X-Git-Url: http://review.tizen.org/git/?a=blobdiff_plain;f=modules%2Fimgproc%2Fsrc%2Fsmooth.cpp;h=f4077521e15703da7ec264bdd81273a7c07fe03e;hb=4560909a5e5cb284cdfd5619cdf4cf3622410388;hp=10d14e398156947a92adb7099a8a824491b0eab3;hpb=8f04c8b16c8a1b38feba733ed5a69743fc968478;p=platform%2Fupstream%2Fopencv.git diff --git a/modules/imgproc/src/smooth.cpp b/modules/imgproc/src/smooth.cpp index 10d14e3..f407752 100644 --- a/modules/imgproc/src/smooth.cpp +++ b/modules/imgproc/src/smooth.cpp @@ -88,7 +88,7 @@ struct RowSum : anchor = _anchor; } - virtual void operator()(const uchar* src, uchar* dst, int width, int cn) + virtual void operator()(const uchar* src, uchar* dst, int width, int cn) CV_OVERRIDE { const T* S = (const T*)src; ST* D = (ST*)dst; @@ -199,9 +199,9 @@ struct ColumnSum : sumCount = 0; } - virtual void reset() { sumCount = 0; } + virtual void reset() CV_OVERRIDE { sumCount = 0; } - virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) + virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE { int i; ST* SUM; @@ -297,9 +297,9 @@ struct ColumnSum : sumCount = 0; } - virtual void reset() { sumCount = 0; } + virtual void reset() CV_OVERRIDE { sumCount = 0; } - virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) + virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE { int* SUM; bool haveScale = scale != 1; @@ -444,9 +444,9 @@ public BaseColumnFilter } } - virtual void reset() { sumCount = 0; } + virtual void reset() CV_OVERRIDE { sumCount = 0; } - virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) + virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE { const int ds = divScale; const int dd = divDelta; @@ -573,9 +573,9 @@ struct ColumnSum : sumCount = 0; } - virtual void reset() { sumCount = 0; } + virtual void reset() CV_OVERRIDE { sumCount = 0; } - virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) + virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE { int i; int* SUM; @@ -701,9 +701,9 @@ struct ColumnSum : sumCount = 0; } - virtual void reset() { sumCount = 0; } + virtual void reset() CV_OVERRIDE { sumCount = 0; } - virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) + virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE { int* SUM; bool haveScale = scale != 1; @@ -826,9 +826,9 @@ struct ColumnSum : sumCount = 0; } - virtual void reset() { sumCount = 0; } + virtual void reset() CV_OVERRIDE { sumCount = 0; } - virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) + virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE { int* SUM; bool haveScale = scale != 1; @@ -945,9 +945,9 @@ struct ColumnSum : sumCount = 0; } - virtual void reset() { sumCount = 0; } + virtual void reset() CV_OVERRIDE { sumCount = 0; } - virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) + virtual void operator()(const uchar** src, uchar* dst, int dststep, int count, int width) CV_OVERRIDE { int* SUM; bool haveScale = scale != 1; @@ -1332,8 +1332,6 @@ cv::Ptr cv::getRowSumFilter(int srcType, int sumType, int ksi CV_Error_( CV_StsNotImplemented, ("Unsupported combination of source format (=%d), and buffer format (=%d)", srcType, sumType)); - - return Ptr(); } @@ -1374,8 +1372,6 @@ cv::Ptr cv::getColumnSumFilter(int sumType, int dstType, i CV_Error_( CV_StsNotImplemented, ("Unsupported combination of sum format (=%d), and destination format (=%d)", sumType, dstType)); - - return Ptr(); } @@ -1458,7 +1454,7 @@ namespace cv ivx::Image::createAddressing(dst.cols, dst.rows, 1, (vx_int32)(dst.step)), dst.data); //ATTENTION: VX_CONTEXT_IMMEDIATE_BORDER attribute change could lead to strange issues in multi-threaded environments - //since OpenVX standart says nothing about thread-safety for now + //since OpenVX standard says nothing about thread-safety for now ivx::border_t prevBorder = ctx.immediateBorder(); ctx.setImmediateBorder(border, (vx_uint8)(0)); ivx::IVX_CHECK_STATUS(vxuBox3x3(ctx, ia, ib)); @@ -1606,7 +1602,7 @@ struct SqrRowSum : anchor = _anchor; } - virtual void operator()(const uchar* src, uchar* dst, int width, int cn) + virtual void operator()(const uchar* src, uchar* dst, int width, int cn) CV_OVERRIDE { const T* S = (const T*)src; ST* D = (ST*)dst; @@ -1656,8 +1652,6 @@ static Ptr getSqrRowSumFilter(int srcType, int sumType, int ksize CV_Error_( CV_StsNotImplemented, ("Unsupported combination of source format (=%d), and buffer format (=%d)", srcType, sumType)); - - return Ptr(); } } @@ -1715,6 +1709,7 @@ void cv::sqrBoxFilter( InputArray _src, OutputArray _dst, int ddepth, cv::Mat cv::getGaussianKernel( int n, double sigma, int ktype ) { + CV_Assert(n > 0); const int SMALL_GAUSSIAN_SIZE = 7; static const float small_gaussian_tab[][SMALL_GAUSSIAN_SIZE] = { @@ -1753,6 +1748,7 @@ cv::Mat cv::getGaussianKernel( int n, double sigma, int ktype ) } } + CV_DbgAssert(fabs(sum) > 0); sum = 1./sum; for( i = 0; i < n; i++ ) { @@ -1826,7 +1822,7 @@ void hlineSmooth1N(const uint8_t* src, int cn, const ufi int lencn = len*cn; v_uint16x8 v_mul = v_setall_u16(*((uint16_t*)m)); int i = 0; - for (; i < lencn - 15; i += 16) + for (; i <= lencn - 16; i += 16) { v_uint8x16 v_src = v_load(src + i); v_uint16x8 v_tmp0, v_tmp1; @@ -1834,7 +1830,7 @@ void hlineSmooth1N(const uint8_t* src, int cn, const ufi 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) + if (i <= lencn - 8) { v_uint16x8 v_src = v_load_expand(src + i); v_store((uint16_t*)dst + i, v_mul*v_src); @@ -1854,7 +1850,7 @@ void hlineSmooth1N1(const uint8_t* src, int cn, const uf { int lencn = len*cn; int i = 0; - for (; i < lencn - 15; i += 16) + for (; i <= lencn - 16; i += 16) { v_uint8x16 v_src = v_load(src + i); v_uint16x8 v_tmp0, v_tmp1; @@ -1862,7 +1858,7 @@ void hlineSmooth1N1(const uint8_t* src, int cn, const uf 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) + if (i <= lencn - 8) { v_uint16x8 v_src = v_load_expand(src + i); v_store((uint16_t*)dst + i, v_shl<8>(v_src)); @@ -1930,33 +1926,17 @@ void hlineSmooth3N(const uint8_t* src, int cn, const ufi 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_mul0 = v_setall_u16(*((uint16_t*)m)); + v_uint16x8 v_mul1 = v_setall_u16(*((uint16_t*)(m + 1))); + v_uint16x8 v_mul2 = v_setall_u16(*((uint16_t*)(m + 2))); + for (; i <= lencn - 16; i += 16, src += 16, dst += 16) { - v_uint16x8 v_src00, v_src01, v_src10, v_src11; - v_int16x8 v_tmp0, v_tmp1; - + 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_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))); + v_expand(v_load(src + cn), v_src20, v_src21); + v_store((uint16_t*)dst, v_src00 * v_mul0 + v_src10 * v_mul1 + v_src20 * v_mul2); + v_store((uint16_t*)dst + 8, v_src01 * v_mul0 + v_src11 * v_mul1 + v_src21 * v_mul2); } for (; i < lencn; i++, src++, dst++) *dst = m[0] * src[-cn] + m[1] * src[0] + m[2] * src[cn]; @@ -1998,7 +1978,7 @@ void hlineSmooth3N121(const ET* src, int cn, const FT*, int, FT* dst, int len, i 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); + *dst = (FT(src[-cn])>>2) + (FT(src[cn])>>2) + (FT(src[0])>>1); // Point that fall right from border for (int k = 0; k < cn; k++) @@ -2037,7 +2017,7 @@ void hlineSmooth3N121(const uint8_t* src, int cn, const src += cn; dst += cn; int i = cn, lencn = (len - 1)*cn; - for (; i < lencn - 15; i += 16, src += 16, dst += 16) + for (; i <= lencn - 16; 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); @@ -2061,11 +2041,107 @@ void hlineSmooth3N121(const uint8_t* src, int cn, const } } template +void hlineSmooth3Naba(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]<<1) + m[1] : m[1]; + for (int k = 0; k < cn; k++) + dst[k] = msum * src[k]; + } + else + { + // Point that fall left from border + 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] = m[1] * src[k] + m[0] * src[cn + k] + m[0] * src[src_idx*cn + k]; + } + else + { + for (int k = 0; k < cn; k++) + dst[k] = m[1] * src[k] + m[0] * src[cn + k]; + } + + src += cn; dst += cn; + for (int i = cn; i < (len - 1)*cn; i++, src++, dst++) + *dst = m[1] * src[0] + m[0] * src[-cn] + m[0] * src[cn]; + + // Point that fall right from border + 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] = m[1] * src[k] + m[0] * src[k - cn] + m[0] * src[src_idx + k]; + } + else + { + for (int k = 0; k < cn; k++) + dst[k] = m[0] * src[k - cn] + m[1] * src[k]; + } + } +} +template <> +void hlineSmooth3Naba(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]<<1) + m[1] : m[1]; + for (int k = 0; k < cn; k++) + dst[k] = msum * src[k]; + } + else + { + // Point that fall left from border + 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++) + ((uint16_t*)dst)[k] = ((uint16_t*)m)[1] * src[k] + ((uint16_t*)m)[0] * ((uint16_t)(src[cn + k]) + (uint16_t)(src[src_idx*cn + k])); + } + else + { + for (int k = 0; k < cn; k++) + dst[k] = m[1] * src[k] + m[0] * src[cn + k]; + } + + src += cn; dst += cn; + int i = cn, lencn = (len - 1)*cn; + v_uint16x8 v_mul0 = v_setall_u16(*((uint16_t*)m)); + v_uint16x8 v_mul1 = v_setall_u16(*((uint16_t*)m+1)); + for (; i <= lencn - 16; 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_mul0 + v_src10 * v_mul1); + v_store((uint16_t*)dst + 8, (v_src01 + v_src21) * v_mul0 + v_src11 * v_mul1); + } + for (; i < lencn; i++, src++, dst++) + *((uint16_t*)dst) = ((uint16_t*)m)[1] * src[0] + ((uint16_t*)m)[0] * ((uint16_t)(src[-cn]) + (uint16_t)(src[cn])); + + // Point that fall right from border + 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++) + ((uint16_t*)dst)[k] = ((uint16_t*)m)[1] * src[k] + ((uint16_t*)m)[0] * ((uint16_t)(src[k - cn]) + (uint16_t)(src[src_idx + k])); + } + else + { + for (int k = 0; k < cn; k++) + dst[k] = m[0] * src[k - cn] + m[1] * src[k]; + } + } +} +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]; + FT 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]; } @@ -2228,44 +2304,21 @@ void hlineSmooth5N(const uint8_t* src, int cn, const ufi 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_mul0 = v_setall_u16(*((uint16_t*)m)); + v_uint16x8 v_mul1 = v_setall_u16(*((uint16_t*)(m + 1))); + v_uint16x8 v_mul2 = v_setall_u16(*((uint16_t*)(m + 2))); + v_uint16x8 v_mul3 = v_setall_u16(*((uint16_t*)(m + 3))); + v_uint16x8 v_mul4 = v_setall_u16(*((uint16_t*)(m + 4))); + for (; i <= lencn - 16; i += 16, src += 16, dst += 16) { - v_uint16x8 v_src00, v_src01, v_src10, v_src11; - v_int16x8 v_tmp0, v_tmp1; - + 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_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))); + 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_src00 * v_mul0 + v_src10 * v_mul1 + v_src20 * v_mul2 + v_src30 * v_mul3 + v_src40 * v_mul4); + v_store((uint16_t*)dst + 8, v_src01 * v_mul0 + v_src11 * v_mul1 + v_src21 * v_mul2 + v_src31 * v_mul3 + v_src41 * v_mul4); } for (; i < lencn; i++, src++, dst++) *dst = m[0] * src[-2*cn] + m[1] * src[-cn] + m[2] * src[0] + m[3] * src[cn] + m[4] * src[2*cn]; @@ -2295,7 +2348,7 @@ void hlineSmooth5N14641(const ET* src, int cn, const FT*, int, FT* dst, int len, { if (borderType == BORDER_CONSTANT) for (int k = 0; k < cn; k++) - dst[k] = (FT(src[k])>>3)*3; + dst[k] = (FT(src[k])>>3)*(uint8_t)3; else for (int k = 0; k < cn; k++) dst[k] = src[k]; @@ -2305,8 +2358,8 @@ void hlineSmooth5N14641(const ET* src, int cn, const FT*, int, FT* dst, int len, 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; + dst[k] = (FT(src[k])>>4)*(uint8_t)6 + (FT(src[k + cn])>>2); + dst[k + cn] = (FT(src[k]) >> 2) + (FT(src[k + cn])>>4)*(uint8_t)6; } else { @@ -2316,8 +2369,8 @@ void hlineSmooth5N14641(const ET* src, int cn, const FT*, int, FT* dst, int len, 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); + dst[k] = (FT(src[k])>>4)*(uint8_t)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)*(uint8_t)6 + (FT(src[k])>>2) + (FT(src[k + idxp1])>>2) + (FT(src[k + idxm1])>>4) + (FT(src[k + idxp2])>>4); } } } @@ -2326,9 +2379,9 @@ void hlineSmooth5N14641(const ET* src, int cn, const FT*, int, FT* dst, int len, 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); + dst[k] = (FT(src[k])>>4)*(uint8_t)6 + (FT(src[k + cn])>>2) + (FT(src[k + 2 * cn])>>4); + dst[k + cn] = (FT(src[k + cn])>>4)*(uint8_t)6 + (FT(src[k])>>2) + (FT(src[k + 2 * cn])>>2); + dst[k + 2 * cn] = (FT(src[k + 2 * cn])>>4)*(uint8_t)6 + (FT(src[k + cn])>>2) + (FT(src[k])>>4); } else { @@ -2338,9 +2391,9 @@ void hlineSmooth5N14641(const ET* src, int cn, const FT*, int, FT* dst, int len, 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); + dst[k] = (FT(src[k])>>4)*(uint8_t)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)*(uint8_t)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)*(uint8_t)6 + (FT(src[k + cn])>>2) + (FT(src[k + idxp1])>>2) + (FT(src[k])>>4) + (FT(src[k + idxp2])>>4); } } } @@ -2349,8 +2402,8 @@ void hlineSmooth5N14641(const ET* src, int cn, const FT*, int, FT* dst, int len, // 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); + dst[k] = (FT(src[k])>>4)*(uint8_t)6 + (FT(src[cn + k])>>2) + (FT(src[2 * cn + k])>>4); + dst[k + cn] = (FT(src[cn + k])>>4)*(uint8_t)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 { @@ -2365,13 +2418,13 @@ void hlineSmooth5N14641(const ET* src, int cn, const FT*, int, FT* dst, int len, 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); + *dst = (FT(src[0])>>4)*(uint8_t)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); + dst[k] = (FT(src[k])>>4)*(uint8_t)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)*(uint8_t)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 { @@ -2392,7 +2445,7 @@ void hlineSmooth5N14641(const uint8_t* src, int cn, cons { if (borderType == BORDER_CONSTANT) for (int k = 0; k < cn; k++) - dst[k] = (ufixedpoint16(src[k])>>3) * 3; + dst[k] = (ufixedpoint16(src[k])>>3) * (uint8_t)3; else { for (int k = 0; k < cn; k++) @@ -2404,8 +2457,8 @@ void hlineSmooth5N14641(const uint8_t* src, int cn, cons 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; + dst[k] = (ufixedpoint16(src[k]) >> 4) * (uint8_t)6 + (ufixedpoint16(src[k + cn]) >> 2); + dst[k + cn] = (ufixedpoint16(src[k]) >> 2) + (ufixedpoint16(src[k + cn]) >> 4) * (uint8_t)6; } else { @@ -2415,8 +2468,8 @@ void hlineSmooth5N14641(const uint8_t* src, int cn, cons 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); + dst[k] = (ufixedpoint16(src[k]) >> 4) * (uint8_t)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) * (uint8_t)6 + (ufixedpoint16(src[k]) >> 2) + (ufixedpoint16(src[k + idxp1]) >> 2) + (ufixedpoint16(src[k + idxm1]) >> 4) + (ufixedpoint16(src[k + idxp2]) >> 4); } } } @@ -2425,9 +2478,9 @@ void hlineSmooth5N14641(const uint8_t* src, int cn, cons 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); + dst[k] = (ufixedpoint16(src[k]) >> 4) * (uint8_t)6 + (ufixedpoint16(src[k + cn]) >> 2) + (ufixedpoint16(src[k + 2 * cn]) >> 4); + dst[k + cn] = (ufixedpoint16(src[k + cn]) >> 4) * (uint8_t)6 + (ufixedpoint16(src[k]) >> 2) + (ufixedpoint16(src[k + 2 * cn]) >> 2); + dst[k + 2 * cn] = (ufixedpoint16(src[k + 2 * cn]) >> 4) * (uint8_t)6 + (ufixedpoint16(src[k + cn]) >> 2) + (ufixedpoint16(src[k]) >> 4); } else { @@ -2437,9 +2490,9 @@ void hlineSmooth5N14641(const uint8_t* src, int cn, cons 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); + dst[k] = (ufixedpoint16(src[k]) >> 4) * (uint8_t)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) * (uint8_t)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) * (uint8_t)6 + (ufixedpoint16(src[k + cn]) >> 2) + (ufixedpoint16(src[k + idxp1]) >> 2) + (ufixedpoint16(src[k]) >> 4) + (ufixedpoint16(src[k + idxp2]) >> 4); } } } @@ -2448,8 +2501,8 @@ void hlineSmooth5N14641(const uint8_t* src, int cn, cons // 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); + dst[k] = (ufixedpoint16(src[k]) >> 4) * (uint8_t)6 + (ufixedpoint16(src[cn + k]) >> 2) + (ufixedpoint16(src[2 * cn + k]) >> 4); + dst[k + cn] = (ufixedpoint16(src[cn + k]) >> 4) * (uint8_t)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 { @@ -2465,7 +2518,7 @@ void hlineSmooth5N14641(const uint8_t* src, int cn, cons 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) + for (; i <= lencn - 16; 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); @@ -2482,8 +2535,8 @@ void hlineSmooth5N14641(const uint8_t* src, int cn, cons // 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); + dst[k] = (ufixedpoint16(src[k]) >> 4) * (uint8_t)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) * (uint8_t)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 { @@ -2498,6 +2551,215 @@ void hlineSmooth5N14641(const uint8_t* src, int cn, cons } } template +void hlineSmooth5Nabcba(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])<<1) + m[2] : 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[1] * 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[1] * src[k + cn] + m[0] * 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[1] * src[k + idxp1] + m[0] * 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[1] * src[k + cn] + m[0] * src[k + 2 * cn]; + dst[k + cn] = m[1] * src[k] + m[2] * src[k + cn] + m[1] * 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[1] * src[k + cn] + m[0] * 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[1] * src[k + 2 * cn] + m[0] * src[k + idxm1] + m[0] * src[k + idxp1]; + dst[k + 2 * cn] = m[0] * src[k] + m[1] * src[k + cn] + m[2] * src[k + 2 * cn] + m[1] * src[k + idxp1] + m[0] * src[k + idxp2]; + } + } + } + else + { + // Points that fall left from border + for (int k = 0; k < cn; k++) + { + dst[k] = m[2] * src[k] + m[1] * src[cn + k] + m[0] * src[2 * cn + k]; + dst[k + cn] = m[1] * src[k] + m[2] * src[cn + k] + m[1] * src[2 * cn + k] + m[0] * 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[0] * src[idxp1 + k]; + dst[k + cn] = dst[k + cn] + m[1] * src[idxp1 + k] + m[0] * src[idxp2 + k]; + } + } + } +} +template <> +void hlineSmooth5Nabcba(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]) << 1) + m[2] : 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[1] * 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++) + { + ((uint16_t*)dst)[k] = ((uint16_t*)m)[1] * ((uint16_t)(src[k + idxm1]) + (uint16_t)(src[k + cn])) + ((uint16_t*)m)[2] * src[k] + ((uint16_t*)m)[0] * ((uint16_t)(src[k + idxp1]) + (uint16_t)(src[k + idxm2])); + ((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[0] * ((uint16_t)(src[k + idxm1]) + (uint16_t)(src[k + idxp2])) + ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[k + idxp1])) + ((uint16_t*)m)[2] * src[k + cn]; + } + } + } + else if (len == 3) + { + if (borderType == BORDER_CONSTANT) + for (int k = 0; k < cn; k++) + { + dst[k] = m[2] * src[k] + m[1] * src[k + cn] + m[0] * src[k + 2 * cn]; + ((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[k + 2 * cn])) + ((uint16_t*)m)[2] * src[k + 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++) + { + ((uint16_t*)dst)[k] = ((uint16_t*)m)[2] * src[k] + ((uint16_t*)m)[1] * ((uint16_t)(src[k + cn]) + (uint16_t)(src[k + idxm1])) + ((uint16_t*)m)[0] * ((uint16_t)(src[k + 2 * cn]) + (uint16_t)(src[k + idxm2])); + ((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[2] * src[k + cn] + ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[k + 2 * cn])) + ((uint16_t*)m)[0] * ((uint16_t)(src[k + idxm1]) + (uint16_t)(src[k + idxp1])); + ((uint16_t*)dst)[k + 2 * cn] = ((uint16_t*)m)[0] * ((uint16_t)(src[k]) + (uint16_t)(src[k + idxp2])) + ((uint16_t*)m)[1] * ((uint16_t)(src[k + cn]) + (uint16_t)(src[k + idxp1])) + ((uint16_t*)m)[2] * src[k + 2 * cn]; + } + } + } + else + { + // Points that fall left from border + 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++) + { + ((uint16_t*)dst)[k] = ((uint16_t*)m)[2] * src[k] + ((uint16_t*)m)[1] * ((uint16_t)(src[cn + k]) + (uint16_t)(src[idxm1 + k])) + ((uint16_t*)m)[0] * ((uint16_t)(src[2 * cn + k]) + (uint16_t)(src[idxm2 + k])); + ((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[2 * cn + k])) + ((uint16_t*)m)[2] * src[cn + k] + ((uint16_t*)m)[0] * ((uint16_t)(src[3 * cn + k]) + (uint16_t)(src[idxm1 + k])); + } + } + else + { + for (int k = 0; k < cn; k++) + { + dst[k] = m[2] * src[k] + m[1] * src[cn + k] + m[0] * src[2 * cn + k]; + ((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[2 * cn + k])) + ((uint16_t*)m)[2] * src[cn + k] + ((uint16_t*)m)[0] * src[3 * cn + k]; + } + } + + src += 2 * cn; dst += 2 * cn; + int i = 2 * cn, lencn = (len - 2)*cn; + v_uint16x8 v_mul0 = v_setall_u16(*((uint16_t*)m)); + v_uint16x8 v_mul1 = v_setall_u16(*((uint16_t*)(m + 1))); + v_uint16x8 v_mul2 = v_setall_u16(*((uint16_t*)(m + 2))); + for (; i <= lencn - 16; 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_src00 + v_src40) * v_mul0 + (v_src10 + v_src30)* v_mul1 + v_src20 * v_mul2); + v_store((uint16_t*)dst + 8, (v_src01 + v_src41) * v_mul0 + (v_src11 + v_src31) * v_mul1 + v_src21 * v_mul2); + } + for (; i < lencn; i++, src++, dst++) + *((uint16_t*)dst) = ((uint16_t*)m)[0] * ((uint16_t)(src[-2 * cn]) + (uint16_t)(src[2 * cn])) + ((uint16_t*)m)[1] * ((uint16_t)(src[-cn]) + (uint16_t)(src[cn])) + ((uint16_t*)m)[2] * src[0]; + + // Points that fall right from border + 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++) + { + ((uint16_t*)dst)[k] = ((uint16_t*)m)[0] * ((uint16_t)(src[k - 2 * cn]) + (uint16_t)(src[idxp1 + k])) + ((uint16_t*)m)[1] * ((uint16_t)(src[k - cn]) + (uint16_t)(src[k + cn])) + ((uint16_t*)m)[2] * src[k]; + ((uint16_t*)dst)[k + cn] = ((uint16_t*)m)[0] * ((uint16_t)(src[k - cn]) + (uint16_t)(src[idxp2 + k])) + ((uint16_t*)m)[1] * ((uint16_t)(src[k]) + (uint16_t)(src[idxp1 + k])) + ((uint16_t*)m)[2] * src[k + cn]; + } + } + else + { + for (int k = 0; k < cn; k++) + { + ((uint16_t*)dst)[k] = ((uint16_t*)m)[0] * src[k - 2 * cn] + ((uint16_t*)m)[1] * ((uint16_t)(src[k - cn]) + (uint16_t)(src[k + cn])) + ((uint16_t*)m)[2] * src[k]; + dst[k + cn] = m[0] * src[k - cn] + m[1] * src[k] + m[2] * src[k + cn]; + } + } + } +} +template void hlineSmooth(const ET* src, int cn, const FT* m, int n, FT* dst, int len, int borderType) { int pre_shift = n / 2; @@ -2582,57 +2844,157 @@ void hlineSmooth(const uint8_t* src, int cn, const ufixe } i *= cn; int lencn = (len - post_shift + 1)*cn; - for (; i < lencn - 15; i+=16, src+=16, dst+=16) + for (; i <= lencn - 16; 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_uint16x8 v_src0, v_src1; + v_uint16x8 v_mul = v_setall_u16(*((uint16_t*)m)); + v_expand(v_load(src), v_src0, v_src1); + v_uint16x8 v_res0 = v_src0 * v_mul; + v_uint16x8 v_res1 = v_src1 * v_mul; + for (int j = 1; j < n; j++) { - 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); + v_mul = v_setall_u16(*((uint16_t*)(m + j))); + v_expand(v_load(src + j * cn), v_src0, v_src1); + v_res0 += v_src0 * v_mul; + v_res1 += v_src1 * v_mul; } - if (j < n) + v_store((uint16_t*)dst, v_res0); + v_store((uint16_t*)dst+8, v_res1); + } + 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 hlineSmoothONa_yzy_a(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[pre_shift] * src[pre_shift*cn]; + for (int j = 0; j < pre_shift; j++) + *dst = *dst + m[j] * src[j*cn] + m[j] * src[(n-1-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 hlineSmoothONa_yzy_a(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 - 16; i += 16, src += 16, dst += 16) + { + v_uint16x8 v_src00, v_src01, v_srcN00, v_srcN01; + + v_uint16x8 v_mul = v_setall_u16(*((uint16_t*)(m + pre_shift))); + v_expand(v_load(src + pre_shift * cn), v_src00, v_src01); + v_uint16x8 v_res0 = v_src00 * v_mul; + v_uint16x8 v_res1 = v_src01 * v_mul; + for (int j = 0; j < pre_shift; j ++) { - v_int32x4 v_resj0, v_resj1, v_resj2, v_resj3; - v_mul = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + j)))); + v_mul = v_setall_u16(*((uint16_t*)(m + j))); v_expand(v_load(src + j * cn), v_src00, v_src01); - v_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_expand(v_load(src + (n - 1 - j)*cn), v_srcN00, v_srcN01); + v_res0 += (v_src00 + v_srcN00) * v_mul; + v_res1 += (v_src01 + v_srcN01) * v_mul; } - v_store((uint16_t*)dst, v_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))); + v_store((uint16_t*)dst, v_res0); + v_store((uint16_t*)dst + 8, v_res1); } 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]; + *dst = m[pre_shift] * src[pre_shift*cn]; + for (int j = 0; j < pre_shift; j++) + *dst = *dst + m[j] * src[j*cn] + m[j] * src[(n - 1 - j)*cn]; } i /= cn; for (i -= pre_shift; i < len - pre_shift; i++, src += cn, dst += cn) // Points that fall right from border @@ -2657,20 +3019,33 @@ 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]; + 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)); +#if CV_SSE2 + v_uint16x8 v_1 = v_setall_u16(1); + v_mul += v_mul; +#endif int i = 0; - for (; i < len - 7; i += 8) + for (; i <= len - 16; i += 16) { v_uint16x8 v_src0 = v_load((uint16_t*)src0 + i); - v_uint32x4 v_res0, v_res1; + v_uint16x8 v_src1 = v_load((uint16_t*)src0 + i + 8); + v_uint8x16 v_res; +#if CV_SSE2 + v_res.val = _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(v_1.val, _mm_mulhi_epu16(v_src0.val, v_mul.val)),1), + _mm_srli_epi16(_mm_add_epi16(v_1.val, _mm_mulhi_epu16(v_src1.val, v_mul.val)),1)); +#else + v_uint32x4 v_res0, v_res1, v_res2, v_res3; v_mul_expand(v_src0, v_mul, v_res0, v_res1); - v_pack_store(dst + i, v_rshr_pack<16>(v_res0, v_res1)); + v_mul_expand(v_src1, v_mul, v_res2, v_res3); + v_res = v_pack(v_rshr_pack<16>(v_res0, v_res1), v_rshr_pack<16>(v_res2, v_res3)); +#endif + v_store(dst + i, v_res); } for (; i < len; i++) dst[i] = m[0] * src0[i]; @@ -2687,7 +3062,7 @@ void vlineSmooth1N1(const ufixedpoint16* const * src, co { const ufixedpoint16* src0 = src[0]; int i = 0; - for (; i < len - 7; i += 8) + for (; i <= len - 8; i += 8) v_rshr_pack_store<8>(dst + i, v_load((uint16_t*)(src0 + i))); for (; i < len; i++) dst[i] = src0[i]; @@ -2701,40 +3076,73 @@ void vlineSmooth3N(const FT* const * src, const FT* m, int, ET* dst, int len) template <> void vlineSmooth3N(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len) { + int i = 0; 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) + for (; i <= len - 32; i += 32) { - v_int16x8 v_src0, v_src1; + v_int16x8 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13; 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_src00 = v_load((int16_t*)(src[0]) + i); + v_src01 = v_load((int16_t*)(src[0]) + i + 8); + v_src02 = v_load((int16_t*)(src[0]) + i + 16); + v_src03 = v_load((int16_t*)(src[0]) + i + 24); + v_src10 = v_load((int16_t*)(src[1]) + i); + v_src11 = v_load((int16_t*)(src[1]) + i + 8); + v_src12 = v_load((int16_t*)(src[1]) + i + 16); + v_src13 = v_load((int16_t*)(src[1]) + i + 24); + v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul01); v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul01); + v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res2 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res3 = v_dotprod(v_tmp1, v_mul01); + v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res4 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res5 = v_dotprod(v_tmp1, v_mul01); + v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res6 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res7 = 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_src00 = v_load((int16_t*)(src[2]) + i); + v_src01 = v_load((int16_t*)(src[2]) + i + 8); + v_src02 = v_load((int16_t*)(src[2]) + i + 16); + v_src03 = v_load((int16_t*)(src[2]) + i + 24); + v_mul_expand(v_add_wrap(v_src00, v_128), v_mul2, v_resj0, v_resj1); v_res0 += v_resj0; v_res1 += v_resj1; + v_mul_expand(v_add_wrap(v_src01, v_128), v_mul2, v_resj0, v_resj1); + v_res2 += v_resj0; + v_res3 += v_resj1; + v_mul_expand(v_add_wrap(v_src02, v_128), v_mul2, v_resj0, v_resj1); + v_res4 += v_resj0; + v_res5 += v_resj1; + v_mul_expand(v_add_wrap(v_src03, v_128), v_mul2, v_resj0, v_resj1); + v_res6 += v_resj0; + v_res7 += 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); + v_res2 += v_128_4; + v_res3 += v_128_4; + v_res4 += v_128_4; + v_res5 += v_128_4; + v_res6 += v_128_4; + v_res7 += v_128_4; + + v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)), + v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3)))); + v_store(dst + i + 16, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)), + v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7)))); } for (; i < len; i++) dst[i] = m[0] * src[0][i] + m[1] * src[1][i] + m[2] * src[2][i]; @@ -2743,20 +3151,23 @@ 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); + dst[i] = (FT::WT(src[0][i]) >> 2) + (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) + for (; i <= len - 16; i += 16) { - v_uint32x4 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21; + v_uint32x4 v_src00, v_src01, v_src02, v_src03, v_src10, v_src11, v_src12, v_src13, v_src20, v_src21, v_src22, v_src23; v_expand(v_load((uint16_t*)(src[0]) + i), v_src00, v_src01); + v_expand(v_load((uint16_t*)(src[0]) + i + 8), v_src02, v_src03); v_expand(v_load((uint16_t*)(src[1]) + i), v_src10, v_src11); + v_expand(v_load((uint16_t*)(src[1]) + i + 8), v_src12, v_src13); 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); + v_expand(v_load((uint16_t*)(src[2]) + i + 8), v_src22, v_src23); + v_store(dst + i, v_pack(v_rshr_pack<10>(v_src00 + v_src20 + (v_src10 + v_src10), v_src01 + v_src21 + (v_src11 + v_src11)), + v_rshr_pack<10>(v_src02 + v_src22 + (v_src12 + v_src12), v_src03 + v_src23 + (v_src13 + v_src13)))); } 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; @@ -2770,47 +3181,95 @@ void vlineSmooth5N(const FT* const * src, const FT* m, int, ET* dst, int len) template <> void vlineSmooth5N(const ufixedpoint16* const * src, const ufixedpoint16* m, int, uint8_t* dst, int len) { + int i = 0; 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) + for (; i <= len - 32; i += 32) { - v_int16x8 v_src0, v_src1; + v_int16x8 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13; 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_src00 = v_load((int16_t*)(src[0]) + i); + v_src01 = v_load((int16_t*)(src[0]) + i + 8); + v_src02 = v_load((int16_t*)(src[0]) + i + 16); + v_src03 = v_load((int16_t*)(src[0]) + i + 24); + v_src10 = v_load((int16_t*)(src[1]) + i); + v_src11 = v_load((int16_t*)(src[1]) + i + 8); + v_src12 = v_load((int16_t*)(src[1]) + i + 16); + v_src13 = v_load((int16_t*)(src[1]) + i + 24); + v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul01); v_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_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res2 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res3 = v_dotprod(v_tmp1, v_mul01); + v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res4 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res5 = v_dotprod(v_tmp1, v_mul01); + v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res6 = v_dotprod(v_tmp0, v_mul01); + v_int32x4 v_res7 = v_dotprod(v_tmp1, v_mul01); + + v_src00 = v_load((int16_t*)(src[2]) + i); + v_src01 = v_load((int16_t*)(src[2]) + i + 8); + v_src02 = v_load((int16_t*)(src[2]) + i + 16); + v_src03 = v_load((int16_t*)(src[2]) + i + 24); + v_src10 = v_load((int16_t*)(src[3]) + i); + v_src11 = v_load((int16_t*)(src[3]) + i + 8); + v_src12 = v_load((int16_t*)(src[3]) + i + 16); + v_src13 = v_load((int16_t*)(src[3]) + i + 24); + v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_res0 += v_dotprod(v_tmp0, v_mul23); v_res1 += v_dotprod(v_tmp1, v_mul23); + v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1); + v_res2 += v_dotprod(v_tmp0, v_mul23); + v_res3 += v_dotprod(v_tmp1, v_mul23); + v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1); + v_res4 += v_dotprod(v_tmp0, v_mul23); + v_res5 += v_dotprod(v_tmp1, v_mul23); + v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1); + v_res6 += v_dotprod(v_tmp0, v_mul23); + v_res7 += 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_src00 = v_load((int16_t*)(src[4]) + i); + v_src01 = v_load((int16_t*)(src[4]) + i + 8); + v_src02 = v_load((int16_t*)(src[4]) + i + 16); + v_src03 = v_load((int16_t*)(src[4]) + i + 24); + v_mul_expand(v_add_wrap(v_src00, v_128), v_mul4, v_resj0, v_resj1); v_res0 += v_resj0; v_res1 += v_resj1; + v_mul_expand(v_add_wrap(v_src01, v_128), v_mul4, v_resj0, v_resj1); + v_res2 += v_resj0; + v_res3 += v_resj1; + v_mul_expand(v_add_wrap(v_src02, v_128), v_mul4, v_resj0, v_resj1); + v_res4 += v_resj0; + v_res5 += v_resj1; + v_mul_expand(v_add_wrap(v_src03, v_128), v_mul4, v_resj0, v_resj1); + v_res6 += v_resj0; + v_res7 += 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); + v_res2 += v_128_4; + v_res3 += v_128_4; + v_res4 += v_128_4; + v_res5 += v_128_4; + v_res6 += v_128_4; + v_res7 += v_128_4; + + v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)), + v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3)))); + v_store(dst + i + 16, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)), + v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7)))); } 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]; @@ -2819,24 +3278,33 @@ 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; + dst[i] = (FT::WT(src[2][i])*(uint8_t)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) + for (; i <= len - 16; i += 16) { - v_uint32x4 v_src00, v_src01, v_src10, v_src11, v_src20, v_src21, v_src30, v_src31, v_src40, v_src41; + v_uint32x4 v_src00, v_src10, v_src20, v_src30, v_src40; + v_uint32x4 v_src01, v_src11, v_src21, v_src31, v_src41; + v_uint32x4 v_src02, v_src12, v_src22, v_src32, v_src42; + v_uint32x4 v_src03, v_src13, v_src23, v_src33, v_src43; v_expand(v_load((uint16_t*)(src[0]) + i), v_src00, v_src01); + v_expand(v_load((uint16_t*)(src[0]) + i + 8), v_src02, v_src03); v_expand(v_load((uint16_t*)(src[1]) + i), v_src10, v_src11); + v_expand(v_load((uint16_t*)(src[1]) + i + 8), v_src12, v_src13); v_expand(v_load((uint16_t*)(src[2]) + i), v_src20, v_src21); + v_expand(v_load((uint16_t*)(src[2]) + i + 8), v_src22, v_src23); v_expand(v_load((uint16_t*)(src[3]) + i), v_src30, v_src31); + v_expand(v_load((uint16_t*)(src[3]) + i + 8), v_src32, v_src33); 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); + v_expand(v_load((uint16_t*)(src[4]) + i + 8), v_src42, v_src43); + v_store(dst + i, v_pack(v_rshr_pack<12>(v_src20*v_6 + ((v_src10 + v_src30) << 2) + v_src00 + v_src40, + v_src21*v_6 + ((v_src11 + v_src31) << 2) + v_src01 + v_src41), + v_rshr_pack<12>(v_src22*v_6 + ((v_src12 + v_src32) << 2) + v_src02 + v_src42, + v_src23*v_6 + ((v_src13 + v_src33) << 2) + v_src03 + v_src43))); } for (; i < len; i++) dst[i] = ((uint32_t)(((uint16_t*)(src[2]))[i]) * 6 + @@ -2857,8 +3325,8 @@ void vlineSmooth(const FT* const * src, const FT* m, int n, ET* dst, int len) template <> void vlineSmooth(const ufixedpoint16* const * src, const ufixedpoint16* m, int n, uint8_t* dst, int len) { + int i = 0; 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) { @@ -2868,46 +3336,188 @@ void vlineSmooth(const ufixedpoint16* const * src, const 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) + for (; i <= len - 32; i += 32) { - v_int16x8 v_src0, v_src1; + v_int16x8 v_src00, v_src10, v_src01, v_src11, v_src02, v_src12, v_src03, v_src13; 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_src00 = v_load((int16_t*)(src[0]) + i); + v_src01 = v_load((int16_t*)(src[0]) + i + 8); + v_src02 = v_load((int16_t*)(src[0]) + i + 16); + v_src03 = v_load((int16_t*)(src[0]) + i + 24); + v_src10 = v_load((int16_t*)(src[1]) + i); + v_src11 = v_load((int16_t*)(src[1]) + i + 8); + v_src12 = v_load((int16_t*)(src[1]) + i + 16); + v_src13 = v_load((int16_t*)(src[1]) + i + 24); + v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_int32x4 v_res0 = v_dotprod(v_tmp0, v_mul); v_int32x4 v_res1 = v_dotprod(v_tmp1, v_mul); + v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res2 = v_dotprod(v_tmp0, v_mul); + v_int32x4 v_res3 = v_dotprod(v_tmp1, v_mul); + v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res4 = v_dotprod(v_tmp0, v_mul); + v_int32x4 v_res5 = v_dotprod(v_tmp1, v_mul); + v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1); + v_int32x4 v_res6 = v_dotprod(v_tmp0, v_mul); + v_int32x4 v_res7 = 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_src00 = v_load((int16_t*)(src[j]) + i); + v_src01 = v_load((int16_t*)(src[j]) + i + 8); + v_src02 = v_load((int16_t*)(src[j]) + i + 16); + v_src03 = v_load((int16_t*)(src[j]) + i + 24); + v_src10 = v_load((int16_t*)(src[j+1]) + i); + v_src11 = v_load((int16_t*)(src[j+1]) + i + 8); + v_src12 = v_load((int16_t*)(src[j+1]) + i + 16); + v_src13 = v_load((int16_t*)(src[j+1]) + i + 24); + v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src10, v_128), v_tmp0, v_tmp1); v_res0 += v_dotprod(v_tmp0, v_mul); v_res1 += v_dotprod(v_tmp1, v_mul); + v_zip(v_add_wrap(v_src01, v_128), v_add_wrap(v_src11, v_128), v_tmp0, v_tmp1); + v_res2 += v_dotprod(v_tmp0, v_mul); + v_res3 += v_dotprod(v_tmp1, v_mul); + v_zip(v_add_wrap(v_src02, v_128), v_add_wrap(v_src12, v_128), v_tmp0, v_tmp1); + v_res4 += v_dotprod(v_tmp0, v_mul); + v_res5 += v_dotprod(v_tmp1, v_mul); + v_zip(v_add_wrap(v_src03, v_128), v_add_wrap(v_src13, v_128), v_tmp0, v_tmp1); + v_res6 += v_dotprod(v_tmp0, v_mul); + v_res7 += 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_src00 = v_load((int16_t*)(src[j]) + i); + v_src01 = v_load((int16_t*)(src[j]) + i + 8); + v_src02 = v_load((int16_t*)(src[j]) + i + 16); + v_src03 = v_load((int16_t*)(src[j]) + i + 24); + v_mul_expand(v_add_wrap(v_src00, v_128), v_mul, v_resj0, v_resj1); v_res0 += v_resj0; v_res1 += v_resj1; + v_mul_expand(v_add_wrap(v_src01, v_128), v_mul, v_resj0, v_resj1); + v_res2 += v_resj0; + v_res3 += v_resj1; + v_mul_expand(v_add_wrap(v_src02, v_128), v_mul, v_resj0, v_resj1); + v_res4 += v_resj0; + v_res5 += v_resj1; + v_mul_expand(v_add_wrap(v_src03, v_128), v_mul, v_resj0, v_resj1); + v_res6 += v_resj0; + v_res7 += v_resj1; } v_res0 += v_128_4; v_res1 += v_128_4; + v_res2 += v_128_4; + v_res3 += v_128_4; + v_res4 += v_128_4; + v_res5 += v_128_4; + v_res6 += v_128_4; + v_res7 += v_128_4; + + v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)), + v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3)))); + v_store(dst + i + 16, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)), + v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7)))); + } + 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 +void vlineSmoothONa_yzy_a(const FT* const * src, const FT* m, int n, ET* dst, int len) +{ + int pre_shift = n / 2; + for (int i = 0; i < len; i++) + { + typename FT::WT val = m[pre_shift] * src[pre_shift][i]; + for (int j = 0; j < pre_shift; j++) + val = val + m[j] * src[j][i] + m[j] * src[(n - 1 - j)][i]; + dst[i] = val; + } +} +template <> +void vlineSmoothONa_yzy_a(const ufixedpoint16* const * src, const ufixedpoint16* m, int n, uint8_t* dst, int len) +{ + int pre_shift = n / 2; + int i = 0; + 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[pre_shift] + m[n - 1]; + for (int j = 1; j < pre_shift; j++) + msum = msum + m[j] + m[n - 1 - j]; + ufixedpoint32 val[] = { msum * ufixedpoint16((uint8_t)128) }; + v_128_4 = v_setall_s32(*((int32_t*)val)); + } + for (; i <= len - 32; i += 32) + { + v_int16x8 v_src00, v_src10, v_src20, v_src30, v_src01, v_src11, v_src21, v_src31; + v_int32x4 v_res0, v_res1, v_res2, v_res3, v_res4, v_res5, v_res6, v_res7; + v_int16x8 v_tmp0, v_tmp1, v_tmp2, v_tmp3, v_tmp4, v_tmp5, v_tmp6, v_tmp7; + + v_int16x8 v_mul = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + pre_shift)))); + v_src00 = v_load((int16_t*)(src[pre_shift]) + i); + v_src10 = v_load((int16_t*)(src[pre_shift]) + i + 8); + v_src20 = v_load((int16_t*)(src[pre_shift]) + i + 16); + v_src30 = v_load((int16_t*)(src[pre_shift]) + i + 24); + v_mul_expand(v_add_wrap(v_src00, v_128), v_mul, v_res0, v_res1); + v_mul_expand(v_add_wrap(v_src10, v_128), v_mul, v_res2, v_res3); + v_mul_expand(v_add_wrap(v_src20, v_128), v_mul, v_res4, v_res5); + v_mul_expand(v_add_wrap(v_src30, v_128), v_mul, v_res6, v_res7); + + int j = 0; + for (; j < pre_shift; j++) + { + v_mul = v_reinterpret_as_s16(v_setall_u16(*((uint16_t*)(m + j)))); - v_uint16x8 v_res = v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)); - v_pack_store(dst + i, v_res); + v_src00 = v_load((int16_t*)(src[j]) + i); + v_src10 = v_load((int16_t*)(src[j]) + i + 8); + v_src20 = v_load((int16_t*)(src[j]) + i + 16); + v_src30 = v_load((int16_t*)(src[j]) + i + 24); + v_src01 = v_load((int16_t*)(src[n - 1 - j]) + i); + v_src11 = v_load((int16_t*)(src[n - 1 - j]) + i + 8); + v_src21 = v_load((int16_t*)(src[n - 1 - j]) + i + 16); + v_src31 = v_load((int16_t*)(src[n - 1 - j]) + i + 24); + v_zip(v_add_wrap(v_src00, v_128), v_add_wrap(v_src01, v_128), v_tmp0, v_tmp1); + v_res0 += v_dotprod(v_tmp0, v_mul); + v_res1 += v_dotprod(v_tmp1, v_mul); + v_zip(v_add_wrap(v_src10, v_128), v_add_wrap(v_src11, v_128), v_tmp2, v_tmp3); + v_res2 += v_dotprod(v_tmp2, v_mul); + v_res3 += v_dotprod(v_tmp3, v_mul); + v_zip(v_add_wrap(v_src20, v_128), v_add_wrap(v_src21, v_128), v_tmp4, v_tmp5); + v_res4 += v_dotprod(v_tmp4, v_mul); + v_res5 += v_dotprod(v_tmp5, v_mul); + v_zip(v_add_wrap(v_src30, v_128), v_add_wrap(v_src31, v_128), v_tmp6, v_tmp7); + v_res6 += v_dotprod(v_tmp6, v_mul); + v_res7 += v_dotprod(v_tmp7, v_mul); + } + + v_res0 += v_128_4; + v_res1 += v_128_4; + v_res2 += v_128_4; + v_res3 += v_128_4; + v_res4 += v_128_4; + v_res5 += v_128_4; + v_res6 += v_128_4; + v_res7 += v_128_4; + + v_store(dst + i , v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res0, v_res1)), + v_reinterpret_as_u16(v_rshr_pack<16>(v_res2, v_res3)))); + v_store(dst + i + 16, v_pack(v_reinterpret_as_u16(v_rshr_pack<16>(v_res4, v_res5)), + v_reinterpret_as_u16(v_rshr_pack<16>(v_res6, v_res7)))); } for (; i < len; i++) { @@ -2930,61 +3540,85 @@ public: { if (kxlen == 1) { - if ((kx[0] - FT::one()).isZero()) + if (kx[0] == FT::one()) 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()) + if (kx[0] == (FT::one()>>2)&&kx[1] == (FT::one()>>1)&&kx[2] == (FT::one()>>2)) hlineSmoothFunc = hlineSmooth3N121; + else if ((kx[0] - kx[2]).isZero()) + hlineSmoothFunc = hlineSmooth3Naba; 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()) + if (kx[2] == (FT::one()*(uint8_t)3>>3) && + kx[1] == (FT::one()>>2) && kx[3] == (FT::one()>>2) && + kx[0] == (FT::one()>>4) && kx[4] == (FT::one()>>4)) hlineSmoothFunc = hlineSmooth5N14641; + else if (kx[0] == kx[4] && kx[1] == kx[3]) + hlineSmoothFunc = hlineSmooth5Nabcba; else hlineSmoothFunc = hlineSmooth5N; } + else if (kxlen % 2 == 1) + { + hlineSmoothFunc = hlineSmoothONa_yzy_a; + for (int i = 0; i < kxlen / 2; i++) + if (!(kx[i] == kx[kxlen - 1 - i])) + { + hlineSmoothFunc = hlineSmooth; + break; + } + } else hlineSmoothFunc = hlineSmooth; if (kylen == 1) { - if ((ky[0] - FT::one()).isZero()) + if (ky[0] == FT::one()) 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()) + if (ky[0] == (FT::one() >> 2) && ky[1] == (FT::one() >> 1) && ky[2] == (FT::one() >> 2)) 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()) + if (ky[2] == (FT::one() * (uint8_t)3 >> 3) && + ky[1] == (FT::one() >> 2) && ky[3] == (FT::one() >> 2) && + ky[0] == (FT::one() >> 4) && ky[4] == (FT::one() >> 4)) vlineSmoothFunc = vlineSmooth5N14641; else vlineSmoothFunc = vlineSmooth5N; } + else if (kylen % 2 == 1) + { + vlineSmoothFunc = vlineSmoothONa_yzy_a; + for (int i = 0; i < kylen / 2; i++) + if (!(ky[i] == ky[kylen - 1 - i])) + { + vlineSmoothFunc = vlineSmooth; + break; + } + } else vlineSmoothFunc = vlineSmooth; } - virtual void operator() (const Range& range) const + virtual void operator() (const Range& range) const CV_OVERRIDE { AutoBuffer _buf(width*cn*kylen); - FT* buf = _buf; + FT* buf = _buf.data(); AutoBuffer _ptrs(kylen*2); - FT** ptrs = _ptrs; + FT** ptrs = _ptrs.data(); if (kylen == 1) { @@ -3345,7 +3979,7 @@ static bool openvx_gaussianBlur(InputArray _src, OutputArray _dst, Size ksize, ivx::Image::createAddressing(dst.cols, dst.rows, 1, (vx_int32)(dst.step)), dst.data); //ATTENTION: VX_CONTEXT_IMMEDIATE_BORDER attribute change could lead to strange issues in multi-threaded environments - //since OpenVX standart says nothing about thread-safety for now + //since OpenVX standard says nothing about thread-safety for now ivx::border_t prevBorder = ctx.immediateBorder(); ctx.setImmediateBorder(border, (vx_uint8)(0)); ivx::IVX_CHECK_STATUS(vxuGaussian3x3(ctx, ia, ib)); @@ -3365,7 +3999,8 @@ static bool openvx_gaussianBlur(InputArray _src, OutputArray _dst, Size ksize, #endif #ifdef HAVE_IPP -#if IPP_VERSION_X100 == 201702 // IW 2017u2 has bug which doesn't allow use of partial inMem with tiling +// IW 2017u2 has bug which doesn't allow use of partial inMem with tiling +#if IPP_DISABLE_GAUSSIANBLUR_PARALLEL #define IPP_GAUSSIANBLUR_PARALLEL 0 #else #define IPP_GAUSSIANBLUR_PARALLEL 1 @@ -3384,7 +4019,7 @@ public: { } - virtual void operator() (const Range& range) const + virtual void operator() (const Range& range) const CV_OVERRIDE { CV_INSTRUMENT_REGION_IPP() @@ -3520,7 +4155,7 @@ void cv::GaussianBlur( InputArray _src, OutputArray _dst, Size ksize, 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)); + parallel_for_(Range(0, dst.rows), invoker, std::max(1, std::min(getNumThreads(), getNumberOfCPUs()))); return; } @@ -4416,7 +5051,7 @@ namespace cv ivx::Image::createAddressing(dst.cols, dst.rows, 1, (vx_int32)(dst.step)), dst.data); //ATTENTION: VX_CONTEXT_IMMEDIATE_BORDER attribute change could lead to strange issues in multi-threaded environments - //since OpenVX standart says nothing about thread-safety for now + //since OpenVX standard says nothing about thread-safety for now ivx::border_t prevBorder = ctx.immediateBorder(); ctx.setImmediateBorder(VX_BORDER_REPLICATE); #ifdef VX_VERSION_1_1 @@ -4566,11 +5201,6 @@ void cv::medianBlur( InputArray _src0, OutputArray _dst, int ksize ) CV_IPP_RUN_FAST(ipp_medianFilter(src0, dst, ksize)); -#ifdef HAVE_TEGRA_OPTIMIZATION - if (tegra::useTegra() && tegra::medianBlur(src0, dst, ksize)) - return; -#endif - bool useSortNet = ksize == 3 || (ksize == 5 #if !(CV_SIMD128) && ( src0.depth() > CV_8U || src0.channels() == 2 || src0.channels() > 4 ) @@ -4632,7 +5262,7 @@ public: { } - virtual void operator() (const Range& range) const + virtual void operator() (const Range& range) const CV_OVERRIDE { int i, j, cn = dest->channels(), k; Size size = dest->size(); @@ -4674,11 +5304,17 @@ public: color_weight[buf[2]], color_weight[buf[3]]); v_float32x4 _sw = v_load(space_weight+k); +#if defined(_MSC_VER) && _MSC_VER == 1700/* MSVS 2012 */ && CV_AVX + // details: https://github.com/opencv/opencv/issues/11004 + vsumw += _cw * _sw; + vsumc += _cw * _sw * _valF; +#else v_float32x4 _w = _cw * _sw; _cw = _w * _valF; vsumw += _w; vsumc += _cw; +#endif } float *bufFloat = (float*)buf; v_float32x4 sum4 = v_reduce_sum4(vsumw, vsumc, vsumw, vsumc); @@ -4695,6 +5331,7 @@ public: wsum += w; } // overflow is not possible here => there is no need to use cv::saturate_cast + CV_DbgAssert(fabs(wsum) > 0); dptr[j] = (uchar)cvRound(sum/wsum); } } @@ -4743,6 +5380,13 @@ public: color_weight[buf[2]],color_weight[buf[3]]); v_float32x4 _sw = v_load(space_weight+k); +#if defined(_MSC_VER) && _MSC_VER == 1700/* MSVS 2012 */ && CV_AVX + // details: https://github.com/opencv/opencv/issues/11004 + vsumw += _w * _sw; + vsumb += _w * _sw * _b; + vsumg += _w * _sw * _g; + vsumr += _w * _sw * _r; +#else _w *= _sw; _b *= _w; _g *= _w; @@ -4752,6 +5396,7 @@ public: vsumb += _b; vsumg += _g; vsumr += _r; +#endif } float *bufFloat = (float*)buf; v_float32x4 sum4 = v_reduce_sum4(vsumw, vsumb, vsumg, vsumr); @@ -4772,6 +5417,7 @@ public: sum_b += b*w; sum_g += g*w; sum_r += r*w; wsum += w; } + CV_DbgAssert(fabs(wsum) > 0); wsum = 1.f/wsum; b0 = cvRound(sum_b*wsum); g0 = cvRound(sum_g*wsum); @@ -4956,7 +5602,7 @@ public: { } - virtual void operator() (const Range& range) const + virtual void operator() (const Range& range) const CV_OVERRIDE { int i, j, k; Size size = dest->size(); @@ -5031,6 +5677,7 @@ public: sum += val*w; wsum += w; } + CV_DbgAssert(fabs(wsum) > 0); dptr[j] = (float)(sum/wsum); } } @@ -5121,6 +5768,7 @@ public: sum_b += b*w; sum_g += g*w; sum_r += r*w; wsum += w; } + CV_DbgAssert(fabs(wsum) > 0); wsum = 1.f/wsum; b0 = sum_b*wsum; g0 = sum_g*wsum; @@ -5249,7 +5897,7 @@ public: } ~ipp_bilateralFilterParallel() {} - virtual void operator() (const Range& range) const + virtual void operator() (const Range& range) const CV_OVERRIDE { if(*pOk == false) return;