Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into unsqueeze_op

Dockerfile

@@ -23,7 +23,7 @@ ENV HOME /root
 COPY ./paddle/scripts/docker/root/ /root/
 
 RUN apt-get update && \
-    apt-get install -y --allow-downgrades \
+    apt-get install -y --allow-downgrades patchelf \
     git python-pip python-dev python-opencv openssh-server bison \
     libnccl2=2.1.2-1+cuda8.0 libnccl-dev=2.1.2-1+cuda8.0 \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \

cmake/generic.cmake

@@ -257,8 +257,8 @@ function(cc_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS ARGS)
     cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_executable(${TARGET_NAME} ${cc_test_SRCS})
-    target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog)
-    add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main memory gtest gflags glog)
+    target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
+    add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     add_test(NAME ${TARGET_NAME}
              COMMAND ${TARGET_NAME} ${cc_test_ARGS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
@@ -324,8 +324,8 @@ function(nv_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS)
     cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
-    target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
-    add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main memory gtest gflags glog)
+    target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
+    add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     add_test(${TARGET_NAME} ${TARGET_NAME})
     if (nv_test_SERIAL)
         set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)

doc/fluid/api/transpiler.rst

@@ -14,6 +14,15 @@ DistributeTranspiler
     :members:
     :noindex:
 
+.. _api_fluid_transpiler_InferenceTranspiler:
+
+InferenceTranspiler
+-------------------
+
+..  autoclass:: paddle.fluid.transpiler.InferenceTranspiler
+    :members:
+    :noindex:
+
 .. _api_fluid_transpiler_memory_optimize:
 
 memory_optimize

doc/fluid/design/dist_train/dist_train_nccl2.md

@@ -0,0 +1,35 @@
+# Distributed Training with NCCL2
+
+We design a pattern that can enable training with `ParallelExecutor` and
+using [NCCL2](https://developer.nvidia.com/nccl) as it's collective
+communication library.
+
+In `ParallelExecutor` we can use `AllReduce` or `Reduce` and `Broadcast`
+to do multi GPU training. And if we initialize NCCL2 communicators as
+ranks in a distributed environment, we can simply run the `ParallelExecutor`
+as a distributed program! The only thing that may be different than in
+the single node version is that we need to broadcast the NCCL unique ID
+to all the nodes, and initialize communicators using that ID, so NCCL2
+will know each other as ranks.
+
+To achieve this feature, we introduce a new operator: `gen_nccl_id` op,
+so we are ***not*** "bind to" running NCCL2 with MPI, we can run it in
+what ever platform you like.
+
+It have two running modes:
+
+1. Generate and broadcast mode, which should be used on trainer 0;
+1. Listen and fetch mode, which should be used on trainers other than 0.
+
+In both two modes, this op can save the NCCL ID into current scope as a
+persistable variable, Then we can insert this op at the end of
+"startup program" of fluid, so that all workers can get the same ID to
+initialize NCCL communicator objects.
+
+<img src="src/ncc2_design.png">
+
+The above figure indicates the general process when training with NCCL2
+distributed. Each trainer have the number of communicators equal to the
+number of GPUs, but the ranks should match the global ranks number: here
+we have total 8 GPUs, so `nranks==8`, for each trainer, the ranks should
+be from 0 ~ 3 on trainer 0 and 4 ~ 7 on trainer 1.

doc/fluid/design/dist_train/distributed_lookup_table_design.md

@@ -119,6 +119,32 @@ optimization algorithm $f$ runs on the storage service.
 - Con: the storage service needs to be able to run the optimization
   algorithm.
 
+## Distributed Sparse Table in Fluid
+
+For another design, we can implement a distributed sparse table in Fluid,
+and don't need to maintain an external storage component while training.
+
+You may need to read Fluid [Distributed Training Architecture](./distributed_architecture.md)
+and [Parameter Server](./parameter_server.md) before going on.
+
+![fluid lookup remote table](./src/fluid_lookup_remote_table.png)
+
+Partition a large table into multiple pserver instances
+1. `DistributeTranspiler` would split the table partitioned into some small
+table blocks with some partitioned algorithms such as
+[RoundRobin](https://en.wikipedia.org/wiki/Round-robin_scheduling),
+[Hash](https://en.wikipedia.org/wiki/Hash) and etc...
+1. For some cases, the range of input `Ids` is very wide and unpredictable, so the sparse
+table would be able to fill a new value for the id that didn't appear before with
+zero, uniform random or Gaussian distribution.
+
+For each Trainer's training process:
+1. In the forward pass, we use `pre-fetch` op to pre-fetch parameter blocks according to the
+input `Ids` from PServers instead of the local `lookup_table` op, and then merge the blocks
+into a parameter `W`.
+1. Compute `GRAD@W'` in the backward pass using the pre-fetched `W` and send it to PServer to
+execute the optimize pass.
+
 ## Conclusion
 
 Let us do the "storage service does not optimize" solution first, as a

paddle/contrib/inference/paddle_inference_api_impl.h

@@ -22,9 +22,9 @@
 #include "paddle/contrib/inference/paddle_inference_api.h"
 
 #include "paddle/fluid/framework/ddim.h"
-#include "paddle/fluid/framework/init.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/io.h"
+#include "paddle/fluid/platform/init.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {

paddle/fluid/framework/CMakeLists.txt

@@ -21,10 +21,10 @@ endif()
 
 cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
 
-nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place memory device_context init)
+nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place memory device_context tensor)
 cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio)
 cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
-nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor init)
+nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
 
 cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
 
@@ -38,7 +38,7 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
 cc_library(data_device_transform SRCS data_device_transform.cc DEPS tensor)
 nv_test(data_device_transform_test SRCS data_device_transform_test.cu
-        DEPS operator op_registry init math_function)
+        DEPS operator op_registry device_context math_function)
 
 if(WITH_GPU)
   nv_library(data_type_transform SRCS data_type_transform.cu DEPS tensor)
@@ -63,7 +63,7 @@ cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
     shape_inference data_transform lod_tensor profiler)
-cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry init)
+cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog)
 
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
@@ -101,14 +101,14 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
-cc_library(init SRCS init.cc DEPS gflags device_context place stringpiece operator)
-cc_test(init_test SRCS init_test.cc DEPS init)
-
 cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
       
 # cc_test(channel_test SRCS channel_test.cc)
 cc_test(tuple_test SRCS tuple_test.cc )
-cc_test(concurrency_test SRCS concurrency_test.cc DEPS go_op channel_close_op channel_create_op
-        channel_send_op channel_recv_op sum_op select_op elementwise_add_op compare_op
-        conditional_block_op while_op assign_op print_op executor proto_desc)
+
+# disable test temporarily.
+# TODO https://github.com/PaddlePaddle/Paddle/issues/11971
+# cc_test(concurrency_test SRCS concurrency_test.cc DEPS go_op channel_close_op channel_create_op
+#         channel_send_op channel_recv_op sum_op select_op elementwise_add_op compare_op
+#         conditional_block_op while_op assign_op print_op executor proto_desc)

paddle/fluid/framework/data_device_transform_test.cu

@@ -14,13 +14,13 @@ limitations under the License. */
 
 #include "gtest/gtest.h"
 
-#include "paddle/fluid/framework/init.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/elementwise_op_function.h"
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/platform/init.h"
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/lod_tensor_test.cu

@@ -17,9 +17,9 @@
 #include <stdio.h>
 
 #include "gtest/gtest.h"
-#include "paddle/fluid/framework/init.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/platform/assert.h"
+#include "paddle/fluid/platform/init.h"
 #include "paddle/fluid/platform/place.h"
 
 __global__ void test(size_t* a, int size) {

paddle/fluid/framework/operator_test.cc

@@ -13,10 +13,10 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #include "gtest/gtest.h"
 
-#include "paddle/fluid/framework/init.h"
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/platform/init.h"
 
 namespace paddle {
 namespace framework {

paddle/fluid/inference/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor init)
+set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor )
 
 # TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal?
 cc_library(paddle_fluid_api

paddle/fluid/inference/analysis/README.md

@@ -54,4 +54,5 @@ It can be used as a helper class that draws the modified graph after each pass.
 There is some helper legacy/function/class for analysis.
 
 - [dot.h](./dot.h) give a easy to use interface for generating `DOT` codes,
-- [graph_traits.h](./graph_traits.h) contains the graph traversal algorithms, it uses `iterator` to make the algorithms easy to share across different passes.
+- [graph_traits.h](./graph_traits.h) contains the interfaces of the graph traversal algorithms, it uses `iterator`to make the algorithms easy to share across different passes,
+there are some implementations in  [data_flow_graph.cc](./data_flow_graph.cc) , such as BFS and DFS..

paddle/fluid/inference/analysis/pass.h

@@ -32,19 +32,6 @@ class Pass {
  public:
   Pass() = default;
   virtual ~Pass() = default;
-  // Virtual method overridden by subclasses to do only necessary initialization
-  // before any pass is run.
-  // virtual bool Initialize() { return false; }
-  // There is some passes such as FlowToDataFlowGraphPass that needs a
-  // ProgramDesc. Here use the native ProgramDesc ProtoBuf message, so that it
-  // only couple with the proto file.
-  // virtual bool Initialize(const framework::proto::ProgramDesc &desc) { return
-  // false; }
-  // There are some Passes such as DataFlowGraphToFluidPass that will output a
-  // ProgramDesc.
-  // virtual bool Initialize(framework::proto::ProgramDesc *desc) { return
-  // false; }
-
   // Mutable Pass.
   virtual bool Initialize(Argument *argument) { return false; }
   // Readonly Pass.

paddle/fluid/inference/io.cc

@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/feed_fetch_type.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/platform/cpu_helper.h"
 #include "paddle/fluid/pybind/pybind.h"
 
 DEFINE_string(devices, "", "The devices to be used which is joined by comma.");
@@ -33,7 +33,7 @@ namespace inference {
 
 void Init(const std::vector<std::string> argv) {
   framework::InitGflags(argv);
-  operators::math::SetNumThreads(FLAGS_math_num_threads);
+  platform::SetNumThreads(FLAGS_math_num_threads);
   // init devices
   std::vector<int> devices;
   std::string token;

paddle/fluid/inference/io.h

@@ -18,9 +18,9 @@ limitations under the License. */
 #include <string>
 #include <vector>
 #include "paddle/fluid/framework/executor.h"
-#include "paddle/fluid/framework/init.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/platform/init.h"
 
 namespace paddle {
 namespace inference {

paddle/fluid/inference/tests/book/test_inference_nlp.cc

@@ -19,7 +19,7 @@ limitations under the License. */
 #include "gflags/gflags.h"
 #include "gtest/gtest.h"
 #include "paddle/fluid/inference/tests/test_helper.h"
-#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/platform/cpu_helper.h"
 #ifdef PADDLE_WITH_MKLML
 #include <omp.h>
 #endif
@@ -164,7 +164,7 @@ TEST(inference, nlp) {
   // only use 1 thread number per std::thread
   omp_set_dynamic(0);
   omp_set_num_threads(1);
-  paddle::operators::math::SetNumThreads(1);
+  paddle::platform::SetNumThreads(1);
 #endif
 
   double start_ms = 0, stop_ms = 0;

paddle/fluid/operators/distributed/CMakeLists.txt

@@ -5,7 +5,7 @@ if(WITH_GRPC)
   set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
   set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
   cc_test(serde_test SRCS grpc_serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr
-          cares zlib protobuf sendrecvop_grpc SERIAL)
+          cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL)
   cc_test(grpc_server_test SRCS rpc_server_test.cc DEPS sendrecvop_grpc
           grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor
           proto_desc lookup_table_op SERIAL)

paddle/fluid/operators/math/CMakeLists.txt

@@ -54,13 +54,13 @@ math_library(softmax DEPS math_function)
 math_library(unpooling)
 math_library(vol2col)
 
-cc_test(math_function_test SRCS math_function_test.cc)
+cc_test(math_function_test SRCS math_function_test.cc DEPS math_function)
 cc_test(selected_rows_functor_test SRCS selected_rows_functor_test.cc DEPS selected_rows_functor)
 cc_test(im2col_test SRCS im2col_test.cc DEPS im2col)
 cc_test(vol2col_test SRCS vol2col_test.cc DEPS vol2col)
 cc_test(sequence_padding_test SRCS sequence_padding_test.cc DEPS sequence_padding)
 if(WITH_GPU)
-    nv_test(math_function_gpu_test SRCS math_function_test.cu)
-    nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor)
+    nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function)
+    nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor math_function)
 endif()
 cc_test(concat_test SRCS concat_test.cc DEPS concat)

paddle/fluid/operators/math/blas.h

@@ -23,41 +23,12 @@
 
 #ifdef PADDLE_USE_OPENBLAS
 #include <cblas.h>
-#ifdef LAPACK_FOUND
-#include <lapacke.h>
-#endif
-#endif
-
-#ifndef LAPACK_FOUND
-extern "C" {
-#include <cblas.h>  // NOLINT
-int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda,
-                   int* ipiv);
-int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda,
-                   int* ipiv);
-int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda,
-                   const int* ipiv);
-int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda,
-                   const int* ipiv);
-}
 #endif
 
 namespace paddle {
 namespace operators {
 namespace math {
 
-static void SetNumThreads(int num_threads) {
-#ifdef PADDLE_USE_OPENBLAS
-  int real_num_threads = num_threads > 1 ? num_threads : 1;
-  openblas_set_num_threads(real_num_threads);
-#elif defined(PADDLE_WITH_MKLML)
-  int real_num_threads = num_threads > 1 ? num_threads : 1;
-  platform::dynload::MKL_Set_Num_Threads(real_num_threads);
-#else
-  PADDLE_ENFORCE(false, "To be implemented.");
-#endif
-}
-
 /**
  * Matrix Descriptor of a memory buffer.
  *

paddle/fluid/operators/math/math_function.h

@@ -19,23 +19,6 @@ limitations under the License. */
 
 #ifdef PADDLE_USE_OPENBLAS
 #include <cblas.h>
-#ifdef LAPACK_FOUND
-#include <lapacke.h>
-#endif
-#endif
-
-#ifndef LAPACK_FOUND
-extern "C" {
-#include <cblas.h>  // NOLINT
-int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda,
-                   int* ipiv);
-int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda,
-                   int* ipiv);
-int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda,
-                   const int* ipiv);
-int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda,
-                   const int* ipiv);
-}
 #endif
 
 #include <cmath>

paddle/fluid/operators/nccl_op_test.cu.cc

@@ -19,7 +19,6 @@ limitations under the License. */
 #include <thread>  // NOLINT
 #include <vector>
 
-#include "paddle/fluid/framework/init.h"
 #include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/program_desc.h"
@@ -27,6 +26,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/gpu_info.h"
+#include "paddle/fluid/platform/init.h"
 #include "paddle/fluid/platform/place.h"
 
 USE_NO_KERNEL_OP(ncclInit);