diff --git a/modules/core/include/opencv2/core/hal/intrin.hpp b/modules/core/include/opencv2/core/hal/intrin.hpp index 4f10584713fa0211ddadb904777e3d6f2071dffc..1971c91e1d305a7a803daa5a1d59b98ea1834418 100644 --- a/modules/core/include/opencv2/core/hal/intrin.hpp +++ b/modules/core/include/opencv2/core/hal/intrin.hpp @@ -785,7 +785,6 @@ namespace CV__SIMD_NAMESPACE { OPENCV_HAL_WRAP_BIN_OP_LOGIC(v_int32) OPENCV_HAL_WRAP_BIN_OP_LOGIC(v_int64) - #define OPENCV_HAL_WRAP_BIN_OP_MUL(_Tpvec) \ inline _Tpvec v_mul(const _Tpvec& a, const _Tpvec& b) \ { \ diff --git a/modules/core/include/opencv2/core/hal/intrin_rvv_scalable.hpp b/modules/core/include/opencv2/core/hal/intrin_rvv_scalable.hpp index 5073e85d8d745f9ac6ce2d2f541df8e8b9796028..c38f1f57dbad97b0b06d066971984c151a46e8d3 100644 --- a/modules/core/include/opencv2/core/hal/intrin_rvv_scalable.hpp +++ b/modules/core/include/opencv2/core/hal/intrin_rvv_scalable.hpp @@ -639,6 +639,35 @@ OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_int32, VTraits::vlanes()) OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_uint64, VTraits::vlanes()) OPENCV_HAL_IMPL_RVV_LOGIC_OP(v_int64, VTraits::vlanes()) +#define OPENCV_HAL_IMPL_RVV_FLT_BIT_OP(intrin) \ +inline v_float32 intrin (const v_float32& a, const v_float32& b) \ +{ \ + return vreinterpret_f32m1(intrin(vreinterpret_i32m1(a), vreinterpret_i32m1(b))); \ +} +OPENCV_HAL_IMPL_RVV_FLT_BIT_OP(v_and) +OPENCV_HAL_IMPL_RVV_FLT_BIT_OP(v_or) +OPENCV_HAL_IMPL_RVV_FLT_BIT_OP(v_xor) + +inline v_float32 v_not (const v_float32& a) \ +{ \ + return vreinterpret_f32m1(v_not(vreinterpret_i32m1(a))); \ +} + +#if CV_SIMD_SCALABLE_64F +#define OPENCV_HAL_IMPL_RVV_FLT64_BIT_OP(intrin) \ +inline v_float64 intrin (const v_float64& a, const v_float64& b) \ +{ \ + return vreinterpret_f64m1(intrin(vreinterpret_i64m1(a), vreinterpret_i64m1(b))); \ +} +OPENCV_HAL_IMPL_RVV_FLT64_BIT_OP(v_and) +OPENCV_HAL_IMPL_RVV_FLT64_BIT_OP(v_or) +OPENCV_HAL_IMPL_RVV_FLT64_BIT_OP(v_xor) + +inline v_float64 v_not (const v_float64& a) \ +{ \ + return vreinterpret_f64m1(v_not(vreinterpret_i64m1(a))); \ +} +#endif ////////////// Bitwise shifts ////////////// diff --git a/modules/imgproc/src/color_hsv.simd.hpp b/modules/imgproc/src/color_hsv.simd.hpp index 9837af4acb794047bdf80cb7d03b099f1a04d7cc..218103c4fe32565f9b9168141f2d17d557fbd66e 100644 --- a/modules/imgproc/src/color_hsv.simd.hpp +++ b/modules/imgproc/src/color_hsv.simd.hpp @@ -5,6 +5,19 @@ #include "precomp.hpp" #include "opencv2/core/hal/intrin.hpp" +#if CV_SIMD_SCALABLE +/* FIX IT: +// std::swap(a, b) is not available for RVV vector types, +// and CV_SWAP needs another "t" as input, +// For compatibility, we swap RVV vector manually by using this macro. + +// If others scalable types (e.g. type in ARM SVE) can use std::swap, +// then replace CV_SIMD_SCALABLE with CV_RVV. +// If std::swap is available for RVV vector types in future, remove this macro. +*/ +#define swap(a, b) {auto t = a; a = b; b = t;} +#endif + namespace cv { namespace hal { CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN @@ -85,8 +98,8 @@ struct RGB2HSV_b int i = 0; -#if CV_SIMD - const int vsize = v_uint8::nlanes; +#if CV_SIMD || CV_SIMD_SCALABLE + const int vsize = VTraits::vlanes(); for ( ; i <= n - vsize; i += vsize, src += scn*vsize, dst += 3*vsize) { @@ -110,108 +123,109 @@ struct RGB2HSV_b vmin = v_min(b, v_min(g, r)); v_uint8 diff, vr, vg; - diff = v - vmin; + diff = v_sub(v, vmin); v_uint8 v255 = vx_setall_u8(0xff), vz = vx_setzero_u8(); - vr = v_select(v == r, v255, vz); - vg = v_select(v == g, v255, vz); + vr = v_select(v_eq(v, r), v255, vz); + vg = v_select(v_eq(v, g), v255, vz); // sdiv = sdiv_table[v] - v_int32 sdiv[4]; - v_uint16 vd[2]; - v_expand(v, vd[0], vd[1]); - v_int32 vq[4]; - v_expand(v_reinterpret_as_s16(vd[0]), vq[0], vq[1]); - v_expand(v_reinterpret_as_s16(vd[1]), vq[2], vq[3]); + v_int32 sdiv0, sdiv1, sdiv2, sdiv3;; + v_uint16 vd0, vd1, vd2; + v_expand(v, vd0, vd1); + v_int32 vq0, vq1, vq2, vq3; + v_expand(v_reinterpret_as_s16(vd0), vq0, vq1); + v_expand(v_reinterpret_as_s16(vd1), vq2, vq3); { - int32_t CV_DECL_ALIGNED(CV_SIMD_WIDTH) storevq[vsize]; - for (int k = 0; k < 4; k++) - { - v_store_aligned(storevq + k*vsize/4, vq[k]); - } - - for(int k = 0; k < 4; k++) - { - sdiv[k] = vx_lut(sdiv_table, storevq + k*vsize/4); - } + int32_t CV_DECL_ALIGNED(CV_SIMD_WIDTH) storevq[VTraits::max_nlanes]; + v_store_aligned(storevq, vq0); + v_store_aligned(storevq + vsize/4, vq1); + v_store_aligned(storevq + 2*vsize/4, vq2); + v_store_aligned(storevq + 3*vsize/4, vq3); + + sdiv0 = vx_lut(sdiv_table, storevq); + sdiv1 = vx_lut(sdiv_table, storevq + vsize/4); + sdiv2 = vx_lut(sdiv_table, storevq + 2*vsize/4); + sdiv3 = vx_lut(sdiv_table, storevq + 3*vsize/4); } // hdiv = hdiv_table[diff] - v_int32 hdiv[4]; - v_uint16 diffd[2]; - v_expand(diff, diffd[0], diffd[1]); - v_int32 diffq[4]; - v_expand(v_reinterpret_as_s16(diffd[0]), diffq[0], diffq[1]); - v_expand(v_reinterpret_as_s16(diffd[1]), diffq[2], diffq[3]); + v_int32 hdiv0, hdiv1, hdiv2, hdiv3; + v_uint16 diffd0, diffd1, diffd2; + v_expand(diff, diffd0, diffd1); + v_int32 diffq0, diffq1, diffq2, diffq3; + v_expand(v_reinterpret_as_s16(diffd0), diffq0, diffq1); + v_expand(v_reinterpret_as_s16(diffd1), diffq2, diffq3); { - int32_t CV_DECL_ALIGNED(CV_SIMD_WIDTH) storediffq[vsize]; - for (int k = 0; k < 4; k++) - { - v_store_aligned(storediffq + k*vsize/4, diffq[k]); - } - - for (int k = 0; k < 4; k++) - { - hdiv[k] = vx_lut((int32_t*)hdiv_table, storediffq + k*vsize/4); - } + int32_t CV_DECL_ALIGNED(CV_SIMD_WIDTH) storediffq[VTraits::max_nlanes]; + v_store_aligned(storediffq, diffq0); + v_store_aligned(storediffq + vsize/4, diffq1); + v_store_aligned(storediffq + 2*vsize/4, diffq2); + v_store_aligned(storediffq + 3*vsize/4, diffq3); + hdiv0 = vx_lut((int32_t*)hdiv_table, storediffq + 0*vsize/4); + hdiv1 = vx_lut((int32_t*)hdiv_table, storediffq + 1*vsize/4); + hdiv2 = vx_lut((int32_t*)hdiv_table, storediffq + 2*vsize/4); + hdiv3 = vx_lut((int32_t*)hdiv_table, storediffq + 3*vsize/4); } // s = (diff * sdiv + (1 << (hsv_shift-1))) >> hsv_shift; - v_int32 sq[4]; + v_int32 sq0, sq1, sq2, sq3; v_int32 vdescale = vx_setall_s32(1 << (hsv_shift-1)); - for (int k = 0; k < 4; k++) - { - sq[k] = (diffq[k]*sdiv[k] + vdescale) >> hsv_shift; - } - v_int16 sd[2]; - sd[0] = v_pack(sq[0], sq[1]); - sd[1] = v_pack(sq[2], sq[3]); - s = v_pack_u(sd[0], sd[1]); + sq0 = v_shr(v_add(v_mul(diffq0, sdiv0), vdescale)); + sq1 = v_shr(v_add(v_mul(diffq1, sdiv1), vdescale)); + sq2 = v_shr(v_add(v_mul(diffq2, sdiv2), vdescale)); + sq3 = v_shr(v_add(v_mul(diffq3, sdiv3), vdescale)); + v_int16 sd0, sd1; + sd0 = v_pack(sq0, sq1); + sd1 = v_pack(sq2, sq3); + s = v_pack_u(sd0, sd1); // expand all to 16 bits - v_uint16 bdu[2], gdu[2], rdu[2]; - v_expand(b, bdu[0], bdu[1]); - v_expand(g, gdu[0], gdu[1]); - v_expand(r, rdu[0], rdu[1]); - v_int16 bd[2], gd[2], rd[2]; - bd[0] = v_reinterpret_as_s16(bdu[0]); - bd[1] = v_reinterpret_as_s16(bdu[1]); - gd[0] = v_reinterpret_as_s16(gdu[0]); - gd[1] = v_reinterpret_as_s16(gdu[1]); - rd[0] = v_reinterpret_as_s16(rdu[0]); - rd[1] = v_reinterpret_as_s16(rdu[1]); - - v_int16 vrd[2], vgd[2]; - v_expand(v_reinterpret_as_s8(vr), vrd[0], vrd[1]); - v_expand(v_reinterpret_as_s8(vg), vgd[0], vgd[1]); - v_int16 diffsd[2]; - diffsd[0] = v_reinterpret_as_s16(diffd[0]); - diffsd[1] = v_reinterpret_as_s16(diffd[1]); - - v_int16 hd[2]; + v_uint16 bdu0, bdu1, gdu0, gdu1, rdu0, rdu1; + v_expand(b, bdu0, bdu1); + v_expand(g, gdu0, gdu1); + v_expand(r, rdu0, rdu1); + v_int16 bd0, bd1, gd0, gd1, rd0, rd1; + bd0 = v_reinterpret_as_s16(bdu0); + bd1 = v_reinterpret_as_s16(bdu1); + gd0 = v_reinterpret_as_s16(gdu0); + gd1 = v_reinterpret_as_s16(gdu1); + rd0 = v_reinterpret_as_s16(rdu0); + rd1 = v_reinterpret_as_s16(rdu1); + + v_int16 vrd0, vrd1, vgd0, vgd1; + v_expand(v_reinterpret_as_s8(vr), vrd0, vrd1); + v_expand(v_reinterpret_as_s8(vg), vgd0, vgd1); + v_int16 diffsd0, diffsd1; + diffsd0 = v_reinterpret_as_s16(diffd0); + diffsd1 = v_reinterpret_as_s16(diffd1); + + v_int16 hd0, hd1; // h before division - for (int k = 0; k < 2; k++) - { - v_int16 gb = gd[k] - bd[k]; - v_int16 br = bd[k] - rd[k] + (diffsd[k] << 1); - v_int16 rg = rd[k] - gd[k] + (diffsd[k] << 2); - hd[k] = (vrd[k] & gb) + ((~vrd[k]) & ((vgd[k] & br) + ((~vgd[k]) & rg))); - } + v_int16 gb = v_sub(gd0 ,bd0); + v_int16 br = v_add(v_sub(bd0 ,rd0), v_shl<1>(diffsd0)); + v_int16 rg = v_add(v_sub(rd0 ,gd0), v_shl<2>(diffsd0)); + hd0 = v_add(v_and(vrd0, gb), v_and(v_not(vrd0), v_add(v_and(vgd0, br), v_and(v_not(vgd0), rg)))); + gb = v_sub(gd1, bd1); + br = v_add(v_sub(bd1, rd1), v_shl<1>(diffsd1)); + rg = v_add(v_sub(rd1, gd1), v_shl<2>(diffsd1)); + hd1 = v_add(v_and(vrd1, gb), v_and(v_not(vrd1), v_add(v_and(vgd1, br), v_and(v_not(vgd1), rg)))); // h div and fix - v_int32 hq[4]; - v_expand(hd[0], hq[0], hq[1]); - v_expand(hd[1], hq[2], hq[3]); - for(int k = 0; k < 4; k++) - { - hq[k] = (hq[k]*hdiv[k] + vdescale) >> hsv_shift; - } - hd[0] = v_pack(hq[0], hq[1]); - hd[1] = v_pack(hq[2], hq[3]); + v_int32 hq0, hq1, hq2, hq3; + v_expand(hd0, hq0, hq1); + v_expand(hd1, hq2, hq3); + hq0 = v_shr(v_add(v_mul(hq0, hdiv0), vdescale)); + hq1 = v_shr(v_add(v_mul(hq1, hdiv1), vdescale)); + hq2 = v_shr(v_add(v_mul(hq2, hdiv2), vdescale)); + hq3 = v_shr(v_add(v_mul(hq3, hdiv3), vdescale)); + + hd0 = v_pack(hq0, hq1); + hd1 = v_pack(hq2, hq3); v_int16 vhr = vx_setall_s16((short)hr); v_int16 vzd = vx_setzero_s16(); - hd[0] += v_select(hd[0] < vzd, vhr, vzd); - hd[1] += v_select(hd[1] < vzd, vhr, vzd); - h = v_pack_u(hd[0], hd[1]); + hd0 = v_add(hd0 ,v_select(v_lt(hd0, vzd), vhr, vzd)); + hd1 = v_add(hd1 ,v_select(v_lt(hd1, vzd), vhr, vzd)); + h = v_pack_u(hd0, hd1); v_store_interleave(dst, h, s, v); } @@ -260,7 +274,7 @@ struct RGB2HSV_f : srccn(_srccn), blueIdx(_blueIdx), hrange(_hrange) { } - #if CV_SIMD + #if CV_SIMD || CV_SIMD_SCALABLE inline void process(const v_float32& v_r, const v_float32& v_g, const v_float32& v_b, v_float32& v_h, v_float32& v_s, v_float32& v_v, float hscale) const @@ -269,17 +283,18 @@ struct RGB2HSV_f v_float32 v_max_rgb = v_max(v_max(v_r, v_g), v_b); v_float32 v_eps = vx_setall_f32(FLT_EPSILON); - v_float32 v_diff = v_max_rgb - v_min_rgb; - v_s = v_diff / (v_abs(v_max_rgb) + v_eps); - - v_float32 v_r_eq_max = v_r == v_max_rgb; - v_float32 v_g_eq_max = v_g == v_max_rgb; - v_h = v_select(v_r_eq_max, v_g - v_b, - v_select(v_g_eq_max, v_b - v_r, v_r - v_g)); - v_float32 v_res = v_select(v_r_eq_max, (v_g < v_b) & vx_setall_f32(360.0f), - v_select(v_g_eq_max, vx_setall_f32(120.0f), vx_setall_f32(240.0f))); - v_float32 v_rev_diff = vx_setall_f32(60.0f) / (v_diff + v_eps); - v_h = v_muladd(v_h, v_rev_diff, v_res) * vx_setall_f32(hscale); + v_float32 v_diff = v_sub(v_max_rgb, v_min_rgb); + v_s = v_div(v_diff, v_add(v_abs(v_max_rgb), v_eps)); + + v_float32 v_r_eq_max = v_eq(v_r, v_max_rgb); + v_float32 v_g_eq_max = v_eq(v_g, v_max_rgb); + v_h = v_select(v_r_eq_max, v_sub(v_g, v_b), + v_select(v_g_eq_max, v_sub(v_b, v_r), v_sub(v_r, v_g))); + v_float32 v_res = v_select(v_r_eq_max, + v_select(v_lt(v_g, v_b), vx_setall_f32(360.0f), vx_setall_f32(0.0f)), + v_select(v_g_eq_max, vx_setall_f32(120.0f), vx_setall_f32(240.0f))); + v_float32 v_rev_diff = v_div(vx_setall_f32(60.0f), v_add(v_diff, v_eps)); + v_h = v_mul(v_muladd(v_h, v_rev_diff, v_res), vx_setall_f32(hscale)); v_v = v_max_rgb; } @@ -293,8 +308,8 @@ struct RGB2HSV_f float hscale = hrange*(1.f/360.f); n *= 3; -#if CV_SIMD - const int vsize = v_float32::nlanes; +#if CV_SIMD || CV_SIMD_SCALABLE + const int vsize = VTraits::vlanes(); for ( ; i <= n - 3*vsize; i += 3*vsize, src += scn * vsize) { v_float32 r, g, b, a; @@ -353,7 +368,7 @@ struct RGB2HSV_f }; -#if CV_SIMD +#if CV_SIMD || CV_SIMD_SCALABLE inline void HSV2RGB_simd(const v_float32& h, const v_float32& s, const v_float32& v, v_float32& b, v_float32& g, v_float32& r, float hscale) { @@ -361,43 +376,43 @@ inline void HSV2RGB_simd(const v_float32& h, const v_float32& s, const v_float32 v_float32 v_s = s; v_float32 v_v = v; - v_h = v_h * vx_setall_f32(hscale); + v_h = v_mul(v_h, vx_setall_f32(hscale)); v_float32 v_pre_sector = v_cvt_f32(v_trunc(v_h)); - v_h = v_h - v_pre_sector; + v_h = v_sub(v_h, v_pre_sector); v_float32 v_tab0 = v_v; v_float32 v_one = vx_setall_f32(1.0f); - v_float32 v_tab1 = v_v * (v_one - v_s); - v_float32 v_tab2 = v_v * (v_one - (v_s * v_h)); - v_float32 v_tab3 = v_v * (v_one - (v_s * (v_one - v_h))); + v_float32 v_tab1 = v_mul(v_v, v_sub(v_one, v_s)); + v_float32 v_tab2 = v_mul(v_v, v_sub(v_one, v_mul(v_s, v_h))); + v_float32 v_tab3 = v_mul(v_v, v_sub(v_one, v_mul(v_s, v_sub(v_one, v_h)))); v_float32 v_one_sixth = vx_setall_f32(1.0f / 6.0f); - v_float32 v_sector = v_pre_sector * v_one_sixth; + v_float32 v_sector = v_mul(v_pre_sector, v_one_sixth); v_sector = v_cvt_f32(v_trunc(v_sector)); v_float32 v_six = vx_setall_f32(6.0f); - v_sector = v_pre_sector - (v_sector * v_six); + v_sector = v_sub(v_pre_sector, v_mul(v_sector, v_six)); v_float32 v_two = vx_setall_f32(2.0f); - v_h = v_tab1 & (v_sector < v_two); - v_h = v_h | (v_tab3 & (v_sector == v_two)); + v_h = v_select(v_lt(v_sector, v_two), v_tab1, vx_setall_f32(0.0f)); + v_h = v_select(v_eq(v_sector, v_two), v_tab3, v_h); v_float32 v_three = vx_setall_f32(3.0f); - v_h = v_h | (v_tab0 & (v_sector == v_three)); + v_h = v_select(v_eq(v_sector, v_three), v_tab0, v_h); v_float32 v_four = vx_setall_f32(4.0f); - v_h = v_h | (v_tab0 & (v_sector == v_four)); - v_h = v_h | (v_tab2 & (v_sector > v_four)); - - v_s = v_tab3 & (v_sector < v_one); - v_s = v_s | (v_tab0 & (v_sector == v_one)); - v_s = v_s | (v_tab0 & (v_sector == v_two)); - v_s = v_s | (v_tab2 & (v_sector == v_three)); - v_s = v_s | (v_tab1 & (v_sector > v_three)); - - v_v = v_tab0 & (v_sector < v_one); - v_v = v_v | (v_tab2 & (v_sector == v_one)); - v_v = v_v | (v_tab1 & (v_sector == v_two)); - v_v = v_v | (v_tab1 & (v_sector == v_three)); - v_v = v_v | (v_tab3 & (v_sector == v_four)); - v_v = v_v | (v_tab0 & (v_sector > v_four)); + v_h = v_select(v_eq(v_sector, v_four), v_tab0, v_h); + v_h = v_select(v_gt(v_sector, v_four), v_tab2, v_h); + + v_s = v_select(v_lt(v_sector, v_one), v_tab3, v_s); + v_s = v_select(v_eq(v_sector, v_one), v_tab0, v_s); + v_s = v_select(v_eq(v_sector, v_two), v_tab0, v_s); + v_s = v_select(v_eq(v_sector, v_three), v_tab2, v_s); + v_s = v_select(v_gt(v_sector, v_three), v_tab1, v_s); + + v_v = v_select(v_lt(v_sector, v_one), v_tab0, v_v); + v_v = v_select(v_eq(v_sector, v_one), v_tab2, v_v); + v_v = v_select(v_eq(v_sector, v_two), v_tab1, v_v); + v_v = v_select(v_eq(v_sector, v_three), v_tab1, v_v); + v_v = v_select(v_eq(v_sector, v_four), v_tab3, v_v); + v_v = v_select(v_gt(v_sector, v_four), v_tab0, v_v); b = v_h; g = v_s; @@ -457,8 +472,8 @@ struct HSV2RGB_f float hs = hscale; n *= 3; -#if CV_SIMD - const int vsize = v_float32::nlanes; +#if CV_SIMD || CV_SIMD_SCALABLE + const int vsize = VTraits::vlanes(); v_float32 valpha = vx_setall_f32(alpha); for (; i <= n - vsize*3; i += vsize*3, dst += dcn * vsize) { @@ -514,60 +529,102 @@ struct HSV2RGB_b int j = 0, dcn = dstcn; uchar alpha = ColorChannel::max(); -#if CV_SIMD - const int vsize = v_float32::nlanes; +#if CV_SIMD || CV_SIMD_SCALABLE + const int vsize = VTraits::vlanes(); for (j = 0; j <= (n - vsize*4) * 3; j += 3 * 4 * vsize, dst += dcn * 4 * vsize) { v_uint8 h_b, s_b, v_b; - v_uint16 h_w[2], s_w[2], v_w[2]; - v_uint32 h_u[4], s_u[4], v_u[4]; + v_uint16 h_w0, h_w1, s_w0, s_w1, v_w0, v_w1; + v_uint32 h_u0, h_u1, h_u2, h_u3, s_u0, s_u1, s_u2, s_u3, v_u0, v_u1, v_u2, v_u3; v_load_deinterleave(src + j, h_b, s_b, v_b); - v_expand(h_b, h_w[0], h_w[1]); - v_expand(s_b, s_w[0], s_w[1]); - v_expand(v_b, v_w[0], v_w[1]); - v_expand(h_w[0], h_u[0], h_u[1]); - v_expand(h_w[1], h_u[2], h_u[3]); - v_expand(s_w[0], s_u[0], s_u[1]); - v_expand(s_w[1], s_u[2], s_u[3]); - v_expand(v_w[0], v_u[0], v_u[1]); - v_expand(v_w[1], v_u[2], v_u[3]); - - v_int32 b_i[4], g_i[4], r_i[4]; + v_expand(h_b, h_w0, h_w1); + v_expand(s_b, s_w0, s_w1); + v_expand(v_b, v_w0, v_w1); + v_expand(h_w0, h_u0, h_u1); + v_expand(h_w1, h_u2, h_u3); + v_expand(s_w0, s_u0, s_u1); + v_expand(s_w1, s_u2, s_u3); + v_expand(v_w0, v_u0, v_u1); + v_expand(v_w1, v_u2, v_u3); + + v_int32 b_i0, b_i1, b_i2, b_i3, g_i0, g_i1, g_i2, g_i3, r_i0, r_i1, r_i2, r_i3; v_float32 v_coeff0 = vx_setall_f32(1.0f / 255.0f); v_float32 v_coeff1 = vx_setall_f32(255.0f); - for( int k = 0; k < 4; k++ ) - { - v_float32 h = v_cvt_f32(v_reinterpret_as_s32(h_u[k])); - v_float32 s = v_cvt_f32(v_reinterpret_as_s32(s_u[k])); - v_float32 v = v_cvt_f32(v_reinterpret_as_s32(v_u[k])); - - s *= v_coeff0; - v *= v_coeff0; - v_float32 b, g, r; - HSV2RGB_simd(h, s, v, b, g, r, hscale); - - b *= v_coeff1; - g *= v_coeff1; - r *= v_coeff1; - b_i[k] = v_trunc(b); - g_i[k] = v_trunc(g); - r_i[k] = v_trunc(r); - } - - v_uint16 r_w[2], g_w[2], b_w[2]; + v_float32 h = v_cvt_f32(v_reinterpret_as_s32(h_u0)); + v_float32 s = v_cvt_f32(v_reinterpret_as_s32(s_u0)); + v_float32 v = v_cvt_f32(v_reinterpret_as_s32(v_u0)); + + s = v_mul(s, v_coeff0); + v = v_mul(v, v_coeff0); + v_float32 b, g, r; + HSV2RGB_simd(h, s, v, b, g, r, hscale); + + b = v_mul(b, v_coeff1); + g = v_mul(g, v_coeff1); + r = v_mul(r, v_coeff1); + b_i0 = v_trunc(b); + g_i0 = v_trunc(g); + r_i0 = v_trunc(r); + + h = v_cvt_f32(v_reinterpret_as_s32(h_u1)); + s = v_cvt_f32(v_reinterpret_as_s32(s_u1)); + v = v_cvt_f32(v_reinterpret_as_s32(v_u1)); + + s = v_mul(s, v_coeff0); + v = v_mul(v, v_coeff0); + HSV2RGB_simd(h, s, v, b, g, r, hscale); + + b = v_mul(b, v_coeff1); + g = v_mul(g, v_coeff1); + r = v_mul(r, v_coeff1); + b_i1 = v_trunc(b); + g_i1 = v_trunc(g); + r_i1 = v_trunc(r); + + h = v_cvt_f32(v_reinterpret_as_s32(h_u2)); + s = v_cvt_f32(v_reinterpret_as_s32(s_u2)); + v = v_cvt_f32(v_reinterpret_as_s32(v_u2)); + + s = v_mul(s, v_coeff0); + v = v_mul(v, v_coeff0); + HSV2RGB_simd(h, s, v, b, g, r, hscale); + + b = v_mul(b, v_coeff1); + g = v_mul(g, v_coeff1); + r = v_mul(r, v_coeff1); + b_i2 = v_trunc(b); + g_i2 = v_trunc(g); + r_i2 = v_trunc(r); + + h = v_cvt_f32(v_reinterpret_as_s32(h_u3)); + s = v_cvt_f32(v_reinterpret_as_s32(s_u3)); + v = v_cvt_f32(v_reinterpret_as_s32(v_u3)); + + s = v_mul(s, v_coeff0); + v = v_mul(v, v_coeff0); + HSV2RGB_simd(h, s, v, b, g, r, hscale); + + b = v_mul(b, v_coeff1); + g = v_mul(g, v_coeff1); + r = v_mul(r, v_coeff1); + b_i3 = v_trunc(b); + g_i3 = v_trunc(g); + r_i3 = v_trunc(r); + + v_uint16 r_w0, r_w1, g_w0, g_w1, b_w0, b_w1; v_uint8 r_b, g_b, b_b; - r_w[0] = v_pack_u(r_i[0], r_i[1]); - r_w[1] = v_pack_u(r_i[2], r_i[3]); - r_b = v_pack(r_w[0], r_w[1]); - g_w[0] = v_pack_u(g_i[0], g_i[1]); - g_w[1] = v_pack_u(g_i[2], g_i[3]); - g_b = v_pack(g_w[0], g_w[1]); - b_w[0] = v_pack_u(b_i[0], b_i[1]); - b_w[1] = v_pack_u(b_i[2], b_i[3]); - b_b = v_pack(b_w[0], b_w[1]); + r_w0 = v_pack_u(r_i0, r_i1); + r_w1 = v_pack_u(r_i2, r_i3); + r_b = v_pack(r_w0, r_w1); + g_w0 = v_pack_u(g_i0, g_i1); + g_w1 = v_pack_u(g_i2, g_i3); + g_b = v_pack(g_w0, g_w1); + b_w0 = v_pack_u(b_i0, b_i1); + b_w1 = v_pack_u(b_i2, b_i3); + b_b = v_pack(b_w0, b_w1); if( dcn == 3 ) { @@ -621,7 +678,7 @@ struct RGB2HLS_f { } -#if CV_SIMD +#if CV_SIMD || CV_SIMD_SCALABLE inline void process(const v_float32& r, const v_float32& g, const v_float32& b, const v_float32& vhscale, v_float32& h, v_float32& l, v_float32& s) const @@ -629,28 +686,28 @@ struct RGB2HLS_f v_float32 maxRgb = v_max(v_max(r, g), b); v_float32 minRgb = v_min(v_min(r, g), b); - v_float32 diff = maxRgb - minRgb; - v_float32 msum = maxRgb + minRgb; + v_float32 diff = v_sub(maxRgb, minRgb); + v_float32 msum = v_add(maxRgb, minRgb); v_float32 vhalf = vx_setall_f32(0.5f); - l = msum * vhalf; + l = v_mul(msum, vhalf); - s = diff / v_select(l < vhalf, msum, vx_setall_f32(2.0f) - msum); + s = v_div(diff, v_select(v_lt(l, vhalf), msum, v_sub(vx_setall_f32(2.0f), msum))); - v_float32 rMaxMask = maxRgb == r; - v_float32 gMaxMask = maxRgb == g; + v_float32 rMaxMask = v_eq(maxRgb, r); + v_float32 gMaxMask = v_eq(maxRgb, g); - h = v_select(rMaxMask, g - b, v_select(gMaxMask, b - r, r - g)); - v_float32 hpart = v_select(rMaxMask, (g < b) & vx_setall_f32(360.0f), + h = v_select(rMaxMask, v_sub(g, b), v_select(gMaxMask, v_sub(b, r), v_sub(r, g))); + v_float32 hpart = v_select(rMaxMask, v_select(v_lt(g, b), vx_setall_f32(360.0f), vx_setall_f32(0.0f)), v_select(gMaxMask, vx_setall_f32(120.0f), vx_setall_f32(240.0f))); - v_float32 invDiff = vx_setall_f32(60.0f) / diff; - h = v_muladd(h, invDiff, hpart) * vhscale; + v_float32 invDiff = v_div(vx_setall_f32(60.0f), diff); + h = v_mul(v_muladd(h, invDiff, hpart), vhscale); - v_float32 diffEpsMask = diff > vx_setall_f32(FLT_EPSILON); + v_float32 diffEpsMask = v_gt(diff, vx_setall_f32(FLT_EPSILON)); - h = diffEpsMask & h; + h = v_select(diffEpsMask, h, vx_setall_f32(0.0f)); // l = l; - s = diffEpsMask & s; + s = v_select(diffEpsMask, s, vx_setall_f32(0.0f)); } #endif @@ -660,8 +717,8 @@ struct RGB2HLS_f int i = 0, bidx = blueIdx, scn = srccn; -#if CV_SIMD - const int vsize = v_float32::nlanes; +#if CV_SIMD || CV_SIMD_SCALABLE + const int vsize = VTraits::vlanes(); v_float32 vhscale = vx_setall_f32(hscale); for ( ; i <= n - vsize; @@ -744,29 +801,29 @@ struct RGB2HLS_b int scn = srccn; -#if CV_SIMD +#if CV_SIMD || CV_SIMD_SCALABLE float CV_DECL_ALIGNED(CV_SIMD_WIDTH) buf[bufChannels*BLOCK_SIZE]; #else float CV_DECL_ALIGNED(16) buf[bufChannels*BLOCK_SIZE]; #endif -#if CV_SIMD - static const int fsize = v_float32::nlanes; +#if CV_SIMD || CV_SIMD_SCALABLE + static const int fsize = VTraits::vlanes(); //TODO: fix that when v_interleave is available - float CV_DECL_ALIGNED(CV_SIMD_WIDTH) interTmpM[fsize*3]; + float CV_DECL_ALIGNED(CV_SIMD_WIDTH) interTmpM[VTraits::max_nlanes*3]; v_store_interleave(interTmpM, vx_setall_f32(1.f), vx_setall_f32(255.f), vx_setall_f32(255.f)); - v_float32 mhls[3]; - for(int k = 0; k < 3; k++) - { - mhls[k] = vx_load_aligned(interTmpM + k*fsize); - } + v_float32 mhls0, mhls1, mhls2, mhls3; + mhls0 = vx_load_aligned(interTmpM); + mhls1 = vx_load_aligned(interTmpM + fsize); + mhls2 = vx_load_aligned(interTmpM + 2*fsize); + mhls3 = vx_load_aligned(interTmpM + 3*fsize); #endif for(int i = 0; i < n; i += BLOCK_SIZE, dst += BLOCK_SIZE*3 ) { int dn = std::min(n - i, (int)BLOCK_SIZE); -#if CV_SIMD +#if CV_SIMD || CV_SIMD_SCALABLE v_float32 v255inv = vx_setall_f32(1.f/255.f); if (scn == 3) { @@ -778,8 +835,8 @@ struct RGB2HLS_b v_uint16 drgb = vx_load_expand(src); v_int32 qrgb0, qrgb1; v_expand(v_reinterpret_as_s16(drgb), qrgb0, qrgb1); - v_store_aligned(buf + j + 0*fsize, v_cvt_f32(qrgb0)*v255inv); - v_store_aligned(buf + j + 1*fsize, v_cvt_f32(qrgb1)*v255inv); + v_store_aligned(buf + j + 0*fsize, v_mul(v_cvt_f32(qrgb0),v255inv)); + v_store_aligned(buf + j + 1*fsize, v_mul(v_cvt_f32(qrgb1),v255inv)); } for( ; j < dn*3; j++, src++ ) { @@ -793,30 +850,39 @@ struct RGB2HLS_b for ( ; j <= dn*bufChannels - nBlock*bufChannels; j += nBlock*bufChannels, src += nBlock*4) { - v_uint8 rgb[3], dummy; - v_load_deinterleave(src, rgb[0], rgb[1], rgb[2], dummy); - - v_uint16 d[3*2]; - for(int k = 0; k < 3; k++) - { - v_expand(rgb[k], d[k*2+0], d[k*2+1]); - } - v_int32 q[3*4]; - for(int k = 0; k < 3*2; k++) - { - v_expand(v_reinterpret_as_s16(d[k]), q[k*2+0], q[k*2+1]); - } - - v_float32 f[3*4]; - for(int k = 0; k < 3*4; k++) - { - f[k] = v_cvt_f32(q[k])*v255inv; - } - - for(int k = 0; k < 4; k++) - { - v_store_interleave(buf + j + k*bufChannels*fsize, f[0*4+k], f[1*4+k], f[2*4+k]); - } + v_uint8 rgb0, rgb1, rgb2, rgb3, dummy; + v_load_deinterleave(src, rgb0, rgb1, rgb2, dummy); + + v_uint16 d0,d1,d2,d3,d4,d5; + v_expand(rgb0, d0, d1); + v_expand(rgb1, d2, d3); + v_expand(rgb2, d4, d5); + v_int32 q0,q1,q2,q3,q4,q5,q6,q7,q8,q9,q10,q11; + + v_expand(v_reinterpret_as_s16(d0), q0, q1); + v_expand(v_reinterpret_as_s16(d1), q2, q3); + v_expand(v_reinterpret_as_s16(d2), q4, q5); + v_expand(v_reinterpret_as_s16(d3), q6, q7); + v_expand(v_reinterpret_as_s16(d4), q8, q9); + v_expand(v_reinterpret_as_s16(d5), q10, q11); + v_float32 f0,f1,f2,f3,f4,f5,f6,f7,f8,f9,f10,f11; + f0 = v_mul(v_cvt_f32(q0), v255inv); + f1 = v_mul(v_cvt_f32(q1), v255inv); + f2 = v_mul(v_cvt_f32(q2), v255inv); + f3 = v_mul(v_cvt_f32(q3), v255inv); + f4 = v_mul(v_cvt_f32(q4), v255inv); + f5 = v_mul(v_cvt_f32(q5), v255inv); + f6 = v_mul(v_cvt_f32(q6), v255inv); + f7 = v_mul(v_cvt_f32(q7), v255inv); + f8 = v_mul(v_cvt_f32(q8), v255inv); + f9 = v_mul(v_cvt_f32(q9), v255inv); + f10 = v_mul(v_cvt_f32(q10), v255inv); + f11 = v_mul(v_cvt_f32(q11), v255inv); + + v_store_interleave(buf + j, f0, f4, f8); + v_store_interleave(buf + j + bufChannels*fsize, f1, f5, f9); + v_store_interleave(buf + j + 2*bufChannels*fsize, f2, f6, f10); + v_store_interleave(buf + j + 3*bufChannels*fsize, f3, f7, f11); } for( ; j < dn*3; j += 3, src += 4 ) { @@ -836,34 +902,53 @@ struct RGB2HLS_b cvt(buf, buf, dn); int j = 0; -#if CV_SIMD +#if CV_SIMD || CV_SIMD_SCALABLE for( ; j <= dn*3 - fsize*3*4; j += fsize*3*4) { - v_float32 f[3*4]; - for(int k = 0; k < 3*4; k++) - { - f[k] = vx_load_aligned(buf + j + k*fsize); - } - - for(int k = 0; k < 4; k++) - { - for(int l = 0; l < 3; l++) - { - f[k*3+l] = f[k*3+l] * mhls[l]; - } - } - - v_int32 q[3*4]; - for(int k = 0; k < 3*4; k++) - { - q[k] = v_round(f[k]); - } - - for(int k = 0; k < 3; k++) - { - v_store(dst + j + k*fsize*4, v_pack_u(v_pack(q[k*4+0], q[k*4+1]), - v_pack(q[k*4+2], q[k*4+3]))); - } + v_float32 f0,f1,f2,f3,f4,f5,f6,f7,f8,f9,f10,f11; + f0 = vx_load_aligned(buf + j + 0*fsize); + f1 = vx_load_aligned(buf + j + 1*fsize); + f2 = vx_load_aligned(buf + j + 2*fsize); + f3 = vx_load_aligned(buf + j + 3*fsize); + f4 = vx_load_aligned(buf + j + 4*fsize); + f5 = vx_load_aligned(buf + j + 5*fsize); + f6 = vx_load_aligned(buf + j + 6*fsize); + f7 = vx_load_aligned(buf + j + 7*fsize); + f8 = vx_load_aligned(buf + j + 8*fsize); + f9 = vx_load_aligned(buf + j + 9*fsize); + f10 = vx_load_aligned(buf + j + 10*fsize); + f11 = vx_load_aligned(buf + j + 11*fsize); + + f0 = v_mul(f0, mhls0); + f1 = v_mul(f1, mhls1); + f2 = v_mul(f2, mhls2); + f3 = v_mul(f3, mhls0); + f4 = v_mul(f4, mhls1); + f5 = v_mul(f5, mhls2); + f6 = v_mul(f6, mhls0); + f7 = v_mul(f7, mhls1); + f8 = v_mul(f8, mhls2); + f9 = v_mul(f9, mhls0); + f10 = v_mul(f10, mhls1); + f11 = v_mul(f11, mhls2); + + v_int32 q0,q1,q2,q3,q4,q5,q6,q7,q8,q9,q10,q11; + q0 = v_round(f0); + q1 = v_round(f1); + q2 = v_round(f2); + q3 = v_round(f3); + q4 = v_round(f4); + q5 = v_round(f5); + q6 = v_round(f6); + q7 = v_round(f7); + q8 = v_round(f8); + q9 = v_round(f9); + q10 = v_round(f10); + q11 = v_round(f11); + + v_store(dst + j + 0*fsize*4, v_pack_u(v_pack(q0, q1),v_pack(q2, q3))); + v_store(dst + j + 1*fsize*4, v_pack_u(v_pack(q4, q5),v_pack(q6, q7))); + v_store(dst + j + 2*fsize*4, v_pack_u(v_pack(q8, q9),v_pack(q10, q11))); } #endif for( ; j < dn*3; j += 3 ) @@ -888,39 +973,39 @@ struct HLS2RGB_f : dstcn(_dstcn), blueIdx(_blueIdx), hscale(6.f/_hrange) { } -#if CV_SIMD +#if CV_SIMD || CV_SIMD_SCALABLE inline void process(const v_float32& h, const v_float32& l, const v_float32& s, v_float32& b, v_float32& g, v_float32& r) const { v_float32 v1 = vx_setall_f32(1.0f), v2 = vx_setall_f32(2.0f), v4 = vx_setall_f32(4.0f); - v_float32 lBelowHalfMask = l <= vx_setall_f32(0.5f); - v_float32 ls = l * s; - v_float32 elem0 = v_select(lBelowHalfMask, ls, s - ls); + v_float32 lBelowHalfMask = v_le(l, vx_setall_f32(0.5f)); + v_float32 ls = v_mul(l, s); + v_float32 elem0 = v_select(lBelowHalfMask, ls, v_sub(s, ls)); - v_float32 hsRaw = h * vx_setall_f32(hscale); + v_float32 hsRaw = v_mul(h, vx_setall_f32(hscale)); v_float32 preHs = v_cvt_f32(v_trunc(hsRaw)); - v_float32 hs = hsRaw - preHs; - v_float32 sector = preHs - vx_setall_f32(6.0f) * v_cvt_f32(v_trunc(hsRaw * vx_setall_f32(1.0f / 6.0f))); - v_float32 elem1 = hs + hs; - - v_float32 tab0 = l + elem0; - v_float32 tab1 = l - elem0; - v_float32 tab2 = l + elem0 - elem0 * elem1; - v_float32 tab3 = l - elem0 + elem0 * elem1; - - b = v_select(sector < v2, tab1, - v_select(sector <= v2, tab3, - v_select(sector <= v4, tab0, tab2))); - - g = v_select(sector < v1, tab3, - v_select(sector <= v2, tab0, - v_select(sector < v4, tab2, tab1))); - - r = v_select(sector < v1, tab0, - v_select(sector < v2, tab2, - v_select(sector < v4, tab1, - v_select(sector <= v4, tab3, tab0)))); + v_float32 hs = v_sub(hsRaw, preHs); + v_float32 sector = v_sub(preHs, v_mul(vx_setall_f32(6.0f), v_cvt_f32(v_trunc(v_mul(hsRaw, vx_setall_f32(1.0f / 6.0f)))))); + v_float32 elem1 = v_add(hs, hs); + + v_float32 tab0 = v_add(l, elem0); + v_float32 tab1 = v_sub(l, elem0); + v_float32 tab2 = v_sub(v_add(l, elem0), v_mul(elem0, elem1)); + v_float32 tab3 = v_add(v_sub(l, elem0), v_mul(elem0, elem1)); + + b = v_select(v_lt(sector, v2), tab1, + v_select(v_le(sector, v2), tab3, + v_select(v_le(sector, v4), tab0, tab2))); + + g = v_select(v_lt(sector, v1), tab3, + v_select(v_le(sector, v2), tab0, + v_select(v_lt(sector, v4), tab2, tab1))); + + r = v_select(v_lt(sector, v1), tab0, + v_select(v_lt(sector, v2), tab2, + v_select(v_lt(sector, v4), tab1, + v_select(v_le(sector, v4), tab3, tab0)))); } #endif @@ -931,8 +1016,8 @@ struct HLS2RGB_f int i = 0, bidx = blueIdx, dcn = dstcn; float alpha = ColorChannel::max(); -#if CV_SIMD - static const int vsize = v_float32::nlanes; +#if CV_SIMD || CV_SIMD_SCALABLE + static const int vsize = VTraits::vlanes(); for (; i <= n - vsize; i += vsize, src += 3*vsize, dst += dcn*vsize) { v_float32 h, l, s, r, g, b; @@ -1020,23 +1105,22 @@ struct HLS2RGB_b int i, j, dcn = dstcn; uchar alpha = ColorChannel::max(); -#if CV_SIMD +#if CV_SIMD || CV_SIMD_SCALABLE float CV_DECL_ALIGNED(CV_SIMD_WIDTH) buf[bufChannels*BLOCK_SIZE]; #else float CV_DECL_ALIGNED(16) buf[bufChannels*BLOCK_SIZE]; #endif -#if CV_SIMD - static const int fsize = v_float32::nlanes; +#if CV_SIMD || CV_SIMD_SCALABLE + static const int fsize = VTraits::vlanes(); //TODO: fix that when v_interleave is available - float CV_DECL_ALIGNED(CV_SIMD_WIDTH) interTmpM[fsize*3]; + float CV_DECL_ALIGNED(CV_SIMD_WIDTH) interTmpM[VTraits::max_nlanes*3]; v_float32 v255inv = vx_setall_f32(1.f/255.f); v_store_interleave(interTmpM, vx_setall_f32(1.f), v255inv, v255inv); - v_float32 mhls[3]; - for(int k = 0; k < 3; k++) - { - mhls[k] = vx_load_aligned(interTmpM + k*fsize); - } + v_float32 mhls0, mhls1, mhls2; + mhls0 = vx_load_aligned(interTmpM + 0*fsize); + mhls1 = vx_load_aligned(interTmpM + 1*fsize); + mhls2 = vx_load_aligned(interTmpM + 2*fsize); #endif for( i = 0; i < n; i += BLOCK_SIZE, src += BLOCK_SIZE*3 ) @@ -1044,39 +1128,53 @@ struct HLS2RGB_b int dn = std::min(n - i, (int)BLOCK_SIZE); j = 0; -#if CV_SIMD +#if CV_SIMD || CV_SIMD_SCALABLE for( ; j <= dn*3 - 3*4*fsize; j += 3*4*fsize) { // 3x uchar -> 3*4 float - v_uint8 u[3]; - for(int k = 0; k < 3; k++) - { - u[k] = vx_load(src + j + k*4*fsize); - } - v_uint16 d[3*2]; - for(int k = 0; k < 3; k++) - { - v_expand(u[k], d[k*2+0], d[k*2+1]); - } - v_int32 q[3*4]; - for(int k = 0; k < 3*2; k++) - { - v_expand(v_reinterpret_as_s16(d[k]), q[k*2+0], q[k*2+1]); - } - - v_float32 f[3*4]; - for(int k = 0; k < 4; k++) - { - for(int l = 0; l < 3; l++) - { - f[k*3+l] = v_cvt_f32(q[k*3+l])*mhls[l]; - } - } - - for (int k = 0; k < 4*3; k++) - { - v_store_aligned(buf + j + k*fsize, f[k]); - } + v_uint8 u0, u1, u2; + u0 = vx_load(src + j + 0*4*fsize); + u1 = vx_load(src + j + 1*4*fsize); + u2 = vx_load(src + j + 2*4*fsize); + v_uint16 d0, d1, d2, d3, d4, d5; + v_expand(u0, d0, d1); + v_expand(u1, d2, d3); + v_expand(u2, d4, d5); + + v_int32 q0,q1,q2,q3,q4,q5,q6,q7,q8,q9,q10,q11; + v_expand(v_reinterpret_as_s16(d0), q0, q1); + v_expand(v_reinterpret_as_s16(d1), q2, q3); + v_expand(v_reinterpret_as_s16(d2), q4, q5); + v_expand(v_reinterpret_as_s16(d3), q6, q7); + v_expand(v_reinterpret_as_s16(d4), q8, q9); + v_expand(v_reinterpret_as_s16(d5), q10, q11); + + v_float32 f0,f1,f2,f3,f4,f5,f6,f7,f8,f9,f10,f11; + f0 = v_mul(v_cvt_f32(q0),mhls0); + f1 = v_mul(v_cvt_f32(q1),mhls1); + f2 = v_mul(v_cvt_f32(q2),mhls2); + f3 = v_mul(v_cvt_f32(q3),mhls0); + f4 = v_mul(v_cvt_f32(q4),mhls1); + f5 = v_mul(v_cvt_f32(q5),mhls2); + f6 = v_mul(v_cvt_f32(q6),mhls0); + f7 = v_mul(v_cvt_f32(q7),mhls1); + f8 = v_mul(v_cvt_f32(q8),mhls2); + f9 = v_mul(v_cvt_f32(q9),mhls0); + f10 = v_mul(v_cvt_f32(q10),mhls1); + f11 = v_mul(v_cvt_f32(q11),mhls2); + + v_store_aligned(buf + j + 0*fsize, f0); + v_store_aligned(buf + j + 1*fsize, f1); + v_store_aligned(buf + j + 2*fsize, f2); + v_store_aligned(buf + j + 3*fsize, f3); + v_store_aligned(buf + j + 4*fsize, f4); + v_store_aligned(buf + j + 5*fsize, f5); + v_store_aligned(buf + j + 6*fsize, f6); + v_store_aligned(buf + j + 7*fsize, f7); + v_store_aligned(buf + j + 8*fsize, f8); + v_store_aligned(buf + j + 9*fsize, f9); + v_store_aligned(buf + j + 10*fsize, f10); + v_store_aligned(buf + j + 11*fsize, f11); } #endif for( ; j < dn*3; j += 3 ) @@ -1087,7 +1185,7 @@ struct HLS2RGB_b } cvt(buf, buf, dn); -#if CV_SIMD +#if CV_SIMD || CV_SIMD_SCALABLE v_float32 v255 = vx_setall_f32(255.f); if(dcn == 3) { @@ -1095,18 +1193,18 @@ struct HLS2RGB_b float* pbuf = buf; for( ; x <= dn - 4*fsize; x += 4*fsize, dst += 4*fsize, pbuf += 4*fsize) { - v_float32 vf[4]; - vf[0] = vx_load_aligned(pbuf + 0*fsize); - vf[1] = vx_load_aligned(pbuf + 1*fsize); - vf[2] = vx_load_aligned(pbuf + 2*fsize); - vf[3] = vx_load_aligned(pbuf + 3*fsize); - v_int32 vi[4]; - vi[0] = v_round(vf[0]*v255); - vi[1] = v_round(vf[1]*v255); - vi[2] = v_round(vf[2]*v255); - vi[3] = v_round(vf[3]*v255); - v_store(dst, v_pack_u(v_pack(vi[0], vi[1]), - v_pack(vi[2], vi[3]))); + v_float32 vf0, vf1, vf2, vf3; + vf0 = vx_load_aligned(pbuf + 0*fsize); + vf1 = vx_load_aligned(pbuf + 1*fsize); + vf2 = vx_load_aligned(pbuf + 2*fsize); + vf3 = vx_load_aligned(pbuf + 3*fsize); + v_int32 vi0, vi1, vi2, vi3; + vi0 = v_round(v_mul(vf0,v255)); + vi1 = v_round(v_mul(vf1,v255)); + vi2 = v_round(v_mul(vf2,v255)); + vi3 = v_round(v_mul(vf3,v255)); + v_store(dst, v_pack_u(v_pack(vi0, vi1), + v_pack(vi2, vi3))); } for( ; x < dn*3; x++, dst++, pbuf++) { @@ -1119,19 +1217,28 @@ struct HLS2RGB_b float* pbuf = buf; for ( ; x <= dn - 4*fsize; x += fsize, dst += 4*fsize, pbuf += bufChannels*fsize) { - v_float32 r[4], g[4], b[4]; - v_int32 ir[4], ig[4], ib[4]; - for(int k = 0; k < 4; k++) - { - v_load_deinterleave(pbuf, r[k], g[k], b[k]); - ir[k] = v_round(r[k]*v255); - ig[k] = v_round(g[k]*v255); - ib[k] = v_round(b[k]*v255); - } + v_float32 r0, r1, r2, r3, g0, g1, g2, g3, b0, b1, b2, b3; + v_int32 ir0, ir1, ir2, ir3, ig0, ig1, ig2, ig3, ib0, ib1, ib2, ib3; + v_load_deinterleave(pbuf, r0, g0, b0); + ir0 = v_round(v_mul(r0, v255)); + ig0 = v_round(v_mul(g0, v255)); + ib0 = v_round(v_mul(b0, v255)); + v_load_deinterleave(pbuf, r1, g1, b1); + ir1 = v_round(v_mul(r1, v255)); + ig1 = v_round(v_mul(g1, v255)); + ib1 = v_round(v_mul(b1, v255)); + v_load_deinterleave(pbuf, r2, g2, b2); + ir2 = v_round(v_mul(r2, v255)); + ig2 = v_round(v_mul(g2, v255)); + ib2 = v_round(v_mul(b2, v255)); + v_load_deinterleave(pbuf, r3, g3, b3); + ir3 = v_round(v_mul(r3, v255)); + ig3 = v_round(v_mul(g3, v255)); + ib3 = v_round(v_mul(b3, v255)); v_uint8 ur, ug, ub; - ur = v_pack_u(v_pack(ir[0], ir[1]), v_pack(ir[2], ir[3])); - ug = v_pack_u(v_pack(ig[0], ig[1]), v_pack(ig[2], ig[3])); - ub = v_pack_u(v_pack(ib[0], ib[1]), v_pack(ib[2], ib[3])); + ur = v_pack_u(v_pack(ir0, ir1), v_pack(ir2, ir3)); + ug = v_pack_u(v_pack(ig0, ig1), v_pack(ig2, ig3)); + ub = v_pack_u(v_pack(ib0, ib1), v_pack(ib2, ib3)); v_uint8 valpha = vx_setall_u8(alpha); v_store_interleave(dst, ur, ug, ub, valpha);