internal_batch_norm_with_update.h

Go to the documentation of this file.

#ifndef INFINI_OPS_BASE_INTERNAL_BATCH_NORM_WITH_UPDATE_H_
#define INFINI_OPS_BASE_INTERNAL_BATCH_NORM_WITH_UPDATE_H_
 
#include <optional>
 
#include "operator.h"
 
namespace infini::ops::internal {
 
class BatchNormWithUpdate : public Operator<BatchNormWithUpdate> {
 public:
  BatchNormWithUpdate(const Tensor input, const std::optional<Tensor> weight,
                      const std::optional<Tensor> bias, Tensor running_mean,
                      Tensor running_var, const double momentum,
                      const double eps, Tensor out, Tensor save_mean,
                      Tensor save_invstd, Tensor reserve)
      : input_shape_{input.shape()},
        input_strides_{input.strides()},
        input_type_{input.dtype()},
        running_mean_shape_{running_mean.shape()},
        running_mean_strides_{running_mean.strides()},
        running_mean_type_{running_mean.dtype()},
        running_var_shape_{running_var.shape()},
        running_var_strides_{running_var.strides()},
        running_var_type_{running_var.dtype()},
        out_shape_{out.shape()},
        out_strides_{out.strides()},
        out_type_{out.dtype()},
        save_mean_shape_{save_mean.shape()},
        save_mean_strides_{save_mean.strides()},
        save_mean_type_{save_mean.dtype()},
        save_invstd_shape_{save_invstd.shape()},
        save_invstd_strides_{save_invstd.strides()},
        save_invstd_type_{save_invstd.dtype()},
        reserve_shape_{reserve.shape()},
        reserve_strides_{reserve.strides()},
        reserve_type_{reserve.dtype()},
        has_weight_{weight.has_value()},
        weight_shape_{weight ? Tensor::Shape{weight->shape()}
                             : Tensor::Shape{}},
        weight_strides_{weight ? Tensor::Strides{weight->strides()}
                               : Tensor::Strides{}},
        weight_type_{weight ? weight->dtype() : DataType::kFloat32},
        has_bias_{bias.has_value()},
        bias_shape_{bias ? Tensor::Shape{bias->shape()} : Tensor::Shape{}},
        bias_strides_{bias ? Tensor::Strides{bias->strides()}
                           : Tensor::Strides{}},
        bias_type_{bias ? bias->dtype() : DataType::kFloat32},
        momentum_{momentum},
        eps_{eps},
        device_index_{out.device().index()} {}
 
  virtual void operator()(const Tensor input,
                          const std::optional<Tensor> weight,
                          const std::optional<Tensor> bias, Tensor running_mean,
                          Tensor running_var, const double momentum,
                          const double eps, Tensor out, Tensor save_mean,
                          Tensor save_invstd, Tensor reserve) const = 0;
 
 protected:
  Tensor::Shape input_shape_;
 
  Tensor::Strides input_strides_;
 
  DataType input_type_;
 
  Tensor::Shape running_mean_shape_;
 
  Tensor::Strides running_mean_strides_;
 
  DataType running_mean_type_;
 
  Tensor::Shape running_var_shape_;
 
  Tensor::Strides running_var_strides_;
 
  DataType running_var_type_;
 
  Tensor::Shape out_shape_;
 
  Tensor::Strides out_strides_;
 
  DataType out_type_;
 
  Tensor::Shape save_mean_shape_;
 
  Tensor::Strides save_mean_strides_;
 
  DataType save_mean_type_;
 
  Tensor::Shape save_invstd_shape_;
 
  Tensor::Strides save_invstd_strides_;
 
  DataType save_invstd_type_;
 
  Tensor::Shape reserve_shape_;
 
  Tensor::Strides reserve_strides_;
 
  DataType reserve_type_;
 
  bool has_weight_{false};
 
  Tensor::Shape weight_shape_;
 
  Tensor::Strides weight_strides_;
 
  DataType weight_type_{DataType::kFloat32};
 
  bool has_bias_{false};
 
  Tensor::Shape bias_shape_;
 
  Tensor::Strides bias_strides_;
 
  DataType bias_type_{DataType::kFloat32};
 
  double momentum_{};
 
  double eps_{};
 
  int device_index_{0};
};
 
}  // namespace infini::ops::internal
 
#endif