generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

HAL: improvements

Browse Source 
- added new functions from core module: split, merge, add, sub, mul, div, ...
- added function replacement mechanism
- added example of HAL replacement library
This commit is contained in:
Maksim Shabunin
2015-12-03 14:43:37 +03:00
parent e8742be30b
commit b4bcdd10a1
25 changed files with 5552 additions and 4934 deletions
Download Patch File Download Diff File Expand all files Collapse all files

752
modules/core/src/convert.cpp
View File

@@ -42,6 +42,7 @@
//M*/
#include "precomp.hpp"
#include "opencl_kernels_core.hpp"
#ifdef __APPLE__
@@ -49,776 +50,37 @@
#define CV_NEON 0
#endif
namespace cv
{
/****************************************************************************************\
* split & merge *
\****************************************************************************************/
#if CV_NEON
template<typename T> struct VSplit2;
template<typename T> struct VSplit3;
template<typename T> struct VSplit4;
#define SPLIT2_KERNEL_TEMPLATE(name, data_type, reg_type, load_func, store_func) \
template<> \
struct name<data_type> \
{ \
void operator()(const data_type* src, data_type* dst0, \
data_type* dst1) const \
{ \
reg_type r = load_func(src); \
store_func(dst0, r.val[0]); \
store_func(dst1, r.val[1]); \
} \
}
#define SPLIT3_KERNEL_TEMPLATE(name, data_type, reg_type, load_func, store_func) \
template<> \
struct name<data_type> \
{ \
void operator()(const data_type* src, data_type* dst0, data_type* dst1, \
data_type* dst2) const \
{ \
reg_type r = load_func(src); \
store_func(dst0, r.val[0]); \
store_func(dst1, r.val[1]); \
store_func(dst2, r.val[2]); \
} \
}
#define SPLIT4_KERNEL_TEMPLATE(name, data_type, reg_type, load_func, store_func) \
template<> \
struct name<data_type> \
{ \
void operator()(const data_type* src, data_type* dst0, data_type* dst1, \
data_type* dst2, data_type* dst3) const \
{ \
reg_type r = load_func(src); \
store_func(dst0, r.val[0]); \
store_func(dst1, r.val[1]); \
store_func(dst2, r.val[2]); \
store_func(dst3, r.val[3]); \
} \
}
SPLIT2_KERNEL_TEMPLATE(VSplit2, uchar , uint8x16x2_t, vld2q_u8 , vst1q_u8 );
SPLIT2_KERNEL_TEMPLATE(VSplit2, ushort, uint16x8x2_t, vld2q_u16, vst1q_u16);
SPLIT2_KERNEL_TEMPLATE(VSplit2, int , int32x4x2_t, vld2q_s32, vst1q_s32);
SPLIT2_KERNEL_TEMPLATE(VSplit2, int64 , int64x1x2_t, vld2_s64 , vst1_s64 );
SPLIT3_KERNEL_TEMPLATE(VSplit3, uchar , uint8x16x3_t, vld3q_u8 , vst1q_u8 );
SPLIT3_KERNEL_TEMPLATE(VSplit3, ushort, uint16x8x3_t, vld3q_u16, vst1q_u16);
SPLIT3_KERNEL_TEMPLATE(VSplit3, int , int32x4x3_t, vld3q_s32, vst1q_s32);
SPLIT3_KERNEL_TEMPLATE(VSplit3, int64 , int64x1x3_t, vld3_s64 , vst1_s64 );
SPLIT4_KERNEL_TEMPLATE(VSplit4, uchar , uint8x16x4_t, vld4q_u8 , vst1q_u8 );
SPLIT4_KERNEL_TEMPLATE(VSplit4, ushort, uint16x8x4_t, vld4q_u16, vst1q_u16);
SPLIT4_KERNEL_TEMPLATE(VSplit4, int , int32x4x4_t, vld4q_s32, vst1q_s32);
SPLIT4_KERNEL_TEMPLATE(VSplit4, int64 , int64x1x4_t, vld4_s64 , vst1_s64 );
#elif CV_SSE2
template <typename T>
struct VSplit2
{
VSplit2() : support(false) { }
void operator()(const T *, T *, T *) const { }
bool support;
};
template <typename T>
struct VSplit3
{
VSplit3() : support(false) { }
void operator()(const T *, T *, T *, T *) const { }
bool support;
};
template <typename T>
struct VSplit4
{
VSplit4() : support(false) { }
void operator()(const T *, T *, T *, T *, T *) const { }
bool support;
};
#define SPLIT2_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_deinterleave, flavor) \
template <> \
struct VSplit2<data_type> \
{ \
enum \
{ \
ELEMS_IN_VEC = 16 / sizeof(data_type) \
}; \
\
VSplit2() \
{ \
support = checkHardwareSupport(CV_CPU_SSE2); \
} \
\
void operator()(const data_type * src, \
data_type * dst0, data_type * dst1) const \
{ \
reg_type v_src0 = _mm_loadu_##flavor((cast_type const *)(src)); \
reg_type v_src1 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC)); \
reg_type v_src2 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 2)); \
reg_type v_src3 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 3)); \
\
_mm_deinterleave(v_src0, v_src1, v_src2, v_src3); \
\
_mm_storeu_##flavor((cast_type *)(dst0), v_src0); \
_mm_storeu_##flavor((cast_type *)(dst0 + ELEMS_IN_VEC), v_src1); \
_mm_storeu_##flavor((cast_type *)(dst1), v_src2); \
_mm_storeu_##flavor((cast_type *)(dst1 + ELEMS_IN_VEC), v_src3); \
} \
\
bool support; \
}
#define SPLIT3_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_deinterleave, flavor) \
template <> \
struct VSplit3<data_type> \
{ \
enum \
{ \
ELEMS_IN_VEC = 16 / sizeof(data_type) \
}; \
\
VSplit3() \
{ \
support = checkHardwareSupport(CV_CPU_SSE2); \
} \
\
void operator()(const data_type * src, \
data_type * dst0, data_type * dst1, data_type * dst2) const \
{ \
reg_type v_src0 = _mm_loadu_##flavor((cast_type const *)(src)); \
reg_type v_src1 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC)); \
reg_type v_src2 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 2)); \
reg_type v_src3 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 3)); \
reg_type v_src4 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 4)); \
reg_type v_src5 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 5)); \
\
_mm_deinterleave(v_src0, v_src1, v_src2, \
v_src3, v_src4, v_src5); \
\
_mm_storeu_##flavor((cast_type *)(dst0), v_src0); \
_mm_storeu_##flavor((cast_type *)(dst0 + ELEMS_IN_VEC), v_src1); \
_mm_storeu_##flavor((cast_type *)(dst1), v_src2); \
_mm_storeu_##flavor((cast_type *)(dst1 + ELEMS_IN_VEC), v_src3); \
_mm_storeu_##flavor((cast_type *)(dst2), v_src4); \
_mm_storeu_##flavor((cast_type *)(dst2 + ELEMS_IN_VEC), v_src5); \
} \
\
bool support; \
}
#define SPLIT4_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_deinterleave, flavor) \
template <> \
struct VSplit4<data_type> \
{ \
enum \
{ \
ELEMS_IN_VEC = 16 / sizeof(data_type) \
}; \
\
VSplit4() \
{ \
support = checkHardwareSupport(CV_CPU_SSE2); \
} \
\
void operator()(const data_type * src, data_type * dst0, data_type * dst1, \
data_type * dst2, data_type * dst3) const \
{ \
reg_type v_src0 = _mm_loadu_##flavor((cast_type const *)(src)); \
reg_type v_src1 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC)); \
reg_type v_src2 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 2)); \
reg_type v_src3 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 3)); \
reg_type v_src4 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 4)); \
reg_type v_src5 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 5)); \
reg_type v_src6 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 6)); \
reg_type v_src7 = _mm_loadu_##flavor((cast_type const *)(src + ELEMS_IN_VEC * 7)); \
\
_mm_deinterleave(v_src0, v_src1, v_src2, v_src3, \
v_src4, v_src5, v_src6, v_src7); \
\
_mm_storeu_##flavor((cast_type *)(dst0), v_src0); \
_mm_storeu_##flavor((cast_type *)(dst0 + ELEMS_IN_VEC), v_src1); \
_mm_storeu_##flavor((cast_type *)(dst1), v_src2); \
_mm_storeu_##flavor((cast_type *)(dst1 + ELEMS_IN_VEC), v_src3); \
_mm_storeu_##flavor((cast_type *)(dst2), v_src4); \
_mm_storeu_##flavor((cast_type *)(dst2 + ELEMS_IN_VEC), v_src5); \
_mm_storeu_##flavor((cast_type *)(dst3), v_src6); \
_mm_storeu_##flavor((cast_type *)(dst3 + ELEMS_IN_VEC), v_src7); \
} \
\
bool support; \
}
SPLIT2_KERNEL_TEMPLATE( uchar, __m128i, __m128i, _mm_deinterleave_epi8, si128);
SPLIT2_KERNEL_TEMPLATE(ushort, __m128i, __m128i, _mm_deinterleave_epi16, si128);
SPLIT2_KERNEL_TEMPLATE( int, __m128, float, _mm_deinterleave_ps, ps);
SPLIT3_KERNEL_TEMPLATE( uchar, __m128i, __m128i, _mm_deinterleave_epi8, si128);
SPLIT3_KERNEL_TEMPLATE(ushort, __m128i, __m128i, _mm_deinterleave_epi16, si128);
SPLIT3_KERNEL_TEMPLATE( int, __m128, float, _mm_deinterleave_ps, ps);
SPLIT4_KERNEL_TEMPLATE( uchar, __m128i, __m128i, _mm_deinterleave_epi8, si128);
SPLIT4_KERNEL_TEMPLATE(ushort, __m128i, __m128i, _mm_deinterleave_epi16, si128);
SPLIT4_KERNEL_TEMPLATE( int, __m128, float, _mm_deinterleave_ps, ps);
#endif
template<typename T> static void
split_( const T* src, T** dst, int len, int cn )
{
int k = cn % 4 ? cn % 4 : 4;
int i, j;
if( k == 1 )
{
T* dst0 = dst[0];
if(cn == 1)
{
memcpy(dst0, src, len * sizeof(T));
}
else
{
for( i = 0, j = 0 ; i < len; i++, j += cn )
dst0[i] = src[j];
}
}
else if( k == 2 )
{
T *dst0 = dst[0], *dst1 = dst[1];
i = j = 0;
#if CV_NEON
if(cn == 2)
{
int inc_i = (sizeof(T) == 8)? 1: 16/sizeof(T);
int inc_j = 2 * inc_i;
VSplit2<T> vsplit;
for( ; i < len - inc_i; i += inc_i, j += inc_j)
vsplit(src + j, dst0 + i, dst1 + i);
}
#elif CV_SSE2
if (cn == 2)
{
int inc_i = 32/sizeof(T);
int inc_j = 2 * inc_i;
VSplit2<T> vsplit;
if (vsplit.support)
{
for( ; i <= len - inc_i; i += inc_i, j += inc_j)
vsplit(src + j, dst0 + i, dst1 + i);
}
}
#endif
for( ; i < len; i++, j += cn )
{
dst0[i] = src[j];
dst1[i] = src[j+1];
}
}
else if( k == 3 )
{
T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2];
i = j = 0;
#if CV_NEON
if(cn == 3)
{
int inc_i = (sizeof(T) == 8)? 1: 16/sizeof(T);
int inc_j = 3 * inc_i;
VSplit3<T> vsplit;
for( ; i <= len - inc_i; i += inc_i, j += inc_j)
vsplit(src + j, dst0 + i, dst1 + i, dst2 + i);
}
#elif CV_SSE2
if (cn == 3)
{
int inc_i = 32/sizeof(T);
int inc_j = 3 * inc_i;
VSplit3<T> vsplit;
if (vsplit.support)
{
for( ; i <= len - inc_i; i += inc_i, j += inc_j)
vsplit(src + j, dst0 + i, dst1 + i, dst2 + i);
}
}
#endif
for( ; i < len; i++, j += cn )
{
dst0[i] = src[j];
dst1[i] = src[j+1];
dst2[i] = src[j+2];
}
}
else
{
T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2], *dst3 = dst[3];
i = j = 0;
#if CV_NEON
if(cn == 4)
{
int inc_i = (sizeof(T) == 8)? 1: 16/sizeof(T);
int inc_j = 4 * inc_i;
VSplit4<T> vsplit;
for( ; i <= len - inc_i; i += inc_i, j += inc_j)
vsplit(src + j, dst0 + i, dst1 + i, dst2 + i, dst3 + i);
}
#elif CV_SSE2
if (cn == 4)
{
int inc_i = 32/sizeof(T);
int inc_j = 4 * inc_i;
VSplit4<T> vsplit;
if (vsplit.support)
{
for( ; i <= len - inc_i; i += inc_i, j += inc_j)
vsplit(src + j, dst0 + i, dst1 + i, dst2 + i, dst3 + i);
}
}
#endif
for( ; i < len; i++, j += cn )
{
dst0[i] = src[j]; dst1[i] = src[j+1];
dst2[i] = src[j+2]; dst3[i] = src[j+3];
}
}
for( ; k < cn; k += 4 )
{
T *dst0 = dst[k], *dst1 = dst[k+1], *dst2 = dst[k+2], *dst3 = dst[k+3];
for( i = 0, j = k; i < len; i++, j += cn )
{
dst0[i] = src[j]; dst1[i] = src[j+1];
dst2[i] = src[j+2]; dst3[i] = src[j+3];
}
}
}
#if CV_NEON
template<typename T> struct VMerge2;
template<typename T> struct VMerge3;
template<typename T> struct VMerge4;
#define MERGE2_KERNEL_TEMPLATE(name, data_type, reg_type, load_func, store_func) \
template<> \
struct name<data_type>{ \
void operator()(const data_type* src0, const data_type* src1, \
data_type* dst){ \
reg_type r; \
r.val[0] = load_func(src0); \
r.val[1] = load_func(src1); \
store_func(dst, r); \
} \
}
#define MERGE3_KERNEL_TEMPLATE(name, data_type, reg_type, load_func, store_func) \
template<> \
struct name<data_type>{ \
void operator()(const data_type* src0, const data_type* src1, \
const data_type* src2, data_type* dst){ \
reg_type r; \
r.val[0] = load_func(src0); \
r.val[1] = load_func(src1); \
r.val[2] = load_func(src2); \
store_func(dst, r); \
} \
}
#define MERGE4_KERNEL_TEMPLATE(name, data_type, reg_type, load_func, store_func) \
template<> \
struct name<data_type>{ \
void operator()(const data_type* src0, const data_type* src1, \
const data_type* src2, const data_type* src3, \
data_type* dst){ \
reg_type r; \
r.val[0] = load_func(src0); \
r.val[1] = load_func(src1); \
r.val[2] = load_func(src2); \
r.val[3] = load_func(src3); \
store_func(dst, r); \
} \
}
MERGE2_KERNEL_TEMPLATE(VMerge2, uchar , uint8x16x2_t, vld1q_u8 , vst2q_u8 );
MERGE2_KERNEL_TEMPLATE(VMerge2, ushort, uint16x8x2_t, vld1q_u16, vst2q_u16);
MERGE2_KERNEL_TEMPLATE(VMerge2, int , int32x4x2_t, vld1q_s32, vst2q_s32);
MERGE2_KERNEL_TEMPLATE(VMerge2, int64 , int64x1x2_t, vld1_s64 , vst2_s64 );
MERGE3_KERNEL_TEMPLATE(VMerge3, uchar , uint8x16x3_t, vld1q_u8 , vst3q_u8 );
MERGE3_KERNEL_TEMPLATE(VMerge3, ushort, uint16x8x3_t, vld1q_u16, vst3q_u16);
MERGE3_KERNEL_TEMPLATE(VMerge3, int , int32x4x3_t, vld1q_s32, vst3q_s32);
MERGE3_KERNEL_TEMPLATE(VMerge3, int64 , int64x1x3_t, vld1_s64 , vst3_s64 );
MERGE4_KERNEL_TEMPLATE(VMerge4, uchar , uint8x16x4_t, vld1q_u8 , vst4q_u8 );
MERGE4_KERNEL_TEMPLATE(VMerge4, ushort, uint16x8x4_t, vld1q_u16, vst4q_u16);
MERGE4_KERNEL_TEMPLATE(VMerge4, int , int32x4x4_t, vld1q_s32, vst4q_s32);
MERGE4_KERNEL_TEMPLATE(VMerge4, int64 , int64x1x4_t, vld1_s64 , vst4_s64 );
#elif CV_SSE2
template <typename T>
struct VMerge2
{
VMerge2() : support(false) { }
void operator()(const T *, const T *, T *) const { }
bool support;
};
template <typename T>
struct VMerge3
{
VMerge3() : support(false) { }
void operator()(const T *, const T *, const T *, T *) const { }
bool support;
};
template <typename T>
struct VMerge4
{
VMerge4() : support(false) { }
void operator()(const T *, const T *, const T *, const T *, T *) const { }
bool support;
};
#define MERGE2_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_interleave, flavor, se) \
template <> \
struct VMerge2<data_type> \
{ \
enum \
{ \
ELEMS_IN_VEC = 16 / sizeof(data_type) \
}; \
\
VMerge2() \
{ \
support = checkHardwareSupport(se); \
} \
\
void operator()(const data_type * src0, const data_type * src1, \
data_type * dst) const \
{ \
reg_type v_src0 = _mm_loadu_##flavor((const cast_type *)(src0)); \
reg_type v_src1 = _mm_loadu_##flavor((const cast_type *)(src0 + ELEMS_IN_VEC)); \
reg_type v_src2 = _mm_loadu_##flavor((const cast_type *)(src1)); \
reg_type v_src3 = _mm_loadu_##flavor((const cast_type *)(src1 + ELEMS_IN_VEC)); \
\
_mm_interleave(v_src0, v_src1, v_src2, v_src3); \
\
_mm_storeu_##flavor((cast_type *)(dst), v_src0); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC), v_src1); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 2), v_src2); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 3), v_src3); \
} \
\
bool support; \
}
#define MERGE3_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_interleave, flavor, se) \
template <> \
struct VMerge3<data_type> \
{ \
enum \
{ \
ELEMS_IN_VEC = 16 / sizeof(data_type) \
}; \
\
VMerge3() \
{ \
support = checkHardwareSupport(se); \
} \
\
void operator()(const data_type * src0, const data_type * src1, const data_type * src2,\
data_type * dst) const \
{ \
reg_type v_src0 = _mm_loadu_##flavor((const cast_type *)(src0)); \
reg_type v_src1 = _mm_loadu_##flavor((const cast_type *)(src0 + ELEMS_IN_VEC)); \
reg_type v_src2 = _mm_loadu_##flavor((const cast_type *)(src1)); \
reg_type v_src3 = _mm_loadu_##flavor((const cast_type *)(src1 + ELEMS_IN_VEC)); \
reg_type v_src4 = _mm_loadu_##flavor((const cast_type *)(src2)); \
reg_type v_src5 = _mm_loadu_##flavor((const cast_type *)(src2 + ELEMS_IN_VEC)); \
\
_mm_interleave(v_src0, v_src1, v_src2, \
v_src3, v_src4, v_src5); \
\
_mm_storeu_##flavor((cast_type *)(dst), v_src0); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC), v_src1); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 2), v_src2); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 3), v_src3); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 4), v_src4); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 5), v_src5); \
} \
\
bool support; \
}
#define MERGE4_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_interleave, flavor, se) \
template <> \
struct VMerge4<data_type> \
{ \
enum \
{ \
ELEMS_IN_VEC = 16 / sizeof(data_type) \
}; \
\
VMerge4() \
{ \
support = checkHardwareSupport(se); \
} \
\
void operator()(const data_type * src0, const data_type * src1, \
const data_type * src2, const data_type * src3, \
data_type * dst) const \
{ \
reg_type v_src0 = _mm_loadu_##flavor((const cast_type *)(src0)); \
reg_type v_src1 = _mm_loadu_##flavor((const cast_type *)(src0 + ELEMS_IN_VEC)); \
reg_type v_src2 = _mm_loadu_##flavor((const cast_type *)(src1)); \
reg_type v_src3 = _mm_loadu_##flavor((const cast_type *)(src1 + ELEMS_IN_VEC)); \
reg_type v_src4 = _mm_loadu_##flavor((const cast_type *)(src2)); \
reg_type v_src5 = _mm_loadu_##flavor((const cast_type *)(src2 + ELEMS_IN_VEC)); \
reg_type v_src6 = _mm_loadu_##flavor((const cast_type *)(src3)); \
reg_type v_src7 = _mm_loadu_##flavor((const cast_type *)(src3 + ELEMS_IN_VEC)); \
\
_mm_interleave(v_src0, v_src1, v_src2, v_src3, \
v_src4, v_src5, v_src6, v_src7); \
\
_mm_storeu_##flavor((cast_type *)(dst), v_src0); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC), v_src1); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 2), v_src2); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 3), v_src3); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 4), v_src4); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 5), v_src5); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 6), v_src6); \
_mm_storeu_##flavor((cast_type *)(dst + ELEMS_IN_VEC * 7), v_src7); \
} \
\
bool support; \
}
MERGE2_KERNEL_TEMPLATE( uchar, __m128i, __m128i, _mm_interleave_epi8, si128, CV_CPU_SSE2);
MERGE3_KERNEL_TEMPLATE( uchar, __m128i, __m128i, _mm_interleave_epi8, si128, CV_CPU_SSE2);
MERGE4_KERNEL_TEMPLATE( uchar, __m128i, __m128i, _mm_interleave_epi8, si128, CV_CPU_SSE2);
#if CV_SSE4_1
MERGE2_KERNEL_TEMPLATE(ushort, __m128i, __m128i, _mm_interleave_epi16, si128, CV_CPU_SSE4_1);
MERGE3_KERNEL_TEMPLATE(ushort, __m128i, __m128i, _mm_interleave_epi16, si128, CV_CPU_SSE4_1);
MERGE4_KERNEL_TEMPLATE(ushort, __m128i, __m128i, _mm_interleave_epi16, si128, CV_CPU_SSE4_1);
#endif
MERGE2_KERNEL_TEMPLATE( int, __m128, float, _mm_interleave_ps, ps, CV_CPU_SSE2);
MERGE3_KERNEL_TEMPLATE( int, __m128, float, _mm_interleave_ps, ps, CV_CPU_SSE2);
MERGE4_KERNEL_TEMPLATE( int, __m128, float, _mm_interleave_ps, ps, CV_CPU_SSE2);
#endif
template<typename T> static void
merge_( const T** src, T* dst, int len, int cn )
{
int k = cn % 4 ? cn % 4 : 4;
int i, j;
if( k == 1 )
{
const T* src0 = src[0];
for( i = j = 0; i < len; i++, j += cn )
dst[j] = src0[i];
}
else if( k == 2 )
{
const T *src0 = src[0], *src1 = src[1];
i = j = 0;
#if CV_NEON
if(cn == 2)
{
int inc_i = (sizeof(T) == 8)? 1: 16/sizeof(T);
int inc_j = 2 * inc_i;
VMerge2<T> vmerge;
for( ; i < len - inc_i; i += inc_i, j += inc_j)
vmerge(src0 + i, src1 + i, dst + j);
}
#elif CV_SSE2
if(cn == 2)
{
int inc_i = 32/sizeof(T);
int inc_j = 2 * inc_i;
VMerge2<T> vmerge;
if (vmerge.support)
for( ; i < len - inc_i; i += inc_i, j += inc_j)
vmerge(src0 + i, src1 + i, dst + j);
}
#endif
for( ; i < len; i++, j += cn )
{
dst[j] = src0[i];
dst[j+1] = src1[i];
}
}
else if( k == 3 )
{
const T *src0 = src[0], *src1 = src[1], *src2 = src[2];
i = j = 0;
#if CV_NEON
if(cn == 3)
{
int inc_i = (sizeof(T) == 8)? 1: 16/sizeof(T);
int inc_j = 3 * inc_i;
VMerge3<T> vmerge;
for( ; i < len - inc_i; i += inc_i, j += inc_j)
vmerge(src0 + i, src1 + i, src2 + i, dst + j);
}
#elif CV_SSE2
if(cn == 3)
{
int inc_i = 32/sizeof(T);
int inc_j = 3 * inc_i;
VMerge3<T> vmerge;
if (vmerge.support)
for( ; i < len - inc_i; i += inc_i, j += inc_j)
vmerge(src0 + i, src1 + i, src2 + i, dst + j);
}
#endif
for( ; i < len; i++, j += cn )
{
dst[j] = src0[i];
dst[j+1] = src1[i];
dst[j+2] = src2[i];
}
}
else
{
const T *src0 = src[0], *src1 = src[1], *src2 = src[2], *src3 = src[3];
i = j = 0;
#if CV_NEON
if(cn == 4)
{
int inc_i = (sizeof(T) == 8)? 1: 16/sizeof(T);
int inc_j = 4 * inc_i;
VMerge4<T> vmerge;
for( ; i < len - inc_i; i += inc_i, j += inc_j)
vmerge(src0 + i, src1 + i, src2 + i, src3 + i, dst + j);
}
#elif CV_SSE2
if(cn == 4)
{
int inc_i = 32/sizeof(T);
int inc_j = 4 * inc_i;
VMerge4<T> vmerge;
if (vmerge.support)
for( ; i < len - inc_i; i += inc_i, j += inc_j)
vmerge(src0 + i, src1 + i, src2 + i, src3 + i, dst + j);
}
#endif
for( ; i < len; i++, j += cn )
{
dst[j] = src0[i]; dst[j+1] = src1[i];
dst[j+2] = src2[i]; dst[j+3] = src3[i];
}
}
for( ; k < cn; k += 4 )
{
const T *src0 = src[k], *src1 = src[k+1], *src2 = src[k+2], *src3 = src[k+3];
for( i = 0, j = k; i < len; i++, j += cn )
{
dst[j] = src0[i]; dst[j+1] = src1[i];
dst[j+2] = src2[i]; dst[j+3] = src3[i];
}
}
}
static void split8u(const uchar* src, uchar** dst, int len, int cn )
{
split_(src, dst, len, cn);
}
static void split16u(const ushort* src, ushort** dst, int len, int cn )
{
split_(src, dst, len, cn);
}
static void split32s(const int* src, int** dst, int len, int cn )
{
split_(src, dst, len, cn);
}
static void split64s(const int64* src, int64** dst, int len, int cn )
{
split_(src, dst, len, cn);
}
static void merge8u(const uchar** src, uchar* dst, int len, int cn )
{
merge_(src, dst, len, cn);
}
static void merge16u(const ushort** src, ushort* dst, int len, int cn )
{
merge_(src, dst, len, cn);
}
static void merge32s(const int** src, int* dst, int len, int cn )
{
merge_(src, dst, len, cn);
}
static void merge64s(const int64** src, int64* dst, int len, int cn )
{
merge_(src, dst, len, cn);
}
typedef void (*SplitFunc)(const uchar* src, uchar** dst, int len, int cn);
typedef void (*MergeFunc)(const uchar** src, uchar* dst, int len, int cn);
static SplitFunc getSplitFunc(int depth)
{
static SplitFunc splitTab[] =
{
(SplitFunc)GET_OPTIMIZED(split8u), (SplitFunc)GET_OPTIMIZED(split8u), (SplitFunc)GET_OPTIMIZED(split16u), (SplitFunc)GET_OPTIMIZED(split16u),
(SplitFunc)GET_OPTIMIZED(split32s), (SplitFunc)GET_OPTIMIZED(split32s), (SplitFunc)GET_OPTIMIZED(split64s), 0
(SplitFunc)GET_OPTIMIZED(cv::hal::split8u), (SplitFunc)GET_OPTIMIZED(cv::hal::split8u), (SplitFunc)GET_OPTIMIZED(cv::hal::split16u), (SplitFunc)GET_OPTIMIZED(cv::hal::split16u),
(SplitFunc)GET_OPTIMIZED(cv::hal::split32s), (SplitFunc)GET_OPTIMIZED(cv::hal::split32s), (SplitFunc)GET_OPTIMIZED(cv::hal::split64s), 0
};
return splitTab[depth];
}
typedef void (*MergeFunc)(const uchar** src, uchar* dst, int len, int cn);
static MergeFunc getMergeFunc(int depth)
{
static MergeFunc mergeTab[] =
{
(MergeFunc)GET_OPTIMIZED(merge8u), (MergeFunc)GET_OPTIMIZED(merge8u), (MergeFunc)GET_OPTIMIZED(merge16u), (MergeFunc)GET_OPTIMIZED(merge16u),
(MergeFunc)GET_OPTIMIZED(merge32s), (MergeFunc)GET_OPTIMIZED(merge32s), (MergeFunc)GET_OPTIMIZED(merge64s), 0
(MergeFunc)GET_OPTIMIZED(cv::hal::merge8u), (MergeFunc)GET_OPTIMIZED(cv::hal::merge8u), (MergeFunc)GET_OPTIMIZED(cv::hal::merge16u), (MergeFunc)GET_OPTIMIZED(cv::hal::merge16u),
(MergeFunc)GET_OPTIMIZED(cv::hal::merge32s), (MergeFunc)GET_OPTIMIZED(cv::hal::merge32s), (MergeFunc)GET_OPTIMIZED(cv::hal::merge64s), 0
};
return mergeTab[depth];
}
}
void cv::split(const Mat& src, Mat* mv)
{
int k, depth = src.depth(), cn = src.channels();