datamov_utils.hpp File Reference

#include "common.hpp"

Macros
#define	OPENCV_CUDA_ASM_PTR   "r"
#define	OPENCV_CUDA_DEFINE_FORCE_GLOB(base_type, ptx_type, reg_mod)
#define	OPENCV_CUDA_DEFINE_FORCE_GLOB_B(base_type, ptx_type)

Detailed Description

Deprecated:

Use Device layer instead.

Macro Definition Documentation

#define OPENCV_CUDA_ASM_PTR   "r"

#define OPENCV_CUDA_DEFINE_FORCE_GLOB	(		base_type,
			ptx_type,
			reg_mod
	)

Value:

template <> struct ForceGlob<base_type> \
            { \
                __device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
                { \
                    asm("ld.global."#ptx_type" %0, [%1];" : "="#reg_mod(val) : OPENCV_CUDA_ASM_PTR(ptr + offset)); \
                } \
            };

#define OPENCV_CUDA_DEFINE_FORCE_GLOB_B	(		base_type,
			ptx_type
	)

Value:

template <> struct ForceGlob<base_type> \
            { \
                __device__ __forceinline__ static void Load(const base_type* ptr, int offset, base_type& val) \
                { \
                    asm("ld.global."#ptx_type" %0, [%1];" : "=r"(*reinterpret_cast<uint*>(&val)) : OPENCV_CUDA_ASM_PTR(ptr + offset)); \
                } \
            };