48 inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) {  in Transpose()
49     auto b16 = _mm_loadu_si128((const __m128i*)bs);  in Transpose()
68 inline void Sk4x4f::transpose(uint8_t bs[16]) const {  in transpose()
73     _mm_storeu_si128((__m128i*)bs, _mm_or_si128(A, _mm_or_si128(B, _mm_or_si128(G, R))));  in transpose()
91 inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) {  in Transpose()
92     auto b16 = vreinterpretq_u32_u8(vld1q_u8(bs));  in Transpose()
105 inline void Sk4x4f::transpose(uint8_t bs[16]) const {  in transpose()
110     vst1q_u8(bs, vreinterpretq_u8_u32(vorrq_u32(A, vorrq_u32(B, vorrq_u32(G, R)))));  in transpose()
128 inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) {  in Transpose()
130         { (float)bs[0], (float)bs[4], (float)bs[ 8], (float)bs[12] },  in Transpose()
131         { (float)bs[1], (float)bs[5], (float)bs[ 9], (float)bs[13] },  in Transpose()
132         { (float)bs[2], (float)bs[6], (float)bs[10], (float)bs[14] },  in Transpose()
133         { (float)bs[3], (float)bs[7], (float)bs[11], (float)bs[15] },  in Transpose()
146 inline void Sk4x4f::transpose(uint8_t bs[16]) const {  in transpose()
147     bs[ 0] = (uint8_t)r[0]; bs[ 1] = (uint8_t)g[0]; bs[ 2] = (uint8_t)b[0]; bs[ 3] = (uint8_t)a[0];  in transpose()
148     bs[ 4] = (uint8_t)r[1]; bs[ 5] = (uint8_t)g[1]; bs[ 6] = (uint8_t)b[1]; bs[ 7] = (uint8_t)a[1];  in transpose()
149     bs[ 8] = (uint8_t)r[2]; bs[ 9] = (uint8_t)g[2]; bs[10] = (uint8_t)b[2]; bs[11] = (uint8_t)a[2];  in transpose()
150     bs[12] = (uint8_t)r[3]; bs[13] = (uint8_t)g[3]; bs[14] = (uint8_t)b[3]; bs[15] = (uint8_t)a[3];  in transpose()