MCPcopy Create free account
hub / github.com/DeepRec-AI/DeepRec / ExtendConvFwdParams

Method ExtendConvFwdParams

tensorflow/core/kernels/mkl_conv_ops.cc:831–849  ·  view source on GitHub ↗

This method is for the base class MklConvOp, which handles the floating point implementation of Conv. The quantized conv implementations will use overridden versions of this method.

Source from the content-addressed store, hash-verified

829 // floating point implementation of Conv. The quantized conv implementations
830 // will use overridden versions of this method.
831 virtual void ExtendConvFwdParams(OpKernelContext* context,
832 MklConvFwdParams& params) {
833 // Create a string from data types of input, filter, bias, and output.
834 params.dtypes.append(typeid(Tinput).name());
835 params.dtypes.append(typeid(Tfilter).name());
836 params.dtypes.append(typeid(Tbias).name());
837 params.dtypes.append(typeid(Toutput).name());
838
839 // Add fusions as post ops
840 // NOTE: Fusion of BiasAdd is handled directly inside MklConvOp by
841 // checking `fuse_biasadd_` flag.
842 if (fuse_add_) {
843 params.post_op_params.push_back({"sum", ALGORITHM_UNDEF, {1.0}, ""});
844 }
845 if (fuse_activation_) {
846 params.post_op_params.push_back(
847 {"activation", activation_alg_, {1.0, alpha_or_upbound_, 0.0}, ""});
848 }
849 }
850
851 virtual Tbias* GetBiasHandle(OpKernelContext* context,
852 std::shared_ptr<ConvFwdPd>& conv2d_fwd_pd,

Callers 1

ComputeMethod · 0.95

Calls 3

nameMethod · 0.65
appendMethod · 0.45
push_backMethod · 0.45

Tested by

no test coverage detected