mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-01-18 14:13:15 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

divide

Browse Source
This commit is contained in:
Ilya Lavrenov 2015-01-12 10:59:30 +03:00
parent 7b060d9122
commit 44d89638fd
2 changed files with 244 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
modules/core/perf/perf_arithm.cpp
View File

@ -242,3 +242,17 @@ PERF_TEST_P(Size_MatType, multiplyScale, TYPICAL_MATS_CORE_ARITHM)
SANITY_CHECK(c, 1e-8);
}
PERF_TEST_P(Size_MatType, divide, TYPICAL_MATS_CORE_ARITHM)
{
Size sz = get<0>(GetParam());
int type = get<1>(GetParam());
cv::Mat a(sz, type), b(sz, type), c(sz, type);
double scale = 0.5;
declare.in(a, b, WARMUP_RNG).out(c);
TEST_CYCLE() divide(a, b, c, scale);
SANITY_CHECK_NOTHING();
}

231
modules/core/src/arithm.cpp
View File

@ -2610,6 +2610,233 @@ mul_( const T* src1, size_t step1, const T* src2, size_t step2,
}
}
template <typename T>
struct Div_SIMD
{
int operator() (const T *, const T *, T *, int, double) const
{
return 0;
}
};
#if CV_SSE2
#if CV_SSE4_1
template <>
struct Div_SIMD<uchar>
{
int operator() (const uchar * src1, const uchar * src2, uchar * dst, int width, double scale) const
{
int x = 0;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
for ( ; x <= width - 8; x += 8)
{
__m128i v_src1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(src1 + x)), v_zero);
__m128i _v_src2 = _mm_loadl_epi64((const __m128i *)(src2 + x));
__m128i v_src2 = _mm_unpacklo_epi8(_v_src2, v_zero);
__m128i v_src1i = _mm_unpacklo_epi16(v_src1, v_zero);
__m128i v_src2i = _mm_unpacklo_epi16(v_src2, v_zero);
__m128d v_src1d = _mm_cvtepi32_pd(v_src1i);
__m128d v_src2d = _mm_cvtepi32_pd(v_src2i);
__m128i v_dst0 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
v_src1d = _mm_cvtepi32_pd(_mm_srli_si128(v_src1i, 8));
v_src2d = _mm_cvtepi32_pd(_mm_srli_si128(v_src2i, 8));
__m128i v_dst1 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
__m128i v_dst_0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(v_dst0), _mm_castsi128_ps(v_dst1)));
v_src1i = _mm_unpackhi_epi16(v_src1, v_zero);
v_src2i = _mm_unpackhi_epi16(v_src2, v_zero);
v_src1d = _mm_cvtepi32_pd(v_src1i);
v_src2d = _mm_cvtepi32_pd(v_src2i);
v_dst0 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
v_src1d = _mm_cvtepi32_pd(_mm_srli_si128(v_src1i, 8));
v_src2d = _mm_cvtepi32_pd(_mm_srli_si128(v_src2i, 8));
v_dst1 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
__m128i v_dst_1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(v_dst0), _mm_castsi128_ps(v_dst1)));
__m128i v_mask = _mm_cmpeq_epi8(_v_src2, v_zero);
_mm_storel_epi64((__m128i *)(dst + x), _mm_andnot_si128(v_mask, _mm_packus_epi16(_mm_packs_epi32(v_dst_0, v_dst_1), v_zero)));
}
return x;
}
};
#endif // CV_SSE4_1
template <>
struct Div_SIMD<schar>
{
int operator() (const schar * src1, const schar * src2, schar * dst, int width, double scale) const
{
int x = 0;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
for ( ; x <= width - 8; x += 8)
{
__m128i v_src1 = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, _mm_loadl_epi64((const __m128i *)(src1 + x))), 8);
__m128i _v_src2 = _mm_loadl_epi64((const __m128i *)(src2 + x));
__m128i v_src2 = _mm_srai_epi16(_mm_unpacklo_epi8(v_zero, _v_src2), 8);
__m128i v_src1i = _mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src1), 16);
__m128i v_src2i = _mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src2), 16);
__m128d v_src1d = _mm_cvtepi32_pd(v_src1i);
__m128d v_src2d = _mm_cvtepi32_pd(v_src2i);
__m128i v_dst0 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
v_src1d = _mm_cvtepi32_pd(_mm_srli_si128(v_src1i, 8));
v_src2d = _mm_cvtepi32_pd(_mm_srli_si128(v_src2i, 8));
__m128i v_dst1 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
__m128i v_dst_0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(v_dst0), _mm_castsi128_ps(v_dst1)));
v_src1i = _mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src1), 16);
v_src2i = _mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src2), 16);
v_src1d = _mm_cvtepi32_pd(v_src1i);
v_src2d = _mm_cvtepi32_pd(v_src2i);
v_dst0 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
v_src1d = _mm_cvtepi32_pd(_mm_srli_si128(v_src1i, 8));
v_src2d = _mm_cvtepi32_pd(_mm_srli_si128(v_src2i, 8));
v_dst1 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
__m128i v_dst_1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(v_dst0), _mm_castsi128_ps(v_dst1)));
__m128i v_mask = _mm_cmpeq_epi8(_v_src2, v_zero);
_mm_storel_epi64((__m128i *)(dst + x), _mm_andnot_si128(v_mask, _mm_packs_epi16(_mm_packs_epi32(v_dst_0, v_dst_1), v_zero)));
}
return x;
}
};
#if CV_SSE4_1
template <>
struct Div_SIMD<ushort>
{
int operator() (const ushort * src1, const ushort * src2, ushort * dst, int width, double scale) const
{
int x = 0;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
for ( ; x <= width - 8; x += 8)
{
__m128i v_src1 = _mm_loadu_si128((const __m128i *)(src1 + x));
__m128i v_src2 = _mm_loadu_si128((const __m128i *)(src2 + x));
__m128i v_src1i = _mm_unpacklo_epi16(v_src1, v_zero);
__m128i v_src2i = _mm_unpacklo_epi16(v_src2, v_zero);
__m128d v_src1d = _mm_cvtepi32_pd(v_src1i);
__m128d v_src2d = _mm_cvtepi32_pd(v_src2i);
__m128i v_dst0 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
v_src1d = _mm_cvtepi32_pd(_mm_srli_si128(v_src1i, 8));
v_src2d = _mm_cvtepi32_pd(_mm_srli_si128(v_src2i, 8));
__m128i v_dst1 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
__m128i v_dst_0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(v_dst0), _mm_castsi128_ps(v_dst1)));
v_src1i = _mm_unpackhi_epi16(v_src1, v_zero);
v_src2i = _mm_unpackhi_epi16(v_src2, v_zero);
v_src1d = _mm_cvtepi32_pd(v_src1i);
v_src2d = _mm_cvtepi32_pd(v_src2i);
v_dst0 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
v_src1d = _mm_cvtepi32_pd(_mm_srli_si128(v_src1i, 8));
v_src2d = _mm_cvtepi32_pd(_mm_srli_si128(v_src2i, 8));
v_dst1 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
__m128i v_dst_1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(v_dst0), _mm_castsi128_ps(v_dst1)));
__m128i v_mask = _mm_cmpeq_epi16(v_src2, v_zero);
_mm_storeu_si128((__m128i *)(dst + x), _mm_andnot_si128(v_mask, _mm_packus_epi32(v_dst_0, v_dst_1)));
}
return x;
}
};
#endif // CV_SSE4_1
template <>
struct Div_SIMD<short>
{
int operator() (const short * src1, const short * src2, short * dst, int width, double scale) const
{
int x = 0;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
for ( ; x <= width - 8; x += 8)
{
__m128i v_src1 = _mm_loadu_si128((const __m128i *)(src1 + x));
__m128i v_src2 = _mm_loadu_si128((const __m128i *)(src2 + x));
__m128i v_src1i = _mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src1), 16);
__m128i v_src2i = _mm_srai_epi32(_mm_unpacklo_epi16(v_zero, v_src2), 16);
__m128d v_src1d = _mm_cvtepi32_pd(v_src1i);
__m128d v_src2d = _mm_cvtepi32_pd(v_src2i);
__m128i v_dst0 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
v_src1d = _mm_cvtepi32_pd(_mm_srli_si128(v_src1i, 8));
v_src2d = _mm_cvtepi32_pd(_mm_srli_si128(v_src2i, 8));
__m128i v_dst1 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
__m128i v_dst_0 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(v_dst0), _mm_castsi128_ps(v_dst1)));
v_src1i = _mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src1), 16);
v_src2i = _mm_srai_epi32(_mm_unpackhi_epi16(v_zero, v_src2), 16);
v_src1d = _mm_cvtepi32_pd(v_src1i);
v_src2d = _mm_cvtepi32_pd(v_src2i);
v_dst0 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
v_src1d = _mm_cvtepi32_pd(_mm_srli_si128(v_src1i, 8));
v_src2d = _mm_cvtepi32_pd(_mm_srli_si128(v_src2i, 8));
v_dst1 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
__m128i v_dst_1 = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(v_dst0), _mm_castsi128_ps(v_dst1)));
__m128i v_mask = _mm_cmpeq_epi16(v_src2, v_zero);
_mm_storeu_si128((__m128i *)(dst + x), _mm_andnot_si128(v_mask, _mm_packs_epi32(v_dst_0, v_dst_1)));
}
return x;
}
};
template <>
struct Div_SIMD<int>
{
int operator() (const int * src1, const int * src2, int * dst, int width, double scale) const
{
int x = 0;
__m128d v_scale = _mm_set1_pd(scale);
__m128i v_zero = _mm_setzero_si128();
for ( ; x <= width - 4; x += 4)
{
__m128i v_src1 = _mm_loadu_si128((const __m128i *)(src1 + x));
__m128i v_src2 = _mm_loadu_si128((const __m128i *)(src2 + x));
__m128d v_src1d = _mm_cvtepi32_pd(v_src1);
__m128d v_src2d = _mm_cvtepi32_pd(v_src2);
__m128i v_dst0 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
v_src1d = _mm_cvtepi32_pd(_mm_srli_si128(v_src1, 8));
v_src2d = _mm_cvtepi32_pd(_mm_srli_si128(v_src2, 8));
__m128i v_dst1 = _mm_cvtpd_epi32(_mm_div_pd(_mm_mul_pd(v_src1d, v_scale), v_src2d));
__m128i v_dst = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(v_dst0), _mm_castsi128_ps(v_dst1)));
__m128i v_mask = _mm_cmpeq_epi32(v_src2, v_zero);
_mm_storeu_si128((__m128i *)(dst + x), _mm_andnot_si128(v_mask, v_dst));
}
return x;
}
};
#endif
template<typename T> static void
div_( const T* src1, size_t step1, const T* src2, size_t step2,
T* dst, size_t step, Size size, double scale )
@ -2618,9 +2845,11 @@ div_( const T* src1, size_t step1, const T* src2, size_t step2,
step2 /= sizeof(src2[0]);
step /= sizeof(dst[0]);
Div_SIMD<T> vop;
for( ; size.height--; src1 += step1, src2 += step2, dst += step )
{
int i = 0;
int i = vop(src1, src2, dst, size.width, scale);
#if CV_ENABLE_UNROLLED
for( ; i <= size.width - 4; i += 4 )
{