diff --git a/modules/imgproc/src/fixedpoint.inl.hpp b/modules/imgproc/src/fixedpoint.inl.hpp index 3183c72feaba65af495bb963ceaf3d74e80737e6..1c5c88a480a4cc779a427e49241b47876e121b44 100644 --- a/modules/imgproc/src/fixedpoint.inl.hpp +++ b/modules/imgproc/src/fixedpoint.inl.hpp @@ -76,6 +76,8 @@ public: // int64_t nval = (int64_t)val + val2.val + nfrac >> 32; // return nval > MAXINT32 ? beConv(MAXINT32, MAXINT32) : beConv((int32_t)(nval), 0); // } + CV_ALWAYS_INLINE fixedpoint64 operator >> (int n) const { return fixedpoint64(val >> n); } + CV_ALWAYS_INLINE fixedpoint64 operator << (int n) const { return fixedpoint64(val << n); } template CV_ALWAYS_INLINE operator ET() const { return cv::saturate_cast((int64_t)fixedround((uint64_t)val) >> fixedShift); } CV_ALWAYS_INLINE operator double() const { return (double)val / (1LL << fixedShift); } @@ -129,6 +131,8 @@ public: // int64_t nval = (int64_t)val + val2.val + nfrac >> 32; // return nval > MAXINT32 ? beConv(MAXINT32, MAXINT32) : beConv((int32_t)(nval), 0); // } + CV_ALWAYS_INLINE ufixedpoint64 operator >> (int n) const { return ufixedpoint64(val >> n); } + CV_ALWAYS_INLINE ufixedpoint64 operator << (int n) const { return ufixedpoint64(val << n); } template CV_ALWAYS_INLINE operator ET() const { return cv::saturate_cast(fixedround(val) >> fixedShift); } CV_ALWAYS_INLINE operator double() const { return (double)val / (1LL << fixedShift); } @@ -170,6 +174,8 @@ public: // int32_t nval = (int32_t)val + val2.val + nfrac >> 32; // return nval > MAXINT32 ? beConv(MAXINT32, MAXINT32) : beConv((int32_t)(nval), 0); // } + CV_ALWAYS_INLINE fixedpoint32 operator >> (int n) const { return fixedpoint32(val >> n); } + CV_ALWAYS_INLINE fixedpoint32 operator << (int n) const { return fixedpoint32(val << n); } template CV_ALWAYS_INLINE operator ET() const { return cv::saturate_cast((int32_t)fixedround((uint32_t)val) >> fixedShift); } CV_ALWAYS_INLINE operator double() const { return (double)val / (1 << fixedShift); } @@ -209,6 +215,8 @@ public: // int32_t nval = (int32_t)val + val2.val + nfrac >> 32; // return nval > MAXINT32 ? beConv(MAXINT32, MAXINT32) : beConv((int32_t)(nval), 0); // } + CV_ALWAYS_INLINE ufixedpoint32 operator >> (int n) const { return ufixedpoint32(val >> n); } + CV_ALWAYS_INLINE ufixedpoint32 operator << (int n) const { return ufixedpoint32(val << n); } template CV_ALWAYS_INLINE operator ET() const { return cv::saturate_cast(fixedround(val) >> fixedShift); } CV_ALWAYS_INLINE operator double() const { return (double)val / (1 << fixedShift); } @@ -241,6 +249,8 @@ public: CV_ALWAYS_INLINE fixedpoint32 operator * (const fixedpoint16& val2) const { return (int32_t)val * (int32_t)(val2.val); } CV_ALWAYS_INLINE fixedpoint16 operator + (const fixedpoint16& val2) const { return fixedpoint16((int16_t)(val + val2.val)); } CV_ALWAYS_INLINE fixedpoint16 operator - (const fixedpoint16& val2) const { return fixedpoint16((int16_t)(val - val2.val)); } + CV_ALWAYS_INLINE fixedpoint16 operator >> (int n) const { return fixedpoint16((int16_t)(val >> n)); } + CV_ALWAYS_INLINE fixedpoint16 operator << (int n) const { return fixedpoint16((int16_t)(val << n)); } template CV_ALWAYS_INLINE operator ET() const { return cv::saturate_cast((int16_t)fixedround((uint16_t)val) >> fixedShift); } CV_ALWAYS_INLINE operator double() const { return (double)val / (1 << fixedShift); } @@ -271,12 +281,8 @@ public: CV_ALWAYS_INLINE ufixedpoint32 operator * (const ufixedpoint16& val2) const { return ((uint32_t)val * (uint32_t)(val2.val)); } CV_ALWAYS_INLINE ufixedpoint16 operator + (const ufixedpoint16& val2) const { return ufixedpoint16((uint16_t)(val + val2.val)); } CV_ALWAYS_INLINE ufixedpoint16 operator - (const ufixedpoint16& val2) const { return ufixedpoint16((uint16_t)(val - val2.val)); } - // CV_ALWAYS_INLINE fixedpoint16 operator + (const fixedpoint16& val2) const - // { - // int16_t nfrac = (int32_t)frac + val2.frac; - // int16_t nval = (int32_t)val + val2.val + nfrac >> 16; - // return nval > MAXINT32 ? beConv(MAXINT16, MAXINT16) : beConv((int16_t)(nval), 0); - // } + CV_ALWAYS_INLINE ufixedpoint16 operator >> (int n) const { return ufixedpoint16((uint16_t)(val >> n)); } + CV_ALWAYS_INLINE ufixedpoint16 operator << (int n) const { return ufixedpoint16((uint16_t)(val << n)); } template CV_ALWAYS_INLINE operator ET() const { return cv::saturate_cast(fixedround(val) >> fixedShift); } CV_ALWAYS_INLINE operator double() const { return (double)val / (1 << fixedShift); } diff --git a/modules/imgproc/src/smooth.cpp b/modules/imgproc/src/smooth.cpp index 8ff072eb2be298b21e306f41bb963e1e2d4750d6..10d14e398156947a92adb7099a8a824491b0eab3 100644 --- a/modules/imgproc/src/smooth.cpp +++ b/modules/imgproc/src/smooth.cpp @@ -42,6 +42,9 @@ //M*/ #include "precomp.hpp" + +#include + #include "opencv2/core/hal/intrin.hpp" #include "opencl_kernels_imgproc.hpp" @@ -49,6 +52,7 @@ #include "filter.hpp" +#include "fixedpoint.inl.hpp" /* * This file includes the code, contributed by Simon Perreault * (the function icvMedianBlur_8u_O1) @@ -1763,7 +1767,1409 @@ cv::Mat cv::getGaussianKernel( int n, double sigma, int ktype ) namespace cv { -static void createGaussianKernels( Mat & kx, Mat & ky, int type, Size & ksize, +template +static std::vector getFixedpointGaussianKernel( int n, double sigma ) +{ + if (sigma <= 0) + { + if(n == 1) + return std::vector(1, softdouble(1.0)); + else if(n == 3) + { + T v3[] = { softdouble(0.25), softdouble(0.5), softdouble(0.25) }; + return std::vector(v3, v3 + 3); + } + else if(n == 5) + { + T v5[] = { softdouble(0.0625), softdouble(0.25), softdouble(0.375), softdouble(0.25), softdouble(0.0625) }; + return std::vector(v5, v5 + 5); + } + else if(n == 7) + { + T v7[] = { softdouble(0.03125), softdouble(0.109375), softdouble(0.21875), softdouble(0.28125), softdouble(0.21875), softdouble(0.109375), softdouble(0.03125) }; + return std::vector(v7, v7 + 7); + } + } + + + softdouble sigmaX = sigma > 0 ? softdouble(sigma) : mulAdd(softdouble(n),softdouble(0.15),softdouble(0.35));// softdouble(((n-1)*0.5 - 1)*0.3 + 0.8) + softdouble scale2X = softdouble(-0.5*0.25)/(sigmaX*sigmaX); + std::vector values(n); + softdouble sum(0.); + for(int i = 0, x = 1 - n; i < n; i++, x+=2 ) + { + // x = i - (n - 1)*0.5 + // t = std::exp(scale2X*x*x) + values[i] = exp(softdouble(x*x)*scale2X); + sum += values[i]; + } + sum = softdouble::one()/sum; + + std::vector kernel(n); + for(int i = 0; i < n; i++ ) + { + kernel[i] = values[i] * sum; + } + + return kernel; +}; + +template +void hlineSmooth1N(const ET* src, int cn, const FT* m, int, FT* dst, int len, int) +{ + for (int i = 0; i < len*cn; i++, src++, dst++) + *dst = (*m) * (*src); +} +template <> +void hlineSmooth1N(const uint8_t* src, int cn, const ufixedpoint16* m, int, ufixedpoint16* dst, int len, int) +{ + int lencn = len*cn; + v_uint16x8 v_mul = v_setall_u16(*((uint16_t*)m)); + int i = 0; + for (; i < lencn - 15; i += 16) + { + v_uint8x16 v_src = v_load(src + i); + v_uint16x8 v_tmp0, v_tmp1; + v_expand(v_src, v_tmp0, v_tmp1); + v_store((uint16_t*)dst + i, v_mul*v_tmp0); + v_store((uint16_t*)dst + i + 8, v_mul*v_tmp1); + } + if (i < lencn - 7) + { + 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++) + dst[i] = m[0] * src[i]; +} +template +void hlineSmooth1N1(const ET* src, int cn, const FT*, int, FT* dst, int len, int) +{ + for (int i = 0; i < len*cn; i++, src++, dst++) + *dst = *src; +} +template <> +void hlineSmooth1N1(const uint8_t* src, int cn, const ufixedpoint16*, int, ufixedpoint16* dst, int len, int) +{ + int lencn = len*cn; + int i = 0; + for (; i < lencn - 15; i += 16) + { + v_uint8x16 v_src = v_load(src + i); + v_uint16x8 v_tmp0, v_tmp1; + v_expand(v_src, v_tmp0, v_tmp1); + 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 - 7) + { + 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++) + dst[i] = src[i]; +} +template +void hlineSmooth3N(const ET* src, int cn, const FT* m, int, FT* dst, int len, int borderType) +{ + if (len == 1) + { + FT msum = borderType != BORDER_CONSTANT ? m[0] + m[1] + m[2] : m[1]; + for (int k = 0; k < cn; k++) + dst[k] = msum * src[k]; + } + else + { + // Point that fall left from border + for (int k = 0; k < cn; k++) + dst[k] = m[1] * src[k] + m[2] * src[cn + k]; + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int src_idx = borderInterpolate(-1, len, borderType); + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[0] * src[src_idx*cn + k]; + } + + src += cn; dst += cn; + for (int i = cn; i < (len - 1)*cn; i++, src++, dst++) + *dst = m[0] * src[-cn] + m[1] * src[0] + m[2] * src[cn]; + + // Point that fall right from border + for (int k = 0; k < cn; k++) + dst[k] = m[0] * src[k - cn] + m[1] * src[k]; + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int src_idx = (borderInterpolate(len, len, borderType) - (len - 1))*cn; + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[2] * src[src_idx + k]; + } + } +} +template <> +void hlineSmooth3N(const uint8_t* src, int cn, const ufixedpoint16* m, int, ufixedpoint16* dst, int len, int borderType) +{ + if (len == 1) + { + ufixedpoint16 msum = borderType != BORDER_CONSTANT ? m[0] + m[1] + m[2] : m[1]; + for (int k = 0; k < cn; k++) + dst[k] = msum * src[k]; + } + else + { + // Point that fall left from border + for (int k = 0; k < cn; k++) + dst[k] = m[1] * src[k] + m[2] * src[cn + k]; + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int src_idx = borderInterpolate(-1, len, borderType); + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[0] * src[src_idx*cn + k]; + } + + src += cn; dst += cn; + int i = cn, lencn = (len - 1)*cn; + v_int16x8 v_mul01 = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)m))); + v_int16x8 v_mul2 = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + 2)))); + for (; i < lencn - 15; i += 16, src += 16, dst += 16) + { + v_uint16x8 v_src00, v_src01, v_src10, v_src11; + v_int16x8 v_tmp0, v_tmp1; + + v_expand(v_load(src - cn), v_src00, v_src01); + v_expand(v_load(src), v_src10, v_src11); + v_zip(v_reinterpret_as_s16(v_src00), v_reinterpret_as_s16(v_src10), v_tmp0, v_tmp1); + v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul01); + v_zip(v_reinterpret_as_s16(v_src01), v_reinterpret_as_s16(v_src11), v_tmp0, v_tmp1); + v_int32x4 v_res2 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res3 = v_dotprod(v_tmp1, v_mul01); + + v_int32x4 v_resj0, v_resj1, v_resj2, v_resj3; + v_expand(v_load(src + cn), v_src00, v_src01); + v_mul_expand(v_reinterpret_as_s16(v_src00), v_mul2, v_resj0, v_resj1); + v_mul_expand(v_reinterpret_as_s16(v_src01), v_mul2, v_resj2, v_resj3); + v_res0 += v_resj0; + v_res1 += v_resj1; + v_res2 += v_resj2; + v_res3 += v_resj3; + + v_store((uint16_t*)dst, v_pack(v_reinterpret_as_u32(v_res0), v_reinterpret_as_u32(v_res1))); + v_store((uint16_t*)dst + 8, v_pack(v_reinterpret_as_u32(v_res2), v_reinterpret_as_u32(v_res3))); + } + for (; i < lencn; i++, src++, dst++) + *dst = m[0] * src[-cn] + m[1] * src[0] + m[2] * src[cn]; + + // Point that fall right from border + for (int k = 0; k < cn; k++) + dst[k] = m[0] * src[k - cn] + m[1] * src[k]; + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int src_idx = (borderInterpolate(len, len, borderType) - (len - 1))*cn; + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[2] * src[src_idx + k]; + } + } +} +template +void hlineSmooth3N121(const ET* src, int cn, const FT*, int, FT* dst, int len, int borderType) +{ + if (len == 1) + { + if(borderType != BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + dst[k] = FT(src[k]); + else + for (int k = 0; k < cn; k++) + dst[k] = FT(src[k])>>1; + } + else + { + // Point that fall left from border + for (int k = 0; k < cn; k++) + dst[k] = (FT(src[k])>>1) + (FT(src[cn + k])>>2); + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int src_idx = borderInterpolate(-1, len, borderType); + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + (FT(src[src_idx*cn + k])>>2); + } + + src += cn; dst += cn; + for (int i = cn; i < (len - 1)*cn; i++, src++, dst++) + *dst = ((FT(src[-cn]) + FT(src[cn]))>>2) + (FT(src[0])>>1); + + // Point that fall right from border + for (int k = 0; k < cn; k++) + dst[k] = (FT(src[k - cn])>>2) + (FT(src[k])>>1); + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int src_idx = (borderInterpolate(len, len, borderType) - (len - 1))*cn; + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + (FT(src[src_idx + k])>>2); + } + } +} +template <> +void hlineSmooth3N121(const uint8_t* src, int cn, const ufixedpoint16*, int, ufixedpoint16* dst, int len, int borderType) +{ + if (len == 1) + { + if (borderType != BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + dst[k] = ufixedpoint16(src[k]); + else + for (int k = 0; k < cn; k++) + dst[k] = ufixedpoint16(src[k]) >> 1; + } + else + { + // Point that fall left from border + for (int k = 0; k < cn; k++) + dst[k] = (ufixedpoint16(src[k])>>1) + (ufixedpoint16(src[cn + k])>>2); + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int src_idx = borderInterpolate(-1, len, borderType); + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + (ufixedpoint16(src[src_idx*cn + k])>>2); + } + + src += cn; dst += cn; + int i = cn, lencn = (len - 1)*cn; + for (; i < lencn - 15; i += 16, src += 16, dst += 16) + { + v_uint16x8 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21; + v_expand(v_load(src - cn), v_src00, v_src01); + v_expand(v_load(src), v_src10, v_src11); + 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++) + *((uint16_t*)dst) = (uint16_t(src[-cn]) + uint16_t(src[cn]) + (uint16_t(src[0]) << 1)) << 6; + + // Point that fall right from border + for (int k = 0; k < cn; k++) + dst[k] = (ufixedpoint16(src[k - cn])>>2) + (ufixedpoint16(src[k])>>1); + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int src_idx = (borderInterpolate(len, len, borderType) - (len - 1))*cn; + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + (ufixedpoint16(src[src_idx + k])>>2); + } + } +} +template +void hlineSmooth5N(const ET* src, int cn, const FT* m, int, FT* dst, int len, int borderType) +{ + if (len == 1) + { + ufixedpoint16 msum = borderType != BORDER_CONSTANT ? m[0] + m[1] + m[2] + m[3] + m[4] : m[2]; + for (int k = 0; k < cn; k++) + dst[k] = msum * src[k]; + } + else if (len == 2) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + { + dst[k ] = m[2] * src[k] + m[3] * src[k+cn]; + dst[k+cn] = m[1] * src[k] + m[2] * src[k+cn]; + } + else + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + int idxp1 = borderInterpolate(2, len, borderType)*cn; + int idxp2 = borderInterpolate(3, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k ] = m[1] * src[k + idxm1] + m[2] * src[k] + m[3] * src[k + cn] + m[4] * src[k + idxp1] + m[0] * src[k + idxm2]; + dst[k + cn] = m[0] * src[k + idxm1] + m[1] * src[k] + m[2] * src[k + cn] + m[3] * src[k + idxp1] + m[4] * src[k + idxp2]; + } + } + } + else if (len == 3) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + { + dst[k ] = m[2] * src[k] + m[3] * src[k + cn] + m[4] * src[k + 2*cn]; + dst[k + cn] = m[1] * src[k] + m[2] * src[k + cn] + m[3] * src[k + 2*cn]; + dst[k + 2*cn] = m[0] * src[k] + m[1] * src[k + cn] + m[2] * src[k + 2*cn]; + } + else + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + int idxp1 = borderInterpolate(3, len, borderType)*cn; + int idxp2 = borderInterpolate(4, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k ] = m[2] * src[k] + m[3] * src[k + cn] + m[4] * src[k + 2*cn] + m[0] * src[k + idxm2] + m[1] * src[k + idxm1]; + dst[k + cn] = m[1] * src[k] + m[2] * src[k + cn] + m[3] * src[k + 2*cn] + m[0] * src[k + idxm1] + m[4] * src[k + idxp1]; + dst[k + 2*cn] = m[0] * src[k] + m[1] * src[k + cn] + m[2] * src[k + 2*cn] + m[3] * src[k + idxp1] + m[4] * src[k + idxp2]; + } + } + } + else + { + // Points that fall left from border + for (int k = 0; k < cn; k++) + { + dst[k] = m[2] * src[k] + m[3] * src[cn + k] + m[4] * src[2*cn + k]; + dst[k + cn] = m[1] * src[k] + m[2] * src[cn + k] + m[3] * src[2*cn + k] + m[4] * src[3*cn + k]; + } + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = dst[k] + m[0] * src[idxm2 + k] + m[1] * src[idxm1 + k]; + dst[k + cn] = dst[k + cn] + m[0] * src[idxm1 + k]; + } + } + + src += 2*cn; dst += 2*cn; + for (int i = 2*cn; i < (len - 2)*cn; 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]; + + // Points that fall right from border + for (int k = 0; k < cn; k++) + { + dst[k] = m[0] * src[k - 2*cn] + m[1] * src[k - cn] + m[2] * src[k] + m[3] * src[k + cn]; + dst[k + cn] = m[0] * src[k - cn] + m[1] * src[k] + m[2] * src[k + cn]; + } + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int idxp1 = (borderInterpolate(len, len, borderType) - (len - 2))*cn; + int idxp2 = (borderInterpolate(len+1, len, borderType) - (len - 2))*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = dst[k] + m[4] * src[idxp1 + k]; + dst[k + cn] = dst[k + cn] + m[3] * src[idxp1 + k] + m[4] * src[idxp2 + k]; + } + } + } +} +template <> +void hlineSmooth5N(const uint8_t* src, int cn, const ufixedpoint16* m, int, ufixedpoint16* dst, int len, int borderType) +{ + if (len == 1) + { + ufixedpoint16 msum = borderType != BORDER_CONSTANT ? m[0] + m[1] + m[2] + m[3] + m[4] : m[2]; + for (int k = 0; k < cn; k++) + dst[k] = msum * src[k]; + } + else if (len == 2) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + { + dst[k] = m[2] * src[k] + m[3] * src[k + cn]; + dst[k + cn] = m[1] * src[k] + m[2] * src[k + cn]; + } + else + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + int idxp1 = borderInterpolate(2, len, borderType)*cn; + int idxp2 = borderInterpolate(3, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = m[1] * src[k + idxm1] + m[2] * src[k] + m[3] * src[k + cn] + m[4] * src[k + idxp1] + m[0] * src[k + idxm2]; + dst[k + cn] = m[0] * src[k + idxm1] + m[1] * src[k] + m[2] * src[k + cn] + m[3] * src[k + idxp1] + m[4] * src[k + idxp2]; + } + } + } + else if (len == 3) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + { + dst[k] = m[2] * src[k] + m[3] * src[k + cn] + m[4] * src[k + 2 * cn]; + dst[k + cn] = m[1] * src[k] + m[2] * src[k + cn] + m[3] * src[k + 2 * cn]; + dst[k + 2 * cn] = m[0] * src[k] + m[1] * src[k + cn] + m[2] * src[k + 2 * cn]; + } + else + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + int idxp1 = borderInterpolate(3, len, borderType)*cn; + int idxp2 = borderInterpolate(4, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = m[2] * src[k] + m[3] * src[k + cn] + m[4] * src[k + 2 * cn] + m[0] * src[k + idxm2] + m[1] * src[k + idxm1]; + dst[k + cn] = m[1] * src[k] + m[2] * src[k + cn] + m[3] * src[k + 2 * cn] + m[0] * src[k + idxm1] + m[4] * src[k + idxp1]; + dst[k + 2 * cn] = m[0] * src[k] + m[1] * src[k + cn] + m[2] * src[k + 2 * cn] + m[3] * src[k + idxp1] + m[4] * src[k + idxp2]; + } + } + } + else + { + // Points that fall left from border + for (int k = 0; k < cn; k++) + { + dst[k] = m[2] * src[k] + m[3] * src[cn + k] + m[4] * src[2 * cn + k]; + dst[k + cn] = m[1] * src[k] + m[2] * src[cn + k] + m[3] * src[2 * cn + k] + m[4] * src[3 * cn + k]; + } + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = dst[k] + m[0] * src[idxm2 + k] + m[1] * src[idxm1 + k]; + dst[k + cn] = dst[k + cn] + m[0] * src[idxm1 + k]; + } + } + + src += 2 * cn; dst += 2 * cn; + int i = 2*cn, lencn = (len - 2)*cn; + v_int16x8 v_mul01 = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)m))); + v_int16x8 v_mul23 = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)(m + 2)))); + v_int16x8 v_mul4 = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + 4)))); + for (; i < lencn - 15; i += 16, src += 16, dst += 16) + { + v_uint16x8 v_src00, v_src01, v_src10, v_src11; + v_int16x8 v_tmp0, v_tmp1; + + v_expand(v_load(src - 2*cn), v_src00, v_src01); + v_expand(v_load(src - cn), v_src10, v_src11); + v_zip(v_reinterpret_as_s16(v_src00), v_reinterpret_as_s16(v_src10), v_tmp0, v_tmp1); + v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul01); + v_zip(v_reinterpret_as_s16(v_src01), v_reinterpret_as_s16(v_src11), v_tmp0, v_tmp1); + v_int32x4 v_res2 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res3 = v_dotprod(v_tmp1, v_mul01); + + + v_expand(v_load(src), v_src00, v_src01); + v_expand(v_load(src + cn), v_src10, v_src11); + v_zip(v_reinterpret_as_s16(v_src00), v_reinterpret_as_s16(v_src10), v_tmp0, v_tmp1); + v_res0 += v_dotprod(v_tmp0, v_mul23); + v_res1 += v_dotprod(v_tmp1, v_mul23); + v_zip(v_reinterpret_as_s16(v_src01), v_reinterpret_as_s16(v_src11), v_tmp0, v_tmp1); + v_res2 += v_dotprod(v_tmp0, v_mul23); + v_res3 += v_dotprod(v_tmp1, v_mul23); + + v_int32x4 v_resj0, v_resj1, v_resj2, v_resj3; + v_expand(v_load(src + 2*cn), v_src00, v_src01); + v_mul_expand(v_reinterpret_as_s16(v_src00), v_mul4, v_resj0, v_resj1); + v_mul_expand(v_reinterpret_as_s16(v_src01), v_mul4, v_resj2, v_resj3); + v_res0 += v_resj0; + v_res1 += v_resj1; + v_res2 += v_resj2; + v_res3 += v_resj3; + + v_store((uint16_t*)dst, v_pack(v_reinterpret_as_u32(v_res0), v_reinterpret_as_u32(v_res1))); + v_store((uint16_t*)dst + 8, v_pack(v_reinterpret_as_u32(v_res2), v_reinterpret_as_u32(v_res3))); + } + 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]; + + // Points that fall right from border + for (int k = 0; k < cn; k++) + { + dst[k] = m[0] * src[k - 2 * cn] + m[1] * src[k - cn] + m[2] * src[k] + m[3] * src[k + cn]; + dst[k + cn] = m[0] * src[k - cn] + m[1] * src[k] + m[2] * src[k + cn]; + } + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int idxp1 = (borderInterpolate(len, len, borderType) - (len - 2))*cn; + int idxp2 = (borderInterpolate(len + 1, len, borderType) - (len - 2))*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = dst[k] + m[4] * src[idxp1 + k]; + dst[k + cn] = dst[k + cn] + m[3] * src[idxp1 + k] + m[4] * src[idxp2 + k]; + } + } + } +} +template +void hlineSmooth5N14641(const ET* src, int cn, const FT*, int, FT* dst, int len, int borderType) +{ + if (len == 1) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + dst[k] = (FT(src[k])>>3)*3; + else + for (int k = 0; k < cn; k++) + dst[k] = src[k]; + } + else if (len == 2) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + { + dst[k] = (FT(src[k])>>4)*6 + (FT(src[k + cn])>>2); + dst[k + cn] = (FT(src[k]) >> 2) + (FT(src[k + cn])>>4)*6; + } + else + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + int idxp1 = borderInterpolate(2, len, borderType)*cn; + int idxp2 = borderInterpolate(3, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = (FT(src[k])>>4)*6 + (FT(src[k + idxm1])>>2) + (FT(src[k + cn])>>2) + (FT(src[k + idxp1])>>4) + (FT(src[k + idxm2])>>4); + dst[k + cn] = (FT(src[k + cn])>>4)*6 + (FT(src[k])>>2) + (FT(src[k + idxp1])>>2) + (FT(src[k + idxm1])>>4) + (FT(src[k + idxp2])>>4); + } + } + } + else if (len == 3) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + { + dst[k] = (FT(src[k])>>4)*6 + (FT(src[k + cn])>>2) + (FT(src[k + 2 * cn])>>4); + dst[k + cn] = (FT(src[k + cn])>>4)*6 + (FT(src[k])>>2) + (FT(src[k + 2 * cn])>>2); + dst[k + 2 * cn] = (FT(src[k + 2 * cn])>>4)*6 + (FT(src[k + cn])>>2) + (FT(src[k])>>4); + } + else + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + int idxp1 = borderInterpolate(3, len, borderType)*cn; + int idxp2 = borderInterpolate(4, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = (FT(src[k])>>4)*6 + (FT(src[k + cn])>>2) + (FT(src[k + idxm1])>>2) + (FT(src[k + 2 * cn])>>4) + (FT(src[k + idxm2])>>4); + dst[k + cn] = (FT(src[k + cn])>>4)*6 + (FT(src[k])>>2) + (FT(src[k + 2 * cn])>>2) + (FT(src[k + idxm1])>>4) + (FT(src[k + idxp1])>>4); + dst[k + 2 * cn] = (FT(src[k + 2 * cn])>>4)*6 + (FT(src[k + cn])>>2) + (FT(src[k + idxp1])>>2) + (FT(src[k])>>4) + (FT(src[k + idxp2])>>4); + } + } + } + else + { + // Points that fall left from border + for (int k = 0; k < cn; k++) + { + dst[k] = (FT(src[k])>>4)*6 + (FT(src[cn + k])>>2) + (FT(src[2 * cn + k])>>4); + dst[k + cn] = (FT(src[cn + k])>>4)*6 + (FT(src[k])>>2) + (FT(src[2 * cn + k])>>2) + (FT(src[3 * cn + k])>>4); + } + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = dst[k] + (FT(src[idxm2 + k])>>4) + (FT(src[idxm1 + k])>>2); + dst[k + cn] = dst[k + cn] + (FT(src[idxm1 + k])>>4); + } + } + + src += 2 * cn; dst += 2 * cn; + for (int i = 2 * cn; i < (len - 2)*cn; i++, src++, dst++) + *dst = (FT(src[0])>>4)*6 + (FT(src[-cn])>>2) + (FT(src[cn])>>2) + (FT(src[-2 * cn])>>4) + (FT(src[2 * cn])>>4); + + // Points that fall right from border + for (int k = 0; k < cn; k++) + { + dst[k] = (FT(src[k])>>4)*6 + (FT(src[k - cn])>>2) + (FT(src[k + cn])>>2) + (FT(src[k - 2 * cn])>>4); + dst[k + cn] = (FT(src[k + cn])>>4)*6 + (FT(src[k])>>2) + (FT(src[k - cn])>>4); + } + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int idxp1 = (borderInterpolate(len, len, borderType) - (len - 2))*cn; + int idxp2 = (borderInterpolate(len + 1, len, borderType) - (len - 2))*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = dst[k] + (FT(src[idxp1 + k])>>4); + dst[k + cn] = dst[k + cn] + (FT(src[idxp1 + k])>>2) + (FT(src[idxp2 + k])>>4); + } + } + } +} +template <> +void hlineSmooth5N14641(const uint8_t* src, int cn, const ufixedpoint16*, int, ufixedpoint16* dst, int len, int borderType) +{ + if (len == 1) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + dst[k] = (ufixedpoint16(src[k])>>3) * 3; + else + { + for (int k = 0; k < cn; k++) + dst[k] = src[k]; + } + } + else if (len == 2) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + { + dst[k] = (ufixedpoint16(src[k]) >> 4) * 6 + (ufixedpoint16(src[k + cn]) >> 2); + dst[k + cn] = (ufixedpoint16(src[k]) >> 2) + (ufixedpoint16(src[k + cn]) >> 4) * 6; + } + else + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + int idxp1 = borderInterpolate(2, len, borderType)*cn; + int idxp2 = borderInterpolate(3, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = (ufixedpoint16(src[k]) >> 4) * 6 + (ufixedpoint16(src[k + idxm1]) >> 2) + (ufixedpoint16(src[k + cn]) >> 2) + (ufixedpoint16(src[k + idxp1]) >> 4) + (ufixedpoint16(src[k + idxm2]) >> 4); + dst[k + cn] = (ufixedpoint16(src[k + cn]) >> 4) * 6 + (ufixedpoint16(src[k]) >> 2) + (ufixedpoint16(src[k + idxp1]) >> 2) + (ufixedpoint16(src[k + idxm1]) >> 4) + (ufixedpoint16(src[k + idxp2]) >> 4); + } + } + } + else if (len == 3) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + { + dst[k] = (ufixedpoint16(src[k]) >> 4) * 6 + (ufixedpoint16(src[k + cn]) >> 2) + (ufixedpoint16(src[k + 2 * cn]) >> 4); + dst[k + cn] = (ufixedpoint16(src[k + cn]) >> 4) * 6 + (ufixedpoint16(src[k]) >> 2) + (ufixedpoint16(src[k + 2 * cn]) >> 2); + dst[k + 2 * cn] = (ufixedpoint16(src[k + 2 * cn]) >> 4) * 6 + (ufixedpoint16(src[k + cn]) >> 2) + (ufixedpoint16(src[k]) >> 4); + } + else + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + int idxp1 = borderInterpolate(3, len, borderType)*cn; + int idxp2 = borderInterpolate(4, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = (ufixedpoint16(src[k]) >> 4) * 6 + (ufixedpoint16(src[k + cn]) >> 2) + (ufixedpoint16(src[k + idxm1]) >> 2) + (ufixedpoint16(src[k + 2 * cn]) >> 4) + (ufixedpoint16(src[k + idxm2]) >> 4); + dst[k + cn] = (ufixedpoint16(src[k + cn]) >> 4) * 6 + (ufixedpoint16(src[k]) >> 2) + (ufixedpoint16(src[k + 2 * cn]) >> 2) + (ufixedpoint16(src[k + idxm1]) >> 4) + (ufixedpoint16(src[k + idxp1]) >> 4); + dst[k + 2 * cn] = (ufixedpoint16(src[k + 2 * cn]) >> 4) * 6 + (ufixedpoint16(src[k + cn]) >> 2) + (ufixedpoint16(src[k + idxp1]) >> 2) + (ufixedpoint16(src[k]) >> 4) + (ufixedpoint16(src[k + idxp2]) >> 4); + } + } + } + else + { + // Points that fall left from border + for (int k = 0; k < cn; k++) + { + dst[k] = (ufixedpoint16(src[k]) >> 4) * 6 + (ufixedpoint16(src[cn + k]) >> 2) + (ufixedpoint16(src[2 * cn + k]) >> 4); + dst[k + cn] = (ufixedpoint16(src[cn + k]) >> 4) * 6 + (ufixedpoint16(src[k]) >> 2) + (ufixedpoint16(src[2 * cn + k]) >> 2) + (ufixedpoint16(src[3 * cn + k]) >> 4); + } + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int idxm2 = borderInterpolate(-2, len, borderType)*cn; + int idxm1 = borderInterpolate(-1, len, borderType)*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = dst[k] + (ufixedpoint16(src[idxm2 + k]) >> 4) + (ufixedpoint16(src[idxm1 + k]) >> 2); + dst[k + cn] = dst[k + cn] + (ufixedpoint16(src[idxm1 + k]) >> 4); + } + } + + src += 2 * cn; dst += 2 * cn; + int i = 2 * cn, lencn = (len - 2)*cn; + v_uint16x8 v_6 = v_setall_u16(6); + for (; i < lencn - 15; i += 16, src += 16, dst += 16) + { + v_uint16x8 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21, v_src30, v_src31, v_src40, v_src41; + v_expand(v_load(src - 2*cn), v_src00, v_src01); + v_expand(v_load(src - cn), v_src10, v_src11); + 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++) + *((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; + + // Points that fall right from border + for (int k = 0; k < cn; k++) + { + dst[k] = (ufixedpoint16(src[k]) >> 4) * 6 + (ufixedpoint16(src[k - cn]) >> 2) + (ufixedpoint16(src[k + cn]) >> 2) + (ufixedpoint16(src[k - 2 * cn]) >> 4); + dst[k + cn] = (ufixedpoint16(src[k + cn]) >> 4) * 6 + (ufixedpoint16(src[k]) >> 2) + (ufixedpoint16(src[k - cn]) >> 4); + } + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int idxp1 = (borderInterpolate(len, len, borderType) - (len - 2))*cn; + int idxp2 = (borderInterpolate(len + 1, len, borderType) - (len - 2))*cn; + for (int k = 0; k < cn; k++) + { + dst[k] = dst[k] + (ufixedpoint16(src[idxp1 + k]) >> 4); + dst[k + cn] = dst[k + cn] + (ufixedpoint16(src[idxp1 + k]) >> 2) + (ufixedpoint16(src[idxp2 + k]) >> 4); + } + } + } +} +template +void hlineSmooth(const ET* src, int cn, const FT* m, int n, FT* dst, int len, int borderType) +{ + int pre_shift = n / 2; + int post_shift = n - pre_shift; + int i = 0; + for (; i < min(pre_shift, len); i++, dst += cn) // Points that fall left from border + { + for (int k = 0; k < cn; k++) + dst[k] = m[pre_shift-i] * src[k]; + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + for (int j = i - pre_shift, mid = 0; j < 0; j++, mid++) + { + int src_idx = borderInterpolate(j, len, borderType); + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[mid] * src[src_idx*cn + k]; + } + int j, mid; + for (j = 1, mid = pre_shift - i + 1; j < min(i + post_shift, len); j++, mid++) + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[mid] * src[j*cn + k]; + if (borderType != BORDER_CONSTANT) + for (; j < i + post_shift; j++, mid++) + { + int src_idx = borderInterpolate(j, len, borderType); + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[mid] * src[src_idx*cn + k]; + } + } + i *= cn; + for (; i < (len - post_shift + 1)*cn; i++, src++, dst++) + { + *dst = m[0] * src[0]; + for (int j = 1; j < n; j++) + *dst = *dst + m[j] * src[j*cn]; + } + i /= cn; + for (i -= pre_shift; i < len - pre_shift; i++, src += cn, dst += cn) // Points that fall right from border + { + for (int k = 0; k < cn; k++) + dst[k] = m[0] * src[k]; + int j = 1; + for (; j < len - i; j++) + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[j] * src[j*cn + k]; + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + for (; j < n; j++) + { + int src_idx = borderInterpolate(i + j, len, borderType) - i; + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[j] * src[src_idx*cn + k]; + } + } +} +template <> +void hlineSmooth(const uint8_t* src, int cn, const ufixedpoint16* m, int n, ufixedpoint16* dst, int len, int borderType) +{ + int pre_shift = n / 2; + int post_shift = n - pre_shift; + int i = 0; + for (; i < min(pre_shift, len); i++, dst += cn) // Points that fall left from border + { + for (int k = 0; k < cn; k++) + dst[k] = m[pre_shift - i] * src[k]; + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + for (int j = i - pre_shift, mid = 0; j < 0; j++, mid++) + { + int src_idx = borderInterpolate(j, len, borderType); + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[mid] * src[src_idx*cn + k]; + } + int j, mid; + for (j = 1, mid = pre_shift - i + 1; j < min(i + post_shift, len); j++, mid++) + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[mid] * src[j*cn + k]; + if (borderType != BORDER_CONSTANT) + for (; j < i + post_shift; j++, mid++) + { + int src_idx = borderInterpolate(j, len, borderType); + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[mid] * src[src_idx*cn + k]; + } + } + i *= cn; + int lencn = (len - post_shift + 1)*cn; + for (; i < lencn - 15; i+=16, src+=16, dst+=16) + { + v_uint16x8 v_src00, v_src01, v_src10, v_src11; + v_int16x8 v_tmp0, v_tmp1; + + v_int16x8 v_mul = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)m))); + + v_expand(v_load(src), v_src00, v_src01); + v_expand(v_load(src+cn), v_src10, v_src11); + v_zip(v_reinterpret_as_s16(v_src00), v_reinterpret_as_s16(v_src10), v_tmp0, v_tmp1); + v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul); + v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul); + v_zip(v_reinterpret_as_s16(v_src01), v_reinterpret_as_s16(v_src11), v_tmp0, v_tmp1); + v_int32x4 v_res2 = v_dotprod(v_tmp0, v_mul); + v_int32x4 v_res3 = v_dotprod(v_tmp1, v_mul); + + int j = 2; + for (; j < n - 1; j += 2) + { + v_mul = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)(m + j)))); + + v_expand(v_load(src + j * cn), v_src00, v_src01); + v_expand(v_load(src + (j + 1) * cn), v_src10, v_src11); + v_zip(v_reinterpret_as_s16(v_src00), v_reinterpret_as_s16(v_src10), v_tmp0, v_tmp1); + v_res0 += v_dotprod(v_tmp0, v_mul); + v_res1 += v_dotprod(v_tmp1, v_mul); + v_zip(v_reinterpret_as_s16(v_src01), v_reinterpret_as_s16(v_src11), v_tmp0, v_tmp1); + v_res2 += v_dotprod(v_tmp0, v_mul); + v_res3 += v_dotprod(v_tmp1, v_mul); + } + if (j < n) + { + v_int32x4 v_resj0, v_resj1, v_resj2, v_resj3; + v_mul = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + j)))); + v_expand(v_load(src + j * cn), v_src00, v_src01); + v_mul_expand(v_reinterpret_as_s16(v_src00), v_mul, v_resj0, v_resj1); + v_mul_expand(v_reinterpret_as_s16(v_src01), v_mul, v_resj2, v_resj3); + v_res0 += v_resj0; + v_res1 += v_resj1; + v_res2 += v_resj2; + v_res3 += v_resj3; + } + + v_store((uint16_t*)dst, v_pack(v_reinterpret_as_u32(v_res0), v_reinterpret_as_u32(v_res1))); + v_store((uint16_t*)dst+8, v_pack(v_reinterpret_as_u32(v_res2), v_reinterpret_as_u32(v_res3))); + } + for (; i < lencn; i++, src++, dst++) + { + *dst = m[0] * src[0]; + for (int j = 1; j < n; j++) + *dst = *dst + m[j] * src[j*cn]; + } + i /= cn; + for (i -= pre_shift; i < len - pre_shift; i++, src += cn, dst += cn) // Points that fall right from border + { + for (int k = 0; k < cn; k++) + dst[k] = m[0] * src[k]; + int j = 1; + for (; j < len - i; j++) + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[j] * src[j*cn + k]; + if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + for (; j < n; j++) + { + int src_idx = borderInterpolate(i + j, len, borderType) - i; + for (int k = 0; k < cn; k++) + dst[k] = dst[k] + m[j] * src[src_idx*cn + k]; + } + } +} +template +void vlineSmooth1N(const FT* const * src, const FT* m, int, ET* dst, int len) +{ + const FT* src0 = src[0]; + for (int i = 0; i < len; i++) + dst[i] = m * src0[i]; +} +template <> +void vlineSmooth1N(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len) +{ + const ufixedpoint16* src0 = src[0]; + v_uint16x8 v_mul = v_setall_u16(*((uint16_t*)m)); + int i = 0; + for (; i < len - 7; i += 8) + { + v_uint16x8 v_src0 = v_load((uint16_t*)src0 + i); + v_uint32x4 v_res0, v_res1; + v_mul_expand(v_src0, v_mul, v_res0, v_res1); + v_pack_store(dst + i, v_rshr_pack<16>(v_res0, v_res1)); + } + for (; i < len; i++) + dst[i] = m[0] * src0[i]; +} +template +void vlineSmooth1N1(const FT* const * src, const FT*, int, ET* dst, int len) +{ + const FT* src0 = src[0]; + for (int i = 0; i < len; i++) + dst[i] = src0[i]; +} +template <> +void vlineSmooth1N1(const ufixedpoint16* const * src, const ufixedpoint16*, int, uint8_t* dst, int len) +{ + const ufixedpoint16* src0 = src[0]; + int i = 0; + for (; i < len - 7; i += 8) + v_rshr_pack_store<8>(dst + i, v_load((uint16_t*)(src0 + i))); + for (; i < len; i++) + dst[i] = src0[i]; +} +template +void vlineSmooth3N(const FT* const * src, const FT* m, int, ET* dst, int len) +{ + for (int i = 0; i < len; i++) + dst[i] = m[0] * src[0][i] + m[1] * src[1][i] + m[2] * src[2][i]; +} +template <> +void vlineSmooth3N(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len) +{ + static const v_int16x8 v_128 = v_reinterpret_as_s16(v_setall_u16((uint16_t)1 << 15)); + + v_int32x4 v_128_4 = v_setall_s32(128 << 16); + if (len > 7) + { + ufixedpoint32 val[] = { (m[0] + m[1] + m[2]) * ufixedpoint16((uint8_t)128) }; + v_128_4 = v_setall_s32(*((int32_t*)val)); + } + + int i = 0; + v_int16x8 v_mul01 = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)m))); + v_int16x8 v_mul2 = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + 2)))); + for (; i < len - 7; i += 8) + { + v_int16x8 v_src0, v_src1; + v_int16x8 v_tmp0, v_tmp1; + + v_src0 = v_load((int16_t*)(src[0]) + i); + v_src1 = v_load((int16_t*)(src[1]) + i); + v_zip(v_add_wrap(v_src0, v_128), v_add_wrap(v_src1, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul01); + + v_int32x4 v_resj0, v_resj1; + v_src0 = v_load((int16_t*)(src[2]) + i); + v_mul_expand(v_add_wrap(v_src0, v_128), v_mul2, v_resj0, v_resj1); + v_res0 += v_resj0; + v_res1 += v_resj1; + + v_res0 += v_128_4; + v_res1 += v_128_4; + + v_uint16x8 v_res = v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)); + v_pack_store(dst + i, v_res); + } + for (; i < len; i++) + dst[i] = m[0] * src[0][i] + m[1] * src[1][i] + m[2] * src[2][i]; +} +template +void vlineSmooth3N121(const FT* const * src, const FT*, int, ET* dst, int len) +{ + for (int i = 0; i < len; i++) + dst[i] = ((FT::WT(src[0][i]) + FT::WT(src[2][i])) >> 2) + (FT::WT(src[1][i]) >> 1); +} +template <> +void vlineSmooth3N121(const ufixedpoint16* const * src, const ufixedpoint16*, int, uint8_t* dst, int len) +{ + int i = 0; + for (; i < len - 7; i += 8) + { + v_uint32x4 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21; + v_expand(v_load((uint16_t*)(src[0]) + i), v_src00, v_src01); + v_expand(v_load((uint16_t*)(src[1]) + i), v_src10, v_src11); + v_expand(v_load((uint16_t*)(src[2]) + i), v_src20, v_src21); + v_uint16x8 v_res = v_rshr_pack<10>(v_src00 + v_src20 + (v_src10 << 1), v_src01 + v_src21 + (v_src11 << 1)); + v_pack_store(dst + i, v_res); + } + 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; +} +template +void vlineSmooth5N(const FT* const * src, const FT* m, int, ET* dst, int len) +{ + for (int i = 0; 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]; +} +template <> +void vlineSmooth5N(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len) +{ + static const v_int16x8 v_128 = v_reinterpret_as_s16(v_setall_u16((uint16_t)1 << 15)); + + v_int32x4 v_128_4 = v_setall_s32(128 << 16); + if (len > 7) + { + 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)); + } + + int i = 0; + v_int16x8 v_mul01 = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)m))); + v_int16x8 v_mul23 = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)(m + 2)))); + v_int16x8 v_mul4 = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + 4)))); + for (; i < len - 7; i += 8) + { + v_int16x8 v_src0, v_src1; + v_int16x8 v_tmp0, v_tmp1; + + v_src0 = v_load((int16_t*)(src[0]) + i); + v_src1 = v_load((int16_t*)(src[1]) + i); + v_zip(v_add_wrap(v_src0, v_128), v_add_wrap(v_src1, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul01); + + v_src0 = v_load((int16_t*)(src[2]) + i); + v_src1 = v_load((int16_t*)(src[3]) + i); + v_zip(v_add_wrap(v_src0, v_128), v_add_wrap(v_src1, v_128), v_tmp0, v_tmp1); + v_res0 += v_dotprod(v_tmp0, v_mul23); + v_res1 += v_dotprod(v_tmp1, v_mul23); + + v_int32x4 v_resj0, v_resj1; + v_src0 = v_load((int16_t*)(src[4]) + i); + v_mul_expand(v_add_wrap(v_src0, v_128), v_mul4, v_resj0, v_resj1); + v_res0 += v_resj0; + v_res1 += v_resj1; + + v_res0 += v_128_4; + v_res1 += v_128_4; + + v_uint16x8 v_res = v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)); + v_pack_store(dst + i, v_res); + } + 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]; +} +template +void vlineSmooth5N14641(const FT* const * src, const FT*, int, ET* dst, int len) +{ + for (int i = 0; i < len; i++) + dst[i] = (FT::WT(src[2][i])*6 + ((FT::WT(src[1][i]) + FT::WT(src[3][i]))<<2) + FT::WT(src[0][i]) + FT::WT(src[4][i])) >> 4; +} +template <> +void vlineSmooth5N14641(const ufixedpoint16* const * src, const ufixedpoint16*, int, uint8_t* dst, int len) +{ + int i = 0; + v_uint32x4 v_6 = v_setall_u32(6); + for (; i < len - 7; i += 8) + { + v_uint32x4 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21, v_src30, v_src31, v_src40, v_src41; + v_expand(v_load((uint16_t*)(src[0]) + i), v_src00, v_src01); + v_expand(v_load((uint16_t*)(src[1]) + i), v_src10, v_src11); + v_expand(v_load((uint16_t*)(src[2]) + i), v_src20, v_src21); + v_expand(v_load((uint16_t*)(src[3]) + i), v_src30, v_src31); + v_expand(v_load((uint16_t*)(src[4]) + i), v_src40, v_src41); + v_uint16x8 v_res = 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_pack_store(dst + i, v_res); + } + for (; i < len; i++) + 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[0]))[i]) + (uint32_t)(((uint16_t*)(src[4]))[i]) + (1 << 11)) >> 12; +} +template +void vlineSmooth(const FT* const * src, const FT* m, int n, ET* dst, int len) +{ + for (int i = 0; i < len; i++) + { + typename FT::WT val = m[0] * src[0][i]; + for (int j = 1; j < n; j++) + val = val + m[j] * src[j][i]; + dst[i] = val; + } +} +template <> +void vlineSmooth(const ufixedpoint16* const * src, const ufixedpoint16* m, int n, uint8_t* dst, int len) +{ + static const v_int16x8 v_128 = v_reinterpret_as_s16(v_setall_u16((uint16_t)1 << 15)); + + v_int32x4 v_128_4 = v_setall_s32(128 << 16); + if (len > 7) + { + ufixedpoint16 msum = m[0] + m[1]; + for (int j = 2; j < n; j++) + msum = msum + m[j]; + ufixedpoint32 val[] = { msum * ufixedpoint16((uint8_t)128) }; + v_128_4 = v_setall_s32(*((int32_t*)val)); + } + + int i = 0; + for (; i < len - 7; i += 8) + { + v_int16x8 v_src0, v_src1; + v_int16x8 v_tmp0, v_tmp1; + + v_int16x8 v_mul = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)m))); + + v_src0 = v_load((int16_t*)(src[0]) + i); + v_src1 = v_load((int16_t*)(src[1]) + i); + v_zip(v_add_wrap(v_src0, v_128), v_add_wrap(v_src1, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul); + v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul); + + int j = 2; + for (; j < n - 1; j+=2) + { + v_mul = v_reinterpret_as_s16(v_setall_u32(*((uint32_t*)(m+j)))); + + v_src0 = v_load((int16_t*)(src[j]) + i); + v_src1 = v_load((int16_t*)(src[j+1]) + i); + v_zip(v_add_wrap(v_src0, v_128), v_add_wrap(v_src1, v_128), v_tmp0, v_tmp1); + v_res0 += v_dotprod(v_tmp0, v_mul); + v_res1 += v_dotprod(v_tmp1, v_mul); + } + if(j < n) + { + v_int32x4 v_resj0, v_resj1; + v_mul = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + j)))); + v_src0 = v_load((int16_t*)(src[j]) + i); + v_mul_expand(v_add_wrap(v_src0, v_128), v_mul, v_resj0, v_resj1); + v_res0 += v_resj0; + v_res1 += v_resj1; + } + v_res0 += v_128_4; + v_res1 += v_128_4; + + v_uint16x8 v_res = v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)); + v_pack_store(dst + i, v_res); + } + for (; i < len; i++) + { + ufixedpoint32 val = m[0] * src[0][i]; + for (int j = 1; j < n; j++) + { + val = val + m[j] * src[j][i]; + } + dst[i] = val; + } +} +template +class fixedSmoothInvoker : public ParallelLoopBody +{ +public: + fixedSmoothInvoker(const ET* _src, size_t _src_stride, ET* _dst, size_t _dst_stride, + int _width, int _height, int _cn, const FT* _kx, int _kxlen, const FT* _ky, int _kylen, int _borderType) : ParallelLoopBody(), + src(_src), dst(_dst), src_stride(_src_stride), dst_stride(_dst_stride), + width(_width), height(_height), cn(_cn), kx(_kx), ky(_ky), kxlen(_kxlen), kylen(_kylen), borderType(_borderType) + { + if (kxlen == 1) + { + if ((kx[0] - FT::one()).isZero()) + hlineSmoothFunc = hlineSmooth1N1; + else + hlineSmoothFunc = hlineSmooth1N; + } + else if (kxlen == 3) + { + if ((kx[0] - (FT::one()>>2)).isZero()&&(kx[1] - (FT::one()>>1)).isZero()&&(kx[2] - (FT::one()>>2)).isZero()) + hlineSmoothFunc = hlineSmooth3N121; + else + hlineSmoothFunc = hlineSmooth3N; + } + else if (kxlen == 5) + { + if ((kx[2] - (FT::one()*3>>3)).isZero()&& + (kx[1] - (FT::one()>>2)).isZero()&&(kx[3] - (FT::one()>>2)).isZero()&& + (kx[0] - (FT::one()>>4)).isZero()&&(kx[4] - (FT::one()>>4)).isZero()) + hlineSmoothFunc = hlineSmooth5N14641; + else + hlineSmoothFunc = hlineSmooth5N; + } + else + hlineSmoothFunc = hlineSmooth; + if (kylen == 1) + { + if ((ky[0] - FT::one()).isZero()) + vlineSmoothFunc = vlineSmooth1N1; + else + vlineSmoothFunc = vlineSmooth1N; + } + else if (kylen == 3) + { + if ((ky[0] - (FT::one() >> 2)).isZero() && (ky[1] - (FT::one() >> 1)).isZero() && (ky[2] - (FT::one() >> 2)).isZero()) + vlineSmoothFunc = vlineSmooth3N121; + else + vlineSmoothFunc = vlineSmooth3N; + } + else if (kylen == 5) + { + if ((ky[2] - (FT::one() * 3 >> 3)).isZero() && + (ky[1] - (FT::one() >> 2)).isZero() && (ky[3] - (FT::one() >> 2)).isZero() && + (ky[0] - (FT::one() >> 4)).isZero() && (ky[4] - (FT::one() >> 4)).isZero()) + vlineSmoothFunc = vlineSmooth5N14641; + else + vlineSmoothFunc = vlineSmooth5N; + } + else + vlineSmoothFunc = vlineSmooth; + } + virtual void operator() (const Range& range) const + { + AutoBuffer _buf(width*cn*kylen); + FT* buf = _buf; + AutoBuffer _ptrs(kylen*2); + FT** ptrs = _ptrs; + + if (kylen == 1) + { + ptrs[0] = buf; + for (int i = range.start; i < range.end; i++) + { + hlineSmoothFunc(src + i * src_stride, cn, kx, kxlen, ptrs[0], width, borderType); + vlineSmoothFunc(ptrs, ky, kylen, dst + i * dst_stride, width*cn); + } + } + else if (borderType != BORDER_CONSTANT)// If BORDER_CONSTANT out of border values are equal to zero and could be skipped + { + int pre_shift = kylen / 2; + int post_shift = kylen - pre_shift - 1; + // First line evaluation + int idst = range.start; + int ifrom = max(0, idst - pre_shift); + int ito = idst + post_shift + 1; + int i = ifrom; + int bufline = 0; + for (; i < min(ito, height); i++, bufline++) + { + ptrs[bufline+kylen] = ptrs[bufline] = buf + bufline * width*cn; + hlineSmoothFunc(src + i * src_stride, cn, kx, kxlen, ptrs[bufline], width, borderType); + } + for (; i < ito; i++, bufline++) + { + int src_idx = borderInterpolate(i, height, borderType); + if (src_idx < ifrom) + { + ptrs[bufline + kylen] = ptrs[bufline] = buf + bufline * width*cn; + hlineSmoothFunc(src + src_idx * src_stride, cn, kx, kxlen, ptrs[bufline], width, borderType); + } + else + { + ptrs[bufline + kylen] = ptrs[bufline] = ptrs[src_idx - ifrom]; + } + } + for (int j = idst - pre_shift; j < 0; j++) + { + int src_idx = borderInterpolate(j, height, borderType); + if (src_idx >= ito) + { + ptrs[2*kylen + j] = ptrs[kylen + j] = buf + (kylen + j) * width*cn; + hlineSmoothFunc(src + src_idx * src_stride, cn, kx, kxlen, ptrs[kylen + j], width, borderType); + } + else + { + ptrs[2*kylen + j] = ptrs[kylen + j] = ptrs[src_idx]; + } + } + vlineSmoothFunc(ptrs + bufline, ky, kylen, dst + idst*dst_stride, width*cn); idst++; + + // border mode dependent part evaluation + // i points to last src row to evaluate in convolution + bufline %= kylen; ito = min(height, range.end + post_shift); + for (; i < min(kylen, ito); i++, idst++) + { + ptrs[bufline + kylen] = ptrs[bufline] = buf + bufline * width*cn; + hlineSmoothFunc(src + i * src_stride, cn, kx, kxlen, ptrs[bufline], width, borderType); + bufline = (bufline + 1) % kylen; + vlineSmoothFunc(ptrs + bufline, ky, kylen, dst + idst*dst_stride, width*cn); + } + // Points inside the border + for (; i < ito; i++, idst++) + { + hlineSmoothFunc(src + i * src_stride, cn, kx, kxlen, ptrs[bufline], width, borderType); + bufline = (bufline + 1) % kylen; + vlineSmoothFunc(ptrs + bufline, ky, kylen, dst + idst*dst_stride, width*cn); + } + // Points that could fall below border + for (; i < range.end + post_shift; i++, idst++) + { + int src_idx = borderInterpolate(i, height, borderType); + if ((i - src_idx) > kylen) + hlineSmoothFunc(src + src_idx * src_stride, cn, kx, kxlen, ptrs[bufline], width, borderType); + else + ptrs[bufline + kylen] = ptrs[bufline] = ptrs[(bufline + kylen - (i - src_idx)) % kylen]; + bufline = (bufline + 1) % kylen; + vlineSmoothFunc(ptrs + bufline, ky, kylen, dst + idst*dst_stride, width*cn); + } + } + else + { + int pre_shift = kylen / 2; + int post_shift = kylen - pre_shift - 1; + // First line evaluation + int idst = range.start; + int ifrom = idst - pre_shift; + int ito = min(idst + post_shift + 1, height); + int i = max(0, ifrom); + int bufline = 0; + for (; i < ito; i++, bufline++) + { + ptrs[bufline + kylen] = ptrs[bufline] = buf + bufline * width*cn; + hlineSmoothFunc(src + i * src_stride, cn, kx, kxlen, ptrs[bufline], width, borderType); + } + + if (bufline == 1) + vlineSmooth1N(ptrs, ky - min(ifrom, 0), bufline, dst + idst*dst_stride, width*cn); + else if (bufline == 3) + vlineSmooth3N(ptrs, ky - min(ifrom, 0), bufline, dst + idst*dst_stride, width*cn); + else if (bufline == 5) + vlineSmooth5N(ptrs, ky - min(ifrom, 0), bufline, dst + idst*dst_stride, width*cn); + else + vlineSmooth(ptrs, ky - min(ifrom, 0), bufline, dst + idst*dst_stride, width*cn); + idst++; + + // border mode dependent part evaluation + // i points to last src row to evaluate in convolution + bufline %= kylen; ito = min(height, range.end + post_shift); + for (; i < min(kylen, ito); i++, idst++) + { + ptrs[bufline + kylen] = ptrs[bufline] = buf + bufline * width*cn; + hlineSmoothFunc(src + i * src_stride, cn, kx, kxlen, ptrs[bufline], width, borderType); + bufline++; + if (bufline == 3) + vlineSmooth3N(ptrs, ky + kylen - bufline, i + 1, dst + idst*dst_stride, width*cn); + else if (bufline == 5) + vlineSmooth5N(ptrs, ky + kylen - bufline, i + 1, dst + idst*dst_stride, width*cn); + else + vlineSmooth(ptrs, ky + kylen - bufline, i + 1, dst + idst*dst_stride, width*cn); + bufline %= kylen; + } + // Points inside the border + if (i - max(0, ifrom) >= kylen) + { + for (; i < ito; i++, idst++) + { + hlineSmoothFunc(src + i * src_stride, cn, kx, kxlen, ptrs[bufline], width, borderType); + bufline = (bufline + 1) % kylen; + vlineSmoothFunc(ptrs + bufline, ky, kylen, dst + idst*dst_stride, width*cn); + } + + // Points that could fall below border + // i points to first src row to evaluate in convolution + bufline = (bufline + 1) % kylen; + for (i = idst - pre_shift; i < range.end - pre_shift; i++, idst++, bufline++) + if (height - i == 3) + vlineSmooth3N(ptrs + bufline, ky, height - i, dst + idst*dst_stride, width*cn); + else if (height - i == 5) + vlineSmooth5N(ptrs + bufline, ky, height - i, dst + idst*dst_stride, width*cn); + else + vlineSmooth(ptrs + bufline, ky, height - i, dst + idst*dst_stride, width*cn); + } + else + { + // i points to first src row to evaluate in convolution + for (i = idst - pre_shift; i < min(range.end - pre_shift, 0); i++, idst++) + if (height == 3) + vlineSmooth3N(ptrs, ky - i, height, dst + idst*dst_stride, width*cn); + else if (height == 5) + vlineSmooth5N(ptrs, ky - i, height, dst + idst*dst_stride, width*cn); + else + vlineSmooth(ptrs, ky - i, height, dst + idst*dst_stride, width*cn); + for (; i < range.end - pre_shift; i++, idst++) + if (height - i == 3) + vlineSmooth3N(ptrs + i - max(0, ifrom), ky, height - i, dst + idst*dst_stride, width*cn); + else if (height - i == 5) + vlineSmooth5N(ptrs + i - max(0, ifrom), ky, height - i, dst + idst*dst_stride, width*cn); + else + vlineSmooth(ptrs + i - max(0, ifrom), ky, height - i, dst + idst*dst_stride, width*cn); + } + } + } +private: + const ET* src; + ET* dst; + size_t src_stride, dst_stride; + int width, height, cn; + const FT *kx, *ky; + int kxlen, kylen; + int borderType; + void(*hlineSmoothFunc)(const ET* src, int cn, const FT* m, int n, FT* dst, int len, int borderType); + void(*vlineSmoothFunc)(const FT* const * src, const FT* m, int n, ET* dst, int len); + + fixedSmoothInvoker(const fixedSmoothInvoker&); + fixedSmoothInvoker& operator=(const fixedSmoothInvoker&); +}; + +static void getGaussianKernel(int n, double sigma, int ktype, Mat& res) { res = getGaussianKernel(n, sigma, ktype); } +template static void getGaussianKernel(int n, double sigma, int, std::vector& res) { res = getFixedpointGaussianKernel(n, sigma); } + +template +static void createGaussianKernels( T & kx, T & ky, int type, Size &ksize, double sigma1, double sigma2 ) { int depth = CV_MAT_DEPTH(type); @@ -1782,11 +3188,11 @@ static void createGaussianKernels( Mat & kx, Mat & ky, int type, Size & ksize, sigma1 = std::max( sigma1, 0. ); sigma2 = std::max( sigma2, 0. ); - kx = getGaussianKernel( ksize.width, sigma1, std::max(depth, CV_32F) ); + getGaussianKernel( ksize.width, sigma1, std::max(depth, CV_32F), kx ); if( ksize.height == ksize.width && std::abs(sigma1 - sigma2) < DBL_EPSILON ) ky = kx; else - ky = getGaussianKernel( ksize.height, sigma2, std::max(depth, CV_32F) ); + getGaussianKernel( ksize.height, sigma2, std::max(depth, CV_32F), ky ); } } @@ -2082,7 +3488,8 @@ void cv::GaussianBlur( InputArray _src, OutputArray _dst, Size ksize, Size size = _src.size(); _dst.create( size, type ); - if( borderType != BORDER_CONSTANT && (borderType & BORDER_ISOLATED) != 0 ) + if( (borderType & ~BORDER_ISOLATED) != BORDER_CONSTANT && + ((borderType & BORDER_ISOLATED) != 0 || !_src.getMat().isSubmatrix()) ) { if( size.height == 1 ) ksize.height = 1; @@ -2104,6 +3511,20 @@ void cv::GaussianBlur( InputArray _src, OutputArray _dst, Size ksize, int sdepth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type); + if(sdepth == CV_8U && ((borderType & BORDER_ISOLATED) || !_src.getMat().isSubmatrix())) + { + std::vector fkx, fky; + createGaussianKernels(fkx, fky, type, ksize, sigma1, sigma2); + Mat src = _src.getMat(); + Mat dst = _dst.getMat(); + if (src.data == dst.data) + src = src.clone(); + fixedSmoothInvoker invoker(src.ptr(), src.step1(), dst.ptr(), dst.step1(), dst.cols, dst.rows, dst.channels(), &fkx[0], (int)fkx.size(), &fky[0], (int)fky.size(), borderType & ~BORDER_ISOLATED); + parallel_for_(Range(0, dst.rows), invoker, dst.total() * cn / (double)(1 << 13)); + return; + } + + Mat kx, ky; createGaussianKernels(kx, ky, type, ksize, sigma1, sigma2); diff --git a/modules/imgproc/test/test_smooth_bitexact.cpp b/modules/imgproc/test/test_smooth_bitexact.cpp new file mode 100644 index 0000000000000000000000000000000000000000..449cea32090b7a46ceb192397883f95e410b31d3 --- /dev/null +++ b/modules/imgproc/test/test_smooth_bitexact.cpp @@ -0,0 +1,167 @@ +// This file is part of OpenCV project. +// It is subject to the license terms in the LICENSE file found in the top-level directory +// of this distribution and at http://opencv.org/license.html. + +#include "test_precomp.hpp" + +#include + +using namespace cv; +using namespace std; + +namespace +{ + static const int fixedShiftU8 = 8; + static const int64_t fixedOne = (1L << fixedShiftU8); + + int64_t v[][9] = { + { fixedOne }, // size 1, sigma 0 + { fixedOne >> 2, fixedOne >> 1, fixedOne >> 2 }, // size 3, sigma 0 + { fixedOne >> 4, fixedOne >> 2, 6 * (fixedOne >> 4), fixedOne >> 2, fixedOne >> 4 }, // size 5, sigma 0 + { fixedOne >> 5, 7 * (fixedOne >> 6), 7 * (fixedOne >> 5), 9 * (fixedOne >> 5), 7 * (fixedOne >> 5), 7 * (fixedOne >> 6), fixedOne >> 5 }, // size 7, sigma 0 + { 4, 13, 30, 51, 61, 51, 30, 13, 4 }, // size 9, sigma 0 + { 81, 95, 81 }, // size 3, sigma 1.75 + { 65, 125, 65 }, // size 3, sigma 0.875 + { 0, 7, 242, 7, 0 }, // size 5, sigma 0.375 + { 4, 56, 136, 56, 4 } // size 5, sigma 0.75 + }; + + template + T eval(Mat src, vector kernelx, vector kernely) + { + static const int64_t fixedRound = ((1LL << (fixedShift * 2)) >> 1); + int64_t val = 0; + for (size_t j = 0; j < kernely.size(); j++) + { + int64_t lineval = 0; + for (size_t i = 0; i < kernelx.size(); i++) + lineval += src.at((int)j, (int)i) * kernelx[i]; + val += lineval * kernely[j]; + } + return saturate_cast((val + fixedRound) >> (fixedShift * 2)); + } +} + +TEST(GaussianBlur_Bitexact, Linear8U) +{ + struct testmode + { + int type; + Size sz; + Size kernel; + double sigma_x; + double sigma_y; + vector kernel_x; + vector kernel_y; + } modes[] = { + { CV_8UC1, Size( 1, 1), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC1, Size( 2, 2), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC1, Size( 3, 1), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC1, Size( 1, 3), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC1, Size( 3, 3), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC1, Size( 3, 3), Size(5, 5), 0, 0, vector(v[2], v[2]+5), vector(v[2], v[2]+5) }, + { CV_8UC1, Size( 3, 3), Size(7, 7), 0, 0, vector(v[3], v[3]+7), vector(v[3], v[3]+7) }, + { CV_8UC1, Size( 5, 5), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC1, Size( 5, 5), Size(5, 5), 0, 0, vector(v[2], v[2]+5), vector(v[2], v[2]+5) }, + { CV_8UC1, Size( 3, 5), Size(5, 5), 0, 0, vector(v[2], v[2]+5), vector(v[2], v[2]+5) }, + { CV_8UC1, Size( 5, 5), Size(5, 5), 0, 0, vector(v[2], v[2]+5), vector(v[2], v[2]+5) }, + { CV_8UC1, Size( 5, 5), Size(7, 7), 0, 0, vector(v[3], v[3]+7), vector(v[3], v[3]+7) }, + { CV_8UC1, Size( 7, 7), Size(7, 7), 0, 0, vector(v[3], v[3]+7), vector(v[3], v[3]+7) }, + { CV_8UC1, Size( 256, 128), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC2, Size( 256, 128), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC3, Size( 256, 128), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC4, Size( 256, 128), Size(3, 3), 0, 0, vector(v[1], v[1]+3), vector(v[1], v[1]+3) }, + { CV_8UC1, Size( 256, 128), Size(5, 5), 0, 0, vector(v[2], v[2]+5), vector(v[2], v[2]+5) }, + { CV_8UC1, Size( 256, 128), Size(7, 7), 0, 0, vector(v[3], v[3]+7), vector(v[3], v[3]+7) }, + { CV_8UC1, Size( 256, 128), Size(9, 9), 0, 0, vector(v[4], v[4]+9), vector(v[4], v[4]+9) }, + { CV_8UC1, Size( 256, 128), Size(3, 3), 1.75, 0.875, vector(v[5], v[5]+3), vector(v[6], v[6]+3) }, + { CV_8UC2, Size( 256, 128), Size(3, 3), 1.75, 0.875, vector(v[5], v[5]+3), vector(v[6], v[6]+3) }, + { CV_8UC3, Size( 256, 128), Size(3, 3), 1.75, 0.875, vector(v[5], v[5]+3), vector(v[6], v[6]+3) }, + { CV_8UC4, Size( 256, 128), Size(3, 3), 1.75, 0.875, vector(v[5], v[5]+3), vector(v[6], v[6]+3) }, + { CV_8UC1, Size( 256, 128), Size(5, 5), 0.375, 0.75, vector(v[7], v[7]+5), vector(v[8], v[8]+5) } + }; + + int bordermodes[] = { + BORDER_CONSTANT | BORDER_ISOLATED, + BORDER_REPLICATE | BORDER_ISOLATED, + BORDER_REFLECT | BORDER_ISOLATED, + BORDER_WRAP | BORDER_ISOLATED, + BORDER_REFLECT_101 | BORDER_ISOLATED +// BORDER_CONSTANT, +// BORDER_REPLICATE, +// BORDER_REFLECT, +// BORDER_WRAP, +// BORDER_REFLECT_101 + }; + + for (int modeind = 0, _modecnt = sizeof(modes) / sizeof(modes[0]); modeind < _modecnt; ++modeind) + { + int type = modes[modeind].type, depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type); + int dcols = modes[modeind].sz.width, drows = modes[modeind].sz.height; + Size kernel = modes[modeind].kernel; + + int rows = drows + 20, cols = dcols + 20; + Mat src(rows, cols, type), refdst(drows, dcols, type), dst; + for (int j = 0; j < rows; j++) + { + uint8_t* line = src.ptr(j); + for (int i = 0; i < cols; i++) + for (int c = 0; c < cn; c++) + { + RNG rnd(0x123456789abcdefULL); + double val = j < rows / 2 ? (i < cols / 2 ? ((sin((i + 1)*CV_PI / 256.)*sin((j + 1)*CV_PI / 256.)*sin((cn + 4)*CV_PI / 8.) + 1.)*128.) : + (((i / 128 + j / 128) % 2) * 250 + (j / 128) % 2)) : + (i < cols / 2 ? ((i / 128) * (85 - j / 256 * 40) * ((j / 128) % 2) + (7 - i / 128) * (85 - j / 256 * 40) * ((j / 128 + 1) % 2)) : + ((uchar)rnd)); + if (depth == CV_8U) + line[i*cn + c] = (uint8_t)val; + else if (depth == CV_16U) + ((uint16_t*)line)[i*cn + c] = (uint16_t)val; + else if (depth == CV_16S) + ((int16_t*)line)[i*cn + c] = (int16_t)val; + else if (depth == CV_32S) + ((int32_t*)line)[i*cn + c] = (int32_t)val; + else + CV_Assert(0); + } + } + Mat src_roi = src(Rect(10, 10, dcols, drows)); + + + for (int borderind = 0, _bordercnt = sizeof(bordermodes) / sizeof(bordermodes[0]); borderind < _bordercnt; ++borderind) + { + Mat src_border; + copyMakeBorder(src_roi, src_border, kernel.height / 2, kernel.height / 2, kernel.width / 2, kernel.width / 2, bordermodes[borderind]); + for (int c = 0; c < src_border.channels(); c++) + { + int fromTo[2] = { c, 0 }; + int toFrom[2] = { 0, c }; + Mat src_chan(src_border.size(), CV_MAKETYPE(src_border.depth(),1)); + Mat dst_chan(refdst.size(), CV_MAKETYPE(refdst.depth(), 1)); + mixChannels(src_border, src_chan, fromTo, 1); + for (int j = 0; j < drows; j++) + for (int i = 0; i < dcols; i++) + { + if (depth == CV_8U) + dst_chan.at(j, i) = eval(src_chan(Rect(i,j,kernel.width,kernel.height)), modes[modeind].kernel_x, modes[modeind].kernel_y); + else if (depth == CV_16U) + dst_chan.at(j, i) = eval(src_chan(Rect(i, j, kernel.width, kernel.height)), modes[modeind].kernel_x, modes[modeind].kernel_y); + else if (depth == CV_16S) + dst_chan.at(j, i) = eval(src_chan(Rect(i, j, kernel.width, kernel.height)), modes[modeind].kernel_x, modes[modeind].kernel_y); + else if (depth == CV_32S) + dst_chan.at(j, i) = eval(src_chan(Rect(i, j, kernel.width, kernel.height)), modes[modeind].kernel_x, modes[modeind].kernel_y); + else + CV_Assert(0); + } + mixChannels(dst_chan, refdst, toFrom, 1); + } + + GaussianBlur(src_roi, dst, kernel, modes[modeind].sigma_x, modes[modeind].sigma_y, bordermodes[borderind]); + + EXPECT_GE(0, cvtest::norm(refdst, dst, cv::NORM_L1)) + << "GaussianBlur " << cn << "-chan mat " << drows << "x" << dcols << " by kernel " << kernel << " sigma(" << modes[modeind].sigma_x << ";" << modes[modeind].sigma_y << ") failed with max diff " << cvtest::norm(refdst, dst, cv::NORM_INF); + } + } +} + +///* End of file. */