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

… del_final_state_pre_name2
PaddlePaddle · Aug 26, 2022 · b516f08 · b516f08
2 parents 2ebf43e + efab2eb
commit b516f08
Show file tree

Hide file tree

Showing 78 changed files with 1,460 additions and 462 deletions.
diff --git a/cmake/external/brpc.cmake b/cmake/external/brpc.cmake
@@ -45,9 +45,8 @@ set(prefix_path
 ExternalProject_Add(
   extern_brpc
   ${EXTERNAL_PROJECT_LOG_ARGS}
-  # TODO(gongwb): change to de newst repo when they changed
-  GIT_REPOSITORY "https://github.com/wangjiawei04/brpc"
-  GIT_TAG "e203afb794caf027da0f1e0776443e7d20c0c28e"
+  GIT_REPOSITORY "https://github.com/apache/incubator-brpc"
+  GIT_TAG 1.2.0
   PREFIX ${BRPC_PREFIX_DIR}
   UPDATE_COMMAND ""
   CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@@ -60,8 +59,8 @@ ExternalProject_Add(
              -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
              -DCMAKE_PREFIX_PATH=${prefix_path}
              -DWITH_GLOG=ON
-             -DIOBUF_WITH_HUGE_BLOCK=ON
-             -DBRPC_WITH_RDMA=${WITH_BRPC_RDMA}
+             -DBUILD_BRPC_TOOLS=ON
+             -DBUILD_SHARED_LIBS=ON
              ${EXTERNAL_OPTIONAL_ARGS}
   LIST_SEPARATOR |
   CMAKE_CACHE_ARGS

diff --git a/paddle/fluid/distributed/collective/CMakeLists.txt b/paddle/fluid/distributed/collective/CMakeLists.txt
@@ -21,6 +21,11 @@ if(WITH_NCCL OR WITH_RCCL)
     DEPS processgroup place enforce collective_helper device_context
          dense_tensor)
   if(WITH_DISTRIBUTE AND WITH_PSCORE)
+    if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 7.0)
+      set(DISTRIBUTE_COMPILE_FLAGS "${DISTRIBUTE_COMPILE_FLAGS} -faligned-new")
+    endif()
+    set_source_files_properties(
+      ProcessGroupHeter.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     cc_library(
       processgroup_heter
       SRCS ProcessGroupHeter.cc NCCLTools.cc Common.cc
@@ -40,6 +45,11 @@ if(WITH_ASCEND_CL)
          phi_api
          eager_api)
   if(WITH_DISTRIBUTE AND WITH_PSCORE)
+    if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 7.0)
+      set(DISTRIBUTE_COMPILE_FLAGS "${DISTRIBUTE_COMPILE_FLAGS} -faligned-new")
+    endif()
+    set_source_files_properties(
+      ProcessGroupHeter.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     cc_library(
       processgroup_heter
       SRCS ProcessGroupHeter.cc HCCLTools.cc Common.cc

diff --git a/paddle/fluid/distributed/ps/service/heter_client.h b/paddle/fluid/distributed/ps/service/heter_client.h
@@ -100,7 +100,9 @@ class HeterClient {
       options.connection_type = "";
       VLOG(4) << "ssl enabled in arm";
 #else
-      options.ssl_options.enable = need_encrypt;
+      if (need_encrypt) {
+        options.mutable_ssl_options();
+      }
 #endif
       client_channels = &peer_switch_channels_;
     } else if (peer_role == PEER_ROLE_IS_WORKER) {

diff --git a/paddle/fluid/distributed/ps/service/heter_server.cc b/paddle/fluid/distributed/ps/service/heter_server.cc
@@ -32,13 +32,8 @@ void HeterServer::StartHeterService(bool neeed_encrypt) {
   server_.AddService(&service_, brpc::SERVER_DOESNT_OWN_SERVICE);
   brpc::ServerOptions options;
   if (neeed_encrypt) {
-#ifdef PADDLE_WITH_ARM_BRPC
     options.mutable_ssl_options()->default_cert.certificate = "/cert.pem";
     options.mutable_ssl_options()->default_cert.private_key = "/key.pem";
-#else
-    options.ssl_options.default_cert.certificate = "/cert.pem";
-    options.ssl_options.default_cert.private_key = "/key.pem";
-#endif
   }
   if (server_.Start(endpoint_.c_str(), &options) != 0) {
     VLOG(0) << "HeterServer start fail. Try again.";
@@ -72,13 +67,8 @@ void HeterServer::StartHeterInterService(bool neeed_encrypt) {
   server_inter_.AddService(&service_, brpc::SERVER_DOESNT_OWN_SERVICE);
   brpc::ServerOptions options;
   if (neeed_encrypt) {
-#ifdef PADDLE_WITH_ARM_BRPC
     options.mutable_ssl_options()->default_cert.certificate = "/cert.pem";
     options.mutable_ssl_options()->default_cert.private_key = "/key.pem";
-#else
-    options.ssl_options.default_cert.certificate = "/cert.pem";
-    options.ssl_options.default_cert.private_key = "/key.pem";
-#endif
   }
   if (server_inter_.Start(endpoint_inter_.c_str(), &options) != 0) {
     VLOG(4) << "switch inter server start fail. Try again.";

diff --git a/paddle/fluid/eager/auto_code_generator/generator/python_c_gen.py b/paddle/fluid/eager/auto_code_generator/generator/python_c_gen.py
@@ -36,6 +36,7 @@ def SkipAPIGeneration(forward_api_name):
     "long": "CastPyArg2Long",
     "int64_t": "CastPyArg2Long",
     "float": "CastPyArg2Float",
+    "double": "CastPyArg2Double",
     "std::string": "CastPyArg2String",
     "std::vector<bool>": "CastPyArg2Booleans",
     "std::vector<int>": "CastPyArg2Ints",

diff --git a/paddle/fluid/framework/attribute.h b/paddle/fluid/framework/attribute.h
@@ -180,6 +180,37 @@ struct ExtractAttribute<float> {
   const std::string& attr_name_;
 };
 
+template <>
+struct ExtractAttribute<double> {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  double* operator()(Attribute& attr) const {
+    if (attr.type() == typeid(int)) {  // NOLINT
+      int val = PADDLE_GET_CONST(int, attr);
+      attr = static_cast<double>(val);
+    } else if (attr.type() == typeid(int64_t)) {  // NOLINT
+      int64_t val = PADDLE_GET_CONST(int64_t, attr);
+      attr = static_cast<double>(val);
+    } else if (attr.type() == typeid(float)) {  // NOLINT
+      int64_t val = PADDLE_GET_CONST(float, attr);
+      attr = static_cast<double>(val);
+    }
+    double* attr_value = nullptr;
+    try {
+      attr_value = &paddle::get<double>(attr);
+    } catch (paddle::bad_variant_access const& bad_get) {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Cannot get attribute (%s) by type double, its type is %s.",
+          attr_name_,
+          paddle::platform::demangle(attr.type().name())));
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
 template <>
 struct ExtractAttribute<std::vector<double>> {
   explicit ExtractAttribute(const std::string& attr_name)

diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt
@@ -54,6 +54,9 @@ if(WITH_PSCORE)
   set(DISTRIBUTE_COMPILE_FLAGS
       "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor"
   )
+  if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 7.0)
+    set(DISTRIBUTE_COMPILE_FLAGS "${DISTRIBUTE_COMPILE_FLAGS} -faligned-new")
+  endif()
   set_source_files_properties(
     reduce_op_handle.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
   set_source_files_properties(

diff --git a/paddle/fluid/framework/executor_cache.cc b/paddle/fluid/framework/executor_cache.cc
@@ -50,6 +50,10 @@ static ExecutionStrategy GetExecutionStrategy(const platform::Place &place) {
       execution_strategy.num_threads_ = 1;
       break;
     }
+    case platform::DeviceType::NPU: {
+      execution_strategy.num_threads_ = 1;
+      break;
+    }
     default:
       PADDLE_THROW(platform::errors::Unavailable("Unsupported Device type %d.",
                                                  device_type));

diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto
@@ -38,6 +38,7 @@ enum AttrType {
   FLOAT64S = 12;
   VAR = 13;
   VARS = 14;
+  FLOAT64 = 15;
 }
 
 // OpDesc describes an instance of a C++ framework::OperatorBase
@@ -62,6 +63,7 @@ message OpDesc {
     repeated double float64s = 16;
     optional string var_name = 17;
     repeated string vars_name = 18;
+    optional double float64 = 19;
   };
 
   message Var {

diff --git a/paddle/fluid/framework/infershape_utils.cc b/paddle/fluid/framework/infershape_utils.cc
@@ -482,6 +482,10 @@ CompatInferMetaContext BuildInferMetaContext(InferShapeContext* ctx,
               infer_meta_context.EmplaceBackAttr(
                   phi::Scalar(PADDLE_GET_CONST(float, attr)));
               break;
+            case framework::proto::AttrType::FLOAT64:
+              infer_meta_context.EmplaceBackAttr(
+                  phi::Scalar(PADDLE_GET_CONST(double, attr)));
+              break;
             case framework::proto::AttrType::INT:
               infer_meta_context.EmplaceBackAttr(
                   phi::Scalar(PADDLE_GET_CONST(int, attr)));
@@ -651,6 +655,10 @@ CompatInferMetaContext BuildInferMetaContext(InferShapeContext* ctx,
             case phi::AttributeType::FLOAT32:
               infer_meta_context.EmplaceBackAttr(PADDLE_GET_CONST(float, attr));
               break;
+            case phi::AttributeType::FLOAT64:
+              infer_meta_context.EmplaceBackAttr(
+                  PADDLE_GET_CONST(double, attr));
+              break;
             case phi::AttributeType::INT32:
               infer_meta_context.EmplaceBackAttr(PADDLE_GET_CONST(int, attr));
               break;

diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc
@@ -668,6 +668,12 @@ void OpDesc::SetAttr(const std::string &name, const Attribute &v) {
         this->attrs_[name] = std::vector<float>();
         break;
       }
+      case proto::AttrType::FLOAT64S: {
+        VLOG(11) << "SetAttr: " << Type() << ", " << name
+                 << " from INTS to FLOAT64S";
+        this->attrs_[name] = std::vector<double>();
+        break;
+      }
       case proto::AttrType::STRINGS: {
         VLOG(11) << "SetAttr: " << Type() << ", " << name
                  << " from INTS to STRINGS";
@@ -838,6 +844,7 @@ struct SetAttrDescVisitor {
   mutable proto::OpDesc::Attr *attr_;
   void operator()(int v) const { attr_->set_i(v); }
   void operator()(float v) const { attr_->set_f(v); }
+  void operator()(double v) const { attr_->set_float64(v); }
   void operator()(const std::string &v) const { attr_->set_s(v); }
 
   // Please refer to https://github.com/PaddlePaddle/Paddle/issues/7162

diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
@@ -2745,6 +2745,10 @@ void OperatorWithKernel::BuildPhiKernelContext(
               phi_kernel_context->EmplaceBackAttr(std::move(
                   phi::Scalar(PADDLE_GET_CONST(float, attr_iter->second))));
               break;
+            case proto::AttrType::FLOAT64:
+              phi_kernel_context->EmplaceBackAttr(std::move(
+                  phi::Scalar(PADDLE_GET_CONST(double, attr_iter->second))));
+              break;
             case proto::AttrType::INT:
               phi_kernel_context->EmplaceBackAttr(std::move(
                   phi::Scalar(PADDLE_GET_CONST(int, attr_iter->second))));
@@ -2884,6 +2888,10 @@ void OperatorWithKernel::BuildPhiKernelContext(
             phi_kernel_context->EmplaceBackAttr(
                 PADDLE_GET_CONST(float, attr_iter->second));
             break;
+          case phi::AttributeType::FLOAT64:
+            phi_kernel_context->EmplaceBackAttr(
+                PADDLE_GET_CONST(double, attr_iter->second));
+            break;
           case phi::AttributeType::INT32:
             phi_kernel_context->EmplaceBackAttr(
                 PADDLE_GET_CONST(int, attr_iter->second));

diff --git a/paddle/fluid/framework/type_defs.h b/paddle/fluid/framework/type_defs.h
@@ -58,7 +58,8 @@ using Attribute = paddle::variant<paddle::blank,
                                   std::vector<int64_t>,
                                   std::vector<double>,
                                   VarDesc*,
-                                  std::vector<VarDesc*>>;
+                                  std::vector<VarDesc*>,
+                                  double>;
 using AttributeMap = std::unordered_map<std::string, Attribute>;
 
 #ifdef PADDLE_WITH_ASCEND_CL

diff --git a/paddle/fluid/imperative/prepared_operator.h b/paddle/fluid/imperative/prepared_operator.h
@@ -412,6 +412,10 @@ void BuildDygraphPhiKernelContext(const phi::KernelSignature& kernel_signature,
               kernel_ctx->EmplaceBackAttr(
                   std::move(phi::Scalar(PADDLE_GET_CONST(float, attr))));
               break;
+            case framework::proto::AttrType::FLOAT64:
+              kernel_ctx->EmplaceBackAttr(
+                  std::move(phi::Scalar(PADDLE_GET_CONST(double, attr))));
+              break;
             case framework::proto::AttrType::INT:
               kernel_ctx->EmplaceBackAttr(
                   std::move(phi::Scalar(PADDLE_GET_CONST(int, attr))));
@@ -549,6 +553,9 @@ void BuildDygraphPhiKernelContext(const phi::KernelSignature& kernel_signature,
           case phi::AttributeType::FLOAT32:
             kernel_ctx->EmplaceBackAttr(PADDLE_GET_CONST(float, attr));
             break;
+          case phi::AttributeType::FLOAT64:
+            kernel_ctx->EmplaceBackAttr(PADDLE_GET_CONST(double, attr));
+            break;
           case phi::AttributeType::INT32:
             kernel_ctx->EmplaceBackAttr(PADDLE_GET_CONST(int, attr));
             break;

diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt
@@ -106,7 +106,7 @@ set(OP_HEADER_DEPS ${OP_HEADER_DEPS} phi phi_api_utils gather_scatter_kernel bac
 register_operators(EXCLUDES py_layer_op py_func_op warpctc_op dgc_op load_combine_op lstm_op run_program_op eye_op quantize_linear_op
         recurrent_op save_combine_op sparse_attention_op sync_batch_norm_op ${OP_MKL_DEPS} DEPS ${OP_HEADER_DEPS})
 
-op_library(run_program_op SRCS run_program_op.cc run_program_op.cu.cc DEPS executor_cache ${OP_HEADER_DEPS})
+op_library(run_program_op SRCS run_program_op.cc run_program_op.cu.cc run_program_op_npu.cc DEPS executor_cache ${OP_HEADER_DEPS})
 target_link_libraries(run_program_op cuda_graph_with_memory_pool)
 op_library(quantize_linear_op DEPS phi)
 op_library(save_combine_op DEPS string_array)

diff --git a/paddle/fluid/operators/conv_base_helper.h b/paddle/fluid/operators/conv_base_helper.h
@@ -44,7 +44,13 @@ struct SearchAlgorithm {};
 template <typename AlgoT>
 struct SearchResult {
   SearchResult() {}
+  explicit SearchResult(const phi::autotune::DnnNode& node)
+      : algo(static_cast<AlgoT>(node.algo)),
+        workspace_size(node.workspace_size) {}
+
   explicit SearchResult(AlgoT a) : algo(a) {}
+  explicit SearchResult(AlgoT a, float t, size_t size)
+      : algo(a), time(t), workspace_size(size) {}
 
   AlgoT algo = static_cast<AlgoT>(0);
   float time = -1.f;
@@ -76,28 +82,50 @@ struct ConvArgsBase {
   // dilations
   std::vector<int> d;
 
+  // groups
+  int group;
+
+  // data foramt
+  DataLayout data_layout;
+
   ConvArgsBase(const framework::Tensor* x,
                const framework::Tensor* w,
                const framework::Tensor* o,
                const std::vector<int> s,
                const std::vector<int> p,
                const std::vector<int> d,
-               DataT dtype)
-      : x(x), w(w), o(o), s(s), p(p), d(d), cudnn_dtype(dtype) {}
+               DataT dtype,
+               int g,
+               DataLayout layout)
+      : x(x),
+        w(w),
+        o(o),
+        s(s),
+        p(p),
+        d(d),
+        cudnn_dtype(dtype),
+        group(g),
+        data_layout(layout) {}
 
   template <typename T>
-  size_t GetCacheKey() const {
+  phi::autotune::ConvCacheKey Convert2ConvCacheKey() const {
     auto x_shape = phi::vectorize(x->dims());
     auto w_shape = phi::vectorize(w->dims());
     VLOG(10) << "[ConvArgs] x_dims=" << x_shape << ", w_dims=" << w_shape
-             << ", strides=" << s << ", paddings=" << p << ", dilations=" << d;
-    return phi::autotune::ConvKey(
+             << ", strides=" << s << ", paddings=" << p << ", dilations=" << d
+             << ",data= " << paddle::experimental::CppTypeToDataType<T>::Type()
+             << ", group=" << group
+             << ", data layout=" << static_cast<int64_t>(data_layout);
+
+    return phi::autotune::ConvCacheKey(
         x_shape,
         w_shape,
         p,
         s,
         d,
-        paddle::experimental::CppTypeToDataType<T>::Type());
+        paddle::experimental::CppTypeToDataType<T>::Type(),
+        group,
+        static_cast<int64_t>(data_layout));
   }
 };