106   __m128i row0, tmp1, tmp2, tmp3, row2, tmp5, tmp6, tmp7;  in jpeg_idct_intelsse()  local
122     row0 = _mm_shufflelo_epi16(row0, 0xD8); /*x7, x6, x5, x4, x3, x1, x2, x0*/    \  in jpeg_idct_intelsse()
124     tmp1 = _mm_shuffle_epi32(row0, 0);      /*x2, x0, x2, x0, x2, x0, x2, x0*/    \  in jpeg_idct_intelsse()
127     tmp3 = _mm_shuffle_epi32(row0, 0x55);   /*x3, x1, x3, x1, x3, x1, x3, x1*/    \  in jpeg_idct_intelsse()
129     row0 = _mm_shufflehi_epi16(row0, 0xD8); /*x7, x5, x6, x4, x3, x1, x2, x0*/    \  in jpeg_idct_intelsse()
135     tmp2 =  _mm_shuffle_epi32(row0, 0xAA);  /*x6, x4, x6, x4, x6, x4, x6, x4*/    \  in jpeg_idct_intelsse()
137     row0 = _mm_shuffle_epi32(row0, 0xFF);   /*x7, x5, x7, x5, x7, x5, x7, x5*/    \  in jpeg_idct_intelsse()
142 …row0 = _mm_madd_epi16(row0, * ( __m128i*)(table1+24)); /*x7*w31+x5*w30, x7*w27+x5*w26, x7*w23+x5*w…  in jpeg_idct_intelsse()
149     row0 = _mm_add_epi32(row0, tmp3);    /*b3, b2, b1, b0*/  \  in jpeg_idct_intelsse()
156     tmp2 = _mm_sub_epi32(tmp2, row0); /*for row0. y4= a3-b3, y5=a2-b2, y6=a1-b1, y7=a0-b0 */   \  in jpeg_idct_intelsse()
158     row0 = _mm_add_epi32(row0, tmp1); /*y3=a3+b3,y2=a2+b2,y1=a1+b1,y0=a0+b0*/   \  in jpeg_idct_intelsse()
162     row0 = _mm_srai_epi32(row0, SHIFT_INV_ROW);             \  in jpeg_idct_intelsse()
166     row0 = _mm_packs_epi32(row0, tmp2); /*row0 = y7,y6,y5,y4,y3,y2,y1,y0*/  \  in jpeg_idct_intelsse()
173     x5 = row0;\  in jpeg_idct_intelsse()
281   row0 = _mm_load_si128((__m128i const*)(coef_blockSSE));  in jpeg_idct_intelsse()
283   row0 = _mm_mullo_epi16( row0, *(__m128i const*)quantptrSSE );  in jpeg_idct_intelsse()
288   _mm_store_si128((__m128i*)(wsptr), row0);  in jpeg_idct_intelsse()
292   row0 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*4));  in jpeg_idct_intelsse()
294   row0 = _mm_mullo_epi16(row0, *(__m128i const*)(quantptrSSE+8*4) );  in jpeg_idct_intelsse()
299   _mm_store_si128((__m128i*)(wsptr+32), row0);  in jpeg_idct_intelsse()
303   row0 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*3));  in jpeg_idct_intelsse()
305   row0 = _mm_mullo_epi16(row0, *(__m128i const*)(quantptrSSE+24) );  in jpeg_idct_intelsse()
310   _mm_store_si128((__m128i*)(wsptr+24), row0);  in jpeg_idct_intelsse()
314   row0 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*5));  in jpeg_idct_intelsse()
316   row0 = _mm_mullo_epi16(row0, *(__m128i const*)(quantptrSSE+40) );  in jpeg_idct_intelsse()