Add StringTensor (#39830)

* add string tensor and case convert kernels * Add strings empty kernel; Reorganize the structure of case convert kernel * Add string infermeta * Update mutable_data of string tensor * rename kernel name * add string copy tmp * Fix strings copy device bug * add utf8 gpu converter * add string tensor c++ api * Remove mutable_data of string tensor * update string tensor interface * remove charcases_flag.h * remove some fluid headers * Add make_ddim * __HIPCC__ -> PADDLE_WITH_HIP * remove fluid headers * fix cpu compile * remove std::hash * Fix cudaMalloc * Remove strings/impl directory * Fix infrt/get_phi_kernel_info.py;Add custom_kernels deps * Add empty kernel test * Remove some comments * Modify lower/upper api encoding type: string->bool * STRING->PSTRING; Add CreateInferLikeMeta * Add code gen for C++ String API * remove strings_api_utils.h * Add ignore file (strings_api.h, strings_api.cc) * update strings gen script * change args order of case convert kernels * Add comments for pstring, StringTensor * cpstring_internal.h -> cpstring_impl.h * Update accordding to comments: 1. Remove fluid headers 2. paddle::platform::errors -> phi::errors 3. Use 'place.GetType() == phi::AllocationType::GPU' instead of 'paddle::platform::is_cpu_space()' 4. Use camel code style * Remove all singletons in strings kernels * fix rocm compile * Fix py3 compile * Fix c++ coverage * 1. Add pstring proto type 2. Add StringTensor debug info 3. Rename case_convert_kernel to strings_lower_upper 4. Remove serialize derialize strings kernel * DataLayout::PSTRING -> DataLayout::PSTRING_UNION * Register pstring data type * Fix strings api gen * Fix dense tensor register pstring dtype * Fix error messages * remove line * add pstring unittest * remove test string api unitest * remove empty line * Remove some headers to decrease the size of executable file

Author: joey12300

.gitignore

@@ -9,10 +9,12 @@ paddle/phi/api/backward/backward_api.h
 paddle/phi/api/backward/sparse_bw_api.h
 paddle/phi/api/include/api.h
 paddle/phi/api/include/sparse_api.h
+paddle/phi/api/include/strings_api.h
 paddle/phi/api/lib/api.cc
 paddle/phi/api/lib/dygraph_api.*
 paddle/phi/api/lib/backward_api.cc
 paddle/phi/api/lib/sparse_api.cc
+paddle/phi/api/lib/strings_api.cc
 paddle/phi/api/lib/sparse_bw_api.cc
 paddle/phi/extension.h
 paddle/phi/include/*

paddle/fluid/framework/convert_utils.cc

@@ -47,6 +47,8 @@ paddle::experimental::DataType TransToPhiDataType(
  return DataType::BFLOAT16;
  case paddle::framework::proto::VarType::BOOL:
  return DataType::BOOL;
+ case paddle::framework::proto::VarType::PSTRING:
+ return DataType::PSTRING;
  default:
  return DataType::UNDEFINED;
  }
@@ -81,6 +83,8 @@ paddle::framework::proto::VarType::Type TransToProtoVarType(
  return paddle::framework::proto::VarType::BF16;
  case DataType::BOOL:
  return paddle::framework::proto::VarType::BOOL;
+ case DataType::PSTRING:
+ return paddle::framework::proto::VarType::PSTRING;
  default:
  PADDLE_THROW(paddle::platform::errors::Unimplemented(
  "Unsupported data type `%s` when casting it into "
@@ -117,6 +121,8 @@ size_t DataTypeSize(DataType dtype) {
  return sizeof(paddle::platform::complex<float>);
  case DataType::COMPLEX128:
  return sizeof(paddle::platform::complex<double>);
+ case DataType::PSTRING:
+ return sizeof(paddle::platform::pstring);
  default:
  return 0;
  }
@@ -145,6 +151,8 @@ DataType String2DataType(const std::string& str) {
  return DataType::COMPLEX64;
  } else if (str == "complex128") {
  return DataType::COMPLEX128;
+ } else if (str == "pstring") {
+ return DataType::PSTRING;
  } else if (str == "bfloat16") {
  return DataType::BFLOAT16;
  } else {
@@ -176,6 +184,8 @@ std::string DataType2String(DataType dtype) {
  return "complex64";
  case DataType::COMPLEX128:
  return "complex128";
+ case DataType::PSTRING:
+ return "pstring";
  case DataType::BFLOAT16:
  return "bfloat16";
  default:

paddle/fluid/framework/data_type.cc

@@ -18,9 +18,11 @@
 
 #include "paddle/fluid/platform/bfloat16.h"
 #include "paddle/fluid/platform/float16.h"
+#include "paddle/phi/common/pstring.h"
 
 using float16 = paddle::platform::float16;
 using bfloat16 = paddle::platform::bfloat16;
+using pstring = phi::dtype::pstring;
 
 namespace paddle {
 namespace framework {
@@ -58,7 +60,8 @@ static DataTypeMap* InitDataTypeMap() {
  RegisterType<cc_type>(retv, proto_type, #cc_type)
 
  _ForEachDataType_(RegType);
-
+ // Register pstring individually
+ RegType(pstring, proto::VarType::PSTRING);
 #undef RegType
  return retv;
 }

paddle/fluid/framework/framework.proto

@@ -152,6 +152,8 @@ message VarType {
  STRINGS = 26;
  VOCAB = 27;
  FEED_LIST = 28;
+ // The data type of phi::StringTensor
+ PSTRING = 29;
  }
 
  required Type type = 1;

paddle/phi/CMakeLists.txt

@@ -23,7 +23,7 @@ add_subdirectory(tools)
 add_subdirectory(tests)
 
 # make an unity target for compile deps
-set(PHI_DEPS convert_utils dense_tensor phi_context kernel_factory kernel_context arg_map_context infermeta lod_utils op_compat_infos sparse_csr_tensor sparse_coo_tensor)
+set(PHI_DEPS convert_utils dense_tensor phi_context kernel_factory kernel_context arg_map_context infermeta lod_utils op_compat_infos sparse_csr_tensor sparse_coo_tensor string_tensor)
 get_property(phi_kernels GLOBAL PROPERTY PHI_KERNELS)
 set(PHI_DEPS ${PHI_DEPS} ${phi_kernels})
 

paddle/phi/api/CMakeLists.txt

@@ -1,2 +1,2 @@
 add_subdirectory(lib)
-cc_library(phi_api SRCS all.cc DEPS phi_function_api phi_bw_function_api sparse_api sparse_bw_api)
+cc_library(phi_api SRCS all.cc DEPS phi_function_api phi_bw_function_api sparse_api sparse_bw_api strings_api)

paddle/phi/api/lib/CMakeLists.txt

@@ -51,6 +51,14 @@ set(sparse_bw_api_source_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/sparse_bw_a
 set(sparse_bw_api_header_file_tmp ${sparse_bw_api_header_file}.tmp)
 set(sparse_bw_api_source_file_tmp ${sparse_bw_api_source_file}.tmp)
 
+# strings api file
+set(strings_api_gen_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/strings_api_gen.py)
+set(strings_api_yaml_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/strings_api.yaml)
+set(strings_api_header_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/include/strings_api.h)
+set(strings_api_source_file ${CMAKE_SOURCE_DIR}/paddle/phi/api/lib/strings_api.cc)
+set(strings_api_header_file_tmp ${strings_api_header_file}.tmp)
+set(strings_api_source_file_tmp ${strings_api_source_file}.tmp)
+
 # wrapped infermeta file
 set(wrapped_infermeta_gen_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/wrapped_infermeta_gen.py)
 set(api_yaml_file ${CMAKE_SOURCE_DIR}/python/paddle/utils/code_gen/api.yaml)
@@ -114,6 +122,19 @@ add_custom_command(
  DEPENDS ${sparse_bw_api_yaml_file} ${sparse_bw_api_gen_file} ${api_gen_base} ${api_gen_file} ${sparse_api_gen_file} ${bw_api_gen_file}
  VERBATIM)
 
+# generate strings api
+add_custom_command(
+ OUTPUT ${strings_api_header_file} ${strings_api_source_file}
+ COMMAND ${PYTHON_EXECUTABLE} ${strings_api_gen_file}
+ --api_yaml_path ${strings_api_yaml_file}
+ --api_header_path ${strings_api_header_file_tmp}
+ --api_source_path ${strings_api_source_file_tmp}
+ COMMAND ${CMAKE_COMMAND} -E copy_if_different ${strings_api_header_file_tmp} ${strings_api_header_file}
+ COMMAND ${CMAKE_COMMAND} -E copy_if_different ${strings_api_source_file_tmp} ${strings_api_source_file}
+ COMMENT "copy_if_different ${strings_api_header_file} ${strings_strings_api_source_file}"
+ DEPENDS ${strings_api_yaml_file} ${strings_api_gen_file} ${api_gen_base} ${api_gen_file}
+ VERBATIM)
+
 # generate dygraph(intermediate) api
 add_custom_command(
  OUTPUT ${dygraph_api_header_file} ${dygraph_api_source_file}
@@ -152,5 +173,5 @@ cc_library(phi_bw_function_api SRCS ${bw_api_source_file} DEPS phi_tensor_raw ph
 cc_library(sparse_api SRCS ${sparse_api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils sparse_api_custom_impl)
 cc_library(sparse_bw_api SRCS ${sparse_bw_api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils sparse_api sparse_api_custom_impl)
 cc_library(phi_dygraph_api SRCS ${dygraph_api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils phi_data_transform phi_function_api sparse_api)
-
-cc_library(phi_tensor SRCS tensor_method.cc DEPS phi_tensor_raw phi_function_api api_gen_utils kernel_dispatch infermeta sparse_api)
+cc_library(strings_api SRCS ${strings_api_source_file} DEPS phi_tensor_raw phi kernel_dispatch api_gen_utils)
+cc_library(phi_tensor SRCS tensor_method.cc DEPS phi_tensor_raw phi_function_api api_gen_utils kernel_dispatch infermeta sparse_api strings_api)

paddle/phi/api/lib/api_declare.h

@@ -19,3 +19,4 @@ limitations under the License. */
 
 // PD_DECLARE_API(Math);
 // PD_DECLARE_API(SparseApi);
+// PD_DECLARE_API(StringsApi);

paddle/phi/api/lib/api_gen_utils.cc

@@ -56,6 +56,10 @@ std::shared_ptr<phi::SelectedRows> TensorToSelectedRows(
  return nullptr;
 }
 
+std::shared_ptr<phi::StringTensor> TensorToStringTensor(const Tensor& tensor) {
+ return std::dynamic_pointer_cast<phi::StringTensor>(tensor.impl());
+}
+
 /* ----------------- for infer_meta --------------------- */
 
 phi::MetaTensor MakeMetaTensor(const phi::DenseTensor& tensor) {
@@ -92,6 +96,10 @@ paddle::optional<phi::MetaTensor> MakeMetaTensor(
  return {paddle::none};
 }
 
+phi::MetaTensor MakeMetaTensor(const phi::StringTensor& tensor) {
+ return phi::MetaTensor(tensor);
+}
+
 /* ------------------ for output ----------------------- */
 
 phi::DenseTensor* SetKernelOutput(Backend backend, Tensor* out) {
@@ -148,5 +156,20 @@ phi::TensorBase* SetSparseKernelOutput(Tensor* out, TensorType type) {
  return out->impl().get();
 }
 
+phi::TensorBase* SetStringsKernelOutput(Backend backend,
+ Tensor* out,
+ TensorType type) {
+ if (!out->initialized()) {
+ if (type == TensorType::STRING_TENSOR) {
+ if (out->impl() == nullptr) {
+ auto strings_tensor = std::make_shared<phi::StringTensor>();
+ out->set_impl(strings_tensor);
+ }
+ return out->impl().get();
+ }
+ }
+ return out->impl().get();
+}
+
 } // namespace experimental
 } // namespace paddle

paddle/phi/api/lib/api_gen_utils.h

@@ -22,11 +22,12 @@ limitations under the License. */
 #include "paddle/phi/core/selected_rows.h"
 #include "paddle/phi/core/sparse_coo_tensor.h"
 #include "paddle/phi/core/sparse_csr_tensor.h"
+#include "paddle/phi/core/string_tensor.h"
 
 namespace paddle {
 namespace experimental {
 
-enum class TensorType { DENSE_TENSOR, SPARSE_CSR, SPARSE_COO };
+enum class TensorType { DENSE_TENSOR, SPARSE_CSR, SPARSE_COO, STRING_TENSOR };
 
 /* ------------------ for input ----------------------- */
 
@@ -43,6 +44,8 @@ std::shared_ptr<phi::SelectedRows> TensorToSelectedRows(const Tensor& tensor);
 std::shared_ptr<phi::SelectedRows> TensorToSelectedRows(
  const paddle::optional<Tensor>& tensor);
 
+std::shared_ptr<phi::StringTensor> TensorToStringTensor(const Tensor& tensor);
+
 /* ----------------- for infer_meta --------------------- */
 
 phi::MetaTensor MakeMetaTensor(const phi::DenseTensor& tensor);
@@ -58,6 +61,8 @@ phi::MetaTensor MakeMetaTensor(const phi::SelectedRows& tensor);
 paddle::optional<phi::MetaTensor> MakeMetaTensor(
  const paddle::optional<const phi::SelectedRows&>& tensor);
 
+phi::MetaTensor MakeMetaTensor(const phi::StringTensor& tensor);
+
 /* ------------------ for output ----------------------- */
 
 phi::DenseTensor* SetKernelOutput(Backend backend, Tensor* out);
@@ -70,5 +75,9 @@ phi::SelectedRows* SetSelectedRowsKernelOutput(Backend backend, Tensor* out);
 
 phi::TensorBase* SetSparseKernelOutput(Tensor* out, TensorType type);
 
+phi::TensorBase* SetStringsKernelOutput(Backend backend,
+ Tensor* out,
+ TensorType type);
+
 } // namespace experimental
 } // namespace paddle