Skip to content

O
opencv
Commit 0f2b535f authored by Vitaly Tuzov's avatar Vitaly Tuzov Committed by Vadim Pisarevsky
Browse files
Options

Bit-exact GaussianBlur reworked to use wide intrinsics (#12073) 

* Bit-exact GaussianBlur reworked to use wide intrinsics

* Added v_mul_hi universal intrinsic

* Removed custom SSE2 branch from bit-exact GaussianBlur

* Removed loop unrolling for gaussianBlur horizontal smoothing
parent e345cb03
Show whitespace changes
Inline Side-by-side
Showing with 334 additions and 325 deletions
modules/core/include/opencv2/core/hal/intrin_avx.hpp
View file @ 0f2b535f
...@@ -664,6 +664,8 @@ inline void v_mul_expand(const v_uint32x8& a, const v_uint32x8& b, ...@@ -664,6 +664,8 @@ inline void v_mul_expand(const v_uint32x8& a, const v_uint32x8& b,
v_zip(v_uint64x4(v0), v_uint64x4(v1), c, d); v_zip(v_uint64x4(v0), v_uint64x4(v1), c, d);
} }
inline v_int16x16 v_mul_hi(const v_int16x16& a, const v_int16x16& b) { return v_int16x16(_mm256_mulhi_epi16(a.val, b.val)); }
inline v_uint16x16 v_mul_hi(const v_uint16x16& a, const v_uint16x16& b) { return v_uint16x16(_mm256_mulhi_epu16(a.val, b.val)); }
/** Non-saturating arithmetics **/ /** Non-saturating arithmetics **/
#define OPENCV_HAL_IMPL_AVX_BIN_FUNC(func, _Tpvec, intrin) \ #define OPENCV_HAL_IMPL_AVX_BIN_FUNC(func, _Tpvec, intrin) \
......
modules/core/include/opencv2/core/hal/intrin_cpp.hpp
View file @ 0f2b535f
...@@ -891,6 +891,20 @@ template<typename _Tp, int n> inline void v_mul_expand(const v_reg<_Tp, n>& a, c ...@@ -891,6 +891,20 @@ template<typename _Tp, int n> inline void v_mul_expand(const v_reg<_Tp, n>& a, c
} }
} }
/** @brief Multiply and extract high part
Multiply values two registers and store high part of the results.
Implemented only for 16-bit source types (v_int16x8, v_uint16x8). Returns \f$ a*b >> 16 \f$
*/
template<typename _Tp, int n> inline v_reg<_Tp, n> v_mul_hi(const v_reg<_Tp, n>& a, const v_reg<_Tp, n>& b)
{
typedef typename V_TypeTraits<_Tp>::w_type w_type;
v_reg<_Tp, n> c;
for (int i = 0; i < n; i++)
c.s[i] = (_Tp)(((w_type)a.s[i] * b.s[i]) >> sizeof(_Tp)*8);
return c;
}
//! @cond IGNORED //! @cond IGNORED
template<typename _Tp, int n> inline void v_hsum(const v_reg<_Tp, n>& a, template<typename _Tp, int n> inline void v_hsum(const v_reg<_Tp, n>& a,
v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>& c) v_reg<typename V_TypeTraits<_Tp>::w_type, n/2>& c)
......
modules/core/include/opencv2/core/hal/intrin_neon.hpp
View file @ 0f2b535f
...@@ -553,6 +553,21 @@ inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b, ...@@ -553,6 +553,21 @@ inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b,
d.val = vmull_u32(vget_high_u32(a.val), vget_high_u32(b.val)); d.val = vmull_u32(vget_high_u32(a.val), vget_high_u32(b.val));
} }
inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b)
{
return v_int16x8(vcombine_s16(
vshrn_n_s32(vmull_s16( vget_low_s16(a.val), vget_low_s16(b.val)), 16),
vshrn_n_s32(vmull_s16(vget_high_s16(a.val), vget_high_s16(b.val)), 16)
));
}
inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b)
{
return v_uint16x8(vcombine_u16(
vshrn_n_u32(vmull_u16( vget_low_u16(a.val), vget_low_u16(b.val)), 16),
vshrn_n_u32(vmull_u16(vget_high_u16(a.val), vget_high_u16(b.val)), 16)
));
}
inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b) inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
{ {
int32x4_t c = vmull_s16(vget_low_s16(a.val), vget_low_s16(b.val)); int32x4_t c = vmull_s16(vget_low_s16(a.val), vget_low_s16(b.val));
......
modules/core/include/opencv2/core/hal/intrin_sse.hpp
View file @ 0f2b535f
...@@ -737,6 +737,9 @@ inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b, ...@@ -737,6 +737,9 @@ inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b,
d.val = _mm_unpackhi_epi64(c0, c1); d.val = _mm_unpackhi_epi64(c0, c1);
} }
inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b) { return v_int16x8(_mm_mulhi_epi16(a.val, b.val)); }
inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b) { return v_uint16x8(_mm_mulhi_epu16(a.val, b.val)); }
inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b) inline v_int32x4 v_dotprod(const v_int16x8& a, const v_int16x8& b)
{ {
return v_int32x4(_mm_madd_epi16(a.val, b.val)); return v_int32x4(_mm_madd_epi16(a.val, b.val));
......
modules/core/include/opencv2/core/hal/intrin_vsx.hpp
View file @ 0f2b535f
...@@ -457,6 +457,21 @@ inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b, v_uint64x2& c ...@@ -457,6 +457,21 @@ inline void v_mul_expand(const v_uint32x4& a, const v_uint32x4& b, v_uint64x2& c
d.val = vec_mul(vec_unpacklu(a.val), vec_unpacklu(b.val)); d.val = vec_mul(vec_unpacklu(a.val), vec_unpacklu(b.val));
} }
inline v_int16x8 v_mul_hi(const v_int16x8& a, const v_int16x8& b)
{
return v_int16x8(vec_packs(
vec_sra(vec_mul(vec_unpackh(a.val), vec_unpackh(b.val)), vec_uint4_sp(16)),
vec_sra(vec_mul(vec_unpackl(a.val), vec_unpackl(b.val)), vec_uint4_sp(16))
));
}
inline v_uint16x8 v_mul_hi(const v_uint16x8& a, const v_uint16x8& b)
{
return v_uint16x8(vec_packs(
vec_sr(vec_mul(vec_unpackhu(a.val), vec_unpackhu(b.val)), vec_uint4_sp(16)),
vec_sr(vec_mul(vec_unpacklu(a.val), vec_unpacklu(b.val)), vec_uint4_sp(16))
));
}
/** Non-saturating arithmetics **/ /** Non-saturating arithmetics **/
#define OPENCV_HAL_IMPL_VSX_BIN_FUNC(func, intrin) \ #define OPENCV_HAL_IMPL_VSX_BIN_FUNC(func, intrin) \
template<typename _Tpvec> \ template<typename _Tpvec> \
......
modules/imgproc/src/smooth.cpp
View file @ 0f2b535f
...@@ -1820,22 +1820,13 @@ template <> ...@@ -1820,22 +1820,13 @@ template <>
void hlineSmooth1N<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ufixedpoint16* m, int, ufixedpoint16* dst, int len, int) void hlineSmooth1N<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ufixedpoint16* m, int, ufixedpoint16* dst, int len, int)
{ {
int lencn = len*cn; int lencn = len*cn;
v_uint16x8 v_mul = v_setall_u16(*((uint16_t*)m));
int i = 0; int i = 0;
for (; i <= lencn - 16; i += 16) #if CV_SIMD
{ const int VECSZ = v_uint16::nlanes;
v_uint8x16 v_src = v_load(src + i); v_uint16 v_mul = vx_setall_u16(*((uint16_t*)m));
v_uint16x8 v_tmp0, v_tmp1; for (; i <= lencn - VECSZ; i += VECSZ)
v_expand(v_src, v_tmp0, v_tmp1); v_store((uint16_t*)dst + i, v_mul*vx_load_expand(src + i));
v_store((uint16_t*)dst + i, v_mul*v_tmp0); #endif
v_store((uint16_t*)dst + i + 8, v_mul*v_tmp1);
}
if (i <= lencn - 8)
{
v_uint16x8 v_src = v_load_expand(src + i);
v_store((uint16_t*)dst + i, v_mul*v_src);
i += 8;
}
for (; i < lencn; i++) for (; i < lencn; i++)
dst[i] = m[0] * src[i]; dst[i] = m[0] * src[i];
} }
...@@ -1850,20 +1841,11 @@ void hlineSmooth1N1<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const uf ...@@ -1850,20 +1841,11 @@ void hlineSmooth1N1<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const uf
{ {
int lencn = len*cn; int lencn = len*cn;
int i = 0; int i = 0;
for (; i <= lencn - 16; i += 16) #if CV_SIMD
{ const int VECSZ = v_uint16::nlanes;
v_uint8x16 v_src = v_load(src + i); for (; i <= lencn - VECSZ; i += VECSZ)
v_uint16x8 v_tmp0, v_tmp1; v_store((uint16_t*)dst + i, v_shl<8>(vx_load_expand(src + i)));
v_expand(v_src, v_tmp0, v_tmp1); #endif
v_store((uint16_t*)dst + i, v_shl<8>(v_tmp0));
v_store((uint16_t*)dst + i + 8, v_shl<8>(v_tmp1));
}
if (i <= lencn - 8)
{
v_uint16x8 v_src = v_load_expand(src + i);
v_store((uint16_t*)dst + i, v_shl<8>(v_src));
i += 8;
}
for (; i < lencn; i++) for (; i < lencn; i++)
dst[i] = src[i]; dst[i] = src[i];
} }
...@@ -1926,18 +1908,15 @@ void hlineSmooth3N<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ufi ...@@ -1926,18 +1908,15 @@ void hlineSmooth3N<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ufi
src += cn; dst += cn; src += cn; dst += cn;
int i = cn, lencn = (len - 1)*cn; int i = cn, lencn = (len - 1)*cn;
v_uint16x8 v_mul0 = v_setall_u16(*((uint16_t*)m)); #if CV_SIMD
v_uint16x8 v_mul1 = v_setall_u16(*((uint16_t*)(m + 1))); const uint16_t* _m = (const uint16_t*)m;
v_uint16x8 v_mul2 = v_setall_u16(*((uint16_t*)(m + 2))); const int VECSZ = v_uint16::nlanes;
for (; i <= lencn - 16; i += 16, src += 16, dst += 16) v_uint16 v_mul0 = vx_setall_u16(_m[0]);
{ v_uint16 v_mul1 = vx_setall_u16(_m[1]);
v_uint16x8 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21; v_uint16 v_mul2 = vx_setall_u16(_m[2]);
v_expand(v_load(src - cn), v_src00, v_src01); for (; i <= lencn - VECSZ; i += VECSZ, src += VECSZ, dst += VECSZ)
v_expand(v_load(src), v_src10, v_src11); v_store((uint16_t*)dst, vx_load_expand(src - cn) * v_mul0 + vx_load_expand(src) * v_mul1 + vx_load_expand(src + cn) * v_mul2);
v_expand(v_load(src + cn), v_src20, v_src21); #endif
v_store((uint16_t*)dst, v_src00 * v_mul0 + v_src10 * v_mul1 + v_src20 * v_mul2);
v_store((uint16_t*)dst + 8, v_src01 * v_mul0 + v_src11 * v_mul1 + v_src21 * v_mul2);
}
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
*dst = m[0] * src[-cn] + m[1] * src[0] + m[2] * src[cn]; *dst = m[0] * src[-cn] + m[1] * src[0] + m[2] * src[cn];
...@@ -2017,15 +1996,11 @@ void hlineSmooth3N121<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ...@@ -2017,15 +1996,11 @@ void hlineSmooth3N121<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const
src += cn; dst += cn; src += cn; dst += cn;
int i = cn, lencn = (len - 1)*cn; int i = cn, lencn = (len - 1)*cn;
for (; i <= lencn - 16; i += 16, src += 16, dst += 16) #if CV_SIMD
{ const int VECSZ = v_uint16::nlanes;
v_uint16x8 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21; for (; i <= lencn - VECSZ; i += VECSZ, src += VECSZ, dst += VECSZ)
v_expand(v_load(src - cn), v_src00, v_src01); v_store((uint16_t*)dst, (vx_load_expand(src - cn) + vx_load_expand(src + cn) + (vx_load_expand(src) << 1)) << 6);
v_expand(v_load(src), v_src10, v_src11); #endif
v_expand(v_load(src + cn), v_src20, v_src21);
v_store((uint16_t*)dst, (v_src00 + v_src20 + (v_src10 << 1)) << 6);
v_store((uint16_t*)dst + 8, (v_src01 + v_src21 + (v_src11 << 1)) << 6);
}
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
*((uint16_t*)dst) = (uint16_t(src[-cn]) + uint16_t(src[cn]) + (uint16_t(src[0]) << 1)) << 6; *((uint16_t*)dst) = (uint16_t(src[-cn]) + uint16_t(src[cn]) + (uint16_t(src[0]) << 1)) << 6;
...@@ -2108,17 +2083,14 @@ void hlineSmooth3Naba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ...@@ -2108,17 +2083,14 @@ void hlineSmooth3Naba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const
src += cn; dst += cn; src += cn; dst += cn;
int i = cn, lencn = (len - 1)*cn; int i = cn, lencn = (len - 1)*cn;
v_uint16x8 v_mul0 = v_setall_u16(*((uint16_t*)m)); #if CV_SIMD
v_uint16x8 v_mul1 = v_setall_u16(*((uint16_t*)m+1)); const uint16_t* _m = (const uint16_t*)m;
for (; i <= lencn - 16; i += 16, src += 16, dst += 16) const int VECSZ = v_uint16::nlanes;
{ v_uint16 v_mul0 = vx_setall_u16(_m[0]);
v_uint16x8 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21; v_uint16 v_mul1 = vx_setall_u16(_m[1]);
v_expand(v_load(src - cn), v_src00, v_src01); for (; i <= lencn - VECSZ; i += VECSZ, src += VECSZ, dst += VECSZ)
v_expand(v_load(src), v_src10, v_src11); v_store((uint16_t*)dst, (vx_load_expand(src - cn) + vx_load_expand(src + cn)) * v_mul0 + vx_load_expand(src) * v_mul1);
v_expand(v_load(src + cn), v_src20, v_src21); #endif
v_store((uint16_t*)dst, (v_src00 + v_src20) * v_mul0 + v_src10 * v_mul1);
v_store((uint16_t*)dst + 8, (v_src01 + v_src21) * v_mul0 + v_src11 * v_mul1);
}
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
*((uint16_t*)dst) = ((uint16_t*)m)[1] * src[0] + ((uint16_t*)m)[0] * ((uint16_t)(src[-cn]) + (uint16_t)(src[cn])); *((uint16_t*)dst) = ((uint16_t*)m)[1] * src[0] + ((uint16_t*)m)[0] * ((uint16_t)(src[-cn]) + (uint16_t)(src[cn]));
...@@ -2304,22 +2276,17 @@ void hlineSmooth5N<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ufi ...@@ -2304,22 +2276,17 @@ void hlineSmooth5N<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ufi
src += 2 * cn; dst += 2 * cn; src += 2 * cn; dst += 2 * cn;
int i = 2*cn, lencn = (len - 2)*cn; int i = 2*cn, lencn = (len - 2)*cn;
v_uint16x8 v_mul0 = v_setall_u16(*((uint16_t*)m)); #if CV_SIMD
v_uint16x8 v_mul1 = v_setall_u16(*((uint16_t*)(m + 1))); const uint16_t* _m = (const uint16_t*)m;
v_uint16x8 v_mul2 = v_setall_u16(*((uint16_t*)(m + 2))); const int VECSZ = v_uint16::nlanes;
v_uint16x8 v_mul3 = v_setall_u16(*((uint16_t*)(m + 3))); v_uint16 v_mul0 = vx_setall_u16(_m[0]);
v_uint16x8 v_mul4 = v_setall_u16(*((uint16_t*)(m + 4))); v_uint16 v_mul1 = vx_setall_u16(_m[1]);
for (; i <= lencn - 16; i += 16, src += 16, dst += 16) v_uint16 v_mul2 = vx_setall_u16(_m[2]);
{ v_uint16 v_mul3 = vx_setall_u16(_m[3]);
v_uint16x8 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21, v_src30, v_src31, v_src40, v_src41; v_uint16 v_mul4 = vx_setall_u16(_m[4]);
v_expand(v_load(src - 2*cn), v_src00, v_src01); for (; i <= lencn - VECSZ; i += VECSZ, src += VECSZ, dst += VECSZ)
v_expand(v_load(src - cn), v_src10, v_src11); v_store((uint16_t*)dst, vx_load_expand(src - 2 * cn) * v_mul0 + vx_load_expand(src - cn) * v_mul1 + vx_load_expand(src) * v_mul2 + vx_load_expand(src + cn) * v_mul3 + vx_load_expand(src + 2 * cn) * v_mul4);
v_expand(v_load(src), v_src20, v_src21); #endif
v_expand(v_load(src + cn), v_src30, v_src31);
v_expand(v_load(src + 2*cn), v_src40, v_src41);
v_store((uint16_t*)dst, v_src00 * v_mul0 + v_src10 * v_mul1 + v_src20 * v_mul2 + v_src30 * v_mul3 + v_src40 * v_mul4);
v_store((uint16_t*)dst + 8, v_src01 * v_mul0 + v_src11 * v_mul1 + v_src21 * v_mul2 + v_src31 * v_mul3 + v_src41 * v_mul4);
}
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
*dst = m[0] * src[-2*cn] + m[1] * src[-cn] + m[2] * src[0] + m[3] * src[cn] + m[4] * src[2*cn]; *dst = m[0] * src[-2*cn] + m[1] * src[-cn] + m[2] * src[0] + m[3] * src[cn] + m[4] * src[2*cn];
...@@ -2517,18 +2484,12 @@ void hlineSmooth5N14641<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons ...@@ -2517,18 +2484,12 @@ void hlineSmooth5N14641<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons
src += 2 * cn; dst += 2 * cn; src += 2 * cn; dst += 2 * cn;
int i = 2 * cn, lencn = (len - 2)*cn; int i = 2 * cn, lencn = (len - 2)*cn;
v_uint16x8 v_6 = v_setall_u16(6); #if CV_SIMD
for (; i <= lencn - 16; i += 16, src += 16, dst += 16) const int VECSZ = v_uint16::nlanes;
{ v_uint16 v_6 = vx_setall_u16(6);
v_uint16x8 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21, v_src30, v_src31, v_src40, v_src41; for (; i <= lencn - VECSZ; i += VECSZ, src += VECSZ, dst += VECSZ)
v_expand(v_load(src - 2*cn), v_src00, v_src01); v_store((uint16_t*)dst, (vx_load_expand(src) * v_6 + ((vx_load_expand(src - cn) + vx_load_expand(src + cn)) << 2) + vx_load_expand(src - 2 * cn) + vx_load_expand(src + 2 * cn)) << 4);
v_expand(v_load(src - cn), v_src10, v_src11); #endif
v_expand(v_load(src), v_src20, v_src21);
v_expand(v_load(src + cn), v_src30, v_src31);
v_expand(v_load(src + 2*cn), v_src40, v_src41);
v_store((uint16_t*)dst, (v_src20 * v_6 + ((v_src10 + v_src30) << 2) + v_src00 + v_src40) << 4);
v_store((uint16_t*)dst + 8, (v_src21 * v_6 + ((v_src11 + v_src31) << 2) + v_src01 + v_src41) << 4);
}
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
*((uint16_t*)dst) = (uint16_t(src[0]) * 6 + ((uint16_t(src[-cn]) + uint16_t(src[cn])) << 2) + uint16_t(src[-2 * cn]) + uint16_t(src[2 * cn])) << 4; *((uint16_t*)dst) = (uint16_t(src[0]) * 6 + ((uint16_t(src[-cn]) + uint16_t(src[cn])) << 2) + uint16_t(src[-2 * cn]) + uint16_t(src[2 * cn])) << 4;
...@@ -2721,20 +2682,15 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons ...@@ -2721,20 +2682,15 @@ void hlineSmooth5Nabcba<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, cons
src += 2 * cn; dst += 2 * cn; src += 2 * cn; dst += 2 * cn;
int i = 2 * cn, lencn = (len - 2)*cn; int i = 2 * cn, lencn = (len - 2)*cn;
v_uint16x8 v_mul0 = v_setall_u16(*((uint16_t*)m)); #if CV_SIMD
v_uint16x8 v_mul1 = v_setall_u16(*((uint16_t*)(m + 1))); const uint16_t* _m = (const uint16_t*)m;
v_uint16x8 v_mul2 = v_setall_u16(*((uint16_t*)(m + 2))); const int VECSZ = v_uint16::nlanes;
for (; i <= lencn - 16; i += 16, src += 16, dst += 16) v_uint16 v_mul0 = vx_setall_u16(_m[0]);
{ v_uint16 v_mul1 = vx_setall_u16(_m[1]);
v_uint16x8 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21, v_src30, v_src31, v_src40, v_src41; v_uint16 v_mul2 = vx_setall_u16(_m[2]);
v_expand(v_load(src - 2 * cn), v_src00, v_src01); for (; i <= lencn - VECSZ; i += VECSZ, src += VECSZ, dst += VECSZ)
v_expand(v_load(src - cn), v_src10, v_src11); v_store((uint16_t*)dst, (vx_load_expand(src - 2 * cn) + vx_load_expand(src + 2 * cn)) * v_mul0 + (vx_load_expand(src - cn) + vx_load_expand(src + cn))* v_mul1 + vx_load_expand(src) * v_mul2);
v_expand(v_load(src), v_src20, v_src21); #endif
v_expand(v_load(src + cn), v_src30, v_src31);
v_expand(v_load(src + 2 * cn), v_src40, v_src41);
v_store((uint16_t*)dst, (v_src00 + v_src40) * v_mul0 + (v_src10 + v_src30)* v_mul1 + v_src20 * v_mul2);
v_store((uint16_t*)dst + 8, (v_src01 + v_src41) * v_mul0 + (v_src11 + v_src31) * v_mul1 + v_src21 * v_mul2);
}
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
*((uint16_t*)dst) = ((uint16_t*)m)[0] * ((uint16_t)(src[-2 * cn]) + (uint16_t)(src[2 * cn])) + ((uint16_t*)m)[1] * ((uint16_t)(src[-cn]) + (uint16_t)(src[cn])) + ((uint16_t*)m)[2] * src[0]; *((uint16_t*)dst) = ((uint16_t*)m)[0] * ((uint16_t)(src[-2 * cn]) + (uint16_t)(src[2 * cn])) + ((uint16_t*)m)[1] * ((uint16_t)(src[-cn]) + (uint16_t)(src[cn])) + ((uint16_t*)m)[2] * src[0];
...@@ -2844,23 +2800,16 @@ void hlineSmooth<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ufixe ...@@ -2844,23 +2800,16 @@ void hlineSmooth<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, const ufixe
} }
i *= cn; i *= cn;
int lencn = (len - post_shift + 1)*cn; int lencn = (len - post_shift + 1)*cn;
for (; i <= lencn - 16; i+=16, src+=16, dst+=16) #if CV_SIMD
const int VECSZ = v_uint16::nlanes;
for (; i <= lencn - VECSZ; i+=VECSZ, src+=VECSZ, dst+=VECSZ)
{ {
v_uint16x8 v_src0, v_src1; v_uint16 v_res0 = vx_load_expand(src) * vx_setall_u16(*((uint16_t*)m));
v_uint16x8 v_mul = v_setall_u16(*((uint16_t*)m));
v_expand(v_load(src), v_src0, v_src1);
v_uint16x8 v_res0 = v_src0 * v_mul;
v_uint16x8 v_res1 = v_src1 * v_mul;
for (int j = 1; j < n; j++) for (int j = 1; j < n; j++)
{ v_res0 += vx_load_expand(src + j * cn) * vx_setall_u16(*((uint16_t*)(m + j)));
v_mul = v_setall_u16(*((uint16_t*)(m + j)));
v_expand(v_load(src + j * cn), v_src0, v_src1);
v_res0 += v_src0 * v_mul;
v_res1 += v_src1 * v_mul;
}
v_store((uint16_t*)dst, v_res0); v_store((uint16_t*)dst, v_res0);
v_store((uint16_t*)dst+8, v_res1);
} }
#endif
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
{ {
*dst = m[0] * src[0]; *dst = m[0] * src[0];
...@@ -2970,26 +2919,16 @@ void hlineSmoothONa_yzy_a<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, co ...@@ -2970,26 +2919,16 @@ void hlineSmoothONa_yzy_a<uint8_t, ufixedpoint16>(const uint8_t* src, int cn, co
} }
i *= cn; i *= cn;
int lencn = (len - post_shift + 1)*cn; int lencn = (len - post_shift + 1)*cn;
for (; i <= lencn - 16; i += 16, src += 16, dst += 16) #if CV_SIMD
const int VECSZ = v_uint16::nlanes;
for (; i <= lencn - VECSZ; i += VECSZ, src += VECSZ, dst += VECSZ)
{ {
v_uint16x8 v_src00, v_src01, v_srcN00, v_srcN01; v_uint16 v_res0 = vx_load_expand(src + pre_shift * cn) * vx_setall_u16(*((uint16_t*)(m + pre_shift)));
v_uint16x8 v_mul = v_setall_u16(*((uint16_t*)(m + pre_shift)));
v_expand(v_load(src + pre_shift * cn), v_src00, v_src01);
v_uint16x8 v_res0 = v_src00 * v_mul;
v_uint16x8 v_res1 = v_src01 * v_mul;
for (int j = 0; j < pre_shift; j ++) for (int j = 0; j < pre_shift; j ++)
{ v_res0 += (vx_load_expand(src + j * cn) + vx_load_expand(src + (n - 1 - j)*cn)) * vx_setall_u16(*((uint16_t*)(m + j)));
v_mul = v_setall_u16(*((uint16_t*)(m + j)));
v_expand(v_load(src + j * cn), v_src00, v_src01);
v_expand(v_load(src + (n - 1 - j)*cn), v_srcN00, v_srcN01);
v_res0 += (v_src00 + v_srcN00) * v_mul;
v_res1 += (v_src01 + v_srcN01) * v_mul;
}
v_store((uint16_t*)dst, v_res0); v_store((uint16_t*)dst, v_res0);
v_store((uint16_t*)dst + 8, v_res1);
} }
#endif
for (; i < lencn; i++, src++, dst++) for (; i < lencn; i++, src++, dst++)
{ {
*dst = m[pre_shift] * src[pre_shift*cn]; *dst = m[pre_shift] * src[pre_shift*cn];
...@@ -3025,28 +2964,13 @@ template <> ...@@ -3025,28 +2964,13 @@ template <>
void vlineSmooth1N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len) void vlineSmooth1N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len)
{ {
const ufixedpoint16* src0 = src[0]; const ufixedpoint16* src0 = src[0];
v_uint16x8 v_mul = v_setall_u16(*((uint16_t*)m));
#if CV_SSE2
v_uint16x8 v_1 = v_setall_u16(1);
v_mul += v_mul;
#endif
int i = 0; int i = 0;
for (; i <= len - 16; i += 16) #if CV_SIMD
{ const int VECSZ = v_uint16::nlanes;
v_uint16x8 v_src0 = v_load((uint16_t*)src0 + i); v_uint16 v_mul = vx_setall_u16(*((uint16_t*)m)<<1);
v_uint16x8 v_src1 = v_load((uint16_t*)src0 + i + 8); for (; i <= len - VECSZ; i += VECSZ)
v_uint8x16 v_res; v_rshr_pack_store<1>(dst + i, v_mul_hi(vx_load((uint16_t*)src0 + i), v_mul));
#if CV_SSE2
v_res.val = _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(v_1.val, _mm_mulhi_epu16(v_src0.val, v_mul.val)),1),
_mm_srli_epi16(_mm_add_epi16(v_1.val, _mm_mulhi_epu16(v_src1.val, v_mul.val)),1));
#else
v_uint32x4 v_res0, v_res1, v_res2, v_res3;
v_mul_expand(v_src0, v_mul, v_res0, v_res1);
v_mul_expand(v_src1, v_mul, v_res2, v_res3);
v_res = v_pack(v_rshr_pack<16>(v_res0, v_res1), v_rshr_pack<16>(v_res2, v_res3));
#endif #endif
v_store(dst + i, v_res);
}
for (; i < len; i++) for (; i < len; i++)
dst[i] = m[0] * src0[i]; dst[i] = m[0] * src0[i];
} }
...@@ -3062,8 +2986,11 @@ void vlineSmooth1N1<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, co ...@@ -3062,8 +2986,11 @@ void vlineSmooth1N1<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, co
{ {
const ufixedpoint16* src0 = src[0]; const ufixedpoint16* src0 = src[0];
int i = 0; int i = 0;
for (; i <= len - 8; i += 8) #if CV_SIMD
v_rshr_pack_store<8>(dst + i, v_load((uint16_t*)(src0 + i))); const int VECSZ = v_uint16::nlanes;
for (; i <= len - VECSZ; i += VECSZ)
v_rshr_pack_store<8>(dst + i, vx_load((uint16_t*)(src0 + i)));
#endif
for (; i < len; i++) for (; i < len; i++)
dst[i] = src0[i]; dst[i] = src0[i];
} }
...@@ -3077,46 +3004,51 @@ template <> ...@@ -3077,46 +3004,51 @@ template <>
void vlineSmooth3N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len) void vlineSmooth3N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len)
{ {
int i = 0; int i = 0;
static const v_int16x8 v_128 = v_reinterpret_as_s16(v_setall_u16((uint16_t)1 << 15)); #if CV_SIMD
v_int32x4 v_128_4 = v_setall_s32(128 << 16); static const v_int16 v_128 = v_reinterpret_as_s16(vx_setall_u16((uint16_t)1 << 15));
if (len > 7) v_int32 v_128_4 = vx_setall_s32(128 << 16);
const int VECSZ = v_uint16::nlanes;
if (len >= VECSZ)
{ {
ufixedpoint32 val[] = { (m[0] + m[1] + m[2]) * ufixedpoint16((uint8_t)128) }; ufixedpoint32 val[] = { (m[0] + m[1] + m[2]) * ufixedpoint16((uint8_t)128) };
v_128_4 = v_setall_s32(*((int32_t*)val)); v_128_4 = vx_setall_s32(*((int32_t*)val));
} }
v_int16x8 v_mul01 = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)m))); v_int16 v_mul01 = v_reinterpret_as_s16(vx_setall_u32(*((uint32_t*)m)));
v_int16x8 v_mul2 = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + 2)))); v_int16 v_mul2 = v_reinterpret_as_s16(vx_setall_u16(*((uint16_t*)(m + 2))));
for (; i <= len - 32; i += 32) for (; i <= len - 4*VECSZ; i += 4*VECSZ)
{ {
v_int16x8 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13; v_int16 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13;
v_int16x8 v_tmp0, v_tmp1; v_int16 v_tmp0, v_tmp1;
v_src00 = v_load((int16_t*)(src[0]) + i); const int16_t* src0 = (const int16_t*)src[0] + i;
v_src01 = v_load((int16_t*)(src[0]) + i + 8); const int16_t* src1 = (const int16_t*)src[1] + i;
v_src02 = v_load((int16_t*)(src[0]) + i + 16); v_src00 = vx_load(src0);
v_src03 = v_load((int16_t*)(src[0]) + i + 24); v_src01 = vx_load(src0 + VECSZ);
v_src10 = v_load((int16_t*)(src[1]) + i); v_src02 = vx_load(src0 + 2*VECSZ);
v_src11 = v_load((int16_t*)(src[1]) + i + 8); v_src03 = vx_load(src0 + 3*VECSZ);
v_src12 = v_load((int16_t*)(src[1]) + i + 16); v_src10 = vx_load(src1);
v_src13 = v_load((int16_t*)(src[1]) + i + 24); v_src11 = vx_load(src1 + VECSZ);
v_src12 = vx_load(src1 + 2*VECSZ);
v_src13 = vx_load(src1 + 3*VECSZ);
v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul01); v_int32 v_res0 = v_dotprod(v_tmp0, v_mul01);
v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul01); v_int32 v_res1 = v_dotprod(v_tmp1, v_mul01);
v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res2 = v_dotprod(v_tmp0, v_mul01); v_int32 v_res2 = v_dotprod(v_tmp0, v_mul01);
v_int32x4 v_res3 = v_dotprod(v_tmp1, v_mul01); v_int32 v_res3 = v_dotprod(v_tmp1, v_mul01);
v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res4 = v_dotprod(v_tmp0, v_mul01); v_int32 v_res4 = v_dotprod(v_tmp0, v_mul01);
v_int32x4 v_res5 = v_dotprod(v_tmp1, v_mul01); v_int32 v_res5 = v_dotprod(v_tmp1, v_mul01);
v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res6 = v_dotprod(v_tmp0, v_mul01); v_int32 v_res6 = v_dotprod(v_tmp0, v_mul01);
v_int32x4 v_res7 = v_dotprod(v_tmp1, v_mul01); v_int32 v_res7 = v_dotprod(v_tmp1, v_mul01);
v_int32x4 v_resj0, v_resj1; v_int32 v_resj0, v_resj1;
v_src00 = v_load((int16_t*)(src[2]) + i); const int16_t* src2 = (const int16_t*)src[2] + i;
v_src01 = v_load((int16_t*)(src[2]) + i + 8); v_src00 = vx_load(src2);
v_src02 = v_load((int16_t*)(src[2]) + i + 16); v_src01 = vx_load(src2 + VECSZ);
v_src03 = v_load((int16_t*)(src[2]) + i + 24); v_src02 = vx_load(src2 + 2*VECSZ);
v_src03 = vx_load(src2 + 3*VECSZ);
v_mul_expand(v_add_wrap(v_src00, v_128), v_mul2, v_resj0, v_resj1); v_mul_expand(v_add_wrap(v_src00, v_128), v_mul2, v_resj0, v_resj1);
v_res0 += v_resj0; v_res0 += v_resj0;
v_res1 += v_resj1; v_res1 += v_resj1;
...@@ -3141,9 +3073,10 @@ void vlineSmooth3N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, con ...@@ -3141,9 +3073,10 @@ void vlineSmooth3N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, con
v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)), v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)),
v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3)))); v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3))));
v_store(dst + i + 16, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)), v_store(dst + i + 2*VECSZ, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)),
v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7)))); v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7))));
} }
#endif
for (; i < len; i++) for (; i < len; i++)
dst[i] = m[0] * src[0][i] + m[1] * src[1][i] + m[2] * src[2][i]; dst[i] = m[0] * src[0][i] + m[1] * src[1][i] + m[2] * src[2][i];
} }
...@@ -3157,18 +3090,21 @@ template <> ...@@ -3157,18 +3090,21 @@ template <>
void vlineSmooth3N121<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16*, int, uint8_t* dst, int len) void vlineSmooth3N121<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16*, int, uint8_t* dst, int len)
{ {
int i = 0; int i = 0;
for (; i <= len - 16; i += 16) #if CV_SIMD
{ const int VECSZ = v_uint16::nlanes;
v_uint32x4 v_src00, v_src01, v_src02, v_src03, v_src10, v_src11, v_src12, v_src13, v_src20, v_src21, v_src22, v_src23; for (; i <= len - 2*VECSZ; i += 2*VECSZ)
v_expand(v_load((uint16_t*)(src[0]) + i), v_src00, v_src01); {
v_expand(v_load((uint16_t*)(src[0]) + i + 8), v_src02, v_src03); v_uint32 v_src00, v_src01, v_src02, v_src03, v_src10, v_src11, v_src12, v_src13, v_src20, v_src21, v_src22, v_src23;
v_expand(v_load((uint16_t*)(src[1]) + i), v_src10, v_src11); v_expand(vx_load((uint16_t*)(src[0]) + i), v_src00, v_src01);
v_expand(v_load((uint16_t*)(src[1]) + i + 8), v_src12, v_src13); v_expand(vx_load((uint16_t*)(src[0]) + i + VECSZ), v_src02, v_src03);
v_expand(v_load((uint16_t*)(src[2]) + i), v_src20, v_src21); v_expand(vx_load((uint16_t*)(src[1]) + i), v_src10, v_src11);
v_expand(v_load((uint16_t*)(src[2]) + i + 8), v_src22, v_src23); v_expand(vx_load((uint16_t*)(src[1]) + i + VECSZ), v_src12, v_src13);
v_expand(vx_load((uint16_t*)(src[2]) + i), v_src20, v_src21);
v_expand(vx_load((uint16_t*)(src[2]) + i + VECSZ), v_src22, v_src23);
v_store(dst + i, v_pack(v_rshr_pack<10>(v_src00 + v_src20 + (v_src10 + v_src10), v_src01 + v_src21 + (v_src11 + v_src11)), v_store(dst + i, v_pack(v_rshr_pack<10>(v_src00 + v_src20 + (v_src10 + v_src10), v_src01 + v_src21 + (v_src11 + v_src11)),
v_rshr_pack<10>(v_src02 + v_src22 + (v_src12 + v_src12), v_src03 + v_src23 + (v_src13 + v_src13)))); v_rshr_pack<10>(v_src02 + v_src22 + (v_src12 + v_src12), v_src03 + v_src23 + (v_src13 + v_src13))));
} }
#endif
for (; i < len; i++) for (; i < len; i++)
dst[i] = (((uint32_t)(((uint16_t*)(src[0]))[i]) + (uint32_t)(((uint16_t*)(src[2]))[i]) + ((uint32_t)(((uint16_t*)(src[1]))[i]) << 1)) + (1 << 9)) >> 10; dst[i] = (((uint32_t)(((uint16_t*)(src[0]))[i]) + (uint32_t)(((uint16_t*)(src[2]))[i]) + ((uint32_t)(((uint16_t*)(src[1]))[i]) << 1)) + (1 << 9)) >> 10;
} }
...@@ -3182,50 +3118,54 @@ template <> ...@@ -3182,50 +3118,54 @@ template <>
void vlineSmooth5N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len) void vlineSmooth5N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len)
{ {
int i = 0; int i = 0;
static const v_int16x8 v_128 = v_reinterpret_as_s16(v_setall_u16((uint16_t)1 << 15)); #if CV_SIMD
v_int32x4 v_128_4 = v_setall_s32(128 << 16); const int VECSZ = v_uint16::nlanes;
if (len > 7) if (len >= 4 * VECSZ)
{ {
ufixedpoint32 val[] = { (m[0] + m[1] + m[2] + m[3] + m[4]) * ufixedpoint16((uint8_t)128) }; ufixedpoint32 val[] = { (m[0] + m[1] + m[2] + m[3] + m[4]) * ufixedpoint16((uint8_t)128) };
v_128_4 = v_setall_s32(*((int32_t*)val)); v_int32 v_128_4 = vx_setall_s32(*((int32_t*)val));
} static const v_int16 v_128 = v_reinterpret_as_s16(vx_setall_u16((uint16_t)1 << 15));
v_int16x8 v_mul01 = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)m))); v_int16 v_mul01 = v_reinterpret_as_s16(vx_setall_u32(*((uint32_t*)m)));
v_int16x8 v_mul23 = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)(m + 2)))); v_int16 v_mul23 = v_reinterpret_as_s16(vx_setall_u32(*((uint32_t*)(m + 2))));
v_int16x8 v_mul4 = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + 4)))); v_int16 v_mul4 = v_reinterpret_as_s16(vx_setall_u16(*((uint16_t*)(m + 4))));
for (; i <= len - 32; i += 32) for (; i <= len - 4*VECSZ; i += 4*VECSZ)
{ {
v_int16x8 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13; v_int16 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13;
v_int16x8 v_tmp0, v_tmp1; v_int16 v_tmp0, v_tmp1;
v_src00 = v_load((int16_t*)(src[0]) + i); const int16_t* src0 = (const int16_t*)src[0] + i;
v_src01 = v_load((int16_t*)(src[0]) + i + 8); const int16_t* src1 = (const int16_t*)src[1] + i;
v_src02 = v_load((int16_t*)(src[0]) + i + 16); v_src00 = vx_load(src0);
v_src03 = v_load((int16_t*)(src[0]) + i + 24); v_src01 = vx_load(src0 + VECSZ);
v_src10 = v_load((int16_t*)(src[1]) + i); v_src02 = vx_load(src0 + 2*VECSZ);
v_src11 = v_load((int16_t*)(src[1]) + i + 8); v_src03 = vx_load(src0 + 3*VECSZ);
v_src12 = v_load((int16_t*)(src[1]) + i + 16); v_src10 = vx_load(src1);
v_src13 = v_load((int16_t*)(src[1]) + i + 24); v_src11 = vx_load(src1 + VECSZ);
v_src12 = vx_load(src1 + 2*VECSZ);
v_src13 = vx_load(src1 + 3*VECSZ);
v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul01); v_int32 v_res0 = v_dotprod(v_tmp0, v_mul01);
v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul01); v_int32 v_res1 = v_dotprod(v_tmp1, v_mul01);
v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res2 = v_dotprod(v_tmp0, v_mul01); v_int32 v_res2 = v_dotprod(v_tmp0, v_mul01);
v_int32x4 v_res3 = v_dotprod(v_tmp1, v_mul01); v_int32 v_res3 = v_dotprod(v_tmp1, v_mul01);
v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res4 = v_dotprod(v_tmp0, v_mul01); v_int32 v_res4 = v_dotprod(v_tmp0, v_mul01);
v_int32x4 v_res5 = v_dotprod(v_tmp1, v_mul01); v_int32 v_res5 = v_dotprod(v_tmp1, v_mul01);
v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res6 = v_dotprod(v_tmp0, v_mul01); v_int32 v_res6 = v_dotprod(v_tmp0, v_mul01);
v_int32x4 v_res7 = v_dotprod(v_tmp1, v_mul01); v_int32 v_res7 = v_dotprod(v_tmp1, v_mul01);
v_src00 = v_load((int16_t*)(src[2]) + i); const int16_t* src2 = (const int16_t*)src[2] + i;
v_src01 = v_load((int16_t*)(src[2]) + i + 8); const int16_t* src3 = (const int16_t*)src[3] + i;
v_src02 = v_load((int16_t*)(src[2]) + i + 16); v_src00 = vx_load(src2);
v_src03 = v_load((int16_t*)(src[2]) + i + 24); v_src01 = vx_load(src2 + VECSZ);
v_src10 = v_load((int16_t*)(src[3]) + i); v_src02 = vx_load(src2 + 2*VECSZ);
v_src11 = v_load((int16_t*)(src[3]) + i + 8); v_src03 = vx_load(src2 + 3*VECSZ);
v_src12 = v_load((int16_t*)(src[3]) + i + 16); v_src10 = vx_load(src3);
v_src13 = v_load((int16_t*)(src[3]) + i + 24); v_src11 = vx_load(src3 + VECSZ);
v_src12 = vx_load(src3 + 2*VECSZ);
v_src13 = vx_load(src3 + 3*VECSZ);
v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1);
v_res0 += v_dotprod(v_tmp0, v_mul23); v_res0 += v_dotprod(v_tmp0, v_mul23);
v_res1 += v_dotprod(v_tmp1, v_mul23); v_res1 += v_dotprod(v_tmp1, v_mul23);
...@@ -3239,11 +3179,12 @@ void vlineSmooth5N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, con ...@@ -3239,11 +3179,12 @@ void vlineSmooth5N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, con
v_res6 += v_dotprod(v_tmp0, v_mul23); v_res6 += v_dotprod(v_tmp0, v_mul23);
v_res7 += v_dotprod(v_tmp1, v_mul23); v_res7 += v_dotprod(v_tmp1, v_mul23);
v_int32x4 v_resj0, v_resj1; v_int32 v_resj0, v_resj1;
v_src00 = v_load((int16_t*)(src[4]) + i); const int16_t* src4 = (const int16_t*)src[4] + i;
v_src01 = v_load((int16_t*)(src[4]) + i + 8); v_src00 = vx_load(src4);
v_src02 = v_load((int16_t*)(src[4]) + i + 16); v_src01 = vx_load(src4 + VECSZ);
v_src03 = v_load((int16_t*)(src[4]) + i + 24); v_src02 = vx_load(src4 + 2*VECSZ);
v_src03 = vx_load(src4 + 3*VECSZ);
v_mul_expand(v_add_wrap(v_src00, v_128), v_mul4, v_resj0, v_resj1); v_mul_expand(v_add_wrap(v_src00, v_128), v_mul4, v_resj0, v_resj1);
v_res0 += v_resj0; v_res0 += v_resj0;
v_res1 += v_resj1; v_res1 += v_resj1;
...@@ -3268,9 +3209,11 @@ void vlineSmooth5N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, con ...@@ -3268,9 +3209,11 @@ void vlineSmooth5N<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, con
v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)), v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)),
v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3)))); v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3))));
v_store(dst + i + 16, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)), v_store(dst + i + 2*VECSZ, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)),
v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7)))); v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7))));
} }
}
#endif
for (; i < len; i++) for (; i < len; i++)
dst[i] = m[0] * src[0][i] + m[1] * src[1][i] + m[2] * src[2][i] + m[3] * src[3][i] + m[4] * src[4][i]; dst[i] = m[0] * src[0][i] + m[1] * src[1][i] + m[2] * src[2][i] + m[3] * src[3][i] + m[4] * src[4][i];
} }
...@@ -3284,28 +3227,31 @@ template <> ...@@ -3284,28 +3227,31 @@ template <>
void vlineSmooth5N14641<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16*, int, uint8_t* dst, int len) void vlineSmooth5N14641<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16*, int, uint8_t* dst, int len)
{ {
int i = 0; int i = 0;
v_uint32x4 v_6 = v_setall_u32(6); #if CV_SIMD
for (; i <= len - 16; i += 16) v_uint32 v_6 = vx_setall_u32(6);
{ const int VECSZ = v_uint16::nlanes;
v_uint32x4 v_src00, v_src10, v_src20, v_src30, v_src40; for (; i <= len - 2*VECSZ; i += 2*VECSZ)
v_uint32x4 v_src01, v_src11, v_src21, v_src31, v_src41; {
v_uint32x4 v_src02, v_src12, v_src22, v_src32, v_src42; v_uint32 v_src00, v_src10, v_src20, v_src30, v_src40;
v_uint32x4 v_src03, v_src13, v_src23, v_src33, v_src43; v_uint32 v_src01, v_src11, v_src21, v_src31, v_src41;
v_expand(v_load((uint16_t*)(src[0]) + i), v_src00, v_src01); v_uint32 v_src02, v_src12, v_src22, v_src32, v_src42;
v_expand(v_load((uint16_t*)(src[0]) + i + 8), v_src02, v_src03); v_uint32 v_src03, v_src13, v_src23, v_src33, v_src43;
v_expand(v_load((uint16_t*)(src[1]) + i), v_src10, v_src11); v_expand(vx_load((uint16_t*)(src[0]) + i), v_src00, v_src01);
v_expand(v_load((uint16_t*)(src[1]) + i + 8), v_src12, v_src13); v_expand(vx_load((uint16_t*)(src[0]) + i + VECSZ), v_src02, v_src03);
v_expand(v_load((uint16_t*)(src[2]) + i), v_src20, v_src21); v_expand(vx_load((uint16_t*)(src[1]) + i), v_src10, v_src11);
v_expand(v_load((uint16_t*)(src[2]) + i + 8), v_src22, v_src23); v_expand(vx_load((uint16_t*)(src[1]) + i + VECSZ), v_src12, v_src13);
v_expand(v_load((uint16_t*)(src[3]) + i), v_src30, v_src31); v_expand(vx_load((uint16_t*)(src[2]) + i), v_src20, v_src21);
v_expand(v_load((uint16_t*)(src[3]) + i + 8), v_src32, v_src33); v_expand(vx_load((uint16_t*)(src[2]) + i + VECSZ), v_src22, v_src23);
v_expand(v_load((uint16_t*)(src[4]) + i), v_src40, v_src41); v_expand(vx_load((uint16_t*)(src[3]) + i), v_src30, v_src31);
v_expand(v_load((uint16_t*)(src[4]) + i + 8), v_src42, v_src43); v_expand(vx_load((uint16_t*)(src[3]) + i + VECSZ), v_src32, v_src33);
v_expand(vx_load((uint16_t*)(src[4]) + i), v_src40, v_src41);
v_expand(vx_load((uint16_t*)(src[4]) + i + VECSZ), v_src42, v_src43);
v_store(dst + i, v_pack(v_rshr_pack<12>(v_src20*v_6 + ((v_src10 + v_src30) << 2) + v_src00 + v_src40, v_store(dst + i, v_pack(v_rshr_pack<12>(v_src20*v_6 + ((v_src10 + v_src30) << 2) + v_src00 + v_src40,
v_src21*v_6 + ((v_src11 + v_src31) << 2) + v_src01 + v_src41), v_src21*v_6 + ((v_src11 + v_src31) << 2) + v_src01 + v_src41),
v_rshr_pack<12>(v_src22*v_6 + ((v_src12 + v_src32) << 2) + v_src02 + v_src42, v_rshr_pack<12>(v_src22*v_6 + ((v_src12 + v_src32) << 2) + v_src02 + v_src42,
v_src23*v_6 + ((v_src13 + v_src33) << 2) + v_src03 + v_src43))); v_src23*v_6 + ((v_src13 + v_src33) << 2) + v_src03 + v_src43)));
} }
#endif
for (; i < len; i++) for (; i < len; i++)
dst[i] = ((uint32_t)(((uint16_t*)(src[2]))[i]) * 6 + dst[i] = ((uint32_t)(((uint16_t*)(src[2]))[i]) * 6 +
(((uint32_t)(((uint16_t*)(src[1]))[i]) + (uint32_t)(((uint16_t*)(src[3]))[i])) << 2) + (((uint32_t)(((uint16_t*)(src[1]))[i]) + (uint32_t)(((uint16_t*)(src[3]))[i])) << 2) +
...@@ -3326,57 +3272,63 @@ template <> ...@@ -3326,57 +3272,63 @@ template <>
void vlineSmooth<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int n, uint8_t* dst, int len) void vlineSmooth<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int n, uint8_t* dst, int len)
{ {
int i = 0; int i = 0;
static const v_int16x8 v_128 = v_reinterpret_as_s16(v_setall_u16((uint16_t)1 << 15)); #if CV_SIMD
v_int32x4 v_128_4 = v_setall_s32(128 << 16); static const v_int16 v_128 = v_reinterpret_as_s16(vx_setall_u16((uint16_t)1 << 15));
if (len > 7) v_int32 v_128_4 = vx_setall_s32(128 << 16);
const int VECSZ = v_uint16::nlanes;
if (len >= VECSZ)
{ {
ufixedpoint16 msum = m[0] + m[1]; ufixedpoint16 msum = m[0] + m[1];
for (int j = 2; j < n; j++) for (int j = 2; j < n; j++)
msum = msum + m[j]; msum = msum + m[j];
ufixedpoint32 val[] = { msum * ufixedpoint16((uint8_t)128) }; ufixedpoint32 val[] = { msum * ufixedpoint16((uint8_t)128) };
v_128_4 = v_setall_s32(*((int32_t*)val)); v_128_4 = vx_setall_s32(*((int32_t*)val));
} }
for (; i <= len - 32; i += 32) for (; i <= len - 4*VECSZ; i += 4*VECSZ)
{ {
v_int16x8 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13; v_int16 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13;
v_int16x8 v_tmp0, v_tmp1; v_int16 v_tmp0, v_tmp1;
v_int16x8 v_mul = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)m))); v_int16 v_mul = v_reinterpret_as_s16(vx_setall_u32(*((uint32_t*)m)));
v_src00 = v_load((int16_t*)(src[0]) + i); const int16_t* src0 = (const int16_t*)src[0] + i;
v_src01 = v_load((int16_t*)(src[0]) + i + 8); const int16_t* src1 = (const int16_t*)src[1] + i;
v_src02 = v_load((int16_t*)(src[0]) + i + 16); v_src00 = vx_load(src0);
v_src03 = v_load((int16_t*)(src[0]) + i + 24); v_src01 = vx_load(src0 + VECSZ);
v_src10 = v_load((int16_t*)(src[1]) + i); v_src02 = vx_load(src0 + 2*VECSZ);
v_src11 = v_load((int16_t*)(src[1]) + i + 8); v_src03 = vx_load(src0 + 3*VECSZ);
v_src12 = v_load((int16_t*)(src[1]) + i + 16); v_src10 = vx_load(src1);
v_src13 = v_load((int16_t*)(src[1]) + i + 24); v_src11 = vx_load(src1 + VECSZ);
v_src12 = vx_load(src1 + 2*VECSZ);
v_src13 = vx_load(src1 + 3*VECSZ);
v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul); v_int32 v_res0 = v_dotprod(v_tmp0, v_mul);
v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul); v_int32 v_res1 = v_dotprod(v_tmp1, v_mul);
v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res2 = v_dotprod(v_tmp0, v_mul); v_int32 v_res2 = v_dotprod(v_tmp0, v_mul);
v_int32x4 v_res3 = v_dotprod(v_tmp1, v_mul); v_int32 v_res3 = v_dotprod(v_tmp1, v_mul);
v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res4 = v_dotprod(v_tmp0, v_mul); v_int32 v_res4 = v_dotprod(v_tmp0, v_mul);
v_int32x4 v_res5 = v_dotprod(v_tmp1, v_mul); v_int32 v_res5 = v_dotprod(v_tmp1, v_mul);
v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1);
v_int32x4 v_res6 = v_dotprod(v_tmp0, v_mul); v_int32 v_res6 = v_dotprod(v_tmp0, v_mul);
v_int32x4 v_res7 = v_dotprod(v_tmp1, v_mul); v_int32 v_res7 = v_dotprod(v_tmp1, v_mul);
int j = 2; int j = 2;
for (; j < n - 1; j+=2) for (; j < n - 1; j+=2)
{ {
v_mul = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)(m+j)))); v_mul = v_reinterpret_as_s16(vx_setall_u32(*((uint32_t*)(m+j))));
v_src00 = v_load((int16_t*)(src[j]) + i); const int16_t* srcj0 = (const int16_t*)src[j] + i;
v_src01 = v_load((int16_t*)(src[j]) + i + 8); const int16_t* srcj1 = (const int16_t*)src[j + 1] + i;
v_src02 = v_load((int16_t*)(src[j]) + i + 16); v_src00 = vx_load(srcj0);
v_src03 = v_load((int16_t*)(src[j]) + i + 24); v_src01 = vx_load(srcj0 + VECSZ);
v_src10 = v_load((int16_t*)(src[j+1]) + i); v_src02 = vx_load(srcj0 + 2*VECSZ);
v_src11 = v_load((int16_t*)(src[j+1]) + i + 8); v_src03 = vx_load(srcj0 + 3*VECSZ);
v_src12 = v_load((int16_t*)(src[j+1]) + i + 16); v_src10 = vx_load(srcj1);
v_src13 = v_load((int16_t*)(src[j+1]) + i + 24); v_src11 = vx_load(srcj1 + VECSZ);
v_src12 = vx_load(srcj1 + 2*VECSZ);
v_src13 = vx_load(srcj1 + 3*VECSZ);
v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1);
v_res0 += v_dotprod(v_tmp0, v_mul); v_res0 += v_dotprod(v_tmp0, v_mul);
v_res1 += v_dotprod(v_tmp1, v_mul); v_res1 += v_dotprod(v_tmp1, v_mul);
...@@ -3392,12 +3344,13 @@ void vlineSmooth<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ...@@ -3392,12 +3344,13 @@ void vlineSmooth<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const
} }
if(j < n) if(j < n)
{ {
v_int32x4 v_resj0, v_resj1; v_int32 v_resj0, v_resj1;
v_mul = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + j)))); v_mul = v_reinterpret_as_s16(vx_setall_u16(*((uint16_t*)(m + j))));
v_src00 = v_load((int16_t*)(src[j]) + i); const int16_t* srcj = (const int16_t*)src[j] + i;
v_src01 = v_load((int16_t*)(src[j]) + i + 8); v_src00 = vx_load(srcj);
v_src02 = v_load((int16_t*)(src[j]) + i + 16); v_src01 = vx_load(srcj + VECSZ);
v_src03 = v_load((int16_t*)(src[j]) + i + 24); v_src02 = vx_load(srcj + 2*VECSZ);
v_src03 = vx_load(srcj + 3*VECSZ);
v_mul_expand(v_add_wrap(v_src00, v_128), v_mul, v_resj0, v_resj1); v_mul_expand(v_add_wrap(v_src00, v_128), v_mul, v_resj0, v_resj1);
v_res0 += v_resj0; v_res0 += v_resj0;
v_res1 += v_resj1; v_res1 += v_resj1;
...@@ -3422,9 +3375,10 @@ void vlineSmooth<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ...@@ -3422,9 +3375,10 @@ void vlineSmooth<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const
v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)), v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)),
v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3)))); v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3))));
v_store(dst + i + 16, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)), v_store(dst + i + 2*VECSZ, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)),
v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7)))); v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7))));
} }
#endif
for (; i < len; i++) for (; i < len; i++)
{ {
ufixedpoint32 val = m[0] * src[0][i]; ufixedpoint32 val = m[0] * src[0][i];
...@@ -3450,29 +3404,32 @@ void vlineSmoothONa_yzy_a(const FT* const * src, const FT* m, int n, ET* dst, in ...@@ -3450,29 +3404,32 @@ void vlineSmoothONa_yzy_a(const FT* const * src, const FT* m, int n, ET* dst, in
template <> template <>
void vlineSmoothONa_yzy_a<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int n, uint8_t* dst, int len) void vlineSmoothONa_yzy_a<uint8_t, ufixedpoint16>(const ufixedpoint16* const * src, const ufixedpoint16* m, int n, uint8_t* dst, int len)
{ {
int pre_shift = n / 2;
int i = 0; int i = 0;
static const v_int16x8 v_128 = v_reinterpret_as_s16(v_setall_u16((uint16_t)1 << 15)); #if CV_SIMD
v_int32x4 v_128_4 = v_setall_s32(128 << 16); int pre_shift = n / 2;
if (len > 7) static const v_int16 v_128 = v_reinterpret_as_s16(vx_setall_u16((uint16_t)1 << 15));
v_int32 v_128_4 = vx_setall_s32(128 << 16);
const int VECSZ = v_uint16::nlanes;
if (len >= VECSZ)
{ {
ufixedpoint16 msum = m[0] + m[pre_shift] + m[n - 1]; ufixedpoint16 msum = m[0] + m[pre_shift] + m[n - 1];
for (int j = 1; j < pre_shift; j++) for (int j = 1; j < pre_shift; j++)
msum = msum + m[j] + m[n - 1 - j]; msum = msum + m[j] + m[n - 1 - j];
ufixedpoint32 val[] = { msum * ufixedpoint16((uint8_t)128) }; ufixedpoint32 val[] = { msum * ufixedpoint16((uint8_t)128) };
v_128_4 = v_setall_s32(*((int32_t*)val)); v_128_4 = vx_setall_s32(*((int32_t*)val));
} }
for (; i <= len - 32; i += 32) for (; i <= len - 4*VECSZ; i += 4*VECSZ)
{ {
v_int16x8 v_src00, v_src10, v_src20, v_src30, v_src01, v_src11, v_src21, v_src31; v_int16 v_src00, v_src10, v_src20, v_src30, v_src01, v_src11, v_src21, v_src31;
v_int32x4 v_res0, v_res1, v_res2, v_res3, v_res4, v_res5, v_res6, v_res7; v_int32 v_res0, v_res1, v_res2, v_res3, v_res4, v_res5, v_res6, v_res7;
v_int16x8 v_tmp0, v_tmp1, v_tmp2, v_tmp3, v_tmp4, v_tmp5, v_tmp6, v_tmp7; v_int16 v_tmp0, v_tmp1, v_tmp2, v_tmp3, v_tmp4, v_tmp5, v_tmp6, v_tmp7;
v_int16x8 v_mul = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + pre_shift)))); v_int16 v_mul = v_reinterpret_as_s16(vx_setall_u16(*((uint16_t*)(m + pre_shift))));
v_src00 = v_load((int16_t*)(src[pre_shift]) + i); const int16_t* srcp = (const int16_t*)src[pre_shift] + i;
v_src10 = v_load((int16_t*)(src[pre_shift]) + i + 8); v_src00 = vx_load(srcp);
v_src20 = v_load((int16_t*)(src[pre_shift]) + i + 16); v_src10 = vx_load(srcp + VECSZ);
v_src30 = v_load((int16_t*)(src[pre_shift]) + i + 24); v_src20 = vx_load(srcp + 2*VECSZ);
v_src30 = vx_load(srcp + 3*VECSZ);
v_mul_expand(v_add_wrap(v_src00, v_128), v_mul, v_res0, v_res1); v_mul_expand(v_add_wrap(v_src00, v_128), v_mul, v_res0, v_res1);
v_mul_expand(v_add_wrap(v_src10, v_128), v_mul, v_res2, v_res3); v_mul_expand(v_add_wrap(v_src10, v_128), v_mul, v_res2, v_res3);
v_mul_expand(v_add_wrap(v_src20, v_128), v_mul, v_res4, v_res5); v_mul_expand(v_add_wrap(v_src20, v_128), v_mul, v_res4, v_res5);
...@@ -3481,16 +3438,18 @@ void vlineSmoothONa_yzy_a<uint8_t, ufixedpoint16>(const ufixedpoint16* const * s ...@@ -3481,16 +3438,18 @@ void vlineSmoothONa_yzy_a<uint8_t, ufixedpoint16>(const ufixedpoint16* const * s
int j = 0; int j = 0;
for (; j < pre_shift; j++) for (; j < pre_shift; j++)
{ {
v_mul = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + j)))); v_mul = v_reinterpret_as_s16(vx_setall_u16(*((uint16_t*)(m + j))));
v_src00 = v_load((int16_t*)(src[j]) + i); const int16_t* srcj0 = (const int16_t*)src[j] + i;
v_src10 = v_load((int16_t*)(src[j]) + i + 8); const int16_t* srcj1 = (const int16_t*)src[n - 1 - j] + i;
v_src20 = v_load((int16_t*)(src[j]) + i + 16); v_src00 = vx_load(srcj0);
v_src30 = v_load((int16_t*)(src[j]) + i + 24); v_src10 = vx_load(srcj0 + VECSZ);
v_src01 = v_load((int16_t*)(src[n - 1 - j]) + i); v_src20 = vx_load(srcj0 + 2*VECSZ);
v_src11 = v_load((int16_t*)(src[n - 1 - j]) + i + 8); v_src30 = vx_load(srcj0 + 3*VECSZ);
v_src21 = v_load((int16_t*)(src[n - 1 - j]) + i + 16); v_src01 = vx_load(srcj1);
v_src31 = v_load((int16_t*)(src[n - 1 - j]) + i + 24); v_src11 = vx_load(srcj1 + VECSZ);
v_src21 = vx_load(srcj1 + 2*VECSZ);
v_src31 = vx_load(srcj1 + 3*VECSZ);
v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src01, v_128), v_tmp0, v_tmp1); v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src01, v_128), v_tmp0, v_tmp1);
v_res0 += v_dotprod(v_tmp0, v_mul); v_res0 += v_dotprod(v_tmp0, v_mul);
v_res1 += v_dotprod(v_tmp1, v_mul); v_res1 += v_dotprod(v_tmp1, v_mul);
...@@ -3516,9 +3475,10 @@ void vlineSmoothONa_yzy_a<uint8_t, ufixedpoint16>(const ufixedpoint16* const * s ...@@ -3516,9 +3475,10 @@ void vlineSmoothONa_yzy_a<uint8_t, ufixedpoint16>(const ufixedpoint16* const * s
v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)), v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)),
v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3)))); v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3))));
v_store(dst + i + 16, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)), v_store(dst + i + 2*VECSZ, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)),
v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7)))); v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7))));
} }
#endif
for (; i < len; i++) for (; i < len; i++)
{ {
ufixedpoint32 val = m[0] * src[0][i]; ufixedpoint32 val = m[0] * src[0][i];
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment