Merge branch 'develop' into resnet50_ut

paddle/fluid/API.spec

@@ -145,14 +145,14 @@ paddle.fluid.layers.unstack ArgSpec(args=['x', 'axis', 'num'], varargs=None, key
 paddle.fluid.layers.sequence_enumerate ArgSpec(args=['input', 'win_size', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0, None))
 paddle.fluid.layers.expand ArgSpec(args=['x', 'expand_times', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.sequence_concat ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.scale ArgSpec(args=['x', 'scale', 'bias', 'bias_after_scale', 'out', 'act', 'name'], varargs=None, keywords=None, defaults=(1.0, 0.0, True, None, None, None))
+paddle.fluid.layers.scale ArgSpec(args=['x', 'scale', 'bias', 'bias_after_scale', 'act', 'name'], varargs=None, keywords=None, defaults=(1.0, 0.0, True, None, None))
-paddle.fluid.layers.elementwise_add ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
+paddle.fluid.layers.elementwise_add ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
-paddle.fluid.layers.elementwise_div ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
+paddle.fluid.layers.elementwise_div ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
-paddle.fluid.layers.elementwise_sub ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
+paddle.fluid.layers.elementwise_sub ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
-paddle.fluid.layers.elementwise_mul ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
+paddle.fluid.layers.elementwise_mul ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
-paddle.fluid.layers.elementwise_max ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
+paddle.fluid.layers.elementwise_max ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
-paddle.fluid.layers.elementwise_min ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
+paddle.fluid.layers.elementwise_min ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
-paddle.fluid.layers.elementwise_pow ArgSpec(args=['x', 'y', 'out', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, -1, False, None, None))
+paddle.fluid.layers.elementwise_pow ArgSpec(args=['x', 'y', 'axis', 'use_mkldnn', 'act', 'name'], varargs=None, keywords=None, defaults=(-1, False, None, None))
 paddle.fluid.layers.uniform_random_batch_size_like ArgSpec(args=['input', 'shape', 'dtype', 'input_dim_idx', 'output_dim_idx', 'min', 'max', 'seed'], varargs=None, keywords=None, defaults=('float32', 0, 0, -1.0, 1.0, 0))
 paddle.fluid.layers.gaussian_random ArgSpec(args=['shape', 'mean', 'std', 'seed', 'dtype', 'use_mkldnn'], varargs=None, keywords=None, defaults=(0.0, 1.0, 0, 'float32', False))
 paddle.fluid.layers.sampling_id ArgSpec(args=['x', 'min', 'max', 'seed', 'dtype'], varargs=None, keywords=None, defaults=(0.0, 1.0, 0, 'float32'))
@@ -160,6 +160,12 @@ paddle.fluid.layers.gaussian_random_batch_size_like ArgSpec(args=['input', 'shap
 paddle.fluid.layers.sum ArgSpec(args=['x', 'use_mkldnn'], varargs=None, keywords=None, defaults=(False,))
 paddle.fluid.layers.slice ArgSpec(args=['input', 'axes', 'starts', 'ends'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.shape ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.layers.logical_and ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.logical_or ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.logical_xor ArgSpec(args=['x', 'y', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.logical_not ArgSpec(args=['x', 'out', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.clip ArgSpec(args=['x', 'min', 'max', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.clip_by_norm ArgSpec(args=['x', 'max_norm', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
 paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
 paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
@@ -225,12 +231,6 @@ paddle.fluid.layers.is_empty ArgSpec(args=['x', 'cond'], varargs=None, keywords=
 paddle.fluid.layers.mean ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.mul ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.clip ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.clip_by_norm ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.logical_and ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.logical_or ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.logical_xor ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
-paddle.fluid.layers.logical_not ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.maxout ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None)
 paddle.fluid.layers.sigmoid ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.logsigmoid ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))

paddle/fluid/framework/scope.cc

@@ -20,13 +20,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/string/printf.h"
 
-// The mutex is not needed by training and inference, only for distribution.
-#if PADDLE_WITH_DISTRIBUTE
-#define WITH_LOCK 1
-#else
-#define WITH_LOCK 0
-#endif
-
 DEFINE_bool(benchmark, false,
             "Doing memory benchmark. It will make deleting scope synchronized, "
             "and add some memory usage logs."
@@ -56,24 +49,18 @@ int64_t GetEagerDeletionThreshold() {
 Scope::~Scope() { DropKids(); }
 
 Scope& Scope::NewScope() const {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   kids_.push_back(new Scope(this));
   return *kids_.back();
 }
 
 Variable* Scope::Var(const std::string& name) {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   return VarInternal(name);
 }
 
 Variable* Scope::Var(std::string* name) {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   auto new_name = string::Sprintf("%p.%d", this, vars_.size());
   if (name != nullptr) {
     *name = new_name;
@@ -82,39 +69,29 @@ Variable* Scope::Var(std::string* name) {
 }
 
 Variable* Scope::FindVar(const std::string& name) const {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   return FindVarInternal(name);
 }
 
 const Scope* Scope::FindScope(const Variable* var) const {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   return FindScopeInternal(var);
 }
 
 void Scope::DropKids() {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   for (Scope* s : kids_) delete s;
   kids_.clear();
 }
 
 bool Scope::HasKid(const Scope* scope) const {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   return it != this->kids_.end();
 }
 
 std::vector<std::string> Scope::LocalVarNames() const {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   std::vector<std::string> known_vars;
   known_vars.reserve(this->vars_.size());
   for (auto& p : vars_) {
@@ -124,9 +101,7 @@ std::vector<std::string> Scope::LocalVarNames() const {
 }
 
 void Scope::DeleteScope(Scope* scope) const {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   auto it = std::find(this->kids_.begin(), this->kids_.end(), scope);
   PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope);
   this->kids_.erase(it);
@@ -139,9 +114,7 @@ void Scope::DeleteScope(Scope* scope) const {
 }
 
 void Scope::EraseVars(const std::vector<std::string>& var_names) {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   std::set<std::string> var_set(var_names.begin(), var_names.end());
   for (auto it = vars_.begin(); it != vars_.end();) {
     if (var_set.find(it->first) != var_set.end()) {
@@ -154,16 +127,12 @@ void Scope::EraseVars(const std::vector<std::string>& var_names) {
 
 void Scope::Rename(const std::string& origin_name,
                    const std::string& new_name) const {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   RenameInternal(origin_name, new_name);
 }
 
 std::string Scope::Rename(const std::string& origin_name) const {
-#if WITH_LOCK
   std::unique_lock<std::mutex> lock(mutex_);
-#endif
   auto new_name = string::Sprintf("%p.%d", this, vars_.size());
   RenameInternal(origin_name, new_name);
   return new_name;

paddle/fluid/platform/dynload/cublas.h

@@ -55,7 +55,7 @@ extern void *cublas_dso_handle;
   struct DynLoad__##__name {                         \
     template <typename... Args>                      \
     inline cublasStatus_t operator()(Args... args) { \
-      return __name(args...);                        \
+      return ::__name(args...);                      \
     }                                                \
   };                                                 \
   extern DynLoad__##__name __name

paddle/fluid/platform/dynload/cudnn.h

@@ -13,6 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#define GLOG_NO_ABBREVIATED_SEVERITIES
+#define GOOGLE_GLOG_DLL_DECL
+#include <glog/logging.h>
 
 #include <cudnn.h>
 #include <mutex>  // NOLINT
@@ -47,13 +50,13 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
 
 #else
 
-#define DECLARE_DYNAMIC_LOAD_CUDNN_WRAP(__name)                  \
+#define DECLARE_DYNAMIC_LOAD_CUDNN_WRAP(__name)     \
-  struct DynLoad__##__name {                                     \
+  struct DynLoad__##__name {                        \
-    template <typename... Args>                                  \
+    template <typename... Args>                     \
-    auto operator()(Args... args) -> decltype(__name(args...)) { \
+    inline cudnnStatus_t operator()(Args... args) { \
-      return __name(args...);                                    \
+      return ::__name(args...);                     \
-    }                                                            \
+    }                                               \
-  };                                                             \
+  };                                                \
   extern DynLoad__##__name __name
 
 #endif

paddle/fluid/platform/dynload/curand.h

@@ -44,7 +44,7 @@ extern void *curand_dso_handle;
   struct DynLoad__##__name {                     \
     template <typename... Args>                  \
     curandStatus_t operator()(Args... args) {    \
-      return __name(args...);                    \
+      return ::__name(args...);                  \
     }                                            \
   };                                             \
   extern DynLoad__##__name __name

paddle/fluid/platform/dynload/dynamic_loader.cc

@@ -107,7 +107,11 @@ static inline void* GetDsoHandleFromDefaultPath(const std::string& dso_path,
 static inline void* GetDsoHandleFromSearchPath(const std::string& search_root,
                                                const std::string& dso_name,
                                                bool throw_on_error = true) {
+#if !defined(_WIN32)
   int dynload_flags = RTLD_LAZY | RTLD_LOCAL;
+#else
+  int dynload_flags = 0;
+#endif  // !_WIN32
   void* dso_handle = nullptr;
 
   std::string dlPath = dso_name;
@@ -117,10 +121,15 @@ static inline void* GetDsoHandleFromSearchPath(const std::string& search_root,
     // search xxx.so from custom path
     dlPath = join(search_root, dso_name);
     dso_handle = dlopen(dlPath.c_str(), dynload_flags);
+#if !defined(_WIN32)
+    auto errorno = dlerror();
+#else
+    auto errorno = GetLastError();
+#endif  // !_WIN32
     // if not found, search from default path
     if (nullptr == dso_handle) {
       LOG(WARNING) << "Failed to find dynamic library: " << dlPath << " ("
-                   << dlerror() << ")";
+                   << errorno << ")";
       if (dlPath.find("nccl") != std::string::npos) {
         std::cout
             << "You may need to install 'nccl2' from NVIDIA official website: "
@@ -139,10 +148,15 @@ static inline void* GetDsoHandleFromSearchPath(const std::string& search_root,
       "export LD_LIBRARY_PATH=... \n Note: After Mac OS 10.11, "
       "using the DYLD_LIBRARY_PATH is impossible unless System "
       "Integrity Protection (SIP) is disabled.";
+#if !defined(_WIN32)
+  auto errorno = dlerror();
+#else
+  auto errorno = GetLastError();
+#endif  // !_WIN32
   if (throw_on_error) {
-    PADDLE_ENFORCE(nullptr != dso_handle, error_msg, dlPath, dlerror());
+    PADDLE_ENFORCE(nullptr != dso_handle, error_msg, dlPath, errorno);
   } else if (nullptr == dso_handle) {
-    LOG(WARNING) << string::Sprintf(error_msg, dlPath, dlerror());
+    LOG(WARNING) << string::Sprintf(error_msg, dlPath, errorno);
   }
 
   return dso_handle;

paddle/scripts/paddle_build.sh

@@ -395,7 +395,7 @@ EOF
         ctest --output-on-failure -j $1     
         # make install should also be test when unittest 
         make install -j 8
-        pip install /usr/local/opt/paddle/share/wheels/*.whl
+        pip install ${INSTALL_PREFIX:-/paddle/build}/opt/paddle/share/wheels/*.whl
         if [[ ${WITH_FLUID_ONLY:-OFF} == "OFF" ]] ; then
             paddle version
         fi
@@ -750,7 +750,7 @@ function main() {
         cmake_gen ${PYTHON_ABI:-""}
         build
         run_test
-        assert_api_not_changed
+        assert_api_not_changed ${PYTHON_ABI:-""}
         ;;
       *)
         print_usage

python/paddle/fluid/clip.py

@@ -271,7 +271,8 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
                     "All parameters' 'clip_norm' of a same group should be the same"
                 )
 
-        local_norm_var = layers.reduce_sum(input=layers.pow(x=grad, factor=2.0))
+        square = grad * grad
+        local_norm_var = layers.cast(layers.reduce_sum(input=square), 'float64')
         context[self.group_name].append(local_norm_var)
 
         self.context = context
@@ -281,6 +282,7 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
         if group_scale_name not in self.context:
             group_norm_var = layers.sums(input=self.context[self.group_name])
             group_norm_var = layers.sqrt(x=group_norm_var)
+            group_norm_var = layers.cast(group_norm_var, 'float32')
             clip_var = self.context[self.group_name + "_clip"]
             group_scale_var = layers.elementwise_div(
                 x=clip_var,

python/paddle/fluid/layers/control_flow.py

@@ -21,7 +21,7 @@ from .. import core
 from ..framework import Program, Variable, Operator
 from ..layer_helper import LayerHelper, unique_name
 from ..initializer import force_init_on_cpu
-from .ops import logical_and, logical_not, logical_or
+from .nn import logical_and, logical_not, logical_or
 import numpy
 import warnings
 import six

python/paddle/fluid/layers/ops.py

@@ -39,12 +39,6 @@ __all__ = [
     'mean',
     'mul',
     'sigmoid_cross_entropy_with_logits',
-    'clip',
-    'clip_by_norm',
-    'logical_and',
-    'logical_or',
-    'logical_xor',
-    'logical_not',
     'maxout',
 ]
 

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

@@ -22,7 +22,7 @@ class TestDistSeResneXt2x2(TestDistBase):
         self._sync_mode = True
         self._use_reader_alloc = False
 
-    def test_dist_train(self):
+    def no_test_dist_train(self):
         self.check_with_place("dist_se_resnext.py", delta=100)
 
 
@@ -40,7 +40,7 @@ class TestDistSeResneXt2x2Async(TestDistBase):
         self._sync_mode = False
         self._use_reader_alloc = False
 
-    def test_dist_train(self):
+    def no_test_dist_train(self):
         self.check_with_place("dist_se_resnext.py", delta=100)