101 …          uint16x8_t v_src0 = vmovl_u8(vget_low_u8(v_src)), v_src1 = vmovl_u8(vget_high_u8(v_src));  in operator ()()  local
103 … vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(vmovl_u16(vget_low_u16(v_src0)))));  in operator ()()
104 …2(dst + x + 4, vaddq_f32(vld1q_f32(dst + x + 4), vcvtq_f32_u32(vmovl_u16(vget_high_u16(v_src0)))));  in operator ()()
116 …          uint16x8_t v_src0 = vmovl_u8(vget_low_u8(v_src)), v_src1 = vmovl_u8(vget_high_u8(v_src));  in operator ()()  local
118 … vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(vmovl_u16(vget_low_u16(v_src0)))));  in operator ()()
119 …2(dst + x + 4, vaddq_f32(vld1q_f32(dst + x + 4), vcvtq_f32_u32(vmovl_u16(vget_high_u16(v_src0)))));  in operator ()()
142 …      uint32x4_t v_src0 = vmovl_u16(vget_low_u16(v_src)), v_src1 = vmovl_u16(vget_high_u16(v_src));  in operator ()()  local
144                 vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(v_src0)));  in operator ()()
188                 uint16x8_t v_src0 = vmull_u8(v_src_0, v_src_0), v_src1 = vmull_u8(v_src_1, v_src_1);  in operator ()()  local
190 … vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(vmovl_u16(vget_low_u16(v_src0)))));  in operator ()()
191 …2(dst + x + 4, vaddq_f32(vld1q_f32(dst + x + 4), vcvtq_f32_u32(vmovl_u16(vget_high_u16(v_src0)))));  in operator ()()
204                 uint16x8_t v_src0 = vmull_u8(v_src_0, v_src_0), v_src1 = vmull_u8(v_src_1, v_src_1);  in operator ()()  local
206 … vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(vmovl_u16(vget_low_u16(v_src0)))));  in operator ()()
207 …2(dst + x + 4, vaddq_f32(vld1q_f32(dst + x + 4), vcvtq_f32_u32(vmovl_u16(vget_high_u16(v_src0)))));  in operator ()()
231 …             uint32x4_t v_src0 = vmull_u16(v_src_0, v_src_0), v_src1 = vmull_u16(v_src_1, v_src_1);  in operator ()()  local
233                 vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(v_src0)));  in operator ()()
249 …             uint32x4_t v_src0 = vmull_u16(v_src_0, v_src_0), v_src1 = vmull_u16(v_src_1, v_src_1);  in operator ()()  local
251                 vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(v_src0)));  in operator ()()
297                 uint16x8_t v_src0 = vmull_u8(vget_low_u8(v_1src), vget_low_u8(v_2src)),  in operator ()()  local
300 … vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(vmovl_u16(vget_low_u16(v_src0)))));  in operator ()()
301 …2(dst + x + 4, vaddq_f32(vld1q_f32(dst + x + 4), vcvtq_f32_u32(vmovl_u16(vget_high_u16(v_src0)))));  in operator ()()
314                 uint16x8_t v_src0 = vmull_u8(vget_low_u8(v_1src), vget_low_u8(v_2src)),  in operator ()()  local
317 … vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(vmovl_u16(vget_low_u16(v_src0)))));  in operator ()()
318 …2(dst + x + 4, vaddq_f32(vld1q_f32(dst + x + 4), vcvtq_f32_u32(vmovl_u16(vget_high_u16(v_src0)))));  in operator ()()
341                 uint32x4_t v_src0 = vmull_u16(vget_low_u16(v_1src), vget_low_u16(v_2src)),  in operator ()()  local
344                 vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(v_src0)));  in operator ()()
360                 uint32x4_t v_src0 = vmull_u16(vget_low_u16(v_1src), vget_low_u16(v_2src)),  in operator ()()  local
363                 vst1q_f32(dst + x, vaddq_f32(vld1q_f32(dst + x), vcvtq_f32_u32(v_src0)));  in operator ()()
407 …          uint16x8_t v_src0 = vmovl_u8(vget_low_u8(v_src)), v_src1 = vmovl_u8(vget_high_u8(v_src));  in operator ()()  local
410 …                                         vcvtq_f32_u32(vmovl_u16(vget_low_u16(v_src0))), v_alpha));  in operator ()()
412 …                                        vcvtq_f32_u32(vmovl_u16(vget_high_u16(v_src0))), v_alpha));  in operator ()()
438 …      uint32x4_t v_src0 = vmovl_u16(vget_low_u16(v_src)), v_src1 = vmovl_u16(vget_high_u16(v_src));  in operator ()()  local
440 …t1q_f32(dst + x, vmlaq_f32(vmulq_f32(vld1q_f32(dst + x), v_beta), vcvtq_f32_u32(v_src0), v_alpha));  in operator ()()