Administrator/opencv-MIRROR
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-11-09 18:04:29 -06:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #30188 from pratham-mcw:imgproc-pyrup-vec

Browse Source 
imgproc: vectorize pyrUp horizontal pass via PyrUpVecH (cn 1/4)
This commit is contained in:
Alexander Smorkalov
2026-10-12 12:05:14 +03:00
committed by GitHub
parent 3388290736 1f12c99c48
commit 1aab5099d2
1 changed files with 149 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

150
modules/imgproc/src/pyramids.cpp
View File

@@ -859,6 +859,142 @@ template <> int PyrUpVecVOneRow<float, float>(float** src, float* dst, int width
#endif
#if CV_SIMD128
static CV_ALWAYS_INLINE void pyrUpH_u8(const uchar* src, int d, v_uint16x8& t0_lo, v_uint16x8& t0_hi,
v_uint16x8& t1_lo, v_uint16x8& t1_hi)
{
v_uint16x8 a_lo, a_hi, b_lo, b_hi, c_lo, c_hi;
v_expand(v_load(src - d), a_lo, a_hi);
v_expand(v_load(src), b_lo, b_hi);
v_expand(v_load(src + d), c_lo, c_hi);
v_uint16x8 v_6 = v_setall_u16(6);
t0_lo = v_add(v_add(a_lo, v_mul_wrap(b_lo, v_6)), c_lo);
t0_hi = v_add(v_add(a_hi, v_mul_wrap(b_hi, v_6)), c_hi);
t1_lo = v_shl<2>(v_add(b_lo, c_lo));
t1_hi = v_shl<2>(v_add(b_hi, c_hi));
}
static CV_ALWAYS_INLINE v_int32x4 pyrUpH_lo(const v_uint16x8& v) {
return v_reinterpret_as_s32(v_expand_low(v));
}
static CV_ALWAYS_INLINE v_int32x4 pyrUpH_hi(const v_uint16x8& v) {
return v_reinterpret_as_s32(v_expand_high(v));
}
static CV_ALWAYS_INLINE void pyrUpH_s16(const short* src, int d, v_int32x4& t0_lo, v_int32x4& t0_hi,
v_int32x4& t1_lo, v_int32x4& t1_hi)
{
v_int32x4 a_lo, a_hi, b_lo, b_hi, c_lo, c_hi;
v_expand(v_load(src - d), a_lo, a_hi);
v_expand(v_load(src), b_lo, b_hi);
v_expand(v_load(src + d), c_lo, c_hi);
v_int32x4 v_6 = v_setall_s32(6);
t0_lo = v_add(v_add(a_lo, v_mul(b_lo, v_6)), c_lo);
t0_hi = v_add(v_add(a_hi, v_mul(b_hi, v_6)), c_hi);
t1_lo = v_shl<2>(v_add(b_lo, c_lo));
t1_hi = v_shl<2>(v_add(b_hi, c_hi));
}
static CV_ALWAYS_INLINE void pyrUpH_f32(const float* src, int d, v_float32x4& t0, v_float32x4& t1)
{
v_float32x4 a = v_load(src - d), b = v_load(src), c = v_load(src + d);
t0 = v_add(v_add(a, v_mul(b, v_setall_f32(6.f))), c);
t1 = v_mul(v_add(b, c), v_setall_f32(4.f));
}
template<> int PyrUpVecH<uchar, int, 1>(const uchar* src, int* row, int width)
{
int x = 0;
for( ; x <= width - 16; x += 16 )
{
v_uint16x8 t0_lo, t0_hi, t1_lo, t1_hi;
pyrUpH_u8(src + x, 1, t0_lo, t0_hi, t1_lo, t1_hi);
int* r = row + 2*x;
v_store_interleave(r, pyrUpH_lo(t0_lo), pyrUpH_lo(t1_lo));
v_store_interleave(r + 8, pyrUpH_hi(t0_lo), pyrUpH_hi(t1_lo));
v_store_interleave(r + 16, pyrUpH_lo(t0_hi), pyrUpH_lo(t1_hi));
v_store_interleave(r + 24, pyrUpH_hi(t0_hi), pyrUpH_hi(t1_hi));
}
return x;
}
template<> int PyrUpVecH<uchar, int, 4>(const uchar* src, int* row, int width)
{
int x = 0;
for( ; x <= width - 16; x += 16 )
{
v_uint16x8 t0_lo, t0_hi, t1_lo, t1_hi;
pyrUpH_u8(src + x, 4, t0_lo, t0_hi, t1_lo, t1_hi);
int* r = row + 2*x;
v_store(r, pyrUpH_lo(t0_lo));
v_store(r + 4, pyrUpH_lo(t1_lo));
v_store(r + 8, pyrUpH_hi(t0_lo));
v_store(r + 12, pyrUpH_hi(t1_lo));
v_store(r + 16, pyrUpH_lo(t0_hi));
v_store(r + 20, pyrUpH_lo(t1_hi));
v_store(r + 24, pyrUpH_hi(t0_hi));
v_store(r + 28, pyrUpH_hi(t1_hi));
}
return x;
}
template<> int PyrUpVecH<short, int, 1>(const short* src, int* row, int width)
{
int x = 0;
for( ; x <= width - 8; x += 8 )
{
v_int32x4 t0_lo, t0_hi, t1_lo, t1_hi;
pyrUpH_s16(src + x, 1, t0_lo, t0_hi, t1_lo, t1_hi);
v_store_interleave(row + 2*x, t0_lo, t1_lo);
v_store_interleave(row + 2*x + 8, t0_hi, t1_hi);
}
return x;
}
template<> int PyrUpVecH<short, int, 4>(const short* src, int* row, int width)
{
int x = 0;
for( ; x <= width - 8; x += 8 )
{
v_int32x4 t0_lo, t0_hi, t1_lo, t1_hi;
pyrUpH_s16(src + x, 4, t0_lo, t0_hi, t1_lo, t1_hi);
int* r = row + 2*x;
v_store(r, t0_lo);
v_store(r + 4, t1_lo);
v_store(r + 8, t0_hi);
v_store(r + 12, t1_hi);
}
return x;
}
template<> int PyrUpVecH<float, float, 1>(const float* src, float* row, int width)
{
int x = 0;
for( ; x <= width - 4; x += 4 )
{
v_float32x4 t0, t1;
pyrUpH_f32(src + x, 1, t0, t1);
v_store_interleave(row + 2*x, t0, t1);
}
return x;
}
template<> int PyrUpVecH<float, float, 4>(const float* src, float* row, int width)
{
int x = 0;
for( ; x <= width - 4; x += 4 )
{
v_float32x4 t0, t1;
pyrUpH_f32(src + x, 4, t0, t1);
v_store(row + 2*x, t0);
v_store(row + 2*x + 4, t1);
}
return x;
}
#endif // CV_SIMD128
template<class CastOp>
struct PyrDownInvoker : ParallelLoopBody
{
@@ -1107,7 +1243,19 @@ pyrUp_( const Mat& _src, Mat& _dst, int)
}
}
for( x = cn; x < ssize.width - cn; x++ )
x = cn;
switch( cn )
{
case 1:
x += PyrUpVecH<T, WT, 1>(src + cn, row + 2*cn, ssize.width - 2*cn);
break;
case 4:
x += PyrUpVecH<T, WT, 4>(src + cn, row + 2*cn, ssize.width - 2*cn);
break;
default:
break;
}
for( ; x < ssize.width - cn; x++ )
{
int dx = dtab[x];
WT t0 = src[x-cn] + src[x]*6 + src[x+cn];