Merge branch 'develop' into FFN

Author: zhangkaihuo

CMakeLists.txt

@@ -300,10 +300,6 @@ if(WITH_GPU)
     endif()
 endif()
 
-if(WITH_CINN)
-    include(cinn)
-endif()
-
 if(WITH_ROCM)
     include(hip)
     include(miopen) # set miopen libraries, must before configure

cmake/cinn.cmake renamed to cmake/external/cinn.cmake

@@ -27,16 +27,15 @@ add_definitions(-w)
 include(ExternalProject)
 set(CINN_SOURCE_DIR ${THIRD_PARTY_PATH}/CINN)
 # TODO(zhhsplendid): Modify git tag after we have release tag
-set(CINN_GIT_TAG 3f004bfa3ed273ecf1de8e7b946433038c79b84f)
-set(CINN_OPTIONAL_ARGS -DWITH_CUDA=${WITH_GPU} -DWITH_CUDNN=${WITH_GPU} -DPUBLISH_LIBS=ON)
-set(CINN_BUILD_COMMAND $(MAKE) cinncore -j && $(MAKE) cinnapi -j)
+set(CINN_GIT_TAG e422c01b7875301996a2baf67a14ba61b0e6192a)
+set(CINN_OPTIONAL_ARGS -DPY_VERSION=${PY_VERSION} -DWITH_CUDA=${WITH_GPU} -DWITH_CUDNN=${WITH_GPU} -DPUBLISH_LIBS=ON -DWITH_TESTING=ON)
+set(CINN_BUILD_COMMAND $(MAKE) cinnapi -j)
 ExternalProject_Add(
   external_cinn
   ${EXTERNAL_PROJECT_LOG_ARGS}
   GIT_REPOSITORY   "${GIT_URL}/PaddlePaddle/CINN.git"
   GIT_TAG          ${CINN_GIT_TAG}
   PREFIX           ${CINN_SOURCE_DIR}
-  UPDATE_COMMAND   ""
   BUILD_COMMAND    ${CINN_BUILD_COMMAND}
   INSTALL_COMMAND  ""
   CMAKE_ARGS       ${CINN_OPTIONAL_ARGS})
@@ -52,49 +51,20 @@ message(STATUS "CINN BINARY_DIR: ${CINN_BINARY_DIR}")
 message(STATUS "CINN SOURCE_DIR: ${CINN_SOURCE_DIR}")
 
 
-#########################
-# Add CINN's dependencies
-#########################
+######################################
+# Add CINN's dependencies header files
+######################################
 
 # Add absl
-set(ABSL_LIB_NAMES
-  hash
-  wyhash
-  city
-  strings
-  throw_delegate
-  bad_any_cast_impl
-  bad_optional_access
-  bad_variant_access
-  raw_hash_set
-  )
-set(ABSL_LIB_DIR "${CINN_BINARY_DIR}/dist/third_party/absl/lib")
 set(ABSL_INCLUDE_DIR "${CINN_BINARY_DIR}/dist/third_party/absl/include")
-add_library(absl STATIC IMPORTED GLOBAL)
-set_target_properties(absl PROPERTIES IMPORTED_LOCATION ${ABSL_LIB_DIR}/libabsl_base.a)
-foreach(lib_name ${ABSL_LIB_NAMES})
-    target_link_libraries(absl INTERFACE ${ABSL_LIB_DIR}/libabsl_${lib_name}.a)
-endforeach()
 include_directories(${ABSL_INCLUDE_DIR})
 
 # Add isl
-set(ISL_LIB_DIR "${CINN_BINARY_DIR}/dist/third_party/isl/lib")
 set(ISL_INCLUDE_DIR "${CINN_BINARY_DIR}/dist/third_party/isl/include")
-add_library(isl STATIC IMPORTED GLOBAL)
-set_target_properties(isl PROPERTIES IMPORTED_LOCATION ${ISL_LIB_DIR}/libisl.a)
 include_directories(${ISL_INCLUDE_DIR})
 
 # Add LLVM
-set(LLVM_LIB_NAMES
-  ExecutionEngine
-  )
-set(LLVM_LIB_DIR "${CINN_BINARY_DIR}/dist/third_party/llvm/lib")
 set(LLVM_INCLUDE_DIR "${CINN_BINARY_DIR}/dist/third_party/llvm/include")
-add_library(llvm STATIC IMPORTED GLOBAL)
-set_target_properties(llvm PROPERTIES IMPORTED_LOCATION ${LLVM_LIB_DIR}/libLLVMCore.a)
-foreach(lib_name ${LLVM_LIB_NAMES})
-    target_link_libraries(llvm INTERFACE ${LLVM_LIB_DIR}/libLLVM${lib_name}.a)
-endforeach()
 include_directories(${LLVM_INCLUDE_DIR})
 
 ######################################################
@@ -108,5 +78,5 @@ set(CINN_INCLUDE_DIR "${CINN_BINARY_DIR}/dist/cinn/include")
 add_library(cinn SHARED IMPORTED GLOBAL)
 set_target_properties(cinn PROPERTIES IMPORTED_LOCATION "${CINN_LIB_LOCATION}/${CINN_LIB_NAME}")
 include_directories(${CINN_INCLUDE_DIR})
-add_dependencies(cinn external_cinn absl isl llvm glog gflag)
+add_dependencies(cinn external_cinn)
 

cmake/external/xpu.cmake

@@ -35,7 +35,7 @@ ELSE ()
 ENDIF()
 
 SET(XPU_BASE_URL_WITHOUT_DATE "https://baidu-kunlun-product.cdn.bcebos.com/KL-SDK/klsdk-dev")
-SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20210921")
+SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20211020")
 SET(XPU_XRE_URL  "${XPU_BASE_URL}/${XPU_XRE_DIR_NAME}.tar.gz" CACHE STRING "" FORCE)
 SET(XPU_XDNN_URL "${XPU_BASE_URL}/${XPU_XDNN_DIR_NAME}.tar.gz" CACHE STRING "" FORCE)
 SET(XPU_XCCL_URL "${XPU_BASE_URL_WITHOUT_DATE}/20210623/${XPU_XCCL_DIR_NAME}.tar.gz" CACHE STRING "" FORCE)

cmake/operators.cmake

@@ -299,7 +299,7 @@ function(op_library TARGET)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n")
     endif()
 
-    if (WITH_XPU AND ${xpu_cc_srcs_len} GREATER 0)
+    if (WITH_XPU AND ${pybind_flag} EQUAL 0 AND ${xpu_cc_srcs_len} GREATER 0)
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, XPU);\n")
     endif()
 

cmake/third_party.cmake

@@ -360,6 +360,12 @@ if (WITH_LITE)
     include(external/lite)
 endif (WITH_LITE)
 
+if (WITH_CINN)
+    message(STATUS "Compile Paddle with CINN.")
+    include(external/cinn)
+    add_definitions(-DPADDLE_WITH_CINN)
+endif (WITH_CINN)
+
 if (WITH_CRYPTO)
     include(external/cryptopp)   # download, build, install cryptopp
     list(APPEND third_party_deps extern_cryptopp)

paddle/fluid/framework/CMakeLists.txt

@@ -26,7 +26,9 @@ add_subdirectory(details)
 add_subdirectory(fleet)
 add_subdirectory(io)
 add_subdirectory(new_executor)
-add_subdirectory(paddle2cinn)
+if (WITH_CINN)
+  add_subdirectory(paddle2cinn)
+endif()
 #ddim lib
 proto_library(framework_proto SRCS framework.proto)
 proto_library(pass_desc_proto SRCS pass_desc.proto DEPS framework_proto)
@@ -353,7 +355,7 @@ target_link_libraries(executor while_op_helper executor_gc_helper recurrent_op_h
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
         threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor async_ssa_graph_executor
         graph build_strategy bind_threaded_ssa_graph_executor collective_helper
-        fast_threaded_ssa_graph_executor variable_helper cinn_runner)
+        fast_threaded_ssa_graph_executor variable_helper)
 
 cc_library(executor_cache SRCS executor_cache.cc DEPS parallel_executor)
 if(WITH_PSCORE)

paddle/fluid/framework/details/CMakeLists.txt

@@ -139,7 +139,12 @@ set(IR_PASS_DEPS graph_viz_pass multi_devices_graph_pass
     coalesce_grad_tensor_pass fuse_all_reduce_op_pass backward_optimizer_op_deps_pass
     fuse_adam_op_pass fuse_sgd_op_pass fuse_momentum_op_pass
     sync_batch_norm_pass runtime_context_cache_pass graph_to_program_pass
-    fix_op_run_order_pass build_cinn_pass)
+    fix_op_run_order_pass)
+
+if (WITH_CINN)
+  set(IR_PASS_DEPS ${IR_PASS_DEPS} build_cinn_pass)
+endif()
+
 if(NOT APPLE AND NOT WIN32 AND (WITH_GPU OR WITH_ROCM))
   set(IR_PASS_DEPS ${IR_PASS_DEPS} fusion_group_pass)
 endif()

paddle/fluid/framework/details/build_strategy.cc

@@ -20,8 +20,10 @@ limitations under the License. */
 #include "paddle/fluid/framework/ir/multi_devices_graph_pass/multi_devices_graph_pass.h"
 
 DECLARE_bool(convert_all_blocks);
-DECLARE_bool(use_cinn);
 DECLARE_bool(use_mkldnn);
+#ifdef PADDLE_WITH_CINN
+DECLARE_bool(use_cinn);
+#endif
 
 namespace paddle {
 namespace framework {
@@ -72,10 +74,13 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
     // Note: This pass is used to check whether the multi_device_graph is right.
     AppendPass("multi_devices_check_pass");
 
-    // Note: This pass is used to enable cinn.
+#ifdef PADDLE_WITH_CINN
     if (FLAGS_use_cinn) {
+      // Note: This pass is used to enable cinn.
       AppendPass("build_cinn_pass");
     }
+#endif
+
     SetCollectiveContext();
   }
 
@@ -486,7 +491,9 @@ USE_PASS(fuse_momentum_op_pass);
 USE_PASS(fuse_all_reduce_op_pass);
 USE_PASS(runtime_context_cache_pass);
 USE_PASS(add_reader_dependency_pass);
+#ifdef PADDLE_WITH_CINN
 USE_PASS(build_cinn_pass);
+#endif
 #ifdef PADDLE_WITH_MKLDNN
 USE_PASS(mkldnn_placement_pass);
 #endif

paddle/fluid/framework/ir/cinn_lib_test.cc

@@ -24,6 +24,7 @@ limitations under the License. */
 #include <cuda_runtime.h>
 #endif
 
+#include "cinn/cinn.h"
 #include "cinn/common/target.h"
 #include "cinn/frontend/net_builder.h"
 #include "cinn/frontend/syntax.h"

paddle/fluid/framework/new_executor/event_count.h

@@ -50,11 +50,13 @@
 #include <cstdlib>
 #include <mutex>
 #include <vector>
-#include "paddle/fluid/framework/new_executor/workqueue_utils.h"
 
 namespace paddle {
 namespace framework {
 
+void* AlignedMalloc(size_t size, size_t alignment);
+void AlignedFree(void* memory_ptr);
+
 class EventCount {
  public:
   class Waiter;