potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-21 19:33:03 +04:00
Code Packages Releases Wiki Activity

Improved Bilateral Filter level2

Browse Source 
AVX512_SKX and AVX512_ICL dispatch added.
BilateralFilter 32f has been improved for AVX512.
This commit is contained in:
Madan mohan Manokar
2026-01-14 00:58:16 +00:00
parent 105a774720
commit f994a43df2
2 changed files with 57 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/imgproc/CMakeLists.txt
+1 -1
View File
@@ -1,6 +1,6 @@
set(the_description "Image Processing")
ocv_add_dispatched_file(accum SSE4_1 AVX AVX2)
ocv_add_dispatched_file(bilateral_filter SSE2 AVX2)
ocv_add_dispatched_file(bilateral_filter SSE2 AVX2 AVX512_SKX AVX512_ICL)
ocv_add_dispatched_file(box_filter SSE2 SSE4_1 AVX2 AVX512_SKX)
ocv_add_dispatched_file(filter SSE2 SSE4_1 AVX2)
ocv_add_dispatched_file(color_hsv SSE2 SSE4_1 AVX2)
modules/imgproc/src/bilateral_filter.simd.hpp
+56 -48
View File
@@ -512,6 +512,9 @@ public:
int nlanes_2 = 2 * nlanes;
int nlanes_3 = 3 * nlanes;
int nlanes_4 = 4 * nlanes;
v_float32 v_one = vx_setall_f32(1.f);
v_float32 sindex = vx_setall_f32(scale_index);
#endif
for( i = range.start; i < range.end; i++ )
{
@@ -523,9 +526,6 @@ public:
j = 0;
const float* sptr_j = sptr;
#if (CV_SIMD || CV_SIMD_SCALABLE)
v_float32 v_one = vx_setall_f32(1.f);
v_float32 sindex = vx_setall_f32(scale_index);
for(; j <= size.width - nlanes_4; j += nlanes_4, sptr_j += nlanes_4, dptr += nlanes_4)
{
v_float32 v_wsum0 = vx_setzero_f32();
@@ -545,45 +545,37 @@ public:
{
const float* ksptr = sptr_j + space_ofs[k];
v_float32 kweight = vx_setall_f32(space_weight[k]);
//0th
v_float32 val0 = vx_load(ksptr);
v_float32 knan0 = v_not_nan(val0);
v_float32 alpha0 = v_and(v_and(v_mul(v_absdiff(val0, rval0), sindex), v_not_nan(rval0)), knan0);
v_int32 idx0 = v_min(v_trunc(alpha0), vx_setall_s32(kExpNumBins));
alpha0 = v_sub(alpha0, v_cvt_f32(idx0));
v_float32 w0 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx0), alpha0, v_mul(v_lut(this->expLUT, idx0), v_sub(v_one, alpha0)))), knan0);
v_wsum0 = v_add(v_wsum0, w0);
v_sum0 = v_muladd(v_and(val0, knan0), w0, v_sum0);
//1st
v_float32 val1 = vx_load(ksptr + nlanes);
v_float32 knan1 = v_not_nan(val1);
v_float32 alpha1 = v_and(v_and(v_mul(v_absdiff(val1, rval1), sindex), v_not_nan(rval1)), knan1);
v_int32 idx1 = v_min(v_trunc(alpha1), vx_setall_s32(kExpNumBins));
alpha1 = v_sub(alpha1, v_cvt_f32(idx1));
v_float32 w1 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx1), alpha1, v_mul(v_lut(this->expLUT, idx1), v_sub(v_one, alpha1)))), knan1);
v_wsum1 = v_add(v_wsum1, w1);
v_sum1 = v_muladd(v_and(val1, knan1), w1, v_sum1);
//2nd
v_float32 val2 = vx_load(ksptr + nlanes_2);
v_float32 knan2 = v_not_nan(val2);
v_float32 alpha2 = v_and(v_and(v_mul(v_absdiff(val2, rval2), sindex), v_not_nan(rval2)), knan2);
v_int32 idx2 = v_min(v_trunc(alpha2), vx_setall_s32(kExpNumBins));
alpha2 = v_sub(alpha2, v_cvt_f32(idx2));
v_float32 w2 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx2), alpha2, v_mul(v_lut(this->expLUT, idx2), v_sub(v_one, alpha2)))), knan2);
v_wsum2 = v_add(v_wsum2, w2);
v_sum2 = v_muladd(v_and(val2, knan2), w2, v_sum2);
//3rd
v_float32 val3 = vx_load(ksptr + nlanes_3);
v_float32 knan0 = v_not_nan(val0);
v_float32 knan1 = v_not_nan(val1);
v_float32 knan2 = v_not_nan(val2);
v_float32 knan3 = v_not_nan(val3);
v_float32 alpha0 = v_and(v_and(v_mul(v_absdiff(val0, rval0), sindex), v_not_nan(rval0)), knan0);
v_float32 alpha1 = v_and(v_and(v_mul(v_absdiff(val1, rval1), sindex), v_not_nan(rval1)), knan1);
v_float32 alpha2 = v_and(v_and(v_mul(v_absdiff(val2, rval2), sindex), v_not_nan(rval2)), knan2);
v_float32 alpha3 = v_and(v_and(v_mul(v_absdiff(val3, rval3), sindex), v_not_nan(rval3)), knan3);
v_int32 idx0 = v_min(v_trunc(alpha0), vx_setall_s32(kExpNumBins));
v_int32 idx1 = v_min(v_trunc(alpha1), vx_setall_s32(kExpNumBins));
v_int32 idx2 = v_min(v_trunc(alpha2), vx_setall_s32(kExpNumBins));
v_int32 idx3 = v_min(v_trunc(alpha3), vx_setall_s32(kExpNumBins));
alpha0 = v_sub(alpha0, v_cvt_f32(idx0));
alpha1 = v_sub(alpha1, v_cvt_f32(idx1));
alpha2 = v_sub(alpha2, v_cvt_f32(idx2));
alpha3 = v_sub(alpha3, v_cvt_f32(idx3));
v_float32 w0 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx0), alpha0, v_mul(v_lut(this->expLUT, idx0), v_sub(v_one, alpha0)))), knan0);
v_float32 w1 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx1), alpha1, v_mul(v_lut(this->expLUT, idx1), v_sub(v_one, alpha1)))), knan1);
v_float32 w2 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx2), alpha2, v_mul(v_lut(this->expLUT, idx2), v_sub(v_one, alpha2)))), knan2);
v_float32 w3 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx3), alpha3, v_mul(v_lut(this->expLUT, idx3), v_sub(v_one, alpha3)))), knan3);
v_wsum0 = v_add(v_wsum0, w0);
v_wsum1 = v_add(v_wsum1, w1);
v_wsum2 = v_add(v_wsum2, w2);
v_wsum3 = v_add(v_wsum3, w3);
v_sum0 = v_muladd(v_and(val0, knan0), w0, v_sum0);
v_sum1 = v_muladd(v_and(val1, knan1), w1, v_sum1);
v_sum2 = v_muladd(v_and(val2, knan2), w2, v_sum2);
v_sum3 = v_muladd(v_and(val3, knan3), w3, v_sum3);
}
v_store(dptr , v_div(v_add(v_sum0, v_and(rval0, v_not_nan(rval0))), v_add(v_wsum0, v_and(v_one, v_not_nan(rval0)))));
@@ -607,24 +599,21 @@ public:
v_float32 kweight = vx_setall_f32(space_weight[k]);
const float* ksptr = sptr_j + space_ofs[k];
//0th
v_float32 val0 = vx_load(ksptr);
v_float32 knan0 = v_not_nan(val0);
v_float32 alpha0 = v_and(v_and(v_mul(v_absdiff(val0, rval0), sindex), rval0_not_nan), knan0);
v_int32 idx0 = v_min(v_trunc(alpha0), vx_setall_s32(kExpNumBins));
alpha0 = v_sub(alpha0, v_cvt_f32(idx0));
v_float32 w0 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx0), alpha0, v_mul(v_lut(this->expLUT, idx0), v_sub(v_one, alpha0)))), knan0);
v_wsum0 = v_add(v_wsum0, w0);
v_sum0 = v_muladd(v_and(val0, knan0), w0, v_sum0);
//1st
v_float32 val1 = vx_load(ksptr + nlanes);
v_float32 knan0 = v_not_nan(val0);
v_float32 knan1 = v_not_nan(val1);
v_float32 alpha0 = v_and(v_and(v_mul(v_absdiff(val0, rval0), sindex), rval0_not_nan), knan0);
v_float32 alpha1 = v_and(v_and(v_mul(v_absdiff(val1, rval1), sindex), rval1_not_nan), knan1);
v_int32 idx0 = v_min(v_trunc(alpha0), vx_setall_s32(kExpNumBins));
v_int32 idx1 = v_min(v_trunc(alpha1), vx_setall_s32(kExpNumBins));
alpha0 = v_sub(alpha0, v_cvt_f32(idx0));
alpha1 = v_sub(alpha1, v_cvt_f32(idx1));
v_float32 w0 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx0), alpha0, v_mul(v_lut(this->expLUT, idx0), v_sub(v_one, alpha0)))), knan0);
v_float32 w1 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx1), alpha1, v_mul(v_lut(this->expLUT, idx1), v_sub(v_one, alpha1)))), knan1);
v_wsum0 = v_add(v_wsum0, w0);
v_wsum1 = v_add(v_wsum1, w1);
v_sum0 = v_muladd(v_and(val0, knan0), w0, v_sum0);
v_sum1 = v_muladd(v_and(val1, knan1), w1, v_sum1);
}
v_store(dptr, v_div(v_add(v_sum0, v_and(rval0, rval0_not_nan)), v_add(v_wsum0, v_and(v_one, rval0_not_nan))));
@@ -660,42 +649,61 @@ public:
j = 0;
const float* sptr_j = sptr;
#if (CV_SIMD || CV_SIMD_SCALABLE)
v_float32 v_one = vx_setall_f32(1.f);
v_float32 sindex = vx_setall_f32(scale_index);
for (; j <= size.width - nlanes; j += nlanes, sptr_j += nlanes_3, dptr += nlanes_3)
for (; j <= size.width - nlanes_2; j += nlanes_2, sptr_j += (nlanes_3*2), dptr += (nlanes_3*2))
{
v_float32 kb, kg, kr, kb1, kg1, kr1;
v_float32 rb, rg, rr, rb1, rg1, rr1;
const float* rsptr = sptr_j;
v_float32 v_wsum = vx_setzero_f32();
v_float32 v_sum_b = vx_setzero_f32();
v_float32 v_sum_g = vx_setzero_f32();
v_float32 v_sum_r = vx_setzero_f32();
v_float32 v_wsum1 = vx_setzero_f32();
v_float32 v_sum_b1 = vx_setzero_f32();
v_float32 v_sum_g1 = vx_setzero_f32();
v_float32 v_sum_r1 = vx_setzero_f32();
v_load_deinterleave(rsptr, rb, rg, rr);
v_load_deinterleave(rsptr + nlanes_3, rb1, rg1, rr1);
for (k = 0; k < maxk; k++)
{
const float* ksptr = sptr_j + space_ofs[k];
v_float32 kweight = vx_setall_f32(space_weight[k]);
v_float32 kb, kg, kr, rb, rg, rr;
v_load_deinterleave(ksptr, kb, kg, kr);
v_load_deinterleave(rsptr, rb, rg, rr);
v_load_deinterleave(ksptr + nlanes_3, kb1, kg1, kr1);
v_float32 knan = v_and(v_and(v_not_nan(kb), v_not_nan(kg)), v_not_nan(kr));
v_float32 knan1 = v_and(v_and(v_not_nan(kb1), v_not_nan(kg1)), v_not_nan(kr1));
v_float32 alpha = v_and(v_and(v_and(v_and(v_mul(v_add(v_add(v_absdiff(kb, rb), v_absdiff(kg, rg)), v_absdiff(kr, rr)), sindex), v_not_nan(rb)), v_not_nan(rg)), v_not_nan(rr)), knan);
v_float32 alpha1 = v_and(v_and(v_and(v_and(v_mul(v_add(v_add(v_absdiff(kb1, rb1), v_absdiff(kg1, rg1)), v_absdiff(kr1, rr1)), sindex), v_not_nan(rb1)), v_not_nan(rg1)), v_not_nan(rr1)), knan1);
v_int32 idx = v_min(v_trunc(alpha), vx_setall_s32(kExpNumBins));
v_int32 idx1 = v_min(v_trunc(alpha1), vx_setall_s32(kExpNumBins));
alpha = v_sub(alpha, v_cvt_f32(idx));
alpha1 = v_sub(alpha1, v_cvt_f32(idx1));
v_float32 w = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx), alpha, v_mul(v_lut(this->expLUT, idx), v_sub(v_one, alpha)))), knan);
v_float32 w1 = v_and(v_mul(kweight, v_muladd(v_lut(this->expLUT + 1, idx1), alpha1, v_mul(v_lut(this->expLUT, idx1), v_sub(v_one, alpha1)))), knan1);
v_wsum = v_add(v_wsum, w);
v_wsum1 = v_add(v_wsum1, w1);
v_sum_b = v_muladd(v_and(kb, knan), w, v_sum_b);
v_sum_g = v_muladd(v_and(kg, knan), w, v_sum_g);
v_sum_r = v_muladd(v_and(kr, knan), w, v_sum_r);
v_sum_b1 = v_muladd(v_and(kb1, knan1), w1, v_sum_b1);
v_sum_g1 = v_muladd(v_and(kg1, knan1), w1, v_sum_g1);
v_sum_r1 = v_muladd(v_and(kr1, knan1), w1, v_sum_r1);
}
v_float32 b, g, r;
v_float32 b1, g1, r1;
v_load_deinterleave(sptr_j, b, g, r);
v_load_deinterleave(sptr_j + nlanes_3, b1, g1, r1);
v_float32 mask = v_and(v_and(v_not_nan(b), v_not_nan(g)), v_not_nan(r));
v_float32 mask1 = v_and(v_and(v_not_nan(b1), v_not_nan(g1)), v_not_nan(r1));
v_float32 w = v_div(v_one, v_add(v_wsum, v_and(v_one, mask)));
v_float32 w1 = v_div(v_one, v_add(v_wsum1, v_and(v_one, mask1)));
v_store_interleave(dptr, v_mul(v_add(v_sum_b, v_and(b, mask)), w), v_mul(v_add(v_sum_g, v_and(g, mask)), w), v_mul(v_add(v_sum_r, v_and(r, mask)), w));
v_store_interleave(dptr + nlanes_3, v_mul(v_add(v_sum_b1, v_and(b1, mask1)), w1), v_mul(v_add(v_sum_g1, v_and(g1, mask1)), w1), v_mul(v_add(v_sum_r1, v_and(r1, mask1)), w1));
}
#endif
for (; j < size.width; j++, sptr_j += 3)