Restore "Add mkldnn bfloat16 option to C-API " (#26882)

* Add mkldnn bfloat16 option to C-API

* Add test for bfloat16 gpu

* Change coverage test

* Repair capi_gpu test

go/paddle/config.go

@@ -154,10 +154,17 @@ func (config *AnalysisConfig) EnableMkldnnQuantizer() {
 	C.PD_EnableMkldnnQuantizer(config.c)
 }
 
+func (config *AnalysisConfig) EnableMkldnnBfloat16() {
+	C.PD_EnableMkldnnBfloat16(config.c)
+}
+
 func (config *AnalysisConfig) MkldnnQuantizerEnabled() bool {
 	return ConvertCBooleanToGo(C.PD_MkldnnQuantizerEnabled(config.c))
 }
 
+func (config *AnalysisConfig) MkldnnBfloat16Enabled() bool {
+	return ConvertCBooleanToGo(C.PD_MkldnnBfloat16Enabled(config.c))
+}
 // SetModelBuffer
 // ModelFromMemory
 

paddle/fluid/inference/api/analysis_config.cc

@@ -217,6 +217,17 @@ void AnalysisConfig::EnableMkldnnQuantizer() {
   Update();
 }
 
+void AnalysisConfig::EnableMkldnnBfloat16() {
+#ifdef PADDLE_WITH_MKLDNN
+  use_mkldnn_bfloat16_ = true;
+#else
+  LOG(ERROR) << "Please compile with MKLDNN first to use MkldnnBfloat16";
+  use_mkldnn_bfloat16_ = false;
+#endif
+
+  Update();
+}
+
 MkldnnQuantizerConfig *AnalysisConfig::mkldnn_quantizer_config() const {
   PADDLE_ENFORCE_NOT_NULL(mkldnn_quantizer_config_,
                           "MkldnnQuantizer was not enabled yet.");
@@ -330,6 +341,12 @@ void AnalysisConfig::Update() {
 #endif
   }
 
+  if (use_mkldnn_bfloat16_) {
+#ifdef PADDLE_WITH_MKLDNN
+    pass_builder()->EnableMkldnnBfloat16();
+#endif
+  }
+
 #ifdef PADDLE_WITH_MKLDNN
   // Do not optimize when mkldnn is on
   if (enable_memory_optim_ && !use_mkldnn_) {
@@ -398,6 +415,7 @@ std::string AnalysisConfig::SerializeInfoCache() {
   ss << ";";
 
   ss << use_mkldnn_quantizer_;
+  ss << use_mkldnn_bfloat16_;
   ss << model_from_memory_;
 
   ss << with_profile_;

paddle/fluid/inference/api/analysis_predictor_tester.cc

@@ -485,4 +485,25 @@ TEST_F(MkldnnQuantizerTest, kl_scaling_factor_unsigned) {
 }
 #endif
 
+#ifdef PADDLE_WITH_CUDA
+TEST(AnalysisPredictor, bf16_gpu_pass_strategy) {
+  AnalysisConfig config;
+  config.SetModel(FLAGS_dirname);
+  config.SwitchIrOptim(true);
+  config.EnableUseGpu(100, 0);
+  config.EnableMkldnnBfloat16();
+#ifdef PADDLE_WITH_MKLDNN
+  ASSERT_EQ(config.mkldnn_bfloat16_enabled(), true);
+#else
+  ASSERT_EQ(config.mkldnn_bfloat16_enabled(), false);
+#endif
+}
+#endif
+
+TEST(AnalysisPredictor, bf16_pass_strategy) {
+  std::vector<std::string> passes;
+  PassStrategy passStrategy(passes);
+  passStrategy.EnableMkldnnBfloat16();
+}
+
 }  // namespace paddle

paddle/fluid/inference/api/paddle_analysis_config.h

@@ -401,6 +401,19 @@ struct PD_INFER_DECL AnalysisConfig {
   ///
   void EnableMkldnnQuantizer();
 
+  ///
+  /// \brief Turn on MKLDNN bfloat16.
+  ///
+  ///
+  void EnableMkldnnBfloat16();
+
+  ///
+  /// \brief A boolean state telling whether to use the MKLDNN Bfloat16.
+  ///
+  /// \return bool Whether to use the MKLDNN Bfloat16.
+  ///
+  bool mkldnn_bfloat16_enabled() const { return use_mkldnn_bfloat16_; }
+
   ///
   /// \brief A boolean state telling whether the thread local CUDA stream is
   /// enabled.
@@ -592,6 +605,7 @@ struct PD_INFER_DECL AnalysisConfig {
   int mkldnn_cache_capacity_{0};
   bool use_mkldnn_quantizer_{false};
   std::shared_ptr<MkldnnQuantizerConfig> mkldnn_quantizer_config_;
+  bool use_mkldnn_bfloat16_{false};
 
   // If the config is already used on a predictor, it becomes invalid.
   // Any config can only be used with one predictor.

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -143,6 +143,10 @@ void GpuPassStrategy::EnableMkldnnQuantizer() {
   LOG(ERROR) << "GPU not support MKL-DNN quantization";
 }
 
+void GpuPassStrategy::EnableMkldnnBfloat16() {
+  LOG(ERROR) << "GPU not support MKL-DNN bfloat16";
+}
+
 CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
   // NOTE the large fusions should be located in the front, so that they will
   // not be damaged by smaller ones.
@@ -225,4 +229,12 @@ void CpuPassStrategy::EnableMkldnnQuantizer() {
 #endif
 }
 
+void CpuPassStrategy::EnableMkldnnBfloat16() {
+#ifdef PADDLE_WITH_MKLDNN
+  use_mkldnn_bfloat16_ = true;
+#else
+  use_mkldnn_bfloat16_ = false;
+#endif
+}
+
 }  // namespace paddle

paddle/fluid/inference/api/paddle_pass_builder.h

@@ -132,6 +132,9 @@ class PD_INFER_DECL PassStrategy : public PaddlePassBuilder {
   /// \brief Enable MKLDNN quantize optimization.
   virtual void EnableMkldnnQuantizer() {}
 
+  /// \brief Enable MKLDNN bfloat16.
+  virtual void EnableMkldnnBfloat16() {}
+
   /// \brief Check if we are using gpu.
   /// \return A bool variable implying whether we are in gpu mode.
   bool use_gpu() const { return use_gpu_; }
@@ -161,6 +164,7 @@ class PD_INFER_DECL CpuPassStrategy : public PassStrategy {
     use_gpu_ = other.use_gpu_;
     use_mkldnn_ = other.use_mkldnn_;
     use_mkldnn_quantizer_ = other.use_mkldnn_quantizer_;
+    use_mkldnn_bfloat16_ = other.use_mkldnn_bfloat16_;
   }
   /// \brief Default destructor.
   virtual ~CpuPassStrategy() = default;
@@ -174,9 +178,13 @@ class PD_INFER_DECL CpuPassStrategy : public PassStrategy {
   /// \brief Enable MKLDNN quantize optimization.
   void EnableMkldnnQuantizer() override;
 
+  /// \brief Enable MKLDNN bfloat16.
+  void EnableMkldnnBfloat16() override;
+
  protected:
   /// \cond Protected
   bool use_mkldnn_quantizer_{false};
+  bool use_mkldnn_bfloat16_{false};
   /// \endcond
 };
 
@@ -205,6 +213,9 @@ class PD_INFER_DECL GpuPassStrategy : public PassStrategy {
   /// \brief Not supported in GPU mode yet.
   void EnableMkldnnQuantizer() override;
 
+  /// \brief Not supported in GPU mode yet.
+  void EnableMkldnnBfloat16() override;
+
   /// \brief Default destructor.
   virtual ~GpuPassStrategy() = default;
 

paddle/fluid/inference/capi/paddle_c_api.h

@@ -235,6 +235,12 @@ PADDLE_CAPI_EXPORT extern void PD_EnableMkldnnQuantizer(
 PADDLE_CAPI_EXPORT extern bool PD_MkldnnQuantizerEnabled(
     const PD_AnalysisConfig* config);
 
+PADDLE_CAPI_EXPORT extern void PD_EnableMkldnnBfloat16(
+    PD_AnalysisConfig* config);
+
+PADDLE_CAPI_EXPORT extern bool PD_MkldnnBfloat16Enabled(
+    const PD_AnalysisConfig* config);
+
 PADDLE_CAPI_EXPORT extern void PD_SetModelBuffer(PD_AnalysisConfig* config,
                                                  const char* prog_buffer,
                                                  size_t prog_buffer_size,

paddle/fluid/inference/capi/pd_config.cc

@@ -207,6 +207,18 @@ bool PD_MkldnnQuantizerEnabled(const PD_AnalysisConfig* config) {
   return config->config.mkldnn_quantizer_enabled();
 }
 
+void PD_EnableMkldnnBfloat16(PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config, paddle::platform::errors::NotFound(
+                                      "PD_AnalysisConfig should not be null"));
+  config->config.EnableMkldnnBfloat16();
+}
+
+bool PD_MkldnnBfloat16Enabled(const PD_AnalysisConfig* config) {
+  PADDLE_ENFORCE_NOT_NULL(config, paddle::platform::errors::NotFound(
+                                      "PD_AnalysisConfig should not be null"));
+  return config->config.mkldnn_bfloat16_enabled();
+}
+
 void PD_SetModelBuffer(PD_AnalysisConfig* config, const char* prog_buffer,
                        size_t prog_buffer_size, const char* params_buffer,
                        size_t params_buffer_size) {

paddle/fluid/inference/tests/api/analyzer_capi_tester.cc

@@ -88,6 +88,9 @@ TEST(PD_AnalysisConfig, profile_mkldnn) {
   PD_EnableMkldnnQuantizer(config);
   bool quantizer_enable = PD_MkldnnQuantizerEnabled(config);
   CHECK(quantizer_enable) << "NO";
+  PD_EnableMkldnnBfloat16(config);
+  bool bfloat16_enable = PD_MkldnnBfloat16Enabled(config);
+  CHECK(bfloat16_enable) << "NO";
   PD_SetMkldnnCacheCapacity(config, 0);
   PD_SetModel(config, prog_file.c_str(), params_file.c_str());
   PD_DeleteAnalysisConfig(config);

paddle/fluid/pybind/inference_api.cc

@@ -448,6 +448,7 @@ void BindAnalysisConfig(py::module *m) {
            &AnalysisConfig::cpu_math_library_num_threads)
       .def("to_native_config", &AnalysisConfig::ToNativeConfig)
       .def("enable_quantizer", &AnalysisConfig::EnableMkldnnQuantizer)
+      .def("enable_mkldnn_bfloat16", &AnalysisConfig::EnableMkldnnBfloat16)
 #ifdef PADDLE_WITH_MKLDNN
       .def("quantizer_config", &AnalysisConfig::mkldnn_quantizer_config,
            py::return_value_policy::reference)
@@ -565,6 +566,7 @@ void BindPaddlePassBuilder(py::module *m) {
       .def("enable_cudnn", &PassStrategy::EnableCUDNN)
       .def("enable_mkldnn", &PassStrategy::EnableMKLDNN)
       .def("enable_mkldnn_quantizer", &PassStrategy::EnableMkldnnQuantizer)
+      .def("enable_mkldnn_bfloat16", &PassStrategy::EnableMkldnnBfloat16)
       .def("use_gpu", &PassStrategy::use_gpu);
 
   py::class_<CpuPassStrategy, PassStrategy>(*m, "CpuPassStrategy")
@@ -572,14 +574,16 @@ void BindPaddlePassBuilder(py::module *m) {
       .def(py::init<const CpuPassStrategy &>())
       .def("enable_cudnn", &CpuPassStrategy::EnableCUDNN)
       .def("enable_mkldnn", &CpuPassStrategy::EnableMKLDNN)
-      .def("enable_mkldnn_quantizer", &CpuPassStrategy::EnableMkldnnQuantizer);
+      .def("enable_mkldnn_quantizer", &CpuPassStrategy::EnableMkldnnQuantizer)
+      .def("enable_mkldnn_bfloat16", &CpuPassStrategy::EnableMkldnnBfloat16);
 
   py::class_<GpuPassStrategy, PassStrategy>(*m, "GpuPassStrategy")
       .def(py::init<>())
       .def(py::init<const GpuPassStrategy &>())
       .def("enable_cudnn", &GpuPassStrategy::EnableCUDNN)
       .def("enable_mkldnn", &GpuPassStrategy::EnableMKLDNN)
-      .def("enable_mkldnn_quantizer", &GpuPassStrategy::EnableMkldnnQuantizer);
+      .def("enable_mkldnn_quantizer", &GpuPassStrategy::EnableMkldnnQuantizer)
+      .def("enable_mkldnn_bfloat16", &GpuPassStrategy::EnableMkldnnBfloat16);
 }
 }  // namespace
 }  // namespace pybind