template <bool align> SIMD_INLINE void LoadBgr(const __m256i * p, __m256i & blue, __m256i & green, __m256i & red) { __m256i bgr[3]; bgr[0] = Load<align>(p + 0); bgr[1] = Load<align>(p + 1); bgr[2] = Load<align>(p + 2); blue = BgrToBlue(bgr); green = BgrToGreen(bgr); red = BgrToRed(bgr); }
template <bool align> SIMD_INLINE void LoadBgr(const uint8_t * p, v128_u8 & blue, v128_u8 & green, v128_u8 & red) { v128_u8 bgr[3]; bgr[0] = Load<align>(p); bgr[1] = Load<align>(p + A); bgr[2] = Load<align>(p + DA); blue = BgrToBlue(bgr); green = BgrToGreen(bgr); red = BgrToRed(bgr); }