gpu inference mixed precision

include/api/context.h

@@ -116,8 +116,20 @@ class MS_API NvidiaGPUDeviceInfo : public DeviceInfoContext {
 
   void SetGpuTrtInferMode(bool gpu_trt_infer_mode);
   bool GetGpuTrtInferMode() const;
+
+  inline void SetPrecisionMode(const std::string &precison_mode);
+  inline std::string GetPrecisionMode() const;
+
+ private:
+  void SetPrecisionMode(const std::vector<char> &precision_mode);
+  std::vector<char> GetPrecisionModeChar() const;
 };
 
+void NvidiaGPUDeviceInfo::SetPrecisionMode(const std::string &precision_mode) {
+  SetPrecisionMode(StringToChar(precision_mode));
+}
+std::string NvidiaGPUDeviceInfo::GetPrecisionMode() const { return CharToString(GetPrecisionModeChar()); }
+
 class MS_API Ascend910DeviceInfo : public DeviceInfoContext {
  public:
   enum DeviceType GetDeviceType() const override { return DeviceType::kAscend910; };

mindspore/ccsrc/backend/optimizer/trt_pass/trt_converter_context.cc

@@ -198,6 +198,15 @@ bool TrtConverterContext::Serialize(std::string *model) {
   MS_EXCEPTION_IF_NULL(model);
   builder_->setMaxBatchSize(batch_size_);
   config_->setMaxWorkspaceSize(workspace_size_);
+
+  // Set precision mode
+  const auto &context = MsContext::GetInstance();
+  const auto &precision_mode = context->get_param<std::string>(MS_CTX_INFER_PRECISION_MODE);
+  if (precision_mode == "fp16") {
+    MS_LOG(WARNING) << "Inference with mixed precision mode. It will take few minutes for operators selection.";
+    config_->setFlag(nvinfer1::BuilderFlag::kFP16);
+  }
+
   engine_ = TrtPtr(builder_->buildEngineWithConfig(*network_, *config_));
   MS_EXCEPTION_IF_NULL(engine_);
 

mindspore/ccsrc/cxx_api/context.cc

@@ -27,6 +27,7 @@ constexpr auto kModelOptionKirinNpuFrequency = "mindspore.option.kirin_npu.frequ
 constexpr auto kModelOptionDeviceID = "mindspore.option.device_id";
 constexpr auto kModelOptionNvidiaGpuDeviceID = kModelOptionDeviceID;
 constexpr auto kModelOptionNvidiaGpuTrtInferMode = "mindspore.option.nvidia_gpu.trt_infer_mode";
+constexpr auto kModelOptionNvidiaGpuPrecisionMode = "mindspore.option.nvidia_gpu.precision_mode";
 constexpr auto kModelOptionAscend910DeviceID = kModelOptionDeviceID;
 constexpr auto kModelOptionAscend310DeviceID = kModelOptionDeviceID;
 constexpr auto kModelOptionAscend310DumpCfgPath = "mindspore.option.ascend310.dump_config_file_path";
@@ -153,6 +154,16 @@ bool NvidiaGPUDeviceInfo::GetGpuTrtInferMode() const {
   return GetValue<bool>(data_, kModelOptionNvidiaGpuTrtInferMode);
 }
 
+void NvidiaGPUDeviceInfo::SetPrecisionMode(const std::vector<char> &precision_mode) {
+  MS_EXCEPTION_IF_NULL(data_);
+  data_->params[kModelOptionNvidiaGpuPrecisionMode] = CharToString(precision_mode);
+}
+std::vector<char> NvidiaGPUDeviceInfo::GetPrecisionModeChar() const {
+  MS_EXCEPTION_IF_NULL(data_);
+  const std::string &ref = GetValue<std::string>(data_, kModelOptionNvidiaGpuPrecisionMode);
+  return StringToChar(ref);
+}
+
 void Ascend910DeviceInfo::SetDeviceID(uint32_t device_id) {
   MS_EXCEPTION_IF_NULL(data_);
   data_->params[kModelOptionAscend910DeviceID] = device_id;

mindspore/ccsrc/cxx_api/graph/gpu/gpu_graph_impl.cc

@@ -63,8 +63,8 @@ Status GPUGraphImpl::InitEnv() {
   if (gpu_info == nullptr) {
     return kMCDeviceError;
   }
-  auto enable_trt = gpu_info->GetGpuTrtInferMode();
-  ms_context->set_param<bool>(MS_CTX_ENABLE_INFER_OPT, enable_trt);
+  ms_context->set_param<bool>(MS_CTX_ENABLE_INFER_OPT, gpu_info->GetGpuTrtInferMode());
+  ms_context->set_param<std::string>(MS_CTX_INFER_PRECISION_MODE, gpu_info->GetPrecisionMode());
 
   session_impl_ = session::SessionFactory::Get().Create(kGpuInferenceDevice);
   if (session_impl_ == nullptr) {

mindspore/ccsrc/runtime/device/gpu/trt_loader.h

@@ -31,7 +31,7 @@ class TrtLoader {
 
   std::shared_ptr<nvinfer1::IBuilder> CreateInferBuilder(nvinfer1::ILogger *logger);
   std::shared_ptr<nvinfer1::IRuntime> CreateInferRuntime(nvinfer1::ILogger *logger);
-  bool nvinfer_loaded() { return nvinfer_loaded_; }
+  bool nvinfer_loaded() const { return nvinfer_loaded_; }
 
  private:
   bool nvinfer_loaded_;

mindspore/core/utils/ms_context.h

@@ -118,6 +118,7 @@ enum MsCtxParam : unsigned {
   MS_CTX_ENV_CONFIG_PATH,
   MS_CTX_TUNE_MODE,
   MS_CTX_GRAPH_KERNEL_FLAGS,
+  MS_CTX_INFER_PRECISION_MODE,  // GPU inference precision mode configured by Serving or Unify API.
   MS_CTX_TYPE_STRING_END,
 
   // parameter numbers of each type