In the CUDA header file, /usr/local/cuda/targets/x86_64/linux/include/texture_fetch_function.h, there is the following statements:

template<typename T> struct __nv_tex_rmnf_ret{};
template<> struct __nv_tex_rmnf_ret<char> {typedef float type;};

I understand the first statement is the definition of struct template. But what does the latter mean? I have never seen such C syntax before. Could anyony explain it? Thanks in advance.

CodePudding user response：

template<> introduces an explicit specialization of a previously declared (primary) template (here the first line). It has otherwise the same syntax as the primary template uses, except that the name in the declarator (here __nv_tex_rmnf_ret) is replaced by a template-id (here __nv_tex_rmnf_ret<char>) which should be valid for the primary template.

Basically it replaces the definition of the template for the specific specialization determined by the template-id.

__nv_tex_rmnf_ret<char> is now a class containing a typedef for type, but all other specializations __nv_tex_rmnf_ret<T> where T is not char are completely empty classes.

CodePudding user response：

It's called template specialization.

Edit: Removed low-quality link.