From dcce9d7088f1cb9472cc75e4d3e2b62de89d33ec Mon Sep 17 00:00:00 2001 From: Ilya Lavrenov Date: Sun, 1 Dec 2013 18:09:52 +0400 Subject: [PATCH] added cv::warpAffine to T-API --- modules/core/src/ocl.cpp | 3 +- modules/imgproc/src/imgwarp.cpp | 146 +++--- modules/imgproc/src/opencl/warp_affine.cl | 725 ++++-------------------------- 3 files changed, 189 insertions(+), 685 deletions(-) diff --git a/modules/core/src/ocl.cpp b/modules/core/src/ocl.cpp index 8a0cc27..4ef3fda 100644 --- a/modules/core/src/ocl.cpp +++ b/modules/core/src/ocl.cpp @@ -2041,7 +2041,6 @@ struct Kernel::Impl cl_int retval = 0; handle = ph != 0 ? clCreateKernel(ph, kname, &retval) : 0; - printf("kernel creation error code: %d\n", retval); for( int i = 0; i < MAX_ARRS; i++ ) u[i] = 0; haveTempDstUMats = false; @@ -2219,7 +2218,7 @@ int Kernel::set(int i, const KernelArg& arg) else if( arg.m->dims <= 2 ) { UMat2D u2d(*arg.m); - clSetKernelArg(p->handle, (cl_uint)i, sizeof(h), &h)); + clSetKernelArg(p->handle, (cl_uint)i, sizeof(h), &h); clSetKernelArg(p->handle, (cl_uint)(i+1), sizeof(u2d.step), &u2d.step); clSetKernelArg(p->handle, (cl_uint)(i+2), sizeof(u2d.offset), &u2d.offset); i += 3; diff --git a/modules/imgproc/src/imgwarp.cpp b/modules/imgproc/src/imgwarp.cpp index e5383be..69f7d36 100644 --- a/modules/imgproc/src/imgwarp.cpp +++ b/modules/imgproc/src/imgwarp.cpp @@ -3789,6 +3789,87 @@ private: }; #endif +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), wdepth = depth; + 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 || cn == 3) + return false; + + 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); + + const char * const interpolationMap[3] = { "NEAREST", "LINEAR", "CUBIC" }; + ocl::ProgramSource2 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"; + + ocl::Kernel k; + if (interpolation == INTER_NEAREST) + { + k.create(kernelName, program, + format("-D INTER_NEAREST -D T=%s%s", ocl::typeToStr(type), + doubleSupport ? " -D DOUBLE_SUPPORT" : "")); + } + else + { + char cvt[2][50]; + wdepth = std::max(CV_32S, depth); + k.create(kernelName, program, + format("-D INTER_%s -D T=%s -D WT=%s -D depth=%d -D convertToWT=%s -D convertToT=%s%s", + interpolationMap[interpolation], ocl::typeToStr(type), + 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" : "")); + } + + k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst), ocl::KernelArg::PtrOnly(M0), + ocl::KernelArg::Constant(Mat(1, 1, CV_MAKE_TYPE(wdepth, cn), borderValue))); + + size_t globalThreads[2] = { dst.cols, dst.rows }; + return k.run(2, globalThreads, NULL, false); +} + } @@ -3796,6 +3877,11 @@ void cv::warpAffine( InputArray _src, OutputArray _dst, InputArray _M0, Size dsize, int flags, int borderType, const Scalar& borderValue ) { + if (ocl::useOpenCL() && _dst.isUMat() && + ocl_warpTransform(_src, _dst, _M0, dsize, flags, borderType, + borderValue, OCL_OP_AFFINE)) + return; + Mat src = _src.getMat(), M0 = _M0.getMat(); _dst.create( dsize.area() == 0 ? src.size() : dsize, src.type() ); Mat dst = _dst.getMat(); @@ -4030,62 +4116,6 @@ private: }; #endif -static bool ocl_warpPerspective(InputArray _src, OutputArray _dst, InputArray _M0, - Size dsize, int flags, int borderType, const Scalar& borderValue) -{ - int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type), wdepth = depth; - 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 || cn == 3) - return false; - - UMat src = _src.getUMat(), M0; - _dst.create( dsize.area() == 0 ? src.size() : dsize, src.type() ); - UMat dst = _dst.getUMat(); - - double M[9]; - Mat matM(3, 3, doubleSupport ? CV_64F : CV_32F, M), M1 = _M0.getMat(); - CV_Assert( (M1.type() == CV_32F || M1.type() == CV_64F) && M1.rows == 3 && M1.cols == 3 ); - M1.convertTo(matM, matM.type()); - if( !(flags & WARP_INVERSE_MAP) ) - invert(matM, matM); - matM.copyTo(M0); - - const char * const interpolationMap[3] = { "NEAREST", "LINEAR", "CUBIC" }; - ocl::Kernel k; - - if (interpolation == INTER_NEAREST) - { - k.create("warpPerspective", ocl::imgproc::warp_perspective_oclsrc, - format("-D INTER_NEAREST -D T=%s%s", ocl::typeToStr(type), - doubleSupport ? " -D DOUBLE_SUPPORT" : "")); - } - else - { - char cvt[2][50]; - wdepth = std::max(CV_32S, depth); - k.create("warpPerspective", ocl::imgproc::warp_perspective_oclsrc, - format("-D INTER_%s -D T=%s -D WT=%s -D depth=%d -D convertToWT=%s -D convertToT=%s%s", - interpolationMap[interpolation], ocl::typeToStr(type), - 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" : "")); - } - - k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst), - ocl::KernelArg::PtrOnly(M0), ocl::KernelArg::Constant(Mat(1, 1, CV_MAKE_TYPE(wdepth, cn), borderValue))); - - size_t globalThreads[2] = { dst.cols, dst.rows }; - return k.run(2, globalThreads, NULL, false); -} - } void cv::warpPerspective( InputArray _src, OutputArray _dst, InputArray _M0, @@ -4093,7 +4123,9 @@ void cv::warpPerspective( InputArray _src, OutputArray _dst, InputArray _M0, { CV_Assert( _src.total() > 0 ); - if (ocl::useOpenCL() && _dst.isUMat() && ocl_warpPerspective(_src, _dst, _M0, dsize, flags, borderType, borderValue)) + if (ocl::useOpenCL() && _dst.isUMat() && + ocl_warpTransform(_src, _dst, _M0, dsize, flags, borderType, borderValue, + OCL_OP_PERSPECTIVE)) return; Mat src = _src.getMat(), M0 = _M0.getMat(); diff --git a/modules/imgproc/src/opencl/warp_affine.cl b/modules/imgproc/src/opencl/warp_affine.cl index 27f99e0..340cfdd 100644 --- a/modules/imgproc/src/opencl/warp_affine.cl +++ b/modules/imgproc/src/opencl/warp_affine.cl @@ -43,23 +43,15 @@ // //M*/ - -//warpAffine kernel -//support data types: CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4, and three interpolation methods: NN, Linear, Cubic. - #ifdef DOUBLE_SUPPORT #ifdef cl_amd_fp64 #pragma OPENCL EXTENSION cl_amd_fp64:enable #elif defined (cl_khr_fp64) #pragma OPENCL EXTENSION cl_khr_fp64:enable #endif -typedef double F; -typedef double4 F4; -#define convert_F4 convert_double4 +#define CT double #else -typedef float F; -typedef float4 F4; -#define convert_F4 convert_float4 +#define CT float #endif #define INTER_BITS 5 @@ -70,286 +62,56 @@ typedef float4 F4; #define INTER_REMAP_COEF_BITS 15 #define INTER_REMAP_COEF_SCALE (1 << INTER_REMAP_COEF_BITS) -inline void interpolateCubic( float x, float* coeffs ) -{ - const float A = -0.75f; - - coeffs[0] = ((A*(x + 1.f) - 5.0f*A)*(x + 1.f) + 8.0f*A)*(x + 1.f) - 4.0f*A; - coeffs[1] = ((A + 2.f)*x - (A + 3.f))*x*x + 1.f; - coeffs[2] = ((A + 2.f)*(1.f - x) - (A + 3.f))*(1.f - x)*(1.f - x) + 1.f; - coeffs[3] = 1.f - coeffs[0] - coeffs[1] - coeffs[2]; -} - - -/**********************************************8UC1********************************************* -***********************************************************************************************/ -__kernel void warpAffineNN_C1_D0(__global uchar const * restrict src, __global uchar * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) -{ - int dx = get_global_id(0); - int dy = get_global_id(1); - - if( dx < threadCols && dy < dst_rows) - { - dx = (dx<<2) - (dst_offset&3); - - int round_delta = (AB_SCALE>>1); - - int4 X, Y; - int4 sx, sy; - int4 DX = (int4)(dx, dx+1, dx+2, dx+3); - DX = (DX << AB_BITS); - F4 M0DX, M3DX; - M0DX = M[0] * convert_F4(DX); - M3DX = M[3] * convert_F4(DX); - X = convert_int4(rint(M0DX)); - Y = convert_int4(rint(M3DX)); - int tmp1, tmp2; - tmp1 = rint((M[1]*dy + M[2]) * AB_SCALE); - tmp2 = rint((M[4]*dy + M[5]) * AB_SCALE); - - X += tmp1 + round_delta; - Y += tmp2 + round_delta; - - sx = convert_int4(convert_short4(X >> AB_BITS)); - sy = convert_int4(convert_short4(Y >> AB_BITS)); - - __global uchar4 * d = (__global uchar4 *)(dst+dst_offset+dy*dstStep+dx); - uchar4 dval = *d; - DX = (int4)(dx, dx+1, dx+2, dx+3); - int4 dcon = DX >= 0 && DX < dst_cols && dy >= 0 && dy < dst_rows; - int4 scon = sx >= 0 && sx < src_cols && sy >= 0 && sy < src_rows; - int4 spos = src_offset + sy * srcStep + sx; - uchar4 sval; - sval.s0 = scon.s0 ? src[spos.s0] : 0; - sval.s1 = scon.s1 ? src[spos.s1] : 0; - sval.s2 = scon.s2 ? src[spos.s2] : 0; - sval.s3 = scon.s3 ? src[spos.s3] : 0; - dval = convert_uchar4(dcon) != (uchar4)(0,0,0,0) ? sval : dval; - *d = dval; - } -} - -__kernel void warpAffineLinear_C1_D0(__global const uchar * restrict src, __global uchar * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) -{ - int dx = get_global_id(0); - int dy = get_global_id(1); - +#define noconvert - if( dx < threadCols && dy < dst_rows) - { - dx = (dx<<2) - (dst_offset&3); - - int round_delta = ((AB_SCALE >> INTER_BITS) >> 1); - - int4 X, Y; - short4 ax, ay; - int4 sx, sy; - int4 DX = (int4)(dx, dx+1, dx+2, dx+3); - DX = (DX << AB_BITS); - F4 M0DX, M3DX; - M0DX = M[0] * convert_F4(DX); - M3DX = M[3] * convert_F4(DX); - X = convert_int4(rint(M0DX)); - Y = convert_int4(rint(M3DX)); - - int tmp1, tmp2; - tmp1 = rint((M[1]*dy + M[2]) * AB_SCALE); - tmp2 = rint((M[4]*dy + M[5]) * AB_SCALE); - - X += tmp1 + round_delta; - Y += tmp2 + round_delta; - - X = X >> (AB_BITS - INTER_BITS); - Y = Y >> (AB_BITS - INTER_BITS); - - sx = convert_int4(convert_short4(X >> INTER_BITS)); - sy = convert_int4(convert_short4(Y >> INTER_BITS)); - ax = convert_short4(X & (INTER_TAB_SIZE-1)); - ay = convert_short4(Y & (INTER_TAB_SIZE-1)); - - uchar4 v0, v1, v2,v3; - int4 scon0, scon1, scon2, scon3; - int4 spos0, spos1, spos2, spos3; - - scon0 = (sx >= 0 && sx < src_cols && sy >= 0 && sy < src_rows); - scon1 = (sx+1 >= 0 && sx+1 < src_cols && sy >= 0 && sy < src_rows); - scon2 = (sx >= 0 && sx < src_cols && sy+1 >= 0 && sy+1 < src_rows); - scon3 = (sx+1 >= 0 && sx+1 < src_cols && sy+1 >= 0 && sy+1 < src_rows); - spos0 = src_offset + sy * srcStep + sx; - spos1 = src_offset + sy * srcStep + sx + 1; - spos2 = src_offset + (sy+1) * srcStep + sx; - spos3 = src_offset + (sy+1) * srcStep + sx + 1; - - v0.s0 = scon0.s0 ? src[spos0.s0] : 0; - v1.s0 = scon1.s0 ? src[spos1.s0] : 0; - v2.s0 = scon2.s0 ? src[spos2.s0] : 0; - v3.s0 = scon3.s0 ? src[spos3.s0] : 0; - - v0.s1 = scon0.s1 ? src[spos0.s1] : 0; - v1.s1 = scon1.s1 ? src[spos1.s1] : 0; - v2.s1 = scon2.s1 ? src[spos2.s1] : 0; - v3.s1 = scon3.s1 ? src[spos3.s1] : 0; - - v0.s2 = scon0.s2 ? src[spos0.s2] : 0; - v1.s2 = scon1.s2 ? src[spos1.s2] : 0; - v2.s2 = scon2.s2 ? src[spos2.s2] : 0; - v3.s2 = scon3.s2 ? src[spos3.s2] : 0; - - v0.s3 = scon0.s3 ? src[spos0.s3] : 0; - v1.s3 = scon1.s3 ? src[spos1.s3] : 0; - v2.s3 = scon2.s3 ? src[spos2.s3] : 0; - v3.s3 = scon3.s3 ? src[spos3.s3] : 0; - - short4 itab0, itab1, itab2, itab3; - float4 taby, tabx; - taby = INTER_SCALE * convert_float4(ay); - tabx = INTER_SCALE * convert_float4(ax); - - itab0 = convert_short4_sat(( (1.0f-taby)*(1.0f-tabx) * (float4)INTER_REMAP_COEF_SCALE )); - itab1 = convert_short4_sat(( (1.0f-taby)*tabx * (float4)INTER_REMAP_COEF_SCALE )); - itab2 = convert_short4_sat(( taby*(1.0f-tabx) * (float4)INTER_REMAP_COEF_SCALE )); - itab3 = convert_short4_sat(( taby*tabx * (float4)INTER_REMAP_COEF_SCALE )); - - - int4 val; - uchar4 tval; - val = convert_int4(v0) * convert_int4(itab0) + convert_int4(v1) * convert_int4(itab1) - + convert_int4(v2) * convert_int4(itab2) + convert_int4(v3) * convert_int4(itab3); - tval = convert_uchar4_sat ( (val + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS ) ; - - __global uchar4 * d =(__global uchar4 *)(dst+dst_offset+dy*dstStep+dx); - uchar4 dval = *d; - DX = (int4)(dx, dx+1, dx+2, dx+3); - int4 dcon = DX >= 0 && DX < dst_cols && dy >= 0 && dy < dst_rows; - dval = convert_uchar4(dcon != 0) ? tval : dval; - *d = dval; - } -} +#ifdef INTER_NEAREST -__kernel void warpAffineCubic_C1_D0(__global uchar * src, __global uchar * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) +__kernel void warpAffine(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols, + __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols, + __constant CT * M, T scalar) { int dx = get_global_id(0); int dy = get_global_id(1); - if( dx < threadCols && dy < dst_rows) + if (dx < dst_cols && dy < dst_rows) { - int round_delta = ((AB_SCALE>>INTER_BITS)>>1); + int round_delta = (AB_SCALE >> 1); int X0 = rint(M[0] * dx * AB_SCALE); int Y0 = rint(M[3] * dx * AB_SCALE); X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta; Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta; - int X = X0 >> (AB_BITS - INTER_BITS); - int Y = Y0 >> (AB_BITS - INTER_BITS); - - short sx = (short)(X >> INTER_BITS) - 1; - short sy = (short)(Y >> INTER_BITS) - 1; - short ay = (short)(Y & (INTER_TAB_SIZE-1)); - short ax = (short)(X & (INTER_TAB_SIZE-1)); - - uchar v[16]; - int i, j; - -#pragma unroll 4 - for(i=0; i<4; i++) - for(j=0; j<4; j++) - { - v[i*4+j] = (sx+j >= 0 && sx+j < src_cols && sy+i >= 0 && sy+i < src_rows) ? src[src_offset+(sy+i) * srcStep + (sx+j)] : 0; - } - - short itab[16]; - float tab1y[4], tab1x[4]; - float axx, ayy; - ayy = 1.f/INTER_TAB_SIZE * ay; - axx = 1.f/INTER_TAB_SIZE * ax; - interpolateCubic(ayy, tab1y); - interpolateCubic(axx, tab1x); - int isum = 0; + short sx = convert_short_sat(X0 >> AB_BITS); + short sy = convert_short_sat(Y0 >> AB_BITS); -#pragma unroll 16 - for( i=0; i<16; i++ ) - { - F v = tab1y[(i>>2)] * tab1x[(i&3)]; - isum += itab[i] = convert_short_sat( rint( v * INTER_REMAP_COEF_SCALE ) ); - } - - if( isum != INTER_REMAP_COEF_SCALE ) - { - int k1, k2; - int diff = isum - INTER_REMAP_COEF_SCALE; - int Mk1=2, Mk2=2, mk1=2, mk2=2; - for( k1 = 2; k1 < 4; k1++ ) - for( k2 = 2; k2 < 4; k2++ ) - { - if( itab[(k1<<2)+k2] < itab[(mk1<<2)+mk2] ) - mk1 = k1, mk2 = k2; - else if( itab[(k1<<2)+k2] > itab[(Mk1<<2)+Mk2] ) - Mk1 = k1, Mk2 = k2; - } - diff<0 ? (itab[(Mk1<<2)+Mk2]=(short)(itab[(Mk1<<2)+Mk2]-diff)) : (itab[(mk1<<2)+mk2]=(short)(itab[(mk1<<2)+mk2]-diff)); - } + int dst_index = mad24(dy, dst_step, dst_offset + dx * (int)sizeof(T)); + __global T * dst = (__global T *)(dstptr + dst_index); - if( dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) + if (sx >= 0 && sx < src_cols && sy >= 0 && sy < src_rows) { - int sum=0; - for ( i =0; i<16; i++ ) - { - sum += v[i] * itab[i] ; - } - dst[dst_offset+dy*dstStep+dx] = convert_uchar_sat( (sum + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS ) ; + int src_index = mad24(sy, src_step, src_offset + sx * (int)sizeof(T)); + __global const T * src = (__global const T *)(srcptr + src_index); + dst[0] = src[0]; } + else + dst[0] = scalar; } } -/**********************************************8UC4********************************************* -***********************************************************************************************/ +#elif defined INTER_LINEAR -__kernel void warpAffineNN_C4_D0(__global uchar4 const * restrict src, __global uchar4 * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) +__kernel void warpAffine(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols, + __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols, + __constant CT * M, WT scalar) { int dx = get_global_id(0); int dy = get_global_id(1); - if( dx < threadCols && dy < dst_rows) - { - int round_delta = (AB_SCALE >> 1); - - int X0 = rint(M[0] * dx * AB_SCALE); - int Y0 = rint(M[3] * dx * AB_SCALE); - X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta; - Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta; - - int sx0 = (short)(X0 >> AB_BITS); - int sy0 = (short)(Y0 >> AB_BITS); - - if(dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) - dst[(dst_offset>>2)+dy*(dstStep>>2)+dx]= (sx0>=0 && sx0=0 && sy0>2)+sy0*(srcStep>>2)+sx0] : (uchar4)0; - } -} - -__kernel void warpAffineLinear_C4_D0(__global uchar4 const * restrict src, __global uchar4 * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) -{ - int dx = get_global_id(0); - int dy = get_global_id(1); - - - if( dx < threadCols && dy < dst_rows) + if (dx < dst_cols && dy < dst_rows) { int round_delta = AB_SCALE/INTER_TAB_SIZE/2; - src_offset = (src_offset>>2); - srcStep = (srcStep>>2); - int tmp = (dx << AB_BITS); int X0 = rint(M[0] * tmp); int Y0 = rint(M[3] * tmp); @@ -358,52 +120,65 @@ __kernel void warpAffineLinear_C4_D0(__global uchar4 const * restrict src, __glo X0 = X0 >> (AB_BITS - INTER_BITS); Y0 = Y0 >> (AB_BITS - INTER_BITS); - short sx0 = (short)(X0 >> INTER_BITS); - short sy0 = (short)(Y0 >> INTER_BITS); - short ax0 = (short)(X0 & (INTER_TAB_SIZE-1)); - short ay0 = (short)(Y0 & (INTER_TAB_SIZE-1)); - - int4 v0, v1, v2, v3; - - v0 = (sx0 >= 0 && sx0 < src_cols && sy0 >= 0 && sy0 < src_rows) ? convert_int4(src[src_offset+sy0 * srcStep + sx0]) : 0; - v1 = (sx0+1 >= 0 && sx0+1 < src_cols && sy0 >= 0 && sy0 < src_rows) ? convert_int4(src[src_offset+sy0 * srcStep + sx0+1]) : 0; - v2 = (sx0 >= 0 && sx0 < src_cols && sy0+1 >= 0 && sy0+1 < src_rows) ? convert_int4(src[src_offset+(sy0+1) * srcStep + sx0]) : 0; - v3 = (sx0+1 >= 0 && sx0+1 < src_cols && sy0+1 >= 0 && sy0+1 < src_rows) ? convert_int4(src[src_offset+(sy0+1) * srcStep + sx0+1]) : 0; - - int itab0, itab1, itab2, itab3; - float taby, tabx; - taby = 1.f/INTER_TAB_SIZE*ay0; - tabx = 1.f/INTER_TAB_SIZE*ax0; + short sx = convert_short_sat(X0 >> INTER_BITS); + short sy = convert_short_sat(Y0 >> INTER_BITS); + short ax = convert_short(X0 & (INTER_TAB_SIZE-1)); + short ay = convert_short(Y0 & (INTER_TAB_SIZE-1)); + + WT v0 = (sx >= 0 && sx < src_cols && sy >= 0 && sy < src_rows) ? + convertToWT(*(__global const T *)(srcptr + mad24(sy, src_step, src_offset + sx * (int)sizeof(T)))) : scalar; + WT v1 = (sx+1 >= 0 && sx+1 < src_cols && sy >= 0 && sy < src_rows) ? + convertToWT(*(__global const T *)(srcptr + mad24(sy, src_step, src_offset + (sx+1) * (int)sizeof(T)))) : scalar; + WT v2 = (sx >= 0 && sx < src_cols && sy+1 >= 0 && sy+1 < src_rows) ? + convertToWT(*(__global const T *)(srcptr + mad24(sy+1, src_step, src_offset + sx * (int)sizeof(T)))) : scalar; + WT v3 = (sx+1 >= 0 && sx+1 < src_cols && sy+1 >= 0 && sy+1 < src_rows) ? + convertToWT(*(__global const T *)(srcptr + mad24(sy+1, src_step, src_offset + (sx+1) * (int)sizeof(T)))) : scalar; + + float taby = 1.f/INTER_TAB_SIZE*ay; + float tabx = 1.f/INTER_TAB_SIZE*ax; + + int dst_index = mad24(dy, dst_step, dst_offset + dx * (int)sizeof(T)); + __global T * dst = (__global T *)(dstptr + dst_index); + +#if depth <= 4 + int itab0 = convert_short_sat_rte( (1.0f-taby)*(1.0f-tabx) * INTER_REMAP_COEF_SCALE ); + int itab1 = convert_short_sat_rte( (1.0f-taby)*tabx * INTER_REMAP_COEF_SCALE ); + int itab2 = convert_short_sat_rte( taby*(1.0f-tabx) * INTER_REMAP_COEF_SCALE ); + int itab3 = convert_short_sat_rte( taby*tabx * INTER_REMAP_COEF_SCALE ); + + WT val = v0 * itab0 + v1 * itab1 + v2 * itab2 + v3 * itab3; + dst[0] = convertToT((val + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS); +#else + float tabx2 = 1.0f - tabx, taby2 = 1.0f - taby; + WT val = v0 * tabx2 * taby2 + v1 * tabx * taby2 + v2 * tabx2 * taby + v3 * tabx * taby; + dst[0] = convertToT(val); +#endif + } +} - itab0 = convert_short_sat(rint( (1.0f-taby)*(1.0f-tabx) * INTER_REMAP_COEF_SCALE )); - itab1 = convert_short_sat(rint( (1.0f-taby)*tabx * INTER_REMAP_COEF_SCALE )); - itab2 = convert_short_sat(rint( taby*(1.0f-tabx) * INTER_REMAP_COEF_SCALE )); - itab3 = convert_short_sat(rint( taby*tabx * INTER_REMAP_COEF_SCALE )); +#elif defined INTER_CUBIC - int4 val; - val = v0 * itab0 + v1 * itab1 + v2 * itab2 + v3 * itab3; +inline void interpolateCubic( float x, float* coeffs ) +{ + const float A = -0.75f; - if(dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) - dst[(dst_offset>>2)+dy*(dstStep>>2)+dx] = convert_uchar4_sat ( (val + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS ) ; - } + coeffs[0] = ((A*(x + 1.f) - 5.0f*A)*(x + 1.f) + 8.0f*A)*(x + 1.f) - 4.0f*A; + coeffs[1] = ((A + 2.f)*x - (A + 3.f))*x*x + 1.f; + coeffs[2] = ((A + 2.f)*(1.f - x) - (A + 3.f))*(1.f - x)*(1.f - x) + 1.f; + coeffs[3] = 1.f - coeffs[0] - coeffs[1] - coeffs[2]; } -__kernel void warpAffineCubic_C4_D0(__global uchar4 const * restrict src, __global uchar4 * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) +__kernel void warpAffine(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols, + __global uchar * dstptr, int dst_step, int dst_offset, int dst_rows, int dst_cols, + __constant CT * M, WT scalar) { int dx = get_global_id(0); int dy = get_global_id(1); - if( dx < threadCols && dy < dst_rows) + if (dx < dst_cols && dy < dst_rows) { int round_delta = ((AB_SCALE>>INTER_BITS)>>1); - src_offset = (src_offset>>2); - srcStep = (srcStep>>2); - dst_offset = (dst_offset>>2); - dstStep = (dstStep>>2); - int tmp = (dx << AB_BITS); int X0 = rint(M[0] * tmp); int Y0 = rint(M[3] * tmp); @@ -417,345 +192,43 @@ __kernel void warpAffineCubic_C4_D0(__global uchar4 const * restrict src, __glob int ay = (short)(Y0 & (INTER_TAB_SIZE-1)); int ax = (short)(X0 & (INTER_TAB_SIZE-1)); - uchar4 v[16]; - int i,j; -#pragma unroll 4 - for(i=0; i<4; i++) - for(j=0; j<4; j++) - { - v[i*4+j] = (sx+j >= 0 && sx+j < src_cols && sy+i >= 0 && sy+i < src_rows) ? (src[src_offset+(sy+i) * srcStep + (sx+j)]) : (uchar4)0; - } - int itab[16]; - float tab1y[4], tab1x[4]; - float axx, ayy; - - ayy = INTER_SCALE * ay; - axx = INTER_SCALE * ax; - interpolateCubic(ayy, tab1y); - interpolateCubic(axx, tab1x); - int isum = 0; - -#pragma unroll 16 - for( i=0; i<16; i++ ) - { - float tmp; - tmp = tab1y[(i>>2)] * tab1x[(i&3)] * INTER_REMAP_COEF_SCALE; - itab[i] = rint(tmp); - isum += itab[i]; - } + WT v[16]; + #pragma unroll + for (int y = 0; y < 4; y++) + #pragma unroll + for (int x = 0; x < 4; x++) + v[mad24(y, 4, x)] = (sx+x >= 0 && sx+x < src_cols && sy+y >= 0 && sy+y < src_rows) ? + convertToWT(*(__global const T *)(srcptr + mad24(sy+y, src_step, src_offset + (sx+x) * (int)sizeof(T)))) : scalar; - if( isum != INTER_REMAP_COEF_SCALE ) - { - int k1, k2; - int diff = isum - INTER_REMAP_COEF_SCALE; - int Mk1=2, Mk2=2, mk1=2, mk2=2; - - for( k1 = 2; k1 < 4; k1++ ) - for( k2 = 2; k2 < 4; k2++ ) - { - - if( itab[(k1<<2)+k2] < itab[(mk1<<2)+mk2] ) - mk1 = k1, mk2 = k2; - else if( itab[(k1<<2)+k2] > itab[(Mk1<<2)+Mk2] ) - Mk1 = k1, Mk2 = k2; - } - - diff<0 ? (itab[(Mk1<<2)+Mk2]=(short)(itab[(Mk1<<2)+Mk2]-diff)) : (itab[(mk1<<2)+mk2]=(short)(itab[(mk1<<2)+mk2]-diff)); - } - - if( dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) - { - int4 sum=0; - for ( i =0; i<16; i++ ) - { - sum += convert_int4(v[i]) * itab[i]; - } - dst[dst_offset+dy*dstStep+dx] = convert_uchar4_sat( (sum + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS ) ; - } - } -} - - -/**********************************************32FC1******************************************** -***********************************************************************************************/ - -__kernel void warpAffineNN_C1_D5(__global float * src, __global float * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) -{ - int dx = get_global_id(0); - int dy = get_global_id(1); - - if( dx < threadCols && dy < dst_rows) - { - int round_delta = AB_SCALE/2; - - int X0 = rint(M[0] * dx * AB_SCALE); - int Y0 = rint(M[3] * dx * AB_SCALE); - X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta; - Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta; - - short sx0 = (short)(X0 >> AB_BITS); - short sy0 = (short)(Y0 >> AB_BITS); - - if(dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) - dst[(dst_offset>>2)+dy*dstStep+dx]= (sx0>=0 && sx0=0 && sy0>2)+sy0*srcStep+sx0] : 0; - } -} - -__kernel void warpAffineLinear_C1_D5(__global float * src, __global float * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) -{ - int dx = get_global_id(0); - int dy = get_global_id(1); - - if( dx < threadCols && dy < dst_rows) - { - int round_delta = AB_SCALE/INTER_TAB_SIZE/2; - - src_offset = (src_offset>>2); - - int X0 = rint(M[0] * dx * AB_SCALE); - int Y0 = rint(M[3] * dx * AB_SCALE); - X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta; - Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta; - X0 = X0 >> (AB_BITS - INTER_BITS); - Y0 = Y0 >> (AB_BITS - INTER_BITS); - - short sx0 = (short)(X0 >> INTER_BITS); - short sy0 = (short)(Y0 >> INTER_BITS); - short ax0 = (short)(X0 & (INTER_TAB_SIZE-1)); - short ay0 = (short)(Y0 & (INTER_TAB_SIZE-1)); - - float v0, v1, v2, v3; - - v0 = (sx0 >= 0 && sx0 < src_cols && sy0 >= 0 && sy0 < src_rows) ? src[src_offset+sy0 * srcStep + sx0] : 0; - v1 = (sx0+1 >= 0 && sx0+1 < src_cols && sy0 >= 0 && sy0 < src_rows) ? src[src_offset+sy0 * srcStep + sx0+1] : 0; - v2 = (sx0 >= 0 && sx0 < src_cols && sy0+1 >= 0 && sy0+1 < src_rows) ? src[src_offset+(sy0+1) * srcStep + sx0] : 0; - v3 = (sx0+1 >= 0 && sx0+1 < src_cols && sy0+1 >= 0 && sy0+1 < src_rows) ? src[src_offset+(sy0+1) * srcStep + sx0+1] : 0; - - float tab[4]; - float taby[2], tabx[2]; - taby[0] = 1.0f - 1.f/INTER_TAB_SIZE*ay0; - taby[1] = 1.f/INTER_TAB_SIZE*ay0; - tabx[0] = 1.0f - 1.f/INTER_TAB_SIZE*ax0; - tabx[1] = 1.f/INTER_TAB_SIZE*ax0; - - tab[0] = taby[0] * tabx[0]; - tab[1] = taby[0] * tabx[1]; - tab[2] = taby[1] * tabx[0]; - tab[3] = taby[1] * tabx[1]; - - float sum = 0; - sum += v0 * tab[0] + v1 * tab[1] + v2 * tab[2] + v3 * tab[3]; - if(dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) - dst[(dst_offset>>2)+dy*dstStep+dx] = sum; - } -} - -__kernel void warpAffineCubic_C1_D5(__global float * src, __global float * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) -{ - int dx = get_global_id(0); - int dy = get_global_id(1); - - if( dx < threadCols && dy < dst_rows) - { - int round_delta = AB_SCALE/INTER_TAB_SIZE/2; - - src_offset = (src_offset>>2); - dst_offset = (dst_offset>>2); - - int X0 = rint(M[0] * dx * AB_SCALE); - int Y0 = rint(M[3] * dx * AB_SCALE); - X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta; - Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta; - X0 = X0 >> (AB_BITS - INTER_BITS); - Y0 = Y0 >> (AB_BITS - INTER_BITS); - - short sx = (short)(X0 >> INTER_BITS) - 1; - short sy = (short)(Y0 >> INTER_BITS) - 1; - short ay = (short)(Y0 & (INTER_TAB_SIZE-1)); - short ax = (short)(X0 & (INTER_TAB_SIZE-1)); - - float v[16]; - int i; - - for(i=0; i<16; i++) - v[i] = (sx+(i&3) >= 0 && sx+(i&3) < src_cols && sy+(i>>2) >= 0 && sy+(i>>2) < src_rows) ? src[src_offset+(sy+(i>>2)) * srcStep + (sx+(i&3))] : 0; - - float tab[16]; float tab1y[4], tab1x[4]; - float axx, ayy; - ayy = 1.f/INTER_TAB_SIZE * ay; - axx = 1.f/INTER_TAB_SIZE * ax; + float ayy = INTER_SCALE * ay; + float axx = INTER_SCALE * ax; interpolateCubic(ayy, tab1y); interpolateCubic(axx, tab1x); -#pragma unroll 4 - for( i=0; i<16; i++ ) - { - tab[i] = tab1y[(i>>2)] * tab1x[(i&3)]; - } - - if( dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) - { - float sum = 0; -#pragma unroll 4 - for ( i =0; i<16; i++ ) - { - sum += v[i] * tab[i]; - } - dst[dst_offset+dy*dstStep+dx] = sum; - - } - } -} - - -/**********************************************32FC4******************************************** -***********************************************************************************************/ - -__kernel void warpAffineNN_C4_D5(__global float4 * src, __global float4 * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) -{ - int dx = get_global_id(0); - int dy = get_global_id(1); - - if( dx < threadCols && dy < dst_rows) - { - int round_delta = AB_SCALE/2; - - int X0 = rint(M[0] * dx * AB_SCALE); - int Y0 = rint(M[3] * dx * AB_SCALE); - X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta; - Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta; - - short sx0 = (short)(X0 >> AB_BITS); - short sy0 = (short)(Y0 >> AB_BITS); - - if(dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) - dst[(dst_offset>>4)+dy*(dstStep>>2)+dx]= (sx0>=0 && sx0=0 && sy0>4)+sy0*(srcStep>>2)+sx0] : (float4)0; - } -} - -__kernel void warpAffineLinear_C4_D5(__global float4 * src, __global float4 * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) -{ - int dx = get_global_id(0); - int dy = get_global_id(1); - - if( dx < threadCols && dy < dst_rows) - { - int round_delta = AB_SCALE/INTER_TAB_SIZE/2; + int dst_index = mad24(dy, dst_step, dst_offset + dx * (int)sizeof(T)); + __global T * dst = (__global T *)(dstptr + dst_index); - src_offset = (src_offset>>4); - dst_offset = (dst_offset>>4); - srcStep = (srcStep>>2); - dstStep = (dstStep>>2); + WT sum = (WT)(0); +#if depth <= 4 + int itab[16]; - int X0 = rint(M[0] * dx * AB_SCALE); - int Y0 = rint(M[3] * dx * AB_SCALE); - X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta; - Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta; - X0 = X0 >> (AB_BITS - INTER_BITS); - Y0 = Y0 >> (AB_BITS - INTER_BITS); + #pragma unroll + for (int i = 0; i < 16; i++) + itab[i] = rint(tab1y[(i>>2)] * tab1x[(i&3)] * INTER_REMAP_COEF_SCALE); - short sx0 = (short)(X0 >> INTER_BITS); - short sy0 = (short)(Y0 >> INTER_BITS); - short ax0 = (short)(X0 & (INTER_TAB_SIZE-1)); - short ay0 = (short)(Y0 & (INTER_TAB_SIZE-1)); - - float4 v0, v1, v2, v3; - - v0 = (sx0 >= 0 && sx0 < src_cols && sy0 >= 0 && sy0 < src_rows) ? src[src_offset+sy0 * srcStep + sx0] : (float4)0; - v1 = (sx0+1 >= 0 && sx0+1 < src_cols && sy0 >= 0 && sy0 < src_rows) ? src[src_offset+sy0 * srcStep + sx0+1] : (float4)0; - v2 = (sx0 >= 0 && sx0 < src_cols && sy0+1 >= 0 && sy0+1 < src_rows) ? src[src_offset+(sy0+1) * srcStep + sx0] : (float4)0; - v3 = (sx0+1 >= 0 && sx0+1 < src_cols && sy0+1 >= 0 && sy0+1 < src_rows) ? src[src_offset+(sy0+1) * srcStep + sx0+1] : (float4)0; - - float tab[4]; - float taby[2], tabx[2]; - taby[0] = 1.0f - 1.f/INTER_TAB_SIZE*ay0; - taby[1] = 1.f/INTER_TAB_SIZE*ay0; - tabx[0] = 1.0f - 1.f/INTER_TAB_SIZE*ax0; - tabx[1] = 1.f/INTER_TAB_SIZE*ax0; - - tab[0] = taby[0] * tabx[0]; - tab[1] = taby[0] * tabx[1]; - tab[2] = taby[1] * tabx[0]; - tab[3] = taby[1] * tabx[1]; - - float4 sum = 0; - sum += v0 * tab[0] + v1 * tab[1] + v2 * tab[2] + v3 * tab[3]; - if(dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) - dst[dst_offset+dy*dstStep+dx] = sum; + #pragma unroll + for (int i = 0; i < 16; i++) + sum += v[i] * itab[i]; + dst[0] = convertToT( (sum + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS ); +#else + #pragma unroll + for (int i = 0; i < 16; i++) + sum += v[i] * tab1y[(i>>2)] * tab1x[(i&3)]; + dst[0] = convertToT( sum ); +#endif } } -__kernel void warpAffineCubic_C4_D5(__global float4 * src, __global float4 * dst, int src_cols, int src_rows, - int dst_cols, int dst_rows, int srcStep, int dstStep, - int src_offset, int dst_offset, __constant F * M, int threadCols ) -{ - int dx = get_global_id(0); - int dy = get_global_id(1); - - if( dx < threadCols && dy < dst_rows) - { - int round_delta = AB_SCALE/INTER_TAB_SIZE/2; - - src_offset = (src_offset>>4); - dst_offset = (dst_offset>>4); - srcStep = (srcStep>>2); - dstStep = (dstStep>>2); - - int X0 = rint(M[0] * dx * AB_SCALE); - int Y0 = rint(M[3] * dx * AB_SCALE); - X0 += rint((M[1]*dy + M[2]) * AB_SCALE) + round_delta; - Y0 += rint((M[4]*dy + M[5]) * AB_SCALE) + round_delta; - X0 = X0 >> (AB_BITS - INTER_BITS); - Y0 = Y0 >> (AB_BITS - INTER_BITS); - - short sx = (short)(X0 >> INTER_BITS) - 1; - short sy = (short)(Y0 >> INTER_BITS) - 1; - short ay = (short)(Y0 & (INTER_TAB_SIZE-1)); - short ax = (short)(X0 & (INTER_TAB_SIZE-1)); - - float4 v[16]; - int i; - - for(i=0; i<16; i++) - v[i] = (sx+(i&3) >= 0 && sx+(i&3) < src_cols && sy+(i>>2) >= 0 && sy+(i>>2) < src_rows) ? src[src_offset+(sy+(i>>2)) * srcStep + (sx+(i&3))] : (float4)0; - - float tab[16]; - float tab1y[4], tab1x[4]; - float axx, ayy; - - ayy = 1.f/INTER_TAB_SIZE * ay; - axx = 1.f/INTER_TAB_SIZE * ax; - interpolateCubic(ayy, tab1y); - interpolateCubic(axx, tab1x); - -#pragma unroll 4 - for( i=0; i<16; i++ ) - { - tab[i] = tab1y[(i>>2)] * tab1x[(i&3)]; - } - - if( dx >= 0 && dx < dst_cols && dy >= 0 && dy < dst_rows) - { - float4 sum = 0; -#pragma unroll 4 - for ( i =0; i<16; i++ ) - { - sum += v[i] * tab[i]; - } - dst[dst_offset+dy*dstStep+dx] = sum; - - } - } -} +#endif -- 2.7.4