// */
#include "precomp.hpp"
-#include <iostream>
-#include <vector>
+#include "opencl_kernels.hpp"
#if defined (HAVE_IPP) && (IPP_VERSION_MAJOR >= 7)
static IppStatus sts = ippInit();
namespace cv
{
+#if defined (HAVE_IPP) && ((IPP_VERSION_MAJOR == 7 && IPP_VERSION_MINOR >= 1) || IPP_VERSION_MAJOR > 7)
+ typedef IppStatus (CV_STDCALL* ippiResizeFunc)(const void*, int, const void*, int, IppiPoint, IppiSize, IppiBorderType, void*, void*, Ipp8u*);
+#endif
#if defined (HAVE_IPP) && (IPP_VERSION_MAJOR >= 7)
typedef IppStatus (CV_STDCALL* ippiSetFunc)(const void*, void *, int, IppiSize);
typedef IppStatus (CV_STDCALL* ippiWarpPerspectiveBackFunc)(const void*, IppiSize, int, IppiRect, void *, int, IppiRect, double [3][3], int);
typedef IppStatus (CV_STDCALL* ippiWarpAffineBackFunc)(const void*, IppiSize, int, IppiRect, void *, int, IppiRect, double [2][3], int);
- typedef IppStatus (CV_STDCALL* ippiResizeSqrPixelFunc)(const void*, IppiSize, int, IppiRect, void*, int, IppiRect, double, double, double, double, int, Ipp8u *);
template <int channels, typename Type>
bool IPPSetSimple(cv::Scalar value, void *dataPointer, int step, IppiSize &size, ippiSetFunc func)
template <typename T, typename WT>
struct ResizeAreaFastNoVec
{
- ResizeAreaFastNoVec(int /*_scale_x*/, int /*_scale_y*/,
- int /*_cn*/, int /*_step*//*, const int**/ /*_ofs*/) { }
- int operator() (const T* /*S*/, T* /*D*/, int /*w*/) const { return 0; }
+ ResizeAreaFastNoVec(int, int) { }
+ ResizeAreaFastNoVec(int, int, int, int) { }
+ int operator() (const T*, T*, int) const
+ { return 0; }
};
-template<typename T>
+#if CV_SSE2
+class ResizeAreaFastVec_SIMD_8u
+{
+public:
+ ResizeAreaFastVec_SIMD_8u(int _cn, int _step) :
+ cn(_cn), step(_step)
+ {
+ use_simd = checkHardwareSupport(CV_CPU_SSE2);
+ }
+
+ int operator() (const uchar* S, uchar* D, int w) const
+ {
+ if (!use_simd)
+ return 0;
+
+ int dx = 0;
+ const uchar* S0 = S;
+ const uchar* S1 = S0 + step;
+ __m128i zero = _mm_setzero_si128();
+ __m128i delta2 = _mm_set1_epi16(2);
+
+ if (cn == 1)
+ {
+ __m128i masklow = _mm_set1_epi16(0x00ff);
+ for ( ; dx <= w - 8; dx += 8, S0 += 16, S1 += 16, D += 8)
+ {
+ __m128i r0 = _mm_loadu_si128((const __m128i*)S0);
+ __m128i r1 = _mm_loadu_si128((const __m128i*)S1);
+
+ __m128i s0 = _mm_add_epi16(_mm_srli_epi16(r0, 8), _mm_and_si128(r0, masklow));
+ __m128i s1 = _mm_add_epi16(_mm_srli_epi16(r1, 8), _mm_and_si128(r1, masklow));
+ s0 = _mm_add_epi16(_mm_add_epi16(s0, s1), delta2);
+ s0 = _mm_packus_epi16(_mm_srli_epi16(s0, 2), zero);
+
+ _mm_storel_epi64((__m128i*)D, s0);
+ }
+ }
+ else if (cn == 3)
+ for ( ; dx <= w - 11; dx += 6, S0 += 12, S1 += 12, D += 6)
+ {
+ __m128i r0 = _mm_loadu_si128((const __m128i*)S0);
+ __m128i r1 = _mm_loadu_si128((const __m128i*)S1);
+
+ __m128i r0_16l = _mm_unpacklo_epi8(r0, zero);
+ __m128i r0_16h = _mm_unpacklo_epi8(_mm_srli_si128(r0, 6), zero);
+ __m128i r1_16l = _mm_unpacklo_epi8(r1, zero);
+ __m128i r1_16h = _mm_unpacklo_epi8(_mm_srli_si128(r1, 6), zero);
+
+ __m128i s0 = _mm_add_epi16(r0_16l, _mm_srli_si128(r0_16l, 6));
+ __m128i s1 = _mm_add_epi16(r1_16l, _mm_srli_si128(r1_16l, 6));
+ s0 = _mm_add_epi16(s1, _mm_add_epi16(s0, delta2));
+ s0 = _mm_packus_epi16(_mm_srli_epi16(s0, 2), zero);
+ _mm_storel_epi64((__m128i*)D, s0);
+
+ s0 = _mm_add_epi16(r0_16h, _mm_srli_si128(r0_16h, 6));
+ s1 = _mm_add_epi16(r1_16h, _mm_srli_si128(r1_16h, 6));
+ s0 = _mm_add_epi16(s1, _mm_add_epi16(s0, delta2));
+ s0 = _mm_packus_epi16(_mm_srli_epi16(s0, 2), zero);
+ _mm_storel_epi64((__m128i*)(D+3), s0);
+ }
+ else
+ {
+ CV_Assert(cn == 4);
+ int v[] = { 0, 0, -1, -1 };
+ __m128i mask = _mm_loadu_si128((const __m128i*)v);
+
+ for ( ; dx <= w - 8; dx += 8, S0 += 16, S1 += 16, D += 8)
+ {
+ __m128i r0 = _mm_loadu_si128((const __m128i*)S0);
+ __m128i r1 = _mm_loadu_si128((const __m128i*)S1);
+
+ __m128i r0_16l = _mm_unpacklo_epi8(r0, zero);
+ __m128i r0_16h = _mm_unpackhi_epi8(r0, zero);
+ __m128i r1_16l = _mm_unpacklo_epi8(r1, zero);
+ __m128i r1_16h = _mm_unpackhi_epi8(r1, zero);
+
+ __m128i s0 = _mm_add_epi16(r0_16l, _mm_srli_si128(r0_16l, 8));
+ __m128i s1 = _mm_add_epi16(r1_16l, _mm_srli_si128(r1_16l, 8));
+ s0 = _mm_add_epi16(s1, _mm_add_epi16(s0, delta2));
+ __m128i res0 = _mm_srli_epi16(s0, 2);
+
+ s0 = _mm_add_epi16(r0_16h, _mm_srli_si128(r0_16h, 8));
+ s1 = _mm_add_epi16(r1_16h, _mm_srli_si128(r1_16h, 8));
+ s0 = _mm_add_epi16(s1, _mm_add_epi16(s0, delta2));
+ __m128i res1 = _mm_srli_epi16(s0, 2);
+ s0 = _mm_packus_epi16(_mm_or_si128(_mm_andnot_si128(mask, res0),
+ _mm_and_si128(mask, _mm_slli_si128(res1, 8))), zero);
+ _mm_storel_epi64((__m128i*)(D), s0);
+ }
+ }
+
+ return dx;
+ }
+
+private:
+ int cn;
+ bool use_simd;
+ int step;
+};
+
+class ResizeAreaFastVec_SIMD_16u
+{
+public:
+ ResizeAreaFastVec_SIMD_16u(int _cn, int _step) :
+ cn(_cn), step(_step)
+ {
+ use_simd = checkHardwareSupport(CV_CPU_SSE2);
+ }
+
+ int operator() (const ushort* S, ushort* D, int w) const
+ {
+ if (!use_simd)
+ return 0;
+
+ int dx = 0;
+ const ushort* S0 = (const ushort*)S;
+ const ushort* S1 = (const ushort*)((const uchar*)(S) + step);
+ __m128i masklow = _mm_set1_epi32(0x0000ffff);
+ __m128i zero = _mm_setzero_si128();
+ __m128i delta2 = _mm_set1_epi32(2);
+
+#define _mm_packus_epi32(a, zero) _mm_packs_epi32(_mm_srai_epi32(_mm_slli_epi32(a, 16), 16), zero)
+
+ if (cn == 1)
+ {
+ for ( ; dx <= w - 4; dx += 4, S0 += 8, S1 += 8, D += 4)
+ {
+ __m128i r0 = _mm_loadu_si128((const __m128i*)S0);
+ __m128i r1 = _mm_loadu_si128((const __m128i*)S1);
+
+ __m128i s0 = _mm_add_epi32(_mm_srli_epi32(r0, 16), _mm_and_si128(r0, masklow));
+ __m128i s1 = _mm_add_epi32(_mm_srli_epi32(r1, 16), _mm_and_si128(r1, masklow));
+ s0 = _mm_add_epi32(_mm_add_epi32(s0, s1), delta2);
+ s0 = _mm_srli_epi32(s0, 2);
+ s0 = _mm_packus_epi32(s0, zero);
+
+ _mm_storel_epi64((__m128i*)D, s0);
+ }
+ }
+ else if (cn == 3)
+ for ( ; dx <= w - 4; dx += 3, S0 += 6, S1 += 6, D += 3)
+ {
+ __m128i r0 = _mm_loadu_si128((const __m128i*)S0);
+ __m128i r1 = _mm_loadu_si128((const __m128i*)S1);
+
+ __m128i r0_16l = _mm_unpacklo_epi16(r0, zero);
+ __m128i r0_16h = _mm_unpacklo_epi16(_mm_srli_si128(r0, 6), zero);
+ __m128i r1_16l = _mm_unpacklo_epi16(r1, zero);
+ __m128i r1_16h = _mm_unpacklo_epi16(_mm_srli_si128(r1, 6), zero);
+
+ __m128i s0 = _mm_add_epi32(r0_16l, r0_16h);
+ __m128i s1 = _mm_add_epi32(r1_16l, r1_16h);
+ s0 = _mm_add_epi32(delta2, _mm_add_epi32(s0, s1));
+ s0 = _mm_packus_epi32(_mm_srli_epi32(s0, 2), zero);
+ _mm_storel_epi64((__m128i*)D, s0);
+ }
+ else
+ {
+ CV_Assert(cn == 4);
+ for ( ; dx <= w - 4; dx += 4, S0 += 8, S1 += 8, D += 4)
+ {
+ __m128i r0 = _mm_loadu_si128((const __m128i*)S0);
+ __m128i r1 = _mm_loadu_si128((const __m128i*)S1);
+
+ __m128i r0_32l = _mm_unpacklo_epi16(r0, zero);
+ __m128i r0_32h = _mm_unpackhi_epi16(r0, zero);
+ __m128i r1_32l = _mm_unpacklo_epi16(r1, zero);
+ __m128i r1_32h = _mm_unpackhi_epi16(r1, zero);
+
+ __m128i s0 = _mm_add_epi32(r0_32l, r0_32h);
+ __m128i s1 = _mm_add_epi32(r1_32l, r1_32h);
+ s0 = _mm_add_epi32(s1, _mm_add_epi32(s0, delta2));
+ s0 = _mm_packus_epi32(_mm_srli_epi32(s0, 2), zero);
+ _mm_storel_epi64((__m128i*)D, s0);
+ }
+ }
+
+#undef _mm_packus_epi32
+
+ return dx;
+ }
+
+private:
+ int cn;
+ int step;
+ bool use_simd;
+};
+
+#else
+typedef ResizeAreaFastNoVec<uchar, uchar> ResizeAreaFastVec_SIMD_8u;
+typedef ResizeAreaFastNoVec<ushort, ushort> ResizeAreaFastVec_SIMD_16u;
+#endif
+
+template<typename T, typename SIMDVecOp>
struct ResizeAreaFastVec
{
- ResizeAreaFastVec(int _scale_x, int _scale_y, int _cn, int _step/*, const int* _ofs*/) :
- scale_x(_scale_x), scale_y(_scale_y), cn(_cn), step(_step)/*, ofs(_ofs)*/
+ ResizeAreaFastVec(int _scale_x, int _scale_y, int _cn, int _step) :
+ scale_x(_scale_x), scale_y(_scale_y), cn(_cn), step(_step), vecOp(_cn, _step)
{
fast_mode = scale_x == 2 && scale_y == 2 && (cn == 1 || cn == 3 || cn == 4);
}
int operator() (const T* S, T* D, int w) const
{
- if( !fast_mode )
+ if (!fast_mode)
return 0;
const T* nextS = (const T*)((const uchar*)S + step);
- int dx = 0;
+ int dx = vecOp(S, D, w);
if (cn == 1)
for( ; dx < w; ++dx )
}
else
{
- assert(cn == 4);
+ CV_Assert(cn == 4);
for( ; dx < w; dx += 4 )
{
int index = dx*2;
int cn;
bool fast_mode;
int step;
+ SIMDVecOp vecOp;
};
template <typename T, typename WT, typename VecOp>
{
double fsx1 = dx * scale;
double fsx2 = fsx1 + scale;
- double cellWidth = min(scale, ssize - fsx1);
+ double cellWidth = std::min(scale, ssize - fsx1);
int sx1 = cvCeil(fsx1), sx2 = cvFloor(fsx2);
assert( k < ssize*2 );
tab[k].di = dx * cn;
tab[k].si = sx2 * cn;
- tab[k++].alpha = (float)(min(min(fsx2 - sx2, 1.), cellWidth) / cellWidth);
+ tab[k++].alpha = (float)(std::min(std::min(fsx2 - sx2, 1.), cellWidth) / cellWidth);
}
}
return k;
}
-#if defined (HAVE_IPP) && (IPP_VERSION_MAJOR >= 7)
+#if defined (HAVE_IPP) && ((IPP_VERSION_MAJOR == 7 && IPP_VERSION_MINOR >= 1) || IPP_VERSION_MAJOR > 7)
class IPPresizeInvoker :
public ParallelLoopBody
{
public:
- IPPresizeInvoker(Mat &_src, Mat &_dst, double &_inv_scale_x, double &_inv_scale_y, int _mode, ippiResizeSqrPixelFunc _func, bool *_ok) :
- ParallelLoopBody(), src(_src), dst(_dst), inv_scale_x(_inv_scale_x), inv_scale_y(_inv_scale_y), mode(_mode), func(_func), ok(_ok)
+ IPPresizeInvoker(Mat &_src, Mat &_dst, double _inv_scale_x, double _inv_scale_y, int _mode, bool *_ok) :
+ ParallelLoopBody(), src(_src), dst(_dst), inv_scale_x(_inv_scale_x), inv_scale_y(_inv_scale_y), mode(_mode), ok(_ok)
{
+ IppStatus status = ippStsNotSupportedModeErr;
*ok = true;
+ IppiSize srcSize, dstSize;
+ int type = src.type();
+ int specSize = 0, initSize = 0;
+ srcSize.width = src.cols;
+ srcSize.height = src.rows;
+ dstSize.width = dst.cols;
+ dstSize.height = dst.rows;
+
+ if (mode == (int)ippLinear)
+ {
+ func =
+ type == CV_8UC1 ? (ippiResizeFunc)ippiResizeLinear_8u_C1R :
+ type == CV_8UC3 ? (ippiResizeFunc)ippiResizeLinear_8u_C3R :
+ type == CV_8UC4 ? (ippiResizeFunc)ippiResizeLinear_8u_C4R :
+ type == CV_16UC1 ? (ippiResizeFunc)ippiResizeLinear_16u_C1R :
+ type == CV_16UC3 ? (ippiResizeFunc)ippiResizeLinear_16u_C3R :
+ type == CV_16UC4 ? (ippiResizeFunc)ippiResizeLinear_16u_C4R :
+ type == CV_16SC1 ? (ippiResizeFunc)ippiResizeLinear_16s_C1R :
+ type == CV_16SC3 ? (ippiResizeFunc)ippiResizeLinear_16s_C3R :
+ type == CV_16SC4 ? (ippiResizeFunc)ippiResizeLinear_16s_C4R :
+ type == CV_32FC1 ? (ippiResizeFunc)ippiResizeLinear_32f_C1R :
+ type == CV_32FC3 ? (ippiResizeFunc)ippiResizeLinear_32f_C3R :
+ type == CV_32FC4 ? (ippiResizeFunc)ippiResizeLinear_32f_C4R :
+ type == CV_64FC1 ? (ippiResizeFunc)ippiResizeLinear_64f_C1R :
+ type == CV_64FC3 ? (ippiResizeFunc)ippiResizeLinear_64f_C3R :
+ type == CV_64FC4 ? (ippiResizeFunc)ippiResizeLinear_64f_C4R :
+ 0;
+ }
+ else if (mode == (int)ippCubic)
+ {
+ func =
+ type == CV_8UC1 ? (ippiResizeFunc)ippiResizeCubic_8u_C1R :
+ type == CV_8UC3 ? (ippiResizeFunc)ippiResizeCubic_8u_C3R :
+ type == CV_8UC4 ? (ippiResizeFunc)ippiResizeCubic_8u_C4R :
+ type == CV_16UC1 ? (ippiResizeFunc)ippiResizeCubic_16u_C1R :
+ type == CV_16UC3 ? (ippiResizeFunc)ippiResizeCubic_16u_C3R :
+ type == CV_16UC4 ? (ippiResizeFunc)ippiResizeCubic_16u_C4R :
+ type == CV_16SC1 ? (ippiResizeFunc)ippiResizeCubic_16s_C1R :
+ type == CV_16SC3 ? (ippiResizeFunc)ippiResizeCubic_16s_C3R :
+ type == CV_16SC4 ? (ippiResizeFunc)ippiResizeCubic_16s_C4R :
+ type == CV_32FC1 ? (ippiResizeFunc)ippiResizeCubic_32f_C1R :
+ type == CV_32FC3 ? (ippiResizeFunc)ippiResizeCubic_32f_C3R :
+ type == CV_32FC4 ? (ippiResizeFunc)ippiResizeCubic_32f_C4R :
+ 0;
+ }
+
+ switch (src.depth())
+ {
+ case CV_8U:
+ status = ippiResizeGetSize_8u(srcSize, dstSize, (IppiInterpolationType)mode, 0, &specSize, &initSize);
+ break;
+ case CV_16U:
+ status = ippiResizeGetSize_16u(srcSize, dstSize, (IppiInterpolationType)mode, 0, &specSize, &initSize);
+ break;
+ case CV_16S:
+ status = ippiResizeGetSize_16s(srcSize, dstSize, (IppiInterpolationType)mode, 0, &specSize, &initSize);
+ break;
+ case CV_32F:
+ status = ippiResizeGetSize_32f(srcSize, dstSize, (IppiInterpolationType)mode, 0, &specSize, &initSize);
+ break;
+ case CV_64F:
+ status = ippiResizeGetSize_64f(srcSize, dstSize, (IppiInterpolationType)mode, 0, &specSize, &initSize);
+ break;
+ }
+ if (status != ippStsNoErr)
+ {
+ *ok = false;
+ return;
+ }
+
+ specBuf.allocate(specSize);
+ pSpec = (uchar*)specBuf;
+
+ status = ippStsNotSupportedModeErr;
+ if (mode == (int)ippLinear)
+ {
+ switch (src.depth())
+ {
+ case CV_8U:
+ status = ippiResizeLinearInit_8u(srcSize, dstSize, (IppiResizeSpec_32f*)pSpec);
+ break;
+ case CV_16U:
+ status = ippiResizeLinearInit_16u(srcSize, dstSize, (IppiResizeSpec_32f*)pSpec);
+ break;
+ case CV_16S:
+ status = ippiResizeLinearInit_16s(srcSize, dstSize, (IppiResizeSpec_32f*)pSpec);
+ break;
+ case CV_32F:
+ status = ippiResizeLinearInit_32f(srcSize, dstSize, (IppiResizeSpec_32f*)pSpec);
+ break;
+ case CV_64F:
+ status = ippiResizeLinearInit_64f(srcSize, dstSize, (IppiResizeSpec_64f*)pSpec);
+ break;
+ }
+ if (status != ippStsNoErr)
+ {
+ *ok = false;
+ return;
+ }
+ }
+ else if (mode == (int)ippCubic)
+ {
+ AutoBuffer<uchar> buf(initSize);
+ uchar* pInit = (uchar*)buf;
+
+ switch (src.depth())
+ {
+ case CV_8U:
+ status = ippiResizeCubicInit_8u(srcSize, dstSize, 0.f, 0.75f, (IppiResizeSpec_32f*)pSpec, pInit);
+ break;
+ case CV_16U:
+ status = ippiResizeCubicInit_16u(srcSize, dstSize, 0.f, 0.75f, (IppiResizeSpec_32f*)pSpec, pInit);
+ break;
+ case CV_16S:
+ status = ippiResizeCubicInit_16s(srcSize, dstSize, 0.f, 0.75f, (IppiResizeSpec_32f*)pSpec, pInit);
+ break;
+ case CV_32F:
+ status = ippiResizeCubicInit_32f(srcSize, dstSize, 0.f, 0.75f, (IppiResizeSpec_32f*)pSpec, pInit);
+ break;
+ }
+ if (status != ippStsNoErr) *ok = false;
+ }
+ }
+
+ ~IPPresizeInvoker()
+ {
}
virtual void operator() (const Range& range) const
{
+ if (*ok == false) return;
+
int cn = src.channels();
- IppiRect srcroi = { 0, range.start, src.cols, range.end - range.start };
int dsty = CV_IMIN(cvRound(range.start * inv_scale_y), dst.rows);
- int dstwidth = CV_IMIN(cvRound(src.cols * inv_scale_x), dst.cols);
+ int dstwidth = CV_IMIN(cvRound(src.cols * inv_scale_x), dst.cols);
int dstheight = CV_IMIN(cvRound(range.end * inv_scale_y), dst.rows);
- IppiRect dstroi = { 0, dsty, dstwidth, dstheight - dsty };
- int bufsize;
- ippiResizeGetBufSize( srcroi, dstroi, cn, mode, &bufsize );
+
+ IppiPoint dstOffset = { 0, dsty }, srcOffset = {0, 0};
+ IppiSize dstSize = { dstwidth, dstheight - dsty };
+ int bufsize = 0, itemSize = 0;
+
+ IppStatus status = ippStsNotSupportedModeErr;
+
+ switch (src.depth())
+ {
+ case CV_8U:
+ itemSize = 1;
+ status = ippiResizeGetBufferSize_8u((IppiResizeSpec_32f*)pSpec, dstSize, cn, &bufsize);
+ if (status == ippStsNoErr)
+ status = ippiResizeGetSrcOffset_8u((IppiResizeSpec_32f*)pSpec, dstOffset, &srcOffset);
+ break;
+ case CV_16U:
+ itemSize = 2;
+ status = ippiResizeGetBufferSize_16u((IppiResizeSpec_32f*)pSpec, dstSize, cn, &bufsize);
+ if (status == ippStsNoErr)
+ status = ippiResizeGetSrcOffset_16u((IppiResizeSpec_32f*)pSpec, dstOffset, &srcOffset);
+ break;
+ case CV_16S:
+ itemSize = 2;
+ status = ippiResizeGetBufferSize_16s((IppiResizeSpec_32f*)pSpec, dstSize, cn, &bufsize);
+ if (status == ippStsNoErr)
+ status = ippiResizeGetSrcOffset_16s((IppiResizeSpec_32f*)pSpec, dstOffset, &srcOffset);
+ break;
+ case CV_32F:
+ itemSize = 4;
+ status = ippiResizeGetBufferSize_32f((IppiResizeSpec_32f*)pSpec, dstSize, cn, &bufsize);
+ if (status == ippStsNoErr)
+ status = ippiResizeGetSrcOffset_32f((IppiResizeSpec_32f*)pSpec, dstOffset, &srcOffset);
+ break;
+ case CV_64F:
+ itemSize = 4;
+ status = ippiResizeGetBufferSize_64f((IppiResizeSpec_64f*)pSpec, dstSize, cn, &bufsize);
+ if (status == ippStsNoErr)
+ status = ippiResizeGetSrcOffset_64f((IppiResizeSpec_64f*)pSpec, dstOffset, &srcOffset);
+ break;
+ }
+ if (status != ippStsNoErr)
+ {
+ *ok = false;
+ return;
+ }
+
+ Ipp8u* pSrc = (Ipp8u*)src.data + (int)src.step[0] * srcOffset.y + srcOffset.x * cn * itemSize;
+ Ipp8u* pDst = (Ipp8u*)dst.data + (int)dst.step[0] * dstOffset.y + dstOffset.x * cn * itemSize;
+
AutoBuffer<uchar> buf(bufsize + 64);
uchar* bufptr = alignPtr((uchar*)buf, 32);
- if( func( src.data, ippiSize(src.cols, src.rows), (int)src.step[0], srcroi, dst.data, (int)dst.step[0], dstroi, inv_scale_x, inv_scale_y, 0, 0, mode, bufptr ) < 0 )
+ if( func( pSrc, (int)src.step[0], pDst, (int)dst.step[0], dstOffset, dstSize, ippBorderRepl, 0, pSpec, bufptr ) <= 0 )
*ok = false;
}
private:
Mat &dst;
double inv_scale_x;
double inv_scale_y;
+ void *pSpec;
+ AutoBuffer<uchar> specBuf;
int mode;
- ippiResizeSqrPixelFunc func;
+ ippiResizeFunc func;
bool *ok;
const IPPresizeInvoker& operator= (const IPPresizeInvoker&);
};
#endif
+#ifdef HAVE_OPENCL
+
+static void ocl_computeResizeAreaTabs(int ssize, int dsize, double scale, int * const map_tab,
+ float * const alpha_tab, int * const ofs_tab)
+{
+ int k = 0, dx = 0;
+ for ( ; dx < dsize; dx++)
+ {
+ ofs_tab[dx] = k;
+
+ double fsx1 = dx * scale;
+ double fsx2 = fsx1 + scale;
+ double cellWidth = std::min(scale, ssize - fsx1);
+
+ int sx1 = cvCeil(fsx1), sx2 = cvFloor(fsx2);
+
+ sx2 = std::min(sx2, ssize - 1);
+ sx1 = std::min(sx1, sx2);
+
+ if (sx1 - fsx1 > 1e-3)
+ {
+ map_tab[k] = sx1 - 1;
+ alpha_tab[k++] = (float)((sx1 - fsx1) / cellWidth);
+ }
+
+ for (int sx = sx1; sx < sx2; sx++)
+ {
+ map_tab[k] = sx;
+ alpha_tab[k++] = float(1.0 / cellWidth);
+ }
+
+ if (fsx2 - sx2 > 1e-3)
+ {
+ map_tab[k] = sx2;
+ alpha_tab[k++] = (float)(std::min(std::min(fsx2 - sx2, 1.), cellWidth) / cellWidth);
+ }
+ }
+ ofs_tab[dx] = k;
+}
+
+static void ocl_computeResizeAreaFastTabs(int * dmap_tab, int * smap_tab, int scale, int dcols, int scol)
+{
+ for (int i = 0; i < dcols; ++i)
+ dmap_tab[i] = scale * i;
+
+ for (int i = 0, size = dcols * scale; i < size; ++i)
+ smap_tab[i] = std::min(scol - 1, i);
}
+static bool ocl_resize( InputArray _src, OutputArray _dst, Size dsize,
+ double fx, double fy, int interpolation)
+{
+ int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+
+ double inv_fx = 1. / fx, inv_fy = 1. / fy;
+ float inv_fxf = (float)inv_fx, inv_fyf = (float)inv_fy;
+
+ if( !(cn <= 4 &&
+ (interpolation == INTER_NEAREST || interpolation == INTER_LINEAR ||
+ (interpolation == INTER_AREA && inv_fx >= 1 && inv_fy >= 1) )) )
+ return false;
+
+ UMat src = _src.getUMat();
+ _dst.create(dsize, type);
+ UMat dst = _dst.getUMat();
+
+ ocl::Kernel k;
+ size_t globalsize[] = { dst.cols, dst.rows };
+
+ if (interpolation == INTER_LINEAR)
+ {
+ int wdepth = std::max(depth, CV_32S);
+ int wtype = CV_MAKETYPE(wdepth, cn);
+ char buf[2][32];
+ k.create("resizeLN", ocl::imgproc::resize_oclsrc,
+ format("-D INTER_LINEAR -D depth=%d -D PIXTYPE=%s -D PIXTYPE1=%s "
+ "-D WORKTYPE=%s -D convertToWT=%s -D convertToDT=%s -D cn=%d",
+ depth, ocl::typeToStr(type), ocl::typeToStr(depth), ocl::typeToStr(wtype),
+ ocl::convertTypeStr(depth, wdepth, cn, buf[0]),
+ ocl::convertTypeStr(wdepth, depth, cn, buf[1]),
+ cn));
+ }
+ else if (interpolation == INTER_NEAREST)
+ {
+ k.create("resizeNN", ocl::imgproc::resize_oclsrc,
+ format("-D INTER_NEAREST -D PIXTYPE=%s -D PIXTYPE1=%s -D cn=%d",
+ ocl::memopTypeToStr(type), ocl::memopTypeToStr(depth), cn));
+ }
+ else if (interpolation == INTER_AREA)
+ {
+ int iscale_x = saturate_cast<int>(inv_fx);
+ int iscale_y = saturate_cast<int>(inv_fy);
+ bool is_area_fast = std::abs(inv_fx - iscale_x) < DBL_EPSILON &&
+ std::abs(inv_fy - iscale_y) < DBL_EPSILON;
+ int wdepth = std::max(depth, is_area_fast ? CV_32S : CV_32F);
+ int wtype = CV_MAKE_TYPE(wdepth, cn);
+
+ char cvt[2][40];
+ String buildOption = format("-D INTER_AREA -D PIXTYPE=%s -D PIXTYPE1=%s -D WTV=%s -D convertToWTV=%s -D cn=%d",
+ ocl::typeToStr(type), ocl::typeToStr(depth), ocl::typeToStr(wtype),
+ ocl::convertTypeStr(depth, wdepth, cn, cvt[0]), cn);
+
+ UMat alphaOcl, tabofsOcl, mapOcl;
+ UMat dmap, smap;
+
+ if (is_area_fast)
+ {
+ int wdepth2 = std::max(CV_32F, depth), wtype2 = CV_MAKE_TYPE(wdepth2, cn);
+ buildOption = buildOption + format(" -D convertToPIXTYPE=%s -D WT2V=%s -D convertToWT2V=%s -D INTER_AREA_FAST"
+ " -D XSCALE=%d -D YSCALE=%d -D SCALE=%ff",
+ ocl::convertTypeStr(wdepth2, depth, cn, cvt[0]),
+ ocl::typeToStr(wtype2), ocl::convertTypeStr(wdepth, wdepth2, cn, cvt[1]),
+ iscale_x, iscale_y, 1.0f / (iscale_x * iscale_y));
+
+ k.create("resizeAREA_FAST", ocl::imgproc::resize_oclsrc, buildOption);
+ if (k.empty())
+ return false;
+
+ int smap_tab_size = dst.cols * iscale_x + dst.rows * iscale_y;
+ AutoBuffer<int> dmap_tab(dst.cols + dst.rows), smap_tab(smap_tab_size);
+ int * dxmap_tab = dmap_tab, * dymap_tab = dxmap_tab + dst.cols;
+ int * sxmap_tab = smap_tab, * symap_tab = smap_tab + dst.cols * iscale_y;
+
+ ocl_computeResizeAreaFastTabs(dxmap_tab, sxmap_tab, iscale_x, dst.cols, src.cols);
+ ocl_computeResizeAreaFastTabs(dymap_tab, symap_tab, iscale_y, dst.rows, src.rows);
+
+ Mat(1, dst.cols + dst.rows, CV_32SC1, (void *)dmap_tab).copyTo(dmap);
+ Mat(1, smap_tab_size, CV_32SC1, (void *)smap_tab).copyTo(smap);
+ }
+ else
+ {
+ buildOption = buildOption + format(" -D convertToPIXTYPE=%s", ocl::convertTypeStr(wdepth, depth, cn, cvt[0]));
+ k.create("resizeAREA", ocl::imgproc::resize_oclsrc, buildOption);
+ if (k.empty())
+ return false;
+
+ Size ssize = src.size();
+ int xytab_size = (ssize.width + ssize.height) << 1;
+ int tabofs_size = dsize.height + dsize.width + 2;
+
+ AutoBuffer<int> _xymap_tab(xytab_size), _xyofs_tab(tabofs_size);
+ AutoBuffer<float> _xyalpha_tab(xytab_size);
+ int * xmap_tab = _xymap_tab, * ymap_tab = _xymap_tab + (ssize.width << 1);
+ float * xalpha_tab = _xyalpha_tab, * yalpha_tab = _xyalpha_tab + (ssize.width << 1);
+ int * xofs_tab = _xyofs_tab, * yofs_tab = _xyofs_tab + dsize.width + 1;
+
+ ocl_computeResizeAreaTabs(ssize.width, dsize.width, inv_fx, xmap_tab, xalpha_tab, xofs_tab);
+ ocl_computeResizeAreaTabs(ssize.height, dsize.height, inv_fy, ymap_tab, yalpha_tab, yofs_tab);
+
+ // loading precomputed arrays to GPU
+ Mat(1, xytab_size, CV_32FC1, (void *)_xyalpha_tab).copyTo(alphaOcl);
+ Mat(1, xytab_size, CV_32SC1, (void *)_xymap_tab).copyTo(mapOcl);
+ Mat(1, tabofs_size, CV_32SC1, (void *)_xyofs_tab).copyTo(tabofsOcl);
+ }
+
+ ocl::KernelArg srcarg = ocl::KernelArg::ReadOnly(src), dstarg = ocl::KernelArg::WriteOnly(dst);
+
+ if (is_area_fast)
+ k.args(srcarg, dstarg, ocl::KernelArg::PtrReadOnly(dmap), ocl::KernelArg::PtrReadOnly(smap));
+ else
+ k.args(srcarg, dstarg, inv_fxf, inv_fyf, ocl::KernelArg::PtrReadOnly(tabofsOcl),
+ ocl::KernelArg::PtrReadOnly(mapOcl), ocl::KernelArg::PtrReadOnly(alphaOcl));
+
+ return k.run(2, globalsize, NULL, false);
+ }
+
+ if( k.empty() )
+ return false;
+ k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst),
+ (float)inv_fx, (float)inv_fy);
+
+ return k.run(2, globalsize, 0, false);
+}
+
+#endif
+
+}
//////////////////////////////////////////////////////////////////////////////////////////
static ResizeAreaFastFunc areafast_tab[] =
{
- resizeAreaFast_<uchar, int, ResizeAreaFastVec<uchar> >,
+ resizeAreaFast_<uchar, int, ResizeAreaFastVec<uchar, ResizeAreaFastVec_SIMD_8u> >,
0,
- resizeAreaFast_<ushort, float, ResizeAreaFastVec<ushort> >,
- resizeAreaFast_<short, float, ResizeAreaFastVec<short> >,
+ resizeAreaFast_<ushort, float, ResizeAreaFastVec<ushort, ResizeAreaFastVec_SIMD_16u> >,
+ resizeAreaFast_<short, float, ResizeAreaFastVec<short, ResizeAreaFastNoVec<short, float> > >,
0,
resizeAreaFast_<float, float, ResizeAreaFastNoVec<float, float> >,
resizeAreaFast_<double, double, ResizeAreaFastNoVec<double, double> >,
resizeArea_<double, double>, 0
};
- Mat src = _src.getMat();
- Size ssize = src.size();
+ Size ssize = _src.size();
CV_Assert( ssize.area() > 0 );
- CV_Assert( dsize.area() || (inv_scale_x > 0 && inv_scale_y > 0) );
- if( !dsize.area() )
+ CV_Assert( dsize.area() > 0 || (inv_scale_x > 0 && inv_scale_y > 0) );
+ if( dsize.area() == 0 )
{
- dsize = Size(saturate_cast<int>(src.cols*inv_scale_x),
- saturate_cast<int>(src.rows*inv_scale_y));
- CV_Assert( dsize.area() );
+ dsize = Size(saturate_cast<int>(ssize.width*inv_scale_x),
+ saturate_cast<int>(ssize.height*inv_scale_y));
+ CV_Assert( dsize.area() > 0 );
}
else
{
- inv_scale_x = (double)dsize.width/src.cols;
- inv_scale_y = (double)dsize.height/src.rows;
+ inv_scale_x = (double)dsize.width/ssize.width;
+ inv_scale_y = (double)dsize.height/ssize.height;
}
+
+ CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
+ ocl_resize(_src, _dst, dsize, inv_scale_x, inv_scale_y, interpolation))
+
+ Mat src = _src.getMat();
_dst.create(dsize, src.type());
Mat dst = _dst.getMat();
-
#ifdef HAVE_TEGRA_OPTIMIZATION
if (tegra::resize(src, dst, (float)inv_scale_x, (float)inv_scale_y, interpolation))
return;
int depth = src.depth(), cn = src.channels();
double scale_x = 1./inv_scale_x, scale_y = 1./inv_scale_y;
int k, sx, sy, dx, dy;
-/*
-#if defined (HAVE_IPP) && (IPP_VERSION_MAJOR >= 7)
- int mode = interpolation == INTER_LINEAR ? IPPI_INTER_LINEAR : 0;
- int type = src.type();
- ippiResizeSqrPixelFunc ippFunc =
- type == CV_8UC1 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_8u_C1R :
- type == CV_8UC3 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_8u_C3R :
- type == CV_8UC4 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_8u_C4R :
- type == CV_16UC1 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_16u_C1R :
- type == CV_16UC3 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_16u_C3R :
- type == CV_16UC4 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_16u_C4R :
- type == CV_16SC1 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_16s_C1R :
- type == CV_16SC3 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_16s_C3R :
- type == CV_16SC4 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_16s_C4R :
- type == CV_32FC1 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_32f_C1R :
- type == CV_32FC3 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_32f_C3R :
- type == CV_32FC4 ? (ippiResizeSqrPixelFunc)ippiResizeSqrPixel_32f_C4R :
- 0;
- if( ippFunc && mode != 0 )
+
+#if defined (HAVE_IPP) && ((IPP_VERSION_MAJOR == 7 && IPP_VERSION_MINOR >= 1) || IPP_VERSION_MAJOR > 7)
+#define IPP_RESIZE_EPS 1.e-10
+
+ double ex = fabs((double)dsize.width/src.cols - inv_scale_x)/inv_scale_x;
+ double ey = fabs((double)dsize.height/src.rows - inv_scale_y)/inv_scale_y;
+
+ if ((ex < IPP_RESIZE_EPS && ey < IPP_RESIZE_EPS && depth != CV_64F) ||
+ (ex == 0 && ey == 0 && depth == CV_64F))
{
- bool ok;
- Range range(0, src.rows);
- IPPresizeInvoker invoker(src, dst, inv_scale_x, inv_scale_y, mode, ippFunc, &ok);
- parallel_for_(range, invoker, dst.total()/(double)(1<<16));
- if( ok )
- return;
+ int mode = 0;
+ if (interpolation == INTER_LINEAR && src.rows >= 2 && src.cols >= 2)
+ {
+ mode = ippLinear;
+ }
+ else if (interpolation == INTER_CUBIC && src.rows >= 4 && src.cols >= 4)
+ {
+ mode = ippCubic;
+ }
+ if( mode != 0 && (cn == 1 || cn ==3 || cn == 4) &&
+ (depth == CV_8U || depth == CV_16U || depth == CV_16S || depth == CV_32F ||
+ (depth == CV_64F && mode == ippLinear)))
+ {
+ bool ok = true;
+ Range range(0, src.rows);
+ IPPresizeInvoker invoker(src, dst, inv_scale_x, inv_scale_y, mode, &ok);
+ parallel_for_(range, invoker, dst.total()/(double)(1<<16));
+ if( ok )
+ return;
+ }
}
+#undef IPP_RESIZE_EPS
#endif
-*/
+
if( interpolation == INTER_NEAREST )
{
resizeNN( src, dst, inv_scale_x, inv_scale_y );
// in case of scale_x && scale_y is equal to 2
// INTER_AREA (fast) also is equal to INTER_LINEAR
if( interpolation == INTER_LINEAR && is_area_fast && iscale_x == 2 && iscale_y == 2 )
- {
interpolation = INTER_AREA;
- }
// true "area" interpolation is only implemented for the case (scale_x <= 1 && scale_y <= 1).
// In other cases it is emulated using some variant of bilinear interpolation
if( sx < ksize2-1 )
{
xmin = dx+1;
- if( sx < 0 )
+ if( sx < 0 && (interpolation != INTER_CUBIC && interpolation != INTER_LANCZOS4))
fx = 0, sx = 0;
}
if( sx + ksize2 >= ssize.width )
{
xmax = std::min( xmax, dx );
- if( sx >= ssize.width-1 )
+ if( sx >= ssize.width-1 && (interpolation != INTER_CUBIC && interpolation != INTER_LANCZOS4))
fx = 0, sx = ssize.width-1;
}
int operator()( const Mat& _src, void* _dst, const short* XY,
const ushort* FXY, const void* _wtab, int width ) const
{
- int cn = _src.channels();
+ int cn = _src.channels(), x = 0, sstep = (int)_src.step;
- if( (cn != 1 && cn != 3 && cn != 4) || !checkHardwareSupport(CV_CPU_SSE2) )
+ if( (cn != 1 && cn != 3 && cn != 4) || !checkHardwareSupport(CV_CPU_SSE2)||
+ sstep > 0x8000 )
return 0;
const uchar *S0 = _src.data, *S1 = _src.data + _src.step;
const short* wtab = cn == 1 ? (const short*)_wtab : &BilinearTab_iC4[0][0][0];
uchar* D = (uchar*)_dst;
- int x = 0, sstep = (int)_src.step;
__m128i delta = _mm_set1_epi32(INTER_REMAP_COEF_SCALE/2);
__m128i xy2ofs = _mm_set1_epi32(cn + (sstep << 16));
__m128i z = _mm_setzero_si128();
{
public:
RemapInvoker(const Mat& _src, Mat& _dst, const Mat *_m1,
- const Mat *_m2, int _interpolation, int _borderType, const Scalar &_borderValue,
+ const Mat *_m2, int _borderType, const Scalar &_borderValue,
int _planar_input, RemapNNFunc _nnfunc, RemapFunc _ifunc, const void *_ctab) :
ParallelLoopBody(), src(&_src), dst(&_dst), m1(_m1), m2(_m2),
- interpolation(_interpolation), borderType(_borderType), borderValue(_borderValue),
+ borderType(_borderType), borderValue(_borderValue),
planar_input(_planar_input), nnfunc(_nnfunc), ifunc(_ifunc), ctab(_ctab)
{
}
const Mat* src;
Mat* dst;
const Mat *m1, *m2;
- int interpolation, borderType;
+ int borderType;
Scalar borderValue;
int planar_input;
RemapNNFunc nnfunc;
const void *ctab;
};
+#ifdef HAVE_OPENCL
+
+static bool ocl_remap(InputArray _src, OutputArray _dst, InputArray _map1, InputArray _map2,
+ int interpolation, int borderType, const Scalar& borderValue)
+{
+ int cn = _src.channels(), type = _src.type(), depth = _src.depth();
+
+ if (borderType == BORDER_TRANSPARENT || cn == 3 || !(interpolation == INTER_LINEAR || interpolation == INTER_NEAREST)
+ || _map1.type() == CV_16SC1 || _map2.type() == CV_16SC1)
+ return false;
+
+ UMat src = _src.getUMat(), map1 = _map1.getUMat(), map2 = _map2.getUMat();
+
+ if( (map1.type() == CV_16SC2 && (map2.type() == CV_16UC1 || map2.empty())) ||
+ (map2.type() == CV_16SC2 && (map1.type() == CV_16UC1 || map1.empty())) )
+ {
+ if (map1.type() != CV_16SC2)
+ std::swap(map1, map2);
+ }
+ else
+ CV_Assert( map1.type() == CV_32FC2 || (map1.type() == CV_32FC1 && map2.type() == CV_32FC1) );
+
+ _dst.create(map1.size(), type);
+ UMat dst = _dst.getUMat();
+
+ String kernelName = "remap";
+ if (map1.type() == CV_32FC2 && map2.empty())
+ kernelName += "_32FC2";
+ else if (map1.type() == CV_16SC2)
+ {
+ kernelName += "_16SC2";
+ if (!map2.empty())
+ kernelName += "_16UC1";
+ }
+ else if (map1.type() == CV_32FC1 && map2.type() == CV_32FC1)
+ kernelName += "_2_32FC1";
+ else
+ CV_Error(Error::StsBadArg, "Unsupported map types");
+
+ static const char * const interMap[] = { "INTER_NEAREST", "INTER_LINEAR", "INTER_CUBIC", "INTER_LINEAR", "INTER_LANCZOS" };
+ static const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", "BORDER_WRAP",
+ "BORDER_REFLECT_101", "BORDER_TRANSPARENT" };
+ String buildOptions = format("-D %s -D %s -D T=%s", interMap[interpolation], borderMap[borderType], ocl::typeToStr(type));
+
+ if (interpolation != INTER_NEAREST)
+ {
+ char cvt[3][40];
+ int wdepth = std::max(CV_32F, dst.depth());
+ buildOptions = buildOptions
+ + format(" -D WT=%s -D convertToT=%s -D convertToWT=%s"
+ " -D convertToWT2=%s -D WT2=%s",
+ ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)),
+ ocl::convertTypeStr(wdepth, depth, cn, cvt[0]),
+ ocl::convertTypeStr(depth, wdepth, cn, cvt[1]),
+ ocl::convertTypeStr(CV_32S, wdepth, 2, cvt[2]),
+ ocl::typeToStr(CV_MAKE_TYPE(wdepth, 2)));
+ }
+
+ ocl::Kernel k(kernelName.c_str(), ocl::imgproc::remap_oclsrc, buildOptions);
+
+ Mat scalar(1, 1, type, borderValue);
+ ocl::KernelArg srcarg = ocl::KernelArg::ReadOnly(src), dstarg = ocl::KernelArg::WriteOnly(dst),
+ map1arg = ocl::KernelArg::ReadOnlyNoSize(map1),
+ scalararg = ocl::KernelArg::Constant((void*)scalar.data, scalar.elemSize());
+
+ if (map2.empty())
+ k.args(srcarg, dstarg, map1arg, scalararg);
+ else
+ k.args(srcarg, dstarg, map1arg, ocl::KernelArg::ReadOnlyNoSize(map2), scalararg);
+
+ size_t globalThreads[2] = { dst.cols, dst.rows };
+ return k.run(2, globalThreads, NULL, false);
+}
+
+#endif
+
}
void cv::remap( InputArray _src, OutputArray _dst,
remapLanczos4<Cast<double, double>, float, 1>, 0
};
- Mat src = _src.getMat(), map1 = _map1.getMat(), map2 = _map2.getMat();
+ CV_Assert( _map1.size().area() > 0 );
+ CV_Assert( _map2.empty() || (_map2.size() == _map1.size()));
- CV_Assert( map1.size().area() > 0 );
- CV_Assert( !map2.data || (map2.size() == map1.size()));
+ CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
+ ocl_remap(_src, _dst, _map1, _map2, interpolation, borderType, borderValue))
+ Mat src = _src.getMat(), map1 = _map1.getMat(), map2 = _map2.getMat();
_dst.create( map1.size(), src.type() );
Mat dst = _dst.getMat();
if( dst.data == src.data )
planar_input = map1.channels() == 1;
}
- RemapInvoker invoker(src, dst, m1, m2, interpolation,
+ RemapInvoker invoker(src, dst, m1, m2,
borderType, borderValue, planar_input, nnfunc, ifunc,
ctab);
parallel_for_(Range(0, dst.rows), invoker, dst.total()/(double)(1<<16));
};
#endif
+#ifdef HAVE_OPENCL
+
+enum { OCL_OP_PERSPECTIVE = 1, OCL_OP_AFFINE = 0 };
+
+static bool ocl_warpTransform(InputArray _src, OutputArray _dst, InputArray _M0,
+ Size dsize, int flags, int borderType, const Scalar& borderValue,
+ int op_type)
+{
+ CV_Assert(op_type == OCL_OP_AFFINE || op_type == OCL_OP_PERSPECTIVE);
+
+ int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
+ double doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
+
+ int interpolation = flags & INTER_MAX;
+ if( interpolation == INTER_AREA )
+ interpolation = INTER_LINEAR;
+
+ if ( !(borderType == cv::BORDER_CONSTANT &&
+ (interpolation == cv::INTER_NEAREST || interpolation == cv::INTER_LINEAR || interpolation == cv::INTER_CUBIC)) ||
+ (!doubleSupport && depth == CV_64F) || cn > 4)
+ return false;
+
+ const char * const interpolationMap[3] = { "NEAREST", "LINEAR", "CUBIC" };
+ ocl::ProgramSource program = op_type == OCL_OP_AFFINE ?
+ ocl::imgproc::warp_affine_oclsrc : ocl::imgproc::warp_perspective_oclsrc;
+ const char * const kernelName = op_type == OCL_OP_AFFINE ? "warpAffine" : "warpPerspective";
+
+ int scalarcn = cn == 3 ? 4 : cn;
+ int wdepth = interpolation == INTER_NEAREST ? depth : std::max(CV_32S, depth);
+ int sctype = CV_MAKETYPE(wdepth, scalarcn);
+
+ ocl::Kernel k;
+ String opts;
+ if (interpolation == INTER_NEAREST)
+ {
+ opts = format("-D INTER_NEAREST -D T=%s%s -D T1=%s -D ST=%s -D cn=%d", ocl::typeToStr(type),
+ doubleSupport ? " -D DOUBLE_SUPPORT" : "",
+ ocl::typeToStr(CV_MAT_DEPTH(type)),
+ ocl::typeToStr(sctype),
+ cn);
+ }
+ else
+ {
+ char cvt[2][50];
+ opts = format("-D INTER_%s -D T=%s -D T1=%s -D ST=%s -D WT=%s -D depth=%d -D convertToWT=%s -D convertToT=%s%s -D cn=%d",
+ interpolationMap[interpolation], ocl::typeToStr(type),
+ ocl::typeToStr(CV_MAT_DEPTH(type)),
+ ocl::typeToStr(sctype),
+ ocl::typeToStr(CV_MAKE_TYPE(wdepth, cn)), depth,
+ ocl::convertTypeStr(depth, wdepth, cn, cvt[0]),
+ ocl::convertTypeStr(wdepth, depth, cn, cvt[1]),
+ doubleSupport ? " -D DOUBLE_SUPPORT" : "", cn);
+ }
+
+ k.create(kernelName, program, opts);
+ if (k.empty())
+ return false;
+
+ double borderBuf[] = {0, 0, 0, 0};
+ scalarToRawData(borderValue, borderBuf, sctype);
+
+ UMat src = _src.getUMat(), M0;
+ _dst.create( dsize.area() == 0 ? src.size() : dsize, src.type() );
+ UMat dst = _dst.getUMat();
+
+ double M[9];
+ int matRows = (op_type == OCL_OP_AFFINE ? 2 : 3);
+ Mat matM(matRows, 3, CV_64F, M), M1 = _M0.getMat();
+ CV_Assert( (M1.type() == CV_32F || M1.type() == CV_64F) &&
+ M1.rows == matRows && M1.cols == 3 );
+ M1.convertTo(matM, matM.type());
+
+ if( !(flags & WARP_INVERSE_MAP) )
+ {
+ if (op_type == OCL_OP_PERSPECTIVE)
+ invert(matM, matM);
+ else
+ {
+ double D = M[0]*M[4] - M[1]*M[3];
+ D = D != 0 ? 1./D : 0;
+ double A11 = M[4]*D, A22=M[0]*D;
+ M[0] = A11; M[1] *= -D;
+ M[3] *= -D; M[4] = A22;
+ double b1 = -M[0]*M[2] - M[1]*M[5];
+ double b2 = -M[3]*M[2] - M[4]*M[5];
+ M[2] = b1; M[5] = b2;
+ }
+ }
+ matM.convertTo(M0, doubleSupport ? CV_64F : CV_32F);
+
+ k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst), ocl::KernelArg::PtrReadOnly(M0),
+ ocl::KernelArg(0, 0, 0, borderBuf, CV_ELEM_SIZE(sctype)));
+
+ size_t globalThreads[2] = { dst.cols, dst.rows };
+ return k.run(2, globalThreads, NULL, false);
+}
+
+#endif
+
}
InputArray _M0, Size dsize,
int flags, int borderType, const Scalar& borderValue )
{
+ CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
+ ocl_warpTransform(_src, _dst, _M0, dsize, flags, borderType,
+ borderValue, OCL_OP_AFFINE))
+
Mat src = _src.getMat(), M0 = _M0.getMat();
_dst.create( dsize.area() == 0 ? src.size() : dsize, src.type() );
Mat dst = _dst.getMat();
void cv::warpPerspective( InputArray _src, OutputArray _dst, InputArray _M0,
Size dsize, int flags, int borderType, const Scalar& borderValue )
{
+ CV_Assert( _src.total() > 0 );
+
+ CV_OCL_RUN(_src.dims() <= 2 && _dst.isUMat(),
+ ocl_warpTransform(_src, _dst, _M0, dsize, flags, borderType, borderValue,
+ OCL_OP_PERSPECTIVE))
+
Mat src = _src.getMat(), M0 = _M0.getMat();
_dst.create( dsize.area() == 0 ? src.size() : dsize, src.type() );
Mat dst = _dst.getMat();
- CV_Assert( src.cols > 0 && src.rows > 0 );
if( dst.data == src.data )
src = src.clone();
ssize = cvGetMatSize(src);
dsize = cvGetMatSize(dst);
- mapx = cvCreateMat( dsize.height, dsize.width, CV_32F );
- mapy = cvCreateMat( dsize.height, dsize.width, CV_32F );
+ mapx.reset(cvCreateMat( dsize.height, dsize.width, CV_32F ));
+ mapy.reset(cvCreateMat( dsize.height, dsize.width, CV_32F ));
if( !(flags & CV_WARP_INVERSE_MAP) )
{
double xx = bufx.data.fl[x];
double yy = bufy.data.fl[x];
- double p = log(sqrt(xx*xx + yy*yy) + 1.)*M;
+ double p = log(std::sqrt(xx*xx + yy*yy) + 1.)*M;
double a = atan2(yy,xx);
if( a < 0 )
a = 2*CV_PI + a;
cvRemap( src, dst, mapx, mapy, flags, cvScalarAll(0) );
}
+void cv::logPolar( InputArray _src, OutputArray _dst,
+ Point2f center, double M, int flags )
+{
+ Mat src = _src.getMat();
+ _dst.create( src.size(), src.type() );
+ CvMat c_src = src, c_dst = _dst.getMat();
+ cvLogPolar( &c_src, &c_dst, center, M, flags );
+}
/****************************************************************************************
Linear-Polar Transform
dsize.width = dst->cols;
dsize.height = dst->rows;
- mapx = cvCreateMat( dsize.height, dsize.width, CV_32F );
- mapy = cvCreateMat( dsize.height, dsize.width, CV_32F );
+ mapx.reset(cvCreateMat( dsize.height, dsize.width, CV_32F ));
+ mapy.reset(cvCreateMat( dsize.height, dsize.width, CV_32F ));
if( !(flags & CV_WARP_INVERSE_MAP) )
{
cvRemap( src, dst, mapx, mapy, flags, cvScalarAll(0) );
}
+void cv::linearPolar( InputArray _src, OutputArray _dst,
+ Point2f center, double maxRadius, int flags )
+{
+ Mat src = _src.getMat();
+ _dst.create( src.size(), src.type() );
+ CvMat c_src = src, c_dst = _dst.getMat();
+ cvLinearPolar( &c_src, &c_dst, center, maxRadius, flags );
+}
/* End of file. */