Chiebot-Mirror
/
opencv
8
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

checks

Browse Source
pull/3591/head
Ilya Lavrenov 10 years ago
parent 1d3c860411
commit 6bce6ee34a
7 changed files with 872 additions and 663 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 89
      modules/core/src/arithm.cpp
  2. 62
      modules/core/src/convert.cpp
  3. 102
      modules/core/src/mathfuncs.cpp
  4. 6
      modules/core/src/stat.cpp
  5. 595
      modules/imgproc/src/color.cpp
  6. 671
      modules/imgproc/src/imgwarp.cpp
  7. 10
      modules/imgproc/src/pyramids.cpp

89
modules/core/src/arithm.cpp
Unescape View File

@ -64,7 +64,7 @@ FUNCTOR_TEMPLATE(VLoadStore128);
#if CV_SSE2
FUNCTOR_TEMPLATE(VLoadStore64);
FUNCTOR_TEMPLATE(VLoadStore128Aligned);
#if CV_AVX
#if CV_AVX2
FUNCTOR_TEMPLATE(VLoadStore256);
FUNCTOR_TEMPLATE(VLoadStore256Aligned);
#endif
@ -2626,10 +2626,16 @@ struct Div_SIMD
template <>
struct Div_SIMD<uchar>
{
bool haveSIMD;
Div_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE4_1); }
int operator() (const uchar * src1, const uchar * src2, uchar * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -2672,10 +2678,16 @@ struct Div_SIMD<uchar>
template <>
struct Div_SIMD<schar>
{
bool haveSIMD;
Div_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE2); }
int operator() (const schar * src1, const schar * src2, schar * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -2718,10 +2730,16 @@ struct Div_SIMD<schar>
template <>
struct Div_SIMD<ushort>
{
bool haveSIMD;
Div_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE4_1); }
int operator() (const ushort * src1, const ushort * src2, ushort * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -2763,10 +2781,16 @@ struct Div_SIMD<ushort>
template <>
struct Div_SIMD<short>
{
bool haveSIMD;
Div_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE2); }
int operator() (const short * src1, const short * src2, short * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -2806,10 +2830,16 @@ struct Div_SIMD<short>
template <>
struct Div_SIMD<int>
{
bool haveSIMD;
Div_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE2); }
int operator() (const int * src1, const int * src2, int * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -2902,10 +2932,16 @@ struct Recip_SIMD
template <>
struct Recip_SIMD<uchar>
{
bool haveSIMD;
Recip_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE4_1); }
int operator() (const uchar * src2, uchar * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -2941,10 +2977,16 @@ struct Recip_SIMD<uchar>
template <>
struct Recip_SIMD<schar>
{
bool haveSIMD;
Recip_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE2); }
int operator() (const schar * src2, schar * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -2980,10 +3022,16 @@ struct Recip_SIMD<schar>
template <>
struct Recip_SIMD<ushort>
{
bool haveSIMD;
Recip_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE4_1); }
int operator() (const ushort * src2, ushort * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -3018,10 +3066,16 @@ struct Recip_SIMD<ushort>
template <>
struct Recip_SIMD<short>
{
bool haveSIMD;
Recip_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE2); }
int operator() (const short * src2, short * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -3054,10 +3108,16 @@ struct Recip_SIMD<short>
template <>
struct Recip_SIMD<int>
{
bool haveSIMD;
Recip_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE2); }
int operator() (const int * src2, int * dst, int width, double scale) const
{
int x = 0;
if (!haveSIMD)
return x;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
@ -4126,7 +4186,8 @@ static void cmp8u(const uchar* src1, size_t step1, const uchar* src2, size_t ste
{
int x =0;
#if CV_SSE2
if( USE_SSE2 ){
if( USE_SSE2 )
{
__m128i m128 = code == CMP_GT ? _mm_setzero_si128() : _mm_set1_epi8 (-1);
__m128i c128 = _mm_set1_epi8 (-128);
for( ; x <= size.width - 16; x += 16 )
@ -4142,7 +4203,7 @@ static void cmp8u(const uchar* src1, size_t step1, const uchar* src2, size_t ste
}
}
#elif CV_NEON
#elif CV_NEON
uint8x16_t mask = code == CMP_GT ? vdupq_n_u8(0) : vdupq_n_u8(255);
for( ; x <= size.width - 16; x += 16 )
@ -4164,7 +4225,8 @@ static void cmp8u(const uchar* src1, size_t step1, const uchar* src2, size_t ste
{
int x = 0;
#if CV_SSE2
if( USE_SSE2 ){
if( USE_SSE2 )
{
__m128i m128 = code == CMP_EQ ? _mm_setzero_si128() : _mm_set1_epi8 (-1);
for( ; x <= size.width - 16; x += 16 )
{
@ -4174,7 +4236,7 @@ static void cmp8u(const uchar* src1, size_t step1, const uchar* src2, size_t ste
_mm_storeu_si128((__m128i*)(dst + x), r00);
}
}
#elif CV_NEON
#elif CV_NEON
uint8x16_t mask = code == CMP_EQ ? vdupq_n_u8(0) : vdupq_n_u8(255);
for( ; x <= size.width - 16; x += 16 )
@ -4254,7 +4316,8 @@ static void cmp16s(const short* src1, size_t step1, const short* src2, size_t st
{
int x =0;
#if CV_SSE2
if( USE_SSE2){//
if( USE_SSE2)
{
__m128i m128 = code == CMP_GT ? _mm_setzero_si128() : _mm_set1_epi16 (-1);
for( ; x <= size.width - 16; x += 16 )
{
@ -4278,7 +4341,7 @@ static void cmp16s(const short* src1, size_t step1, const short* src2, size_t st
x += 8;
}
}
#elif CV_NEON
#elif CV_NEON
uint8x16_t mask = code == CMP_GT ? vdupq_n_u8(0) : vdupq_n_u8(255);
for( ; x <= size.width - 16; x += 16 )
@ -4293,8 +4356,7 @@ static void cmp16s(const short* src1, size_t step1, const short* src2, size_t st
vst1q_u8(dst+x, veorq_u8(vcombine_u8(t1, t2), mask));
}
#endif
#endif
for( ; x < size.width; x++ ){
dst[x] = (uchar)(-(src1[x] > src2[x]) ^ m);
@ -4308,7 +4370,8 @@ static void cmp16s(const short* src1, size_t step1, const short* src2, size_t st
{
int x = 0;
#if CV_SSE2
if( USE_SSE2 ){
if( USE_SSE2 )
{
__m128i m128 = code == CMP_EQ ? _mm_setzero_si128() : _mm_set1_epi16 (-1);
for( ; x <= size.width - 16; x += 16 )
{
@ -4332,7 +4395,7 @@ static void cmp16s(const short* src1, size_t step1, const short* src2, size_t st
x += 8;
}
}
#elif CV_NEON
#elif CV_NEON
uint8x16_t mask = code == CMP_EQ ? vdupq_n_u8(0) : vdupq_n_u8(255);
for( ; x <= size.width - 16; x += 16 )
@ -4347,8 +4410,8 @@ static void cmp16s(const short* src1, size_t step1, const short* src2, size_t st
vst1q_u8(dst+x, veorq_u8(vcombine_u8(t1, t2), mask));
}
#endif
for( ; x < size.width; x++ )
#endif
for( ; x < size.width; x++ )
dst[x] = (uchar)(-(src1[x] == src2[x]) ^ m);
}
}

62
modules/core/src/convert.cpp
Unescape View File

@ -158,7 +158,7 @@ struct VSplit2<data_type>
\
VSplit2() \
{ \
support = true; \
support = checkHardwareSupport(CV_CPU_SSE2); \
} \
\
void operator()(const data_type * src, \
@ -191,7 +191,7 @@ struct VSplit3<data_type>
\
VSplit3() \
{ \
support = true; \
support = checkHardwareSupport(CV_CPU_SSE2); \
} \
\
void operator()(const data_type * src, \
@ -229,7 +229,7 @@ struct VSplit4<data_type>
\
VSplit4() \
{ \
support = true; \
support = checkHardwareSupport(CV_CPU_SSE2); \
} \
\
void operator()(const data_type * src, data_type * dst0, data_type * dst1, \
@ -502,7 +502,7 @@ struct VMerge4
bool support;
};
#define MERGE2_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_interleave, flavor) \
#define MERGE2_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_interleave, flavor, se) \
template <> \
struct VMerge2<data_type> \
{ \
@ -513,7 +513,7 @@ struct VMerge2<data_type>
\
VMerge2() \
{ \
support = true; \
support = checkHardwareSupport(se); \
} \
\
void operator()(const data_type * src0, const data_type * src1, \
@ -535,7 +535,7 @@ struct VMerge2<data_type>
bool support; \
}
#define MERGE3_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_interleave, flavor) \
#define MERGE3_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_interleave, flavor, se) \
template <> \
struct VMerge3<data_type> \
{ \
@ -546,7 +546,7 @@ struct VMerge3<data_type>
\
VMerge3() \
{ \
support = true; \
support = checkHardwareSupport(se); \
} \
\
void operator()(const data_type * src0, const data_type * src1, const data_type * src2,\
@ -573,7 +573,7 @@ struct VMerge3<data_type>
bool support; \
}
#define MERGE4_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_interleave, flavor) \
#define MERGE4_KERNEL_TEMPLATE(data_type, reg_type, cast_type, _mm_interleave, flavor, se) \
template <> \
struct VMerge4<data_type> \
{ \
@ -584,7 +584,7 @@ struct VMerge4<data_type>
\
VMerge4() \
{ \
support = true; \
support = checkHardwareSupport(se); \
} \
\
void operator()(const data_type * src0, const data_type * src1, \
@ -616,19 +616,19 @@ struct VMerge4<data_type>
bool support; \
}
MERGE2_KERNEL_TEMPLATE( uchar, __m128i, __m128i, _mm_interleave_epi8, si128);
MERGE3_KERNEL_TEMPLATE( uchar, __m128i, __m128i, _mm_interleave_epi8, si128);
MERGE4_KERNEL_TEMPLATE( uchar, __m128i, __m128i, _mm_interleave_epi8, si128);
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);
MERGE3_KERNEL_TEMPLATE(ushort, __m128i, __m128i, _mm_interleave_epi16, si128);
MERGE4_KERNEL_TEMPLATE(ushort, __m128i, __m128i, _mm_interleave_epi16, si128);
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);
MERGE3_KERNEL_TEMPLATE( int, __m128, float, _mm_interleave_ps, ps);
MERGE4_KERNEL_TEMPLATE( int, __m128, float, _mm_interleave_ps, ps);
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
@ -4404,6 +4404,9 @@ struct Cvt_SIMD<double, uchar>
{
int x = 0;
if (!USE_SSE2)
return x;
for ( ; x <= width - 8; x += 8)
{
__m128 v_src0 = _mm_cvtpd_ps(_mm_loadu_pd(src + x));
@ -4430,6 +4433,9 @@ struct Cvt_SIMD<double, schar>
{
int x = 0;
if (!USE_SSE2)
return x;
for ( ; x <= width - 8; x += 8)
{
__m128 v_src0 = _mm_cvtpd_ps(_mm_loadu_pd(src + x));
@ -4454,10 +4460,16 @@ struct Cvt_SIMD<double, schar>
template <>
struct Cvt_SIMD<double, ushort>
{
bool haveSIMD;
Cvt_SIMD() { haveSIMD = checkHardwareSupport(CV_CPU_SSE4_1); }
int operator() (const double * src, ushort * dst, int width) const
{
int x = 0;
if (!haveSIMD)
return x;
for ( ; x <= width - 8; x += 8)
{
__m128 v_src0 = _mm_cvtpd_ps(_mm_loadu_pd(src + x));
@ -4486,6 +4498,9 @@ struct Cvt_SIMD<double, short>
{
int x = 0;
if (!USE_SSE2)
return x;
for ( ; x <= width - 8; x += 8)
{
__m128 v_src0 = _mm_cvtpd_ps(_mm_loadu_pd(src + x));
@ -4512,6 +4527,9 @@ struct Cvt_SIMD<double, int>
{
int x = 0;
if (!USE_SSE2)
return x;
for ( ; x <= width - 4; x += 4)
{
__m128 v_src0 = _mm_cvtpd_ps(_mm_loadu_pd(src + x));
@ -4532,6 +4550,9 @@ struct Cvt_SIMD<double, float>
{
int x = 0;
if (!USE_SSE2)
return x;
for ( ; x <= width - 4; x += 4)
{
__m128 v_src0 = _mm_cvtpd_ps(_mm_loadu_pd(src + x));
@ -5114,8 +5135,9 @@ cvt_<float, short>( const float* src, size_t sstep,
{
int x = 0;
#if CV_SSE2
if(USE_SSE2){
for( ; x <= size.width - 8; x += 8 )
if(USE_SSE2)
{
for( ; x <= size.width - 8; x += 8 )
{
__m128 src128 = _mm_loadu_ps (src + x);
__m128i src_int128 = _mm_cvtps_epi32 (src128);

102
modules/core/src/mathfuncs.cpp
Unescape View File

@ -597,15 +597,18 @@ void phase( InputArray src1, InputArray src2, OutputArray dst, bool angleInDegre
k = 0;
#if CV_SSE2
for ( ; k <= len - 4; k += 4)
if (USE_SSE2)
{
__m128 v_dst0 = _mm_movelh_ps(_mm_cvtpd_ps(_mm_loadu_pd(x + k)),
_mm_cvtpd_ps(_mm_loadu_pd(x + k + 2)));
__m128 v_dst1 = _mm_movelh_ps(_mm_cvtpd_ps(_mm_loadu_pd(y + k)),
_mm_cvtpd_ps(_mm_loadu_pd(y + k + 2)));
for ( ; k <= len - 4; k += 4)
{
__m128 v_dst0 = _mm_movelh_ps(_mm_cvtpd_ps(_mm_loadu_pd(x + k)),
_mm_cvtpd_ps(_mm_loadu_pd(x + k + 2)));
__m128 v_dst1 = _mm_movelh_ps(_mm_cvtpd_ps(_mm_loadu_pd(y + k)),
_mm_cvtpd_ps(_mm_loadu_pd(y + k + 2)));
_mm_storeu_ps(buf[0] + k, v_dst0);
_mm_storeu_ps(buf[1] + k, v_dst1);
_mm_storeu_ps(buf[0] + k, v_dst0);
_mm_storeu_ps(buf[1] + k, v_dst1);
}
}
#endif
@ -619,11 +622,14 @@ void phase( InputArray src1, InputArray src2, OutputArray dst, bool angleInDegre
k = 0;
#if CV_SSE2
for ( ; k <= len - 4; k += 4)
if (USE_SSE2)
{
__m128 v_src = _mm_loadu_ps(buf[0] + k);
_mm_storeu_pd(angle + k, _mm_cvtps_pd(v_src));
_mm_storeu_pd(angle + k + 2, _mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_src), 8))));
for ( ; k <= len - 4; k += 4)
{
__m128 v_src = _mm_loadu_ps(buf[0] + k);
_mm_storeu_pd(angle + k, _mm_cvtps_pd(v_src));
_mm_storeu_pd(angle + k + 2, _mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_src), 8))));
}
}
#endif
@ -728,15 +734,18 @@ void cartToPolar( InputArray src1, InputArray src2,
k = 0;
#if CV_SSE2
for ( ; k <= len - 4; k += 4)
if (USE_SSE2)
{
__m128 v_dst0 = _mm_movelh_ps(_mm_cvtpd_ps(_mm_loadu_pd(x + k)),
_mm_cvtpd_ps(_mm_loadu_pd(x + k + 2)));
__m128 v_dst1 = _mm_movelh_ps(_mm_cvtpd_ps(_mm_loadu_pd(y + k)),
_mm_cvtpd_ps(_mm_loadu_pd(y + k + 2)));
for ( ; k <= len - 4; k += 4)
{
__m128 v_dst0 = _mm_movelh_ps(_mm_cvtpd_ps(_mm_loadu_pd(x + k)),
_mm_cvtpd_ps(_mm_loadu_pd(x + k + 2)));
__m128 v_dst1 = _mm_movelh_ps(_mm_cvtpd_ps(_mm_loadu_pd(y + k)),
_mm_cvtpd_ps(_mm_loadu_pd(y + k + 2)));
_mm_storeu_ps(buf[0] + k, v_dst0);
_mm_storeu_ps(buf[1] + k, v_dst1);
_mm_storeu_ps(buf[0] + k, v_dst0);
_mm_storeu_ps(buf[1] + k, v_dst1);
}
}
#endif
@ -750,11 +759,14 @@ void cartToPolar( InputArray src1, InputArray src2,
k = 0;
#if CV_SSE2
for ( ; k <= len - 4; k += 4)
if (USE_SSE2)
{
__m128 v_src = _mm_loadu_ps(buf[0] + k);
_mm_storeu_pd(angle + k, _mm_cvtps_pd(v_src));
_mm_storeu_pd(angle + k + 2, _mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_src), 8))));
for ( ; k <= len - 4; k += 4)
{
__m128 v_src = _mm_loadu_ps(buf[0] + k);
_mm_storeu_pd(angle + k, _mm_cvtps_pd(v_src));
_mm_storeu_pd(angle + k + 2, _mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_src), 8))));
}
}
#endif
@ -832,17 +844,16 @@ static void SinCos_32f( const float *angle, float *sinval, float* cosval,
k1 = N/360.;
#if CV_AVX2
__m128d v_i = _mm_set_pd(1, 0);
__m128d v_k1 = _mm_set1_pd(k1);
__m128d v_1 = _mm_set1_pd(1);
__m128i v_N1 = _mm_set1_epi32(N - 1);
__m128i v_N4 = _mm_set1_epi32(N >> 2);
__m128d v_sin_a0 = _mm_set1_pd(sin_a0);
__m128d v_sin_a2 = _mm_set1_pd(sin_a2);
__m128d v_cos_a0 = _mm_set1_pd(cos_a0);
if (USE_AVX2)
{
__m128d v_k1 = _mm_set1_pd(k1);
__m128d v_1 = _mm_set1_pd(1);
__m128i v_N1 = _mm_set1_epi32(N - 1);
__m128i v_N4 = _mm_set1_epi32(N >> 2);
__m128d v_sin_a0 = _mm_set1_pd(sin_a0);
__m128d v_sin_a2 = _mm_set1_pd(sin_a2);
__m128d v_cos_a0 = _mm_set1_pd(cos_a0);
for ( ; i <= len - 4; i += 4)
{
__m128 v_angle = _mm_loadu_ps(angle + i);
@ -859,8 +870,8 @@ static void SinCos_32f( const float *angle, float *sinval, float* cosval,
__m128d v_sin_b = _mm_mul_pd(_mm_add_pd(_mm_mul_pd(v_sin_a0, v_t2), v_sin_a2), v_t);
__m128d v_cos_b = _mm_add_pd(_mm_mul_pd(v_cos_a0, v_t2), v_1);
__m128d v_sin_a = _mm_i32gather_pd(sin_table, v_sin_idx, 1);
__m128d v_cos_a = _mm_i32gather_pd(sin_table, v_cos_idx, 1);
__m128d v_sin_a = _mm_i32gather_pd(sin_table, v_sin_idx, 8);
__m128d v_cos_a = _mm_i32gather_pd(sin_table, v_cos_idx, 8);
__m128d v_sin_val_0 = _mm_add_pd(_mm_mul_pd(v_sin_a, v_cos_b),
_mm_mul_pd(v_cos_a, v_sin_b));
@ -868,7 +879,7 @@ static void SinCos_32f( const float *angle, float *sinval, float* cosval,
_mm_mul_pd(v_sin_a, v_sin_b));
// 2-3
v_t = _mm_mul_pd(_mm_cvtps_pd(_mm_castsi128_ps(_mm_slli_si128(_mm_castps_si128(v_angle), 8))), v_k1);
v_t = _mm_mul_pd(_mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(v_angle), 8))), v_k1);
v_it = _mm_cvtpd_epi32(v_t);
v_t = _mm_sub_pd(v_t, _mm_cvtepi32_pd(v_it));
@ -879,8 +890,8 @@ static void SinCos_32f( const float *angle, float *sinval, float* cosval,
v_sin_b = _mm_mul_pd(_mm_add_pd(_mm_mul_pd(v_sin_a0, v_t2), v_sin_a2), v_t);
v_cos_b = _mm_add_pd(_mm_mul_pd(v_cos_a0, v_t2), v_1);
v_sin_a = _mm_i32gather_pd(sin_table, v_sin_idx, 1);
v_cos_a = _mm_i32gather_pd(sin_table, v_cos_idx, 1);
v_sin_a = _mm_i32gather_pd(sin_table, v_sin_idx, 8);
v_cos_a = _mm_i32gather_pd(sin_table, v_cos_idx, 8);
__m128d v_sin_val_1 = _mm_add_pd(_mm_mul_pd(v_sin_a, v_cos_b),
_mm_mul_pd(v_cos_a, v_sin_b));
@ -1032,11 +1043,14 @@ void polarToCart( InputArray src1, InputArray src2,
vst1q_f32(y + k, vmulq_f32(vld1q_f32(y + k), v_m));
}
#elif CV_SSE2
for( ; k <= len - 4; k += 4 )
if (USE_SSE2)
{
__m128 v_m = _mm_loadu_ps(mag + k);
_mm_storeu_ps(x + k, _mm_mul_ps(_mm_loadu_ps(x + k), v_m));
_mm_storeu_ps(y + k, _mm_mul_ps(_mm_loadu_ps(y + k), v_m));
for( ; k <= len - 4; k += 4 )
{
__m128 v_m = _mm_loadu_ps(mag + k);
_mm_storeu_ps(x + k, _mm_mul_ps(_mm_loadu_ps(x + k), v_m));
_mm_storeu_ps(y + k, _mm_mul_ps(_mm_loadu_ps(y + k), v_m));
}
}
#endif
@ -1063,10 +1077,10 @@ void polarToCart( InputArray src1, InputArray src2,
x[k] = buf[0][k]*m; y[k] = buf[1][k]*m;
}
else
for( k = 0; k < len; k++ )
{
x[k] = buf[0][k]; y[k] = buf[1][k];
}
{
std::memcpy(x, buf[0], sizeof(float) * len);
std::memcpy(y, buf[1], sizeof(float) * len);
}
}
if( ptrs[0] )

6
modules/core/src/stat.cpp
Unescape View File

@ -397,6 +397,8 @@ static int countNonZero_(const T* src, int len )
return nz;
}
#if CV_SSE2
static const uchar * initPopcountTable()
{
static uchar tab[256];
@ -425,6 +427,8 @@ static const uchar * initPopcountTable()
return tab;
}
#endif
static int countNonZero8u( const uchar* src, int len )
{
int i=0, nz = 0;
@ -645,7 +649,7 @@ static int countNonZero32f( const float* src, int len )
}
static int countNonZero64f( const double* src, int len )
{
{
int i = 0, nz = 0;
#if CV_SSE2
if (USE_SSE2)

595
modules/imgproc/src/color.cpp
View File

File diff suppressed because it is too large Load Diff

671
modules/imgproc/src/imgwarp.cpp
Unescape View File

@ -1963,9 +1963,9 @@ private:
struct ResizeAreaFastVec_SIMD_32f
{
ResizeAreaFastVec_SIMD_32f(int _scale_x, int _scale_y, int _cn, int _step) :
scale_x(_scale_x), scale_y(_scale_y), cn(_cn), step(_step)
cn(_cn), step(_step)
{
fast_mode = scale_x == 2 && scale_y == 2 && (cn == 1 || cn == 3 || cn == 4);
fast_mode = _scale_x == 2 && _scale_y == 2 && (cn == 1 || cn == 4);
}
int operator() (const float * S, float * D, int w) const
@ -2005,7 +2005,6 @@ struct ResizeAreaFastVec_SIMD_32f
}
private:
int scale_x, scale_y;
int cn;
bool fast_mode;
int step;
@ -2289,9 +2288,10 @@ private:
struct ResizeAreaFastVec_SIMD_32f
{
ResizeAreaFastVec_SIMD_32f(int _scale_x, int _scale_y, int _cn, int _step) :
scale_x(_scale_x), scale_y(_scale_y), cn(_cn), step(_step)
cn(_cn), step(_step)
{
fast_mode = scale_x == 2 && scale_y == 2 && (cn == 1 || cn == 3 || cn == 4);
fast_mode = _scale_x == 2 && _scale_y == 2 && (cn == 1 || cn == 4);
fast_mode = fast_mode && checkHardwareSupport(CV_CPU_SSE2);
}
int operator() (const float * S, float * D, int w) const
@ -2335,7 +2335,6 @@ struct ResizeAreaFastVec_SIMD_32f
}
private:
int scale_x, scale_y;
int cn;
bool fast_mode;
int step;
@ -4817,6 +4816,13 @@ void cv::convertMaps( InputArray _map1, InputArray _map2,
size.height = 1;
}
#if CV_SSE2
bool useSSE2 = checkHardwareSupport(CV_CPU_SSE2);
#endif
#if CV_SSE4_1
bool useSSE4_1 = checkHardwareSupport(CV_CPU_SSE4_1);
#endif
const float scale = 1.f/INTER_TAB_SIZE;
int x, y;
for( y = 0; y < size.height; y++ )
@ -4848,24 +4854,27 @@ void cv::convertMaps( InputArray _map1, InputArray _map2,
vst2q_s16(dst1 + (x << 1), v_dst);
}
#elif CV_SSE4_1
for( ; x <= size.width - 16; x += 16 )
if (useSSE4_1)
{
__m128i v_dst0 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src1f + x)),
_mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 4)));
__m128i v_dst1 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 8)),
_mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 12)));
__m128i v_dst2 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src2f + x)),
_mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 4)));
__m128i v_dst3 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 8)),
_mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 12)));
_mm_interleave_epi16(v_dst0, v_dst1, v_dst2, v_dst3);
_mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst0);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 8), v_dst1);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 16), v_dst2);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 24), v_dst3);
for( ; x <= size.width - 16; x += 16 )
{
__m128i v_dst0 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src1f + x)),
_mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 4)));
__m128i v_dst1 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 8)),
_mm_cvtps_epi32(_mm_loadu_ps(src1f + x + 12)));
__m128i v_dst2 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src2f + x)),
_mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 4)));
__m128i v_dst3 = _mm_packs_epi32(_mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 8)),
_mm_cvtps_epi32(_mm_loadu_ps(src2f + x + 12)));
_mm_interleave_epi16(v_dst0, v_dst1, v_dst2, v_dst3);
_mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst0);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 8), v_dst1);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 16), v_dst2);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 24), v_dst3);
}
}
#endif
for( ; x < size.width; x++ )
@ -4902,47 +4911,50 @@ void cv::convertMaps( InputArray _map1, InputArray _map2,
vst1q_u16(dst2 + x, vcombine_u16(v_dst0, v_dst1));
}
#elif CV_SSE4_1
__m128 v_its = _mm_set1_ps(INTER_TAB_SIZE);
__m128i v_its1 = _mm_set1_epi32(INTER_TAB_SIZE-1);
for( ; x <= size.width - 16; x += 16 )
if (useSSE4_1)
{
__m128i v_ix0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x), v_its));
__m128i v_ix1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 4), v_its));
__m128i v_iy0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x), v_its));
__m128i v_iy1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 4), v_its));
__m128i v_dst10 = _mm_packs_epi32(_mm_srai_epi32(v_ix0, INTER_BITS),
_mm_srai_epi32(v_ix1, INTER_BITS));
__m128i v_dst12 = _mm_packs_epi32(_mm_srai_epi32(v_iy0, INTER_BITS),
_mm_srai_epi32(v_iy1, INTER_BITS));
__m128i v_dst20 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy0, v_its1), INTER_BITS),
_mm_and_si128(v_ix0, v_its1));
__m128i v_dst21 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy1, v_its1), INTER_BITS),
_mm_and_si128(v_ix1, v_its1));
_mm_storeu_si128((__m128i *)(dst2 + x), _mm_packus_epi32(v_dst20, v_dst21));
v_ix0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 8), v_its));
v_ix1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 12), v_its));
v_iy0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 8), v_its));
v_iy1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 12), v_its));
__m128i v_dst11 = _mm_packs_epi32(_mm_srai_epi32(v_ix0, INTER_BITS),
_mm_srai_epi32(v_ix1, INTER_BITS));
__m128i v_dst13 = _mm_packs_epi32(_mm_srai_epi32(v_iy0, INTER_BITS),
_mm_srai_epi32(v_iy1, INTER_BITS));
v_dst20 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy0, v_its1), INTER_BITS),
_mm_and_si128(v_ix0, v_its1));
v_dst21 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy1, v_its1), INTER_BITS),
_mm_and_si128(v_ix1, v_its1));
_mm_storeu_si128((__m128i *)(dst2 + x + 8), _mm_packus_epi32(v_dst20, v_dst21));
_mm_interleave_epi16(v_dst10, v_dst11, v_dst12, v_dst13);
_mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst10);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 8), v_dst11);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 16), v_dst12);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 24), v_dst13);
__m128 v_its = _mm_set1_ps(INTER_TAB_SIZE);
__m128i v_its1 = _mm_set1_epi32(INTER_TAB_SIZE-1);
for( ; x <= size.width - 16; x += 16 )
{
__m128i v_ix0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x), v_its));
__m128i v_ix1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 4), v_its));
__m128i v_iy0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x), v_its));
__m128i v_iy1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 4), v_its));
__m128i v_dst10 = _mm_packs_epi32(_mm_srai_epi32(v_ix0, INTER_BITS),
_mm_srai_epi32(v_ix1, INTER_BITS));
__m128i v_dst12 = _mm_packs_epi32(_mm_srai_epi32(v_iy0, INTER_BITS),
_mm_srai_epi32(v_iy1, INTER_BITS));
__m128i v_dst20 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy0, v_its1), INTER_BITS),
_mm_and_si128(v_ix0, v_its1));
__m128i v_dst21 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy1, v_its1), INTER_BITS),
_mm_and_si128(v_ix1, v_its1));
_mm_storeu_si128((__m128i *)(dst2 + x), _mm_packus_epi32(v_dst20, v_dst21));
v_ix0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 8), v_its));
v_ix1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x + 12), v_its));
v_iy0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 8), v_its));
v_iy1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src2f + x + 12), v_its));
__m128i v_dst11 = _mm_packs_epi32(_mm_srai_epi32(v_ix0, INTER_BITS),
_mm_srai_epi32(v_ix1, INTER_BITS));
__m128i v_dst13 = _mm_packs_epi32(_mm_srai_epi32(v_iy0, INTER_BITS),
_mm_srai_epi32(v_iy1, INTER_BITS));
v_dst20 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy0, v_its1), INTER_BITS),
_mm_and_si128(v_ix0, v_its1));
v_dst21 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_iy1, v_its1), INTER_BITS),
_mm_and_si128(v_ix1, v_its1));
_mm_storeu_si128((__m128i *)(dst2 + x + 8), _mm_packus_epi32(v_dst20, v_dst21));
_mm_interleave_epi16(v_dst10, v_dst11, v_dst12, v_dst13);
_mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst10);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 8), v_dst11);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 16), v_dst12);
_mm_storeu_si128((__m128i *)(dst1 + x * 2 + 24), v_dst13);
}
}
#endif
for( ; x < size.width; x++ )
@ -5005,25 +5017,28 @@ void cv::convertMaps( InputArray _map1, InputArray _map2,
vst1q_u16(dst2 + x, vcombine_u16(v_dst0, v_dst1));
}
#elif CV_SSE2
__m128 v_its = _mm_set1_ps(INTER_TAB_SIZE);
__m128i v_its1 = _mm_set1_epi32(INTER_TAB_SIZE-1);
__m128i v_y_mask = _mm_set1_epi32((INTER_TAB_SIZE-1) << 16);
for( ; x <= size.width - 4; x += 4 )
if (useSSE2)
{
__m128i v_src0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x * 2), v_its));
__m128i v_src1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x * 2 + 4), v_its));
__m128i v_dst1 = _mm_packs_epi32(_mm_srai_epi32(v_src0, INTER_BITS),
_mm_srai_epi32(v_src1, INTER_BITS));
_mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst1);
// x0 y0 x1 y1 . . .
v_src0 = _mm_packs_epi32(_mm_and_si128(v_src0, v_its1),
_mm_and_si128(v_src1, v_its1));
__m128i v_dst2 = _mm_or_si128(_mm_srli_epi32(_mm_and_si128(v_src0, v_y_mask), 16 - INTER_BITS), // y0 0 y1 0 . . .
_mm_and_si128(v_src0, v_its1)); // 0 x0 0 x1 . . .
_mm_storel_epi64((__m128i *)(dst2 + x), _mm_packus_epi32(v_dst2, v_dst2));
__m128 v_its = _mm_set1_ps(INTER_TAB_SIZE);
__m128i v_its1 = _mm_set1_epi32(INTER_TAB_SIZE-1);
__m128i v_y_mask = _mm_set1_epi32((INTER_TAB_SIZE-1) << 16);
for( ; x <= size.width - 4; x += 4 )
{
__m128i v_src0 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x * 2), v_its));
__m128i v_src1 = _mm_cvtps_epi32(_mm_mul_ps(_mm_loadu_ps(src1f + x * 2 + 4), v_its));
__m128i v_dst1 = _mm_packs_epi32(_mm_srai_epi32(v_src0, INTER_BITS),
_mm_srai_epi32(v_src1, INTER_BITS));
_mm_storeu_si128((__m128i *)(dst1 + x * 2), v_dst1);
// x0 y0 x1 y1 . . .
v_src0 = _mm_packs_epi32(_mm_and_si128(v_src0, v_its1),
_mm_and_si128(v_src1, v_its1));
__m128i v_dst2 = _mm_or_si128(_mm_srli_epi32(_mm_and_si128(v_src0, v_y_mask), 16 - INTER_BITS), // y0 0 y1 0 . . .
_mm_and_si128(v_src0, v_its1)); // 0 x0 0 x1 . . .
_mm_storel_epi64((__m128i *)(dst2 + x), _mm_packus_epi32(v_dst2, v_dst2));
}
}
#endif
for( ; x < size.width; x++ )
@ -5150,22 +5165,25 @@ void cv::convertMaps( InputArray _map1, InputArray _map2,
vst2q_f32(dst1f + (x << 1) + 8, v_dst);
}
#elif CV_SSE2
__m128i v_mask2 = _mm_set1_epi16(INTER_TAB_SIZE2-1);
__m128i v_zero = _mm_set1_epi32(0), v_mask = _mm_set1_epi32(INTER_TAB_SIZE-1);
__m128 v_scale = _mm_set1_ps(scale);
for ( ; x <= size.width - 8; x += 8)
if (useSSE2)
{
__m128i v_src = _mm_loadu_si128((__m128i const *)(src1 + x * 2));
__m128i v_fxy = src2 ? _mm_and_si128(_mm_loadu_si128((__m128i const *)(src2 + x)), v_mask2) : v_zero;
__m128i v_fxy1 = _mm_and_si128(v_fxy, v_mask);
__m128i v_fxy2 = _mm_srli_epi16(v_fxy, INTER_BITS);
__m128i v_mask2 = _mm_set1_epi16(INTER_TAB_SIZE2-1);
__m128i v_zero = _mm_set1_epi32(0), v_mask = _mm_set1_epi32(INTER_TAB_SIZE-1);
__m128 v_scale = _mm_set1_ps(scale);
for ( ; x <= size.width - 8; x += 8)
{
__m128i v_src = _mm_loadu_si128((__m128i const *)(src1 + x * 2));
__m128i v_fxy = src2 ? _mm_and_si128(_mm_loadu_si128((__m128i const *)(src2 + x)), v_mask2) : v_zero;
__m128i v_fxy1 = _mm_and_si128(v_fxy, v_mask);
__m128i v_fxy2 = _mm_srli_epi16(v_fxy, INTER_BITS);
__m128 v_add = _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(v_fxy1, v_fxy2)), v_scale);
_mm_storeu_ps(dst1f + x * 2, _mm_add_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(v_src, v_zero)), v_add));
__m128 v_add = _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(v_fxy1, v_fxy2)), v_scale);
_mm_storeu_ps(dst1f + x * 2, _mm_add_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(v_src, v_zero)), v_add));
v_add = _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(v_fxy1, v_fxy2)), v_scale);
_mm_storeu_ps(dst1f + x * 2, _mm_add_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(v_src, v_zero)), v_add));
v_add = _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(v_fxy1, v_fxy2)), v_scale);
_mm_storeu_ps(dst1f + x * 2, _mm_add_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(v_src, v_zero)), v_add));
}
}
#endif
for( ; x < size.width; x++ )
@ -5204,7 +5222,10 @@ public:
const int AB_SCALE = 1 << AB_BITS;
int round_delta = interpolation == INTER_NEAREST ? AB_SCALE/2 : AB_SCALE/INTER_TAB_SIZE/2, x, y, x1, y1;
#if CV_SSE2
bool useSIMD = checkHardwareSupport(CV_CPU_SSE2);
bool useSSE2 = checkHardwareSupport(CV_CPU_SSE2);
#endif
#if CV_SSE4_1
bool useSSE4_1 = checkHardwareSupport(CV_CPU_SSE4_1);
#endif
int bh0 = std::min(BLOCK_SZ/2, dst.rows);
@ -5243,26 +5264,29 @@ public:
vst2q_s16(xy + (x1 << 1), v_dst);
}
#elif CV_SSE4_1
__m128i v_X0 = _mm_set1_epi32(X0);
__m128i v_Y0 = _mm_set1_epi32(Y0);
for ( ; x1 <= bw - 16; x1 += 16)
if (useSSE4_1)
{
__m128i v_x0 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1))), AB_BITS),
_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 4))), AB_BITS));
__m128i v_x1 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 8))), AB_BITS),
_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 12))), AB_BITS));
__m128i v_y0 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1))), AB_BITS),
_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1 + 4))), AB_BITS));
__m128i v_y1 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1 + 8))), AB_BITS),
_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1 + 12))), AB_BITS));
_mm_interleave_epi16(v_x0, v_x1, v_y0, v_y1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2), v_x0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_x1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_y0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_y1);
__m128i v_X0 = _mm_set1_epi32(X0);
__m128i v_Y0 = _mm_set1_epi32(Y0);
for ( ; x1 <= bw - 16; x1 += 16)
{
__m128i v_x0 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1))), AB_BITS),
_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 4))), AB_BITS));
__m128i v_x1 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 8))), AB_BITS),
_mm_srai_epi32(_mm_add_epi32(v_X0, _mm_loadu_si128((__m128i const *)(adelta + x + x1 + 12))), AB_BITS));
__m128i v_y0 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1))), AB_BITS),
_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1 + 4))), AB_BITS));
__m128i v_y1 = _mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1 + 8))), AB_BITS),
_mm_srai_epi32(_mm_add_epi32(v_Y0, _mm_loadu_si128((__m128i const *)(bdelta + x + x1 + 12))), AB_BITS));
_mm_interleave_epi16(v_x0, v_x1, v_y0, v_y1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2), v_x0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_x1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_y0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_y1);
}
}
#endif
for( ; x1 < bw; x1++ )
@ -5278,7 +5302,7 @@ public:
short* alpha = A + y1*bw;
x1 = 0;
#if CV_SSE2
if( useSIMD )
if( useSSE2 )
{
__m128i fxy_mask = _mm_set1_epi32(INTER_TAB_SIZE - 1);
__m128i XX = _mm_set1_epi32(X0), YY = _mm_set1_epi32(Y0);
@ -5672,6 +5696,7 @@ public:
bh0 = std::min(BLOCK_SZ*BLOCK_SZ/bw0, height);
#if CV_SSE4_1
bool haveSSE4_1 = checkHardwareSupport(CV_CPU_SSE4_1);
__m128d v_M0 = _mm_set1_pd(M[0]);
__m128d v_M3 = _mm_set1_pd(M[3]);
__m128d v_M6 = _mm_set1_pd(M[6]);
@ -5706,109 +5731,112 @@ public:
x1 = 0;
#if CV_SSE4_1
__m128d v_X0d = _mm_set1_pd(X0);
__m128d v_Y0d = _mm_set1_pd(Y0);
__m128d v_W0 = _mm_set1_pd(W0);
__m128d v_x1 = _mm_set_pd(1, 0);
for( ; x1 <= bw - 16; x1 += 16 )
if (haveSSE4_1)
{
// 0-3
__m128i v_X0, v_Y0;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
__m128d v_X0d = _mm_set1_pd(X0);
__m128d v_Y0d = _mm_set1_pd(Y0);
__m128d v_W0 = _mm_set1_pd(W0);
__m128d v_x1 = _mm_set_pd(1, 0);
// 4-8
__m128i v_X1, v_Y1;
for( ; x1 <= bw - 16; x1 += 16 )
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 0-3
__m128i v_X0, v_Y0;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 8-11
__m128i v_X2, v_Y2;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 4-8
__m128i v_X1, v_Y1;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 12-15
__m128i v_X3, v_Y3;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 8-11
__m128i v_X2, v_Y2;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 12-15
__m128i v_X3, v_Y3;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_1, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// convert to 16s
v_X0 = _mm_packs_epi32(v_X0, v_X1);
v_X1 = _mm_packs_epi32(v_X2, v_X3);
v_Y0 = _mm_packs_epi32(v_Y0, v_Y1);
v_Y1 = _mm_packs_epi32(v_Y2, v_Y3);
// convert to 16s
v_X0 = _mm_packs_epi32(v_X0, v_X1);
v_X1 = _mm_packs_epi32(v_X2, v_X3);
v_Y0 = _mm_packs_epi32(v_Y0, v_Y1);
v_Y1 = _mm_packs_epi32(v_Y2, v_Y3);
_mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
_mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2), v_X0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_Y0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_Y1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2), v_X0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_Y0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_Y1);
}
}
#endif
@ -5831,122 +5859,125 @@ public:
x1 = 0;
#if CV_SSE4_1
__m128d v_X0d = _mm_set1_pd(X0);
__m128d v_Y0d = _mm_set1_pd(Y0);
__m128d v_W0 = _mm_set1_pd(W0);
__m128d v_x1 = _mm_set_pd(1, 0);
for( ; x1 <= bw - 16; x1 += 16 )
if (haveSSE4_1)
{
// 0-3
__m128i v_X0, v_Y0;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
__m128d v_X0d = _mm_set1_pd(X0);
__m128d v_Y0d = _mm_set1_pd(Y0);
__m128d v_W0 = _mm_set1_pd(W0);
__m128d v_x1 = _mm_set_pd(1, 0);
// 4-8
__m128i v_X1, v_Y1;
for( ; x1 <= bw - 16; x1 += 16 )
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 0-3
__m128i v_X0, v_Y0;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 8-11
__m128i v_X2, v_Y2;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 4-8
__m128i v_X1, v_Y1;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 12-15
__m128i v_X3, v_Y3;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// 8-11
__m128i v_X2, v_Y2;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y2 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// store alpha
__m128i v_alpha0 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y0, v_itsi1), INTER_BITS),
_mm_and_si128(v_X0, v_itsi1));
__m128i v_alpha1 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y1, v_itsi1), INTER_BITS),
_mm_and_si128(v_X1, v_itsi1));
_mm_storeu_si128((__m128i *)(alpha + x1), _mm_packs_epi32(v_alpha0, v_alpha1));
v_alpha0 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y2, v_itsi1), INTER_BITS),
_mm_and_si128(v_X2, v_itsi1));
v_alpha1 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y3, v_itsi1), INTER_BITS),
_mm_and_si128(v_X3, v_itsi1));
_mm_storeu_si128((__m128i *)(alpha + x1 + 8), _mm_packs_epi32(v_alpha0, v_alpha1));
// convert to 16s
v_X0 = _mm_packs_epi32(_mm_srai_epi32(v_X0, INTER_BITS), _mm_srai_epi32(v_X1, INTER_BITS));
v_X1 = _mm_packs_epi32(_mm_srai_epi32(v_X2, INTER_BITS), _mm_srai_epi32(v_X3, INTER_BITS));
v_Y0 = _mm_packs_epi32(_mm_srai_epi32(v_Y0, INTER_BITS), _mm_srai_epi32(v_Y1, INTER_BITS));
v_Y1 = _mm_packs_epi32(_mm_srai_epi32(v_Y2, INTER_BITS), _mm_srai_epi32(v_Y3, INTER_BITS));
_mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2), v_X0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_Y0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_Y1);
// 12-15
__m128i v_X3, v_Y3;
{
__m128d v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY0 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_W = _mm_add_pd(_mm_mul_pd(v_M6, v_x1), v_W0);
v_W = _mm_andnot_pd(_mm_cmpeq_pd(v_W, v_zero), _mm_div_pd(v_its, v_W));
__m128d v_fX1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_X0d, _mm_mul_pd(v_M0, v_x1)), v_W)));
__m128d v_fY1 = _mm_max_pd(v_intmin, _mm_min_pd(v_intmax, _mm_mul_pd(_mm_add_pd(v_Y0d, _mm_mul_pd(v_M3, v_x1)), v_W)));
v_x1 = _mm_add_pd(v_x1, v_2);
v_X3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fX1))));
v_Y3 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY0)),
_mm_castsi128_ps(_mm_cvtpd_epi32(v_fY1))));
}
// store alpha
__m128i v_alpha0 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y0, v_itsi1), INTER_BITS),
_mm_and_si128(v_X0, v_itsi1));
__m128i v_alpha1 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y1, v_itsi1), INTER_BITS),
_mm_and_si128(v_X1, v_itsi1));
_mm_storeu_si128((__m128i *)(alpha + x1), _mm_packs_epi32(v_alpha0, v_alpha1));
v_alpha0 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y2, v_itsi1), INTER_BITS),
_mm_and_si128(v_X2, v_itsi1));
v_alpha1 = _mm_add_epi32(_mm_slli_epi32(_mm_and_si128(v_Y3, v_itsi1), INTER_BITS),
_mm_and_si128(v_X3, v_itsi1));
_mm_storeu_si128((__m128i *)(alpha + x1 + 8), _mm_packs_epi32(v_alpha0, v_alpha1));
// convert to 16s
v_X0 = _mm_packs_epi32(_mm_srai_epi32(v_X0, INTER_BITS), _mm_srai_epi32(v_X1, INTER_BITS));
v_X1 = _mm_packs_epi32(_mm_srai_epi32(v_X2, INTER_BITS), _mm_srai_epi32(v_X3, INTER_BITS));
v_Y0 = _mm_packs_epi32(_mm_srai_epi32(v_Y0, INTER_BITS), _mm_srai_epi32(v_Y1, INTER_BITS));
v_Y1 = _mm_packs_epi32(_mm_srai_epi32(v_Y2, INTER_BITS), _mm_srai_epi32(v_Y3, INTER_BITS));
_mm_interleave_epi16(v_X0, v_X1, v_Y0, v_Y1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2), v_X0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 8), v_X1);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 16), v_Y0);
_mm_storeu_si128((__m128i *)(xy + x1 * 2 + 24), v_Y1);
}
}
#endif

10
modules/imgproc/src/pyramids.cpp
Unescape View File

@ -386,10 +386,10 @@ struct PyrUpVec_32s16s
__m128i v_dst01 = _mm_add_epi32(_mm_add_epi32(v_r0, v_r2), _mm_add_epi32(v_2r1, v_4r1));
__m128i v_dst11 = _mm_slli_epi32(_mm_add_epi32(v_r1, v_r2), 2);
_mm_storeu_si128((__m128i *)(dst0 + x),
_mm_storeu_si128((__m128i *)(dst0 + x),
_mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_dst00, v_delta), 6),
_mm_srai_epi32(_mm_add_epi32(v_dst01, v_delta), 6)));
_mm_storeu_si128((__m128i *)(dst1 + x),
_mm_storeu_si128((__m128i *)(dst1 + x),
_mm_packs_epi32(_mm_srai_epi32(_mm_add_epi32(v_dst10, v_delta), 6),
_mm_srai_epi32(_mm_add_epi32(v_dst11, v_delta), 6)));
}
@ -446,10 +446,10 @@ struct PyrUpVec_32s16u
__m128i v_dst01 = _mm_add_epi32(_mm_add_epi32(v_r0, v_r2), _mm_add_epi32(v_2r1, v_4r1));
__m128i v_dst11 = _mm_slli_epi32(_mm_add_epi32(v_r1, v_r2), 2);
_mm_storeu_si128((__m128i *)(dst0 + x),
_mm_storeu_si128((__m128i *)(dst0 + x),
_mm_packus_epi32(_mm_srli_epi32(_mm_add_epi32(v_dst00, v_delta), 6),
_mm_srli_epi32(_mm_add_epi32(v_dst01, v_delta), 6)));
_mm_storeu_si128((__m128i *)(dst1 + x),
_mm_storeu_si128((__m128i *)(dst1 + x),
_mm_packus_epi32(_mm_srli_epi32(_mm_add_epi32(v_dst10, v_delta), 6),
_mm_srli_epi32(_mm_add_epi32(v_dst11, v_delta), 6)));
}
@ -491,7 +491,7 @@ struct PyrUpVec_32f
const float *row0 = src[0], *row1 = src[1], *row2 = src[2];
float *dst0 = dst[0], *dst1 = dst[1];
__m128 v_6 = _mm_set1_ps(6.0f), v_scale = _mm_set1_ps(1.f/64.0f),
__m128 v_6 = _mm_set1_ps(6.0f), v_scale = _mm_set1_ps(1.f/64.0f),
v_scale4 = _mm_mul_ps(v_scale, _mm_set1_ps(4.0f));
for( ; x <= width - 8; x += 8 )

Write Preview
Loading…
Cancel
Save