From 284006e336a3b0bb3ce3824f508f2d41a2f21039 Mon Sep 17 00:00:00 2001 From: k-shinotsuka Date: Wed, 23 Nov 2016 19:18:41 +0900 Subject: [PATCH] add NEON code for RGB2Luv_f. --- modules/imgproc/src/color.cpp | 97 +++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 94 insertions(+), 3 deletions(-) diff --git a/modules/imgproc/src/color.cpp b/modules/imgproc/src/color.cpp index 9f3e6ac..6c10e37 100644 --- a/modules/imgproc/src/color.cpp +++ b/modules/imgproc/src/color.cpp @@ -141,7 +141,32 @@ template static inline _Tp splineInterpolate(_Tp x, const _Tp* tab return ((tab[3]*x + tab[2])*x + tab[1])*x + tab[0]; } -#if CV_SSE2 +#if CV_NEON +template static inline void splineInterpolate(float32x4_t& v_x, const _Tp* tab, int n) +{ + int32x4_t v_ix = vcvtq_s32_f32(vminq_f32(vmaxq_f32(v_x, vdupq_n_f32(0)), vdupq_n_f32(n - 1))); + v_x = vsubq_f32(v_x, vcvtq_f32_s32(v_ix)); + v_ix = vshlq_n_s32(v_ix, 2); + + int CV_DECL_ALIGNED(16) ix[4]; + vst1q_s32(ix, v_ix); + + float32x4_t v_tab0 = vld1q_f32(tab + ix[0]); + float32x4_t v_tab1 = vld1q_f32(tab + ix[1]); + float32x4_t v_tab2 = vld1q_f32(tab + ix[2]); + float32x4_t v_tab3 = vld1q_f32(tab + ix[3]); + + float32x4x2_t v01 = vtrnq_f32(v_tab0, v_tab1); + float32x4x2_t v23 = vtrnq_f32(v_tab2, v_tab3); + + v_tab0 = vcombine_f32(vget_low_f32(v01.val[0]), vget_low_f32(v23.val[0])); + v_tab1 = vcombine_f32(vget_low_f32(v01.val[1]), vget_low_f32(v23.val[1])); + v_tab2 = vcombine_f32(vget_high_f32(v01.val[0]), vget_high_f32(v23.val[0])); + v_tab3 = vcombine_f32(vget_high_f32(v01.val[1]), vget_high_f32(v23.val[1])); + + v_x = vmlaq_f32(v_tab0, vmlaq_f32(v_tab1, vmlaq_f32(v_tab2, v_tab3, v_x), v_x), v_x); +} +#elif CV_SSE2 template static inline void splineInterpolate(__m128& v_x, const _Tp* tab, int n) { __m128i v_ix = _mm_cvttps_epi32(_mm_min_ps(_mm_max_ps(v_x, _mm_setzero_ps()), _mm_set1_ps(float(n - 1)))); @@ -5968,7 +5993,28 @@ struct RGB2Luv_f CV_Assert(whitept[1] == 1.f); } - #if CV_SSE2 + #if CV_NEON + void process(float32x4x3_t& v_src) const + { + float32x4_t v_x = vmlaq_f32(vmlaq_f32(vmulq_f32(v_src.val[0], vdupq_n_f32(coeffs[0])), v_src.val[1], vdupq_n_f32(coeffs[1])), v_src.val[2], vdupq_n_f32(coeffs[2])); + float32x4_t v_y = vmlaq_f32(vmlaq_f32(vmulq_f32(v_src.val[0], vdupq_n_f32(coeffs[3])), v_src.val[1], vdupq_n_f32(coeffs[4])), v_src.val[2], vdupq_n_f32(coeffs[5])); + float32x4_t v_z = vmlaq_f32(vmlaq_f32(vmulq_f32(v_src.val[0], vdupq_n_f32(coeffs[6])), v_src.val[1], vdupq_n_f32(coeffs[7])), v_src.val[2], vdupq_n_f32(coeffs[8])); + + v_src.val[0] = vmulq_f32(v_y, vdupq_n_f32(LabCbrtTabScale)); + splineInterpolate(v_src.val[0], LabCbrtTab, LAB_CBRT_TAB_SIZE); + + v_src.val[0] = vmlaq_f32(vdupq_n_f32(-16.f), v_src.val[0], vdupq_n_f32(116.f)); + + float32x4_t v_div = vmaxq_f32(vmlaq_f32(vmlaq_f32(v_x, vdupq_n_f32(15.f), v_y), vdupq_n_f32(3.f), v_z), vdupq_n_f32(FLT_EPSILON)); + float32x4_t v_reciprocal = vrecpeq_f32(v_div); + v_reciprocal = vmulq_f32(vrecpsq_f32(v_div, v_reciprocal), v_reciprocal); + v_reciprocal = vmulq_f32(vrecpsq_f32(v_div, v_reciprocal), v_reciprocal); + float32x4_t v_d = vmulq_f32(vdupq_n_f32(52.f), v_reciprocal); + + v_src.val[1] = vmulq_f32(v_src.val[0], vmlaq_f32(vdupq_n_f32(-un), v_x, v_d)); + v_src.val[2] = vmulq_f32(v_src.val[0], vmlaq_f32(vdupq_n_f32(-vn), vmulq_f32(vdupq_n_f32(2.25f), v_y), v_d)); + } + #elif CV_SSE2 void process(__m128& v_r0, __m128& v_r1, __m128& v_g0, __m128& v_g1, __m128& v_b0, __m128& v_b1) const { @@ -6042,7 +6088,52 @@ struct RGB2Luv_f C6 = coeffs[6], C7 = coeffs[7], C8 = coeffs[8]; n *= 3; - #if CV_SSE2 + #if CV_NEON + if (scn == 3) + { + for( ; i <= n - 12; i += 12, src += scn * 4 ) + { + float32x4x3_t v_src = vld3q_f32(src); + if( gammaTab ) + { + v_src.val[0] = vmulq_f32(v_src.val[0], vdupq_n_f32(gscale)); + v_src.val[1] = vmulq_f32(v_src.val[1], vdupq_n_f32(gscale)); + v_src.val[2] = vmulq_f32(v_src.val[2], vdupq_n_f32(gscale)); + splineInterpolate(v_src.val[0], gammaTab, GAMMA_TAB_SIZE); + splineInterpolate(v_src.val[1], gammaTab, GAMMA_TAB_SIZE); + splineInterpolate(v_src.val[2], gammaTab, GAMMA_TAB_SIZE); + } + + process(v_src); + + vst3q_f32(dst + i, v_src); + } + } + else + { + for( ; i <= n - 12; i += 12, src += scn * 4 ) + { + float32x4x4_t v_src = vld4q_f32(src); + if( gammaTab ) + { + v_src.val[0] = vmulq_f32(v_src.val[0], vdupq_n_f32(gscale)); + v_src.val[1] = vmulq_f32(v_src.val[1], vdupq_n_f32(gscale)); + v_src.val[2] = vmulq_f32(v_src.val[2], vdupq_n_f32(gscale)); + splineInterpolate(v_src.val[0], gammaTab, GAMMA_TAB_SIZE); + splineInterpolate(v_src.val[1], gammaTab, GAMMA_TAB_SIZE); + splineInterpolate(v_src.val[2], gammaTab, GAMMA_TAB_SIZE); + } + + float32x4x3_t v_dst; + v_dst.val[0] = v_src.val[0]; + v_dst.val[1] = v_src.val[1]; + v_dst.val[2] = v_src.val[2]; + process(v_dst); + + vst3q_f32(dst + i, v_dst); + } + } + #elif CV_SSE2 if (haveSIMD) { for( ; i <= n - 24; i += 24, src += scn * 8 ) -- 2.7.4