From ed0ce48179663a89e1458a705ec090566c019079 Mon Sep 17 00:00:00 2001 From: orestis Date: Fri, 19 Dec 2014 22:44:39 +0200 Subject: [PATCH] SymmColumnVec_32f16s asymm NEON speedup: 9.46x Auto-vect speedup: 1x Test kernel: [-0.9432, -1.1528, 0, 1.1528, 0.9432] --- modules/imgproc/src/filter.cpp | 48 +++++++++++++++++++++++++++++++++++++++++- 1 file changed, 47 insertions(+), 1 deletion(-) diff --git a/modules/imgproc/src/filter.cpp b/modules/imgproc/src/filter.cpp index 6991a44..d0516f5 100644 --- a/modules/imgproc/src/filter.cpp +++ b/modules/imgproc/src/filter.cpp @@ -2873,7 +2873,53 @@ struct SymmColumnVec_32f16s } else { - return 0; + float32x2_t k32; + k32 = vdup_n_f32(0); + k32 = vld1_lane_f32(ky + 1, k32, 1); + + for( ; i <= width - 8; i += 8 ) + { + float32x4_t x1l, x1h, x2l, x2h; + float32x4_t accl, acch; + + S = src[1] + i; + S2 = src[-1] + i; + + x1l = vld1q_f32(S); + x1h = vld1q_f32(S + 4); + x2l = vld1q_f32(S2); + x2h = vld1q_f32(S2 + 4); + + accl = acch = d4; + accl = vmlaq_lane_f32(accl, vsubq_f32(x1l, x2l), k32, 1); + acch = vmlaq_lane_f32(acch, vsubq_f32(x1h, x2h), k32, 1); + + for( k = 2; k <= ksize2; k++ ) + { + S = src[k] + i; + S2 = src[-k] + i; + + float32x4_t x3l, x3h, x4l, x4h; + x3l = vld1q_f32(S); + x3h = vld1q_f32(S + 4); + x4l = vld1q_f32(S2); + x4h = vld1q_f32(S2 + 4); + + accl = vmlaq_n_f32(accl, vsubq_f32(x3l, x4l), ky[k]); + acch = vmlaq_n_f32(acch, vsubq_f32(x3h, x4h), ky[k]); + } + + int32x4_t s32l, s32h; + s32l = vcvtq_s32_f32(accl); + s32h = vcvtq_s32_f32(acch); + + int16x4_t s16l, s16h; + s16l = vqmovn_s32(s32l); + s16h = vqmovn_s32(s32h); + + vst1_s16((int16_t *)(dst + i), s16l); + vst1_s16((int16_t *)(dst + i + 4), s16h); + } } return i; -- 2.7.4