add gpu ResetResource

mindspore/ccsrc/plugin/device/gpu/kernel/arrays/oneslike_gpu_kernel.h

@@ -25,7 +25,7 @@ namespace kernel {
 template <typename T>
 class OnesLikeGpuKernelMod : public NativeGpuKernelMod {
  public:
-  OnesLikeGpuKernelMod() : input_size_(0), output_size_(0), is_null_input_(false) {}
+  OnesLikeGpuKernelMod() : input_size_(0), output_size_(0), is_null_input_(false) { ResetResource(); }
   ~OnesLikeGpuKernelMod() override = default;
 
   bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
@@ -68,6 +68,15 @@ class OnesLikeGpuKernelMod : public NativeGpuKernelMod {
     return true;
   }
 
+  void ResetResource() noexcept override {
+    input_size_ = 0;
+    output_size_ = 0;
+    is_null_input_ = false;
+    input_size_list_.clear();
+    output_size_list_.clear();
+    workspace_size_list_.clear();
+  }
+
  protected:
   void InitSizeLists() override {
     input_size_list_.push_back(input_size_);

mindspore/ccsrc/plugin/device/gpu/kernel/nn/bias_add_grad_gpu_kenel.h

@@ -44,7 +44,9 @@ class BiasAddGradGpuKernelMod : public NativeGpuKernelMod {
         cudnn_compute_format_(CUDNN_TENSOR_NCHW),
         dy_desc_(nullptr),
         db_desc_(nullptr),
-        op_desc_(nullptr) {}
+        op_desc_(nullptr) {
+    ResetResource();
+  }
   ~BiasAddGradGpuKernelMod() override { DestroyResource(); }
 
   bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
@@ -139,6 +141,29 @@ class BiasAddGradGpuKernelMod : public NativeGpuKernelMod {
     }
   }
 
+  void ResetResource() noexcept override {
+    same_dims_ = true;
+    is_null_input_ = false;
+    kernel_name_ = "BiasAddGrad";
+    use_cudnn_ = false;
+    dy_num_ = 1;
+    db_num_ = 1;
+    num_dims_ = 0;
+    bias_size_ = 0;
+    dy_shape_.clear();
+    db_shape_.clear();
+    data_format_ = kOpFormat_NCHW;
+    cudnn_handle_ = nullptr;
+    cudnn_data_type_ = CUDNN_DATA_FLOAT;
+    cudnn_compute_format_ = CUDNN_TENSOR_NCHW;
+    dy_desc_ = nullptr;
+    db_desc_ = nullptr;
+    op_desc_ = nullptr;
+    input_size_list_.clear();
+    output_size_list_.clear();
+    workspace_size_list_.clear();
+  }
+
  protected:
   void MethodSelection() {
     // opt implementation can only process num_dims_ <= 4

mindspore/ccsrc/plugin/device/gpu/kernel/nn/ctcloss_gpu_kernel.h

@@ -83,7 +83,9 @@ class CtcLossGpuKernelMod : public NativeGpuKernelMod {
         batch_label(0),
         label_value_with_blank(nullptr),
         log_alpha_b(nullptr),
-        log_beta_b(nullptr) {}
+        log_beta_b(nullptr) {
+    ResetResource();
+  }
   ~CtcLossGpuKernelMod() override = default;
 
   bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
@@ -145,6 +147,44 @@ class CtcLossGpuKernelMod : public NativeGpuKernelMod {
     return true;
   }
 
+  void ResetResource() noexcept override {
+    input_size_list_.clear();
+    output_size_list_.clear();
+    workspace_size_list_.clear();
+    label_indice_size_ = 0;
+    label_size_ = 0;
+    sequence_lengths_size_ = 0;
+    preprocess_collapse_repeated_ = false;
+    ctc_merge_repeated_ = true;
+    ignore_longer_outputs_than_inputs_ = false;
+    is_null_input_ = false;
+    kernel_name_ = "CTCLoss";
+    probs = nullptr;
+    label_indices = nullptr;
+    label_values = nullptr;
+    sequence_length = nullptr;
+    costs = nullptr;
+    grads = nullptr;
+    softmax_probs = nullptr;
+    cum_labels_length = nullptr;
+    label_squence_length = nullptr;
+    label_value_sp = nullptr;
+    label_value_pcr = nullptr;
+    prob_num = nullptr;
+    precum_labels_length = nullptr;
+    max_labels_length = nullptr;
+    numclass = 0;
+    batch = 0;
+    max_time = 0;
+    max_sequence = 0;
+    max_labels_length_host = 0;
+    batch_label = 0;
+    label_value_with_blank = nullptr;
+    log_alpha_b = nullptr;
+    log_beta_b = nullptr;
+    workspace_size_list_.clear();
+  }
+
  protected:
   void LaunchInit(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
                   const std::vector<AddressPtr> &outputs) {

mindspore/ccsrc/plugin/device/gpu/kernel/nn/layer_norm_gpu_kernel.h

@@ -27,7 +27,7 @@ namespace kernel {
 template <typename T>
 class LayerNormGpuKernelMod : public NativeGpuKernelMod {
  public:
-  LayerNormGpuKernelMod() : input_row_(1), input_col_(1), param_dim_(1), is_null_input_(false) {}
+  LayerNormGpuKernelMod() : input_row_(1), input_col_(1), param_dim_(1), is_null_input_(false) { ResetResource(); }
   ~LayerNormGpuKernelMod() override = default;
 
   bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
@@ -87,6 +87,16 @@ class LayerNormGpuKernelMod : public NativeGpuKernelMod {
     return true;
   }
 
+  void ResetResource() noexcept override {
+    input_row_ = 1;
+    input_col_ = 1;
+    param_dim_ = 1;
+    is_null_input_ = false;
+    input_size_list_.clear();
+    output_size_list_.clear();
+    workspace_size_list_.clear();
+  }
+
  protected:
   void InitSizeLists() override {
     input_size_list_.push_back(input_row_ * input_col_ * sizeof(T));

mindspore/ccsrc/plugin/device/gpu/kernel/nn/softmax_gpu_kernel.h

@@ -47,7 +47,9 @@ class SoftmaxGpuKernelMod : public NativeGpuKernelMod {
         batch_size_(0),
         channel_size_(0),
         height_(0),
-        width_(0) {}
+        width_(0) {
+    ResetResource();
+  }
   ~SoftmaxGpuKernelMod() override { DestroyResource(); }
 
   bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspace,
@@ -155,6 +157,33 @@ class SoftmaxGpuKernelMod : public NativeGpuKernelMod {
                                "destroy input_descriptor failed");
   }
 
+  void ResetResource() noexcept override {
+    cudnn_handle_ = nullptr;
+    input_descriptor_ = nullptr;
+    output_descriptor_ = nullptr;
+    algo_ = CUDNN_SOFTMAX_ACCURATE;
+    mode_ = CUDNN_SOFTMAX_MODE_INSTANCE;
+    cudnn_data_type_ = CUDNN_DATA_FLOAT;
+    is_null_input_ = false;
+    kernel_name_ = "Softmax";
+    input_size_ = 0;
+    output_size_ = 0;
+    workspace_size_ = 0;
+    input_size_list_.clear();
+    output_size_list_.clear();
+    workspace_size_list_.clear();
+    input_shape_.clear();
+    transpose_shape_.clear();
+    transpose_axis_.clear();
+    need_transpose_ = false;
+    shape_size_ = 0;
+    batch_size_ = 0;
+    channel_size_ = 0;
+    height_ = 0;
+    width_ = 0;
+    workspace_size_list_.clear();
+  }
+
  protected:
   void InitResource() override {
     cudnn_handle_ = device::gpu::GPUDeviceManager::GetInstance().GetCudnnHandle();