Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into del_densetensor_mem_desc_2

Author: wanghuancoder

.gitignore

@@ -108,5 +108,13 @@ paddle/fluid/pir/dialect/operator/ir/pd_api.*
 paddle/fluid/pir/dialect/operator/ir/op_decomp.cc
 paddle/fluid/pir/dialect/operator/ir/pd_op_vjp.cc
 paddle/fluid/pir/dialect/operator/ir/pd_op.*
+paddle/fluid/pir/dialect/operator/ir/pd_op_bwd.*
+paddle/fluid/pir/dialect/operator/ir/pd_op_fused.*
+paddle/fluid/pir/dialect/operator/ir/pd_op_fused_bwd.*
+paddle/fluid/pir/dialect/operator/ir/pd_pir_op.*
+paddle/fluid/pir/dialect/operator/ir/pd_pir_op_bwd.*
+paddle/fluid/pir/dialect/operator/ir/pd_pir_op_update.*
+paddle/fluid/pir/dialect/operator/ir/pd_op_info.*
 paddle/cinn/hlir/dialect/generated/ops.parsed.yaml
 paddle/cinn/hlir/dialect/operator/ir/cinn_op.*
+paddle/cinn/hlir/dialect/operator/ir/cinn_op_info.*

cmake/generic.cmake

@@ -833,7 +833,6 @@ function(hip_test TARGET_NAME)
       ${hip_test_DEPS}
       paddle_gtest_main
       lod_tensor
-      memory
       gtest
       glog
       phi
@@ -843,7 +842,6 @@ function(hip_test TARGET_NAME)
       ${hip_test_DEPS}
       paddle_gtest_main
       lod_tensor
-      memory
       gtest
       phi
       glog)
@@ -940,7 +938,6 @@ function(xpu_test TARGET_NAME)
       ${xpu_test_DEPS}
       paddle_gtest_main
       lod_tensor
-      memory
       gtest
       phi
       glog
@@ -950,7 +947,6 @@ function(xpu_test TARGET_NAME)
       ${xpu_test_DEPS}
       paddle_gtest_main
       lod_tensor
-      memory
       gtest
       phi
       glog)

paddle/cinn/hlir/dialect/operator/ir/CMakeLists.txt

@@ -30,6 +30,9 @@ if(NOT CINN_ONLY)
   set(cinn_op_header_file_tmp ${cinn_op_header_file}.tmp)
   set(cinn_op_source_file_tmp ${cinn_op_source_file}.tmp)
 
+  set(cinn_op_info_file ${CINN_DIALECT_SOURCE_DIR}/cinn_op_info.cc)
+  set(cinn_op_info_file_tmp ${cinn_op_info_file}.tmp)
+
   execute_process(
     COMMAND ${CMAKE_COMMAND} -E make_directory ${parsed_op_dir}
     COMMAND ${PYTHON_EXECUTABLE} ${cinn_op_gen_parsed_yaml_file} --op_yaml_path
@@ -41,9 +44,11 @@ if(NOT CINN_ONLY)
       ${cinn_op_parsed_yaml_files} --op_compat_yaml_file
       ${cinn_op_compat_yaml_file} --namespaces ${cinn_op_namespace}
       --dialect_name ${cinn_op_dialect_name} --op_def_h_file
-      ${cinn_op_header_file_tmp} --op_def_cc_file ${cinn_op_source_file_tmp})
+      ${cinn_op_header_file_tmp} --op_info_file ${cinn_op_info_file_tmp}
+      --op_def_cc_file ${cinn_op_source_file_tmp})
 
-  set(generated_files_cinn_op "${cinn_op_header_file}" "${cinn_op_source_file}")
+  set(generated_files_cinn_op "${cinn_op_header_file}" "${cinn_op_info_file}"
+                              "${cinn_op_source_file}")
   foreach(generated_file ${generated_files_cinn_op})
     if(EXISTS "${generated_file}.tmp" AND EXISTS "${generated_file}")
       execute_process(COMMAND ${CMAKE_COMMAND} -E copy_if_different
@@ -61,6 +66,7 @@ if(NOT CINN_ONLY)
     SRCS
     op_dialect.cc
     ${cinn_op_source_file}
+    ${cinn_op_info_file}
     manual_op.cc
     op_attribute.cc
     DEPS

paddle/cinn/hlir/dialect/operator/ir/op_dialect.cc

@@ -35,7 +35,7 @@ void OperatorDialect::initialize() {
   // paddle/cinn/hlir/dialect/CMakeLists.txt.
   RegisterOps<
 #define GET_OP_LIST
-#include "paddle/cinn/hlir/dialect/operator/ir/cinn_op.cc"  // NOLINT
+#include "paddle/cinn/hlir/dialect/operator/ir/cinn_op_info.cc"  // NOLINT
       >();
   RegisterOp<GroupOp>();
   RegisterOp<ConcatOp>();

paddle/common/errors.cc

@@ -58,6 +58,9 @@ std::string error_name(ErrorCode code) {
     case ErrorCode::EXTERNAL:
       return "ExternalError";
       break;
+    case ErrorCode::INVALID_TYPE:
+      return "InvalidTypeError";
+      break;
     default:
       throw std::invalid_argument("The error type is undefined.");
       break;

paddle/common/errors.h

@@ -85,6 +85,10 @@ enum ErrorCode {
   // Third-party library error.
   // Error type string: "ExternalError"
   EXTERNAL = 12,
+
+  // Client specified an unmatched type.
+  // Error type string: "INVALID_TYPE"
+  INVALID_TYPE = 13,
 };
 
 class ErrorSummary {
@@ -140,6 +144,7 @@ REGISTER_ERROR(Unimplemented, ErrorCode::UNIMPLEMENTED)
 REGISTER_ERROR(Unavailable, ErrorCode::UNAVAILABLE)
 REGISTER_ERROR(Fatal, ErrorCode::FATAL)
 REGISTER_ERROR(External, ErrorCode::EXTERNAL)
+REGISTER_ERROR(InvalidType, ErrorCode::INVALID_TYPE)
 
 #undef REGISTER_ERROR
 

paddle/fluid/eager/CMakeLists.txt

@@ -17,14 +17,7 @@ if(NOT (NOT WITH_PYTHON AND ON_INFER))
   set(eager_deps ${eager_deps} accumulation_node prim_utils)
 endif()
 
-set(fluid_deps
-    tracer
-    layer
-    proto_desc
-    operator
-    op_registry
-    variable_helper
-    memcpy)
+set(fluid_deps tracer layer proto_desc operator op_registry variable_helper)
 set(generated_deps final_dygraph_function final_dygraph_node dygraph_function
                    dygraph_node)
 
@@ -73,7 +66,6 @@ cc_library(
        operator
        op_registry
        variable_helper
-       memcpy
        generated_op
        autograd_meta
        hook_utils)

paddle/fluid/eager/to_static/run_program_op_func.h

@@ -129,6 +129,10 @@ static std::vector<paddle::Tensor> Trans2ContiguousTensors(
   return res;
 }
 
+inline int64_t hash_int_value(int64_t value) {
+  return value + 0x9e3779b9 + (value << 6) + (value >> 2);
+}
+
 inline void run_program_ad_func(
     const std::vector<paddle::Tensor>& x,
     const std::vector<paddle::Tensor>& params,
@@ -155,7 +159,22 @@ inline void run_program_ad_func(
   auto params_tmp = Trans2ContiguousTensors(params);
   // Call forward function
   // if require_any_grad is False, don't save any middle vars.
-  RunProgramAPI(x_tmp, params_tmp, out, step_scope, require_any_grad, attrs);
+  std::vector<int64_t> place_hash_keys = std::vector<int64_t>();
+  for (const paddle::Tensor& tensor : x) {
+    int64_t device_type = static_cast<int64_t>(tensor.place().GetType());
+    place_hash_keys.emplace_back(hash_int_value(device_type));
+  }
+  for (const paddle::Tensor& tensor : params) {
+    int64_t device_type = static_cast<int64_t>(tensor.place().GetType());
+    place_hash_keys.emplace_back(hash_int_value(device_type));
+  }
+  RunProgramAPI(x_tmp,
+                params_tmp,
+                out,
+                step_scope,
+                require_any_grad,
+                attrs,
+                place_hash_keys);
   VLOG(2) << "start run run_program grad";
   auto is_test = false;
   if (attrs.count("is_test")) {
@@ -168,6 +187,9 @@ inline void run_program_ad_func(
     // Create GradOpNode (1 means [out_grad], 2 means [x_grad, paramx_grad])
     auto grad_node = std::make_shared<GradNodeRunProgram>(1, 2);
 
+    // Set place hash keys for backward
+    grad_node->SetPlaceHashKeys(place_hash_keys);
+
     // Set Attributes
     grad_node->SetAttrMap(attrs);
 
@@ -266,9 +288,27 @@ inline void pir_run_program_ad_func(
 
   // Call forward function
   // if require_any_grad is False, don't save any middle vars.
-  PirRunProgramAPI(
-      x, params, out, middles, step_scope, require_any_grad, attrs);
+  std::vector<int64_t> place_hash_keys = std::vector<int64_t>();
+  for (const paddle::Tensor& tensor : x) {
+    int64_t device_type = static_cast<int64_t>(tensor.place().GetType());
+    place_hash_keys.emplace_back(hash_int_value(device_type));
+  }
+  for (const paddle::Tensor& tensor : params) {
+    int64_t device_type = static_cast<int64_t>(tensor.place().GetType());
+    place_hash_keys.emplace_back(hash_int_value(device_type));
+  }
+  PirRunProgramAPI(x,
+                   params,
+                   out,
+                   middles,
+                   step_scope,
+                   require_any_grad,
+                   attrs,
+                   place_hash_keys);
   if (!is_test && require_any_grad) {
+    // Set place hash keys for backward
+    grad_node->SetPlaceHashKeys(place_hash_keys);
+
     // Set Attributes
     grad_node->SetAttrMap(attrs);