mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-07-31 09:57:28 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Merge pull request #27067 from amane-ame:sepfilter_optimize

Browse Source 
Optimize RISC-V HAL cv::sepFilter
This commit is contained in:
Alexander Smorkalov 2025-03-17 09:21:33 +03:00 committed by GitHub
commit 0a39f98bee
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 116 additions and 134 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

250
3rdparty/hal_rvv/hal_rvv_1p0/filter.hpp vendored
View File

@ -41,10 +41,10 @@ private:
};
template<typename... Args>
static inline int invoke(int start, int end, std::function<int(int, int, Args...)> func, Args&&... args)
static inline int invoke(int height, std::function<int(int, int, Args...)> func, Args&&... args)
{
cv::parallel_for_(Range(start + 1, end), FilterInvoker(func, std::forward<Args>(args)...), cv::getNumThreads());
return func(start, start + 1, std::forward<Args>(args)...);
cv::parallel_for_(Range(1, height), FilterInvoker(func, std::forward<Args>(args)...), cv::getNumThreads());
return func(0, 1, std::forward<Args>(args)...);
}
static inline int borderInterpolate( int p, int len, int borderType )
@ -182,10 +182,10 @@ static void process5(int anchor, int left, int right, float delta, const float*
auto v3 = __riscv_vfwcvt_f(__riscv_vwcvtu_x(__riscv_vget_v_u8m1x4_u8m1(src, 3), vl), vl);
const uchar* extra = row + (i + vl - anchor) * 4;
s0 = addshift(s0, v0, k0, k1, k2, k3, k4, *(extra ), *(extra + 4), *(extra + 8), *(extra + 12));
s1 = addshift(s1, v1, k0, k1, k2, k3, k4, *(extra + 1), *(extra + 5), *(extra + 9), *(extra + 13));
s2 = addshift(s2, v2, k0, k1, k2, k3, k4, *(extra + 2), *(extra + 6), *(extra + 10), *(extra + 14));
s3 = addshift(s3, v3, k0, k1, k2, k3, k4, *(extra + 3), *(extra + 7), *(extra + 11), *(extra + 15));
s0 = addshift(s0, v0, k0, k1, k2, k3, k4, extra[0], extra[4], extra[ 8], extra[12]);
s1 = addshift(s1, v1, k0, k1, k2, k3, k4, extra[1], extra[5], extra[ 9], extra[13]);
s2 = addshift(s2, v2, k0, k1, k2, k3, k4, extra[2], extra[6], extra[10], extra[14]);
s3 = addshift(s3, v3, k0, k1, k2, k3, k4, extra[3], extra[7], extra[11], extra[15]);
};
loadsrc(row0, kernel[ 0], kernel[ 1], kernel[ 2], kernel[ 3], kernel[ 4]);
@ -253,9 +253,9 @@ static inline int filter(int start, int end, Filter2D* data, const uchar* src_da
dst_data[(x * width + y) * 4 + 3] = std::max(0, std::min((int)std::round(sum3), (int)std::numeric_limits<uchar>::max()));
};
const int left = data->anchor_x, right = width - (ksize - 1 - data->anchor_x);
for (int i = start; i < end; i++)
{
const int left = ksize - 1, right = width - (ksize - 1);
if (left >= right)
{
for (int j = 0; j < width; j++)
@ -296,10 +296,10 @@ inline int filter(cvhalFilter2D* context, uchar* src_data, size_t src_step, ucha
switch (data->kernel_width)
{
case 3:
res = invoke(0, height, {filter<3>}, data, src_data, src_step, dst.data(), width, height, full_width, full_height, offset_x, offset_y);
res = invoke(height, {filter<3>}, data, src_data, src_step, dst.data(), width, height, full_width, full_height, offset_x, offset_y);
break;
case 5:
res = invoke(0, height, {filter<5>}, data, src_data, src_step, dst.data(), width, height, full_width, full_height, offset_x, offset_y);
res = invoke(height, {filter<5>}, data, src_data, src_step, dst.data(), width, height, full_width, full_height, offset_x, offset_y);
break;
}
@ -358,85 +358,10 @@ inline int sepFilterInit(cvhalFilter2D **context, int src_type, int dst_type, in
// the algorithm is copied from 3rdparty/carotene/src/separable_filter.hpp,
// in the functor RowFilter3x3S16Generic and ColFilter3x3S16Generic
template<int ksize>
static inline int sepFilterRow(int start, int end, sepFilter2D* data, const uchar* src_data, size_t src_step, float* dst_data, int width, int full_width, int offset_x)
static inline int sepFilter(int start, int end, sepFilter2D* data, const uchar* src_data, size_t src_step, uchar* dst_data, size_t dst_step, int width, int height, int full_width, int full_height, int offset_x, int offset_y)
{
constexpr int noval = std::numeric_limits<int>::max();
auto access = [&](int y) {
int pj;
if (data->borderType & BORDER_ISOLATED)
{
pj = filter::borderInterpolate(y - data->anchor_x, width, data->borderType & ~BORDER_ISOLATED);
pj = pj < 0 ? noval : pj;
}
else
{
pj = filter::borderInterpolate(offset_x + y - data->anchor_x, full_width, data->borderType);
pj = pj < 0 ? noval : pj - offset_x;
}
return pj;
};
const float* kx = reinterpret_cast<const float*>(data->kernelx_data);
auto process = [&](int x, int y) {
float sum = 0;
for (int i = 0; i < ksize; i++)
{
int p = access(y + i);
if (p != noval)
{
sum += kx[i] * src_data[x * src_step + p];
}
}
dst_data[x * width + y] = sum;
};
for (int i = start; i < end; i++)
{
const int left = ksize - 1, right = width - (ksize - 1);
if (left >= right)
{
for (int j = 0; j < width; j++)
process(i, j);
}
else
{
for (int j = 0; j < left; j++)
process(i, j);
for (int j = right; j < width; j++)
process(i, j);
int vl;
for (int j = left; j < right; j += vl)
{
vl = __riscv_vsetvl_e8m2(right - j);
const uchar* extra = src_data + i * src_step + j - data->anchor_x;
auto sum = __riscv_vfmv_v_f_f32m8(0, vl);
auto src = __riscv_vfwcvt_f(__riscv_vwcvtu_x(__riscv_vle8_v_u8m2(extra, vl), vl), vl);
sum = __riscv_vfmacc(sum, kx[0], src, vl);
src = __riscv_vfslide1down(src, extra[vl], vl);
sum = __riscv_vfmacc(sum, kx[1], src, vl);
src = __riscv_vfslide1down(src, extra[vl + 1], vl);
sum = __riscv_vfmacc(sum, kx[2], src, vl);
if (ksize == 5)
{
src = __riscv_vfslide1down(src, extra[vl + 2], vl);
sum = __riscv_vfmacc(sum, kx[3], src, vl);
src = __riscv_vfslide1down(src, extra[vl + 3], vl);
sum = __riscv_vfmacc(sum, kx[4], src, vl);
}
__riscv_vse32(dst_data + i * width + j, sum, vl);
}
}
}
return CV_HAL_ERROR_OK;
}
template<int ksize>
static inline int sepFilterCol(int start, int end, sepFilter2D* data, const float* src_data, uchar* dst_data, size_t dst_step, int width, int height, int full_height, int offset_y)
{
constexpr int noval = std::numeric_limits<int>::max();
auto access = [&](int x) {
auto accessX = [&](int x) {
int pi;
if (data->borderType & BORDER_ISOLATED)
{
@ -450,42 +375,115 @@ static inline int sepFilterCol(int start, int end, sepFilter2D* data, const floa
}
return pi;
};
const float* ky = reinterpret_cast<const float*>(data->kernely_data);
for (int i = start; i < end; i++)
{
const float* row0 = access(i ) == noval ? nullptr : src_data + access(i ) * width;
const float* row1 = access(i + 1) == noval ? nullptr : src_data + access(i + 1) * width;
const float* row2 = access(i + 2) == noval ? nullptr : src_data + access(i + 2) * width;
const float* row3, *row4;
if (ksize == 5)
auto accessY = [&](int y) {
int pj;
if (data->borderType & BORDER_ISOLATED)
{
row3 = access(i + 3) == noval ? nullptr : src_data + access(i + 3) * width;
row4 = access(i + 4) == noval ? nullptr : src_data + access(i + 4) * width;
pj = filter::borderInterpolate(y - data->anchor_x, width, data->borderType & ~BORDER_ISOLATED);
pj = pj < 0 ? noval : pj;
}
int vl;
for (int j = 0; j < width; j += vl)
else
{
vl = __riscv_vsetvl_e32m4(width - j);
auto v0 = row0 ? __riscv_vle32_v_f32m4(row0 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
auto v1 = row1 ? __riscv_vle32_v_f32m4(row1 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
auto v2 = row2 ? __riscv_vle32_v_f32m4(row2 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
auto sum = __riscv_vfmacc(__riscv_vfmacc(__riscv_vfmacc(__riscv_vfmv_v_f_f32m4(data->delta, vl), ky[0], v0, vl), ky[1], v1, vl), ky[2], v2, vl);
pj = filter::borderInterpolate(offset_x + y - data->anchor_x, full_width, data->borderType);
pj = pj < 0 ? noval : pj - offset_x;
}
return pj;
};
auto p2idx = [&](int x, int y){ return (x + ksize) % ksize * width + y; };
if (ksize == 5)
const float* kx = reinterpret_cast<const float*>(data->kernelx_data);
const float* ky = reinterpret_cast<const float*>(data->kernely_data);
std::vector<float> res(width * ksize);
auto process = [&](int x, int y) {
float sum = 0;
for (int i = 0; i < ksize; i++)
{
int p = accessY(y + i);
if (p != noval)
{
auto v3 = row3 ? __riscv_vle32_v_f32m4(row3 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
auto v4 = row4 ? __riscv_vle32_v_f32m4(row4 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
sum = __riscv_vfmacc(__riscv_vfmacc(sum, ky[3], v3, vl), ky[4], v4, vl);
sum += kx[i] * src_data[x * src_step + p];
}
if (data->dst_type == CV_16SC1)
}
res[p2idx(x, y)] = sum;
};
const int left = data->anchor_x, right = width - (ksize - 1 - data->anchor_x);
for (int i = start - data->anchor_y; i < end + (ksize - 1 - data->anchor_y); i++)
{
if (i + offset_y >= 0 && i + offset_y < full_height)
{
if (left >= right)
{
__riscv_vse16(reinterpret_cast<short*>(dst_data + i * dst_step) + j, __riscv_vfncvt_x(sum, vl), vl);
for (int j = 0; j < width; j++)
process(i, j);
}
else
{
__riscv_vse32(reinterpret_cast<float*>(dst_data + i * dst_step) + j, sum, vl);
for (int j = 0; j < left; j++)
process(i, j);
for (int j = right; j < width; j++)
process(i, j);
int vl;
for (int j = left; j < right; j += vl)
{
vl = __riscv_vsetvl_e8m2(right - j);
const uchar* extra = src_data + i * src_step + j - data->anchor_x;
auto sum = __riscv_vfmv_v_f_f32m8(0, vl);
auto src = __riscv_vfwcvt_f(__riscv_vwcvtu_x(__riscv_vle8_v_u8m2(extra, vl), vl), vl);
sum = __riscv_vfmacc(sum, kx[0], src, vl);
src = __riscv_vfslide1down(src, extra[vl], vl);
sum = __riscv_vfmacc(sum, kx[1], src, vl);
src = __riscv_vfslide1down(src, extra[vl + 1], vl);
sum = __riscv_vfmacc(sum, kx[2], src, vl);
if (ksize == 5)
{
src = __riscv_vfslide1down(src, extra[vl + 2], vl);
sum = __riscv_vfmacc(sum, kx[3], src, vl);
src = __riscv_vfslide1down(src, extra[vl + 3], vl);
sum = __riscv_vfmacc(sum, kx[4], src, vl);
}
__riscv_vse32(res.data() + p2idx(i, j), sum, vl);
}
}
}
int cur = i - (ksize - 1 - data->anchor_y);
if (cur >= start)
{
const float* row0 = accessX(cur ) == noval ? nullptr : res.data() + p2idx(accessX(cur ), 0);
const float* row1 = accessX(cur + 1) == noval ? nullptr : res.data() + p2idx(accessX(cur + 1), 0);
const float* row2 = accessX(cur + 2) == noval ? nullptr : res.data() + p2idx(accessX(cur + 2), 0);
const float* row3, *row4;
if (ksize == 5)
{
row3 = accessX(cur + 3) == noval ? nullptr : res.data() + p2idx(accessX(cur + 3), 0);
row4 = accessX(cur + 4) == noval ? nullptr : res.data() + p2idx(accessX(cur + 4), 0);
}
int vl;
for (int j = 0; j < width; j += vl)
{
vl = __riscv_vsetvl_e32m4(width - j);
auto v0 = row0 ? __riscv_vle32_v_f32m4(row0 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
auto v1 = row1 ? __riscv_vle32_v_f32m4(row1 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
auto v2 = row2 ? __riscv_vle32_v_f32m4(row2 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
auto sum = __riscv_vfmacc(__riscv_vfmacc(__riscv_vfmacc(__riscv_vfmv_v_f_f32m4(data->delta, vl), ky[0], v0, vl), ky[1], v1, vl), ky[2], v2, vl);
if (ksize == 5)
{
auto v3 = row3 ? __riscv_vle32_v_f32m4(row3 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
auto v4 = row4 ? __riscv_vle32_v_f32m4(row4 + j, vl) : __riscv_vfmv_v_f_f32m4(0, vl);
sum = __riscv_vfmacc(__riscv_vfmacc(sum, ky[3], v3, vl), ky[4], v4, vl);
}
if (data->dst_type == CV_16SC1)
{
__riscv_vse16(reinterpret_cast<short*>(dst_data + cur * dst_step) + j, __riscv_vfncvt_x(sum, vl), vl);
}
else
{
__riscv_vse32(reinterpret_cast<float*>(dst_data + cur * dst_step) + j, sum, vl);
}
}
}
}
@ -496,29 +494,13 @@ static inline int sepFilterCol(int start, int end, sepFilter2D* data, const floa
inline int sepFilter(cvhalFilter2D *context, uchar *src_data, size_t src_step, uchar* dst_data, size_t dst_step, int width, int height, int full_width, int full_height, int offset_x, int offset_y)
{
sepFilter2D* data = reinterpret_cast<sepFilter2D*>(context);
const int padding = data->kernelx_length - 1;
std::vector<float> _result(width * (height + 2 * padding));
float* result = _result.data() + width * padding;
int res = CV_HAL_ERROR_NOT_IMPLEMENTED;
switch (data->kernelx_length)
{
case 3:
res = filter::invoke(-std::min(offset_y, padding), height + std::min(full_height - height - offset_y, padding), {sepFilterRow<3>}, data, src_data, src_step, result, width, full_width, offset_x);
break;
case 5:
res = filter::invoke(-std::min(offset_y, padding), height + std::min(full_height - height - offset_y, padding), {sepFilterRow<5>}, data, src_data, src_step, result, width, full_width, offset_x);
break;
}
if (res == CV_HAL_ERROR_NOT_IMPLEMENTED)
return CV_HAL_ERROR_NOT_IMPLEMENTED;
switch (data->kernelx_length)
{
case 3:
return filter::invoke(0, height, {sepFilterCol<3>}, data, result, dst_data, dst_step, width, height, full_height, offset_y);
return filter::invoke(height, {sepFilter<3>}, data, src_data, src_step, dst_data, dst_step, width, height, full_width, full_height, offset_x, offset_y);
case 5:
return filter::invoke(0, height, {sepFilterCol<5>}, data, result, dst_data, dst_step, width, height, full_height, offset_y);
return filter::invoke(height, {sepFilter<5>}, data, src_data, src_step, dst_data, dst_step, width, height, full_width, full_height, offset_x, offset_y);
}
return CV_HAL_ERROR_NOT_IMPLEMENTED;
@ -699,9 +681,9 @@ static inline int morph(int start, int end, Morph2D* data, const uchar* src_data
}
};
const int left = data->anchor_x, right = width - (2 - data->anchor_x);
for (int i = start; i < end; i++)
{
const int left = 2, right = width - 2;
if (left >= right)
{
for (int j = 0; j < width; j++)
@ -830,10 +812,10 @@ inline int morph(cvhalFilter2D* context, uchar *src_data, size_t src_step, uchar
switch (data->operation)
{
case CV_HAL_MORPH_ERODE:
res = filter::invoke(0, height, {morph<CV_HAL_MORPH_ERODE>}, data, src_data, src_step, dst.data(), width, height, src_full_width, src_full_height, src_roi_x, src_roi_y);
res = filter::invoke(height, {morph<CV_HAL_MORPH_ERODE>}, data, src_data, src_step, dst.data(), width, height, src_full_width, src_full_height, src_roi_x, src_roi_y);
break;
case CV_HAL_MORPH_DILATE:
res = filter::invoke(0, height, {morph<CV_HAL_MORPH_DILATE>}, data, src_data, src_step, dst.data(), width, height, src_full_width, src_full_height, src_roi_x, src_roi_y);
res = filter::invoke(height, {morph<CV_HAL_MORPH_DILATE>}, data, src_data, src_step, dst.data(), width, height, src_full_width, src_full_height, src_roi_x, src_roi_y);
break;
}