[phi decoupling] decouple dependency to device_context in phi (Part 1)

.gitignore

@@ -94,4 +94,6 @@ paddle/fluid/pybind/op_function_impl.h
 paddle/fluid/pybind/*final_state_op_function_impl.h
 paddle/fluid/prim/api/generated/prim_api/*
 paddle/fluid/framework/__init__.py
+paddle/phi/api/profiler/__init__.py
 python/paddle/incubate/fleet/parameter_server/pslib/ps_pb2.py
+python/paddle/fluid/incubate/fleet/parameter_server/pslib/ps_pb2.py

paddle/fluid/operators/fused/fusion_conv_inception_op.cu

@@ -165,7 +165,7 @@ class CUDNNConvInceptionFusionOpKernel : public framework::OpKernel<T> {
       PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
           conv_desc[i], CUDNN_DEFAULT_MATH));
 #if CUDA_VERSION >= 11000 && CUDNN_VERSION >= 8000
-      if (!platform::allow_tf32_cudnn) {
+      if (!phi::allow_tf32_cudnn) {
         PADDLE_ENFORCE_GPU_SUCCESS(
             platform::dynload::cudnnSetConvolutionMathType(conv_desc[i],
                                                            CUDNN_FMA_MATH));

paddle/fluid/operators/math/beam_search_xpu.cc

@@ -19,13 +19,6 @@ namespace phi {
 class DenseTensor;
 }  // namespace phi
 
-namespace paddle {
-namespace framework {}  // namespace framework
-namespace platform {
-class XPUDeviceContext;
-}  // namespace platform
-}  // namespace paddle
-
 namespace paddle {
 namespace operators {
 namespace math {

paddle/fluid/platform/CMakeLists.txt

@@ -159,7 +159,6 @@ cc_library(
        cudnn_workspace_helper
        ${XPU_CTX_DEPS}
        ${MLU_CTX_DEPS}
-       eigen3
        phi_backends
        phi_device_context
        generator)

paddle/fluid/platform/cuda_graph_with_memory_pool.cc

@@ -15,7 +15,7 @@
 #include "paddle/fluid/platform/cuda_graph_with_memory_pool.h"
 
 #include "paddle/fluid/memory/allocation/allocator_facade.h"
-#include "paddle/phi/backends/all_context.h"
+#include "paddle/phi/backends/context_pool.h"
 
 DECLARE_bool(use_stream_safe_cuda_allocator);
 DECLARE_bool(new_executor_use_cuda_graph);

paddle/fluid/platform/device/custom/custom_device_test.cc

@@ -54,7 +54,8 @@ void InitDevice() {
   }
   EXPECT_GT(static_cast<int>(places.size()), 0);
 
-  paddle::platform::DeviceContextPool::Init(places);
+  paddle::platform::DeviceContextPool::Init(
+      places, paddle::platform::EmplaceExternalContext);
 }
 
 void TestDeviceInterface(const paddle::platform::Place& place) {