paddle/fluid/pybind/inference_api.cc add onednn [fluid_ops]

paddle/fluid/pybind/inference_api.cc

@@ -1050,26 +1050,35 @@ void BindAnalysisConfig(py::module *m) {
  &AnalysisConfig::SwitchIrDebug,
  py::arg("x") = true,
  py::arg("passes") = std::vector<std::string>())
- .def("enable_mkldnn", &AnalysisConfig::EnableONEDNN)
- .def("disable_mkldnn", &AnalysisConfig::DisableONEDNN)
- .def("mkldnn_enabled", &AnalysisConfig::onednn_enabled)
+ .def("enable_mkldnn", &AnalysisConfig::EnableONEDNN) // deprecated
+ .def("disable_mkldnn", &AnalysisConfig::DisableONEDNN) // deprecated
+ .def("mkldnn_enabled", &AnalysisConfig::onednn_enabled) // deprecated
+ .def("enable_onednn", &AnalysisConfig::EnableONEDNN)
+ .def("disable_onednn", &AnalysisConfig::DisableONEDNN)
+ .def("onednn_enabled", &AnalysisConfig::onednn_enabled)
  .def("enable_cinn", &AnalysisConfig::EnableCINN)
  .def("set_cpu_math_library_num_threads",
  &AnalysisConfig::SetCpuMathLibraryNumThreads)
  .def("cpu_math_library_num_threads",
  &AnalysisConfig::cpu_math_library_num_threads)
  .def("to_native_config", &AnalysisConfig::ToNativeConfig)
- .def("enable_mkldnn_bfloat16", &AnalysisConfig::EnableOnednnBfloat16)
+ .def("enable_mkldnn_bfloat16",
+ &AnalysisConfig::EnableOnednnBfloat16) // deprecated
+ .def("enable_onednn_bfloat16", &AnalysisConfig::EnableOnednnBfloat16)
 #ifdef PADDLE_WITH_DNNL
  .def("set_mkldnn_cache_capacity",
+ &AnalysisConfig::SetOnednnCacheCapacity,
+ py::arg("capacity") = 0) // deprecated
+ .def("set_onednn_cache_capacity",
  &AnalysisConfig::SetOnednnCacheCapacity,
  py::arg("capacity") = 0)
  .def("set_bfloat16_op", &AnalysisConfig::SetBfloat16Op)
  .def("enable_mkldnn_int8",
  &AnalysisConfig::EnableOnednnInt8,
  py::arg("mkldnn_int8_enabled_op_types") =
- std::unordered_set<std::string>({}))
- .def("mkldnn_int8_enabled", &AnalysisConfig::onednn_int8_enabled)
+ std::unordered_set<std::string>({})) // deprecated
+ .def("mkldnn_int8_enabled",
+ &AnalysisConfig::onednn_int8_enabled) // deprecated
  .def("disable_mkldnn_fc_passes",
  &AnalysisConfig::DisableOnednnFcPasses,
  R"DOC(
@@ -1085,9 +1094,31 @@ void BindAnalysisConfig(py::module *m) {
  >>> config = Config("")
  >>> config.enable_mkldnn()
  >>> config.disable_mkldnn_fc_passes()
+ )DOC") // deprecated
+ .def("enable_onednn_int8",
+ &AnalysisConfig::EnableOnednnInt8,
+ py::arg("onednn_int8_enabled_op_types") =
+ std::unordered_set<std::string>({}))
+ .def("onednn_int8_enabled", &AnalysisConfig::onednn_int8_enabled)
+ .def("disable_onednn_fc_passes",
+ &AnalysisConfig::DisableOnednnFcPasses,
+ R"DOC(
+ Disable Onednn FC
+ Returns:
+ None.
+
+ Examples:
+ .. code-block:: python
+
+ >>> from paddle.inference import Config
+
+ >>> config = Config("")
+ >>> config.enable_onednn()
+ >>> config.disable_onednn_fc_passes()
  )DOC")
 #endif
- .def("set_mkldnn_op", &AnalysisConfig::SetONEDNNOp)
+ .def("set_mkldnn_op", &AnalysisConfig::SetONEDNNOp) // deprecated
+ .def("set_onednn_op", &AnalysisConfig::SetONEDNNOp)
  .def("set_model_buffer", &AnalysisConfig::SetModelBuffer)
  .def("model_from_memory", &AnalysisConfig::model_from_memory)
  .def("delete_pass", &AnalysisConfig::DeletePass)
@@ -1329,23 +1360,32 @@ void BindPaddlePassBuilder(py::module *m) {
  py::class_<PassStrategy, PaddlePassBuilder>(*m, "PassStrategy")
  .def(py::init<const std::vector<std::string> &>())
  .def("enable_cudnn", &PassStrategy::EnableCUDNN)
- .def("enable_mkldnn", &PassStrategy::EnableONEDNN)
- .def("enable_mkldnn_bfloat16", &PassStrategy::EnableMkldnnBfloat16)
+ .def("enable_mkldnn", &PassStrategy::EnableONEDNN) // deprecated
+ .def("enable_mkldnn_bfloat16",
+ &PassStrategy::EnableMkldnnBfloat16) // deprecated
+ .def("enable_onednn", &PassStrategy::EnableONEDNN)
+ .def("enable_onednn_bfloat16", &PassStrategy::EnableOnednnBfloat16)
  .def("use_gpu", &PassStrategy::use_gpu);
 
  py::class_<CpuPassStrategy, PassStrategy>(*m, "CpuPassStrategy")
  .def(py::init<>())
  .def(py::init<const CpuPassStrategy &>())
  .def("enable_cudnn", &CpuPassStrategy::EnableCUDNN)
- .def("enable_mkldnn", &CpuPassStrategy::EnableONEDNN)
- .def("enable_mkldnn_bfloat16", &CpuPassStrategy::EnableMkldnnBfloat16);
+ .def("enable_mkldnn", &CpuPassStrategy::EnableONEDNN) // deprecated
+ .def("enable_mkldnn_bfloat16",
+ &CpuPassStrategy::EnableMkldnnBfloat16) // deprecated
+ .def("enable_onednn", &CpuPassStrategy::EnableONEDNN)
+ .def("enable_onednn_bfloat16", &CpuPassStrategy::EnableOnednnBfloat16);
 
  py::class_<GpuPassStrategy, PassStrategy>(*m, "GpuPassStrategy")
  .def(py::init<>())
  .def(py::init<const GpuPassStrategy &>())
  .def("enable_cudnn", &GpuPassStrategy::EnableCUDNN)
- .def("enable_mkldnn", &GpuPassStrategy::EnableONEDNN)
- .def("enable_mkldnn_bfloat16", &GpuPassStrategy::EnableMkldnnBfloat16);
+ .def("enable_mkldnn", &GpuPassStrategy::EnableONEDNN) // deprecated
+ .def("enable_mkldnn_bfloat16",
+ &GpuPassStrategy::EnableMkldnnBfloat16) // deprecated
+ .def("enable_onednn", &GpuPassStrategy::EnableONEDNN)
+ .def("enable_onednn_bfloat16", &GpuPassStrategy::EnableOnednnBfloat16);
 }
 
 void BindInternalUtils(py::module *m) {

test/deprecated/ir/inference/inference_pass_test.py

@@ -178,9 +178,9 @@ def _get_analysis_config(
  config.enable_tensorrt_varseqlen()
 
  elif use_mkldnn:
- config.enable_mkldnn()
+ config.enable_onednn()
  if self.enable_mkldnn_bfloat16:
- config.enable_mkldnn_bfloat16()
+ config.enable_onednn_bfloat16()
  return config
 
  def check_output(self, atol=1e-3):

test/deprecated/ir/inference/quant_dequant_test.py

@@ -231,9 +231,9 @@ def _get_analysis_config(
  config.enable_tensorrt_varseqlen()
 
  elif use_mkldnn:
- config.enable_mkldnn()
+ config.enable_onednn()
  if self.enable_mkldnn_bfloat16:
- config.enable_mkldnn_bfloat16()
+ config.enable_onednn_bfloat16()
  return config
 
  def check_output_with_option(

test/deprecated/ir/inference/test_trt_inference_predictor_deprecated.py

@@ -98,14 +98,14 @@ def load(self, config_arg, inputs=None, outputs=None):
 
  if self.args.enable_mkldnn and not self.args.enable_gpu:
  config.disable_gpu()
- config.enable_mkldnn()
+ config.enable_onednn()
  if self.args.precision == 'int8':
- config.enable_mkldnn_int8(
+ config.enable_onednn_int8(
  {"conv2d", "depthwise_conv2d", "transpose2", "pool2d"}
  )
  if not self.args.enable_mkldnn and not self.args.enable_gpu:
  config.disable_gpu()
- # config.enable_mkldnn()
+ # config.enable_onednn()
  if self.args.enable_profile:
  config.enable_profile()
  shape_range_file = os.path.join(

test/ir/inference/inference_pass_test.py

@@ -179,9 +179,9 @@ def _get_analysis_config(
  config.enable_tensorrt_varseqlen()
 
  elif use_mkldnn:
- config.enable_mkldnn()
+ config.enable_onednn()
  if self.enable_mkldnn_bfloat16:
- config.enable_mkldnn_bfloat16()
+ config.enable_onednn_bfloat16()
  print('config summary:', config.summary())
  return config
 

test/ir/inference/quant_dequant_test.py

@@ -231,9 +231,9 @@ def _get_analysis_config(
  config.enable_tensorrt_varseqlen()
 
  elif use_mkldnn:
- config.enable_mkldnn()
+ config.enable_onednn()
  if self.enable_mkldnn_bfloat16:
- config.enable_mkldnn_bfloat16()
+ config.enable_onednn_bfloat16()
  print('config summary:', config.summary())
  return config
 

test/ir/inference/test_conv_eltwiseadd_bn_fuse_pass.py

@@ -47,7 +47,7 @@ def sample_predictor_configs(self, program_config):
 
  # MKLDNN
  config = self.create_inference_config(use_gpu=False)
- config.enable_mkldnn()
+ config.enable_onednn()
  yield config, ["conv2d", "elementwise_add"], (1e-4, 1e-5)
 
  # for gpu

test/mkldnn/test_onnx_format_quantization_mobilenetv1.py

@@ -208,12 +208,12 @@ def run_program(
  image_shape = [3, 224, 224]
  config = paddle.inference.Config(model_path)
  config.disable_gpu()
- config.enable_mkldnn()
+ config.enable_onednn()
  config.switch_ir_optim()
  config.set_cpu_math_library_num_threads(1)
  config.disable_glog_info()
  if is_quantized_model:
- config.enable_mkldnn_int8()
+ config.enable_onednn_int8()
  predictor = paddle.inference.create_predictor(config)
 
  input_names = predictor.get_input_names()

test/quantization/quant2_int8_lstm_model.py

@@ -112,7 +112,7 @@ def set_config(
  config.disable_gpu()
  config.switch_use_feed_fetch_ops(True)
  config.switch_ir_optim(True)
- config.enable_mkldnn()
+ config.enable_onednn()
  config.disable_mkldnn_fc_passes() # fc passes caused dnnl error
  config.pass_builder().insert_pass(5, "fc_lstm_fuse_pass")
  config.set_mkldnn_cache_capacity(mkldnn_cache_capacity)
@@ -121,7 +121,7 @@ def set_config(
  config.quantizer_config().set_quant_data(warmup_data)
  config.quantizer_config().set_quant_batch_size(1)
  elif mode == "qat":
- config.enable_mkldnn_int8()
+ config.enable_onednn_int8()
 
  return config
 

test/quantization/quant2_int8_nlp_comparison.py

@@ -158,9 +158,9 @@ def set_config(
  config.switch_specify_input_names(True)
  config.switch_ir_optim(True)
  config.switch_use_feed_fetch_ops(True)
- config.enable_mkldnn()
+ config.enable_onednn()
  if target == 'int8':
- config.enable_mkldnn_int8(self._quantized_ops)
+ config.enable_onednn_int8(self._quantized_ops)
  config.delete_pass(
  "constant_folding_pass"
  ) # same reason as in analyzer_ernie_int8_tester.cc