mirror of
https://github.com/opencv/opencv.git
synced 2026-07-30 07:43:03 +04:00
Merge pull request #21157 from alexgiving:atrutnev/move_resize
GAPI: Move Resize kernel from core to imgproc * Move Resize kernel from core to imgproc * Applied style comments * Adding backward compatibility * Applied Asya PR
This commit is contained in:
@@ -462,30 +462,6 @@ GAPI_OCV_KERNEL(GCPUMerge4, cv::gapi::core::GMerge4)
|
||||
}
|
||||
};
|
||||
|
||||
GAPI_OCV_KERNEL(GCPUResize, cv::gapi::core::GResize)
|
||||
{
|
||||
static void run(const cv::Mat& in, cv::Size sz, double fx, double fy, int interp, cv::Mat &out)
|
||||
{
|
||||
cv::resize(in, out, sz, fx, fy, interp);
|
||||
}
|
||||
};
|
||||
|
||||
GAPI_OCV_KERNEL(GCPUResizeP, cv::gapi::core::GResizeP)
|
||||
{
|
||||
static void run(const cv::Mat& in, cv::Size out_sz, int interp, cv::Mat& out)
|
||||
{
|
||||
int inH = in.rows / 3;
|
||||
int inW = in.cols;
|
||||
int outH = out.rows / 3;
|
||||
int outW = out.cols;
|
||||
for (int i = 0; i < 3; i++) {
|
||||
auto in_plane = in(cv::Rect(0, i*inH, inW, inH));
|
||||
auto out_plane = out(cv::Rect(0, i*outH, outW, outH));
|
||||
cv::resize(in_plane, out_plane, out_sz, 0, 0, interp);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
GAPI_OCV_KERNEL(GCPURemap, cv::gapi::core::GRemap)
|
||||
{
|
||||
static void run(const cv::Mat& in, const cv::Mat& x, const cv::Mat& y, int a, int b, cv::Scalar s, cv::Mat& out)
|
||||
@@ -775,8 +751,6 @@ cv::GKernelPackage cv::gapi::core::cpu::kernels()
|
||||
, GCPUInRange
|
||||
, GCPUSplit3
|
||||
, GCPUSplit4
|
||||
, GCPUResize
|
||||
, GCPUResizeP
|
||||
, GCPUMerge3
|
||||
, GCPUMerge4
|
||||
, GCPURemap
|
||||
|
||||
@@ -28,6 +28,30 @@ namespace {
|
||||
}
|
||||
}
|
||||
|
||||
GAPI_OCV_KERNEL(GCPUResize, cv::gapi::imgproc::GResize)
|
||||
{
|
||||
static void run(const cv::Mat& in, cv::Size sz, double fx, double fy, int interp, cv::Mat &out)
|
||||
{
|
||||
cv::resize(in, out, sz, fx, fy, interp);
|
||||
}
|
||||
};
|
||||
|
||||
GAPI_OCV_KERNEL(GCPUResizeP, cv::gapi::imgproc::GResizeP)
|
||||
{
|
||||
static void run(const cv::Mat& in, cv::Size out_sz, int interp, cv::Mat& out)
|
||||
{
|
||||
int inH = in.rows / 3;
|
||||
int inW = in.cols;
|
||||
int outH = out.rows / 3;
|
||||
int outW = out.cols;
|
||||
for (int i = 0; i < 3; i++) {
|
||||
auto in_plane = in(cv::Rect(0, i*inH, inW, inH));
|
||||
auto out_plane = out(cv::Rect(0, i*outH, outW, outH));
|
||||
cv::resize(in_plane, out_plane, out_sz, 0, 0, interp);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
GAPI_OCV_KERNEL(GCPUSepFilter, cv::gapi::imgproc::GSepFilter)
|
||||
{
|
||||
static void run(const cv::Mat& in, int ddepth, const cv::Mat& kernX, const cv::Mat& kernY, const cv::Point& anchor, const cv::Scalar& delta,
|
||||
@@ -617,6 +641,8 @@ cv::GKernelPackage cv::gapi::imgproc::cpu::kernels()
|
||||
{
|
||||
static auto pkg = cv::gapi::kernels
|
||||
< GCPUFilter2D
|
||||
, GCPUResize
|
||||
, GCPUResizeP
|
||||
, GCPUSepFilter
|
||||
, GCPUBoxFilter
|
||||
, GCPUBlur
|
||||
|
||||
@@ -23,10 +23,6 @@
|
||||
#include <opencv2/gapi/fluid/gfluidkernel.hpp>
|
||||
#include <opencv2/gapi/fluid/core.hpp>
|
||||
|
||||
#if CV_SSE4_1
|
||||
#include "gfluidcore_simd_sse41.hpp"
|
||||
#endif
|
||||
|
||||
#include "gfluidbuffer_priv.hpp"
|
||||
#include "gfluidbackend.hpp"
|
||||
#include "gfluidutils.hpp"
|
||||
@@ -2760,301 +2756,6 @@ GAPI_FLUID_KERNEL(GFluidPhase, cv::gapi::core::GPhase, false)
|
||||
}
|
||||
};
|
||||
|
||||
template<typename T, typename Mapper, int chanNum>
|
||||
struct LinearScratchDesc {
|
||||
using alpha_t = typename Mapper::alpha_type;
|
||||
using index_t = typename Mapper::index_type;
|
||||
|
||||
alpha_t* alpha;
|
||||
alpha_t* clone;
|
||||
index_t* mapsx;
|
||||
alpha_t* beta;
|
||||
index_t* mapsy;
|
||||
T* tmp;
|
||||
|
||||
LinearScratchDesc(int /*inW*/, int /*inH*/, int outW, int outH, void* data) {
|
||||
alpha = reinterpret_cast<alpha_t*>(data);
|
||||
clone = reinterpret_cast<alpha_t*>(alpha + outW);
|
||||
mapsx = reinterpret_cast<index_t*>(clone + outW*4);
|
||||
beta = reinterpret_cast<alpha_t*>(mapsx + outW);
|
||||
mapsy = reinterpret_cast<index_t*>(beta + outH);
|
||||
tmp = reinterpret_cast<T*> (mapsy + outH*2);
|
||||
}
|
||||
|
||||
static int bufSize(int inW, int /*inH*/, int outW, int outH, int lpi) {
|
||||
auto size = outW * sizeof(alpha_t) +
|
||||
outW * sizeof(alpha_t) * 4 + // alpha clones
|
||||
outW * sizeof(index_t) +
|
||||
outH * sizeof(alpha_t) +
|
||||
outH * sizeof(index_t) * 2 +
|
||||
inW * sizeof(T) * lpi * chanNum;
|
||||
|
||||
return static_cast<int>(size);
|
||||
}
|
||||
};
|
||||
static inline double invRatio(int inSz, int outSz) {
|
||||
return static_cast<double>(outSz) / inSz;
|
||||
}
|
||||
|
||||
static inline double ratio(int inSz, int outSz) {
|
||||
return 1 / invRatio(inSz, outSz);
|
||||
}
|
||||
|
||||
template<typename T, typename Mapper, int chanNum = 1>
|
||||
static inline void initScratchLinear(const cv::GMatDesc& in,
|
||||
const Size& outSz,
|
||||
cv::gapi::fluid::Buffer& scratch,
|
||||
int lpi) {
|
||||
using alpha_type = typename Mapper::alpha_type;
|
||||
static const auto unity = Mapper::unity;
|
||||
|
||||
auto inSz = in.size;
|
||||
auto sbufsize = LinearScratchDesc<T, Mapper, chanNum>::bufSize(inSz.width, inSz.height, outSz.width, outSz.height, lpi);
|
||||
|
||||
Size scratch_size{sbufsize, 1};
|
||||
|
||||
cv::GMatDesc desc;
|
||||
desc.chan = 1;
|
||||
desc.depth = CV_8UC1;
|
||||
desc.size = scratch_size;
|
||||
|
||||
cv::gapi::fluid::Buffer buffer(desc);
|
||||
scratch = std::move(buffer);
|
||||
|
||||
double hRatio = ratio(in.size.width, outSz.width);
|
||||
double vRatio = ratio(in.size.height, outSz.height);
|
||||
|
||||
LinearScratchDesc<T, Mapper, chanNum> scr(inSz.width, inSz.height, outSz.width, outSz.height, scratch.OutLineB());
|
||||
|
||||
auto *alpha = scr.alpha;
|
||||
auto *clone = scr.clone;
|
||||
auto *index = scr.mapsx;
|
||||
|
||||
for (int x = 0; x < outSz.width; x++) {
|
||||
auto map = Mapper::map(hRatio, 0, in.size.width, x);
|
||||
auto alpha0 = map.alpha0;
|
||||
auto index0 = map.index0;
|
||||
|
||||
// TRICK:
|
||||
// Algorithm takes pair of input pixels, sx0'th and sx1'th,
|
||||
// and compute result as alpha0*src[sx0] + alpha1*src[sx1].
|
||||
// By definition: sx1 == sx0 + 1 either sx1 == sx0, and
|
||||
// alpha0 + alpha1 == unity (scaled appropriately).
|
||||
// Here we modify formulas for alpha0 and sx1: by assuming
|
||||
// that sx1 == sx0 + 1 always, and patching alpha0 so that
|
||||
// result remains intact.
|
||||
// Note that we need in.size.width >= 2, for both sx0 and
|
||||
// sx0+1 were indexing pixels inside the input's width.
|
||||
if (map.index1 != map.index0 + 1) {
|
||||
GAPI_DbgAssert(map.index1 == map.index0);
|
||||
GAPI_DbgAssert(in.size.width >= 2);
|
||||
if (map.index0 < in.size.width-1) {
|
||||
// sx1=sx0+1 fits inside row,
|
||||
// make sure alpha0=unity and alpha1=0,
|
||||
// so that result equals src[sx0]*unity
|
||||
alpha0 = saturate_cast<alpha_type>(unity);
|
||||
} else {
|
||||
// shift sx0 to left by 1 pixel,
|
||||
// and make sure that alpha0=0 and alpha1==1,
|
||||
// so that result equals to src[sx0+1]*unity
|
||||
alpha0 = 0;
|
||||
index0--;
|
||||
}
|
||||
}
|
||||
|
||||
alpha[x] = alpha0;
|
||||
index[x] = index0;
|
||||
|
||||
for (int l = 0; l < 4; l++) {
|
||||
clone[4*x + l] = alpha0;
|
||||
}
|
||||
}
|
||||
|
||||
auto *beta = scr.beta;
|
||||
auto *index_y = scr.mapsy;
|
||||
|
||||
for (int y = 0; y < outSz.height; y++) {
|
||||
auto mapY = Mapper::map(vRatio, 0, in.size.height, y);
|
||||
beta[y] = mapY.alpha0;
|
||||
index_y[y] = mapY.index0;
|
||||
index_y[outSz.height + y] = mapY.index1;
|
||||
}
|
||||
}
|
||||
|
||||
template<typename F, typename I>
|
||||
struct MapperUnit {
|
||||
F alpha0, alpha1;
|
||||
I index0, index1;
|
||||
};
|
||||
|
||||
inline static uint8_t calc(short alpha0, uint8_t src0, short alpha1, uint8_t src1) {
|
||||
constexpr static const int half = 1 << 14;
|
||||
return (src0 * alpha0 + src1 * alpha1 + half) >> 15;
|
||||
}
|
||||
struct Mapper {
|
||||
constexpr static const int ONE = 1 << 15;
|
||||
typedef short alpha_type;
|
||||
typedef short index_type;
|
||||
constexpr static const int unity = ONE;
|
||||
|
||||
typedef MapperUnit<short, short> Unit;
|
||||
|
||||
static inline Unit map(double ratio, int start, int max, int outCoord) {
|
||||
float f = static_cast<float>((outCoord + 0.5) * ratio - 0.5);
|
||||
int s = cvFloor(f);
|
||||
f -= s;
|
||||
|
||||
Unit u;
|
||||
|
||||
u.index0 = static_cast<short>(std::max(s - start, 0));
|
||||
u.index1 = static_cast<short>(((f == 0.0) || s + 1 >= max) ? s - start : s - start + 1);
|
||||
|
||||
u.alpha0 = saturate_cast<short>(ONE * (1.0f - f));
|
||||
u.alpha1 = saturate_cast<short>(ONE * f);
|
||||
|
||||
return u;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename T, class Mapper, int numChan>
|
||||
static void calcRowLinearC(const cv::gapi::fluid::View & in,
|
||||
cv::gapi::fluid::Buffer& out,
|
||||
cv::gapi::fluid::Buffer& scratch) {
|
||||
using alpha_type = typename Mapper::alpha_type;
|
||||
|
||||
auto inSz = in.meta().size;
|
||||
auto outSz = out.meta().size;
|
||||
|
||||
auto inY = in.y();
|
||||
int outY = out.y();
|
||||
int lpi = out.lpi();
|
||||
|
||||
GAPI_DbgAssert(outY + lpi <= outSz.height);
|
||||
GAPI_DbgAssert(lpi <= 4);
|
||||
|
||||
LinearScratchDesc<T, Mapper, numChan> scr(inSz.width, inSz.height, outSz.width, outSz.height, scratch.OutLineB());
|
||||
|
||||
const auto *alpha = scr.alpha;
|
||||
const auto *mapsx = scr.mapsx;
|
||||
const auto *beta_0 = scr.beta;
|
||||
const auto *mapsy = scr.mapsy;
|
||||
|
||||
const auto *beta = beta_0 + outY;
|
||||
const T *src0[4];
|
||||
const T *src1[4];
|
||||
T* dst[4];
|
||||
|
||||
for (int l = 0; l < lpi; l++) {
|
||||
auto index0 = mapsy[outY + l] - inY;
|
||||
auto index1 = mapsy[outSz.height + outY + l] - inY;
|
||||
src0[l] = in.InLine<const T>(index0);
|
||||
src1[l] = in.InLine<const T>(index1);
|
||||
dst[l] = out.OutLine<T>(l);
|
||||
}
|
||||
|
||||
#if 0 // Disabling SSE4.1 path due to Valgrind issues: https://github.com/opencv/opencv/issues/21097
|
||||
#if CV_SSE4_1
|
||||
const auto* clone = scr.clone;
|
||||
auto* tmp = scr.tmp;
|
||||
|
||||
if (inSz.width >= 16 && outSz.width >= 16)
|
||||
{
|
||||
sse42::calcRowLinear_8UC_Impl_<numChan>(reinterpret_cast<uint8_t**>(dst),
|
||||
reinterpret_cast<const uint8_t**>(src0),
|
||||
reinterpret_cast<const uint8_t**>(src1),
|
||||
reinterpret_cast<const short*>(alpha),
|
||||
reinterpret_cast<const short*>(clone),
|
||||
reinterpret_cast<const short*>(mapsx),
|
||||
reinterpret_cast<const short*>(beta),
|
||||
reinterpret_cast<uint8_t*>(tmp),
|
||||
inSz, outSz, lpi);
|
||||
|
||||
return;
|
||||
}
|
||||
#endif // CV_SSE4_1
|
||||
#endif
|
||||
int length = out.length();
|
||||
for (int l = 0; l < lpi; l++) {
|
||||
constexpr static const auto unity = Mapper::unity;
|
||||
|
||||
auto beta0 = beta[l];
|
||||
auto beta1 = saturate_cast<alpha_type>(unity - beta[l]);
|
||||
|
||||
for (int x = 0; x < length; x++) {
|
||||
auto alpha0 = alpha[x];
|
||||
auto alpha1 = saturate_cast<alpha_type>(unity - alpha[x]);
|
||||
auto sx0 = mapsx[x];
|
||||
auto sx1 = sx0 + 1;
|
||||
|
||||
for (int c = 0; c < numChan; c++) {
|
||||
auto idx0 = numChan*sx0 + c;
|
||||
auto idx1 = numChan*sx1 + c;
|
||||
T tmp0 = calc(beta0, src0[l][idx0], beta1, src1[l][idx0]);
|
||||
T tmp1 = calc(beta0, src0[l][idx1], beta1, src1[l][idx1]);
|
||||
dst[l][numChan * x + c] = calc(alpha0, tmp0, alpha1, tmp1);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
GAPI_FLUID_KERNEL(GFluidResize, cv::gapi::core::GResize, true)
|
||||
{
|
||||
static const int Window = 1;
|
||||
static const int LPI = 4;
|
||||
static const auto Kind = GFluidKernel::Kind::Resize;
|
||||
|
||||
constexpr static const int INTER_RESIZE_COEF_BITS = 11;
|
||||
constexpr static const int INTER_RESIZE_COEF_SCALE = 1 << INTER_RESIZE_COEF_BITS;
|
||||
constexpr static const short ONE = INTER_RESIZE_COEF_SCALE;
|
||||
|
||||
static void initScratch(const cv::GMatDesc& in,
|
||||
cv::Size outSz, double fx, double fy, int /*interp*/,
|
||||
cv::gapi::fluid::Buffer &scratch)
|
||||
{
|
||||
int outSz_w;
|
||||
int outSz_h;
|
||||
if (outSz.width == 0 || outSz.height == 0)
|
||||
{
|
||||
outSz_w = static_cast<int>(round(in.size.width * fx));
|
||||
outSz_h = static_cast<int>(round(in.size.height * fy));
|
||||
}
|
||||
else
|
||||
{
|
||||
outSz_w = outSz.width;
|
||||
outSz_h = outSz.height;
|
||||
}
|
||||
cv::Size outSize(outSz_w, outSz_h);
|
||||
|
||||
if (in.chan == 3)
|
||||
{
|
||||
initScratchLinear<uchar, Mapper, 3>(in, outSize, scratch, LPI);
|
||||
}
|
||||
else if (in.chan == 4)
|
||||
{
|
||||
initScratchLinear<uchar, Mapper, 4>(in, outSize, scratch, LPI);
|
||||
}
|
||||
}
|
||||
|
||||
static void resetScratch(cv::gapi::fluid::Buffer& /*scratch*/)
|
||||
{}
|
||||
|
||||
static void run(const cv::gapi::fluid::View& in, cv::Size /*sz*/, double /*fx*/, double /*fy*/, int interp,
|
||||
cv::gapi::fluid::Buffer& out,
|
||||
cv::gapi::fluid::Buffer& scratch) {
|
||||
const int channels = in.meta().chan;
|
||||
GAPI_Assert((channels == 3 || channels == 4) && (interp == cv::INTER_LINEAR));
|
||||
|
||||
if (channels == 3)
|
||||
{
|
||||
calcRowLinearC<uint8_t, Mapper, 3>(in, out, scratch);
|
||||
}
|
||||
else if (channels == 4)
|
||||
{
|
||||
calcRowLinearC<uint8_t, Mapper, 4>(in, out, scratch);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
GAPI_FLUID_KERNEL(GFluidSqrt, cv::gapi::core::GSqrt, false)
|
||||
{
|
||||
static const int Window = 1;
|
||||
@@ -3134,7 +2835,6 @@ cv::GKernelPackage cv::gapi::core::fluid::kernels()
|
||||
,GFluidCmpNEScalar
|
||||
,GFluidThreshold
|
||||
,GFluidInRange
|
||||
,GFluidResize
|
||||
,GFluidSqrt
|
||||
#if 0
|
||||
,GFluidMean -- not fluid
|
||||
|
||||
@@ -25,6 +25,10 @@
|
||||
|
||||
#include "gfluidimgproc_func.hpp"
|
||||
|
||||
#if CV_SSE4_1
|
||||
#include "gfluidcore_simd_sse41.hpp"
|
||||
#endif
|
||||
|
||||
#include <opencv2/imgproc/hal/hal.hpp>
|
||||
#include <opencv2/core/hal/intrin.hpp>
|
||||
|
||||
@@ -1821,6 +1825,301 @@ GAPI_FLUID_KERNEL(GFluidBayerGR2RGB, cv::gapi::imgproc::GBayerGR2RGB, false)
|
||||
}
|
||||
};
|
||||
|
||||
template<typename T, typename Mapper, int chanNum>
|
||||
struct LinearScratchDesc {
|
||||
using alpha_t = typename Mapper::alpha_type;
|
||||
using index_t = typename Mapper::index_type;
|
||||
|
||||
alpha_t* alpha;
|
||||
alpha_t* clone;
|
||||
index_t* mapsx;
|
||||
alpha_t* beta;
|
||||
index_t* mapsy;
|
||||
T* tmp;
|
||||
|
||||
LinearScratchDesc(int /*inW*/, int /*inH*/, int outW, int outH, void* data) {
|
||||
alpha = reinterpret_cast<alpha_t*>(data);
|
||||
clone = reinterpret_cast<alpha_t*>(alpha + outW);
|
||||
mapsx = reinterpret_cast<index_t*>(clone + outW*4);
|
||||
beta = reinterpret_cast<alpha_t*>(mapsx + outW);
|
||||
mapsy = reinterpret_cast<index_t*>(beta + outH);
|
||||
tmp = reinterpret_cast<T*> (mapsy + outH*2);
|
||||
}
|
||||
|
||||
static int bufSize(int inW, int /*inH*/, int outW, int outH, int lpi) {
|
||||
auto size = outW * sizeof(alpha_t) +
|
||||
outW * sizeof(alpha_t) * 4 + // alpha clones
|
||||
outW * sizeof(index_t) +
|
||||
outH * sizeof(alpha_t) +
|
||||
outH * sizeof(index_t) * 2 +
|
||||
inW * sizeof(T) * lpi * chanNum;
|
||||
|
||||
return static_cast<int>(size);
|
||||
}
|
||||
};
|
||||
static inline double invRatio(int inSz, int outSz) {
|
||||
return static_cast<double>(outSz) / inSz;
|
||||
}
|
||||
|
||||
static inline double ratio(int inSz, int outSz) {
|
||||
return 1 / invRatio(inSz, outSz);
|
||||
}
|
||||
|
||||
template<typename T, typename Mapper, int chanNum = 1>
|
||||
static inline void initScratchLinear(const cv::GMatDesc& in,
|
||||
const Size& outSz,
|
||||
cv::gapi::fluid::Buffer& scratch,
|
||||
int lpi) {
|
||||
using alpha_type = typename Mapper::alpha_type;
|
||||
static const auto unity = Mapper::unity;
|
||||
|
||||
auto inSz = in.size;
|
||||
auto sbufsize = LinearScratchDesc<T, Mapper, chanNum>::bufSize(inSz.width, inSz.height, outSz.width, outSz.height, lpi);
|
||||
|
||||
Size scratch_size{sbufsize, 1};
|
||||
|
||||
cv::GMatDesc desc;
|
||||
desc.chan = 1;
|
||||
desc.depth = CV_8UC1;
|
||||
desc.size = scratch_size;
|
||||
|
||||
cv::gapi::fluid::Buffer buffer(desc);
|
||||
scratch = std::move(buffer);
|
||||
|
||||
double hRatio = ratio(in.size.width, outSz.width);
|
||||
double vRatio = ratio(in.size.height, outSz.height);
|
||||
|
||||
LinearScratchDesc<T, Mapper, chanNum> scr(inSz.width, inSz.height, outSz.width, outSz.height, scratch.OutLineB());
|
||||
|
||||
auto *alpha = scr.alpha;
|
||||
auto *clone = scr.clone;
|
||||
auto *index = scr.mapsx;
|
||||
|
||||
for (int x = 0; x < outSz.width; x++) {
|
||||
auto map = Mapper::map(hRatio, 0, in.size.width, x);
|
||||
auto alpha0 = map.alpha0;
|
||||
auto index0 = map.index0;
|
||||
|
||||
// TRICK:
|
||||
// Algorithm takes pair of input pixels, sx0'th and sx1'th,
|
||||
// and compute result as alpha0*src[sx0] + alpha1*src[sx1].
|
||||
// By definition: sx1 == sx0 + 1 either sx1 == sx0, and
|
||||
// alpha0 + alpha1 == unity (scaled appropriately).
|
||||
// Here we modify formulas for alpha0 and sx1: by assuming
|
||||
// that sx1 == sx0 + 1 always, and patching alpha0 so that
|
||||
// result remains intact.
|
||||
// Note that we need in.size.width >= 2, for both sx0 and
|
||||
// sx0+1 were indexing pixels inside the input's width.
|
||||
if (map.index1 != map.index0 + 1) {
|
||||
GAPI_DbgAssert(map.index1 == map.index0);
|
||||
GAPI_DbgAssert(in.size.width >= 2);
|
||||
if (map.index0 < in.size.width-1) {
|
||||
// sx1=sx0+1 fits inside row,
|
||||
// make sure alpha0=unity and alpha1=0,
|
||||
// so that result equals src[sx0]*unity
|
||||
alpha0 = saturate_cast<alpha_type>(unity);
|
||||
} else {
|
||||
// shift sx0 to left by 1 pixel,
|
||||
// and make sure that alpha0=0 and alpha1==1,
|
||||
// so that result equals to src[sx0+1]*unity
|
||||
alpha0 = 0;
|
||||
index0--;
|
||||
}
|
||||
}
|
||||
|
||||
alpha[x] = alpha0;
|
||||
index[x] = index0;
|
||||
|
||||
for (int l = 0; l < 4; l++) {
|
||||
clone[4*x + l] = alpha0;
|
||||
}
|
||||
}
|
||||
|
||||
auto *beta = scr.beta;
|
||||
auto *index_y = scr.mapsy;
|
||||
|
||||
for (int y = 0; y < outSz.height; y++) {
|
||||
auto mapY = Mapper::map(vRatio, 0, in.size.height, y);
|
||||
beta[y] = mapY.alpha0;
|
||||
index_y[y] = mapY.index0;
|
||||
index_y[outSz.height + y] = mapY.index1;
|
||||
}
|
||||
}
|
||||
|
||||
template<typename F, typename I>
|
||||
struct MapperUnit {
|
||||
F alpha0, alpha1;
|
||||
I index0, index1;
|
||||
};
|
||||
|
||||
inline static uint8_t calc(short alpha0, uint8_t src0, short alpha1, uint8_t src1) {
|
||||
constexpr static const int half = 1 << 14;
|
||||
return (src0 * alpha0 + src1 * alpha1 + half) >> 15;
|
||||
}
|
||||
struct Mapper {
|
||||
constexpr static const int ONE = 1 << 15;
|
||||
typedef short alpha_type;
|
||||
typedef short index_type;
|
||||
constexpr static const int unity = ONE;
|
||||
|
||||
typedef MapperUnit<short, short> Unit;
|
||||
|
||||
static inline Unit map(double ratio, int start, int max, int outCoord) {
|
||||
float f = static_cast<float>((outCoord + 0.5) * ratio - 0.5);
|
||||
int s = cvFloor(f);
|
||||
f -= s;
|
||||
|
||||
Unit u;
|
||||
|
||||
u.index0 = static_cast<short>(std::max(s - start, 0));
|
||||
u.index1 = static_cast<short>(((f == 0.0) || s + 1 >= max) ? s - start : s - start + 1);
|
||||
|
||||
u.alpha0 = saturate_cast<short>(ONE * (1.0f - f));
|
||||
u.alpha1 = saturate_cast<short>(ONE * f);
|
||||
|
||||
return u;
|
||||
}
|
||||
};
|
||||
|
||||
template<typename T, class Mapper, int numChan>
|
||||
static void calcRowLinearC(const cv::gapi::fluid::View & in,
|
||||
cv::gapi::fluid::Buffer& out,
|
||||
cv::gapi::fluid::Buffer& scratch) {
|
||||
using alpha_type = typename Mapper::alpha_type;
|
||||
|
||||
auto inSz = in.meta().size;
|
||||
auto outSz = out.meta().size;
|
||||
|
||||
auto inY = in.y();
|
||||
int outY = out.y();
|
||||
int lpi = out.lpi();
|
||||
|
||||
GAPI_DbgAssert(outY + lpi <= outSz.height);
|
||||
GAPI_DbgAssert(lpi <= 4);
|
||||
|
||||
LinearScratchDesc<T, Mapper, numChan> scr(inSz.width, inSz.height, outSz.width, outSz.height, scratch.OutLineB());
|
||||
|
||||
const auto *alpha = scr.alpha;
|
||||
const auto *mapsx = scr.mapsx;
|
||||
const auto *beta_0 = scr.beta;
|
||||
const auto *mapsy = scr.mapsy;
|
||||
|
||||
const auto *beta = beta_0 + outY;
|
||||
const T *src0[4];
|
||||
const T *src1[4];
|
||||
T* dst[4];
|
||||
|
||||
for (int l = 0; l < lpi; l++) {
|
||||
auto index0 = mapsy[outY + l] - inY;
|
||||
auto index1 = mapsy[outSz.height + outY + l] - inY;
|
||||
src0[l] = in.InLine<const T>(index0);
|
||||
src1[l] = in.InLine<const T>(index1);
|
||||
dst[l] = out.OutLine<T>(l);
|
||||
}
|
||||
|
||||
#if 0 // Disabling SSE4.1 path due to Valgrind issues: https://github.com/opencv/opencv/issues/21097
|
||||
#if CV_SSE4_1
|
||||
const auto* clone = scr.clone;
|
||||
auto* tmp = scr.tmp;
|
||||
|
||||
if (inSz.width >= 16 && outSz.width >= 16)
|
||||
{
|
||||
sse42::calcRowLinear_8UC_Impl_<numChan>(reinterpret_cast<uint8_t**>(dst),
|
||||
reinterpret_cast<const uint8_t**>(src0),
|
||||
reinterpret_cast<const uint8_t**>(src1),
|
||||
reinterpret_cast<const short*>(alpha),
|
||||
reinterpret_cast<const short*>(clone),
|
||||
reinterpret_cast<const short*>(mapsx),
|
||||
reinterpret_cast<const short*>(beta),
|
||||
reinterpret_cast<uint8_t*>(tmp),
|
||||
inSz, outSz, lpi);
|
||||
|
||||
return;
|
||||
}
|
||||
#endif // CV_SSE4_1
|
||||
#endif
|
||||
int length = out.length();
|
||||
for (int l = 0; l < lpi; l++) {
|
||||
constexpr static const auto unity = Mapper::unity;
|
||||
|
||||
auto beta0 = beta[l];
|
||||
auto beta1 = saturate_cast<alpha_type>(unity - beta[l]);
|
||||
|
||||
for (int x = 0; x < length; x++) {
|
||||
auto alpha0 = alpha[x];
|
||||
auto alpha1 = saturate_cast<alpha_type>(unity - alpha[x]);
|
||||
auto sx0 = mapsx[x];
|
||||
auto sx1 = sx0 + 1;
|
||||
|
||||
for (int c = 0; c < numChan; c++) {
|
||||
auto idx0 = numChan*sx0 + c;
|
||||
auto idx1 = numChan*sx1 + c;
|
||||
T tmp0 = calc(beta0, src0[l][idx0], beta1, src1[l][idx0]);
|
||||
T tmp1 = calc(beta0, src0[l][idx1], beta1, src1[l][idx1]);
|
||||
dst[l][numChan * x + c] = calc(alpha0, tmp0, alpha1, tmp1);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
GAPI_FLUID_KERNEL(GFluidResize, cv::gapi::imgproc::GResize, true)
|
||||
{
|
||||
static const int Window = 1;
|
||||
static const int LPI = 4;
|
||||
static const auto Kind = GFluidKernel::Kind::Resize;
|
||||
|
||||
constexpr static const int INTER_RESIZE_COEF_BITS = 11;
|
||||
constexpr static const int INTER_RESIZE_COEF_SCALE = 1 << INTER_RESIZE_COEF_BITS;
|
||||
constexpr static const short ONE = INTER_RESIZE_COEF_SCALE;
|
||||
|
||||
static void initScratch(const cv::GMatDesc& in,
|
||||
cv::Size outSz, double fx, double fy, int /*interp*/,
|
||||
cv::gapi::fluid::Buffer &scratch)
|
||||
{
|
||||
int outSz_w;
|
||||
int outSz_h;
|
||||
if (outSz.width == 0 || outSz.height == 0)
|
||||
{
|
||||
outSz_w = static_cast<int>(round(in.size.width * fx));
|
||||
outSz_h = static_cast<int>(round(in.size.height * fy));
|
||||
}
|
||||
else
|
||||
{
|
||||
outSz_w = outSz.width;
|
||||
outSz_h = outSz.height;
|
||||
}
|
||||
cv::Size outSize(outSz_w, outSz_h);
|
||||
|
||||
if (in.chan == 3)
|
||||
{
|
||||
initScratchLinear<uchar, Mapper, 3>(in, outSize, scratch, LPI);
|
||||
}
|
||||
else if (in.chan == 4)
|
||||
{
|
||||
initScratchLinear<uchar, Mapper, 4>(in, outSize, scratch, LPI);
|
||||
}
|
||||
}
|
||||
|
||||
static void resetScratch(cv::gapi::fluid::Buffer& /*scratch*/)
|
||||
{}
|
||||
|
||||
static void run(const cv::gapi::fluid::View& in, cv::Size /*sz*/, double /*fx*/, double /*fy*/, int interp,
|
||||
cv::gapi::fluid::Buffer& out,
|
||||
cv::gapi::fluid::Buffer& scratch) {
|
||||
const int channels = in.meta().chan;
|
||||
GAPI_Assert((channels == 3 || channels == 4) && (interp == cv::INTER_LINEAR));
|
||||
|
||||
if (channels == 3)
|
||||
{
|
||||
calcRowLinearC<uint8_t, Mapper, 3>(in, out, scratch);
|
||||
}
|
||||
else if (channels == 4)
|
||||
{
|
||||
calcRowLinearC<uint8_t, Mapper, 4>(in, out, scratch);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace fluid
|
||||
} // namespace gapi
|
||||
} // namespace cv
|
||||
@@ -1831,6 +2130,7 @@ cv::GKernelPackage cv::gapi::imgproc::fluid::kernels()
|
||||
|
||||
return cv::gapi::kernels
|
||||
< GFluidBGR2Gray
|
||||
, GFluidResize
|
||||
, GFluidRGB2Gray
|
||||
, GFluidRGB2GrayCustom
|
||||
, GFluidRGB2YUV
|
||||
|
||||
@@ -458,14 +458,6 @@ GAPI_OCL_KERNEL(GOCLMerge4, cv::gapi::core::GMerge4)
|
||||
}
|
||||
};
|
||||
|
||||
GAPI_OCL_KERNEL(GOCLResize, cv::gapi::core::GResize)
|
||||
{
|
||||
static void run(const cv::UMat& in, cv::Size sz, double fx, double fy, int interp, cv::UMat &out)
|
||||
{
|
||||
cv::resize(in, out, sz, fx, fy, interp);
|
||||
}
|
||||
};
|
||||
|
||||
GAPI_OCL_KERNEL(GOCLRemap, cv::gapi::core::GRemap)
|
||||
{
|
||||
static void run(const cv::UMat& in, const cv::Mat& x, const cv::Mat& y, int a, int b, cv::Scalar s, cv::UMat& out)
|
||||
@@ -585,7 +577,6 @@ cv::GKernelPackage cv::gapi::core::ocl::kernels()
|
||||
, GOCLInRange
|
||||
, GOCLSplit3
|
||||
, GOCLSplit4
|
||||
, GOCLResize
|
||||
, GOCLMerge3
|
||||
, GOCLMerge4
|
||||
, GOCLRemap
|
||||
|
||||
@@ -11,6 +11,13 @@
|
||||
#include <opencv2/gapi/ocl/imgproc.hpp>
|
||||
#include "backends/ocl/goclimgproc.hpp"
|
||||
|
||||
GAPI_OCL_KERNEL(GOCLResize, cv::gapi::imgproc::GResize)
|
||||
{
|
||||
static void run(const cv::UMat& in, cv::Size sz, double fx, double fy, int interp, cv::UMat &out)
|
||||
{
|
||||
cv::resize(in, out, sz, fx, fy, interp);
|
||||
}
|
||||
};
|
||||
|
||||
GAPI_OCL_KERNEL(GOCLSepFilter, cv::gapi::imgproc::GSepFilter)
|
||||
{
|
||||
@@ -270,6 +277,7 @@ cv::GKernelPackage cv::gapi::imgproc::ocl::kernels()
|
||||
{
|
||||
static auto pkg = cv::gapi::kernels
|
||||
< GOCLFilter2D
|
||||
, GOCLResize
|
||||
, GOCLSepFilter
|
||||
, GOCLBoxFilter
|
||||
, GOCLBlur
|
||||
|
||||
Reference in New Issue
Block a user