Skip to content

O
opencv
Commit b3755ae6 authored by k-shinotsuka's avatar k-shinotsuka
Browse files
Options

add universal intrinsics for RGB2HLS_f

parent 225c75e4
Hide whitespace changes
Inline Side-by-side
Showing with 77 additions and 94 deletions
modules/imgproc/src/color_hsv.cpp
View file @ b3755ae6
...@@ -523,62 +523,38 @@ struct RGB2HLS_f ...@@ -523,62 +523,38 @@ struct RGB2HLS_f
RGB2HLS_f(int _srccn, int _blueIdx, float _hrange) RGB2HLS_f(int _srccn, int _blueIdx, float _hrange)
: srccn(_srccn), blueIdx(_blueIdx), hscale(_hrange/360.f) { : srccn(_srccn), blueIdx(_blueIdx), hscale(_hrange/360.f) {
#if CV_SSE2 #if CV_SIMD128
haveSIMD = checkHardwareSupport(CV_CPU_SSE2); hasSIMD = hasSIMD128();
#endif #endif
} }
#if CV_SSE2 #if CV_SIMD128
void process(__m128& v_b0, __m128& v_b1, __m128& v_g0, inline void process(v_float32x4& v_r, v_float32x4& v_g,
__m128& v_g1, __m128& v_r0, __m128& v_r1) const v_float32x4& v_b, v_float32x4& v_hscale) const
{ {
__m128 v_max0 = _mm_max_ps(_mm_max_ps(v_b0, v_g0), v_r0); v_float32x4 v_max_rgb = v_max(v_max(v_r, v_g), v_b);
__m128 v_max1 = _mm_max_ps(_mm_max_ps(v_b1, v_g1), v_r1); v_float32x4 v_min_rgb = v_min(v_min(v_r, v_g), v_b);
__m128 v_min0 = _mm_min_ps(_mm_min_ps(v_b0, v_g0), v_r0);
__m128 v_min1 = _mm_min_ps(_mm_min_ps(v_b1, v_g1), v_r1); v_float32x4 v_diff = v_max_rgb - v_min_rgb;
__m128 v_diff0 = _mm_sub_ps(v_max0, v_min0); v_float32x4 v_sum = v_max_rgb + v_min_rgb;
__m128 v_diff1 = _mm_sub_ps(v_max1, v_min1); v_float32x4 v_half = v_setall_f32(0.5f);
__m128 v_sum0 = _mm_add_ps(v_max0, v_min0); v_float32x4 v_l = v_sum * v_half;
__m128 v_sum1 = _mm_add_ps(v_max1, v_min1);
__m128 v_l0 = _mm_mul_ps(v_sum0, _mm_set1_ps(0.5f)); v_float32x4 v_s = v_diff / v_select(v_l < v_half, v_sum, v_setall_f32(2.0f) - v_sum);
__m128 v_l1 = _mm_mul_ps(v_sum1, _mm_set1_ps(0.5f));
v_float32x4 v_r_eq_max = v_max_rgb == v_r;
__m128 v_gel0 = _mm_cmpge_ps(v_l0, _mm_set1_ps(0.5f)); v_float32x4 v_g_eq_max = v_max_rgb == v_g;
__m128 v_gel1 = _mm_cmpge_ps(v_l1, _mm_set1_ps(0.5f)); v_float32x4 v_h = v_select(v_r_eq_max, v_g - v_b,
__m128 v_s0 = _mm_and_ps(v_gel0, _mm_sub_ps(_mm_set1_ps(2.0f), v_sum0)); v_select(v_g_eq_max, v_b - v_r, v_r - v_g));
__m128 v_s1 = _mm_and_ps(v_gel1, _mm_sub_ps(_mm_set1_ps(2.0f), v_sum1)); v_float32x4 v_res = v_select(v_r_eq_max, (v_g < v_b) & v_setall_f32(360.0f),
v_s0 = _mm_or_ps(v_s0, _mm_andnot_ps(v_gel0, v_sum0)); v_select(v_g_eq_max, v_setall_f32(120.0f), v_setall_f32(240.0f)));
v_s1 = _mm_or_ps(v_s1, _mm_andnot_ps(v_gel1, v_sum1)); v_float32x4 v_rev_diff = v_setall_f32(60.0f) / v_diff;
v_s0 = _mm_div_ps(v_diff0, v_s0); v_h = v_muladd(v_h, v_rev_diff, v_res) * v_hscale;
v_s1 = _mm_div_ps(v_diff1, v_s1);
v_float32x4 v_diff_gt_eps = v_diff > v_setall_f32(FLT_EPSILON);
__m128 v_gteps0 = _mm_cmpgt_ps(v_diff0, _mm_set1_ps(FLT_EPSILON)); v_r = v_diff_gt_eps & v_h;
__m128 v_gteps1 = _mm_cmpgt_ps(v_diff1, _mm_set1_ps(FLT_EPSILON)); v_g = v_l;
v_b = v_diff_gt_eps & v_s;
v_diff0 = _mm_div_ps(_mm_set1_ps(60.f), v_diff0);
v_diff1 = _mm_div_ps(_mm_set1_ps(60.f), v_diff1);
__m128 v_eqr0 = _mm_cmpeq_ps(v_max0, v_r0);
__m128 v_eqr1 = _mm_cmpeq_ps(v_max1, v_r1);
__m128 v_h0 = _mm_and_ps(v_eqr0, _mm_mul_ps(_mm_sub_ps(v_g0, v_b0), v_diff0));
__m128 v_h1 = _mm_and_ps(v_eqr1, _mm_mul_ps(_mm_sub_ps(v_g1, v_b1), v_diff1));
__m128 v_eqg0 = _mm_cmpeq_ps(v_max0, v_g0);
__m128 v_eqg1 = _mm_cmpeq_ps(v_max1, v_g1);
v_h0 = _mm_or_ps(v_h0, _mm_and_ps(_mm_andnot_ps(v_eqr0, v_eqg0), _mm_add_ps(_mm_mul_ps(_mm_sub_ps(v_b0, v_r0), v_diff0), _mm_set1_ps(120.f))));
v_h1 = _mm_or_ps(v_h1, _mm_and_ps(_mm_andnot_ps(v_eqr1, v_eqg1), _mm_add_ps(_mm_mul_ps(_mm_sub_ps(v_b1, v_r1), v_diff1), _mm_set1_ps(120.f))));
v_h0 = _mm_or_ps(v_h0, _mm_andnot_ps(_mm_or_ps(v_eqr0, v_eqg0), _mm_add_ps(_mm_mul_ps(_mm_sub_ps(v_r0, v_g0), v_diff0), _mm_set1_ps(240.f))));
v_h1 = _mm_or_ps(v_h1, _mm_andnot_ps(_mm_or_ps(v_eqr1, v_eqg1), _mm_add_ps(_mm_mul_ps(_mm_sub_ps(v_r1, v_g1), v_diff1), _mm_set1_ps(240.f))));
v_h0 = _mm_add_ps(v_h0, _mm_and_ps(_mm_cmplt_ps(v_h0, _mm_setzero_ps()), _mm_set1_ps(360.f)));
v_h1 = _mm_add_ps(v_h1, _mm_and_ps(_mm_cmplt_ps(v_h1, _mm_setzero_ps()), _mm_set1_ps(360.f)));
v_h0 = _mm_mul_ps(v_h0, _mm_set1_ps(hscale));
v_h1 = _mm_mul_ps(v_h1, _mm_set1_ps(hscale));
v_b0 = _mm_and_ps(v_gteps0, v_h0);
v_b1 = _mm_and_ps(v_gteps1, v_h1);
v_g0 = v_l0;
v_g1 = v_l1;
v_r0 = _mm_and_ps(v_gteps0, v_s0);
v_r1 = _mm_and_ps(v_gteps1, v_s1);
} }
#endif #endif
...@@ -587,49 +563,56 @@ struct RGB2HLS_f ...@@ -587,49 +563,56 @@ struct RGB2HLS_f
int i = 0, bidx = blueIdx, scn = srccn; int i = 0, bidx = blueIdx, scn = srccn;
n *= 3; n *= 3;
#if CV_SSE2 #if CV_SIMD128
if (haveSIMD) if (hasSIMD)
{ {
for( ; i <= n - 24; i += 24, src += scn * 8 ) v_float32x4 v_hscale = v_setall_f32(hscale);
{ if (scn == 3) {
__m128 v_b0 = _mm_loadu_ps(src + 0); if (bidx) {
__m128 v_b1 = _mm_loadu_ps(src + 4); for ( ; i <= n - 12; i += 12, src += scn * 4)
__m128 v_g0 = _mm_loadu_ps(src + 8); {
__m128 v_g1 = _mm_loadu_ps(src + 12); v_float32x4 v_r;
__m128 v_r0 = _mm_loadu_ps(src + 16); v_float32x4 v_g;
__m128 v_r1 = _mm_loadu_ps(src + 20); v_float32x4 v_b;
v_load_deinterleave(src, v_r, v_g, v_b);
if (scn == 3) process(v_r, v_g, v_b, v_hscale);
{ v_store_interleave(dst + i, v_r, v_g, v_b);
_mm_deinterleave_ps(v_b0, v_b1, v_g0, v_g1, v_r0, v_r1); }
} } else {
else for ( ; i <= n - 12; i += 12, src += scn * 4)
{ {
__m128 v_a0 = _mm_loadu_ps(src + 24); v_float32x4 v_r;
__m128 v_a1 = _mm_loadu_ps(src + 28); v_float32x4 v_g;
_mm_deinterleave_ps(v_b0, v_b1, v_g0, v_g1, v_r0, v_r1, v_a0, v_a1); v_float32x4 v_b;
v_load_deinterleave(src, v_r, v_g, v_b);
process(v_b, v_g, v_r, v_hscale);
v_store_interleave(dst + i, v_b, v_g, v_r);
}
} }
} else { // scn == 4
if (bidx) if (bidx) {
{ for ( ; i <= n - 12; i += 12, src += scn * 4)
__m128 v_tmp0 = v_b0; {
__m128 v_tmp1 = v_b1; v_float32x4 v_r;
v_b0 = v_r0; v_float32x4 v_g;
v_b1 = v_r1; v_float32x4 v_b;
v_r0 = v_tmp0; v_float32x4 v_a;
v_r1 = v_tmp1; v_load_deinterleave(src, v_r, v_g, v_b, v_a);
process(v_r, v_g, v_b, v_hscale);
v_store_interleave(dst + i, v_r, v_g, v_b);
}
} else {
for ( ; i <= n - 12; i += 12, src += scn * 4)
{
v_float32x4 v_r;
v_float32x4 v_g;
v_float32x4 v_b;
v_float32x4 v_a;
v_load_deinterleave(src, v_r, v_g, v_b, v_a);
process(v_b, v_g, v_r, v_hscale);
v_store_interleave(dst + i, v_b, v_g, v_r);
}
} }
process(v_b0, v_b1, v_g0, v_g1, v_r0, v_r1);
_mm_interleave_ps(v_b0, v_b1, v_g0, v_g1, v_r0, v_r1);
_mm_storeu_ps(dst + i + 0, v_b0);
_mm_storeu_ps(dst + i + 4, v_b1);
_mm_storeu_ps(dst + i + 8, v_g0);
_mm_storeu_ps(dst + i + 12, v_g1);
_mm_storeu_ps(dst + i + 16, v_r0);
_mm_storeu_ps(dst + i + 20, v_r1);
} }
} }
#endif #endif
...@@ -672,8 +655,8 @@ struct RGB2HLS_f ...@@ -672,8 +655,8 @@ struct RGB2HLS_f
int srccn, blueIdx; int srccn, blueIdx;
float hscale; float hscale;
#if CV_SSE2 #if CV_SIMD128
bool haveSIMD; bool hasSIMD;
#endif #endif
}; };
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment