diff --git a/3rdparty/openvx/include/openvx_hal.hpp b/3rdparty/openvx/include/openvx_hal.hpp index 893547dfdf..a6547e12bb 100644 --- a/3rdparty/openvx/include/openvx_hal.hpp +++ b/3rdparty/openvx/include/openvx_hal.hpp @@ -7,6 +7,10 @@ #include "VX/vxu.h" #include +#include + +#include +#include //================================================================================================== // utility @@ -52,7 +56,8 @@ template struct VX_Traits { enum { - Type = 0 + ImgType = 0, + DataType = 0 }; }; @@ -60,7 +65,8 @@ template <> struct VX_Traits { enum { - Type = VX_DF_IMAGE_U8 + ImgType = VX_DF_IMAGE_U8, + DataType = VX_TYPE_UINT8 }; }; @@ -68,7 +74,8 @@ template <> struct VX_Traits { enum { - Type = VX_DF_IMAGE_U16 + ImgType = VX_DF_IMAGE_U16, + DataType = VX_TYPE_UINT16 }; }; @@ -76,7 +83,17 @@ template <> struct VX_Traits { enum { - Type = VX_DF_IMAGE_S16 + ImgType = VX_DF_IMAGE_S16, + DataType = VX_TYPE_INT16 + }; +}; + +template <> +struct VX_Traits +{ + enum { + ImgType = 0, + DataType = VX_TYPE_FLOAT32 }; }; @@ -108,6 +125,14 @@ struct vxErr { vxErr(vxGetStatus((vx_reference)img), "image check").check(); } + static void check(vx_matrix mtx) + { + vxErr(vxGetStatus((vx_reference)mtx), "matrix check").check(); + } + static void check(vx_convolution cnv) + { + vxErr(vxGetStatus((vx_reference)cnv), "convolution check").check(); + } static void check(vx_status s) { vxErr(s, "status check").check(); @@ -145,28 +170,176 @@ struct vxImage addr.dim_x = w; addr.dim_y = h; addr.stride_x = sizeof(T); - addr.stride_y = step; + addr.stride_y = (vx_int32)step; addr.scale_x = VX_SCALE_UNITY; addr.scale_y = VX_SCALE_UNITY; addr.step_x = 1; addr.step_y = 1; void *ptrs[] = { (void*)data }; - img = vxCreateImageFromHandle(ctx.ctx, VX_Traits::Type, &addr, ptrs, VX_MEMORY_TYPE_HOST); + img = vxCreateImageFromHandle(ctx.ctx, VX_Traits::ImgType, &addr, ptrs, VX_MEMORY_TYPE_HOST); + vxErr::check(img); + } + vxImage(vxContext &ctx, int imgType, const uchar *data, size_t step, int w, int h) + { + if (h == 1) + step = w; + vx_imagepatch_addressing_t addr[4]; + void *ptrs[4]; + switch (imgType) + { + case VX_DF_IMAGE_U8: + case VX_DF_IMAGE_U16: + case VX_DF_IMAGE_S16: + case VX_DF_IMAGE_U32: + case VX_DF_IMAGE_S32: + case VX_DF_IMAGE_RGB: + case VX_DF_IMAGE_RGBX: + case VX_DF_IMAGE_UYVY: + case VX_DF_IMAGE_YUYV: + addr[0].dim_x = w; + addr[0].dim_y = h; + addr[0].stride_x = imgType == VX_DF_IMAGE_U8 ? 1 : + imgType == VX_DF_IMAGE_RGB ? 3 : + (imgType == VX_DF_IMAGE_U16 || imgType == VX_DF_IMAGE_S16) ? 2 : 4; + addr[0].stride_y = (vx_int32)step; + addr[0].scale_x = VX_SCALE_UNITY; + addr[0].scale_y = VX_SCALE_UNITY; + addr[0].step_x = (imgType == VX_DF_IMAGE_UYVY || imgType == VX_DF_IMAGE_YUYV) ? 2 : 1; + addr[0].step_y = 1; + ptrs[0] = (void*)data; + break; + case VX_DF_IMAGE_NV12: + case VX_DF_IMAGE_NV21: + addr[0].dim_x = w; + addr[0].dim_y = h; + addr[0].stride_x = 1; + addr[0].stride_y = (vx_int32)step; + addr[0].scale_x = VX_SCALE_UNITY; + addr[0].scale_y = VX_SCALE_UNITY; + addr[0].step_x = 1; + addr[0].step_y = 1; + ptrs[0] = (void*)data; + addr[1].dim_x = (w + 1) / 2; + addr[1].dim_y = (h + 1) / 2; + addr[1].stride_x = 2; + addr[1].stride_y = (vx_int32)step; + addr[1].scale_x = VX_SCALE_UNITY; + addr[1].scale_y = VX_SCALE_UNITY; + addr[1].step_x = 2; + addr[1].step_y = 2; + ptrs[1] = (void*)(data + h * step); + break; + case VX_DF_IMAGE_IYUV: + case VX_DF_IMAGE_YUV4: + addr[0].dim_x = w; + addr[0].dim_y = h; + addr[0].stride_x = 1; + addr[0].stride_y = (vx_int32)step; + addr[0].scale_x = VX_SCALE_UNITY; + addr[0].scale_y = VX_SCALE_UNITY; + addr[0].step_x = 1; + addr[0].step_y = 1; + ptrs[0] = (void*)data; + addr[1].dim_x = imgType == VX_DF_IMAGE_YUV4 ? w : (w + 1) / 2; + addr[1].dim_y = imgType == VX_DF_IMAGE_YUV4 ? h : (h + 1) / 2; + addr[1].stride_x = 1; + addr[1].stride_y = (vx_int32)step; + addr[1].scale_x = VX_SCALE_UNITY; + addr[1].scale_y = VX_SCALE_UNITY; + addr[1].step_x = imgType == VX_DF_IMAGE_YUV4 ? 1 : 2; + addr[1].step_y = imgType == VX_DF_IMAGE_YUV4 ? 1 : 2; + ptrs[1] = (void*)(data + h * step); + addr[2].dim_x = addr[1].dim_x; + addr[2].dim_y = addr[1].dim_y; + addr[2].stride_x = 1; + addr[2].stride_y = (vx_int32)step; + addr[2].scale_x = VX_SCALE_UNITY; + addr[2].scale_y = VX_SCALE_UNITY; + addr[2].step_x = addr[1].step_x; + addr[2].step_y = addr[1].step_y; + ptrs[2] = (void*)(data + (h + addr[1].dim_y) * step); + break; + default: + vxErr(VX_ERROR_INVALID_PARAMETERS, "Bad image format").check(); + } + img = vxCreateImageFromHandle(ctx.ctx, imgType, addr, ptrs, VX_MEMORY_TYPE_HOST); vxErr::check(img); } ~vxImage() { - vxSwapImageHandle(img, NULL, NULL, 1); + vxErr::check(vxSwapImageHandle(img, NULL, NULL, 1)); vxReleaseImage(&img); } }; +struct vxMatrix +{ + vx_matrix mtx; + + template + vxMatrix(vxContext &ctx, const T *data, int w, int h) + { + mtx = vxCreateMatrix(ctx.ctx, VX_Traits::DataType, w, h); + vxErr::check(mtx); + vxErr::check(vxCopyMatrix(mtx, const_cast(data), VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST)); + } + ~vxMatrix() + { + vxReleaseMatrix(&mtx); + } +}; + +struct vxConvolution +{ + vx_convolution cnv; + + vxConvolution(vxContext &ctx, const short *data, int w, int h) + { + cnv = vxCreateConvolution(ctx.ctx, w, h); + vxErr::check(cnv); + vxErr::check(vxCopyConvolutionCoefficients(cnv, const_cast(data), VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST)); + } + ~vxConvolution() + { + vxReleaseConvolution(&cnv); + } +}; + //================================================================================================== // real code starts here // ... +#define OVX_BINARY_OP(hal_func, ovx_call) \ +template \ +inline int ovx_hal_##hal_func(const T *a, size_t astep, const T *b, size_t bstep, T *c, size_t cstep, int w, int h) \ +{ \ + try \ + { \ + vxContext * ctx = vxContext::getContext(); \ + vxImage ia(*ctx, a, astep, w, h); \ + vxImage ib(*ctx, b, bstep, w, h); \ + vxImage ic(*ctx, c, cstep, w, h); \ + ovx_call \ + } \ + catch (vxErr & e) \ + { \ + e.print(); \ + return CV_HAL_ERROR_UNKNOWN; \ + } \ + return CV_HAL_ERROR_OK; \ +} + +OVX_BINARY_OP(add, {vxErr::check(vxuAdd(ctx->ctx, ia.img, ib.img, VX_CONVERT_POLICY_SATURATE, ic.img));}) +OVX_BINARY_OP(sub, {vxErr::check(vxuSubtract(ctx->ctx, ia.img, ib.img, VX_CONVERT_POLICY_SATURATE, ic.img));}) + +OVX_BINARY_OP(absdiff, {vxErr::check(vxuAbsDiff(ctx->ctx, ia.img, ib.img, ic.img));}) + +OVX_BINARY_OP(and, {vxErr::check(vxuAnd(ctx->ctx, ia.img, ib.img, ic.img));}) +OVX_BINARY_OP(or, {vxErr::check(vxuOr(ctx->ctx, ia.img, ib.img, ic.img));}) +OVX_BINARY_OP(xor, {vxErr::check(vxuXor(ctx->ctx, ia.img, ib.img, ic.img));}) + template -inline int ovx_hal_add(const T *a, size_t astep, const T *b, size_t bstep, T *c, size_t cstep, int w, int h) +inline int ovx_hal_mul(const T *a, size_t astep, const T *b, size_t bstep, T *c, size_t cstep, int w, int h, double scale) { try { @@ -174,7 +347,7 @@ inline int ovx_hal_add(const T *a, size_t astep, const T *b, size_t bstep, T *c, vxImage ia(*ctx, a, astep, w, h); vxImage ib(*ctx, b, bstep, w, h); vxImage ic(*ctx, c, cstep, w, h); - vxErr::check(vxuAdd(ctx->ctx, ia.img, ib.img, VX_CONVERT_POLICY_SATURATE, ic.img)); + vxErr::check(vxuMultiply(ctx->ctx, ia.img, ib.img, (float)scale, VX_CONVERT_POLICY_SATURATE, VX_ROUND_POLICY_TO_ZERO, ic.img)); } catch (vxErr & e) { @@ -184,6 +357,637 @@ inline int ovx_hal_add(const T *a, size_t astep, const T *b, size_t bstep, T *c, return CV_HAL_ERROR_OK; } +inline int ovx_hal_not(const uchar *a, size_t astep, uchar *c, size_t cstep, int w, int h) +{ + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, a, astep, w, h); + vxImage ic(*ctx, c, cstep, w, h); + vxErr::check(vxuNot(ctx->ctx, ia.img, ic.img)); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_merge8u(const uchar **src_data, uchar *dst_data, int len, int cn) +{ + if (cn != 3 && cn != 4) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, src_data[0], len, len, 1); + vxImage ib(*ctx, src_data[1], len, len, 1); + vxImage ic(*ctx, src_data[2], len, len, 1); + vxImage id(*ctx, cn == 4 ? VX_DF_IMAGE_RGBX : VX_DF_IMAGE_RGB, dst_data, len, len, 1); + vxErr::check(vxuChannelCombine(ctx->ctx, ia.img, ib.img, ic.img, + cn == 4 ? vxImage(*ctx, src_data[3], len, len, 1).img : NULL, + id.img)); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + + +#if defined OPENCV_IMGPROC_HAL_INTERFACE_H +#define CV_HAL_INTER_NEAREST 0 +#define CV_HAL_INTER_LINEAR 1 +#define CV_HAL_INTER_CUBIC 2 +#define CV_HAL_INTER_AREA 3 +#define CV_HAL_INTER_LANCZOS4 4 +#define MORPH_ERODE 0 +#define MORPH_DILATE 1 + +inline int ovx_hal_resize(int atype, const uchar *a, size_t astep, int aw, int ah, uchar *b, size_t bstep, int bw, int bh, double inv_scale_x, double inv_scale_y, int interpolation) +{ + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, a, astep, aw, ah); + vxImage ib(*ctx, b, bstep, bw, bh); + + if(!((atype == CV_8UC1 || atype == CV_8SC1) && + inv_scale_x > 0 && inv_scale_y > 0 && + (bw - 0.5) / inv_scale_x - 0.5 < aw && (bh - 0.5) / inv_scale_y - 0.5 < ah && + (bw + 0.5) / inv_scale_x + 0.5 >= aw && (bh + 0.5) / inv_scale_y + 0.5 >= ah && + std::abs(bw / inv_scale_x - aw) < 0.1 && std::abs(bh / inv_scale_y - ah) < 0.1 )) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + int mode; + if (interpolation == CV_HAL_INTER_LINEAR) + mode = VX_INTERPOLATION_BILINEAR; + else if (interpolation == CV_HAL_INTER_AREA) + mode = VX_INTERPOLATION_AREA; + else if (interpolation == CV_HAL_INTER_NEAREST) + mode = VX_INTERPOLATION_NEAREST_NEIGHBOR; + else + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + vxErr::check( vxuScaleImage(ctx->ctx, ia.img, ib.img, mode)); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_warpAffine(int atype, const uchar *a, size_t astep, int aw, int ah, uchar *b, size_t bstep, int bw, int bh, const double M[6], int interpolation, int borderType, const double borderValue[4]) +{ + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, a, astep, aw, ah); + vxImage ib(*ctx, b, bstep, bw, bh); + + if (!(atype == CV_8UC1 || atype == CV_8SC1)) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + vx_border_t border; + switch (borderType) + { + case CV_HAL_BORDER_CONSTANT: + border.mode = VX_BORDER_CONSTANT; + border.constant_value.U8 = (vx_uint8)(borderValue[0]); + break; + case CV_HAL_BORDER_REPLICATE: + border.mode = VX_BORDER_REPLICATE; + break; + default: + return CV_HAL_ERROR_NOT_IMPLEMENTED; + } + + int mode; + if (interpolation == CV_HAL_INTER_LINEAR) + mode = VX_INTERPOLATION_BILINEAR; + else if (interpolation == CV_HAL_INTER_AREA) + mode = VX_INTERPOLATION_AREA; + else if (interpolation == CV_HAL_INTER_NEAREST) + mode = VX_INTERPOLATION_NEAREST_NEIGHBOR; + else + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + std::vector data; + data.reserve(6); + for (int j = 0; j < 3; ++j) + for (int i = 0; i < 2; ++i) + data.push_back((float)(M[i*3+j])); + + vxMatrix mtx(*ctx, data.data(), 2, 3); + //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 + vx_border_t prevBorder; + vxErr::check(vxQueryContext(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &prevBorder, sizeof(prevBorder))); + vxErr::check(vxSetContextAttribute(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &border, sizeof(border))); + vxErr::check(vxuWarpAffine(ctx->ctx, ia.img, mtx.mtx, mode, ib.img)); + vxErr::check(vxSetContextAttribute(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &prevBorder, sizeof(prevBorder))); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_warpPerspectve(int atype, const uchar *a, size_t astep, int aw, int ah, uchar *b, size_t bstep, int bw, int bh, const double M[9], int interpolation, int borderType, const double borderValue[4]) +{ + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, a, astep, aw, ah); + vxImage ib(*ctx, b, bstep, bw, bh); + + if (!(atype == CV_8UC1 || atype == CV_8SC1)) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + vx_border_t border; + switch (borderType) + { + case CV_HAL_BORDER_CONSTANT: + border.mode = VX_BORDER_CONSTANT; + border.constant_value.U8 = (vx_uint8)(borderValue[0]); + break; + case CV_HAL_BORDER_REPLICATE: + border.mode = VX_BORDER_REPLICATE; + break; + default: + return CV_HAL_ERROR_NOT_IMPLEMENTED; + } + + int mode; + if (interpolation == CV_HAL_INTER_LINEAR) + mode = VX_INTERPOLATION_BILINEAR; + else if (interpolation == CV_HAL_INTER_AREA) + mode = VX_INTERPOLATION_AREA; + else if (interpolation == CV_HAL_INTER_NEAREST) + mode = VX_INTERPOLATION_NEAREST_NEIGHBOR; + else + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + std::vector data; + data.reserve(9); + for (int j = 0; j < 3; ++j) + for (int i = 0; i < 3; ++i) + data.push_back((float)(M[i * 3 + j])); + + vxMatrix mtx(*ctx, data.data(), 3, 3); + //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 + vx_border_t prevBorder; + vxErr::check(vxQueryContext(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &prevBorder, sizeof(prevBorder))); + vxErr::check(vxSetContextAttribute(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &border, sizeof(border))); + vxErr::check(vxuWarpPerspective(ctx->ctx, ia.img, mtx.mtx, mode, ib.img)); + vxErr::check(vxSetContextAttribute(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &prevBorder, sizeof(prevBorder))); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +struct cvhalFilter2D; + +struct FilterCtx +{ + vxConvolution cnv; + vx_border_t border; + int dst_type; + FilterCtx(vxContext &ctx, const short *data, int w, int h, int _dst_type, vx_border_t & _border) : + cnv(ctx, data, w, h), dst_type(_dst_type), border(_border) {} +}; + +inline int ovx_hal_filterInit(cvhalFilter2D **filter_context, uchar *kernel_data, size_t kernel_step, int kernel_type, int kernel_width, int kernel_height, + int , int , int src_type, int dst_type, int borderType, double delta, int anchor_x, int anchor_y, bool allowSubmatrix, bool allowInplace) +{ + if (!filter_context || !kernel_data || allowSubmatrix || allowInplace || delta != 0 || + src_type != CV_8UC1 || (dst_type != CV_8UC1 && dst_type != CV_16SC1) || + kernel_width % 2 == 0 || kernel_height % 2 == 0 || anchor_x != kernel_width / 2 || anchor_y != kernel_height / 2) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + vx_border_t border; + switch (borderType) + { + case CV_HAL_BORDER_CONSTANT: + border.mode = VX_BORDER_CONSTANT; + border.constant_value.U8 = 0; + break; + case CV_HAL_BORDER_REPLICATE: + border.mode = VX_BORDER_REPLICATE; + break; + default: + return CV_HAL_ERROR_NOT_IMPLEMENTED; + } + + vxContext * ctx = vxContext::getContext(); + + std::vector data; + data.reserve(kernel_width*kernel_height); + switch (kernel_type) + { + case CV_8UC1: + for (int j = 0; j < kernel_height; ++j) + { + uchar * row = (uchar*)(kernel_data + kernel_step*j); + for (int i = 0; i < kernel_width; ++i) + data.push_back(row[i]); + } + break; + case CV_8SC1: + for (int j = 0; j < kernel_height; ++j) + { + schar * row = (schar*)(kernel_data + kernel_step*j); + for (int i = 0; i < kernel_width; ++i) + data.push_back(row[i]); + } + break; + case CV_16SC1: + for (int j = 0; j < kernel_height; ++j) + { + short * row = (short*)(kernel_data + kernel_step*j); + for (int i = 0; i < kernel_width; ++i) + data.push_back(row[i]); + } + default: + return CV_HAL_ERROR_NOT_IMPLEMENTED; + } + + FilterCtx* cnv = new FilterCtx(*ctx, data.data(), kernel_width, kernel_height, dst_type, border); + if (!cnv) + return CV_HAL_ERROR_UNKNOWN; + + *filter_context = (cvhalFilter2D*)(cnv); + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_filterFree(cvhalFilter2D *filter_context) +{ + if (filter_context) + { + delete (FilterCtx*)filter_context; + return CV_HAL_ERROR_OK; + } + else + { + return CV_HAL_ERROR_UNKNOWN; + } +} + +inline int ovx_hal_filter(cvhalFilter2D *filter_context, uchar *a, size_t astep, uchar *b, size_t bstep, int w, int h, int , int , int , int ) +{ + try + { + FilterCtx* cnv = (FilterCtx*)filter_context; + if(!cnv) + vxErr(VX_ERROR_INVALID_PARAMETERS, "Bad HAL context").check(); + + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, a, astep, w, h); + + //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 + vx_border_t prevBorder; + vxErr::check(vxQueryContext(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &prevBorder, sizeof(prevBorder))); + vxErr::check(vxSetContextAttribute(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &(cnv->border), sizeof(cnv->border))); + if (cnv->dst_type == CV_16SC1) + { + vxImage ib(*ctx, (short*)b, bstep, w, h); + vxErr::check(vxuConvolve(ctx->ctx, ia.img, cnv->cnv.cnv, ib.img)); + } + else + { + vxImage ib(*ctx, b, bstep, w, h); + vxErr::check(vxuConvolve(ctx->ctx, ia.img, cnv->cnv.cnv, ib.img)); + } + vxErr::check(vxSetContextAttribute(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &prevBorder, sizeof(prevBorder))); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_sepFilterInit(cvhalFilter2D **filter_context, int src_type, int dst_type, + int kernel_type, uchar *kernelx_data, int kernelx_length, uchar *kernely_data, int kernely_length, + int anchor_x, int anchor_y, double delta, int borderType) +{ + if (!filter_context || !kernelx_data || !kernely_data || delta != 0 || + src_type != CV_8UC1 || (dst_type != CV_8UC1 && dst_type != CV_16SC1) || + kernelx_length % 2 == 0 || kernely_length % 2 == 0 || anchor_x != kernelx_length / 2 || anchor_y != kernely_length / 2) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + vx_border_t border; + switch (borderType) + { + case CV_HAL_BORDER_CONSTANT: + border.mode = VX_BORDER_CONSTANT; + border.constant_value.U8 = 0; + break; + case CV_HAL_BORDER_REPLICATE: + border.mode = VX_BORDER_REPLICATE; + break; + default: + return CV_HAL_ERROR_NOT_IMPLEMENTED; + } + + vxContext * ctx = vxContext::getContext(); + + //At the moment OpenVX doesn't support separable filters natively so combine kernels to generic convolution + std::vector data; + data.reserve(kernelx_length*kernely_length); + switch (kernel_type) + { + case CV_8UC1: + for (int j = 0; j < kernely_length; ++j) + for (int i = 0; i < kernelx_length; ++i) + data.push_back((short)(kernely_data[j]) * kernelx_data[i]); + break; + case CV_8SC1: + for (int j = 0; j < kernely_length; ++j) + for (int i = 0; i < kernelx_length; ++i) + data.push_back((short)(((schar*)kernely_data)[j]) * ((schar*)kernelx_data)[i]); + break; + default: + return CV_HAL_ERROR_NOT_IMPLEMENTED; + } + + FilterCtx* cnv = new FilterCtx(*ctx, data.data(), kernelx_length, kernely_length, dst_type, border); + if (!cnv) + return CV_HAL_ERROR_UNKNOWN; + + *filter_context = (cvhalFilter2D*)(cnv); + return CV_HAL_ERROR_OK; +} + +struct MorphCtx +{ + vxMatrix mask; + int operation; + vx_border_t border; + MorphCtx(vxContext &ctx, const uchar *data, int w, int h, int _operation, vx_border_t & _border) : + mask(ctx, data, w, h), operation(_operation), border(_border) {} +}; + +inline int ovx_hal_morphInit(cvhalFilter2D **filter_context, int operation, int src_type, int dst_type, int , int , + int kernel_type, uchar *kernel_data, size_t kernel_step, int kernel_width, int kernel_height, int anchor_x, int anchor_y, + int borderType, const double borderValue[4], int iterations, bool allowSubmatrix, bool allowInplace) +{ + if (!filter_context || !kernel_data || allowSubmatrix || allowInplace || iterations != 1 || + src_type != CV_8UC1 || dst_type != CV_8UC1 || + kernel_width % 2 == 0 || kernel_height % 2 == 0 || anchor_x != kernel_width / 2 || anchor_y != kernel_height / 2) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + vx_border_t border; + switch (borderType) + { + case CV_HAL_BORDER_CONSTANT: + border.mode = VX_BORDER_CONSTANT; + if (borderValue[0] == DBL_MAX && borderValue[1] == DBL_MAX && borderValue[2] == DBL_MAX && borderValue[3] == DBL_MAX) + { + if (operation == MORPH_ERODE) + border.constant_value.U8 = UCHAR_MAX; + else + border.constant_value.U8 = 0; + } + else + { + border.constant_value.U8 = cv::saturate_cast(borderValue[0]); + } + break; + case CV_HAL_BORDER_REPLICATE: + border.mode = VX_BORDER_REPLICATE; + break; + default: + return CV_HAL_ERROR_NOT_IMPLEMENTED; + } + + vxContext * ctx = vxContext::getContext(); + + std::vector kernel_mat; + kernel_mat.resize(kernel_width * kernel_height); + switch (CV_MAT_DEPTH(kernel_type)) + { + case CV_8U: + case CV_8S: + for (int j = 0; j < kernel_height; ++j) + { + uchar * kernel_row = kernel_data + j * kernel_step; + for (int i = 0; i < kernel_height; ++i) + kernel_mat.push_back(kernel_row[i] ? 255 : 0); + } + break; + case CV_16U: + case CV_16S: + for (int j = 0; j < kernel_height; ++j) + { + short * kernel_row = (short*)(kernel_data + j * kernel_step); + for (int i = 0; i < kernel_height; ++i) + kernel_mat.push_back(kernel_row[i] ? 255 : 0); + } + break; + case CV_32S: + for (int j = 0; j < kernel_height; ++j) + { + int * kernel_row = (int*)(kernel_data + j * kernel_step); + for (int i = 0; i < kernel_height; ++i) + kernel_mat.push_back(kernel_row[i] ? 255 : 0); + } + break; + case CV_32F: + for (int j = 0; j < kernel_height; ++j) + { + float * kernel_row = (float*)(kernel_data + j * kernel_step); + for (int i = 0; i < kernel_height; ++i) + kernel_mat.push_back(kernel_row[i] ? 255 : 0); + } + break; + case CV_64F: + for (int j = 0; j < kernel_height; ++j) + { + double * kernel_row = (double*)(kernel_data + j * kernel_step); + for (int i = 0; i < kernel_height; ++i) + kernel_mat.push_back(kernel_row[i] ? 255 : 0); + } + break; + default: + return CV_HAL_ERROR_NOT_IMPLEMENTED; + } + + MorphCtx* mat; + switch (operation) + { + case MORPH_ERODE: + mat = new MorphCtx(*ctx, kernel_mat.data(), kernel_width, kernel_height, VX_NONLINEAR_FILTER_MIN, border); + case MORPH_DILATE: + mat = new MorphCtx(*ctx, kernel_mat.data(), kernel_width, kernel_height, VX_NONLINEAR_FILTER_MAX, border); + break; + default: + return CV_HAL_ERROR_NOT_IMPLEMENTED; + } + if (!mat) + return CV_HAL_ERROR_UNKNOWN; + + *filter_context = (cvhalFilter2D*)(mat); + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_morphFree(cvhalFilter2D *filter_context) +{ + if (filter_context) + { + delete (MorphCtx*)filter_context; + return CV_HAL_ERROR_OK; + } + else + { + return CV_HAL_ERROR_UNKNOWN; + } +} + +inline int ovx_hal_morph(cvhalFilter2D *filter_context, uchar *a, size_t astep, uchar *b, size_t bstep, int w, int h, int , int , int , int , int , int , int , int ) +{ + try + { + MorphCtx* mat = (MorphCtx*)filter_context; + if (!mat) + vxErr(VX_ERROR_INVALID_PARAMETERS, "Bad HAL context").check(); + + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, a, astep, w, h); + vxImage ib(*ctx, b, bstep, w, h); + + //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 + vx_border_t prevBorder; + vxErr::check(vxQueryContext(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &prevBorder, sizeof(prevBorder))); + vxErr::check(vxSetContextAttribute(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &(mat->border), sizeof(mat->border))); + vxErr::check(vxuNonLinearFilter(ctx->ctx, mat->operation, ia.img, mat->mask.mtx, ib.img)); + vxErr::check(vxSetContextAttribute(ctx->ctx, VX_CONTEXT_IMMEDIATE_BORDER, &prevBorder, sizeof(prevBorder))); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_cvtBGRtoBGR(const uchar * a, size_t astep, uchar * b, size_t bstep, int w, int h, int depth, int acn, int bcn, bool swapBlue) +{ + if (depth != CV_8U || swapBlue || acn == bcn || (acn != 3 && acn != 4) || (bcn != 3 && bcn != 4)) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, acn == 3 ? VX_DF_IMAGE_RGB : VX_DF_IMAGE_RGBX, a, astep, w, h); + vxImage ib(*ctx, bcn == 3 ? VX_DF_IMAGE_RGB : VX_DF_IMAGE_RGBX, b, bstep, w, h); + vxErr::check(vxuColorConvert(ctx->ctx, ia.img, ib.img)); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_cvtTwoPlaneYUVtoBGR(const uchar * a, size_t astep, uchar * b, size_t bstep, int w, int h, int bcn, bool swapBlue, int uIdx) +{ + if (!swapBlue || (bcn != 3 && bcn != 4)) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, uIdx ? VX_DF_IMAGE_NV21 : VX_DF_IMAGE_NV12, a, astep, w, h); + vxImage ib(*ctx, bcn == 3 ? VX_DF_IMAGE_RGB : VX_DF_IMAGE_RGBX, b, bstep, w, h); + vxErr::check(vxuColorConvert(ctx->ctx, ia.img, ib.img)); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_cvtThreePlaneYUVtoBGR(const uchar * a, size_t astep, uchar * b, size_t bstep, int w, int h, int bcn, bool swapBlue, int uIdx) +{ + if (!swapBlue || (bcn != 3 && bcn != 4) || uIdx) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, VX_DF_IMAGE_IYUV, a, astep, w, h); + vxImage ib(*ctx, bcn == 3 ? VX_DF_IMAGE_RGB : VX_DF_IMAGE_RGBX, b, bstep, w, h); + vxErr::check(vxuColorConvert(ctx->ctx, ia.img, ib.img)); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_cvtBGRtoThreePlaneYUV(const uchar * a, size_t astep, uchar * b, size_t bstep, int w, int h, int acn, bool swapBlue, int uIdx) +{ + if (!swapBlue || (acn != 3 && acn != 4) || uIdx) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, acn == 3 ? VX_DF_IMAGE_RGB : VX_DF_IMAGE_RGBX, a, astep, w, h); + vxImage ib(*ctx, VX_DF_IMAGE_IYUV, b, bstep, w, h); + vxErr::check(vxuColorConvert(ctx->ctx, ia.img, ib.img)); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +inline int ovx_hal_cvtOnePlaneYUVtoBGR(const uchar * a, size_t astep, uchar * b, size_t bstep, int w, int h, int bcn, bool swapBlue, int uIdx, int ycn) +{ + if (!swapBlue || (bcn != 3 && bcn != 4) || uIdx) + return CV_HAL_ERROR_NOT_IMPLEMENTED; + + try + { + vxContext * ctx = vxContext::getContext(); + vxImage ia(*ctx, ycn ? VX_DF_IMAGE_UYVY : VX_DF_IMAGE_YUYV, a, astep, w, h); + vxImage ib(*ctx, bcn == 3 ? VX_DF_IMAGE_RGB : VX_DF_IMAGE_RGBX, b, bstep, w, h); + vxErr::check(vxuColorConvert(ctx->ctx, ia.img, ib.img)); + } + catch (vxErr & e) + { + e.print(); + return CV_HAL_ERROR_UNKNOWN; + } + return CV_HAL_ERROR_OK; +} + +#endif + //================================================================================================== // functions redefinition // ... @@ -192,5 +996,74 @@ inline int ovx_hal_add(const T *a, size_t astep, const T *b, size_t bstep, T *c, #define cv_hal_add8u ovx_hal_add #undef cv_hal_add16s #define cv_hal_add16s ovx_hal_add +#undef cv_hal_sub8u +#define cv_hal_sub8u ovx_hal_sub +#undef cv_hal_sub16s +#define cv_hal_sub16s ovx_hal_sub + +#undef cv_hal_absdiff8u +#define cv_hal_absdiff8u ovx_hal_absdiff +#undef cv_hal_absdiff16s +#define cv_hal_absdiff16s ovx_hal_absdiff + +#undef cv_hal_and8u +#define cv_hal_and8u ovx_hal_and +#undef cv_hal_or8u +#define cv_hal_or8u ovx_hal_or +#undef cv_hal_xor8u +#define cv_hal_xor8u ovx_hal_xor +#undef cv_hal_not8u +#define cv_hal_not8u ovx_hal_not + +#undef cv_hal_mul8u +#define cv_hal_mul8u ovx_hal_mul +#undef cv_hal_mul16s +#define cv_hal_mul16s ovx_hal_mul + +#undef cv_hal_merge8u +#define cv_hal_merge8u ovx_hal_merge8u + +#if defined OPENCV_IMGPROC_HAL_INTERFACE_H + +#undef cv_hal_resize +#define cv_hal_resize ovx_hal_resize +#undef cv_hal_warpAffine +#define cv_hal_warpAffine ovx_hal_warpAffine +#undef cv_hal_warpPerspective +#define cv_hal_warpPerspective ovx_hal_warpPerspectve + +#undef cv_hal_filterInit +#define cv_hal_filterInit ovx_hal_filterInit +#undef cv_hal_filter +#define cv_hal_filter ovx_hal_filter +#undef cv_hal_filterFree +#define cv_hal_filterFree ovx_hal_filterFree + +#undef cv_hal_sepFilterInit +#define cv_hal_sepFilterInit ovx_hal_sepFilterInit +#undef cv_hal_sepFilter +#define cv_hal_sepFilter ovx_hal_filter +#undef cv_hal_sepFilterFree +#define cv_hal_sepFilterFree ovx_hal_filterFree + +#undef cv_hal_morphInit +#define cv_hal_morphInit ovx_hal_morphInit +#undef cv_hal_morph +#define cv_hal_morph ovx_hal_morph +#undef cv_hal_morphFree +#define cv_hal_morphFree ovx_hal_morphFree + +#undef cv_hal_cvtBGRtoBGR +#define cv_hal_cvtBGRtoBGR ovx_hal_cvtBGRtoBGR +#undef cv_hal_cvtTwoPlaneYUVtoBGR +#define cv_hal_cvtTwoPlaneYUVtoBGR ovx_hal_cvtTwoPlaneYUVtoBGR +#undef cv_hal_cvtThreePlaneYUVtoBGR +#define cv_hal_cvtThreePlaneYUVtoBGR ovx_hal_cvtThreePlaneYUVtoBGR +#undef cv_hal_cvtBGRtoThreePlaneYUV +#define cv_hal_cvtBGRtoThreePlaneYUV ovx_hal_cvtBGRtoThreePlaneYUV +#undef cv_hal_cvtOnePlaneYUVtoBGR +#define cv_hal_cvtOnePlaneYUVtoBGR ovx_hal_cvtOnePlaneYUVtoBGR + +#endif #endif diff --git a/modules/imgproc/src/hal_replacement.hpp b/modules/imgproc/src/hal_replacement.hpp index fdf3ab72e8..3196cd88eb 100644 --- a/modules/imgproc/src/hal_replacement.hpp +++ b/modules/imgproc/src/hal_replacement.hpp @@ -263,7 +263,7 @@ inline int hal_ni_resize(int src_type, const uchar *src_data, size_t src_step, i @param dst_step destination image step @param dst_width destination image width @param dst_height destination image height - @param M 3x2 matrix with transform coefficients + @param M 2x3 matrix with transform coefficients @param interpolation interpolation mode (CV_HAL_INTER_NEAREST, ...) @param borderType border processing mode (CV_HAL_BORDER_REFLECT, ...) @param borderValue values to use for CV_HAL_BORDER_CONSTANT mode