159     const __m128i a = _mm_add_epi16(in0, in2);  in ITransform()  local
217     const __m128i a =  _mm_add_epi16(dc, T2);  in ITransform()  local
480     const __m128i b3 = _mm_srai_epi32(_mm_sub_epi32(a0, a1), 1);  in FTransformWHT()  local
491 static int SSE_Nx4(const uint8_t* a, const uint8_t* b,  in SSE_Nx4()
507     const __m128i b3 = _mm_loadu_si128((__m128i*)&b[BPS * 3]);  in SSE_Nx4()  local
569 static int SSE16x16(const uint8_t* a, const uint8_t* b) {  in SSE16x16()
573 static int SSE16x8(const uint8_t* a, const uint8_t* b) {  in SSE16x8()
577 static int SSE8x8(const uint8_t* a, const uint8_t* b) {  in SSE8x8()
581 static int SSE4x4(const uint8_t* a, const uint8_t* b) {  in SSE4x4()
593   const __m128i b3 = _mm_loadl_epi64((__m128i*)&b[BPS * 3]);  in SSE4x4()  local
696     const __m128i b3 = _mm_sub_epi16(a0, a1);  in TTransform()  local
745     const __m128i b3 = _mm_sub_epi16(a0, a1);  in TTransform()  local
786 static int Disto4x4(const uint8_t* const a, const uint8_t* const b,  in Disto4x4()
792 static int Disto16x16(const uint8_t* const a, const uint8_t* const b,  in Disto16x16()