149         __m128i leftB = _mm_loadu_si128((__m128i *)(refSamples+8));  in IntraModeDCLuma16bit_SSE4_1_INTRIN()  local
157         sum = _mm_add_epi16(sum, leftB);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
173 …leftB = _mm_add_epi16(leftB, temp0);  // refSamples[leftOffset+rowIndex] + 3*predictionPtr[writeIn…  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
178 …leftB = _mm_add_epi16(leftB, const2); // refSamples[leftOffset+rowIndex] + 3*predictionPtr[writeIn…  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
183 …leftB = _mm_srli_epi16(leftB, 2);     // (refSamples[leftOffset+rowIndex] + 3*predictionPtr[writeI…  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
206             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
210             leftB = _mm_srli_si128(leftB, 4);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
211             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
215             leftB = _mm_srli_si128(leftB, 4);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
216             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
220             leftB = _mm_srli_si128(leftB, 4);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
221             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
261             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
265             leftB = _mm_srli_si128(leftB, 2);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
266             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
270             leftB = _mm_srli_si128(leftB, 2);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
271             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
275             leftB = _mm_srli_si128(leftB, 2);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
276             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
280             leftB = _mm_srli_si128(leftB, 2);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
281             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
285             leftB = _mm_srli_si128(leftB, 2);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
286             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
290             leftB = _mm_srli_si128(leftB, 2);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
291             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
295             leftB = _mm_srli_si128(leftB, 2);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()
296             topL = _mm_blend_epi16(sum, leftB, 1);  in IntraModeDCLuma16bit_SSE4_1_INTRIN()