Pointwise.md

Table of Contents

1. Pointwise
2. Reduction

Pointwise

Pointwise performs an elementwise operation between two tensors. The operation used is controlled by pointwise mode cudnn_frontend::PointwiseMode_t.

The API to achieve above is:

std::shared_ptr<Tensor_attributes>
pointwise(std::shared_ptr<Tensor_attributes>,
          Pointwise_attributes);

std::shared_ptr<Tensor_attributes>
pointwise(std::shared_ptr<Tensor_attributes>,
          std::shared_ptr<Tensor_attributes>,
          Pointwise_attributes);

std::shared_ptr<Tensor_attributes>
pointwise(std::shared_ptr<Tensor_attributes>,
          std::shared_ptr<Tensor_attributes>,
          std::shared_ptr<Tensor_attributes>,
          Pointwise_attributes);

where the pointwise mode dictates the API among the choices above. Please refer to documentation of cudnn_frontend::PointwiseMode_t for details.

Pointwise attributes is a lightweight structure with setters:

Pointwise_attributes&
set_mode(PointwiseMode_t)

Pointwise_attributes&
set_axis(int64_t)

Pointwise_attributes&
set_relu_lower_clip(float)

Pointwise_attributes&
set_relu_upper_clip(float)

Pointwise_attributes&
set_relu_lower_clip_slope(float)

Pointwise_attributes&
set_name(std::string const&)

Pointwise_attributes&
set_compute_data_type(DataType_t value)

Python API:

• add
  • a
  • b
  • compute_data_type
  • name
• bias
  • input
  • bias
  • compute_data_type
  • name
• rsqrt
  • input
  • compute_data_type
  • name
• sub
  • a
  • b
  • compute_data_type
  • name
• mul
  • a
  • b
  • compute_data_type
  • name
• scale
  • input
  • scale
  • compute_data_type
  • name
• relu
  • input
  • compute_data_type
  • name
• gelu
  • input
  • compute_data_type
  • name
• elu
  • input
  • compute_data_type
  • name
• cmp_gt
  • input
  • comparison
  • compute_data_type
  • name

Reduction

Reduction operation reduces an input tensor using an operation controlled by cudnn_frontend::ReductionMode_t. The dimensions in input tensors to reduce are deduced using output tensor dimensions.

The API to achieve above is:

std::shared_ptr<Tensor_attributes>
reduction(std::shared_ptr<Tensor_attributes> input, Reduction_attributes);

Reduction attributes is a lightweight structure with setters:

Reduction_attributes&
set_mode(ReductionMode_t)

Reduction_attributes&
set_name(std::string const&)

Reduction_attributes&
set_compute_data_type(DataType_t value)