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

… remove_convert_tensor
PaddlePaddle · Jan 15, 2022 · 46efe37 · 46efe37 · paddle-bot-old · Jan 15, 2022
2 parents 66e8e22 + a887914
commit 46efe37
Show file tree

Hide file tree

Showing 310 changed files with 5,806 additions and 3,456 deletions.
diff --git a/cmake/configure.cmake b/cmake/configure.cmake
@@ -31,10 +31,12 @@ endif(NOT WITH_PROFILER)
 if(WITH_AVX AND AVX_FOUND)
     set(SIMD_FLAG ${AVX_FLAG})
     add_definitions(-DPADDLE_WITH_AVX)
-elseif(SSE3_FOUND)
-    if(NOT WIN32)
-        set(SIMD_FLAG ${SSE3_FLAG})
-    endif()
+elseif(SSE3_FOUND AND NOT WIN32)
+    set(SIMD_FLAG ${SSE3_FLAG})
+endif()
+
+if (SSE3_FOUND)
+    # TODO: Runtime detection should be used here.
     add_definitions(-DPADDLE_WITH_SSE3)
 endif()
 

diff --git a/cmake/external/llvm.cmake b/cmake/external/llvm.cmake
@@ -1,7 +1,7 @@
 include(FetchContent)
 
-set(LLVM_DOWNLOAD_URL https://paddle-inference-dist.bj.bcebos.com/CINN/llvm11.tar.gz)
-set(LLVM_MD5 39d32b6be466781dddf5869318dcba53)
+set(LLVM_DOWNLOAD_URL https://paddle-inference-dist.bj.bcebos.com/infrt/llvm_b5149f4e66a49a98b67e8e2de4e24a4af8e2781b.tar.gz)
+set(LLVM_MD5 022819bb5760817013cf4b8a37e97d5e)
 
 set(FETCHCONTENT_BASE_DIR ${THIRD_PARTY_PATH}/llvm)
 set(FETCHCONTENT_QUIET OFF)
@@ -51,16 +51,18 @@ message(STATUS "Using LLVMConfig.cmake in: ${LLVM_DIR}")
 # To build with MLIR, the LLVM is build from source code using the following flags:
 
 #[==[
-cmake -G Ninja ../llvm \
+cmake ../llvm  -G "Unix Makefiles" \
   -DLLVM_ENABLE_PROJECTS="mlir;clang" \
   -DLLVM_BUILD_EXAMPLES=OFF \
   -DLLVM_TARGETS_TO_BUILD="X86" \
   -DCMAKE_BUILD_TYPE=Release \
   -DLLVM_ENABLE_ASSERTIONS=ON \
   -DLLVM_ENABLE_ZLIB=OFF \
   -DLLVM_ENABLE_RTTI=ON \
+  -DLLVM_INSTALL_UTILS=ON \
+  -DCMAKE_INSTALL_PREFIX=./install
 #]==]
-# The matched llvm-project version is f9dc2b7079350d0fed3bb3775f496b90483c9e42 (currently a temporary commit)
+# The matched llvm-project version is b5149f4e66a49a98b67e8e2de4e24a4af8e2781b (currently a temporary commit)
 
 add_definitions(${LLVM_DEFINITIONS})
 
@@ -75,7 +77,7 @@ add_definitions(${LLVM_DEFINITIONS})
 
 
 # The minimum needed libraries for MLIR IR parse and transform.
-set(MLIR_IR_LIBS MLIRAnalysis MLIRStandardOps MLIRPass MLIRParser MLIRDialect MLIRIR MLIROptLib)
+set(MLIR_IR_LIBS MLIRAnalysis MLIRPass MLIRParser MLIRDialect MLIRIR MLIROptLib)
 
 
 # tb_base is the name of a xxx.td file (without the .td suffix)
@@ -89,6 +91,7 @@ function(mlir_tablegen_on td_base)
   mlir_tablegen(${td_base}.cpp.inc -gen-op-defs)
   if (mlir_tablegen_on_DIALECT)
     mlir_tablegen(${td_base}_dialect.hpp.inc --gen-dialect-decls -dialect=${mlir_tablegen_on_DIALECT})
+    mlir_tablegen(${td_base}_dialect.cpp.inc --gen-dialect-defs -dialect=${mlir_tablegen_on_DIALECT})
   endif()
   add_public_tablegen_target(${td_base}_IncGen)
   add_custom_target(${td_base}_inc DEPENDS ${td_base}_IncGen)

diff --git a/cmake/pten_kernel.cmake b/cmake/pten_kernel.cmake
@@ -16,12 +16,12 @@
 function(kernel_declare TARGET_LIST)
     foreach(kernel_path ${TARGET_LIST})
         file(READ ${kernel_path} kernel_impl)
-        # TODO(chenweihang): rename PT_REGISTER_CTX_KERNEL to PT_REGISTER_KERNEL
+        # TODO(chenweihang): rename PT_REGISTER_KERNEL to PT_REGISTER_KERNEL
         # NOTE(chenweihang): now we don't recommend to use digit in kernel name
-        string(REGEX MATCH "(PT_REGISTER_CTX_KERNEL|PT_REGISTER_GENERAL_KERNEL)\\([ \t\r\n]*[a-z0-9_]*," first_registry "${kernel_impl}")
+        string(REGEX MATCH "(PT_REGISTER_KERNEL|PT_REGISTER_GENERAL_KERNEL)\\([ \t\r\n]*[a-z0-9_]*," first_registry "${kernel_impl}")
         if (NOT first_registry STREQUAL "")
             # parse the first kernel name
-            string(REPLACE "PT_REGISTER_CTX_KERNEL(" "" kernel_name "${first_registry}")
+            string(REPLACE "PT_REGISTER_KERNEL(" "" kernel_name "${first_registry}")
             string(REPLACE "PT_REGISTER_GENERAL_KERNEL(" "" kernel_name "${kernel_name}")
             string(REPLACE "," "" kernel_name "${kernel_name}")
             string(REGEX REPLACE "[ \t\r\n]+" "" kernel_name "${kernel_name}")

diff --git a/paddle/fluid/distributed/fleet_executor/carrier.h b/paddle/fluid/distributed/fleet_executor/carrier.h
@@ -101,8 +101,8 @@ class Carrier final {
   std::mutex running_mutex_;
   std::condition_variable cond_var_;
   std::vector<framework::Scope*> microbatch_scopes_;
-  framework::Scope* root_scope_;
-  framework::Scope* minibatch_scope_;
+  framework::Scope* root_scope_{nullptr};
+  framework::Scope* minibatch_scope_{nullptr};
   paddle::platform::Place place_;
   paddle::platform::DeviceContext* dev_ctx_{nullptr};
   int64_t rank_;

diff --git a/paddle/fluid/eager/accumulation/accumulation_node.cc b/paddle/fluid/eager/accumulation/accumulation_node.cc
@@ -18,7 +18,6 @@
 
 #include "paddle/pten/api/all.h"
 #include "paddle/pten/core/dense_tensor.h"
-#include "paddle/pten/include/core.h"
 
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"

diff --git a/paddle/fluid/eager/accumulation/gradient_accumulation.cc b/paddle/fluid/eager/accumulation/gradient_accumulation.cc
@@ -28,7 +28,6 @@
 #include "paddle/fluid/platform/float16.h"
 #include "paddle/pten/api/all.h"
 #include "paddle/pten/core/convert_utils.h"
-#include "paddle/pten/include/core.h"
 #include "unsupported/Eigen/CXX11/Tensor"
 #ifdef PADDLE_WITH_XPU
 #include "xpu/refactor/math.h"

diff --git a/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.cc b/paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.cc
@@ -16,7 +16,7 @@
 #include "paddle/fluid/eager/api/utils/global_utils.h"
 #include "paddle/fluid/eager/eager_tensor.h"
 
-#include "paddle/pten/api/all.h"
+#include "paddle/pten/kernels/scale_kernel.h"
 
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
@@ -33,28 +33,28 @@ static void ScaleDeviceDispatch(const pten::DenseTensor& dense_tensor,
                                 pten::DenseTensor* dense_out) {
   switch (dense_tensor.dtype()) {
     case pten::DataType::FLOAT64: {
-      pten::Scale<double, DeviceContext>(
+      pten::ScaleKernel<double, DeviceContext>(
           dev_ctx, dense_tensor /* tensor */, scale /* scale */,
           bias /* bias */, bias_after_scale /* bias_after_scale */,
           dense_out /* out tensor */);
       break;
     }
     case pten::DataType::FLOAT32: {
-      pten::Scale<float, DeviceContext>(dev_ctx, dense_tensor /* tensor */,
-                                        scale /* scale */, bias /* bias */,
-                                        bias_after_scale /* bias_after_scale */,
-                                        dense_out /* out tensor */);
+      pten::ScaleKernel<float, DeviceContext>(
+          dev_ctx, dense_tensor /* tensor */, scale /* scale */,
+          bias /* bias */, bias_after_scale /* bias_after_scale */,
+          dense_out /* out tensor */);
       break;
     }
     case pten::DataType::INT64: {
-      pten::Scale<int64_t, DeviceContext>(
+      pten::ScaleKernel<int64_t, DeviceContext>(
           dev_ctx, dense_tensor /* tensor */, scale /* scale */,
           bias /* bias */, bias_after_scale /* bias_after_scale */,
           dense_out /* out tensor */);
       break;
     }
     case pten::DataType::INT32: {
-      pten::Scale<int32_t, DeviceContext>(
+      pten::ScaleKernel<int32_t, DeviceContext>(
           dev_ctx, dense_tensor /* tensor */, scale /* scale */,
           bias /* bias */, bias_after_scale /* bias_after_scale */,
           dense_out /* out tensor */);

diff --git a/paddle/fluid/eager/api/generated/eager_generated/forwards/scale.cc b/paddle/fluid/eager/api/generated/eager_generated/forwards/scale.cc
@@ -29,7 +29,6 @@
 #include "paddle/fluid/eager/utils.h"
 
 #include "paddle/pten/api/all.h"
-#include "paddle/pten/include/core.h"
 
 namespace egr {
 

diff --git a/paddle/fluid/eager/eager_tensor.h b/paddle/fluid/eager/eager_tensor.h
@@ -18,7 +18,6 @@
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/variable.h"
 // pten deps
-#include "paddle/pten/all.h"
 #include "paddle/pten/api/all.h"
 #include "paddle/pten/api/lib/api_declare.h"
 #include "paddle/pten/api/lib/utils/tensor_utils.h"

diff --git a/paddle/fluid/eager/grad_node_info.h b/paddle/fluid/eager/grad_node_info.h
@@ -16,7 +16,6 @@
 
 #include "paddle/fluid/eager/eager_tensor.h"
 #include "paddle/pten/api/all.h"
-#include "paddle/pten/include/core.h"
 
 namespace egr {
 /**

diff --git a/paddle/fluid/eager/legacy/infer_shape_context.h b/paddle/fluid/eager/legacy/infer_shape_context.h
@@ -31,15 +31,18 @@ class EagerInferShapeContext : public paddle::framework::InferShapeContext {
   using DDim = paddle::framework::DDim;
 
  public:
-  EagerInferShapeContext(const NameTensorMap* in, const NameTensorMap* out,
-                         const paddle::framework::AttributeMap* attr,
-                         const paddle::framework::AttributeMap* default_attr,
-                         const std::string op_type)
+  EagerInferShapeContext(
+      const NameTensorMap* in, const NameTensorMap* out,
+      const paddle::framework::AttributeMap* attr,
+      const paddle::framework::AttributeMap* default_attr,
+      const std::string op_type,
+      const paddle::framework::OpKernelType* op_kernel_type = nullptr)
       : tensor_in_(in),
         tensor_out_(out),
         attrs_(attr),
         default_attrs_(default_attr),
-        op_type_(op_type) {}
+        op_type_(op_type),
+        op_kernel_type_(op_kernel_type) {}
 
   bool HasInput(const std::string& name) const override {
     // has only one input
@@ -214,6 +217,11 @@ class EagerInferShapeContext : public paddle::framework::InferShapeContext {
 
   bool IsRuntime() const override { return true; }
 
+  bool IsRunMKLDNNKernel() const override {
+    return (op_kernel_type_ && (op_kernel_type_->data_layout_ ==
+                                paddle::framework::DataLayout::kMKLDNN));
+  }
+
   // TODO(paddle-dev): Can this be template?
   std::vector<paddle::framework::InferShapeVarPtr> GetInputVarPtrs(
       const std::string& name) const override {
@@ -400,6 +408,7 @@ class EagerInferShapeContext : public paddle::framework::InferShapeContext {
   const paddle::framework::AttributeMap* attrs_;
   const paddle::framework::AttributeMap* default_attrs_;
   const std::string op_type_;
+  const paddle::framework::OpKernelType* op_kernel_type_;
 };
 
 }  // namespace legacy

diff --git a/paddle/fluid/eager/legacy/infer_var_type_context.h b/paddle/fluid/eager/legacy/infer_var_type_context.h
@@ -26,7 +26,6 @@
 #include "paddle/fluid/framework/var_type_inference.h"
 #include "paddle/fluid/framework/var_type_traits.h"
 #include "paddle/pten/api/all.h"
-#include "paddle/pten/include/core.h"
 
 namespace egr {
 namespace legacy {

diff --git a/paddle/fluid/eager/legacy/prepared_operator.cc b/paddle/fluid/eager/legacy/prepared_operator.cc
@@ -173,7 +173,7 @@ static void PreparedOpRunImpl(
   paddle::framework::Scope scope;
 
   EagerInferShapeContext infer_shape_ctx(&ins, &outs, &attrs, &default_attrs,
-                                         op.Type());
+                                         op.Type(), &kernel_type);
   op.Info().infer_shape_(&infer_shape_ctx);
 
   func(EagerExecutionContext(op, scope, *dev_ctx, ctx, ins, outs, attrs,

diff --git a/paddle/fluid/eager/legacy/prepared_operator.h b/paddle/fluid/eager/legacy/prepared_operator.h
@@ -25,8 +25,6 @@
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/type_defs.h"
 
-#include "paddle/pten/include/core.h"
-
 DECLARE_bool(use_mkldnn);
 
 namespace paddle {

diff --git a/paddle/fluid/eager/legacy/tensor_helper.h b/paddle/fluid/eager/legacy/tensor_helper.h
@@ -17,7 +17,6 @@
 #include <vector>
 #include "paddle/fluid/eager/eager_tensor.h"
 #include "paddle/pten/api/all.h"
-#include "paddle/pten/include/core.h"
 namespace egr {
 namespace legacy {
 

diff --git a/paddle/fluid/framework/data_device_transform_test.cu b/paddle/fluid/framework/data_device_transform_test.cu
@@ -24,7 +24,6 @@ limitations under the License. */
 #include "paddle/fluid/platform/init.h"
 
 #include "paddle/fluid/framework/pten_utils.h"
-#include "paddle/pten/include/core.h"
 
 namespace paddle {
 namespace framework {

diff --git a/paddle/fluid/framework/new_executor/new_executor_defs.cc b/paddle/fluid/framework/new_executor/new_executor_defs.cc
@@ -307,6 +307,17 @@ void InterpretercoreInferShapeContext::SetLoDLevel(const std::string& out,
 
 bool InterpretercoreInferShapeContext::IsRuntime() const { return true; }
 
+bool InterpretercoreInferShapeContext::IsRunMKLDNNKernel() const {
+  try {
+    auto& op_with_kernel = dynamic_cast<const OperatorWithKernel&>(op_);
+    return ((op_with_kernel.kernel_type()) &&
+            (op_with_kernel.kernel_type()->data_layout_ ==
+             framework::DataLayout::kMKLDNN));
+  } catch (std::bad_cast exp) {
+    return false;
+  }
+}
+
 // TODO(paddle-dev): Can this be template?
 std::vector<InferShapeVarPtr> InterpretercoreInferShapeContext::GetInputVarPtrs(
     const std::string& name) const {

diff --git a/paddle/fluid/framework/new_executor/new_executor_defs.h b/paddle/fluid/framework/new_executor/new_executor_defs.h
@@ -84,6 +84,8 @@ class InterpretercoreInferShapeContext : public InferShapeContext {
 
   bool IsRuntime() const override;
 
+  bool IsRunMKLDNNKernel() const override;
+
   // TODO(paddle-dev): Can this be template?
   std::vector<InferShapeVarPtr> GetInputVarPtrs(
       const std::string& name) const override;

diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc
@@ -240,6 +240,8 @@ class CompileTimeInferShapeContext : public InferShapeContext {
 
   bool IsRuntime() const override;
 
+  bool IsRunMKLDNNKernel() const override;
+
   std::vector<proto::VarType::Type> GetInputsVarType(
       const std::string &name) const override {
     return GetVarTypes(Inputs(name));
@@ -930,6 +932,8 @@ void CompileTimeInferShapeContext::SetRepeatedDims(
 
 bool CompileTimeInferShapeContext::IsRuntime() const { return false; }
 
+bool CompileTimeInferShapeContext::IsRunMKLDNNKernel() const { return false; }
+
 proto::VarType::Type CompileTimeInferShapeContext::GetVarType(
     const std::string &name) const {
   return block_.FindVarRecursive(name)->GetType();

diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
@@ -884,6 +884,17 @@ class RuntimeInferShapeContext : public InferShapeContext {
 
   bool IsRuntime() const override { return true; }
 
+  bool IsRunMKLDNNKernel() const override {
+    try {
+      auto& op_with_kernel = dynamic_cast<const OperatorWithKernel&>(op_);
+      return ((op_with_kernel.kernel_type()) &&
+              (op_with_kernel.kernel_type()->data_layout_ ==
+               framework::DataLayout::kMKLDNN));
+    } catch (std::bad_cast exp) {
+      return false;
+    }
+  }
+
   // TODO(paddle-dev): Can this be template?
   std::vector<InferShapeVarPtr> GetInputVarPtrs(
       const std::string& name) const override {
@@ -1178,9 +1189,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
     platform::RecordEvent record_event("infer_shape",
                                        platform::EventRole::kInnerOp);
     RuntimeInferShapeContext infer_shape_ctx(*this, *runtime_ctx);
-    // TODO(chenweihang): replace this after removing `this->IsMKLDNNType()`
-    // in some mkldnn infershape functions, such conv2d infershape
-    this->InferShape(&infer_shape_ctx);
+    this->Info().infer_shape_(&infer_shape_ctx);
   }
 
   if (FLAGS_enable_unused_var_check) {

diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h
@@ -41,7 +41,8 @@ limitations under the License. */
 #include "paddle/utils/flat_hash_map.h"
 
 #include "paddle/pten/core/arg_map_context.h"
-#include "paddle/pten/include/core.h"
+#include "paddle/pten/core/kernel_context.h"
+#include "paddle/pten/core/kernel_factory.h"
 
 namespace paddle {
 namespace framework {
@@ -410,8 +411,8 @@ class ExecutionContext {
     auto tmp_allocation_ptr = memory::Alloc(dev_ctx, product(dim) * sizeof(T));
     auto& deleter = tmp_allocation_ptr.get_deleter();
     auto* allocation_ptr = tmp_allocation_ptr.release();
-    auto shared_allocation = std::shared_ptr<memory::allocation::Allocation>(
-        allocation_ptr, deleter);
+    auto shared_allocation =
+        std::shared_ptr<pten::Allocation>(allocation_ptr, deleter);
 
     PADDLE_ENFORCE_GE(
         allocation_ptr->size(), framework::product(dim) * sizeof(T),
@@ -528,11 +529,6 @@ class OperatorWithKernel : public OperatorBase {
     return g_all_op_kernels;
   }
 
-  bool IsMKLDNNType() const {
-    return ((this->kernel_type_) && (this->kernel_type_->data_layout_ ==
-                                     framework::DataLayout::kMKLDNN));
-  }
-
   bool SupportGPU() const override {
     auto& op_kernels = OperatorWithKernel::AllOpKernels().at(type_);
     return std::any_of(op_kernels.begin(), op_kernels.end(),
@@ -609,6 +605,8 @@ class OperatorWithKernel : public OperatorBase {
     return pt_kernel_context_.get();
   }
 
+  const OpKernelType* kernel_type() const { return kernel_type_.get(); }
+
  private:
   void RunImpl(const Scope& scope, const platform::Place& place) const final;
   void RunImpl(const Scope& scope, const platform::Place& place,

diff --git a/paddle/fluid/framework/shape_inference.h b/paddle/fluid/framework/shape_inference.h
@@ -102,6 +102,8 @@ class InferShapeContext {
 
   virtual bool IsRuntime() const = 0;
 
+  virtual bool IsRunMKLDNNKernel() const = 0;
+
   virtual std::vector<InferShapeVarPtr> GetInputVarPtrs(
       const std::string &name) const = 0;
   virtual std::vector<InferShapeVarPtr> GetOutputVarPtrs(