Perform InnerProduct/FullyConnected/Dense forward calulations, allow fusing bias and activation into same GPU kernel. More...
#include <include/dlprim/core/conv.hpp>
Inheritance diagram for dlprim::core::Conv2DForward:
Public Member Functions | |
| virtual void | enqueue (Tensor &x, Tensor &w, Tensor *bias, Tensor &y, Tensor &ws, float factor, ExecutionContext const &e)=0 |
 Public Member Functions inherited from dlprim::core::Conv2DBase | |
| virtual char const * | algo () const =0 |
| virtual size_t | workspace () |
Static Public Member Functions | |
| static std::unique_ptr< Conv2DForward > | create (Context &ctx, Conv2DSettings const &config, bool bias, StandardActivations activation=StandardActivations::identity, std::string const &algo=std::string()) |
| Create optimal object for conv2d. More... | |
| Static Public Member Functions inherited from dlprim::core::Conv2DBase | |
| static Shape | get_output_shape (Convolution2DConfigBase const &config, Shape const &in) |
| static Shape | get_output_shape_transposed (Convolution2DConfigBase const &config, Shape const &in, int output_pad[2]) |
Detailed Description
Perform InnerProduct/FullyConnected/Dense forward calulations, allow fusing bias and activation into same GPU kernel.
Member Function Documentation
|
static |
Create optimal object for conv2d.
algo is one of "" or "auto" - automatic selection, "gemm" - use fused GEMM based algo "winograd" - use Winograd convoltion - suitable for non strided, non dilated, non grouped 3x3 with pad=1 conv "depthwise_separable"
The documentation for this class was generated from the following file:
- include/dlprim/core/conv.hpp
