jit::kVExp, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kVSigmoid) { BenchXYNKernel<jit::kVSigmoid, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kVTanh) { BenchXYNKernel<jit::kVTanh, T, PlaceType>(); }
+BENCH_FP32_CPU(kVRelu) { BenchXYNKernel<jit::kVRelu, T, CPUPlace>(); }
+BENCH_FP32_CPU(kVIdentity) { BenchXYNKernel<jit::kVIdentity, T, CPUPlace>(); }
+BENCH_FP32_CPU(kVSquare) { BenchXYNKernel<jit::kVSquare, T, CPUPlace>(); }
+BENCH_FP32_CPU(kVExp) { BenchXYNKernel<jit::kVExp, T, CPUPlace>(); }
+BENCH_FP32_CPU(kVSigmoid) { BenchXYNKernel<jit::kVSigmoid, T, CPUPlace>(); }
+BENCH_FP32_CPU(kVTanh) { BenchXYNKernel<jit::kVTanh, T, CPUPlace>(); }
 
 // lstm and peephole
-BENCH_FP32_CPU(kLSTMCtHt) { BenchLSTMKernel<jit::kLSTMCtHt, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kLSTMC1H1) { BenchLSTMKernel<jit::kLSTMC1H1, T, PlaceType>(); }
+BENCH_FP32_CPU(kLSTMCtHt) { BenchLSTMKernel<jit::kLSTMCtHt, T, CPUPlace>(); }
+BENCH_FP32_CPU(kLSTMC1H1) { BenchLSTMKernel<jit::kLSTMC1H1, T, CPUPlace>(); }
 
 // gru functions
-BENCH_FP32_CPU(kGRUH1) { BenchGRUKernel<jit::kGRUH1, T, PlaceType>(); }
-
-BENCH_FP32_CPU(kGRUHtPart1) {
-  BenchGRUKernel<jit::kGRUHtPart1, T, PlaceType>();
-}
-
-BENCH_FP32_CPU(kGRUHtPart2) {
-  BenchGRUKernel<jit::kGRUHtPart2, T, PlaceType>();
-}
+BENCH_FP32_CPU(kGRUH1) { BenchGRUKernel<jit::kGRUH1, T, CPUPlace>(); }
+BENCH_FP32_CPU(kGRUHtPart1) { BenchGRUKernel<jit::kGRUHtPart1, T, CPUPlace>(); }
+BENCH_FP32_CPU(kGRUHtPart2) { BenchGRUKernel<jit::kGRUHtPart2, T, CPUPlace>(); }
 
 // seq pool function
-BENCH_FP32_CPU(kSeqPool) { BenchSeqPoolKernel<jit::kSeqPool, T, PlaceType>(); }
+BENCH_FP32_CPU(kSeqPool) { BenchSeqPoolKernel<jit::kSeqPool, T, CPUPlace>(); }
 
 // matmul
-BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel<jit::kMatMul, T, PlaceType>(); }
+BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel<jit::kMatMul, T, CPUPlace>(); }
+
+// softmax
+BENCH_FP32_CPU(kSoftmax) { BenchSoftmaxKernel<jit::kSoftmax, T, CPUPlace>(); }
 
 // Benchmark all jit kernels including jitcode, mkl and refer.
 // To use this tool, run command: ./benchmark [options...]

paddle/fluid/operators/jit/gen/CMakeLists.txt

@@ -28,3 +28,5 @@ USE_JITKERNEL_GEN(kGRUHtPart1)
 USE_JITKERNEL_GEN(kGRUHtPart2)
 USE_JITKERNEL_GEN(kNCHW16CMulNC)
 USE_JITKERNEL_GEN(kSeqPool)
+USE_JITKERNEL_GEN(kHMax)
+USE_JITKERNEL_GEN(kHSum)

paddle/fluid/operators/jit/gen/act.cc

@@ -81,9 +81,7 @@ void VActJitCode::genCode() {
 #define DECLARE_ACT_CREATOR(name)                                            \
   class name##Creator : public JitCodeCreator<int> {                         \
    public:                                                                   \
-    bool UseMe(const int& attr) const override {                             \
-      return platform::MayIUse(platform::avx);                               \
-    }                                                                        \
+    bool UseMe(const int& attr) const override;                              \
     size_t CodeSize(const int& d) const override;                            \
     std::unique_ptr<GenBase> CreateJitCode(const int& attr) const override { \
       return make_unique<name##JitCode>(attr, CodeSize(attr));               \
@@ -98,6 +96,30 @@ DECLARE_ACT_CREATOR(VSigmoid);
 DECLARE_ACT_CREATOR(VTanh);
 
 // TODO(TJ): tuning use me
+bool VReluCreator::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx);
+}
+
+bool VSquareCreator::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx);
+}
+
+bool VIdentityCreator::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx);
+}
+
+bool VExpCreator::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx) && d < 32;
+}
+
+bool VSigmoidCreator::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx);
+}
+
+bool VTanhCreator::UseMe(const int& d) const {
+  return platform::MayIUse(platform::avx);
+}
+
 size_t VReluCreator::CodeSize(const int& d) const {
   return 96 /* init size */ +
          (d / YMM_FLOAT_BLOCK + 3) * 4 /* instructions */ *

paddle/fluid/operators/jit/gen/hopv.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License. */
+
+#include "paddle/fluid/operators/jit/gen/hopv.h"
+#include "paddle/fluid/operators/jit/registry.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+void HOPVJitCode::genCode() {
+  const int num_blocks = num_ / YMM_FLOAT_BLOCK;
+  int offset = 0;
+
+  if (num_blocks > 0) {
+    // load one firstly
+    vmovups(ymm_tmp, ptr[param_src]);
+    offset += sizeof(float) * YMM_FLOAT_BLOCK;
+    for (int i = 1; i < num_blocks; ++i) {
+      vmovups(ymm_src, ptr[param_src + offset]);
+      process(ymm_tmp, ymm_src, ymm_tmp);
+      offset += sizeof(float) * YMM_FLOAT_BLOCK;
+    }
+    vextractf128(xmm_dst, ymm_tmp, 1);
+    process(xmm_dst, xmm_dst, xmm_tmp);
+  } else {
+    if (type_ == operand_type::MAX) {
+      vbroadcastss(ymm_dst, ptr[param_src]);
+    } else if (type_ == operand_type::ADD) {
+      vxorps(ymm_dst, ymm_dst, ymm_dst);
+    }
+  }
+
+  int rest = num_ % YMM_FLOAT_BLOCK;
+  if (rest >= 4) {
+    vmovups(xmm_src, ptr[param_src + offset]);
+    offset += sizeof(float) * 4;
+    rest -= 4;
+    process(xmm_dst, xmm_dst, xmm_src);
+  }
+
+  vpermilps(xmm_tmp, xmm_dst, 16 + 8 + 3);
+  process(xmm_dst, xmm_dst, xmm_tmp);
+
+  if (rest >= 2) {
+    vmovq(xmm_src, ptr[param_src + offset]);
+    offset += sizeof(float) * 2;
+    rest -= 2;
+    process(xmm_dst, xmm_dst, xmm_src);
+  }
+
+  vpermilps(xmm_tmp, xmm_dst, 1);
+  process(xmm_dst, xmm_dst, xmm_tmp);
+
+  if (rest >= 1) {
+    vmovss(xmm_src, ptr[param_src + offset]);
+    process(xmm_dst, xmm_dst, xmm_src);
+  }
+  vmovss(ptr[param_dst], xmm_dst);
+  ret();
+}
+
+#define DECLARE_HOP_CREATOR(name)                                            \
+  class name##Creator : public JitCodeCreator<int> {                         \
+   public:                                                                   \
+    bool UseMe(const int& attr) const override {                             \
+      return platform::MayIUse(platform::avx);                               \
+    }                                                                        \
+    size_t CodeSize(const int& d) const override {                           \
+      return 96 + d / YMM_FLOAT_BLOCK * 4 * 8;                               \
+    }                                                                        \
+    std::unique_ptr<GenBase> CreateJitCode(const int& attr) const override { \
+      return make_unique<name##JitCode>(attr, CodeSize(attr));               \
+    }                                                                        \
+  }
+
+DECLARE_HOP_CREATOR(HMax);
+DECLARE_HOP_CREATOR(HSum);
+
+#undef DECLARE_HOP_CREATOR
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle
+
+namespace gen = paddle::operators::jit::gen;
+
+REGISTER_JITKERNEL_GEN(kHMax, gen::HMaxCreator);
+REGISTER_JITKERNEL_GEN(kHSum, gen::HSumCreator);

paddle/fluid/operators/jit/gen/hopv.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License. */
+
+#pragma once
+
+#include <string>
+#include "glog/logging.h"
+#include "paddle/fluid/operators/jit/gen/jitcode.h"
+
+namespace paddle {
+namespace operators {
+namespace jit {
+namespace gen {
+
+// horizontal operand vector
+class HOPVJitCode : public JitCode {
+ public:
+  explicit HOPVJitCode(int d, operand_type type, size_t code_size = 256 * 1024,
+                       void* code_ptr = nullptr)
+      : JitCode(code_size, code_ptr), num_(d), type_(type) {
+    if (!(type_ == operand_type::MAX || type_ == operand_type::ADD)) {
+      LOG(FATAL) << "Do not support this operand type: " << type_;
+    }
+    this->genCode();
+  }
+
+  virtual const char* name() const {
+    std::string base = "VXXJitCode";
+    if (type_ == operand_type::MAX) {
+      base += "_MAX";
+    } else {
+      base += "_SUM";
+    }
+    return base.c_str();
+  }
+  void genCode() override;
+
+ protected:
+  template <typename JMM>
+  void process(JMM& dst, JMM& src1, JMM& src2) {  // NOLINT
+    if (type_ == operand_type::MAX) {
+      vmaxps(dst, src1, src2);
+    } else if (type_ == operand_type::ADD) {
+      vaddps(dst, src1, src2);
+    }
+  }
+
+ private:
+  int num_;
+  operand_type type_;
+  reg64_t param_src{abi_param1};
+  reg64_t param_dst{abi_param2};
+  reg64_t param_attr{abi_param3};
+
+  ymm_t ymm_tmp = ymm_t(0);
+  ymm_t ymm_src = ymm_t(1);
+  ymm_t ymm_dst = ymm_t(2);
+
+  xmm_t xmm_tmp = xmm_t(0);
+  xmm_t xmm_src = xmm_t(1);
+  xmm_t xmm_dst = xmm_t(2);
+};
+
+#define DECLARE_HOP_JITCODE(name, op_type)                                    \
+  class name##JitCode : public HOPVJitCode {                                  \
+   public:                                                                    \
+    explicit name##JitCode(int d, size_t code_size, void* code_ptr = nullptr) \
+        : HOPVJitCode(d, op_type, code_size, code_ptr) {}                     \
+  };
+
+DECLARE_HOP_JITCODE(HMax, operand_type::MAX);
+DECLARE_HOP_JITCODE(HSum, operand_type::ADD);
+
+#undef DECLARE_HOP_JITCODE
+
+}  // namespace gen
+}  // namespace jit
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/jit/gen/jitcode.h

@@ -47,6 +47,7 @@ using Label = Xbyak::Label;
 
 typedef enum {
   MUL = 0,
+  MAX,
   ADD,
   SUB,
   RELU,

paddle/fluid/operators/jit/helper.cc

@@ -49,6 +49,9 @@ const char* to_string(KernelType kt) {
     ONE_CASE(kNCHW16CMulNC);
     ONE_CASE(kSeqPool);
     ONE_CASE(kMatMul);
+    ONE_CASE(kHMax);
+    ONE_CASE(kHSum);
+    ONE_CASE(kSoftmax);
     default:
       PADDLE_THROW("Not support type: %d, or forget to add it.", kt);
       return "NOT JITKernel";

paddle/fluid/operators/jit/helper.h

@@ -118,6 +118,28 @@ typename KernelTuples::func_type Get(
   return GetRefer<KT, KernelTuples>();
 }
 
+template <KernelType KT, typename KernelTuples>
+class KernelFuncsCache {
+ public:
+  KernelFuncsCache() = default;
+  static KernelFuncsCache& Instance() {
+    static thread_local KernelFuncsCache<KT, KernelTuples> g_func_cache;
+    return g_func_cache;
+  }
+
+  bool Has(int key) const { return funcs_.find(key) != funcs_.end(); }
+
+  typename KernelTuples::func_type At(int key) { return funcs_.at(key); }
+
+  void Insert(int key, typename KernelTuples::func_type func) {
+    funcs_.emplace(key, func);
+  }
+
+ private:
+  std::unordered_map<int, typename KernelTuples::func_type> funcs_;
+  DISABLE_COPY_AND_ASSIGN(KernelFuncsCache);
+};
+
 const char* to_string(KernelType kt);
 const char* to_string(SeqPoolType kt);
 

paddle/fluid/operators/jit/kernel_base.h

@@ -20,6 +20,7 @@ namespace paddle {
 namespace operators {
 namespace jit {
 
+// TODO(TJ): reorder by alphabet
 typedef enum {
   kNone = 0,
   kVMul = 1,
@@ -44,6 +45,9 @@ typedef enum {
   kNCHW16CMulNC,
   kSeqPool,
   kMatMul,
+  kHSum,  // horizontal max
+  kHMax,  // horizontal sum
+  kSoftmax,
 } KernelType;
 
 typedef enum {
@@ -70,6 +74,10 @@ struct XYNTuples {
   typedef void (*func_type)(const T*, T*, int);
 };
 
+// x, return and int
+template <typename T>
+struct XRNTuples : public XYNTuples<T> {};
+
 typedef struct {
   void* gates;  // gates: x_ch, x_ih, x_fh, x_oh
   const void* ct_1;
@@ -159,6 +167,13 @@ struct LayerNormTuples {
                             const float, int);
 };
 
+template <typename T>
+struct SoftmaxTuples {
+  typedef T data_type;
+  typedef int attr_type;
+  typedef void (*func_type)(const T*, T*, int, int);
+};
+
 // nChw16c = nChw16c .* NC
 template <typename T>
 struct NCHW16CMulNCTuples {

paddle/fluid/operators/jit/more/mix/CMakeLists.txt

@@ -12,3 +12,4 @@ USE_JITKERNEL_MORE(kLSTMC1H1, mix)
 USE_JITKERNEL_MORE(kGRUH1, mix)
 USE_JITKERNEL_MORE(kGRUHtPart1, mix)
 USE_JITKERNEL_MORE(kGRUHtPart2, mix)
+USE_JITKERNEL_MORE(kSoftmax, mix)

paddle/fluid/operators/jit/more/mix/mix.cc

@@ -48,6 +48,65 @@ void VTanh(const T* x, T* y, int n) {
   compute_addbias(&b, y, y, n);
 }
 
+void Softmax(const T* x, T* y, int n, int bs) {
+  typename XRNTuples<T>::func_type compute_hmax{nullptr};
+  typename XRNTuples<T>::func_type compute_hsum{nullptr};
+  typename AXYNTuples<T>::func_type compute_vscal{nullptr};
+  typename AXYNTuples<T>::func_type compute_vaddbias{nullptr};
+  typename XYNTuples<T>::func_type compute_vexp{nullptr};
+
+  if (!KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().Has(n)) {
+    compute_hmax = Get<kHMax, XRNTuples<T>, platform::CPUPlace>(n);
+    KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().Insert(n, compute_hmax);
+  } else {
+    compute_hmax = KernelFuncsCache<kHMax, XRNTuples<T>>::Instance().At(n);
+  }
+
+  if (!KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().Has(n)) {
+    compute_hsum = Get<kHSum, XRNTuples<T>, platform::CPUPlace>(n);
+    KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().Insert(n, compute_hsum);
+  } else {
+    compute_hsum = KernelFuncsCache<kHSum, XRNTuples<T>>::Instance().At(n);
+  }
+
+  if (!KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().Has(n)) {
+    compute_vscal = Get<kVScal, AXYNTuples<T>, platform::CPUPlace>(n);
+    KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().Insert(n,
+                                                               compute_vscal);
+  } else {
+    compute_vscal = KernelFuncsCache<kVScal, AXYNTuples<T>>::Instance().At(n);
+  }
+
+  if (!KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().Has(n)) {
+    compute_vaddbias = Get<kVAddBias, AXYNTuples<T>, platform::CPUPlace>(n);
+    KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().Insert(
+        n, compute_vaddbias);
+  } else {
+    compute_vaddbias =
+        KernelFuncsCache<kVAddBias, AXYNTuples<T>>::Instance().At(n);
+  }
+
+  if (!KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().Has(n)) {
+    compute_vexp = Get<KernelType::kVExp, XYNTuples<T>, platform::CPUPlace>(n);
+    KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().Insert(n, compute_vexp);
+  } else {
+    compute_vexp = KernelFuncsCache<kVExp, XYNTuples<T>>::Instance().At(n);
+  }
+
+  for (int i = 0; i < bs; ++i) {
+    T scalar;
+    compute_hmax(x, &scalar, n);
+    scalar = static_cast<T>(0) - scalar;
+    compute_vaddbias(&scalar, x, y, n);  // x - max
+    compute_vexp(y, y, n);
+    compute_hsum(y, &scalar, n);
+    scalar = static_cast<T>(1) / scalar;
+    compute_vscal(&scalar, y, y, n);
+    x += n;
+    y += n;
+  }
+}
+
 void (*getActFunc(KernelType type, int d))(const T*, T*, int) {  // NOLINT
   if (type == kVSigmoid) {
     return Get<kVSigmoid, XYNTuples<T>, platform::CPUPlace>(d);
@@ -184,6 +243,8 @@ bool VSigmoidKernel::UseMe(const int& d) const { return true; }
 
 bool VTanhKernel::UseMe(const int& d) const { return true; }
 
+bool SoftmaxKernel::UseMe(const int& d) const { return true; }
+
 bool LSTMCtHtKernel::UseMe(const lstm_attr_t& attr) const { return true; }
 
 bool LSTMC1H1Kernel::UseMe(const lstm_attr_t& attr) const { return true; }
@@ -207,6 +268,7 @@ namespace mix = paddle::operators::jit::more::mix;
 
 REGISTER_MORE_KERNEL(kVSigmoid, VSigmoid);
 REGISTER_MORE_KERNEL(kVTanh, VTanh);
+REGISTER_MORE_KERNEL(kSoftmax, Softmax);
 REGISTER_MORE_KERNEL(kLSTMCtHt, LSTMCtHt);
 REGISTER_MORE_KERNEL(kLSTMC1H1, LSTMC1H1);
 REGISTER_MORE_KERNEL(kGRUH1, GRUH1);

paddle/fluid/operators/jit/more/mix/mix.h

@@ -26,6 +26,7 @@ using T = float;
 
 void VSigmoid(const T* x, T* y, int n);
 void VTanh(const T* x, T* y, int n);
+void Softmax(const T* x, T* y, int n, int bs);
 
 void LSTMCtHt(lstm_t* step, const lstm_attr_t* attr);
 void LSTMC1H1(lstm_t* step, const lstm_attr_t* attr);
@@ -45,6 +46,9 @@ void GRUHtPart2(gru_t* step, const gru_attr_t* attr);
 DECLARE_MORE_KERNEL(VSigmoid, XYNTuples);
 DECLARE_MORE_KERNEL(VTanh, XYNTuples);
 
+// XRN
+DECLARE_MORE_KERNEL(Softmax, SoftmaxTuples);
+
 DECLARE_MORE_KERNEL(LSTMCtHt, LSTMTuples);
 DECLARE_MORE_KERNEL(LSTMC1H1, LSTMTuples);
 

paddle/fluid/operators/jit/more/mkl/CMakeLists.txt

@@ -12,3 +12,4 @@ USE_JITKERNEL_MORE(kVSquare, mkl)
 USE_JITKERNEL_MORE(kVSigmoid, mkl)
 USE_JITKERNEL_MORE(kVTanh, mkl)
 USE_JITKERNEL_MORE(kSeqPool, mkl)
+USE_JITKERNEL_MORE(kSoftmax, mkl)

paddle/fluid/operators/jit/more/mkl/mkl.cc

@@ -116,6 +116,16 @@ void VAXPY<double>(double a, const double* x, double* y, int n) {
   platform::dynload::cblas_daxpy(n, a, x, 1, y, 1);
 }
 
+template <>
+void ASum<float>(const float* x, float* res, int n) {
+  res[0] = platform::dynload::cblas_sasum(n, x, 1);
+}
+
+template <>
+void ASum<double>(const double* x, double* res, int n) {
+  res[0] = platform::dynload::cblas_dasum(n, x, 1);
+}
+
 // TODO(TJ): tuning me carefully on AVX, AVX2 and AVX512
 template <>
 bool MatMulKernel<float>::UseMe(const int& d) const {
@@ -167,6 +177,12 @@ bool SeqPoolKernel<double>::UseMe(const seq_pool_attr_t& attr) const {
   return true;
 }
 
+template <>
+bool SoftmaxKernel<float>::UseMe(const int& d) const {
+  // tuned on avx2
+  return platform::MayIUse(platform::avx) && d < 60;
+}
+
 #define AWALYS_USE_ME_WITH_DOUBLE(func)                  \
   template <>                                            \
   bool func##Kernel<double>::UseMe(const int& d) const { \
@@ -181,6 +197,7 @@ AWALYS_USE_ME_WITH_DOUBLE(VExp);
 AWALYS_USE_ME_WITH_DOUBLE(VSigmoid);
 AWALYS_USE_ME_WITH_DOUBLE(VTanh);
 AWALYS_USE_ME_WITH_DOUBLE(VSquare);
+AWALYS_USE_ME_WITH_DOUBLE(Softmax);
 
 #undef AWALYS_USE_ME_WITH_DOUBLE
 }  // namespace mkl
@@ -204,5 +221,6 @@ REGISTER_MKL_KERNEL(kVSquare, VSquare);
 REGISTER_MKL_KERNEL(kVSigmoid, VSigmoid);
 REGISTER_MKL_KERNEL(kVTanh, VTanh);
 REGISTER_MKL_KERNEL(kSeqPool, SeqPool);
+REGISTER_MKL_KERNEL(kSoftmax, Softmax);
 
 #undef REGISTER_MKL_KERNEL

paddle/fluid/operators/jit/more/mkl/mkl.h

@@ -16,6 +16,7 @@
 
 #include <cmath>
 #include <type_traits>
+#include <vector>
 #include "paddle/fluid/operators/jit/kernel_base.h"
 
 namespace paddle {
@@ -90,6 +91,30 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) {
   }
 }
 
+template <typename T>
+void ASum(const T* x, T* res, int n);
+
+template <typename T>
+void Softmax(const T* x, T* y, int n, int bs) {
+  std::vector<T> entities(bs);
+  for (int i = 0; i < bs; ++i) {
+    entities[i] = x[i * n];
+    for (int c = 1; c < n; ++c) {
+      entities[i] = x[i * n + c] > entities[i] ? x[i * n + c] : entities[i];
+    }
+    for (int c = 0; c < n; ++c) {
+      y[i * n + c] = x[i * n + c] - entities[i];
+    }
+  }
+  VExp(y, y, n * bs);
+  for (int i = 0; i < bs; ++i) {
+    T sum;
+    ASum(&y[i * n], &sum, n);
+    sum = static_cast<T>(1) / sum;
+    VScal(&sum, &y[i * n], &y[i * n], n);
+  }
+}
+
 #define DECLARE_MKL_KERNEL(name, tuples)                             \
   template <typename T>                                              \
   class name##Kernel : public KernelMore<tuples<T>> {                \
@@ -117,6 +142,8 @@ DECLARE_MKL_KERNEL(VSquare, XYNTuples);
 
 DECLARE_MKL_KERNEL(SeqPool, SeqPoolTuples);
 
+DECLARE_MKL_KERNEL(Softmax, SoftmaxTuples);
+
 #undef DECLARE_MKL_KERNEL
 
 }  // namespace mkl

paddle/fluid/operators/jit/refer/CMakeLists.txt

@@ -29,3 +29,6 @@ USE_JITKERNEL_REFER(kNCHW16CMulNC)
 USE_JITKERNEL_REFER(kSeqPool)
 USE_JITKERNEL_REFER(kMatMul)
 USE_JITKERNEL_REFER(kVSquare)
+USE_JITKERNEL_REFER(kHSum)
+USE_JITKERNEL_REFER(kHMax)
+USE_JITKERNEL_REFER(kSoftmax)

paddle/fluid/operators/jit/refer/refer.cc

@@ -52,4 +52,9 @@ REGISTER_REFER_KERNEL(kSeqPool, SeqPool);
 
 REGISTER_REFER_KERNEL(kMatMul, MatMul);
 
+REGISTER_REFER_KERNEL(kHMax, HMax);
+REGISTER_REFER_KERNEL(kHSum, HSum);
+
+REGISTER_REFER_KERNEL(kSoftmax, Softmax);
+
 #undef REGISTER_REFER_KERNEL

paddle/fluid/operators/jit/refer/refer.h

@@ -378,6 +378,40 @@ void MatMul(const T* A, const T* B, T* C, int M, int N, int K) {
   }
 }
 
+template <typename T>
+void HMax(const T* x, T* res, int n) {
+  res[0] = x[0];
+  for (int i = 1; i < n; ++i) {
+    res[0] = res[0] < x[i] ? x[i] : res[0];
+  }
+}
+
+template <typename T>
+void HSum(const T* x, T* res, int n) {
+  res[0] = x[0];
+  for (int i = 1; i < n; ++i) {
+    res[0] += x[i];
+  }
+}
+
+// y = e^(x - max(x))
+// y = y / sum(y)
+template <typename T>
+void Softmax(const T* x, T* y, int n, int bs = 1) {
+  for (int i = 0; i < bs; ++i) {
+    T scalar;
+    HMax(x, &scalar, n);
+    scalar = static_cast<T>(0) - scalar;
+    VAddBias(&scalar, x, y, n);  // x - max
+    VExp(y, y, n);
+    HSum(y, &scalar, n);
+    scalar = static_cast<T>(1) / scalar;
+    VScal(&scalar, y, y, n);
+    x += n;
+    y += n;
+  }
+}
+
 #define DECLARE_REFER_KERNEL(name, tuples)             \
   template <typename T>                                \
   class name##Kernel : public ReferKernel<tuples<T>> { \
@@ -421,6 +455,11 @@ DECLARE_REFER_KERNEL(SeqPool, SeqPoolTuples);
 
 DECLARE_REFER_KERNEL(MatMul, MatMulTuples);
 
+DECLARE_REFER_KERNEL(HMax, XRNTuples);
+DECLARE_REFER_KERNEL(HSum, XRNTuples);
+
+DECLARE_REFER_KERNEL(Softmax, SoftmaxTuples);
+
 #undef DECLARE_REFER_KERNEL
 
 }  // namespace refer

paddle/fluid/operators/jit/test.cc

@@ -61,6 +61,7 @@ std::vector<int> TestSizes() {
 }
 
 namespace jit = paddle::operators::jit;
+using CPUPlace = paddle::platform::CPUPlace;
 
 template <typename KernelTuples, typename... Args>
 struct TestFuncWithRefer {
@@ -121,6 +122,40 @@ struct TestFuncWithRefer<jit::AXYNTuples<T>, T, std::vector<T>,
   }
 };
 
+template <typename T>
+struct TestFuncWithRefer<jit::SoftmaxTuples<T>, std::vector<T>, std::vector<T>,
+                         int, int> {
+  void operator()(const typename jit::SoftmaxTuples<T>::func_type tgt,
+                  const std::vector<T>& x, const std::vector<T>& yref, int n,
+                  int bs) {
+    EXPECT_TRUE(tgt != nullptr);
+    EXPECT_EQ(yref.size(), x.size());
+    EXPECT_EQ(x.size(), static_cast<size_t>(n * bs));
+    const T* x_data = x.data();
+    const T* yref_data = yref.data();
+    std::vector<T> ytgt(n * bs);
+    T* ytgt_data = ytgt.data();
+    // test normal
+    tgt(x_data, ytgt_data, n, bs);
+    ExpectEQ<T>(ytgt_data, yref_data, n * bs);
+    // test inplace x
+    std::copy(x.begin(), x.end(), ytgt.begin());
+    tgt(ytgt_data, ytgt_data, n, bs);
+    ExpectEQ<T>(ytgt_data, yref_data, n * bs);
+  }
+};
+
+template <typename T>
+struct TestFuncWithRefer<jit::XRNTuples<T>, std::vector<T>, T> {
+  void operator()(const typename jit::XRNTuples<T>::func_type tgt,
+                  const std::vector<T>& x, const T ref_res) {
+    EXPECT_TRUE(tgt != nullptr);
+    T tgt_res;
+    tgt(x.data(), &tgt_res, x.size());
+    ExpectEQ<T>(&tgt_res, &ref_res, 1);
+  }
+};
+
 template <typename T>
 struct TestFuncWithRefer<jit::XYNTuples<T>, std::vector<T>, std::vector<T>> {
   void operator()(const typename jit::XYNTuples<T>::func_type tgt,
@@ -172,7 +207,7 @@ struct TestFuncWithRefer<jit::LSTMTuples<T>, std::vector<T>, std::vector<T>,
     T* ht_data = ht.data();
     T* checked_data = checked.data();
 
-    paddle::operators::jit::lstm_t step;
+    jit::lstm_t step;
     step.gates = x_data;
     step.ct_1 = ct_1_data;
     step.ct = ct_data;
@@ -208,7 +243,7 @@ struct TestFuncWithRefer<jit::GRUTuples<T>, std::vector<T>, std::vector<T>,
     const T* ht_ref_data = ht_ref.data();
     T* x_data = x.data();
     T* ht_data = ht.data();
-    paddle::operators::jit::gru_t step;
+    jit::gru_t step;
     step.gates = x_data;
     step.ht_1 = ht_1_data;
     step.ht = ht_data;
@@ -255,8 +290,8 @@ struct TestFuncWithRefer<jit::MatMulTuples<T>, std::vector<T>, std::vector<T>,
   }
 };
 
-template <paddle::operators::jit::KernelType KT, typename KernelTuples,
-          typename PlaceType, typename... Args>
+template <jit::KernelType KT, typename KernelTuples, typename PlaceType,
+          typename... Args>
 void TestAllImpls(const typename KernelTuples::attr_type& attr, Args... args) {
   TestFuncWithRefer<KernelTuples, Args...> test;
   // test jitcode
@@ -286,9 +321,8 @@ void TestAllImpls(const typename KernelTuples::attr_type& attr, Args... args) {
   test(tgt, args...);
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void TestXYZNKernel() {
-  namespace jit = paddle::operators::jit;
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
   for (int d : TestSizes()) {
     auto ref = jit::GetRefer<KT, jit::XYZNTuples<T>>();
@@ -320,9 +354,8 @@ void TestXYZNKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void TestAXYNKernel() {
-  namespace jit = paddle::operators::jit;
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
   for (int d : TestSizes()) {
     auto ref = jit::GetRefer<KT, jit::AXYNTuples<T>>();
@@ -347,9 +380,26 @@ void TestAXYNKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
+void TestXRNKernel() {
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  auto last_acc = acc;
+  acc = 1e-4;
+  for (int d : TestSizes()) {
+    auto ref = jit::GetRefer<KT, jit::XRNTuples<T>>();
+    EXPECT_TRUE(ref != nullptr);
+    std::vector<T> x(d);
+    RandomVec<T>(d, x.data(), -2.f, 2.f);
+    T ref_res;
+    ref(x.data(), &ref_res, d);
+    TestAllImpls<KT, jit::XRNTuples<T>, PlaceType, std::vector<T>, T>(d, x,
+                                                                      ref_res);
+  }
+  acc = last_acc;
+}
+
+template <jit::KernelType KT, typename T, typename PlaceType>
 void TestXYNKernel() {
-  namespace jit = paddle::operators::jit;
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
   for (int d : TestSizes()) {
     auto ref = jit::GetRefer<KT, jit::XYNTuples<T>>();
@@ -373,9 +423,8 @@ void TestXYNKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void TestLSTMKernel() {
-  namespace jit = paddle::operators::jit;
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
   std::vector<std::string> all_acts = {"sigmoid", "tanh", "relu", "identity"};
   for (int d : TestSizes()) {
@@ -424,9 +473,8 @@ void TestLSTMKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void TestGRUKernel() {
-  namespace jit = paddle::operators::jit;
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
   std::vector<std::string> all_acts = {"sigmoid", "tanh", "relu", "identity"};
   for (int d : TestSizes()) {
@@ -459,7 +507,7 @@ void TestGRUKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void TestSeqPoolKernel() {
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
   std::vector<jit::SeqPoolType> pool_types = {
@@ -484,7 +532,7 @@ void TestSeqPoolKernel() {
   }
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
 void TestMatMulKernel() {
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
   auto last_acc = acc;
@@ -510,7 +558,32 @@ void TestMatMulKernel() {
   acc = last_acc;
 }
 
-template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+template <jit::KernelType KT, typename T, typename PlaceType>
+void TestSoftmaxKernel() {
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  for (int bs : {1, 2, 10}) {
+    for (int n : TestSizes()) {
+      auto ref = jit::GetRefer<KT, jit::SoftmaxTuples<T>>();
+      EXPECT_TRUE(ref != nullptr);
+      std::vector<T> x(bs * n), y(bs * n);
+      RandomVec<T>(bs * n, x.data(), -2.f, 2.f);
+      const T* x_data = x.data();
+      T* y_data = y.data();
+
+      std::vector<T> xinp(x.size());  // inplace test
+      std::copy(x.begin(), x.end(), xinp.begin());
+      ref(x_data, y_data, n, bs);
+      T* xinp_data = xinp.data();
+      ref(xinp_data, xinp_data, n, bs);
+      ExpectEQ<T>(xinp_data, y_data, n * bs);
+
+      TestAllImpls<KT, jit::SoftmaxTuples<T>, PlaceType, std::vector<T>,
+                   std::vector<T>>(n, x, y, n, bs);
+    }
+  }
+}
+
+template <jit::KernelType KT, typename T, typename PlaceType>
 void TestNCHW16CMulNCKernel() {
   VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
   const int n = 3, c = 16 * 4, h = 10, w = 10;
@@ -565,129 +638,123 @@ void TestNCHW16CMulNCKernel() {
 
 // XYZNTuple
 TEST(JITKernel, kVMul) {
-  namespace jit = paddle::operators::jit;
-  TestXYZNKernel<jit::kVMul, float, paddle::platform::CPUPlace>();
-  TestXYZNKernel<jit::kVMul, double, paddle::platform::CPUPlace>();
+  TestXYZNKernel<jit::kVMul, float, CPUPlace>();
+  TestXYZNKernel<jit::kVMul, double, CPUPlace>();
 }
 
 TEST(JITKernel, kVAdd) {
-  namespace jit = paddle::operators::jit;
-  TestXYZNKernel<jit::kVAdd, float, paddle::platform::CPUPlace>();
-  TestXYZNKernel<jit::kVAdd, double, paddle::platform::CPUPlace>();
+  TestXYZNKernel<jit::kVAdd, float, CPUPlace>();
+  TestXYZNKernel<jit::kVAdd, double, CPUPlace>();
 }
 
 TEST(JITKernel, kVAddRelu) {
-  namespace jit = paddle::operators::jit;
-  TestXYZNKernel<jit::kVAddRelu, float, paddle::platform::CPUPlace>();
-  TestXYZNKernel<jit::kVAddRelu, double, paddle::platform::CPUPlace>();
+  TestXYZNKernel<jit::kVAddRelu, float, CPUPlace>();
+  TestXYZNKernel<jit::kVAddRelu, double, CPUPlace>();
 }
 
 TEST(JITKernel, kVSub) {
-  namespace jit = paddle::operators::jit;
-  TestXYZNKernel<jit::kVSub, float, paddle::platform::CPUPlace>();
-  TestXYZNKernel<jit::kVSub, double, paddle::platform::CPUPlace>();
+  TestXYZNKernel<jit::kVSub, float, CPUPlace>();
+  TestXYZNKernel<jit::kVSub, double, CPUPlace>();
 }
 
 // AXYNTuples
 TEST(JITKernel, kVScal) {
-  namespace jit = paddle::operators::jit;
-  TestAXYNKernel<jit::kVScal, float, paddle::platform::CPUPlace>();
-  TestAXYNKernel<jit::kVScal, double, paddle::platform::CPUPlace>();
+  TestAXYNKernel<jit::kVScal, float, CPUPlace>();
+  TestAXYNKernel<jit::kVScal, double, CPUPlace>();
 }
 
 TEST(JITKernel, kVAddBias) {
-  namespace jit = paddle::operators::jit;
-  TestAXYNKernel<jit::kVAddBias, float, paddle::platform::CPUPlace>();
-  TestAXYNKernel<jit::kVAddBias, double, paddle::platform::CPUPlace>();
+  TestAXYNKernel<jit::kVAddBias, float, CPUPlace>();
+  TestAXYNKernel<jit::kVAddBias, double, CPUPlace>();
+}
+
+// XRNTuples
+TEST(JITKernel, kHMax) {
+  TestXRNKernel<jit::kHMax, float, CPUPlace>();
+  TestXRNKernel<jit::kHMax, double, CPUPlace>();
+}
+
+TEST(JITKernel, kHSum) {
+  TestXRNKernel<jit::kHSum, float, CPUPlace>();
+  TestXRNKernel<jit::kHSum, double, CPUPlace>();
 }
 
 // XYNTuples
 TEST(JITKernel, kVRelu) {
-  namespace jit = paddle::operators::jit;
-  TestXYNKernel<jit::kVRelu, float, paddle::platform::CPUPlace>();
-  TestXYNKernel<jit::kVRelu, double, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVRelu, float, CPUPlace>();
+  TestXYNKernel<jit::kVRelu, double, CPUPlace>();
 }
 
 TEST(JITKernel, kVIdentity) {
-  namespace jit = paddle::operators::jit;
-  TestXYNKernel<jit::kVIdentity, float, paddle::platform::CPUPlace>();
-  TestXYNKernel<jit::kVIdentity, double, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVIdentity, float, CPUPlace>();
+  TestXYNKernel<jit::kVIdentity, double, CPUPlace>();
 }
 
 TEST(JITKernel, kVSquare) {
-  namespace jit = paddle::operators::jit;
-  TestXYNKernel<jit::kVSquare, float, paddle::platform::CPUPlace>();
-  TestXYNKernel<jit::kVSquare, double, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVSquare, float, CPUPlace>();
+  TestXYNKernel<jit::kVSquare, double, CPUPlace>();
 }
 
 TEST(JITKernel, kVExp) {
-  namespace jit = paddle::operators::jit;
-  TestXYNKernel<jit::kVExp, float, paddle::platform::CPUPlace>();
-  TestXYNKernel<jit::kVExp, double, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVExp, float, CPUPlace>();
+  TestXYNKernel<jit::kVExp, double, CPUPlace>();
 }
 
 TEST(JITKernel, kVSigmoid) {
-  namespace jit = paddle::operators::jit;
-  TestXYNKernel<jit::kVSigmoid, float, paddle::platform::CPUPlace>();
-  TestXYNKernel<jit::kVSigmoid, double, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVSigmoid, float, CPUPlace>();
+  TestXYNKernel<jit::kVSigmoid, double, CPUPlace>();
 }
 
 TEST(JITKernel, kVTanh) {
-  namespace jit = paddle::operators::jit;
-  TestXYNKernel<jit::kVTanh, float, paddle::platform::CPUPlace>();
-  TestXYNKernel<jit::kVTanh, double, paddle::platform::CPUPlace>();
+  TestXYNKernel<jit::kVTanh, float, CPUPlace>();
+  TestXYNKernel<jit::kVTanh, double, CPUPlace>();
 }
 
 // LSTM
 TEST(JITKernel, kLSTMCtHt) {
-  namespace jit = paddle::operators::jit;
-  TestLSTMKernel<jit::kLSTMCtHt, float, paddle::platform::CPUPlace>();
-  TestLSTMKernel<jit::kLSTMCtHt, double, paddle::platform::CPUPlace>();
+  TestLSTMKernel<jit::kLSTMCtHt, float, CPUPlace>();
+  TestLSTMKernel<jit::kLSTMCtHt, double, CPUPlace>();
 }
 
 TEST(JITKernel, kLSTMC1H1) {
-  namespace jit = paddle::operators::jit;
-  TestLSTMKernel<jit::kLSTMC1H1, float, paddle::platform::CPUPlace>();
-  TestLSTMKernel<jit::kLSTMC1H1, double, paddle::platform::CPUPlace>();
+  TestLSTMKernel<jit::kLSTMC1H1, float, CPUPlace>();
+  TestLSTMKernel<jit::kLSTMC1H1, double, CPUPlace>();
 }
 
 // GRU
 TEST(JITKernel, kGRUH1) {
-  namespace jit = paddle::operators::jit;
-  TestGRUKernel<jit::kGRUH1, float, paddle::platform::CPUPlace>();
-  TestGRUKernel<jit::kGRUH1, double, paddle::platform::CPUPlace>();
+  TestGRUKernel<jit::kGRUH1, float, CPUPlace>();
+  TestGRUKernel<jit::kGRUH1, double, CPUPlace>();
 }
 
 TEST(JITKernel, kGRUHtPart1) {
-  namespace jit = paddle::operators::jit;
-  TestGRUKernel<jit::kGRUHtPart1, float, paddle::platform::CPUPlace>();
-  TestGRUKernel<jit::kGRUHtPart1, double, paddle::platform::CPUPlace>();
+  TestGRUKernel<jit::kGRUHtPart1, float, CPUPlace>();
+  TestGRUKernel<jit::kGRUHtPart1, double, CPUPlace>();
 }
 
 TEST(JITKernel, kGRUHtPart2) {
-  namespace jit = paddle::operators::jit;
-  TestGRUKernel<jit::kGRUHtPart2, float, paddle::platform::CPUPlace>();
-  TestGRUKernel<jit::kGRUHtPart2, double, paddle::platform::CPUPlace>();
+  TestGRUKernel<jit::kGRUHtPart2, float, CPUPlace>();
+  TestGRUKernel<jit::kGRUHtPart2, double, CPUPlace>();
 }
 
 TEST(JITKernel, kSeqPool) {
-  namespace jit = paddle::operators::jit;
-  TestSeqPoolKernel<jit::kSeqPool, float, paddle::platform::CPUPlace>();
-  TestSeqPoolKernel<jit::kSeqPool, double, paddle::platform::CPUPlace>();
+  TestSeqPoolKernel<jit::kSeqPool, float, CPUPlace>();
+  TestSeqPoolKernel<jit::kSeqPool, double, CPUPlace>();
 }
 
 TEST(JITKernel, kMatMul) {
-  namespace jit = paddle::operators::jit;
-  TestMatMulKernel<jit::kMatMul, float, paddle::platform::CPUPlace>();
-  TestMatMulKernel<jit::kMatMul, double, paddle::platform::CPUPlace>();
+  TestMatMulKernel<jit::kMatMul, float, CPUPlace>();
+  TestMatMulKernel<jit::kMatMul, double, CPUPlace>();
+}
+
+TEST(JITKernel, kSoftmax) {
+  TestSoftmaxKernel<jit::kSoftmax, float, CPUPlace>();
+  TestSoftmaxKernel<jit::kSoftmax, double, CPUPlace>();
 }
 
 TEST(JITKernel, kNCHW16CMulNC) {
-  namespace jit = paddle::operators::jit;
-  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, float,
-                         paddle::platform::CPUPlace>();
-  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, double,
-                         paddle::platform::CPUPlace>();
+  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, float, CPUPlace>();
+  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, double, CPUPlace>();
 }
 
 // TODO(yihua/TJ): add crf decoding and layer norm unit tests

paddle/fluid/operators/lookup_table_op.cu

@@ -17,6 +17,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/lookup_table_op.h"
 #include "paddle/fluid/platform/assert.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
+#include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace operators {
@@ -193,8 +194,11 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(lookup_table, ops::LookupTableCUDAKernel<float>,
-                        ops::LookupTableCUDAKernel<double>);
+                        ops::LookupTableCUDAKernel<double>,
+                        ops::LookupTableCUDAKernel<plat::float16>);
 REGISTER_OP_CUDA_KERNEL(lookup_table_grad,
                         ops::LookupTableGradCUDAKernel<float>,
-                        ops::LookupTableGradCUDAKernel<double>);
+                        ops::LookupTableGradCUDAKernel<double>,
+                        ops::LookupTableGradCUDAKernel<plat::float16>);

paddle/fluid/operators/math/CMakeLists.txt

@@ -53,7 +53,7 @@ math_library(sequence2batch)
 math_library(sequence_padding)
 math_library(sequence_pooling DEPS math_function jit_kernel_helper)
 math_library(sequence_scale)
-math_library(softmax DEPS math_function)
+math_library(softmax DEPS math_function jit_kernel_helper)
 math_library(beam_search DEPS math_function)
 
 math_library(matrix_bit_code)

paddle/fluid/operators/math/beam_search.cc

@@ -29,8 +29,9 @@ class BeamSearchFunctor<platform::CPUDeviceContext, T> {
                   const framework::LoDTensor *ids,
                   const framework::LoDTensor *scores,
                   framework::LoDTensor *selected_ids,
-                  framework::LoDTensor *selected_scores, size_t level,
-                  size_t beam_size, int end_id, bool is_accumulated) {
+                  framework::LoDTensor *selected_scores,
+                  framework::Tensor *parent_idx, size_t level, size_t beam_size,
+                  int end_id, bool is_accumulated) {
     auto abs_lod = framework::ToAbsOffset(scores->lod());
     auto &high_level = abs_lod[level];
 
@@ -57,11 +58,13 @@ class BeamSearchFunctor<platform::CPUDeviceContext, T> {
         std::vector<int64_t>({static_cast<int>(num_instances), 1}));
     selected_ids->Resize(dims);
     selected_scores->Resize(dims);
+    parent_idx->Resize({static_cast<int64_t>(num_instances)});
 
     auto *selected_ids_data =
         selected_ids->mutable_data<int64_t>(platform::CPUPlace());
     auto *selected_scores_data =
         selected_scores->mutable_data<float>(platform::CPUPlace());
+    auto *parent_idx_data = parent_idx->mutable_data<int>(platform::CPUPlace());
 
     // fill in data
     std::vector<size_t> low_level;
@@ -69,6 +72,7 @@ class BeamSearchFunctor<platform::CPUDeviceContext, T> {
     for (auto &items : selected_items) {
       low_level.push_back(low_offset);
       for (auto &item : items) {
+        parent_idx_data[low_offset] = static_cast<int>(low_level.size() - 1);
         selected_ids_data[low_offset] = item.id;
         selected_scores_data[low_offset] = item.score;
         low_offset++;

paddle/fluid/operators/math/beam_search.cu

@@ -157,10 +157,10 @@ __device__ __forceinline__ bool PruneEndBeams(Triple* top_beam_local,
 }
 
 __device__ __forceinline__ void WriteBack(
-    int64_t* selected_ids, float* selected_scores, size_t* selected_offsets,
-    Triple* top_beam_local, const int seq_offset_start,
-    const int seq_offset_end, const int selected_seq_start,
-    const int selected_seq_length) {
+    int64_t* selected_ids, float* selected_scores, int* parent_idx,
+    size_t* selected_offsets, Triple* top_beam_local,
+    const int seq_offset_start, const int seq_offset_end,
+    const int selected_seq_start, const int selected_seq_length) {
   const int tid = threadIdx.x;  // use 1 thread only for each sequence
   int global_index = selected_seq_start;
   for (int global_offset = seq_offset_start; global_offset < seq_offset_end;
@@ -171,6 +171,7 @@ __device__ __forceinline__ void WriteBack(
         selected_ids[global_index] =
             static_cast<int64_t>(top_beam_local[local_index].id);
         selected_scores[global_index] = top_beam_local[local_index].score;
+        parent_idx[global_index] = static_cast<int>(global_offset);
         global_index++;
       }
     }
@@ -180,11 +181,11 @@ __device__ __forceinline__ void WriteBack(
 
 template <int MaxLength, int MaxThreadsPerSeq, int MaxSeqs>
 __device__ void BeamSearchDetails(
-    int64_t* selected_ids, float* selected_scores, size_t* selected_offsets,
-    const int64_t* pre_ids, const float* pre_scores, const int64_t* ids,
-    const float* scores, const int seq_offset_start, const int seq_offset_end,
-    const int seq_width, int beam_size, int end_id, bool is_accumulated,
-    int num_used_threads) {
+    int64_t* selected_ids, float* selected_scores, int* parent_idx,
+    size_t* selected_offsets, const int64_t* pre_ids, const float* pre_scores,
+    const int64_t* ids, const float* scores, const int seq_offset_start,
+    const int seq_offset_end, const int seq_width, int beam_size, int end_id,
+    bool is_accumulated, int num_used_threads) {
   __shared__ Triple top_beam[MaxLength];
 
   int num_items = 0;
@@ -228,15 +229,15 @@ __device__ void BeamSearchDetails(
       selected_offsets[0] = 0;
     }
 
-    WriteBack(selected_ids, selected_scores, selected_offsets, top_beam_local,
-              seq_offset_start, seq_offset_end, selected_seq_start,
-              selected_seq_length);
+    WriteBack(selected_ids, selected_scores, parent_idx, selected_offsets,
+              top_beam_local, seq_offset_start, seq_offset_end,
+              selected_seq_start, selected_seq_length);
   }
 }
 
 template <int MaxLength, int MaxThreadsPerSeq, int MaxSeqs>
 __global__ void BeamSearchKernel(int64_t* selected_ids, float* selected_scores,
-                                 size_t* selected_offsets,
+                                 int* parent_idx, size_t* selected_offsets,
                                  const int64_t* pre_ids,
                                  const float* pre_scores, const int64_t* ids,
                                  const float* scores, const size_t* seq_offsets,
@@ -250,24 +251,25 @@ __global__ void BeamSearchKernel(int64_t* selected_ids, float* selected_scores,
   int seq_offset_end = static_cast<int>(seq_offsets[seq_id + 1]);
 
   BeamSearchDetails<MaxLength, MaxThreadsPerSeq, MaxSeqs>(
-      selected_ids, selected_scores, selected_offsets, pre_ids, pre_scores, ids,
-      scores, seq_offset_start, seq_offset_end, seq_width, beam_size, end_id,
-      is_accumulated, num_used_threads);
+      selected_ids, selected_scores, parent_idx, selected_offsets, pre_ids,
+      pre_scores, ids, scores, seq_offset_start, seq_offset_end, seq_width,
+      beam_size, end_id, is_accumulated, num_used_threads);
 }
 
 template <int MaxLength, int MaxThreadsPerSeq>
 __global__ void BeamSearchKernelSingle(
-    int64_t* selected_ids, float* selected_scores, size_t* selected_offsets,
-    const int64_t* pre_ids, const float* pre_scores, const int64_t* ids,
-    const float* scores, const int seq_length, const int seq_width,
-    int beam_size, int end_id, bool is_accumulated, int num_used_threads) {
+    int64_t* selected_ids, float* selected_scores, int* parent_idx,
+    size_t* selected_offsets, const int64_t* pre_ids, const float* pre_scores,
+    const int64_t* ids, const float* scores, const int seq_length,
+    const int seq_width, int beam_size, int end_id, bool is_accumulated,
+    int num_used_threads) {
   const int seq_offset_start = 0;
   const int seq_offset_end = seq_length;
 
   BeamSearchDetails<MaxLength, MaxThreadsPerSeq, 1>(
-      selected_ids, selected_scores, selected_offsets, pre_ids, pre_scores, ids,
-      scores, seq_offset_start, seq_offset_end, seq_width, beam_size, end_id,
-      is_accumulated, num_used_threads);
+      selected_ids, selected_scores, parent_idx, selected_offsets, pre_ids,
+      pre_scores, ids, scores, seq_offset_start, seq_offset_end, seq_width,
+      beam_size, end_id, is_accumulated, num_used_threads);
 }
 
 static inline int GetNumUsedThreads(const int max_threads_per_seq,
@@ -300,8 +302,9 @@ class BeamSearchFunctor<platform::CUDADeviceContext, T> {
                   const framework::LoDTensor* ids,
                   const framework::LoDTensor* scores,
                   framework::LoDTensor* selected_ids,
-                  framework::LoDTensor* selected_scores, size_t level,
-                  size_t beam_size, int end_id, bool is_accumulated) {
+                  framework::LoDTensor* selected_scores,
+                  framework::Tensor* parent_idx, size_t level, size_t beam_size,
+                  int end_id, bool is_accumulated) {
     auto abs_lod = framework::ToAbsOffset(scores->lod());
 
     const int64_t* pre_ids_data = pre_ids->data<int64_t>();
@@ -322,6 +325,8 @@ class BeamSearchFunctor<platform::CUDADeviceContext, T> {
         selected_ids->mutable_data<int64_t>(selected_dims, context.GetPlace());
     float* selected_scores_data =
         selected_scores->mutable_data<float>(selected_dims, context.GetPlace());
+    int* parent_idx_data = parent_idx->mutable_data<int>(
+        {static_cast<int64_t>(num_seqs * beam_size)}, context.GetPlace());
 
     framework::LoD selected_lod(2);
     selected_lod[0].assign(abs_lod[level].begin(), abs_lod[level].end());
@@ -339,9 +344,9 @@ class BeamSearchFunctor<platform::CUDADeviceContext, T> {
         CUDA_LAUNCH_KERNEL_HELPER(
             BeamSearchKernelSingle<kPowerOfTwoDim, kMaxThreadsPerSeq><<<
                 1, kMaxThreadsPerSeq, 0, context.stream()>>>(
-                selected_ids_data, selected_scores_data, selected_offsets,
-                pre_ids_data, pre_scores_data, ids_data, scores_data,
-                seq_length, static_cast<int>(seq_width),
+                selected_ids_data, selected_scores_data, parent_idx_data,
+                selected_offsets, pre_ids_data, pre_scores_data, ids_data,
+                scores_data, seq_length, static_cast<int>(seq_width),
                 static_cast<int>(beam_size), static_cast<int>(end_id),
                 is_accumulated, num_used_threads));
       }
@@ -357,9 +362,9 @@ class BeamSearchFunctor<platform::CUDADeviceContext, T> {
         CUDA_LAUNCH_KERNEL_HELPER(
             BeamSearchKernel<kPowerOfTwoDim, kMaxThreadsPerSeq, kMaxSeqs><<<
                 1, num_seqs * kMaxThreadsPerSeq, 0, context.stream()>>>(
-                selected_ids_data, selected_scores_data, selected_offsets,
-                pre_ids_data, pre_scores_data, ids_data, scores_data,
-                seq_offsets, static_cast<int>(num_seqs),
+                selected_ids_data, selected_scores_data, parent_idx_data,
+                selected_offsets, pre_ids_data, pre_scores_data, ids_data,
+                scores_data, seq_offsets, static_cast<int>(num_seqs),
                 static_cast<int>(seq_width), static_cast<int>(beam_size),
                 end_id, is_accumulated, num_used_threads));
       }
@@ -379,6 +384,7 @@ class BeamSearchFunctor<platform::CUDADeviceContext, T> {
           {static_cast<int64_t>(selected_lod[1].back()), 1});
       selected_ids->Resize(final_selected_dims);
       selected_scores->Resize(final_selected_dims);
+      parent_idx->Resize({static_cast<int64_t>(selected_lod[1].back())});
     }
   }
 };

paddle/fluid/operators/math/beam_search.h

@@ -104,14 +104,12 @@ class BeamSearchFunctor {
    * Return false if all the input tensor is empty, in machine translation task
    * that means no candidates is provided, and the task will stop running.
    */
-  void operator()(const DeviceContext& context,
-                  const framework::LoDTensor* pre_ids,
-                  const framework::LoDTensor* pre_scores,
-                  const framework::LoDTensor* ids,
-                  const framework::LoDTensor* scores,
-                  framework::LoDTensor* selected_ids,
-                  framework::LoDTensor* selected_scores, size_t level,
-                  size_t beam_size, int end_id, bool is_accumulated);
+  void operator()(
+      const DeviceContext& context, const framework::LoDTensor* pre_ids,
+      const framework::LoDTensor* pre_scores, const framework::LoDTensor* ids,
+      const framework::LoDTensor* scores, framework::LoDTensor* selected_ids,
+      framework::LoDTensor* selected_scores, framework::Tensor* parent_idx,
+      size_t level, size_t beam_size, int end_id, bool is_accumulated);
 };
 
 }  // namespace math

paddle/fluid/operators/math/beam_search_test.cc

@@ -93,13 +93,14 @@ void TestBeamSearch() {
 
   paddle::framework::LoDTensor selected_ids;
   paddle::framework::LoDTensor selected_scores;
+  paddle::framework::LoDTensor parent_idx;
 
   size_t level = 0;
   size_t beam_size = 2;
   int end_id = 0;
   paddle::operators::math::BeamSearchFunctor<DeviceContext, float> beamsearch;
   beamsearch(*context, &pre_ids, &pre_scores, &ids, &scores, &selected_ids,
-             &selected_scores, level, beam_size, end_id, true);
+             &selected_scores, &parent_idx, level, beam_size, end_id, true);
 
   ASSERT_EQ(selected_ids.lod(), selected_scores.lod());
 

paddle/fluid/operators/math/softmax_impl.h

@@ -16,8 +16,8 @@ limitations under the License. */
 #include <vector>
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/tensor.h"
+#include "paddle/fluid/operators/jit/kernels.h"
 
-#include "paddle/fluid/operators/math/blas.h"
 namespace paddle {
 namespace operators {
 namespace math {
@@ -81,28 +81,10 @@ class SoftmaxFunctor<DeviceContext, float, true, enable_if_CPU<DeviceContext>> {
     const int kBatchDim = 0;
     const int kClassDim = 1;
     // 2D data. Batch x C
-    const int batch_size = in_dims[kBatchDim];
-    const int num_classes = in_dims[kClassDim];
-    std::vector<float> entities(batch_size);
-    auto blas = math::GetBlas<DeviceContext, float>(context);
-    for (int n = 0; n < batch_size; ++n) {
-      entities[n] = in_data[n * num_classes];
-      for (int c = 1; c < num_classes; ++c) {
-        entities[n] = in_data[n * num_classes + c] > entities[n]
-                          ? in_data[n * num_classes + c]
-                          : entities[n];
-      }
-      for (int c = 0; c < num_classes; ++c) {
-        out_data[n * num_classes + c] =
-            in_data[n * num_classes + c] - entities[n];
-      }
-    }
-
-    blas.VEXP(num_classes * batch_size, out_data, out_data);
-    for (int n = 0; n < batch_size; ++n) {
-      auto sum = blas.ASUM(num_classes, &out_data[n * num_classes], 1);
-      blas.SCAL(num_classes, 1.0f / sum, &out_data[n * num_classes]);
-    }
+    auto compute_softmax =
+        jit::Get<jit::kSoftmax, jit::SoftmaxTuples<float>, platform::CPUPlace>(
+            in_dims[kClassDim]);
+    compute_softmax(in_data, out_data, in_dims[kClassDim], in_dims[kBatchDim]);
   }
 };
 

paddle/fluid/operators/lrn_mkldnn_op.cc → paddle/fluid/operators/mkldnn/lrn_mkldnn_op.cc

@@ -67,7 +67,13 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     mid->mutable_data<T>(ctx.GetPlace());
 
     const int n = ctx.Attr<int>("n");
-    const float alpha = ctx.Attr<float>("alpha");
+    // MKL-DNN implements LRN in a caffe way:
+    // http://caffe.berkeleyvision.org/tutorial/layers/lrn.html
+    // Where sum of squares is divided by size of normalization window
+    // this is not the case for PaddlePaddle LRN.
+    // Hence we need to compensate for this diffrence by
+    // multipliing alpha by size of window(n)
+    const float alpha = ctx.Attr<float>("alpha") * static_cast<float>(n);
     const float beta = ctx.Attr<float>("beta");
     const float k = ctx.Attr<float>("k");
     const bool is_test = ctx.Attr<bool>("is_test");
@@ -78,10 +84,7 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
     auto dims = paddle::framework::vectorize2int(x->dims());
 
     auto src_md = paddle::platform::MKLDNNMemDesc(
-        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
-
-    auto dst_md = paddle::platform::MKLDNNMemDesc(
-        dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw);
+        dims, mkldnn::memory::data_type::f32, x->format());
 
     auto forward_desc = mkldnn::lrn_forward::desc{mkldnn::prop_kind::forward,
                                                   mkldnn::lrn_across_channels,
@@ -92,8 +95,6 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
                                                   k};
 
     auto src_memory_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine};
-    auto dst_memory = mkldnn::memory{{dst_md, mkldnn_engine},
-                                     static_cast<void*>(output_data)};
 
     if (!is_test) {
       const std::string key = ctx.op().Output("Out");
@@ -110,11 +111,16 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
       src_memory->set_data_handle(
           static_cast<void*>(const_cast<T*>(input_data)));
 
+      auto dst_memory = mkldnn::memory(forward_pd->dst_primitive_desc(),
+                                       static_cast<void*>(output_data));
       auto workspace_memory = insert_to_context<mkldnn::memory>(
           key_workspace_memory, dev_ctx,
           forward_pd->workspace_primitive_desc());
 
       run_primitive(*forward_pd, *src_memory, *workspace_memory, dst_memory);
+
+      out->set_layout(framework::DataLayout::kMKLDNN);
+      out->set_format(platform::GetMKLDNNFormat(dst_memory));
     } else {
       auto forward_pd =
           mkldnn::lrn_forward::primitive_desc{forward_desc, mkldnn_engine};
@@ -122,8 +128,13 @@ class LRNMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
           src_memory_pd, static_cast<void*>(const_cast<T*>(input_data))};
       auto workspace_memory =
           mkldnn::memory{forward_pd.workspace_primitive_desc()};
+      auto dst_memory = mkldnn::memory(forward_pd.dst_primitive_desc(),
+                                       static_cast<void*>(output_data));
 
       run_primitive(forward_pd, src_memory, workspace_memory, dst_memory);
+
+      out->set_layout(framework::DataLayout::kMKLDNN);
+      out->set_format(platform::GetMKLDNNFormat(dst_memory));
     }
   }
 };
@@ -151,7 +162,7 @@ class LRNMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
     const std::string key_workspace_memory = key + "@lrn_workspace_memory";
 
     const int n = ctx.Attr<int>("n");
-    const float alpha = ctx.Attr<float>("alpha");
+    const float alpha = ctx.Attr<float>("alpha") * static_cast<float>(n);
     const float beta = ctx.Attr<float>("beta");
     const float k = ctx.Attr<float>("k");
 

paddle/fluid/operators/ngraph/ngraph_bridge.cc

@@ -38,6 +38,8 @@ std::map<std::string,
         {"mean_grad", NG_OPS::BuildMeanGradNode},
         {"mul", NG_OPS::BuildMulNode},
         {"mul_grad", NG_OPS::BuildMulGradNode},
+        {"pool2d", NG_OPS::BuildPool2dNode},
+        {"pool2d_grad", NG_OPS::BuildPool2dGradNode},
         {"softmax", NG_OPS::BuildSoftmaxNode},
         {"softmax_grad", NG_OPS::BuildSoftmaxGradNode},
         {"scale", NG_OPS::BuildScaleNode},

paddle/fluid/operators/ngraph/ngraph_ops.h

@@ -26,6 +26,7 @@ limitations under the License. */
 #include "ops/fill_constant_op.h"
 #include "ops/mean_op.h"
 #include "ops/mul_op.h"
+#include "ops/pool2d_op.h"
 #include "ops/scale_op.h"
 #include "ops/softmax_op.h"
 #include "ops/top_k_op.h"

paddle/fluid/operators/ngraph/ops/pool2d_op.h

+/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <string>
+#include <vector>
+
+#include "ngraph/ngraph.hpp"
+#include "paddle/fluid/platform/ngraph_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace ngraphs {
+
+void BuildPool2dNode(
+    const std::shared_ptr<paddle::framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto op_attrs = paddle::framework::AttrReader(op->Attrs());
+  auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map);
+  auto x_shape = x->get_shape();
+
+  std::string pooling_type = op_attrs.Get<std::string>("pooling_type");
+  std::vector<int> ksize = op_attrs.Get<std::vector<int>>("ksize");
+  std::vector<int> strides = op_attrs.Get<std::vector<int>>("strides");
+  std::vector<int> paddings = op_attrs.Get<std::vector<int>>("paddings");
+
+  PADDLE_ENFORCE_EQ(x_shape.size() - 2, ksize.size(),
+                    "Handling 2d pooling only");
+
+  if (op_attrs.Get<bool>("global_pooling")) {
+    for (size_t i = 0; i < ksize.size(); ++i) {
+      paddings[i] = 0;
+      ksize[i] = static_cast<int>(x_shape.at(i + 2));
+    }
+  }
+
+  ngraph::Shape ng_padding_below{static_cast<size_t>(paddings.at(0)),
+                                 static_cast<size_t>(paddings.at(1))};
+  ngraph::Shape ng_padding_above{static_cast<size_t>(paddings.at(0)),
+                                 static_cast<size_t>(paddings.at(1))};
+  ngraph::Shape ng_ksize_shape{static_cast<size_t>(ksize.at(0)),
+                               static_cast<size_t>(ksize.at(1))};
+  ngraph::Strides ng_strides{static_cast<size_t>(strides.at(0)),
+                             static_cast<size_t>(strides.at(1))};
+
+  auto ComputeCeiledOutput = [](size_t in, size_t k, size_t p, size_t s) {
+    return (in - k + 2 * p) / s + 1;
+  };
+
+  if (op_attrs.Get<bool>("ceil_mode")) {
+    auto dummy_out = paddle::platform::GetOutputNode(op, "Out", ngb_node_map);
+    auto dummpy_shape = dummy_out->get_shape();
+    for (size_t i = 0; i < ng_padding_above.size(); ++i) {
+      auto desired_size = ComputeCeiledOutput(x_shape[i + 2], ksize[i],
+                                              paddings[i], strides[i]);
+      if (desired_size != dummpy_shape[i + 2]) {
+        ng_padding_above[i] += strides[i];
+      }
+    }
+  }
+
+  bool padding_exclusive = op_attrs.Get<bool>("exclusive");
+  if (pooling_type == "max") {
+    auto pool2d = std::make_shared<ngraph::op::MaxPool>(
+        x, ng_ksize_shape, ng_strides, ng_padding_below, ng_padding_above);
+    paddle::platform::SetOutputNode(op, "Out", pool2d, ngb_node_map);
+  } else if (pooling_type == "avg") {
+    std::shared_ptr<ngraph::Node> pool2d;
+    if (op_attrs.Get<bool>("adaptive")) {
+      auto ComputeAdaptive = [](size_t in, size_t k) {
+        return std::floor(in / k);
+      };
+      ng_strides[0] = x_shape.size() == 4
+                          ? ComputeAdaptive(x_shape[3], ksize[0])
+                          : ng_strides[0];
+      ng_strides[1] = x_shape.size() == 4
+                          ? ComputeAdaptive(x_shape[3], ksize[0])
+                          : ng_strides[1];
+      pool2d =
+          std::make_shared<ngraph::op::AvgPool>(x, ng_ksize_shape, ng_strides);
+    } else {
+      pool2d = std::make_shared<ngraph::op::AvgPool>(
+          x, ng_ksize_shape, ng_strides, ng_padding_below, ng_padding_above,
+          !padding_exclusive);
+    }
+    paddle::platform::SetOutputNode(op, "Out", pool2d, ngb_node_map);
+  } else {
+    PADDLE_THROW("Support max and avg pooling only");
+  }
+}
+
+void BuildPool2dGradNode(
+    const std::shared_ptr<paddle::framework::OperatorBase>& op,
+    std::shared_ptr<
+        std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
+        ngb_node_map) {
+  auto op_attrs = paddle::framework::AttrReader(op->Attrs());
+  auto out = paddle::platform::GetInputNode(op, "Out", ngb_node_map);
+  auto dout = paddle::platform::GetInputNode(op, "Out@GRAD", ngb_node_map);
+  auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map);
+  auto x_shape = x->get_shape();
+
+  std::string pooling_type = op_attrs.Get<std::string>("pooling_type");
+  std::vector<int> ksize = op_attrs.Get<std::vector<int>>("ksize");
+  std::vector<int> strides = op_attrs.Get<std::vector<int>>("strides");
+  std::vector<int> paddings = op_attrs.Get<std::vector<int>>("paddings");
+
+  PADDLE_ENFORCE_EQ(x_shape.size() - 2, ksize.size(),
+                    "Handling 2d pooling only");
+
+  if (op_attrs.Get<bool>("global_pooling")) {
+    for (size_t i = 0; i < ksize.size(); ++i) {
+      paddings[i] = 0;
+      ksize[i] = static_cast<int>(x_shape.at(i + 2));
+    }
+  }
+
+  ngraph::Shape ng_padding_below{static_cast<size_t>(paddings.at(0)),
+                                 static_cast<size_t>(paddings.at(1))};
+  ngraph::Shape ng_padding_above{static_cast<size_t>(paddings.at(0)),
+                                 static_cast<size_t>(paddings.at(1))};
+  ngraph::Shape ng_ksize_shape{static_cast<size_t>(ksize.at(0)),
+                               static_cast<size_t>(ksize.at(1))};
+  ngraph::Strides ng_strides{static_cast<size_t>(strides.at(0)),
+                             static_cast<size_t>(strides.at(1))};
+
+  bool padding_exclusive = op_attrs.Get<bool>("exclusive");
+  if (pooling_type == "max") {
+    auto pool2d_grad = std::make_shared<ngraph::op::MaxPoolBackprop>(
+        x, dout, out, ng_ksize_shape, ng_strides, ng_padding_below,
+        ng_padding_above);
+    paddle::platform::SetOutputNode(op, "X@GRAD", pool2d_grad, ngb_node_map);
+  } else if (pooling_type == "avg") {
+    std::shared_ptr<ngraph::Node> pool2d_grad;
+    if (op_attrs.Get<bool>("adaptive")) {
+      auto ComputeAdaptive = [](size_t in, size_t k) {
+        return std::floor(in / k);
+      };
+      ng_strides[0] = x_shape.size() == 4
+                          ? ComputeAdaptive(x_shape[3], ksize[0])
+                          : ng_strides[0];
+      ng_strides[1] = x_shape.size() == 4
+                          ? ComputeAdaptive(x_shape[3], ksize[0])
+                          : ng_strides[1];
+      pool2d_grad = std::make_shared<ngraph::op::AvgPoolBackprop>(
+          x->get_shape(), dout, ng_ksize_shape, ng_strides, ng_padding_below,
+          ng_padding_above, !padding_exclusive);
+    } else {
+      pool2d_grad = std::make_shared<ngraph::op::AvgPoolBackprop>(
+          x->get_shape(), dout, ng_ksize_shape, ng_strides, ng_padding_below,
+          ng_padding_above, !padding_exclusive);
+    }
+    paddle::platform::SetOutputNode(op, "X@GRAD", pool2d_grad, ngb_node_map);
+  } else {
+    PADDLE_THROW("Support max and avg pooling only");
+  }
+}
+}  // namespace ngraphs
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/operators/reshape_op.cc

@@ -330,6 +330,7 @@ class Reshape2GradOp : public framework::OperatorWithKernel {
 }  // namespace operators
 }  // namespace paddle
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 
 REGISTER_OPERATOR(reshape, ops::ReshapeOp, ops::ReshapeOpMaker,
                   paddle::framework::DefaultGradOpDescMaker<true>);
@@ -356,16 +357,20 @@ REGISTER_OP_CPU_KERNEL_FUNCTOR(reshape2_grad, float, ops::ReshapeGradKernel,
 #ifdef PADDLE_WITH_CUDA
 REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape, float, ops::ReshapeKernel, double,
                                 ops::ReshapeKernel, int, ops::ReshapeKernel,
-                                int64_t, ops::ReshapeKernel);
+                                int64_t, ops::ReshapeKernel, plat::float16,
+                                ops::ReshapeKernel);
 REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape_grad, float, ops::ReshapeGradKernel,
                                 double, ops::ReshapeGradKernel, int,
                                 ops::ReshapeGradKernel, int64_t,
+                                ops::ReshapeGradKernel, plat::float16,
                                 ops::ReshapeGradKernel);
 REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape2, float, ops::ReshapeKernel, double,
                                 ops::ReshapeKernel, int, ops::ReshapeKernel,
-                                int64_t, ops::ReshapeKernel);
+                                int64_t, ops::ReshapeKernel, plat::float16,
+                                ops::ReshapeKernel);
 REGISTER_OP_CUDA_KERNEL_FUNCTOR(reshape2_grad, float, ops::ReshapeGradKernel,
                                 double, ops::ReshapeGradKernel, int,
                                 ops::ReshapeGradKernel, int64_t,
+                                ops::ReshapeGradKernel, plat::float16,
                                 ops::ReshapeGradKernel);
 #endif

paddle/fluid/operators/slice_op.cc

@@ -54,6 +54,9 @@ class SliceOp : public framework::OperatorWithKernel {
       out_dims[axes[i]] = end - start;
     }
     ctx->SetOutputDim("Out", out_dims);
+    if (axes[0] != 0) {
+      ctx->ShareLoD("Input", /*->*/ "Out");
+    }
   }
 
  protected:

paddle/fluid/operators/stack_op.cu

@@ -17,13 +17,16 @@
 namespace plat = paddle::platform;
 namespace ops = paddle::operators;
 
-REGISTER_OP_CUDA_KERNEL(stack, ops::StackKernel<plat::CUDADeviceContext, float>,
-                        ops::StackKernel<plat::CUDADeviceContext, double>,
-                        ops::StackKernel<plat::CUDADeviceContext, int>,
-                        ops::StackKernel<plat::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    stack, ops::StackKernel<plat::CUDADeviceContext, float>,
+    ops::StackKernel<plat::CUDADeviceContext, double>,
+    ops::StackKernel<plat::CUDADeviceContext, int>,
+    ops::StackKernel<plat::CUDADeviceContext, int64_t>,
+    ops::StackKernel<plat::CUDADeviceContext, plat::float16>);
 
-REGISTER_OP_CUDA_KERNEL(stack_grad,
-                        ops::StackGradKernel<plat::CUDADeviceContext, float>,
-                        ops::StackGradKernel<plat::CUDADeviceContext, double>,
-                        ops::StackGradKernel<plat::CUDADeviceContext, int>,
-                        ops::StackGradKernel<plat::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    stack_grad, ops::StackGradKernel<plat::CUDADeviceContext, float>,
+    ops::StackGradKernel<plat::CUDADeviceContext, double>,
+    ops::StackGradKernel<plat::CUDADeviceContext, int>,
+    ops::StackGradKernel<plat::CUDADeviceContext, int64_t>,
+    ops::StackGradKernel<plat::CUDADeviceContext, plat::float16>);

paddle/fluid/operators/tensorrt/tensorrt_engine_op.cc

@@ -29,8 +29,14 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Xs", "A list of inputs.").AsDuplicable();
     AddOutput("Ys", "A list of outputs").AsDuplicable();
     AddAttr<std::string>("subgraph", "the subgraph.");
+    AddAttr<std::string>("calibration_data", "the calibration data for int8");
+    AddAttr<std::string>(
+        "engine_key",
+        "The engine_key here is used to distinguish different TRT Engines");
     AddAttr<int>("max_batch_size", "the maximum batch size.");
     AddAttr<int>("workspace_size", "the workspace size.");
+    AddAttr<framework::BlockDesc *>("sub_block", "the trt block");
+    AddAttr<bool>("enable_int8", "whether swith to int8 mode");
     AddComment("TensorRT engine operator.");
   }
 };
@@ -47,6 +53,6 @@ class TensorRTEngineInferVarType : public framework::VarTypeInference {
 namespace ops = paddle::operators;
 
 REGISTER_OPERATOR(tensorrt_engine, ops::TensorRTEngineOp,
-                  ops::TensorRTEngineOpMaker);
+                  ops::TensorRTEngineOpMaker, ops::TensorRTEngineOpMaker);
 
 #endif  // PADDLE_WITH_CUDA

paddle/fluid/operators/tensorrt/tensorrt_engine_op.h

@@ -17,8 +17,10 @@
 #ifdef PADDLE_WITH_CUDA
 
 #include <string>
+#include <unordered_map>
 #include <vector>
 
+#include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/inference/analysis/helper.h"
@@ -62,6 +64,9 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t> &shape) {
 
 using inference::Singleton;
 using inference::tensorrt::TensorRTEngine;
+using inference::tensorrt::TRTInt8Calibrator;
+using inference::tensorrt::TRTCalibratorEngine;
+using inference::tensorrt::TRTCalibratorEngineManager;
 
 class TensorRTEngineOp : public framework::OperatorBase {
  private:
@@ -70,6 +75,11 @@ class TensorRTEngineOp : public framework::OperatorBase {
   mutable std::unique_ptr<TensorRTEngine> trt_engine_;
   int max_batch_size_;
   int workspace_size_;
+  std::unique_ptr<TRTInt8Calibrator> calibrator_;
+  bool enable_int8_;
+  std::string calibration_data_;
+  std::string engine_key_;
+  bool calibration_mode_;
 
  public:
   TensorRTEngineOp(const std::string &type,
@@ -80,19 +90,96 @@ class TensorRTEngineOp : public framework::OperatorBase {
     input_names_ = Inputs("Xs");
     max_batch_size_ = Attr<int>("max_batch_size");
     workspace_size_ = Attr<int>("workspace_size");
+    enable_int8_ = Attr<bool>("enable_int8");
+    calibration_data_ = Attr<std::string>("calibration_data");
+    engine_key_ = Attr<std::string>("engine_key");
 
     auto params = Attr<std::vector<std::string>>("parameters");
     for (const auto &param : params) {
       param_names_.insert(param);
     }
+    // calibration_mode is ture represents we need to
+    // generate the calibration table data.
+    calibration_mode_ = (enable_int8_ && calibration_data_.size() == 0);
+
+    VLOG(4) << "calibration_mode: " << calibration_mode_;
+    if (enable_int8_ && calibration_data_.size()) {
+      calibrator_.reset(new TRTInt8Calibrator(calibration_data_));
+    }
   }
 
  protected:
+  void RunNativeImpl(const framework::Scope &scope,
+                     const platform::Place &dev_place) const {
+    framework::Executor executor(dev_place);
+    auto *block = Attr<framework::BlockDesc *>("sub_block");
+    auto *program = block->Program();
+    auto &current_scope = scope.NewScope();
+    auto ctx = executor.Prepare(*program, block->ID());
+    executor.RunPreparedContext(ctx.get(), &current_scope, false, true, true);
+  }
+
   void RunImpl(const framework::Scope &scope,
                const platform::Place &dev_place) const override {
+    if (calibration_mode_ == true) {
+      RunCalibration(scope, dev_place);
+      return;
+    }
     RunTrt(scope, dev_place);
   }
 
+  void RunCalibration(const framework::Scope &scope,
+                      const platform::Place &dev_place) const {
+    // This process will builds a 32-bit trt engine, runs it on the calibration
+    // set, and records a histogram for each
+    // tensor of the distribution of activation values.
+    LOG_FIRST_N(INFO, 1) << "The TRT engine: " << engine_key_
+                         << " is running calibration trt int8... ";
+    int runtime_batch = 1;
+    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+    auto &dev_ctx = *pool.Get(dev_place);
+    auto stream =
+        reinterpret_cast<const platform::CUDADeviceContext &>(dev_ctx).stream();
+    if (!Singleton<TRTCalibratorEngineManager>::Global().Has(engine_key_)) {
+      TRTCalibratorEngine *calib_res =
+          Singleton<TRTCalibratorEngineManager>::Global().Create(engine_key_);
+      std::unordered_map<std::string, size_t> calib_buffers;
+      for (auto &x : input_names_) {
+        if (param_names_.count(x)) continue;
+        auto &t =
+            inference::analysis::GetFromScope<framework::LoDTensor>(scope, x);
+        calib_buffers[x] = t.memory_size();
+        auto t_shape = framework::vectorize(t.dims());
+        runtime_batch = t_shape[0];
+      }
+      calib_res->calib_.reset(new TRTInt8Calibrator(
+          calib_buffers, runtime_batch, engine_key_, dev_place));
+      calib_res->thr_.reset(new std::thread([&]() {
+        calib_res->engine_.reset(new TensorRTEngine(
+            max_batch_size_, workspace_size_, stream,
+            boost::get<platform::CUDAPlace>(dev_place).device, enable_int8_,
+            calib_res->calib_.get()));
+        VLOG(3) << "start the calib trt engine thread";
+        Prepare(scope, dev_place, calib_res->engine_.get());
+      }));
+    }
+
+    TRTInt8Calibrator *temp_calibrator =
+        Singleton<TRTCalibratorEngineManager>::Global()
+            .Get(engine_key_)
+            ->calib_.get();
+    std::unordered_map<std::string, void *> calib_data;
+
+    for (auto &x : Inputs("Xs")) {
+      if (param_names_.count(x)) continue;
+      auto &t =
+          inference::analysis::GetFromScope<framework::LoDTensor>(scope, x);
+      calib_data.emplace(x, t.data<void>());
+    }
+    temp_calibrator->setBatch(calib_data);
+    RunNativeImpl(scope, dev_place);
+  }
+
   void RunTrt(const framework::Scope &scope,
               const platform::Place &dev_place) const {
     int runtime_batch = 1;
@@ -101,9 +188,10 @@ class TensorRTEngineOp : public framework::OperatorBase {
     auto stream =
         reinterpret_cast<const platform::CUDADeviceContext &>(dev_ctx).stream();
     if (trt_engine_.get() == nullptr) {
-      trt_engine_.reset(new TensorRTEngine(
-          max_batch_size_, workspace_size_, stream,
-          boost::get<platform::CUDAPlace>(dev_place).device));
+      trt_engine_.reset(
+          new TensorRTEngine(max_batch_size_, workspace_size_, stream,
+                             boost::get<platform::CUDAPlace>(dev_place).device,
+                             enable_int8_, calibrator_.get()));
       Prepare(scope, dev_place, trt_engine_.get());
     }
 
@@ -173,7 +261,8 @@ class TensorRTEngineOp : public framework::OperatorBase {
 
   void Prepare(const framework::Scope &scope, const platform::Place &dev_place,
                TensorRTEngine *engine) const {
-    VLOG(4) << "Prepare engine";
+    LOG(INFO) << "Prepare TRT engine (Optimize model structure, Select OP "
+                 "kernel etc). This process may cost a lot of time.";
     framework::proto::BlockDesc block_desc;
     block_desc.ParseFromString(Attr<std::string>("subgraph"));
 

paddle/fluid/operators/tensorrt/tensorrt_engine_op_test.cc

@@ -96,19 +96,20 @@ TEST(TensorRTEngineOp, manual) {
   engine_op_desc.SetType("tensorrt_engine");
   engine_op_desc.SetInput("Xs", std::vector<std::string>({"x"}));
   engine_op_desc.SetOutput("Ys", std::vector<std::string>({"z0"}));
-  SetAttr<std::string>(engine_op_desc.Proto(), "subgraph",
-                       block_->SerializeAsString());
-  SetAttr<int>(engine_op_desc.Proto(), "max_batch_size", 2);
-  SetAttr<int>(engine_op_desc.Proto(), "workspace_size", 1 << 20);
-  SetAttr<std::string>(engine_op_desc.Proto(), "engine_uniq_key", "a_engine");
-  SetAttr<std::vector<std::string>>(engine_op_desc.Proto(), "parameters",
-                                    std::vector<std::string>({}));
-  SetAttr<std::vector<std::string>>(engine_op_desc.Proto(),
-                                    "output_name_mapping",
-                                    std::vector<std::string>({"z0"}));
+
+  engine_op_desc.SetBlockAttr("sub_block", &block_desc);
+  engine_op_desc.SetAttr("max_batch_size", static_cast<int>(2));
+  engine_op_desc.SetAttr("workspace_size", static_cast<int>(1 << 20));
+  engine_op_desc.SetAttr("parameters", std::vector<std::string>({}));
+  engine_op_desc.SetAttr("engine_key", std::string("a_engine"));
+  engine_op_desc.SetAttr("calibration_data", std::string(""));
+  engine_op_desc.SetAttr("enable_int8", static_cast<bool>(false));
+  engine_op_desc.SetAttr("output_name_mapping",
+                         std::vector<std::string>({"z0"}));
+  engine_op_desc.SetAttr("subgraph", std::string(block_->SerializeAsString()));
 
   LOG(INFO) << "create engine op";
-  auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto());
+  auto engine_op = framework::OpRegistry::CreateOp(engine_op_desc);
   LOG(INFO) << "engine_op " << engine_op.get();
 
   framework::Scope scope;
@@ -190,20 +191,19 @@ void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) {
   engine_op_desc.SetInput("Xs", std::vector<std::string>({"x0"}));
   engine_op_desc.SetOutput("Ys", std::vector<std::string>({"z3"}));
 
-  SetAttr<std::string>(engine_op_desc.Proto(), "subgraph",
-                       block_->SerializeAsString());
-  SetAttr<int>(engine_op_desc.Proto(), "max_batch_size", batch_size);
-  SetAttr<int>(engine_op_desc.Proto(), "workspace_size", 1 << 20);
-  SetAttr<std::vector<std::string>>(
-      engine_op_desc.Proto(), "parameters",
-      std::vector<std::string>({"y0", "y1", "y2", "y3"}));
-  SetAttr<std::string>(engine_op_desc.Proto(), "engine_uniq_key", "b_engine");
-
-  SetAttr<std::vector<std::string>>(engine_op_desc.Proto(),
-                                    "output_name_mapping",
-                                    std::vector<std::string>({"z3"}));
-
-  auto engine_op = framework::OpRegistry::CreateOp(*engine_op_desc.Proto());
+  engine_op_desc.SetBlockAttr("sub_block", &block_desc);
+  engine_op_desc.SetAttr("max_batch_size", static_cast<int>(batch_size));
+  engine_op_desc.SetAttr("workspace_size", static_cast<int>(1 << 20));
+  engine_op_desc.SetAttr("parameters",
+                         std::vector<std::string>({"y0", "y1", "y2", "y3"}));
+  engine_op_desc.SetAttr("engine_key", std::string("b_engine"));
+  engine_op_desc.SetAttr("calibration_data", std::string(""));
+  engine_op_desc.SetAttr("enable_int8", static_cast<bool>(false));
+  engine_op_desc.SetAttr("output_name_mapping",
+                         std::vector<std::string>({"z3"}));
+  engine_op_desc.SetAttr("subgraph", std::string(block_->SerializeAsString()));
+
+  auto engine_op = framework::OpRegistry::CreateOp(engine_op_desc);
 
   // Execute them.
   engine_op->Run(scope, place);

paddle/fluid/operators/transpose_op.cu.cc

@@ -15,19 +15,27 @@ limitations under the License. */
 #include "paddle/fluid/operators/transpose_op.h"
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
 REGISTER_OP_CUDA_KERNEL(
     transpose, ops::TransposeKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::TransposeKernel<paddle::platform::CUDADeviceContext, double>);
+    ops::TransposeKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::TransposeKernel<paddle::platform::CUDADeviceContext, plat::float16>);
 REGISTER_OP_CUDA_KERNEL(
     transpose_grad,
     ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, double>);
+    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext,
+                             plat::float16>);
 
 REGISTER_OP_CUDA_KERNEL(
     transpose2,
     ops::TransposeKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::TransposeKernel<paddle::platform::CUDADeviceContext, double>);
+    ops::TransposeKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::TransposeKernel<paddle::platform::CUDADeviceContext, plat::float16>);
 REGISTER_OP_CUDA_KERNEL(
     transpose2_grad,
     ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, double>);
+    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::TransposeGradKernel<paddle::platform::CUDADeviceContext,
+                             plat::float16>);

paddle/fluid/platform/dynload/mklml.h

@@ -70,6 +70,8 @@ extern void* mklml_dso_handle;
   __macro(cblas_ddot);              \
   __macro(cblas_sasum);             \
   __macro(cblas_dasum);             \
+  __macro(cblas_isamax);            \
+  __macro(cblas_idamax);            \
   __macro(cblas_sscal);             \
   __macro(cblas_dscal);             \
   __macro(vsAdd);                   \

paddle/fluid/platform/enforce.h

@@ -71,9 +71,8 @@ struct EnforceNotMet : public std::exception {
     }
   }
 
-  template <typename... ARGS>
-  EnforceNotMet(const char* f, int l, ARGS... args) {
-    Init(string::Sprintf(args...), f, l);
+  EnforceNotMet(const std::string& str, const char* f, int l) {
+    Init(str, f, l);
   }
 
   const char* what() const noexcept override { return err_str_.c_str(); }
@@ -142,28 +141,23 @@ struct EOFException : public std::exception {
 
 inline bool is_error(bool stat) { return !stat; }
 
-template <typename... Args>
-inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
-    bool stat, const Args&... args) {
+inline void throw_on_error(bool stat, const std::string& msg) {
 #ifndef REPLACE_ENFORCE_GLOG
-  throw std::runtime_error(string::Sprintf(args...));
+  throw std::runtime_error(msg);
 #else
-  LOG(FATAL) << string::Sprintf(args...);
+  LOG(FATAL) << msg;
 #endif
 }
 
 #ifdef PADDLE_WITH_CUDA
 
-inline bool is_error(cudaError_t e) { return UNLIKELY(e); }
+inline bool is_error(cudaError_t e) { return e != cudaSuccess; }
 
-template <typename... Args>
-inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
-    cudaError_t e, const Args&... args) {
+inline void throw_on_error(cudaError_t e, const std::string& msg) {
 #ifndef REPLACE_ENFORCE_GLOG
-  throw thrust::system_error(e, thrust::cuda_category(),
-                             string::Sprintf(args...));
+  throw thrust::system_error(e, thrust::cuda_category(), msg);
 #else
-  LOG(FATAL) << string::Sprintf(args...);
+  LOG(FATAL) << msg;
 #endif
 }
 
@@ -171,14 +165,12 @@ inline bool is_error(curandStatus_t stat) {
   return stat != CURAND_STATUS_SUCCESS;
 }
 
-template <typename... Args>
-inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
-    curandStatus_t stat, const Args&... args) {
+inline void throw_on_error(curandStatus_t stat, const std::string& msg) {
 #ifndef REPLACE_ENFORCE_GLOG
   throw thrust::system_error(cudaErrorLaunchFailure, thrust::cuda_category(),
-                             string::Sprintf(args...));
+                             msg);
 #else
-  LOG(FATAL) << string::Sprintf(args...);
+  LOG(FATAL) << msg;
 #endif
 }
 
@@ -186,14 +178,11 @@ inline bool is_error(cudnnStatus_t stat) {
   return stat != CUDNN_STATUS_SUCCESS;
 }
 
-template <typename... Args>
-inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
-    cudnnStatus_t stat, const Args&... args) {
+inline void throw_on_error(cudnnStatus_t stat, const std::string& msg) {
 #ifndef REPLACE_ENFORCE_GLOG
-  throw std::runtime_error(platform::dynload::cudnnGetErrorString(stat) +
-                           string::Sprintf(args...));
+  throw std::runtime_error(platform::dynload::cudnnGetErrorString(stat) + msg);
 #else
-  LOG(FATAL) << string::Sprintf(args...);
+  LOG(FATAL) << platform::dynload::cudnnGetErrorString(stat) << msg;
 #endif
 }
 
@@ -201,9 +190,7 @@ inline bool is_error(cublasStatus_t stat) {
   return stat != CUBLAS_STATUS_SUCCESS;
 }
 
-template <typename... Args>
-inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
-    cublasStatus_t stat, const Args&... args) {
+inline void throw_on_error(cublasStatus_t stat, const std::string& msg) {
   std::string err;
   if (stat == CUBLAS_STATUS_NOT_INITIALIZED) {
     err = "CUBLAS: not initialized, ";
@@ -225,87 +212,45 @@ inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
     err = "CUBLAS: license error, ";
   }
 #ifndef REPLACE_ENFORCE_GLOG
-  throw std::runtime_error(err + string::Sprintf(args...));
+  throw std::runtime_error(err + msg);
 #else
-  LOG(FATAL) << err << string::Sprintf(args...);
+  LOG(FATAL) << err << msg;
 #endif
 }
 
 #if !defined(__APPLE__) && !defined(_WIN32)
-template <typename... Args>
-inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
-    ncclResult_t stat, const Args&... args) {
-  if (stat == ncclSuccess) {
-    return;
-  } else {
+inline bool is_error(ncclResult_t nccl_result) {
+  return nccl_result != ncclSuccess;
+}
+
+inline void throw_on_error(ncclResult_t stat, const std::string& msg) {
 #ifndef REPLACE_ENFORCE_GLOG
-    throw std::runtime_error(platform::dynload::ncclGetErrorString(stat) +
-                             string::Sprintf(args...));
+  throw std::runtime_error(platform::dynload::ncclGetErrorString(stat) + msg);
 #else
-    LOG(FATAL) << platform::dynload::ncclGetErrorString(stat)
-               << string::Sprintf(args...);
+  LOG(FATAL) << platform::dynload::ncclGetErrorString(stat) << msg;
 #endif
-  }
 }
 #endif  // __APPLE__ and windows
 #endif  // PADDLE_WITH_CUDA
 
-template <typename T>
-inline void throw_on_error(T e) {
-  throw_on_error(e, "");
-}
-
-#define PADDLE_THROW(...) \
-  throw ::paddle::platform::EnforceNotMet(__FILE__, __LINE__, __VA_ARGS__)
-
-#define __PADDLE_THROW_ERROR_I(_, _9, _8, _7, _6, _5, _4, _3, _2, X_, ...) X_;
-
-#define __THROW_ON_ERROR_ONE_ARG(COND, ARG) \
-  ::paddle::platform::throw_on_error(COND, ::paddle::string::Sprintf(ARG));
-
-#ifdef _WIN32
-#define __PADDLE_THROW_ON_ERROR(COND, ...) \
-  __THROW_ON_ERROR_ONE_ARG(COND, __VA_ARGS__)
-#else  // _WIN32
-#define __PADDLE_THROW_ON_ERROR(COND, ...)                                \
-  __PADDLE_THROW_ERROR_I(                                                 \
-      __VA_ARGS__, ::paddle::platform::throw_on_error(COND, __VA_ARGS__), \
-      ::paddle::platform::throw_on_error(COND, __VA_ARGS__),              \
-      ::paddle::platform::throw_on_error(COND, __VA_ARGS__),              \
-      ::paddle::platform::throw_on_error(COND, __VA_ARGS__),              \
-      ::paddle::platform::throw_on_error(COND, __VA_ARGS__),              \
-      ::paddle::platform::throw_on_error(COND, __VA_ARGS__),              \
-      ::paddle::platform::throw_on_error(COND, __VA_ARGS__),              \
-      ::paddle::platform::throw_on_error(COND, __VA_ARGS__),              \
-      __THROW_ON_ERROR_ONE_ARG(COND, __VA_ARGS__))
-#endif  // _WIN32
-
-#define __PADDLE_UNARY_COMPARE(COND, ...)                 \
-  do {                                                    \
-    auto __cond = COND;                                   \
-    if (UNLIKELY(::paddle::platform::is_error(__cond))) { \
-      __PADDLE_THROW_ON_ERROR(__cond, __VA_ARGS__);       \
-    }                                                     \
+#define PADDLE_THROW(...)                  \
+  throw ::paddle::platform::EnforceNotMet( \
+      ::paddle::string::Sprintf(__VA_ARGS__), __FILE__, __LINE__)
+
+#define PADDLE_ENFORCE(COND, ...)                                         \
+  do {                                                                    \
+    auto __cond__ = (COND);                                               \
+    if (UNLIKELY(::paddle::platform::is_error(__cond__))) {               \
+      try {                                                               \
+        ::paddle::platform::throw_on_error(                               \
+            __cond__, ::paddle::string::Sprintf(__VA_ARGS__));            \
+      } catch (...) {                                                     \
+        throw ::paddle::platform::EnforceNotMet(std::current_exception(), \
+                                                __FILE__, __LINE__);      \
+      }                                                                   \
+    }                                                                     \
   } while (0)
 
-#ifndef REPLACE_ENFORCE_GLOG
-#define __PADDLE_ENFORCE_I(COND, ...)                                   \
-  do {                                                                  \
-    try {                                                               \
-      __PADDLE_UNARY_COMPARE(COND, __VA_ARGS__);                        \
-    } catch (...) {                                                     \
-      throw ::paddle::platform::EnforceNotMet(std::current_exception(), \
-                                              __FILE__, __LINE__);      \
-    }                                                                   \
-  } while (0)
-
-#else
-#define __PADDLE_ENFORCE_I(COND, ...) __PADDLE_UNARY_COMPARE(COND, __VA_ARGS__);
-#endif  // REPLACE_ENFORCE_GLOG
-
-#define __PADDLE_ENFORCE(__args) __PADDLE_ENFORCE_I __args
-#define PADDLE_ENFORCE(...) __PADDLE_ENFORCE((__VA_ARGS__))
-
 #define PADDLE_THROW_EOF()                                                     \
   do {                                                                         \
     throw ::paddle::platform::EOFException("There is no next data.", __FILE__, \

paddle/fluid/platform/nccl_helper.h

@@ -64,7 +64,7 @@ class NCCLGroupGuard {
   }
 
   inline ~NCCLGroupGuard() {
-    CHECK_EQ(dynload::ncclGroupEnd(), ncclSuccess);
+    PADDLE_ENFORCE(dynload::ncclGroupEnd());
     NCCLMutex().unlock();
   }
 };

paddle/fluid/pybind/inference_api.cc

@@ -33,7 +33,6 @@ using paddle::PaddlePredictor;
 using paddle::NativeConfig;
 using paddle::NativePaddlePredictor;
 using paddle::AnalysisPredictor;
-using paddle::contrib::AnalysisConfig;
 
 static void BindPaddleDType(py::module *m);
 static void BindPaddleBuf(py::module *m);
@@ -180,8 +179,14 @@ void BindNativePredictor(py::module *m) {
 }
 
 void BindAnalysisConfig(py::module *m) {
-  py::class_<AnalysisConfig>(*m, "AnalysisConfig")
-      .def(py::init<const AnalysisConfig &>())
+  py::class_<AnalysisConfig> analysis_config(*m, "AnalysisConfig");
+
+  py::enum_<AnalysisConfig::Precision>(analysis_config, "Precision")
+      .value("Float32", AnalysisConfig::Precision::kFloat32)
+      .value("Int8", AnalysisConfig::Precision::kInt8)
+      .export_values();
+
+  analysis_config.def(py::init<const AnalysisConfig &>())
       .def(py::init<const std::string &>())
       .def(py::init<const std::string &, const std::string &>())
       .def("set_model", (void (AnalysisConfig::*)(const std::string &)) &
@@ -215,7 +220,8 @@ void BindAnalysisConfig(py::module *m) {
       .def("specify_input_name", &AnalysisConfig::specify_input_name)
       .def("enable_tensorrt_engine", &AnalysisConfig::EnableTensorRtEngine,
            py::arg("workspace_size") = 1 << 20, py::arg("max_batch_size") = 1,
-           py::arg("min_subgraph_size") = 3)
+           py::arg("min_subgraph_size") = 3,
+           py::arg("precision_mode") = AnalysisConfig::Precision::kFloat32)
       .def("tensorrt_engine_enabled", &AnalysisConfig::tensorrt_engine_enabled)
       .def("switch_ir_debug", &AnalysisConfig::SwitchIrDebug,
            py::arg("x") = true)

paddle/fluid/string/printf.h

@@ -84,6 +84,8 @@ void Fprintf(std::ostream& out, const char* fmt, const Args&... args) {
   tinyformat::vformat(out, fmt, tinyformat::makeFormatList(args...));
 }
 
+inline std::string Sprintf() { return ""; }
+
 template <typename... Args>
 std::string Sprintf(const Args&... args) {
   std::ostringstream oss;

paddle/scripts/fast_install.sh

@@ -40,9 +40,11 @@ function checkMacPython2(){
           else
             python_version=""
           fi
+          check_python=`echo $python_version | grep "Python 2"`
+          echo $check_python
           if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then
                python_version=""
-          else
+          elif [ -n "$check_python" ];then
               while true
                 do
                   read -p "找到：$python_version, 是否使用：(y/n)，输入n来输入自定义使用的python路径，或者按ctrl + c退出： " use_python
@@ -60,6 +62,9 @@ function checkMacPython2(){
               if [ "$use_python" == "y" ];then
                 break
               fi
+          else
+               echo "您输入Python的不是Python2"
+               python_version=""
           fi
        done
 }
@@ -77,9 +82,10 @@ function checkMacPython3(){
           else
               python_version=""
           fi
+          check_python=`echo $python_version | grep "Python 3"`
           if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then
                python_version=""
-          else
+          elif [ -n "$check_python" ] ;then
               while true
                 do
                   read -p "找到：$python_version, 是否使用：(y/n)，输入n来输入自定义使用的python路径，或者按ctrl + c退出： " use_python
@@ -97,6 +103,9 @@ function checkMacPython3(){
               if [ "$use_python" == "y" ];then
                     break
               fi
+          else
+              echo "您输入Python的不是Python2"
+              python_version=""
           fi
        done
 }

paddle/scripts/paddle_build.sh

@@ -173,7 +173,6 @@ function cmake_gen() {
         -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release}
         ${PYTHON_FLAGS}
         -DWITH_DSO=ON
-        -DWITH_DOC=${WITH_DOC:-OFF}
         -DWITH_GPU=${WITH_GPU:-OFF}
         -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF}
         -DWITH_DISTRIBUTE=${distibuted_flag}
@@ -208,7 +207,6 @@ EOF
         -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release} \
         ${PYTHON_FLAGS} \
         -DWITH_DSO=ON \
-        -DWITH_DOC=${WITH_DOC:-OFF} \
         -DWITH_GPU=${WITH_GPU:-OFF} \
         -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF} \
         -DWITH_DISTRIBUTE=${distibuted_flag} \
@@ -528,31 +526,6 @@ function bind_test() {
     wait
 }
 
-
-function gen_docs() {
-    mkdir -p ${PADDLE_ROOT}/build
-    cd ${PADDLE_ROOT}/build
-    cat <<EOF
-    ========================================
-    Building documentation ...
-    In /paddle/build
-    ========================================
-EOF
-    cmake .. \
-        -DCMAKE_BUILD_TYPE=Release \
-        -DWITH_DOC=ON \
-        -DWITH_GPU=OFF \
-        -DWITH_MKL=OFF
-
-    make -j `nproc` paddle_docs paddle_apis
-
-    # check websites for broken links
-    linkchecker doc/v2/en/html/index.html
-    linkchecker doc/v2/cn/html/index.html
-    linkchecker doc/v2/api/en/html/index.html
-
-}
-
 function gen_doc_lib() {
     mkdir -p ${PADDLE_ROOT}/build
     cd ${PADDLE_ROOT}/build
@@ -564,7 +537,6 @@ function gen_doc_lib() {
 EOF
     cmake .. \
         -DCMAKE_BUILD_TYPE=Release \
-        -DWITH_DOC=ON \
         -DWITH_GPU=OFF \
         -DWITH_MKL=OFF \
         -DWITH_FLUID_ONLY=ON
@@ -803,9 +775,6 @@ function main() {
       bind_test)
         bind_test
         ;;
-      doc)
-        gen_docs
-        ;;
       gen_doc_lib)
         gen_doc_lib $2
         ;;

python/paddle/fluid/contrib/tests/test_calibration.py

@@ -19,10 +19,9 @@ import sys
 import random
 import paddle
 import paddle.fluid as fluid
-import argparse
 import functools
 import contextlib
-import paddle.fluid.profiler as profiler
+from paddle.dataset.common import download
 from PIL import Image, ImageEnhance
 import math
 sys.path.append('..')
@@ -42,7 +41,7 @@ img_mean = np.array([0.485, 0.456, 0.406]).reshape((3, 1, 1))
 img_std = np.array([0.229, 0.224, 0.225]).reshape((3, 1, 1))
 
 
-# TODO(guomingz): Remove duplicated code from line 45 ~ line 114
+# TODO(guomingz): Remove duplicated code from resize_short, crop_image, process_image, _reader_creator
 def resize_short(img, target_size):
     percent = float(target_size) / min(img.size[0], img.size[1])
     resized_width = int(round(img.size[0] * percent))
@@ -116,26 +115,88 @@ def val(data_dir=DATA_DIR):
     return _reader_creator(file_list, 'val', shuffle=False, data_dir=data_dir)
 
 
-class TestCalibration(unittest.TestCase):
+class TestCalibrationForResnet50(unittest.TestCase):
     def setUp(self):
-        # TODO(guomingz): Put the download process in the cmake.
-        # Download and unzip test data set
-        imagenet_dl_url = 'http://paddle-inference-dist.cdn.bcebos.com/int8/calibration_test_data.tar.gz'
-        zip_file_name = imagenet_dl_url.split('/')[-1]
-        cmd = 'rm -rf data {}  && mkdir data && wget {} && tar xvf {} -C data'.format(
-            zip_file_name, imagenet_dl_url, zip_file_name)
-        os.system(cmd)
-        # resnet50 fp32 data
-        resnet50_fp32_model_url = 'http://paddle-inference-dist.cdn.bcebos.com/int8/resnet50_int8_model.tar.gz'
-        resnet50_zip_name = resnet50_fp32_model_url.split('/')[-1]
-        resnet50_unzip_folder_name = 'resnet50_fp32'
-        cmd = 'rm -rf {} {} && mkdir {} && wget {} && tar xvf {} -C {}'.format(
-            resnet50_unzip_folder_name, resnet50_zip_name,
-            resnet50_unzip_folder_name, resnet50_fp32_model_url,
-            resnet50_zip_name, resnet50_unzip_folder_name)
+        self.int8_download = 'int8/download'
+        self.cache_folder = os.path.expanduser('~/.cache/paddle/dataset/' +
+                                               self.int8_download)
+
+        data_urls = []
+        data_md5s = []
+        self.data_cache_folder = ''
+        if os.environ.get('DATASET') == 'full':
+            data_urls.append(
+                'https://paddle-inference-dist.bj.bcebos.com/int8/ILSVRC2012_img_val.tar.gz.partaa'
+            )
+            data_md5s.append('60f6525b0e1d127f345641d75d41f0a8')
+            data_urls.append(
+                'https://paddle-inference-dist.bj.bcebos.com/int8/ILSVRC2012_img_val.tar.gz.partab'
+            )
+            data_md5s.append('1e9f15f64e015e58d6f9ec3210ed18b5')
+            self.data_cache_folder = self.download_data(data_urls, data_md5s,
+                                                        "full_data", False)
+        else:
+            data_urls.append(
+                'http://paddle-inference-dist.cdn.bcebos.com/int8/calibration_test_data.tar.gz'
+            )
+            data_md5s.append('1b6c1c434172cca1bf9ba1e4d7a3157d')
+            self.data_cache_folder = self.download_data(data_urls, data_md5s,
+                                                        "small_data", False)
+
+        # reader/decorator.py requires the relative path to the data folder
+        cmd = 'rm -rf {0} && ln -s {1} {0}'.format("data",
+                                                   self.data_cache_folder)
         os.system(cmd)
 
-        self.iterations = 50
+        self.batch_size = 1
+        self.sample_iterations = 50
+        self.infer_iterations = 50000 if os.environ.get(
+            'DATASET') == 'full' else 50
+
+    def cache_unzipping(self, target_folder, zip_path):
+        if not os.path.exists(target_folder):
+            cmd = 'mkdir {0} && tar xf {1} -C {0}'.format(target_folder,
+                                                          zip_path)
+            os.system(cmd)
+
+    def download_data(self, data_urls, data_md5s, folder_name, is_model=True):
+        data_cache_folder = os.path.join(self.cache_folder, folder_name)
+        zip_path = ''
+        if os.environ.get('DATASET') == 'full':
+            file_names = []
+            for i in range(0, len(data_urls)):
+                download(data_urls[i], self.int8_download, data_md5s[i])
+                file_names.append(data_urls[i].split('/')[-1])
+
+            zip_path = os.path.join(self.cache_folder,
+                                    'full_imagenet_val.tar.gz')
+            if not os.path.exists(zip_path):
+                cat_command = 'cat'
+                for file_name in file_names:
+                    cat_command += ' ' + os.path.join(self.cache_folder,
+                                                      file_name)
+                cat_command += ' > ' + zip_path
+                os.system(cat_command)
+
+        if os.environ.get('DATASET') != 'full' or is_model:
+            download(data_urls[0], self.int8_download, data_md5s[0])
+            file_name = data_urls[0].split('/')[-1]
+            zip_path = os.path.join(self.cache_folder, file_name)
+
+        print('Data is downloaded at {0}').format(zip_path)
+        self.cache_unzipping(data_cache_folder, zip_path)
+        return data_cache_folder
+
+    def download_model(self):
+        # resnet50 fp32 data
+        data_urls = [
+            'http://paddle-inference-dist.cdn.bcebos.com/int8/resnet50_int8_model.tar.gz'
+        ]
+        data_md5s = ['4a5194524823d9b76da6e738e1367881']
+        self.model_cache_folder = self.download_data(data_urls, data_md5s,
+                                                     "resnet50_fp32")
+        self.model = "ResNet-50"
+        self.algo = "direct"
 
     def run_program(self, model_path, generate_int8=False, algo='direct'):
         image_shape = [3, 224, 224]
@@ -151,17 +212,17 @@ class TestCalibration(unittest.TestCase):
         t = fluid.transpiler.InferenceTranspiler()
         t.transpile(infer_program, fluid.CPUPlace())
 
-        val_reader = paddle.batch(val(), batch_size=1)
+        val_reader = paddle.batch(val(), self.batch_size)
+        iterations = self.infer_iterations
 
         if generate_int8:
             int8_model = os.path.join(os.getcwd(), "calibration_out")
+            iterations = self.sample_iterations
 
             if os.path.exists(int8_model):
                 os.system("rm -rf " + int8_model)
                 os.system("mkdir " + int8_model)
 
-            print("Start calibration ...")
-
             calibrator = int8_utility.Calibrator(
                 program=infer_program,
                 pretrained_model=model_path,
@@ -173,6 +234,7 @@ class TestCalibration(unittest.TestCase):
 
         test_info = []
         cnt = 0
+        periods = []
         for batch_id, data in enumerate(val_reader()):
             image = np.array(
                 [x[0].reshape(image_shape) for x in data]).astype("float32")
@@ -184,37 +246,79 @@ class TestCalibration(unittest.TestCase):
                 if op.has_attr("use_mkldnn"):
                     op._set_attr("use_mkldnn", True)
 
+            t1 = time.time()
             _, acc1, _ = exe.run(
                 running_program,
                 feed={feed_dict[0]: image,
                       feed_dict[1]: label},
                 fetch_list=fetch_targets)
+            t2 = time.time()
+            period = t2 - t1
+            periods.append(period)
+
             if generate_int8:
                 calibrator.sample_data()
 
             test_info.append(np.mean(acc1) * len(data))
             cnt += len(data)
 
-            if batch_id != self.iterations - 1:
-                continue
+            if (batch_id + 1) % 100 == 0:
+                print("{0} images,".format(batch_id + 1))
+                sys.stdout.flush()
 
-            break
+            if (batch_id + 1) == iterations:
+                break
 
         if generate_int8:
             calibrator.save_int8_model()
 
             print(
-                "Calibration is done and the corresponding files were generated at {}".
+                "Calibration is done and the corresponding files are generated at {}".
                 format(os.path.abspath("calibration_out")))
         else:
-            return np.sum(test_info) / cnt
-
-    def test_calibration_for_resnet50(self):
-        fp32_acc1 = self.run_program("resnet50_fp32/model")
-        self.run_program("resnet50_fp32/model", True)
-        int8_acc1 = self.run_program("calibration_out")
+            throughput = cnt / np.sum(periods)
+            latency = np.average(periods)
+            acc1 = np.sum(test_info) / cnt
+            return (throughput, latency, acc1)
+
+    def test_calibration(self):
+        self.download_model()
+        print("Start FP32 inference for {0} on {1} images ...").format(
+            self.model, self.infer_iterations)
+        (fp32_throughput, fp32_latency,
+         fp32_acc1) = self.run_program(self.model_cache_folder + "/model")
+        print("Start INT8 calibration for {0} on {1} images ...").format(
+            self.model, self.sample_iterations)
+        self.run_program(
+            self.model_cache_folder + "/model", True, algo=self.algo)
+        print("Start INT8 inference for {0} on {1} images ...").format(
+            self.model, self.infer_iterations)
+        (int8_throughput, int8_latency,
+         int8_acc1) = self.run_program("calibration_out")
         delta_value = np.abs(fp32_acc1 - int8_acc1)
         self.assertLess(delta_value, 0.01)
+        print(
+            "FP32 {0}: batch_size {1}, throughput {2} images/second, latency {3} second, accuracy {4}".
+            format(self.model, self.batch_size, fp32_throughput, fp32_latency,
+                   fp32_acc1))
+        print(
+            "INT8 {0}: batch_size {1}, throughput {2} images/second, latency {3} second, accuracy {4}".
+            format(self.model, self.batch_size, int8_throughput, int8_latency,
+                   int8_acc1))
+        sys.stdout.flush()
+
+
+class TestCalibrationForMobilenetv1(TestCalibrationForResnet50):
+    def download_model(self):
+        # mobilenetv1 fp32 data
+        data_urls = [
+            'http://paddle-inference-dist.cdn.bcebos.com/int8/mobilenetv1_int8_model.tar.gz'
+        ]
+        data_md5s = ['13892b0716d26443a8cdea15b3c6438b']
+        self.model_cache_folder = self.download_data(data_urls, data_md5s,
+                                                     "mobilenetv1_fp32")
+        self.model = "MobileNet-V1"
+        self.algo = "KL"
 
 
 if __name__ == '__main__':

python/paddle/fluid/framework.py

@@ -445,11 +445,16 @@ class Variable(object):
 
     @property
     def _stop_gradient(self):
-        return self._ivar.stop_gradient
+        if _in_imperative_mode():
+            return self._ivar.stop_gradient
+        else:
+            return self.stop_gradient
 
     @_stop_gradient.setter
     def _stop_gradient(self, s):
-        self._ivar.stop_gradient = s
+        if _in_imperative_mode():
+            self._ivar.stop_gradient = s
+        self.stop_gradient = s
 
     @property
     def persistable(self):
@@ -1310,6 +1315,9 @@ class Block(object):
             outputs=kwargs.get("outputs", None),
             attrs=kwargs.get("attrs", None))
         self.ops.append(op)
+
+        # TODO(minqiyang): add stop_gradient support in static mode too.
+        # currently, we only support stop_gradient in imperative mode.
         self._trace_op(op, kwargs.get("stop_gradient", False))
         return op
 

python/paddle/fluid/imperative/layers.py

@@ -15,6 +15,7 @@
 import contextlib
 import sys
 import numpy as np
+import collections
 
 from paddle.fluid import core
 from paddle.fluid import framework
@@ -31,7 +32,23 @@ class Layer(core.Layer):
         self._dtype = dtype
 
     def parameters(self):
-        return []
+        params = []
+        for key in self.__dict__.keys():
+            value = self.__dict__[key]
+            if isinstance(value, framework.Parameter):
+                params.append(value)
+            elif isinstance(value, core.Layer):
+                params.extend(value.parameters())
+            elif isinstance(value, collections.Container):
+                if len(value) == 0:
+                    continue
+                if isinstance(value[0], framework.Parameter):
+                    params.extend(value)
+                elif isinstance(value[0], core.Layer):
+                    for v in value:
+                        params.extend(v.parameters())
+
+        return params
 
     def clear_gradients(self):
         for p in self.parameters():

python/paddle/fluid/imperative/nn.py

@@ -22,13 +22,7 @@ from . import layers
 from ..framework import Variable, OpProtoHolder
 from ..param_attr import ParamAttr
 from ..initializer import Normal, Constant
-
-__all__ = [
-    'Conv2D',
-    'Pool2D',
-    'FC',
-    'BatchNorm',
-]
+__all__ = ['Conv2D', 'Pool2D', 'FC', 'BatchNorm', 'Embedding']
 
 
 class Conv2D(layers.Layer):
@@ -332,21 +326,16 @@ class BatchNorm(layers.Layer):
             shape=param_shape,
             dtype=self._dtype,
             default_initializer=Constant(1.0))
-
-        # TODO(minqiyang): change stop_gradient sign to trainable to align with static graph
-        #  # setting stop_gradient=True to reduce computation
-        #  if use_global_stats and self._helper.param_attr.learning_rate == 0.:
-        #  self._scale.stop_gradient = True
+        if use_global_stats and self._helper.param_attr.learning_rate == 0.:
+            self._scale._stop_gradient = True
 
         self._bias = self._helper.create_parameter(
             attr=self._helper.bias_attr,
             shape=param_shape,
             dtype=self._dtype,
             is_bias=True)
-        # TODO(minqiyang): change stop_gradient sign to trainable to align with static graph
-        #  # setting stop_gradient=True to reduce computation
-        #  if use_global_stats and self._helper.bias_attr.learning_rate == 0.:
-        #  self._bias.stop_gradient = True
+        if use_global_stats and self._helper.bias_attr.learning_rate == 0.:
+            self._bias._stop_gradient = True
 
         self._mean = self._helper.create_parameter(
             attr=ParamAttr(
@@ -356,7 +345,7 @@ class BatchNorm(layers.Layer):
                 do_model_average=do_model_average_for_mean_and_var),
             shape=param_shape,
             dtype=self._dtype)
-        self._mean.stop_gradient = True
+        self._mean._stop_gradient = True
 
         self._variance = self._helper.create_parameter(
             attr=ParamAttr(
@@ -366,7 +355,7 @@ class BatchNorm(layers.Layer):
                 do_model_average=do_model_average_for_mean_and_var),
             shape=param_shape,
             dtype=self._dtype)
-        self._variance.stop_gradient = True
+        self._variance._stop_gradient = True
 
         self._in_place = in_place
         self._momentum = momentum
@@ -419,3 +408,91 @@ class BatchNorm(layers.Layer):
 
         # Currently, we don't support inplace in imperative mode
         return self._helper.append_activation(batch_norm_out)
+
+
+class Embedding(layers.Layer):
+    """
+    **Embedding Layer**
+
+    This layer is used to lookup embeddings of IDs, provided by :attr:`input`, in
+    a lookup table. The result of this lookup is the embedding of each ID in the
+    :attr:`input`.
+
+    All the input variables are passed in as local variables to the LayerHelper
+    constructor.
+
+    Args:
+        size(tuple|list): The shape of the look up table parameter. It should
+            have two elements which indicate the size of the dictionary of
+            embeddings and the size of each embedding vector respectively.
+        is_sparse(bool): The flag indicating whether to use sparse update.
+        is_distributed(bool): Whether to run lookup table from remote parameter server.
+        padding_idx(int|long|None): If :attr:`None`, it makes no effect to lookup.
+            Otherwise the given :attr:`padding_idx` indicates padding the output
+            with zeros whenever lookup encounters it in :attr:`input`. If
+            :math:`padding_idx < 0`, the :attr:`padding_idx` to use in lookup is
+            :math:`size[0] + dim`.
+        param_attr(ParamAttr): Parameters for this layer
+        dtype(np.dtype|core.VarDesc.VarType|str): The type of data : float32, float_16, int etc
+
+    Returns:
+        Variable: The tensor variable storing the embeddings of the \
+                  supplied inputs.
+
+    Examples:
+        .. code-block:: python
+
+          dict_size = len(dataset.ids)
+          input = fluid.layers.data(name='ids', shape=[32, 32], dtype='float32')
+          embedding = fluid.imperative.Embedding(size=[dict_size, 16])
+          fc = embedding(input)
+    """
+
+    def __init__(self,
+                 size,
+                 is_sparse=False,
+                 is_distributed=False,
+                 padding_idx=None,
+                 param_attr=None,
+                 dtype='float32'):
+
+        super(Embedding, self).__init__()
+        self._size = size
+        self._is_sparse = is_sparse
+        self._is_distributed = is_distributed
+
+        self._padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else (
+            size[0] + padding_idx)
+
+        self._param_attr = param_attr
+        self._dtype = dtype
+        self._remote_prefetch = self._is_sparse and (not self._is_distributed)
+        if self._remote_prefetch:
+            assert self._is_sparse is True and self._is_distributed is False
+
+        from ..layer_helper import LayerHelper
+        self._helper = LayerHelper('embedding', param_attr=param_attr)
+        self._w = self._helper.create_parameter(
+            attr=self._param_attr,
+            shape=self._size,
+            dtype=self._dtype,
+            is_bias=False)
+
+    def parameters(self):
+        return [self._w]
+
+    def forward(self, input):
+        out = self._helper.create_variable_for_type_inference(self._dtype)
+        self._helper.append_op(
+            type='lookup_table',
+            inputs={'Ids': input,
+                    'W': self._w},
+            outputs={'Out': out},
+            attrs={
+                'is_sparse': self._is_sparse,
+                'is_distributed': self._is_distributed,
+                'remote_prefetch': self._remote_prefetch,
+                'padding_idx': self._padding_idx
+            })
+
+        return out

python/paddle/fluid/initializer.py

@@ -366,17 +366,40 @@ class TruncatedNormalInitializer(Initializer):
         # Initialization Ops should be prepended and not appended
         if self._seed == 0:
             self._seed = block.program.random_seed
+
+        # to be compatible of fp16 initalizers
+        if var.dtype == VarDesc.VarType.FP16:
+            out_dtype = VarDesc.VarType.FP32
+            out_var = block.create_var(
+                name=unique_name.generate(".".join(
+                    ['truncated_gaussian_random', 'tmp'])),
+                shape=var.shape,
+                dtype=out_dtype,
+                type=VarDesc.VarType.LOD_TENSOR,
+                persistable=False)
+        else:
+            out_dtype = var.dtype
+            out_var = var
+
         op = block._prepend_op(
             type="truncated_gaussian_random",
-            outputs={"Out": var},
+            outputs={"Out": out_var},
             attrs={
                 "shape": var.shape,
-                "dtype": int(var.dtype),
+                "dtype": out_dtype,
                 "mean": self._mean,
                 "std": self._std_dev,
                 "seed": self._seed
             },
             stop_gradient=True)
+
+        if var.dtype == VarDesc.VarType.FP16:
+            block.append_op(
+                type="cast",
+                inputs={"X": out_var},
+                outputs={"Out": var},
+                attrs={"in_dtype": out_var.dtype,
+                       "out_dtype": var.dtype})
         var.op = op
         return op
 

python/paddle/fluid/layer_helper.py

@@ -300,6 +300,17 @@ class LayerHelper(object):
             attr.name = unique_name.generate(".".join([self.name, suffix]))
 
         if default_initializer is None and attr.initializer is None:
+            if isinstance(dtype, core.VarDesc.VarType):
+                if dtype != core.VarDesc.VarType.FP32 and \
+                    dtype != core.VarDesc.VarType.FP64:
+                    raise TypeError(
+                        "Can not create parameter with default initializer when dtype is not float type. Set default_initializer to fit the parameter dtype!"
+                    )
+            else:
+                if not (dtype.startswith("float") or dtype == "double"):
+                    raise TypeError(
+                        "Can not create parameter with default initializer when dtype is not float type. Set default_initializer to fit the parameter dtype!"
+                    )
             if is_bias:
                 attr._set_default_bias_initializer()
             else:

python/paddle/fluid/layers/detection.py

@@ -49,6 +49,7 @@ __all__ = [
     'box_coder',
     'polygon_box_transform',
     'yolov3_loss',
+    'multiclass_nms',
 ]
 
 
@@ -262,8 +263,10 @@ def detection_output(loc,
             number is N + 1, N is the batch size. The i-th image has
             `LoD[i + 1] - LoD[i]` detected results, if it is 0, the i-th image
             has no detected results. If all images have not detected results,
-            all the elements in LoD are 0, and output tensor only contains one
+            LoD will be set to {1}, and output tensor only contains one
             value, which is -1.
+            (After version 1.3, when no boxes detected, the lod is changed
+             from {0} to {1}.)
 
     Examples:
         .. code-block:: python
@@ -343,19 +346,107 @@ def box_coder(prior_box,
               target_box,
               code_type="encode_center_size",
               box_normalized=True,
-              name=None):
+              name=None,
+              axis=0):
     """
-    ${comment}
+    **Box Coder Layer**
+
+    Encode/Decode the target bounding box with the priorbox information.
+    
+    The Encoding schema described below:
+
+    .. math::
+
+        ox = (tx - px) / pw / pxv
+
+        oy = (ty - py) / ph / pyv
+
+        ow = \log(\abs(tw / pw)) / pwv 
+
+        oh = \log(\abs(th / ph)) / phv 
+
+    The Decoding schema described below:
+    
+    .. math::
+  
+        ox = (pw * pxv * tx * + px) - tw / 2
+
+        oy = (ph * pyv * ty * + py) - th / 2
+
+        ow = \exp(pwv * tw) * pw + tw / 2
+
+        oh = \exp(phv * th) * ph + th / 2   
+
+    where `tx`, `ty`, `tw`, `th` denote the target box's center coordinates, 
+    width and height respectively. Similarly, `px`, `py`, `pw`, `ph` denote 
+    the priorbox's (anchor) center coordinates, width and height. `pxv`, 
+    `pyv`, `pwv`, `phv` denote the variance of the priorbox and `ox`, `oy`, 
+    `ow`, `oh` denote the encoded/decoded coordinates, width and height. 
+
+    During Box Decoding, two modes for broadcast are supported. Say target 
+    box has shape [N, M, 4], and the shape of prior box can be [N, 4] or 
+    [M, 4]. Then prior box will broadcast to target box along the 
+    assigned axis. 
 
     Args:
-        prior_box(${prior_box_type}): ${prior_box_comment}
-        prior_box_var(${prior_box_var_type}): ${prior_box_var_comment}
-        target_box(${target_box_type}): ${target_box_comment}
-        code_type(${code_type_type}): ${code_type_comment}
-        box_normalized(${box_normalized_type}): ${box_normalized_comment}
+        prior_box(Variable): Box list prior_box is a 2-D Tensor with shape 
+                             [M, 4] holds M boxes, each box is represented as
+                             [xmin, ymin, xmax, ymax], [xmin, ymin] is the 
+                             left top coordinate of the anchor box, if the 
+                             input is image feature map, they are close to 
+                             the origin of the coordinate system. [xmax, ymax]
+                             is the right bottom coordinate of the anchor box.       
+        prior_box_var(Variable|list): prior_box_var supports two types of input. 
+                              One is variable with shape [M, 4] holds M group.
+                              The other one is list consist of 4 elements 
+                              shared by all boxes. 
+        target_box(Variable): This input can be a 2-D LoDTensor with shape 
+                              [N, 4] when code_type is 'encode_center_size'. 
+                              This input also can be a 3-D Tensor with shape 
+                              [N, M, 4] when code_type is 'decode_center_size'. 
+                              Each box is represented as  
+                              [xmin, ymin, xmax, ymax]. This tensor can 
+                              contain LoD information to represent a batch 
+                              of inputs. 
+        code_type(string): The code type used with the target box. It can be
+                           encode_center_size or decode_center_size
+        box_normalized(int): Whether treat the priorbox as a noramlized box.
+                             Set true by default.
+        name(string): The name of box coder.
+        axis(int): Which axis in PriorBox to broadcast for box decode, 
+                   for example, if axis is 0 and TargetBox has shape
+                   [N, M, 4] and PriorBox has shape [M, 4], then PriorBox
+                   will broadcast to [N, M, 4] for decoding. It is only valid
+                   when code type is decode_center_size. Set 0 by default. 
 
     Returns:
-        output_box(${output_box_type}): ${output_box_comment}
+        output_box(Variable): When code_type is 'encode_center_size', the 
+                              output tensor of box_coder_op with shape 
+                              [N, M, 4] representing the result of N target 
+                              boxes encoded with M Prior boxes and variances. 
+                              When code_type is 'decode_center_size', 
+                              N represents the batch size and M represents 
+                              the number of deocded boxes.
+
+    Examples:
+ 
+        .. code-block:: python
+ 
+            prior_box = fluid.layers.data(name='prior_box', 
+                                          shape=[512, 4], 
+                                          dtype='float32',
+                                          append_batch_size=False)
+            target_box = fluid.layers.data(name='target_box',
+                                           shape=[512,81,4],
+                                           dtype='float32',
+                                           append_batch_size=False)
+            output = fluid.layers.box_coder(prior_box=prior_box,
+                                            prior_box_var=[0.1,0.1,0.2,0.2],
+                                            target_box=target_box,
+                                            code_type="decode_center_size",
+                                            box_normalized=False,
+                                            axis=1)
+
     """
     helper = LayerHelper("box_coder", **locals())
 
@@ -366,15 +457,22 @@ def box_coder(prior_box,
         output_box = helper.create_variable(
             name=name, dtype=prior_box.dtype, persistable=False)
 
+    inputs = {"PriorBox": prior_box, "TargetBox": target_box}
+    attrs = {
+        "code_type": code_type,
+        "box_normalized": box_normalized,
+        "axis": axis
+    }
+    if isinstance(prior_box_var, Variable):
+        inputs['PriorBoxVar'] = prior_box_var
+    elif isinstance(prior_box_var, list):
+        attrs['variance'] = prior_box_var
+    else:
+        raise TypeError("Input variance of box_coder must be Variable or lisz")
     helper.append_op(
         type="box_coder",
-        inputs={
-            "PriorBox": prior_box,
-            "PriorBoxVar": prior_box_var,
-            "TargetBox": target_box
-        },
-        attrs={"code_type": code_type,
-               "box_normalized": box_normalized},
+        inputs=inputs,
+        attrs=attrs,
         outputs={"OutputBox": output_box})
     return output_box
 
@@ -1960,3 +2058,119 @@ def generate_proposals(scores,
     rpn_roi_probs.stop_gradient = True
 
     return rpn_rois, rpn_roi_probs
+
+
+def multiclass_nms(bboxes,
+                   scores,
+                   score_threshold,
+                   nms_top_k,
+                   keep_top_k,
+                   nms_threshold=0.3,
+                   normalized=True,
+                   nms_eta=1.,
+                   background_label=0,
+                   name=None):
+    """
+    **Multiclass NMS**
+    
+    This operator is to do multi-class non maximum suppression (NMS) on
+    boxes and scores.
+
+    In the NMS step, this operator greedily selects a subset of detection bounding
+    boxes that have high scores larger than score_threshold, if providing this
+    threshold, then selects the largest nms_top_k confidences scores if nms_top_k
+    is larger than -1. Then this operator pruns away boxes that have high IOU
+    (intersection over union) overlap with already selected boxes by adaptive
+    threshold NMS based on parameters of nms_threshold and nms_eta.
+
+    Aftern NMS step, at most keep_top_k number of total bboxes are to be kept
+    per image if keep_top_k is larger than -1.
+
+    Args:
+        bboxes (Variable): Two types of bboxes are supported:
+                           1. (Tensor) A 3-D Tensor with shape
+                           [N, M, 4 or 8 16 24 32] represents the
+                           predicted locations of M bounding bboxes,
+                           N is the batch size. Each bounding box has four
+                           coordinate values and the layout is 
+                           [xmin, ymin, xmax, ymax], when box size equals to 4.
+                           2. (LoDTensor) A 3-D Tensor with shape [M, C, 4]
+                           M is the number of bounding boxes, C is the 
+                           class number   
+        scores (Variable): Two types of scores are supported:
+                           1. (Tensor) A 3-D Tensor with shape [N, C, M]
+                           represents the predicted confidence predictions.
+                           N is the batch size, C is the class number, M is 
+                           number of bounding boxes. For each category there 
+                           are total M scores which corresponding M bounding
+                           boxes. Please note, M is equal to the 2nd dimension
+                           of BBoxes.
+                           2. (LoDTensor) A 2-D LoDTensor with shape [M, C].
+                           M is the number of bbox, C is the class number.
+                           In this case, input BBoxes should be the second
+                           case with shape [M, C, 4].
+        background_label (int): The index of background label, the background 
+                                label will be ignored. If set to -1, then all
+                                categories will be considered. Default: 0
+        score_threshold (float): Threshold to filter out bounding boxes with
+                                 low confidence score. If not provided, 
+                                 consider all boxes.
+        nms_top_k (int): Maximum number of detections to be kept according to
+                         the confidences aftern the filtering detections based
+                         on score_threshold.
+        nms_threshold (float): The threshold to be used in NMS. Default: 0.3
+        nms_eta (float): The threshold to be used in NMS. Default: 1.0
+        keep_top_k (int): Number of total bboxes to be kept per image after NMS
+                          step. -1 means keeping all bboxes after NMS step.
+        normalized (bool): Whether detections are normalized. Default: True
+        name(str): Name of the multiclass nms op. Default: None.
+
+    Returns:
+        Out: A 2-D LoDTensor with shape [No, 6] represents the detections.
+             Each row has 6 values: [label, confidence, xmin, ymin, xmax, ymax]
+             or A 2-D LoDTensor with shape [No, 10] represents the detections.
+             Each row has 10 values: 
+             [label, confidence, x1, y1, x2, y2, x3, y3, x4, y4]. No is the 
+             total number of detections. If there is no detected boxes for all
+             images, lod will be set to {1} and Out only contains one value
+             which is -1.
+             (After version 1.3, when no boxes detected, the lod is changed 
+             from {0} to {1}) 
+
+    Examples:
+        .. code-block:: python
+
+            boxes = fluid.layers.data(name='bboxes', shape=[81, 4],
+                                      dtype='float32', lod_level=1)
+            scores = fluid.layers.data(name='scores', shape=[81],
+                                      dtype='float32', lod_level=1)
+            out = fluid.layers.multiclass_nms(bboxes=boxes,
+                                              scores=scores,
+                                              background_label=0,
+                                              score_threshold=0.5,
+                                              nms_top_k=400,
+                                              nms_threshold=0.3,
+                                              keep_top_k=200,
+                                              normalized=False)
+    """
+    helper = LayerHelper('multiclass_nms', **locals())
+
+    output = helper.create_variable_for_type_inference(dtype=bboxes.dtype)
+    helper.append_op(
+        type="multiclass_nms",
+        inputs={'BBoxes': bboxes,
+                'Scores': scores},
+        attrs={
+            'background_label': background_label,
+            'score_threshold': score_threshold,
+            'nms_top_k': nms_top_k,
+            'nms_threshold': nms_threshold,
+            'nms_eta': nms_eta,
+            'keep_top_k': keep_top_k,
+            'nms_eta': nms_eta,
+            'normalized': normalized
+        },
+        outputs={'Out': output})
+    output.stop_gradient = True
+
+    return output

python/paddle/fluid/layers/nn.py

@@ -932,7 +932,7 @@ def dynamic_gru(input,
             create ParamAttr as param_attr. If the Initializer of the param_attr
             is not set, the parameter is initialized with Xavier. Default: None.
         bias_attr (ParamAttr|bool|None): The parameter attribute for the bias
-            of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates
+            of GRU.Note that the bias with :math:`(1 \\times 3D)` concatenates
             the bias in the update gate, reset gate and candidate calculations.
             If it is set to False, no bias will be applied to the update gate,
             reset gate and candidate calculations. If it is set to None or one
@@ -1073,7 +1073,7 @@ def gru_unit(input,
             create ParamAttr as param_attr. If the Initializer of the param_attr
             is not set, the parameter is initialized with Xavier. Default: None.
         bias_attr (ParamAttr|bool|None): The parameter attribute for the bias
-            of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates
+            of GRU.Note that the bias with :math:`(1 \\times 3D)` concatenates
             the bias in the update gate, reset gate and candidate calculations.
             If it is set to False, no bias will be applied to the update gate,
             reset gate and candidate calculations. If it is set to None or one
@@ -3877,7 +3877,8 @@ def beam_search(pre_ids,
                 end_id,
                 level=0,
                 is_accumulated=True,
-                name=None):
+                name=None,
+                return_parent_idx=False):
     """
     Beam search is a classical algorithm for selecting candidate words in a
     machine translation task.
@@ -3933,10 +3934,16 @@ def beam_search(pre_ids,
              accumulated scores.
         name(str|None): A name for this layer(optional). If set None, the layer
                         will be named automatically.
+        return_parent_idx(bool): Whether to return an extra Tensor variable 
+                        preserving the selected_ids' parent indice in pre_ids
+                        in output, which can be used to gather cell states at
+                        the next time step.
 
     Returns:
-        Variable: The LodTensor pair containing the selected ids and the \
-            corresponding scores.
+        Variable: The LodTensor tuple containing the selected ids and the \
+            corresponding scores. If :attr:`return_parent_idx` is :attr:`True`, \
+            an extra Tensor variable preserving the selected_ids' parent indice \
+            is included.
 
     Examples:
         .. code-block:: python
@@ -3969,6 +3976,11 @@ def beam_search(pre_ids,
     selected_scores = helper.create_variable_for_type_inference(
         dtype=score_type)
     selected_ids = helper.create_variable_for_type_inference(dtype=id_type)
+    # parent_idx is a tensor used to gather cell states at the next time
+    # step. Though lod in selected_ids can also be used to gather by
+    # sequence_expand, it is not efficient.
+    # gather_op's index input only supports int32 dtype currently
+    parent_idx = helper.create_variable_for_type_inference(dtype="int32")
 
     helper.append_op(
         type='beam_search',
@@ -3976,6 +3988,7 @@ def beam_search(pre_ids,
         outputs={
             'selected_ids': selected_ids,
             'selected_scores': selected_scores,
+            'parent_idx': parent_idx
         },
         attrs={
             # TODO(ChunweiYan) to assure other value support
@@ -3984,8 +3997,10 @@ def beam_search(pre_ids,
             'end_id': end_id,
             'is_accumulated': is_accumulated,
         })
-
-    return selected_ids, selected_scores
+    if return_parent_idx:
+        return selected_ids, selected_scores, parent_idx
+    else:
+        return selected_ids, selected_scores
 
 
 def beam_search_decode(ids, scores, beam_size, end_id, name=None):
@@ -5403,7 +5418,7 @@ def transpose(x, perm, name=None):
     Examples:
         .. code-block:: python
 
-            # use append_batch_size=False to avoid prepending extra
+            # use append_batch_size=False to avoid prepending extra
             # batch size in shape
             x = fluid.layers.data(name='x', shape=[5, 10, 15],
                             dtype='float32', append_batch_size=False)
@@ -5920,7 +5935,7 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
                                 than :attr:`shape`.
         act (str): The non-linear activation to be applied to the reshaped tensor
                    variable.
-        inplace(bool): Must use :attr:`False` if :attr:`x` is used in multiple
+        inplace(bool): Must use :attr:`False` if :attr:`x` is used in multiple
                        operators. If this flag is set :attr:`True`, reuse input
                        :attr:`x` to reshape, which will change the shape of
                        tensor variable :attr:`x` and might cause errors when
@@ -6581,7 +6596,9 @@ def image_resize(input,
                  scale=None,
                  name=None,
                  resample='BILINEAR',
-                 actual_shape=None):
+                 actual_shape=None,
+                 align_corners=True,
+                 align_mode=1):
     """
     **Resize a Batch of Images**
 
@@ -6594,6 +6611,80 @@ def image_resize(input,
 
         'NEAREST' : Nearest neighbor interpolation
 
+    Nearest neighbor interpolation is to perform nearest neighbor interpolation
+    in both the 3rd dimention(in height direction) and the 4th dimention(in width 
+    direction) on input tensor.
+            
+    Bilinear interpolation is an extension of linear interpolation for 
+    interpolating functions of two variables (e.g. H-direction and 
+    W-direction in this op) on a rectilinear 2D grid. The key idea is 
+    to perform linear interpolation first in one direction, and then 
+    again in the other direction.
+
+    Align_corners and align_mode are optinal parameters,the calculation method 
+    of interpolation can be selected by them.
+
+    Example:
+
+      For scale:
+      
+        if align_corners = True && out_size > 1 :
+
+          scale_factor = (in_size-1.0)/(out_size-1.0)
+        
+        else:
+          
+          scale_factor = float(in_size/out_size)
+        
+      
+      Nearest neighbor interpolation:
+      
+      if:
+          align_corners = False
+
+          input : (N,C,H_in,W_in)
+          output: (N,C,H_out,W_out) where:
+
+          H_out = \left \lfloor {H_{in} * scale_{}factor}} \right \rfloor
+          W_out = \left \lfloor {W_{in} * scale_{}factor}} \right \rfloor
+
+      else:
+          align_corners = True
+
+          input : (N,C,H_in,W_in)
+          output: (N,C,H_out,W_out) where:
+
+          H_out = round(H_{in} * scale_{factor})
+          W_out = round(W_{in} * scale_{factor})
+
+      Bilinear interpolation:
+
+      if:
+          align_corners = False , align_mode = 0
+          
+          input : (N,C,H_in,W_in)
+          output: (N,C,H_out,W_out) where:
+          
+          H_out = (H_{in}+0.5) * scale_{factor} - 0.5
+          W_out = (W_{in}+0.5) * scale_{factor} - 0.5
+
+
+      else:
+       
+          input : (N,C,H_in,W_in)
+          output: (N,C,H_out,W_out) where:
+
+          H_out = H_{in} * scale_{factor}
+          W_out = W_{in} * scale_{factor}
+
+    For details of nearest neighbor interpolation, please refer to Wikipedia: 
+    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation.
+
+    For details of bilinear interpolation, please refer to Wikipedia: 
+    https://en.wikipedia.org/wiki/Bilinear_interpolation.
+
+
+
     Args:
         input (Variable): The input tensor of image resize layer,
                           This is a 4-D tensor of the shape
@@ -6623,6 +6714,13 @@ def image_resize(input,
                                 set, otherwise errors would be occured in graph
                                 constructing stage.
                                 Default: None
+        align_corners(bool) :  An optional bool, If True, the centers of the 4 corner pixels of the 
+                               input and output tensors are aligned, preserving the values at the 
+                               corner pixels.
+                               Default: True
+        align_mode(int)  :  An optional for bilinear interpolation. can be \'0\' 
+                            for src_idx = scale*(dst_indx+0.5)-0.5 , can be \'1\' for 
+                            src_idx = scale*dst_index .
 
     Returns:
         Variable: The output is a 4-D tensor of the shape
@@ -6635,6 +6733,8 @@ def image_resize(input,
                     or 'NEAREST' currently.
         ValueError: One of out_shape and scale must not be None.
         ValueError: out_shape length should be 2.
+        TypeError: align_corners shoule be a bool value
+        ValueError: align_mode can only be '0' or '1'
 
     Examples:
         .. code-block:: python
@@ -6650,6 +6750,12 @@ def image_resize(input,
             "The 'resample' of image_resize can only be 'BILINEAR' or 'NEAREST' currently."
         )
     resample_type = resample_methods[resample]
+
+    if not isinstance(align_corners, bool):
+        raise TypeError("Attr align_corners should be a bool value")
+    if align_mode != 0 and align_mode != 1:
+        raise ValueError("align_mode can only be 0 or 1")
+
     if out_shape is None and scale is None:
         raise ValueError("One of out_shape and scale must not be None.")
     helper = LayerHelper('{}_interp'.format(resample_type), **locals())
@@ -6689,9 +6795,13 @@ def image_resize(input,
         type='{}_interp'.format(resample_type),
         inputs=inputs,
         outputs={"Out": out},
-        attrs={"out_h": out_h,
-               "out_w": out_w,
-               "interp_method": resample_type})
+        attrs={
+            "out_h": out_h,
+            "out_w": out_w,
+            "interp_method": resample_type,
+            "align_corners": align_corners,
+            "align_mode": align_mode
+        })
     return out
 
 
@@ -6700,7 +6810,9 @@ def resize_bilinear(input,
                     out_shape=None,
                     scale=None,
                     name=None,
-                    actual_shape=None):
+                    actual_shape=None,
+                    align_corners=True,
+                    align_mode=1):
     """
     Resize input by performing bilinear interpolation based on given
     output shape which specified by actual_shape, out_shape and scale
@@ -6715,6 +6827,47 @@ def resize_bilinear(input,
     For details of bilinear interpolation, please refer to Wikipedia:
     https://en.wikipedia.org/wiki/Bilinear_interpolation
 
+    Align_corners and align_mode are optinal parameters,the calculation 
+    method of interpolation can be selected by them.
+
+
+    Align_corners and align_mode are optinal parameters,the calculation method 
+    of interpolation can be selected by them.
+
+    Example:
+
+      For scale:
+      
+        if align_corners = True && out_size > 1 :
+
+          scale_factor = (in_size-1.0)/(out_size-1.0)
+        
+        else:
+          
+          scale_factor = float(in_size/out_size)     
+
+    Bilinear interpolation:
+
+      if:
+          align_corners = False , align_mode = 0
+          
+          input : (N,C,H_in,W_in)
+          output: (N,C,H_out,W_out) where:
+          
+          H_out = (H_{in}+0.5) * scale_{factor} - 0.5
+          W_out = (W_{in}+0.5) * scale_{factor} - 0.5
+
+
+      else:
+
+          input : (N,C,H_in,W_in)
+          output: (N,C,H_out,W_out) where:
+
+          H_out = H_{in} * scale_{factor}
+          W_out = W_{in} * scale_{factor}
+
+
+
     Args:
         input(${x_type}): ${x_comment}.
 
@@ -6738,6 +6891,8 @@ def resize_bilinear(input,
                                 set, otherwise errors would be occured in graph
                                 constructing stage.
                                 Default: None
+        align_corners(bool): ${align_corners_comment}
+        align_mode(bool): ${align_mode_comment}
 
     Returns:
         ${out_comment}.
@@ -6748,7 +6903,8 @@ def resize_bilinear(input,
             out = fluid.layers.resize_bilinear(input, out_shape=[12, 12])
     """
 
-    return image_resize(input, out_shape, scale, name, 'BILINEAR', actual_shape)
+    return image_resize(input, out_shape, scale, name, 'BILINEAR', actual_shape,
+                        align_corners, align_mode)
 
 
 @templatedoc(op_type="nearest_interp")
@@ -6756,13 +6912,48 @@ def resize_nearest(input,
                    out_shape=None,
                    scale=None,
                    name=None,
-                   actual_shape=None):
+                   actual_shape=None,
+                   align_corners=True):
     """
     Resize input by performing nearest neighbor interpolation in both the
     3rd dimention(in height direction) and the 4th dimention(in width
     direction) based on given output shape which specified by actual_shape,
     out_shape and scale in priority order.
 
+    Example:
+
+      For scale:
+      
+        if align_corners = True && out_size > 1 :
+
+          scale_factor = (in_size-1.0)/(out_size-1.0)
+        
+        else:
+          
+          scale_factor = float(in_size/out_size)
+        
+      
+      Nearest neighbor interpolation:
+      
+      if:
+          align_corners = False
+
+          input : (N,C,H_in,W_in)
+          output: (N,C,H_out,W_out) where:
+
+          H_out = \left \lfloor {H_{in} * scale_{}factor}} \right \rfloor
+          W_out = \left \lfloor {W_{in} * scale_{}factor}} \right \rfloor
+
+      else:
+          align_corners = True
+
+          input : (N,C,H_in,W_in)
+          output: (N,C,H_out,W_out) where:
+
+          H_out = round(H_{in} * scale_{factor})
+          W_out = round(W_{in} * scale_{factor})
+
+
     For details of nearest neighbor interpolation, please refer to Wikipedia:
     https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation
 
@@ -6789,6 +6980,7 @@ def resize_nearest(input,
                                 set, otherwise errors would be occured in graph
                                 constructing stage.
                                 Default: None
+        align_corners(bool): ${align_corners_comment}
 
     Returns:
         ${out_comment}.
@@ -6799,7 +6991,8 @@ def resize_nearest(input,
             out = fluid.layers.resize_nearest(input, out_shape=[12, 12])
     """
 
-    return image_resize(input, out_shape, scale, name, 'NEAREST', actual_shape)
+    return image_resize(input, out_shape, scale, name, 'NEAREST', actual_shape,
+                        align_corners)
 
 
 def image_resize_short(input, out_short_len, resample='BILINEAR'):

python/paddle/fluid/layers/ops.py

@@ -135,7 +135,7 @@ def thresholded_relu(x, threshold=None):
         if val is not None:
             kwargs[name] = val
 
-    _thresholded_relu_(**kwargs)
+    return _thresholded_relu_(**kwargs)
 
 
 thresholded_relu.__doc__ = _thresholded_relu_.__doc__ + """

python/paddle/fluid/optimizer.py

@@ -387,7 +387,7 @@ class Optimizer(object):
 
             params_grads = []
             for param in parameters:
-                if param.stop_gradient:
+                if param.stop_gradient or not param.trainable:
                     continue
                 # create gradient variable
                 grad_var = Variable(

python/paddle/fluid/tests/test_detection.py

@@ -50,6 +50,19 @@ class TestDetection(unittest.TestCase):
             self.assertEqual(out.shape[-1], 6)
         print(str(program))
 
+    def test_box_coder_api(self):
+        program = Program()
+        with program_guard(program):
+            x = layers.data(name='x', shape=[4], dtype='float32')
+            y = layers.data(name='z', shape=[4], dtype='float32', lod_level=1)
+            bcoder = layers.box_coder(
+                prior_box=x,
+                prior_box_var=[0.1, 0.2, 0.1, 0.2],
+                target_box=y,
+                code_type='encode_center_size')
+            self.assertIsNotNone(bcoder)
+        print(str(program))
+
     def test_detection_api(self):
         program = Program()
         with program_guard(program):
@@ -469,5 +482,16 @@ class TestYoloDetection(unittest.TestCase):
             self.assertIsNotNone(loss)
 
 
+class TestMulticlassNMS(unittest.TestCase):
+    def test_multiclass_nms(self):
+        program = Program()
+        with program_guard(program):
+            bboxes = layers.data(
+                name='bboxes', shape=[-1, 10, 4], dtype='float32')
+            scores = layers.data(name='scores', shape=[-1, 10], dtype='float32')
+            output = layers.multiclass_nms(bboxes, scores, 0.3, 400, 200, 0.7)
+            self.assertIsNotNone(output)
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/CMakeLists.txt

 file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
 string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
 
-# The MKLDNN tests are skiped when the MKLDNN flag is OFF
-if(NOT WITH_MKLDNN)
-    foreach(src ${TEST_OPS})
-        if(${src} MATCHES ".*_mkldnn_op$")
-            list(REMOVE_ITEM TEST_OPS ${src})
-        endif()
-    endforeach()
-endif(NOT WITH_MKLDNN)
-
 if(NOT WITH_DISTRIBUTE)
     list(REMOVE_ITEM TEST_OPS test_recv_op)
     list(REMOVE_ITEM TEST_OPS test_dist_transpiler)
@@ -85,6 +76,7 @@ list(REMOVE_ITEM TEST_OPS test_image_classification_resnet)
 list(REMOVE_ITEM TEST_OPS test_bilinear_interp_op)
 list(REMOVE_ITEM TEST_OPS test_nearest_interp_op)
 list(REMOVE_ITEM TEST_OPS test_imperative_resnet)
+list(REMOVE_ITEM TEST_OPS test_imperative_optimizer)
 foreach(TEST_OP ${TEST_OPS})
     py_test_modules(${TEST_OP} MODULES ${TEST_OP})
 endforeach(TEST_OP)
@@ -94,6 +86,8 @@ py_test_modules(test_bilinear_interp_op MODULES test_bilinear_interp_op SERIAL)
 py_test_modules(test_nearest_interp_op MODULES test_nearest_interp_op SERIAL)
 py_test_modules(test_imperative_resnet MODULES test_imperative_resnet ENVS
   FLAGS_cudnn_deterministic=1)
+py_test_modules(test_imperative_optimizer MODULES test_imperative_optimizer ENVS
+  FLAGS_cudnn_deterministic=1)
 if(WITH_DISTRIBUTE)
     py_test_modules(test_dist_train MODULES test_dist_train SERIAL)
     set_tests_properties(test_listen_and_serv_op PROPERTIES TIMEOUT 20)
@@ -120,3 +114,7 @@ endif()
 if (WITH_NGRAPH)
     add_subdirectory(ngraph)
 endif()
+
+if (WITH_MKLDNN)
+    add_subdirectory(mkldnn)
+endif()

python/paddle/fluid/tests/unittests/mkldnn/CMakeLists.txt

+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+foreach(TEST_OP ${TEST_OPS})
+    py_test_modules(${TEST_OP} MODULES ${TEST_OP})
+endforeach(TEST_OP)

python/paddle/fluid/tests/unittests/mkldnn/__init__.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

python/paddle/fluid/tests/unittests/test_activation_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py

@@ -17,9 +17,9 @@ from __future__ import print_function
 import unittest
 import numpy as np
 import paddle.fluid.core as core
-from op_test import OpTest
+from paddle.fluid.tests.unittests.op_test import OpTest
 from scipy.special import expit
-from test_activation_op import TestRelu, TestTanh, TestSqrt, TestAbs
+from paddle.fluid.tests.unittests.test_activation_op import TestRelu, TestTanh, TestSqrt, TestAbs
 
 
 class TestMKLDNNReluDim2(TestRelu):

python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_batch_norm_mkldnn_op.py

@@ -19,9 +19,9 @@ import numpy as np
 import paddle.fluid.core as core
 from paddle.fluid.op import Operator
 import paddle.fluid as fluid
-from op_test import OpTest
+from paddle.fluid.tests.unittests.op_test import OpTest
 from paddle.fluid.framework import grad_var_name
-from test_batch_norm_op import TestBatchNormOpInference, TestBatchNormOpTraining, _reference_training, _reference_grad
+from paddle.fluid.tests.unittests.test_batch_norm_op import TestBatchNormOpInference, TestBatchNormOpTraining, _reference_training, _reference_grad
 
 
 class TestMKLDNNBatchNormOpTraining(TestBatchNormOpTraining):

python/paddle/fluid/tests/unittests/test_concat_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_concat_mkldnn_op.py

@@ -15,7 +15,7 @@
 from __future__ import print_function
 
 import unittest
-from test_concat_op import TestConcatOp, TestConcatOp2, TestConcatOp3
+from paddle.fluid.tests.unittests.test_concat_op import TestConcatOp, TestConcatOp2, TestConcatOp3
 
 
 class TestMKLDNNConcatOp(TestConcatOp):

python/paddle/fluid/tests/unittests/test_conv2d_int8_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_int8_mkldnn_op.py

@@ -18,8 +18,8 @@ import unittest
 import numpy as np
 
 import paddle.fluid.core as core
-from op_test import OpTest
-from test_conv2d_op import conv2d_forward_naive, TestConv2dOp
+from paddle.fluid.tests.unittests.op_test import OpTest
+from paddle.fluid.tests.unittests.test_conv2d_op import conv2d_forward_naive, TestConv2dOp
 
 
 def conv2d_forward_refer(input, filter, group, conv_param):

python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 
-from test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride, TestWithGroup, TestWith1x1, TestWithInput1x1Filter1x1
+from paddle.fluid.tests.unittests.test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride, TestWithGroup, TestWith1x1, TestWithInput1x1Filter1x1
 
 
 class TestMKLDNN(TestConv2dOp):

python/paddle/fluid/tests/unittests/test_conv2d_transpose_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_conv2d_transpose_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 
-from test_conv2d_transpose_op import TestConv2dTransposeOp, TestWithPad, TestWithStride
+from paddle.fluid.tests.unittests.test_conv2d_transpose_op import TestConv2dTransposeOp, TestWithPad, TestWithStride
 
 
 class TestMKLDNN(TestConv2dTransposeOp):

python/paddle/fluid/tests/unittests/test_conv3d_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_conv3d_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 
-from test_conv3d_op import TestConv3dOp, TestCase1, TestWithGroup1, TestWithGroup2, TestWith1x1, TestWithInput1x1Filter1x1
+from paddle.fluid.tests.unittests.test_conv3d_op import TestConv3dOp, TestCase1, TestWithGroup1, TestWithGroup2, TestWith1x1, TestWithInput1x1Filter1x1
 
 
 class TestMKLDNN(TestConv3dOp):

python/paddle/fluid/tests/unittests/test_dequantize_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_dequantize_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 import numpy as np
-from op_test import OpTest
+from paddle.fluid.tests.unittests.op_test import OpTest
 
 
 class TestDeQuantizeOp(OpTest):

python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_add_mkldnn_op.py

@@ -16,8 +16,8 @@ from __future__ import print_function
 import unittest
 import numpy as np
 import paddle.fluid.core as core
-from op_test import OpTest
-from test_elementwise_add_op import *
+from paddle.fluid.tests.unittests.op_test import OpTest
+from paddle.fluid.tests.unittests.test_elementwise_add_op import *
 '''
 Some tests differ from the tests defined in test_elementwise_add_op.py
 because MKLDNN does not support tensors of number of dimensions 3.

python/paddle/fluid/tests/unittests/test_elementwise_mul_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_elementwise_mul_mkldnn_op.py

@@ -15,10 +15,10 @@
 from __future__ import print_function
 import unittest
 import numpy as np
-from op_test import OpTest
+from paddle.fluid.tests.unittests.op_test import OpTest
 import paddle.fluid.core as core
 from paddle.fluid.op import Operator
-from test_elementwise_mul_op import *
+from paddle.fluid.tests.unittests.test_elementwise_mul_op import *
 
 
 class TestElementwiseMulMKLDNNOp_BroadcastNCHW16c(ElementwiseMulOp):

python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_fc_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 import numpy as np
-from op_test import OpTest
+from paddle.fluid.tests.unittests.op_test import OpTest
 
 
 def fully_connected_naive(input, weights, bias_data=None):

python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_gaussian_random_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 
-from test_gaussian_random_op import TestGaussianRandomOp
+from paddle.fluid.tests.unittests.test_gaussian_random_op import TestGaussianRandomOp
 
 
 class TestMKLDNN(TestGaussianRandomOp):

python/paddle/fluid/tests/unittests/test_lrn_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_lrn_mkldnn_op.py

@@ -15,7 +15,7 @@
 from __future__ import print_function
 
 import unittest
-from test_lrn_op import TestLRNOp
+from paddle.fluid.tests.unittests.test_lrn_op import TestLRNOp
 
 
 class TestLRNMKLDNNOp(TestLRNOp):

python/paddle/fluid/tests/unittests/test_pool2d_int8_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_pool2d_int8_mkldnn_op.py

@@ -19,8 +19,8 @@ import unittest
 import numpy as np
 
 import paddle.fluid.core as core
-from op_test import OpTest
-from test_pool2d_op import TestPool2D_Op, avg_pool2D_forward_naive, max_pool2D_forward_naive
+from paddle.fluid.tests.unittests.op_test import OpTest
+from paddle.fluid.tests.unittests.test_pool2d_op import TestPool2D_Op, avg_pool2D_forward_naive, max_pool2D_forward_naive
 
 
 class TestPool2dMKLDNNInt8_Op(TestPool2D_Op):

python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_pool2d_mkldnn_op.py

@@ -15,7 +15,7 @@
 from __future__ import print_function
 
 import unittest
-from test_pool2d_op import TestPool2D_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5
+from paddle.fluid.tests.unittests.test_pool2d_op import TestPool2D_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5
 
 
 def create_test_mkldnn_class(parent):

python/paddle/fluid/tests/unittests/test_quantize_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_quantize_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 import numpy as np
-from op_test import OpTest
+from paddle.fluid.tests.unittests.op_test import OpTest
 
 
 class TestQuantizeOp(OpTest):

python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_sum_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 
-from test_sum_op import TestSumOp
+from paddle.fluid.tests.unittests.test_sum_op import TestSumOp
 
 
 class TestMKLDNN(TestSumOp):

python/paddle/fluid/tests/unittests/test_transpose_mkldnn_op.py → python/paddle/fluid/tests/unittests/mkldnn/test_transpose_mkldnn_op.py

@@ -16,7 +16,7 @@ from __future__ import print_function
 
 import unittest
 
-from test_transpose_op import TestTransposeOp
+from paddle.fluid.tests.unittests.test_transpose_op import TestTransposeOp
 
 
 class TestTransposeMKLDNN(TestTransposeOp):

python/paddle/fluid/tests/unittests/ngraph/test_pool2d_ngraph_op.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+from paddle.fluid.tests.unittests.test_pool2d_op import *
+
+
+class TestNGRAPHPool2D_Op(TestPool2D_Op):
+    def init_test_case(self):
+        super(TestNGRAPHPool2D_Op, self).init_test_case()
+
+
+class TestNGRAPHCase1(TestCase1):
+    def init_test_case(self):
+        super(TestNGRAPHCase1, self).init_test_case()
+
+
+class TestNGRAPHCase2(TestCase2):
+    def init_test_case(self):
+        super(TestNGRAPHCase2, self).init_test_case()
+
+
+class TestNGRAPHCase3(TestCase3):
+    def init_pool_type(self):
+        super(TestNGRAPHCase3, self).init_pool_type()
+
+
+class TestNGRAPHCase4(TestCase4):
+    def init_pool_type(self):
+        super(TestNGRAPHCase4, self).init_pool_type()
+
+
+class TestNGRAPHCase5(TestCase5):
+    def init_pool_type(self):
+        super(TestNGRAPHCase5, self).init_pool_type()
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_beam_search_op.py

@@ -38,6 +38,7 @@ class BeamSearchOpTester(unittest.TestCase):
         self._create_pre_ids()
         self.scope.var('selected_ids')
         self.scope.var('selected_scores')
+        self.scope.var('parent_idx')
 
     def test_run(self):
         op = Operator(
@@ -48,12 +49,14 @@ class BeamSearchOpTester(unittest.TestCase):
             scores='scores',
             selected_ids='selected_ids',
             selected_scores='selected_scores',
+            parent_idx='parent_idx',
             level=0,
             beam_size=2,
             end_id=0, )
         op.run(self.scope, core.CPUPlace())
         selected_ids = self.scope.find_var("selected_ids").get_tensor()
         selected_scores = self.scope.find_var("selected_scores").get_tensor()
+        parent_idx = self.scope.find_var("parent_idx").get_tensor()
         self.assertTrue(
             np.allclose(
                 np.array(selected_ids), np.array([4, 2, 3, 8])[:, np.newaxis]))
@@ -62,6 +65,8 @@ class BeamSearchOpTester(unittest.TestCase):
                 np.array(selected_scores),
                 np.array([0.5, 0.6, 0.9, 0.7])[:, np.newaxis]))
         self.assertEqual(selected_ids.lod(), [[0, 2, 4], [0, 1, 2, 3, 4]])
+        self.assertTrue(
+            np.allclose(np.array(parent_idx), np.array([0, 1, 2, 3])))
 
     def _create_pre_ids(self):
         np_data = np.array([[1, 2, 3, 4]], dtype='int64')

python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py

@@ -20,7 +20,13 @@ from op_test import OpTest
 import paddle.fluid.core as core
 
 
-def bilinear_interp_np(input, out_h, out_w, out_size=None, actual_shape=None):
+def bilinear_interp_np(input,
+                       out_h,
+                       out_w,
+                       out_size=None,
+                       actual_shape=None,
+                       align_corners=True,
+                       align_mode=0):
     """bilinear interpolation implement in shape [N, C, H, W]"""
     if out_size is not None:
         out_h = out_size[0]
@@ -29,25 +35,45 @@ def bilinear_interp_np(input, out_h, out_w, out_size=None, actual_shape=None):
         out_h = actual_shape[0]
         out_w = actual_shape[1]
     batch_size, channel, in_h, in_w = input.shape
+
+    ratio_h = ratio_w = 0.0
     if out_h > 1:
-        ratio_h = (in_h - 1.0) / (out_h - 1.0)
-    else:
-        ratio_h = 0.0
+        if (align_corners):
+            ratio_h = (in_h - 1.0) / (out_h - 1.0)
+        else:
+            ratio_h = 1.0 * in_h / out_h
     if out_w > 1:
-        ratio_w = (in_w - 1.0) / (out_w - 1.0)
-    else:
-        ratio_w = 0.0
+        if (align_corners):
+            ratio_w = (in_w - 1.0) / (out_w - 1.0)
+        else:
+            ratio_w = 1.0 * in_w / out_w
 
     out = np.zeros((batch_size, channel, out_h, out_w))
+
     for i in range(out_h):
-        h = int(ratio_h * i)
+        if (align_mode == 0 and not align_corners):
+            h = int(ratio_h * (i + 0.5) - 0.5)
+        else:
+            h = int(ratio_h * i)
+
+        h = max(0, h)
         hid = 1 if h < in_h - 1 else 0
-        h1lambda = ratio_h * i - h
+        if (align_mode == 0 and not align_corners):
+            h1lambda = ratio_h * (i + 0.5) - 0.5 - h
+        else:
+            h1lambda = ratio_h * i - h
         h2lambda = 1.0 - h1lambda
         for j in range(out_w):
-            w = int(ratio_w * j)
+            if (align_mode == 0 and not align_corners):
+                w = int(ratio_w * (j + 0.5) - 0.5)
+            else:
+                w = int(ratio_w * j)
+            w = max(0, w)
             wid = 1 if w < in_w - 1 else 0
-            w1lambda = ratio_w * j - w
+            if (align_mode == 0 and not align_corners):
+                w1lambda = ratio_w * (j + 0.5) - 0.5 - w
+            else:
+                w1lambda = ratio_w * j - w
             w2lambda = 1.0 - w1lambda
 
             out[:, :, i, j] = h2lambda*(w2lambda*input[:, :, h, w] +
@@ -66,7 +92,8 @@ class TestBilinearInterpOp(OpTest):
         input_np = np.random.random(self.input_shape).astype("float32")
 
         output_np = bilinear_interp_np(input_np, self.out_h, self.out_w,
-                                       self.out_size, self.actual_shape)
+                                       self.out_size, self.actual_shape,
+                                       self.align_corners, self.align_mode)
         self.inputs = {'X': input_np}
         if self.out_size is not None:
             self.inputs['OutSize'] = self.out_size
@@ -75,7 +102,9 @@ class TestBilinearInterpOp(OpTest):
         self.attrs = {
             'out_h': self.out_h,
             'out_w': self.out_w,
-            'interp_method': self.interp_method
+            'interp_method': self.interp_method,
+            'align_corners': self.align_corners,
+            'align_mode': self.align_mode
         }
         self.outputs = {'Out': output_np}
 
@@ -91,6 +120,8 @@ class TestBilinearInterpOp(OpTest):
         self.out_h = 2
         self.out_w = 2
         self.out_size = np.array([3, 3]).astype("int32")
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpCase1(TestBilinearInterpOp):
@@ -99,6 +130,8 @@ class TestBilinearInterpCase1(TestBilinearInterpOp):
         self.input_shape = [4, 1, 7, 8]
         self.out_h = 1
         self.out_w = 1
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpCase2(TestBilinearInterpOp):
@@ -107,6 +140,8 @@ class TestBilinearInterpCase2(TestBilinearInterpOp):
         self.input_shape = [3, 3, 9, 6]
         self.out_h = 12
         self.out_w = 12
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpCase3(TestBilinearInterpOp):
@@ -115,6 +150,8 @@ class TestBilinearInterpCase3(TestBilinearInterpOp):
         self.input_shape = [1, 1, 128, 64]
         self.out_h = 64
         self.out_w = 128
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpCase4(TestBilinearInterpOp):
@@ -124,6 +161,8 @@ class TestBilinearInterpCase4(TestBilinearInterpOp):
         self.out_h = 1
         self.out_w = 1
         self.out_size = np.array([2, 2]).astype("int32")
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpCase5(TestBilinearInterpOp):
@@ -133,6 +172,8 @@ class TestBilinearInterpCase5(TestBilinearInterpOp):
         self.out_h = 12
         self.out_w = 12
         self.out_size = np.array([11, 11]).astype("int32")
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpCase6(TestBilinearInterpOp):
@@ -142,6 +183,8 @@ class TestBilinearInterpCase6(TestBilinearInterpOp):
         self.out_h = 64
         self.out_w = 128
         self.out_size = np.array([65, 129]).astype("int32")
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpActualShape(TestBilinearInterpOp):
@@ -151,6 +194,8 @@ class TestBilinearInterpActualShape(TestBilinearInterpOp):
         self.out_h = 64
         self.out_w = 32
         self.out_size = np.array([66, 40]).astype("int32")
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpOpUint8(OpTest):
@@ -162,14 +207,17 @@ class TestBilinearInterpOpUint8(OpTest):
         input_np = np.random.randint(
             low=0, high=256, size=self.input_shape).astype("uint8")
         output_np = bilinear_interp_np(input_np, self.out_h, self.out_w,
-                                       self.out_size, self.actual_shape)
+                                       self.out_size, self.actual_shape,
+                                       self.align_corners, self.align_mode)
         self.inputs = {'X': input_np}
         if self.out_size is not None:
             self.inputs['OutSize'] = self.out_size
         self.attrs = {
             'out_h': self.out_h,
             'out_w': self.out_w,
-            'interp_method': self.interp_method
+            'interp_method': self.interp_method,
+            'align_corners': self.align_corners,
+            'align_mode': self.align_mode
         }
         self.outputs = {'Out': output_np}
 
@@ -181,6 +229,8 @@ class TestBilinearInterpOpUint8(OpTest):
         self.input_shape = [1, 3, 9, 6]
         self.out_h = 10
         self.out_w = 9
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpCase1Uint8(TestBilinearInterpOpUint8):
@@ -189,6 +239,8 @@ class TestBilinearInterpCase1Uint8(TestBilinearInterpOpUint8):
         self.input_shape = [2, 3, 128, 64]
         self.out_h = 120
         self.out_w = 50
+        self.align_corners = True
+        self.align_mode = 1
 
 
 class TestBilinearInterpCase2Uint8(TestBilinearInterpOpUint8):
@@ -198,6 +250,26 @@ class TestBilinearInterpCase2Uint8(TestBilinearInterpOpUint8):
         self.out_h = 5
         self.out_w = 13
         self.out_size = np.array([6, 15]).astype("int32")
+        self.align_corners = True
+        self.align_mode = 1
+
+
+class TestBilinearInterpOtherMethod1(TestBilinearInterpOp):
+    def set_align_mode(self):
+        self.align_corners = False
+        self.align_mode = 1
+
+
+class TestBilinearInterpWithMethod2(TestBilinearInterpOp):
+    def set_align_mode(self):
+        self.align_corners = False
+        self.align_mode = 0
+
+
+class TestBilinearInterpWithMethod3(TestBilinearInterpOp):
+    def set_align_mode(self):
+        self.align_corners = True
+        self.align_mode = 0
 
 
 if __name__ == "__main__":

python/paddle/fluid/tests/unittests/test_box_coder_op.py

@@ -21,80 +21,80 @@ import math
 from op_test import OpTest
 
 
-def box_coder(target_box, prior_box, prior_box_var, output_box, code_type,
-              box_normalized):
-    prior_box_x = (
-        (prior_box[:, 2] + prior_box[:, 0]) / 2).reshape(1, prior_box.shape[0])
-    prior_box_y = (
-        (prior_box[:, 3] + prior_box[:, 1]) / 2).reshape(1, prior_box.shape[0])
-    prior_box_width = (
-        (prior_box[:, 2] - prior_box[:, 0])).reshape(1, prior_box.shape[0])
-    prior_box_height = (
-        (prior_box[:, 3] - prior_box[:, 1])).reshape(1, prior_box.shape[0])
-    prior_box_var = prior_box_var.reshape(1, prior_box_var.shape[0],
-                                          prior_box_var.shape[1])
-    if not box_normalized:
-        prior_box_height = prior_box_height + 1
-        prior_box_width = prior_box_width + 1
-
-    if (code_type == "EncodeCenterSize"):
-        target_box_x = ((target_box[:, 2] + target_box[:, 0]) / 2).reshape(
-            target_box.shape[0], 1)
-        target_box_y = ((target_box[:, 3] + target_box[:, 1]) / 2).reshape(
-            target_box.shape[0], 1)
-        target_box_width = ((target_box[:, 2] - target_box[:, 0])).reshape(
-            target_box.shape[0], 1)
-        target_box_height = ((target_box[:, 3] - target_box[:, 1])).reshape(
-            target_box.shape[0], 1)
-        if not box_normalized:
-            target_box_height = target_box_height + 1
-            target_box_width = target_box_width + 1
-
-        output_box[:,:,0] = (target_box_x - prior_box_x) / prior_box_width / \
-                prior_box_var[:,:,0]
-        output_box[:,:,1] = (target_box_y - prior_box_y) / prior_box_height / \
-                prior_box_var[:,:,1]
-        output_box[:,:,2] = np.log(np.fabs(target_box_width / prior_box_width)) / \
-                prior_box_var[:,:,2]
-        output_box[:,:,3] = np.log(np.fabs(target_box_height / prior_box_height)) / \
-                prior_box_var[:,:,3]
-
-    elif (code_type == "DecodeCenterSize"):
-        target_box_x = prior_box_var[:,:,0] * target_box[:,:,0] * \
-                       prior_box_width + prior_box_x
-        target_box_y = prior_box_var[:,:,1] * target_box[:,:,1] * \
-                       prior_box_height + prior_box_y
-        target_box_width = np.exp(prior_box_var[:,:,2] * target_box[:,:,2]) * \
-                           prior_box_width
-        target_box_height = np.exp(prior_box_var[:,:,3] * target_box[:,:,3]) * \
-                            prior_box_height
-
-        output_box[:, :, 0] = target_box_x - target_box_width / 2
-        output_box[:, :, 1] = target_box_y - target_box_height / 2
-        output_box[:, :, 2] = target_box_x + target_box_width / 2
-        output_box[:, :, 3] = target_box_y + target_box_height / 2
-        if not box_normalized:
-            output_box[:, :, 2] = output_box[:, :, 2] - 1
-            output_box[:, :, 3] = output_box[:, :, 3] - 1
-
-
-def batch_box_coder(prior_box, prior_box_var, target_box, lod, code_type,
-                    box_normalized):
-    n = target_box.shape[0]
-    m = prior_box.shape[0]
+def box_decoder(t_box, p_box, pb_v, output_box, norm, axis=0):
+    pb_w = p_box[:, 2] - p_box[:, 0] + (norm == False)
+    pb_h = p_box[:, 3] - p_box[:, 1] + (norm == False)
+    pb_x = pb_w * 0.5 + p_box[:, 0]
+    pb_y = pb_h * 0.5 + p_box[:, 1]
+    shape = (1, p_box.shape[0]) if axis == 0 else (p_box.shape[0], 1)
+
+    pb_w = pb_w.reshape(shape)
+    pb_h = pb_h.reshape(shape)
+    pb_x = pb_x.reshape(shape)
+    pb_y = pb_y.reshape(shape)
+
+    if pb_v.ndim == 2:
+        pb_v = pb_v.reshape(1, pb_v.shape[0], pb_v.shape[1])
+    if pb_v.ndim == 1:
+        tb_x = pb_v[0] * t_box[:, :, 0] * pb_w + pb_x
+        tb_y = pb_v[1] * t_box[:, :, 1] * pb_h + pb_y
+        tb_w = np.exp(pb_v[2] * t_box[:, :, 2]) * pb_w
+        tb_h = np.exp(pb_v[3] * t_box[:, :, 3]) * pb_h
+    else:
+        tb_x = pb_v[:, :, 0] * t_box[:, :, 0] * pb_w + pb_x
+        tb_y = pb_v[:, :, 1] * t_box[:, :, 1] * pb_h + pb_y
+        tb_w = np.exp(pb_v[:, :, 2] * t_box[:, :, 2]) * pb_w
+        tb_h = np.exp(pb_v[:, :, 3] * t_box[:, :, 3]) * pb_h
+    output_box[:, :, 0] = tb_x - tb_w / 2
+    output_box[:, :, 1] = tb_y - tb_h / 2
+    output_box[:, :, 2] = tb_x + tb_w / 2 - (not norm)
+    output_box[:, :, 3] = tb_y + tb_h / 2 - (not norm)
+
+
+def box_encoder(t_box, p_box, pb_v, output_box, norm):
+    pb_w = p_box[:, 2] - p_box[:, 0] + (norm == False)
+    pb_h = p_box[:, 3] - p_box[:, 1] + (norm == False)
+    pb_x = pb_w * 0.5 + p_box[:, 0]
+    pb_y = pb_h * 0.5 + p_box[:, 1]
+    shape = (1, p_box.shape[0])
+
+    pb_w = pb_w.reshape(shape)
+    pb_h = pb_h.reshape(shape)
+    pb_x = pb_x.reshape(shape)
+    pb_y = pb_y.reshape(shape)
+
+    if pb_v.ndim == 2:
+        pb_v = pb_v.reshape(1, pb_v.shape[0], pb_v.shape[1])
+    tb_x = ((t_box[:, 2] + t_box[:, 0]) / 2).reshape(t_box.shape[0], 1)
+    tb_y = ((t_box[:, 3] + t_box[:, 1]) / 2).reshape(t_box.shape[0], 1)
+    tb_w = (t_box[:, 2] - t_box[:, 0]).reshape(t_box.shape[0], 1) + (not norm)
+    tb_h = (t_box[:, 3] - t_box[:, 1]).reshape(t_box.shape[0], 1) + (not norm)
+    if pb_v.ndim == 1:
+        output_box[:, :, 0] = (tb_x - pb_x) / pb_w / pb_v[0]
+        output_box[:, :, 1] = (tb_y - pb_y) / pb_h / pb_v[1]
+        output_box[:, :, 2] = np.log(np.fabs(tb_w / pb_w)) / pb_v[2]
+        output_box[:, :, 3] = np.log(np.fabs(tb_h / pb_h)) / pb_v[3]
+    else:
+        output_box[:, :, 0] = (tb_x - pb_x) / pb_w / pb_v[:, :, 0]
+        output_box[:, :, 1] = (tb_y - pb_y) / pb_h / pb_v[:, :, 1]
+        output_box[:, :, 2] = np.log(np.fabs(tb_w / pb_w)) / pb_v[:, :, 2]
+        output_box[:, :, 3] = np.log(np.fabs(tb_h / pb_h)) / pb_v[:, :, 3]
+
+
+def batch_box_coder(p_box, pb_v, t_box, lod, code_type, norm, axis=0):
+    n = t_box.shape[0]
+    m = p_box.shape[0]
+    if code_type == "DecodeCenterSize":
+        m = t_box.shape[1]
     output_box = np.zeros((n, m, 4), dtype=np.float32)
     cur_offset = 0
     for i in range(len(lod)):
         if (code_type == "EncodeCenterSize"):
-            box_coder(target_box[cur_offset:(cur_offset + lod[i]), :],
-                      prior_box, prior_box_var,
-                      output_box[cur_offset:(cur_offset + lod[i]), :, :],
-                      code_type, box_normalized)
+            box_encoder(t_box[cur_offset:(cur_offset + lod[i]), :], p_box, pb_v,
+                        output_box[cur_offset:(cur_offset + lod[i]), :, :],
+                        norm)
         elif (code_type == "DecodeCenterSize"):
-            box_coder(target_box[cur_offset:(cur_offset + lod[i]), :, :],
-                      prior_box, prior_box_var,
-                      output_box[cur_offset:(cur_offset + lod[i]), :, :],
-                      code_type, box_normalized)
+            box_decoder(t_box, p_box, pb_v, output_box, norm, axis)
         cur_offset += lod[i]
     return output_box
 
@@ -106,9 +106,35 @@ class TestBoxCoderOp(OpTest):
     def setUp(self):
         self.op_type = "box_coder"
         lod = [[1, 1, 1, 1, 1]]
-        prior_box = np.random.random((10, 4)).astype('float32')
-        prior_box_var = np.random.random((10, 4)).astype('float32')
-        target_box = np.random.random((5, 10, 4)).astype('float32')
+        prior_box = np.random.random((81, 4)).astype('float32')
+        prior_box_var = np.random.random((81, 4)).astype('float32')
+        target_box = np.random.random((20, 81, 4)).astype('float32')
+        code_type = "DecodeCenterSize"
+        box_normalized = False
+        output_box = batch_box_coder(prior_box, prior_box_var, target_box,
+                                     lod[0], code_type, box_normalized)
+        self.inputs = {
+            'PriorBox': prior_box,
+            'PriorBoxVar': prior_box_var,
+            'TargetBox': target_box,
+        }
+        self.attrs = {
+            'code_type': 'decode_center_size',
+            'box_normalized': False
+        }
+        self.outputs = {'OutputBox': output_box}
+
+
+class TestBoxCoderOpWithOneRankVar(OpTest):
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        self.op_type = "box_coder"
+        lod = [[1, 1, 1, 1, 1]]
+        prior_box = np.random.random((81, 4)).astype('float32')
+        prior_box_var = np.random.random((4)).astype('float32')
+        target_box = np.random.random((20, 81, 4)).astype('float32')
         code_type = "DecodeCenterSize"
         box_normalized = False
         output_box = batch_box_coder(prior_box, prior_box_var, target_box,
@@ -133,9 +159,9 @@ class TestBoxCoderOpWithoutBoxVar(OpTest):
     def setUp(self):
         self.op_type = "box_coder"
         lod = [[0, 1, 2, 3, 4, 5]]
-        prior_box = np.random.random((10, 4)).astype('float32')
-        prior_box_var = np.ones((10, 4)).astype('float32')
-        target_box = np.random.random((5, 10, 4)).astype('float32')
+        prior_box = np.random.random((81, 4)).astype('float32')
+        prior_box_var = np.ones((81, 4)).astype('float32')
+        target_box = np.random.random((20, 81, 4)).astype('float32')
         code_type = "DecodeCenterSize"
         box_normalized = False
         output_box = batch_box_coder(prior_box, prior_box_var, target_box,
@@ -158,10 +184,10 @@ class TestBoxCoderOpWithLoD(OpTest):
 
     def setUp(self):
         self.op_type = "box_coder"
-        lod = [[4, 8, 8]]
-        prior_box = np.random.random((10, 4)).astype('float32')
-        prior_box_var = np.random.random((10, 4)).astype('float32')
-        target_box = np.random.random((20, 4)).astype('float32')
+        lod = [[10, 20, 20]]
+        prior_box = np.random.random((20, 4)).astype('float32')
+        prior_box_var = np.random.random((20, 4)).astype('float32')
+        target_box = np.random.random((50, 4)).astype('float32')
         code_type = "EncodeCenterSize"
         box_normalized = True
         output_box = batch_box_coder(prior_box, prior_box_var, target_box,
@@ -176,5 +202,63 @@ class TestBoxCoderOpWithLoD(OpTest):
         self.outputs = {'OutputBox': output_box}
 
 
+class TestBoxCoderOpWithAxis(OpTest):
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        self.op_type = "box_coder"
+        lod = [[1, 1, 1, 1, 1]]
+        prior_box = np.random.random((30, 4)).astype('float32')
+        prior_box_var = np.random.random((4)).astype('float32')
+        target_box = np.random.random((30, 81, 4)).astype('float32')
+        code_type = "DecodeCenterSize"
+        box_normalized = False
+        axis = 1
+        output_box = batch_box_coder(prior_box, prior_box_var, target_box,
+                                     lod[0], code_type, box_normalized, axis)
+
+        self.inputs = {
+            'PriorBox': prior_box,
+            'PriorBoxVar': prior_box_var,
+            'TargetBox': target_box,
+        }
+        self.attrs = {
+            'code_type': 'decode_center_size',
+            'box_normalized': False,
+            'axis': axis
+        }
+        self.outputs = {'OutputBox': output_box}
+
+
+class TestBoxCoderOpWithVariance(OpTest):
+    def test_check_output(self):
+        self.check_output()
+
+    def setUp(self):
+        self.op_type = "box_coder"
+        lod = [[1, 1, 1, 1, 1]]
+        prior_box = np.random.random((30, 4)).astype('float32')
+        prior_box_var = np.random.random((4)).astype('float32')
+        target_box = np.random.random((30, 81, 4)).astype('float32')
+        code_type = "DecodeCenterSize"
+        box_normalized = False
+        axis = 1
+        output_box = batch_box_coder(prior_box, prior_box_var, target_box,
+                                     lod[0], code_type, box_normalized, axis)
+
+        self.inputs = {
+            'PriorBox': prior_box,
+            'TargetBox': target_box,
+        }
+        self.attrs = {
+            'code_type': 'decode_center_size',
+            'box_normalized': False,
+            'variance': prior_box_var.astype(np.float).flatten(),
+            'axis': axis
+        }
+        self.outputs = {'OutputBox': output_box}
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_imperative.py

@@ -66,7 +66,141 @@ class MLP(fluid.imperative.Layer):
         return x
 
 
+class SimpleRNNCell(fluid.imperative.Layer):
+    def __init__(self, step_input_size, hidden_size, output_size, param_attr):
+        super(SimpleRNNCell, self).__init__()
+        self.step_input_size = step_input_size
+        self.hidden_size = hidden_size
+        self.output_size = output_size
+        self._dype = core.VarDesc.VarType.FP32
+        from paddle.fluid.layer_helper import LayerHelper
+        self._helper = LayerHelper(
+            'SimpleRNNCell', act="tanh", param_attr=param_attr)
+
+    def _build_once(self, inputs, pre_hidden):
+        i2h_param_shape = [self.step_input_size, self.hidden_size]
+        h2h_param_shape = [self.hidden_size, self.hidden_size]
+        h2o_param_shape = [self.output_size, self.hidden_size]
+        self._i2h_w = self._helper.create_parameter(
+            attr=self._helper.param_attr,
+            shape=i2h_param_shape,
+            dtype=self._dtype,
+            is_bias=False)
+        self._h2h_w = self._helper.create_parameter(
+            attr=self._helper.param_attr,
+            shape=h2h_param_shape,
+            dtype=self._dtype,
+            is_bias=False)
+        self._h2o_w = self._helper.create_parameter(
+            attr=self._helper.param_attr,
+            shape=h2o_param_shape,
+            dtype=self._dtype,
+            is_bias=False)
+
+    def forward(self, input, pre_hidden):
+
+        tmp_i2h = self._helper.create_variable_for_type_inference(self._dtype)
+        tmp_h2h = self._helper.create_variable_for_type_inference(self._dtype)
+        hidden = self._helper.create_variable_for_type_inference(self._dype)
+        out = self._helper.create_variable_for_type_inference(self._dype)
+        softmax_out = self._helper.create_variable_for_type_inference(
+            self._dtype)
+        reduce_out = self._helper.create_variable_for_type_inference(
+            self._dtype)
+        self._helper.append_op(
+            type="mul",
+            inputs={"X": input,
+                    "Y": self._i2h_w},
+            outputs={"Out": tmp_i2h},
+            attrs={"x_num_col_dims": 1,
+                   "y_num_col_dims": 1})
+
+        self._helper.append_op(
+            type="mul",
+            inputs={"X": pre_hidden,
+                    "Y": self._h2h_w},
+            outputs={"Out": tmp_h2h},
+            attrs={"x_num_col_dims": 1,
+                   "y_num_col_dims": 1})
+
+        self._helper.append_op(
+            type="elementwise_add",
+            inputs={'X': tmp_h2h,
+                    'Y': tmp_i2h},
+            outputs={'Out': hidden},
+            attrs={'axis': -1,
+                   'use_mkldnn': False})
+        hidden = self._helper.append_activation(hidden)
+
+        self._helper.append_op(
+            type="mul",
+            inputs={"X": hidden,
+                    "Y": self._h2o_w},
+            outputs={"Out": out},
+            attrs={"x_num_col_dims": 1,
+                   "y_num_col_dims": 1})
+
+        self._helper.append_op(
+            type="softmax",
+            inputs={"X": out},
+            outputs={"Out": softmax_out},
+            attrs={"use_cudnn": False})
+
+        self._helper.append_op(
+            type='reduce_sum',
+            inputs={'X': softmax_out},
+            outputs={'Out': reduce_out},
+            attrs={'dim': None,
+                   'keep_dim': False,
+                   'reduce_all': True})
+
+        return reduce_out, hidden
+
+
+class SimpleRNN(fluid.imperative.Layer):
+    def __init__(self):
+        super(SimpleRNN, self).__init__()
+        self.seq_len = 4
+        self._cell = SimpleRNNCell(
+            3,
+            3,
+            3,
+            fluid.ParamAttr(initializer=fluid.initializer.Constant(value=0.1)))
+
+    def forward(self, inputs):
+        outs = list()
+        pre_hiddens = list()
+
+        init_hidden = fluid.layers.tensor.create_parameter(
+            attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=0.1)),
+            shape=[1, 3],
+            dtype='float32',
+            is_bias=False)
+        pre_hidden = init_hidden
+        for i in range(self.seq_len):
+            input = fluid.layers.slice(
+                inputs, axes=[1], starts=[i], ends=[i + 1])
+            input = fluid.layers.reshape(input, shape=[1, 3])
+            out_softmax, pre_hidden = self._cell(input, pre_hidden)
+            outs.append(out_softmax)
+
+        return outs, pre_hiddens
+
+
 class TestImperative(unittest.TestCase):
+    def test_sum_op(self):
+        x = np.ones([2, 2], np.float32)
+        with fluid.imperative.guard():
+            inputs = []
+            for _ in range(10):
+                inputs.append(fluid.imperative.base.to_variable(x))
+            ret = fluid.layers.sums(inputs)
+            loss = fluid.layers.reduce_sum(ret)
+            loss._backward()
+            self.assertTrue(np.allclose(ret._numpy(), x * 10))
+            self.assertTrue(np.allclose(inputs[0]._gradient(), x))
+
     def test_layer(self):
         with fluid.imperative.guard():
             cl = core.Layer()
@@ -199,6 +333,41 @@ class TestImperative(unittest.TestCase):
         self.assertTrue(np.allclose(dy_out, static_out))
         self.assertTrue(np.allclose(dy_grad, static_grad))
 
+    def test_rnn(self):
+        np_inp = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0],
+                           [10.0, 11.0, 12.0]])
+        np_inp = np_inp.reshape((1, 4, 3))
+        np_inp = np_inp.astype(np.float32)
+        with fluid.imperative.guard():
+            var_inp = fluid.imperative.base.to_variable(np_inp)
+            var_inp = fluid.layers.reshape(var_inp, shape=[1, 4, 3])
+            simple_rnn = SimpleRNN()
+            outs, pre_hiddens = simple_rnn.forward(var_inp)
+            dy_out = outs[3]._numpy()
+            outs[3]._backward()
+            dy_grad_h2o = simple_rnn._cell._h2o_w._gradient()
+            dy_grad_h2h = simple_rnn._cell._h2h_w._gradient()
+            dy_grad_i2h = simple_rnn._cell._i2h_w._gradient()
+
+        with new_program_scope():
+            inp = fluid.layers.data(
+                name="inp", shape=[1, 4, 3], append_batch_size=False)
+            simple_rnn = SimpleRNN()
+            outs, pre_hiddens = simple_rnn(inp)
+            param_grads = fluid.backward.append_backward(outs[3])
+            exe = fluid.Executor(fluid.CPUPlace())
+            exe.run(fluid.default_startup_program())
+            static_out, static_grad_h2o, static_grad_h2h, static_grad_i2h = exe.run(
+                feed={inp.name: np_inp},
+                fetch_list=[
+                    outs[3].name, param_grads[0][1].name,
+                    param_grads[1][1].name, param_grads[2][1].name
+                ])
+        self.assertTrue(np.allclose(dy_out, static_out))
+        self.assertTrue(np.allclose(dy_grad_h2o, static_grad_h2o))
+        self.assertTrue(np.allclose(dy_grad_h2h, static_grad_h2h))
+        self.assertTrue(np.allclose(dy_grad_i2h, static_grad_i2h))
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_imperative_optimizer.py

@@ -82,13 +82,14 @@ class MNIST(fluid.imperative.Layer):
         self._simple_img_conv_pool_2 = SimpleImgConvPool(
             20, 50, 5, 2, 2, act="relu")
 
-        pool_2_shape = 50 * 8 * 8
+        pool_2_shape = 50 * 4 * 4
         SIZE = 10
         scale = (2.0 / (pool_2_shape**2 * SIZE))**0.5
         self._fc = FC(10,
                       param_attr=fluid.param_attr.ParamAttr(
                           initializer=fluid.initializer.NormalInitializer(
-                              loc=0.0, scale=scale)))
+                              loc=0.0, scale=scale)),
+                      act="softmax")
 
     def forward(self, inputs):
         x = self._simple_img_conv_pool_1(inputs)
@@ -98,9 +99,9 @@ class MNIST(fluid.imperative.Layer):
 
 
 class TestImperativeMnist(unittest.TestCase):
-    def test_mnist_cpu_float32(self):
+    def test_mnist_float32(self):
         seed = 90
-
+        batch_num = 2
         with fluid.imperative.guard():
             fluid.default_startup_program().random_seed = seed
             fluid.default_main_program().random_seed = seed
@@ -112,15 +113,15 @@ class TestImperativeMnist(unittest.TestCase):
 
             dy_param_init_value = {}
             for batch_id, data in enumerate(train_reader()):
-                if batch_id >= 2:
+                if batch_id >= batch_num:
                     break
 
-                x_data = np.array(
+                dy_x_data = np.array(
                     [x[0].reshape(1, 28, 28) for x in data]).astype('float32')
                 y_data = np.array([x[1] for x in data]).astype('int64').reshape(
                     128, 1)
 
-                img = to_variable(x_data)
+                img = to_variable(dy_x_data)
                 label = to_variable(y_data)
                 label._stop_gradient = True
 
@@ -136,6 +137,7 @@ class TestImperativeMnist(unittest.TestCase):
 
                 avg_loss._backward()
                 sgd.minimize(avg_loss)
+                mnist.clear_gradients()
                 dy_param_value = {}
                 for param in fluid.default_main_program().global_block(
                 ).all_parameters():
@@ -175,10 +177,10 @@ class TestImperativeMnist(unittest.TestCase):
                 static_param_init_value[static_param_name_list[i]] = out[i]
 
             for batch_id, data in enumerate(train_reader()):
-                if batch_id >= 2:
+                if batch_id >= batch_num:
                     break
 
-                x_data = np.array(
+                static_x_data = np.array(
                     [x[0].reshape(1, 28, 28) for x in data]).astype('float32')
                 y_data = np.array([x[1] for x in data]).astype('int64').reshape(
                     [128, 1])
@@ -186,7 +188,7 @@ class TestImperativeMnist(unittest.TestCase):
                 fetch_list = [avg_loss.name]
                 fetch_list.extend(static_param_name_list)
                 out = exe.run(fluid.default_main_program(),
-                              feed={"pixel": x_data,
+                              feed={"pixel": static_x_data,
                                     "label": y_data},
                               fetch_list=fetch_list)
 
@@ -196,11 +198,12 @@ class TestImperativeMnist(unittest.TestCase):
                     static_param_value[static_param_name_list[i - 1]] = out[i]
 
         for key, value in six.iteritems(static_param_init_value):
-            self.assertTrue(
-                np.allclose(value.all(), dy_param_init_value[key].all()))
-        self.assertTrue(np.allclose(static_out.all(), dy_out.all()))
+            self.assertTrue(np.allclose(value, dy_param_init_value[key]))
+
+        self.assertTrue(np.allclose(static_out, dy_out))
+
         for key, value in six.iteritems(static_param_value):
-            self.assertTrue(np.allclose(value.all(), dy_param_value[key].all()))
+            self.assertTrue(np.allclose(value, dy_param_value[key]))
 
 
 if __name__ == '__main__':

python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import paddle.fluid as fluid
+from paddle.fluid.imperative.nn import Embedding
+import paddle.fluid.framework as framework
+from paddle.fluid.optimizer import SGDOptimizer
+from paddle.fluid.imperative.base import to_variable
+from test_imperative_base import new_program_scope
+import numpy as np
+import six
+from paddle.fluid.backward import append_backward
+
+
+class SimpleLSTMRNN(fluid.imperative.Layer):
+    def __init__(self,
+                 hidden_size,
+                 num_steps,
+                 num_layers=2,
+                 init_scale=0.1,
+                 dropout=None):
+        super(SimpleLSTMRNN, self).__init__()
+        self._hidden_size = hidden_size
+        self._num_layers = num_layers
+        self._init_scale = init_scale
+        self._dropout = dropout
+        self._input = None
+        self._num_steps = num_steps
+
+    def _build_once(self, input_embedding, init_hidden=None, init_cell=None):
+        self.weight_1_arr = []
+        self.weight_2_arr = []
+        self.bias_arr = []
+        self.hidden_array = []
+        self.cell_array = []
+        self.mask_array = []
+
+        for i in range(self._num_layers):
+            weight_1 = fluid.layers.create_parameter(
+                shape=[self._hidden_size * 2, self._hidden_size * 4],
+                dtype="float32",
+                name="fc_weight1_" + str(i),
+                default_initializer=fluid.initializer.UniformInitializer(
+                    low=-self._init_scale, high=self._init_scale))
+            self.weight_1_arr.append(weight_1)
+            bias_1 = fluid.layers.create_parameter(
+                [self._hidden_size * 4],
+                dtype="float32",
+                name="fc_bias1_" + str(i),
+                default_initializer=fluid.initializer.Constant(0.0))
+            self.bias_arr.append(bias_1)
+
+            pre_hidden = fluid.layers.slice(
+                init_hidden, axes=[0], starts=[i], ends=[i + 1])
+            pre_cell = fluid.layers.slice(
+                init_cell, axes=[0], starts=[i], ends=[i + 1])
+            pre_hidden = fluid.layers.reshape(
+                pre_hidden, shape=[-1, self._hidden_size])
+            pre_cell = fluid.layers.reshape(
+                pre_cell, shape=[-1, self._hidden_size])
+            self.hidden_array.append(pre_hidden)
+            self.cell_array.append(pre_cell)
+
+    def parameters(self):
+        parameters = list()
+        for param in self.weight_1_arr:
+            parameters.append(param)
+        for param in self.weight_2_arr:
+            parameters.append(param)
+        for bias in self.bias_arr:
+            parameters.append(bias)
+        return parameters
+
+    def forward(self, input_embedding, init_hidden=None, init_cell=None):
+        res = []
+        for index in range(self._num_steps):
+            self._input = fluid.layers.slice(
+                input_embedding, axes=[1], starts=[index], ends=[index + 1])
+            self._input = fluid.layers.reshape(
+                self._input, shape=[-1, self._hidden_size])
+            for k in range(self._num_layers):
+                pre_hidden = self.hidden_array[k]
+                pre_cell = self.cell_array[k]
+                weight_1 = self.weight_1_arr[k]
+                bias = self.bias_arr[k]
+
+                nn = fluid.layers.concat([self._input, pre_hidden], 1)
+                gate_input = fluid.layers.matmul(x=nn, y=weight_1)
+
+                gate_input = fluid.layers.elementwise_add(gate_input, bias)
+                i, j, f, o = fluid.layers.split(
+                    gate_input, num_or_sections=4, dim=-1)
+                c = pre_cell * fluid.layers.sigmoid(f) + fluid.layers.sigmoid(
+                    i) * fluid.layers.tanh(j)
+                m = fluid.layers.tanh(c) * fluid.layers.sigmoid(o)
+                self.hidden_array[k] = m
+                self.cell_array[k] = c
+                self._input = m
+
+                if self._dropout is not None and self._dropout > 0.0:
+                    self._input = fluid.layers.dropout(
+                        self._input,
+                        dropout_prob=self._dropout,
+                        dropout_implementation='upscale_in_train')
+            res.append(
+                fluid.layers.reshape(
+                    self._input, shape=[1, -1, self._hidden_size]))
+        real_res = fluid.layers.concat(res, 0)
+        real_res = fluid.layers.transpose(x=real_res, perm=[1, 0, 2])
+        last_hidden = fluid.layers.concat(self.hidden_array, 1)
+        last_hidden = fluid.layers.reshape(
+            last_hidden, shape=[-1, self._num_layers, self._hidden_size])
+        last_hidden = fluid.layers.transpose(x=last_hidden, perm=[1, 0, 2])
+        last_cell = fluid.layers.concat(self.cell_array, 1)
+        last_cell = fluid.layers.reshape(
+            last_cell, shape=[-1, self._num_layers, self._hidden_size])
+        last_cell = fluid.layers.transpose(x=last_cell, perm=[1, 0, 2])
+        return real_res, last_hidden, last_cell
+
+
+class PtbModel(fluid.imperative.Layer):
+    def __init__(self,
+                 hidden_size,
+                 vocab_size,
+                 num_layers=2,
+                 num_steps=20,
+                 init_scale=0.1,
+                 dropout=None):
+        super(PtbModel, self).__init__()
+        self.hidden_size = hidden_size
+        self.vocab_size = vocab_size
+        self.init_scale = init_scale
+        self.num_layers = num_layers
+        self.num_steps = num_steps
+        self.dropout = dropout
+        self.simple_lstm_rnn = SimpleLSTMRNN(
+            hidden_size,
+            num_steps,
+            num_layers=num_layers,
+            init_scale=init_scale,
+            dropout=dropout)
+        self.embedding = Embedding(
+            size=[vocab_size, hidden_size],
+            dtype='float32',
+            is_sparse=False,
+            param_attr=fluid.ParamAttr(
+                name='embedding_para',
+                initializer=fluid.initializer.UniformInitializer(
+                    low=-init_scale, high=init_scale)))
+        self.softmax_weight = fluid.layers.create_parameter(
+            [self.hidden_size, self.vocab_size],
+            dtype="float32",
+            name="softmax_weight",
+            default_initializer=fluid.initializer.UniformInitializer(
+                low=-self.init_scale, high=self.init_scale))
+        self.softmax_bias = fluid.layers.create_parameter(
+            [self.vocab_size],
+            dtype="float32",
+            name='softmax_bias',
+            default_initializer=fluid.initializer.UniformInitializer(
+                low=-self.init_scale, high=self.init_scale))
+
+    def _build_once(self, input, label, init_hidden, init_cell):
+        pass
+
+    def parameters(self):
+        parameters = self.simple_lstm_rnn.parameters() + [
+            self.softmax_weight, self.softmax_bias
+        ] + self.embedding.parameters()
+        return parameters
+
+    def forward(self, input, label, init_hidden, init_cell):
+
+        init_h = fluid.layers.reshape(
+            init_hidden, shape=[self.num_layers, -1, self.hidden_size])
+
+        init_c = fluid.layers.reshape(
+            init_cell, shape=[self.num_layers, -1, self.hidden_size])
+
+        x_emb = self.embedding(input)
+        x_emb = fluid.layers.reshape(
+            x_emb, shape=[-1, self.num_steps, self.hidden_size])
+        if self.dropout is not None and self.dropout > 0.0:
+            x_emb = fluid.layers.dropout(
+                x_emb,
+                dropout_prob=self.drop_out,
+                dropout_implementation='upscale_in_train')
+        rnn_out, last_hidden, last_cell = self.simple_lstm_rnn(x_emb, init_h,
+                                                               init_c)
+        rnn_out = fluid.layers.reshape(
+            rnn_out, shape=[-1, self.num_steps, self.hidden_size])
+        projection = fluid.layers.matmul(rnn_out, self.softmax_weight)
+        projection = fluid.layers.elementwise_add(projection, self.softmax_bias)
+        projection = fluid.layers.reshape(
+            projection, shape=[-1, self.vocab_size])
+        projection = fluid.layers.reshape(
+            projection, shape=[-1, self.vocab_size])
+        loss = fluid.layers.softmax_with_cross_entropy(
+            logits=projection, label=label, soft_label=False)
+        loss = fluid.layers.reshape(loss, shape=[-1, self.num_steps])
+        loss = fluid.layers.reduce_mean(loss, dim=[0])
+        loss = fluid.layers.reduce_sum(loss)
+        loss.permissions = True
+
+        return loss, last_hidden, last_cell
+
+
+class TestImperativePtbRnn(unittest.TestCase):
+    def test_ptb_rnn_cpu_float32(self):
+        seed = 90
+        hidden_size = 10
+        vocab_size = 1000
+        num_layers = 1
+        num_steps = 3
+        init_scale = 0.1
+        batch_size = 4
+
+        with fluid.imperative.guard():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+            # TODO: marsyang1993 Change seed to
+            ptb_model = PtbModel(
+                hidden_size=hidden_size,
+                vocab_size=vocab_size,
+                num_layers=num_layers,
+                num_steps=num_steps,
+                init_scale=init_scale)
+
+            sgd = SGDOptimizer(learning_rate=1e-3)
+            dy_param_updated = dict()
+            dy_param_init = dict()
+            dy_loss = None
+            last_hidden = None
+            last_cell = None
+            for i in range(2):
+                x_data = np.arange(12).reshape(4, 3).astype('int64')
+                y_data = np.arange(1, 13).reshape(4, 3).astype('int64')
+                x_data = x_data.reshape((-1, num_steps, 1))
+                y_data = y_data.reshape((-1, 1))
+                init_hidden_data = np.zeros(
+                    (num_layers, batch_size, hidden_size), dtype='float32')
+                init_cell_data = np.zeros(
+                    (num_layers, batch_size, hidden_size), dtype='float32')
+                x = to_variable(x_data)
+                y = to_variable(y_data)
+                init_hidden = to_variable(init_hidden_data)
+                init_cell = to_variable(init_cell_data)
+                dy_loss, last_hidden, last_cell = ptb_model(x, y, init_hidden,
+                                                            init_cell)
+                if i == 0:
+                    for param in ptb_model.parameters():
+                        dy_param_init[param.name] = param._numpy()
+                dy_loss._backward()
+                sgd.minimize(dy_loss)
+                for param in ptb_model.parameters():
+                    dy_param_updated[param.name] = param._numpy()
+
+        with new_program_scope():
+            fluid.default_startup_program().random_seed = seed
+            fluid.default_main_program().random_seed = seed
+            # TODO: marsyang1993 Change seed to
+            ptb_model = PtbModel(
+                hidden_size=hidden_size,
+                vocab_size=vocab_size,
+                num_layers=num_layers,
+                num_steps=num_steps,
+                init_scale=init_scale)
+
+            exe = fluid.Executor(fluid.CPUPlace())
+            sgd = SGDOptimizer(learning_rate=1e-3)
+            x = fluid.layers.data(name="x", shape=[-1, 3, 1], dtype='int64')
+            y = fluid.layers.data(name="y", shape=[-1, 1], dtype='float32')
+            init_hidden = fluid.layers.data(
+                name="init_hidden", shape=[1], dtype='float32')
+            init_cell = fluid.layers.data(
+                name="init_cell", shape=[1], dtype='float32')
+
+            static_loss, static_last_hidden, static_last_cell = ptb_model(
+                x, y, init_hidden, init_cell)
+            sgd.minimize(static_loss)
+            static_param_updated = dict()
+            static_param_init = dict()
+            static_param_name_list = list()
+            for param in ptb_model.parameters():
+                static_param_name_list.append(param.name)
+
+            out = exe.run(framework.default_startup_program(),
+                          fetch_list=static_param_name_list)
+            for i in range(len(static_param_name_list)):
+                static_param_init[static_param_name_list[i]] = out[i]
+            static_loss_value = None
+            static_last_cell_value = None
+            static_last_hidden_value = None
+            for i in range(2):
+                x_data = np.arange(12).reshape(4, 3).astype('int64')
+                y_data = np.arange(1, 13).reshape(4, 3).astype('int64')
+                x_data = x_data.reshape((-1, num_steps, 1))
+                y_data = y_data.reshape((-1, 1))
+                init_hidden_data = np.zeros(
+                    (num_layers, batch_size, hidden_size), dtype='float32')
+                init_cell_data = np.zeros(
+                    (num_layers, batch_size, hidden_size), dtype='float32')
+                fetch_list = [static_loss, static_last_hidden, static_last_cell]
+                fetch_list.extend(static_param_name_list)
+                out = exe.run(fluid.default_main_program(),
+                              feed={
+                                  "x": x_data,
+                                  "y": y_data,
+                                  "init_hidden": init_hidden_data,
+                                  "init_cell": init_cell_data
+                              },
+                              fetch_list=fetch_list)
+                static_loss_value = out[0]
+                static_last_cell_value = out[1]
+                static_last_hidden_value = out[2]
+                for k in range(3, len(out)):
+                    static_param_updated[static_param_name_list[k - 3]] = out[k]
+
+            self.assertTrue(
+                np.allclose(static_loss_value.all(), dy_loss._numpy().all()))
+            self.assertTrue(
+                np.allclose(static_last_cell_value.all(),
+                            last_cell._numpy().all()))
+            self.assertTrue(
+                np.allclose(static_last_hidden_value.all(),
+                            last_hidden._numpy().all()))
+            for key, value in six.iteritems(static_param_init):
+                self.assertTrue(
+                    np.allclose(value.all(), dy_param_init[key].all()))
+            for key, value in six.iteritems(static_param_updated):
+                self.assertTrue(
+                    np.allclose(value.all(), dy_param_updated[key].all()))
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_imperative_resnet.py

@@ -264,6 +264,7 @@ class TestImperativeResnet(unittest.TestCase):
                         )] = np_array
 
                 optimizer.minimize(avg_loss)
+                resnet.clear_gradients()
 
                 dy_param_value = {}
                 for param in fluid.default_main_program().global_block(

python/paddle/fluid/tests/unittests/test_layers.py

@@ -58,7 +58,8 @@ class TestBook(unittest.TestCase):
     def test_simple_conv2d(self):
         program = Program()
         with program_guard(program, startup_program=Program()):
-            images = layers.data(name='pixel', shape=[3, 48, 48], dtype='int32')
+            images = layers.data(
+                name='pixel', shape=[3, 48, 48], dtype='float32')
             layers.conv2d(input=images, num_filters=3, filter_size=[4, 4])
 
         print(str(program))

python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py

@@ -19,7 +19,7 @@ import copy
 from op_test import OpTest
 
 
-def iou(box_a, box_b):
+def iou(box_a, box_b, norm):
     """Apply intersection-over-union overlap between box_a and box_b
     """
     xmin_a = min(box_a[0], box_a[2])
@@ -32,8 +32,10 @@ def iou(box_a, box_b):
     xmax_b = max(box_b[0], box_b[2])
     ymax_b = max(box_b[1], box_b[3])
 
-    area_a = (ymax_a - ymin_a) * (xmax_a - xmin_a)
-    area_b = (ymax_b - ymin_b) * (xmax_b - xmin_b)
+    area_a = (ymax_a - ymin_a + (norm == False)) * (xmax_a - xmin_a +
+                                                    (norm == False))
+    area_b = (ymax_b - ymin_b + (norm == False)) * (xmax_b - xmin_b +
+                                                    (norm == False))
     if area_a <= 0 and area_b <= 0:
         return 0.0
 
@@ -42,17 +44,21 @@ def iou(box_a, box_b):
     xb = min(xmax_a, xmax_b)
     yb = min(ymax_a, ymax_b)
 
-    inter_area = max(xb - xa, 0.0) * max(yb - ya, 0.0)
-
-    box_a_area = (box_a[2] - box_a[0]) * (box_a[3] - box_a[1])
-    box_b_area = (box_b[2] - box_b[0]) * (box_b[3] - box_b[1])
+    inter_area = max(xb - xa + (norm == False),
+                     0.0) * max(yb - ya + (norm == False), 0.0)
 
     iou_ratio = inter_area / (area_a + area_b - inter_area)
 
     return iou_ratio
 
 
-def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0):
+def nms(boxes,
+        scores,
+        score_threshold,
+        nms_threshold,
+        top_k=200,
+        normalized=True,
+        eta=1.0):
     """Apply non-maximum suppression at test time to avoid detecting too many
     overlapping bounding boxes for a given object.
     Args:
@@ -87,7 +93,7 @@ def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0):
         for k in range(len(selected_indices)):
             if keep:
                 kept_idx = selected_indices[k]
-                overlap = iou(boxes[idx], boxes[kept_idx])
+                overlap = iou(boxes[idx], boxes[kept_idx], normalized)
                 keep = True if overlap <= adaptive_threshold else False
             else:
                 break
@@ -99,16 +105,24 @@ def nms(boxes, scores, score_threshold, nms_threshold, top_k=200, eta=1.0):
 
 
 def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold,
-                   nms_top_k, keep_top_k):
-    class_num = scores.shape[0]
-    priorbox_num = scores.shape[1]
+                   nms_top_k, keep_top_k, normalized, shared):
+    if shared:
+        class_num = scores.shape[0]
+        priorbox_num = scores.shape[1]
+    else:
+        box_num = scores.shape[0]
+        class_num = scores.shape[1]
 
     selected_indices = {}
     num_det = 0
     for c in range(class_num):
         if c == background: continue
-        indices = nms(boxes, scores[c], score_threshold, nms_threshold,
-                      nms_top_k)
+        if shared:
+            indices = nms(boxes, scores[c], score_threshold, nms_threshold,
+                          nms_top_k, normalized)
+        else:
+            indices = nms(boxes[:, c, :], scores[:, c], score_threshold,
+                          nms_threshold, nms_top_k, normalized)
         selected_indices[c] = indices
         num_det += len(indices)
 
@@ -116,7 +130,10 @@ def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold,
         score_index = []
         for c, indices in selected_indices.items():
             for idx in indices:
-                score_index.append((scores[c][idx], c, idx))
+                if shared:
+                    score_index.append((scores[c][idx], c, idx))
+                else:
+                    score_index.append((scores[idx][c], c, idx))
 
         sorted_score_index = sorted(
             score_index, key=lambda tup: tup[0], reverse=True)
@@ -127,24 +144,75 @@ def multiclass_nms(boxes, scores, background, score_threshold, nms_threshold,
             selected_indices[c] = []
         for s, c, idx in sorted_score_index:
             selected_indices[c].append(idx)
+        if not shared:
+            for labels in selected_indices:
+                selected_indices[labels].sort()
         num_det = keep_top_k
 
     return selected_indices, num_det
 
 
-def batched_multiclass_nms(boxes, scores, background, score_threshold,
-                           nms_threshold, nms_top_k, keep_top_k):
+def lod_multiclass_nms(boxes, scores, background, score_threshold,
+                       nms_threshold, nms_top_k, keep_top_k, box_lod,
+                       normalized):
+    det_outs = []
+    lod = []
+    head = 0
+    for n in range(len(box_lod[0])):
+        box = boxes[head:head + box_lod[0][n]]
+        score = scores[head:head + box_lod[0][n]]
+        head = head + box_lod[0][n]
+        nmsed_outs, nmsed_num = multiclass_nms(
+            box,
+            score,
+            background,
+            score_threshold,
+            nms_threshold,
+            nms_top_k,
+            keep_top_k,
+            normalized,
+            shared=False)
+        if nmsed_num == 0:
+            #lod.append(1)
+            continue
+        lod.append(nmsed_num)
+        for c, indices in nmsed_outs.items():
+            for idx in indices:
+                xmin, ymin, xmax, ymax = box[idx, c, :]
+                det_outs.append([c, score[idx][c], xmin, ymin, xmax, ymax])
+    if len(lod) == 0:
+        lod.append(1)
+
+    return det_outs, lod
+
+
+def batched_multiclass_nms(boxes,
+                           scores,
+                           background,
+                           score_threshold,
+                           nms_threshold,
+                           nms_top_k,
+                           keep_top_k,
+                           normalized=True):
     batch_size = scores.shape[0]
 
     det_outs = []
     lod = []
     for n in range(batch_size):
-        nmsed_outs, nmsed_num = multiclass_nms(boxes[n], scores[n], background,
-                                               score_threshold, nms_threshold,
-                                               nms_top_k, keep_top_k)
-        lod.append(nmsed_num)
-        if nmsed_num == 0: continue
+        nmsed_outs, nmsed_num = multiclass_nms(
+            boxes[n],
+            scores[n],
+            background,
+            score_threshold,
+            nms_threshold,
+            nms_top_k,
+            keep_top_k,
+            normalized,
+            shared=True)
+        if nmsed_num == 0:
+            continue
 
+        lod.append(nmsed_num)
         tmp_det_out = []
         for c, indices in nmsed_outs.items():
             for idx in indices:
@@ -154,7 +222,8 @@ def batched_multiclass_nms(boxes, scores, background, score_threshold,
         sorted_det_out = sorted(
             tmp_det_out, key=lambda tup: tup[0], reverse=False)
         det_outs.extend(sorted_det_out)
-
+    if len(lod) == 0:
+        lod += [1]
     return det_outs, lod
 
 
@@ -168,7 +237,6 @@ class TestMulticlassNMSOp(OpTest):
         M = 1200
         C = 21
         BOX_SIZE = 4
-
         background = 0
         nms_threshold = 0.3
         nms_top_k = 400
@@ -206,6 +274,7 @@ class TestMulticlassNMSOp(OpTest):
             'keep_top_k': keep_top_k,
             'score_threshold': score_threshold,
             'nms_eta': 1.0,
+            'normalized': True,
         }
 
     def test_check_output(self):
@@ -219,13 +288,70 @@ class TestMulticlassNMSOpNoOutput(TestMulticlassNMSOp):
         self.score_threshold = 2.0
 
 
+class TestMulticlassNMSLoDInput(OpTest):
+    def set_argument(self):
+        self.score_threshold = 0.01
+
+    def setUp(self):
+        self.set_argument()
+        M = 1200
+        C = 21
+        BOX_SIZE = 4
+        box_lod = [[1200]]
+        background = 0
+        nms_threshold = 0.3
+        nms_top_k = 400
+        keep_top_k = 200
+        score_threshold = self.score_threshold
+        normalized = False
+
+        scores = np.random.random((M, C)).astype('float32')
+
+        def softmax(x):
+            shiftx = x - np.max(x).clip(-64.)
+            exps = np.exp(shiftx)
+            return exps / np.sum(exps)
+
+        scores = np.apply_along_axis(softmax, 1, scores)
+
+        boxes = np.random.random((M, C, BOX_SIZE)).astype('float32')
+        boxes[:, :, 0] = boxes[:, :, 0] * 10
+        boxes[:, :, 1] = boxes[:, :, 1] * 10
+        boxes[:, :, 2] = boxes[:, :, 2] * 10 + 10
+        boxes[:, :, 3] = boxes[:, :, 3] * 10 + 10
+
+        nmsed_outs, lod = lod_multiclass_nms(
+            boxes, scores, background, score_threshold, nms_threshold,
+            nms_top_k, keep_top_k, box_lod, normalized)
+        nmsed_outs = [-1] if not nmsed_outs else nmsed_outs
+        nmsed_outs = np.array(nmsed_outs).astype('float32')
+        self.op_type = 'multiclass_nms'
+        self.inputs = {
+            'BBoxes': (boxes, box_lod),
+            'Scores': (scores, box_lod),
+        }
+        self.outputs = {'Out': (nmsed_outs, [lod])}
+        self.attrs = {
+            'background_label': 0,
+            'nms_threshold': nms_threshold,
+            'nms_top_k': nms_top_k,
+            'keep_top_k': keep_top_k,
+            'score_threshold': score_threshold,
+            'nms_eta': 1.0,
+            'normalized': normalized,
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+
 class TestIOU(unittest.TestCase):
     def test_iou(self):
         box1 = np.array([4.0, 3.0, 7.0, 5.0]).astype('float32')
         box2 = np.array([3.0, 4.0, 6.0, 8.0]).astype('float32')
 
         expt_output = np.array([2.0 / 16.0]).astype('float32')
-        calc_output = np.array([iou(box1, box2)]).astype('float32')
+        calc_output = np.array([iou(box1, box2, True)]).astype('float32')
         self.assertTrue(np.allclose(calc_output, expt_output))
 
 

python/paddle/fluid/tests/unittests/test_nearest_interp_op.py

@@ -24,7 +24,8 @@ def nearest_neighbor_interp_np(X,
                                out_h,
                                out_w,
                                out_size=None,
-                               actual_shape=None):
+                               actual_shape=None,
+                               align_corners=True):
     """nearest neighbor interpolation implement in shape [N, C, H, W]"""
     if out_size is not None:
         out_h = out_size[0]
@@ -35,17 +36,31 @@ def nearest_neighbor_interp_np(X,
     n, c, in_h, in_w = X.shape
 
     ratio_h = ratio_w = 0.0
-    if out_h > 1:
-        ratio_h = (in_h - 1.0) / (out_h - 1.0)
-    if out_w > 1:
-        ratio_w = (in_w - 1.0) / (out_w - 1.0)
+    if (out_h > 1):
+        if (align_corners):
+            ratio_h = (in_h - 1.0) / (out_h - 1.0)
+        else:
+            ratio_h = 1.0 * in_h / out_h
+    if (out_w > 1):
+        if (align_corners):
+            ratio_w = (in_w - 1.0) / (out_w - 1.0)
+        else:
+            ratio_w = 1.0 * in_w / out_w
 
     out = np.zeros((n, c, out_h, out_w))
-    for i in range(out_h):
-        in_i = int(ratio_h * i + 0.5)
-        for j in range(out_w):
-            in_j = int(ratio_w * j + 0.5)
-            out[:, :, i, j] = X[:, :, in_i, in_j]
+
+    if align_corners:
+        for i in range(out_h):
+            in_i = int(ratio_h * i + 0.5)
+            for j in range(out_w):
+                in_j = int(ratio_w * j + 0.5)
+                out[:, :, i, j] = X[:, :, in_i, in_j]
+    else:
+        for i in range(out_h):
+            in_i = int(ratio_h * i)
+            for j in range(out_w):
+                in_j = int(ratio_w * j)
+                out[:, :, i, j] = X[:, :, in_i, in_j]
 
     return out.astype(X.dtype)
 
@@ -59,7 +74,8 @@ class TestNearestInterpOp(OpTest):
         input_np = np.random.random(self.input_shape).astype("float32")
 
         output_np = nearest_neighbor_interp_np(input_np, self.out_h, self.out_w,
-                                               self.out_size, self.actual_shape)
+                                               self.out_size, self.actual_shape,
+                                               self.align_corners)
         self.inputs = {'X': input_np}
         if self.out_size is not None:
             self.inputs['OutSize'] = self.out_size
@@ -68,7 +84,8 @@ class TestNearestInterpOp(OpTest):
         self.attrs = {
             'out_h': self.out_h,
             'out_w': self.out_w,
-            'interp_method': self.interp_method
+            'interp_method': self.interp_method,
+            'align_corners': self.align_corners,
         }
         self.outputs = {'Out': output_np}
 
@@ -84,6 +101,7 @@ class TestNearestInterpOp(OpTest):
         self.out_h = 2
         self.out_w = 2
         self.out_size = np.array([3, 3]).astype("int32")
+        self.align_corners = True
 
 
 class TestNearestNeighborInterpCase1(TestNearestInterpOp):
@@ -92,6 +110,7 @@ class TestNearestNeighborInterpCase1(TestNearestInterpOp):
         self.input_shape = [4, 1, 7, 8]
         self.out_h = 1
         self.out_w = 1
+        self.align_corners = True
 
 
 class TestNearestNeighborInterpCase2(TestNearestInterpOp):
@@ -100,6 +119,7 @@ class TestNearestNeighborInterpCase2(TestNearestInterpOp):
         self.input_shape = [3, 3, 9, 6]
         self.out_h = 12
         self.out_w = 12
+        self.align_corners = True
 
 
 class TestNearestNeighborInterpCase3(TestNearestInterpOp):
@@ -108,6 +128,7 @@ class TestNearestNeighborInterpCase3(TestNearestInterpOp):
         self.input_shape = [1, 1, 128, 64]
         self.out_h = 64
         self.out_w = 128
+        self.align_corners = True
 
 
 class TestNearestNeighborInterpCase4(TestNearestInterpOp):
@@ -117,6 +138,7 @@ class TestNearestNeighborInterpCase4(TestNearestInterpOp):
         self.out_h = 1
         self.out_w = 1
         self.out_size = np.array([2, 2]).astype("int32")
+        self.align_corners = True
 
 
 class TestNearestNeighborInterpCase5(TestNearestInterpOp):
@@ -126,6 +148,7 @@ class TestNearestNeighborInterpCase5(TestNearestInterpOp):
         self.out_h = 12
         self.out_w = 12
         self.out_size = np.array([11, 11]).astype("int32")
+        self.align_corners = True
 
 
 class TestNearestNeighborInterpCase6(TestNearestInterpOp):
@@ -135,6 +158,7 @@ class TestNearestNeighborInterpCase6(TestNearestInterpOp):
         self.out_h = 64
         self.out_w = 128
         self.out_size = np.array([65, 129]).astype("int32")
+        self.align_corners = True
 
 
 class TestNearestNeighborInterpActualShape(TestNearestInterpOp):
@@ -144,6 +168,7 @@ class TestNearestNeighborInterpActualShape(TestNearestInterpOp):
         self.out_h = 64
         self.out_w = 32
         self.out_size = np.array([66, 40]).astype("int32")
+        self.align_corners = True
 
 
 class TestNearestInterpOpUint8(OpTest):
@@ -155,14 +180,16 @@ class TestNearestInterpOpUint8(OpTest):
         input_np = np.random.randint(
             low=0, high=256, size=self.input_shape).astype("uint8")
         output_np = nearest_neighbor_interp_np(input_np, self.out_h, self.out_w,
-                                               self.out_size, self.actual_shape)
+                                               self.out_size, self.actual_shape,
+                                               self.align_corners)
         self.inputs = {'X': input_np}
         if self.out_size is not None:
             self.inputs['OutSize'] = self.out_size
         self.attrs = {
             'out_h': self.out_h,
             'out_w': self.out_w,
-            'interp_method': self.interp_method
+            'interp_method': self.interp_method,
+            'align_corners': self.align_corners
         }
         self.outputs = {'Out': output_np}
 
@@ -174,6 +201,7 @@ class TestNearestInterpOpUint8(OpTest):
         self.input_shape = [1, 3, 9, 6]
         self.out_h = 10
         self.out_w = 9
+        self.align_corners = True
 
 
 class TestNearestNeighborInterpCase1Uint8(TestNearestInterpOpUint8):
@@ -182,6 +210,7 @@ class TestNearestNeighborInterpCase1Uint8(TestNearestInterpOpUint8):
         self.input_shape = [2, 3, 128, 64]
         self.out_h = 120
         self.out_w = 50
+        self.align_corners = True
 
 
 class TestNearestNeighborInterpCase2Uint8(TestNearestInterpOpUint8):
@@ -191,6 +220,12 @@ class TestNearestNeighborInterpCase2Uint8(TestNearestInterpOpUint8):
         self.out_h = 5
         self.out_w = 13
         self.out_size = np.array([6, 15]).astype("int32")
+        self.align_corners = True
+
+
+class TestNearestInterpWithoutCorners(TestNearestInterpOp):
+    def set_align_corners(self):
+        self.align_corners = False
 
 
 if __name__ == "__main__":

python/paddle/fluid/transpiler/details/__init__.py

@@ -17,3 +17,4 @@ from __future__ import print_function
 from .program_utils import *
 from .ufind import *
 from .checkport import *
+from .vars_distributed import *

python/paddle/fluid/transpiler/details/vars_distributed.py

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import print_function
+from paddle.fluid.framework import Variable
+
+
+class VarStruct(object):
+    """
+    record part properties of a Variable in python.
+    """
+
+    def __init__(self, name, shape, dtype, type, lod_level, persistable):
+        self.name = name
+        self.shape = shape
+        self.dtype = dtype
+        self.type = type
+        self.lod_level = lod_level
+        self.persistable = persistable
+
+
+class VarDistributed(object):
+    """
+    a class to record the var distributed on parameter servers.
+    the class will record the relationship between origin var and slice var.
+    the slice var's properties, such as type/shape/offset/endpoint.
+    """
+
+    def __init__(self,
+                 origin_var,
+                 slice_var,
+                 is_slice=None,
+                 block_id=None,
+                 offset=None,
+                 vtype=None,
+                 endpoint=None):
+        """
+        Args:
+            origin_var(Variable|VarStruct): origin var properties
+            slice_var(Variable|VarStruct): slice var properties
+            is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard.
+            block_id(int|None): the number about the slice var.
+            offset(int|None): if the slice var is sliced, offset is the numel before the var.
+            vtype(str|None): a tag, such as Optimizer/Param/RemoteProfetch.
+            endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001"
+        """
+
+        if isinstance(origin_var, Variable):
+            self.origin = self.__create_var_struct(origin_var)
+        else:
+            self.origin = origin_var
+
+        if isinstance(slice_var, Variable):
+            self.slice = self.__create_var_struct(slice_var)
+        else:
+            self.slice = slice_var
+
+        if self.equal(self.origin, self.slice):
+            self.is_slice = False
+            self.block_id = 0
+            self.offset = 0
+        else:
+            self.is_slice = True
+            self.block_id = 0
+            self.offset = 0
+
+        if is_slice is not None:
+            self.is_slice = is_slice
+        if block_id is not None:
+            self.block_id = block_id
+        if offset is not None:
+            self.offset = offset
+
+        self.vtype = vtype
+        self.endpoint = endpoint
+
+    @staticmethod
+    def __create_var_struct(var):
+        return VarStruct(var.name, var.shape, var.dtype, var.type,
+                         var.lod_level, var.persistable)
+
+    @staticmethod
+    def equal(var1, var2):
+        """
+        the two var is equal or not.
+        Returns:
+            bool: equal will return True else False
+        """
+        assert isinstance(var1, VarStruct) and isinstance(var2, VarStruct)
+
+        return var1.name == var2.name and \
+               var1.type == var2.type and \
+               var1.shape == var2.shape and \
+               var1.dtype == var2.dtype and \
+               var1.lod_level == var2.lod_level and \
+               var1.persistable == var2.persistable
+
+    def __str__(self):
+        origin_var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})". \
+            format(i="{", e="}", name=self.origin.name, type=self.origin.type,
+                   shape=self.origin.shape, dtype=self.origin.dtype)
+
+        slice_var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})" \
+                        ".slice({is_slice}).block({block_id}).offset({offset})". \
+            format(i="{", e="}", name=self.slice.name, type=self.slice.type,
+                   shape=self.slice.shape, dtype=self.slice.dtype,
+                   is_slice=self.is_slice, block_id=self.block_id, offset=self.offset)
+
+        return "var owned: {}, origin var: ( {} ), slice var: ( {} ), endpoint: {} ".format(
+            self.vtype, origin_var_str, slice_var_str, self.endpoint)
+
+
+class VarsDistributed(object):
+    """
+    a gather about VarDistributed with many methods to find distributed vars.
+    through the class, we can get overview about the distributed parameters on parameter servers.
+    this class may centralized and convenient for developer to manage and get variable's distribute.
+    other module can also use this to find variables such io.py.
+    """
+
+    def __init__(self):
+        self.distributed_vars = []
+
+    def add_distributed_var(self,
+                            origin_var,
+                            slice_var,
+                            is_slice=None,
+                            block_id=None,
+                            offset=None,
+                            vtype=None,
+                            endpoint=None):
+        """
+        add distributed var in this.
+
+        Args:
+            origin_var(Variable|VarStruct): origin var properties
+            slice_var(Variable|VarStruct): slice var properties
+            is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard.
+            block_id(int|None): the number about the slice var.
+            offset(int|None): if the slice var is sliced, offset is the numel before the var.
+            vtype(str|None): a tag, such as Optimizer/Param/RemoteProfetch.
+            endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001"
+        Returns:
+            None
+        """
+        self.distributed_vars.append(
+            VarDistributed(origin_var, slice_var, is_slice, block_id, offset,
+                           vtype, endpoint))
+
+    def get_distributed_var_by_slice(self, var_name):
+        """
+        get distributed var by conditions.
+
+        Args:
+            var_name(str): slice var name, such as "w.traier0.block1"
+        Returns:
+            VarDistributed: distributed var.
+        """
+        for dist_var in self.distributed_vars:
+            if dist_var.slice.name == var_name:
+                return dist_var
+        return None
+
+    @staticmethod
+    def equal(var1, var2):
+        """
+        the two var is equal or not.
+        Returns:
+            bool: equal will return True else False
+        """
+        return var1.name == var2.name and \
+               var1.type == var2.type and \
+               var1.shape == var2.shape and \
+               var1.dtype == var2.dtype and \
+               var1.lod_level == var2.lod_level and \
+               var1.persistable == var2.persistable
+
+    def get_distributed_var_by_origin_and_ep(self, origin_var_name, endpoint):
+        """
+        get distributed var by conditions.
+
+        Args:
+            origin_var_name(str):
+            endpoint(str): the parameter endpoint, such as "127.0.0.1:1001"
+        Returns:
+            VarDistributed: distributed var.
+        """
+        for dist_var in self.distributed_vars:
+            if dist_var.origin.name == origin_var_name and dist_var.endpoint == endpoint:
+                return dist_var
+        return None
+
+    def get_distributed_vars_by_vtypes(self, vtypes, groupby=False):
+        """
+        get distributed vars by conditions.
+
+        Args:
+            vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch"
+            groupby(bool|False): group by origin var or not.
+
+        Returns:
+            list: distributed var list.
+            dict: distributed var map when groupby=True
+        """
+        vtype_vars = []
+        for var in self.distributed_vars:
+            if var.vtype in vtypes:
+                vtype_vars.append(var)
+        if not groupby:
+            return vtype_vars
+
+        params_map = {}
+        for var in vtype_vars:
+            origin_var_name = var.origin.name
+
+            if origin_var_name in params_map.keys():
+                optimizers = params_map.get(origin_var_name)
+            else:
+                optimizers = []
+            optimizers.append(var)
+            params_map[origin_var_name] = optimizers
+        return params_map
+
+    def get_distributed_vars_by_ep(self, endpoint, vtype=None):
+        """
+        get distributed vars by conditions.
+
+        Args:
+            endpoint(str): the parameter server endpoint, such as "127.0.0.1:2001"
+            vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch"
+
+        Returns:
+            list: distributed var list.
+        """
+        endpoint_vars = []
+        for var in self.distributed_vars:
+            if var.endpoint == endpoint:
+                endpoint_vars.append(var)
+        if not vtype:
+            return endpoint_vars
+
+        vtype_vars = []
+        for var in endpoint_vars:
+            if var.vtype == vtype:
+                vtype_vars.append(var)
+        return vtype_vars
+
+    def overview(self):
+        """
+        get the overview string about all params on all parameter servers.
+
+        Returns:
+            Str: overview string.
+
+        """
+        vars_str = []
+        for var in self.distributed_vars:
+            vars_str.append(str(var))
+        return "\n".join(vars_str)

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -30,19 +30,23 @@ Steps to transpile pserver:
 5. add listen_and_serv op
 """
 
+import sys
 import math
-import numpy as np
+from functools import reduce
+
 import collections
+import six
 import logging
 
+import numpy as np
+
 from .ps_dispatcher import RoundRobin, PSDispatcher
 from .. import core, framework, unique_name
 from ..framework import Program, default_main_program, \
-    default_startup_program, Block, \
-    Parameter, Variable, grad_var_name
-from .details import *
+    default_startup_program, Block, Parameter, grad_var_name
+from .details import wait_server_ready, UnionFind, VarStruct, VarsDistributed
+from .details import delete_ops, find_op_by_output_arg
 from ..distribute_lookup_table import find_distributed_lookup_table
-from functools import reduce
 
 LOOKUP_TABLE_TYPE = "lookup_table"
 LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
@@ -62,260 +66,6 @@ def log(*args):
         print(args)
 
 
-class VarStruct(object):
-    """
-    record part properties of a Variable in python.
-    """
-
-    def __init__(self, name, shape, dtype, type, lod_level, persistable):
-        self.name = name
-        self.shape = shape
-        self.dtype = dtype
-        self.type = type
-        self.lod_level = lod_level
-        self.persistable = persistable
-
-
-class VarDistributed(object):
-    """
-    a class to record the var distributed on parameter servers.
-    the class will record the relationship between origin var and slice var.
-    the slice var's properties, such as type/shape/offset/endpoint.
-    """
-
-    def __init__(self,
-                 origin_var,
-                 slice_var,
-                 is_slice=None,
-                 block_id=None,
-                 offset=None,
-                 vtype=None,
-                 endpoint=None):
-        """
-        Args:
-            origin_var(Variable|VarStruct): origin var properties
-            slice_var(Variable|VarStruct): slice var properties
-            is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard.
-            block_id(int|None): the number about the slice var.
-            offset(int|None): if the slice var is sliced, offset is the numel before the var.
-            vtype(str|None): a tag, such as Optimizer/Param/RemoteProfetch.
-            endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001"
-        """
-
-        if isinstance(origin_var, Variable):
-            self.origin = self.__create_var_struct(origin_var)
-        else:
-            self.origin = origin_var
-
-        if isinstance(slice_var, Variable):
-            self.slice = self.__create_var_struct(slice_var)
-        else:
-            self.slice = slice_var
-
-        if self.equal(self.origin, self.slice):
-            self.is_slice = False
-            self.block_id = 0
-            self.offset = 0
-        else:
-            self.is_slice = True
-            self.block_id = 0
-            self.offset = 0
-
-        if is_slice is not None:
-            self.is_slice = is_slice
-        if block_id is not None:
-            self.block_id = block_id
-        if offset is not None:
-            self.offset = offset
-
-        self.vtype = vtype
-        self.endpoint = endpoint
-
-    @staticmethod
-    def __create_var_struct(var):
-        return VarStruct(var.name, var.shape, var.dtype, var.type,
-                         var.lod_level, var.persistable)
-
-    @staticmethod
-    def equal(var1, var2):
-        """
-        the two var is equal or not.
-        Returns:
-            bool: equal will return True else False
-        """
-        assert isinstance(var1, VarStruct) and isinstance(var2, VarStruct)
-
-        return var1.name == var2.name and \
-               var1.type == var2.type and \
-               var1.shape == var2.shape and \
-               var1.dtype == var2.dtype and \
-               var1.lod_level == var2.lod_level and \
-               var1.persistable == var2.persistable
-
-    def __str__(self):
-        origin_var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})". \
-            format(i="{", e="}", name=self.origin.name, type=self.origin.type,
-                   shape=self.origin.shape, dtype=self.origin.dtype)
-
-        slice_var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})" \
-                        ".slice({is_slice}).block({block_id}).offset({offset})". \
-            format(i="{", e="}", name=self.slice.name, type=self.slice.type,
-                   shape=self.slice.shape, dtype=self.slice.dtype,
-                   is_slice=self.is_slice, block_id=self.block_id, offset=self.offset)
-
-        return "var owned: {}, origin var: ( {} ), slice var: ( {} ), endpoint: {} ".format(
-            self.vtype, origin_var_str, slice_var_str, self.endpoint)
-
-
-class VarsDistributed(object):
-    """
-    a gather about VarDistributed with many methods to find distributed vars.
-    through the class, we can get overview about the distributed parameters on parameter servers.
-    this class may centralized and convenient for developer to manage and get variable's distribute.
-    other module can also use this to find variables such io.py.
-    """
-
-    def __init__(self):
-        self.distributed_vars = []
-
-    def add_distributed_var(self,
-                            origin_var,
-                            slice_var,
-                            is_slice=None,
-                            block_id=None,
-                            offset=None,
-                            vtype=None,
-                            endpoint=None):
-        """
-        add distributed var in this.
-
-        Args:
-            origin_var(Variable|VarStruct): origin var properties
-            slice_var(Variable|VarStruct): slice var properties
-            is_slice(bool|None): slice or not, slice_var=True/False and its block size > 8192 are the judgement standard.
-            block_id(int|None): the number about the slice var.
-            offset(int|None): if the slice var is sliced, offset is the numel before the var.
-            vtype(str|None): a tag, such as Optimizer/Param/RemoteProfetch.
-            endpoint(str|None): which parameter the slice var on, such as "127.0.0.1:1001"
-        Returns:
-            None
-        """
-        self.distributed_vars.append(
-            VarDistributed(origin_var, slice_var, is_slice, block_id, offset,
-                           vtype, endpoint))
-
-    def get_distributed_var_by_slice(self, var_name):
-        """
-        get distributed var by conditions.
-
-        Args:
-            var_name(str): slice var name, such as "w.traier0.block1"
-        Returns:
-            VarDistributed: distributed var.
-        """
-        for dist_var in self.distributed_vars:
-            if dist_var.slice.name == var_name:
-                return dist_var
-        return None
-
-    @staticmethod
-    def equal(var1, var2):
-        """
-        the two var is equal or not.
-        Returns:
-            bool: equal will return True else False
-        """
-        return var1.name == var2.name and \
-               var1.type == var2.type and \
-               var1.shape == var2.shape and \
-               var1.dtype == var2.dtype and \
-               var1.lod_level == var2.lod_level and \
-               var1.persistable == var2.persistable
-
-    def get_distributed_var_by_origin_and_ep(self, origin_var_name, endpoint):
-        """
-        get distributed var by conditions.
-
-        Args:
-            origin_var_name(str):
-            endpoint(str): the parameter endpoint, such as "127.0.0.1:1001"
-        Returns:
-            VarDistributed: distributed var.
-        """
-        for dist_var in self.distributed_vars:
-            if dist_var.origin.name == origin_var_name and dist_var.endpoint == endpoint:
-                return dist_var
-        return None
-
-    def get_distributed_vars_by_vtypes(self, vtypes, groupby=False):
-        """
-        get distributed vars by conditions.
-
-        Args:
-            vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch"
-            groupby(bool|False): group by origin var or not.
-
-        Returns:
-            list: distributed var list.
-            dict: distributed var map when groupby=True
-        """
-        vtype_vars = []
-        for var in self.distributed_vars:
-            if var.vtype in vtypes:
-                vtype_vars.append(var)
-        if not groupby:
-            return vtype_vars
-
-        params_map = {}
-        for var in vtype_vars:
-            origin_var_name = var.origin.name
-
-            if origin_var_name in params_map.keys():
-                optimizers = params_map.get(origin_var_name)
-            else:
-                optimizers = []
-            optimizers.append(var)
-            params_map[origin_var_name] = optimizers
-        return params_map
-
-    def get_distributed_vars_by_ep(self, endpoint, vtype=None):
-        """
-        get distributed vars by conditions.
-
-        Args:
-            endpoint(str): the parameter server endpoint, such as "127.0.0.1:2001"
-            vtype(str|None): distributed var's vtype, such as "Optimizer", "RemotePrefetch"
-
-        Returns:
-            list: distributed var list.
-        """
-        endpoint_vars = []
-        for var in self.distributed_vars:
-            if var.endpoint == endpoint:
-                endpoint_vars.append(var)
-        if not vtype:
-            return endpoint_vars
-
-        vtype_vars = []
-        for var in endpoint_vars:
-            if var.vtype == vtype:
-                vtype_vars.append(var)
-        return vtype_vars
-
-    def overview(self):
-        """
-        get the overview string about all params on all parameter servers.
-
-        Returns:
-            Str: overview string.
-
-        """
-        vars_str = []
-        for var in self.distributed_vars:
-            vars_str.append(str(var))
-        return "\n".join(vars_str)
-
-
 class VarBlock:
     def __init__(self, varname, offset, size):
         self.varname = varname