229 		__m128i lo, hi;  in opj_mct_encode_real()  local
236 		hi = _mm_shuffle_epi32(r, _MM_SHUFFLE(3, 3, 1, 1));  in opj_mct_encode_real()
238 		hi = _mm_mul_epi32(hi, ry);  in opj_mct_encode_real()
240 		hi = _mm_add_epi64(hi, mulround);  in opj_mct_encode_real()
242 		hi = _mm_slli_epi64(hi, 32-13);  in opj_mct_encode_real()
243 		y = _mm_blend_epi16(lo, hi, 0xCC);  in opj_mct_encode_real()
246 		hi = _mm_shuffle_epi32(g, _MM_SHUFFLE(3, 3, 1, 1));  in opj_mct_encode_real()
248 		hi = _mm_mul_epi32(hi, gy);  in opj_mct_encode_real()
250 		hi = _mm_add_epi64(hi, mulround);  in opj_mct_encode_real()
252 		hi = _mm_slli_epi64(hi, 32-13);  in opj_mct_encode_real()
253 		y = _mm_add_epi32(y, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
256 		hi = _mm_shuffle_epi32(b, _MM_SHUFFLE(3, 3, 1, 1));  in opj_mct_encode_real()
258 		hi = _mm_mul_epi32(hi, by);  in opj_mct_encode_real()
260 		hi = _mm_add_epi64(hi, mulround);  in opj_mct_encode_real()
262 		hi = _mm_slli_epi64(hi, 32-13);  in opj_mct_encode_real()
263 		y = _mm_add_epi32(y, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
271 		hi = _mm_cvtepi32_epi64(_mm_shuffle_epi32(b, _MM_SHUFFLE(3, 2, 3, 1)));  in opj_mct_encode_real()
273 		hi = _mm_slli_epi64(hi, 12);  in opj_mct_encode_real()
275 		hi = _mm_add_epi64(hi, mulround);  in opj_mct_encode_real()
277 		hi = _mm_slli_epi64(hi, 32-13);  in opj_mct_encode_real()
278 		u = _mm_blend_epi16(lo, hi, 0xCC);  in opj_mct_encode_real()
281 		hi = _mm_shuffle_epi32(r, _MM_SHUFFLE(3, 3, 1, 1));  in opj_mct_encode_real()
283 		hi = _mm_mul_epi32(hi, ru);  in opj_mct_encode_real()
285 		hi = _mm_add_epi64(hi, mulround);  in opj_mct_encode_real()
287 		hi = _mm_slli_epi64(hi, 32-13);  in opj_mct_encode_real()
288 		u = _mm_sub_epi32(u, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
291 		hi = _mm_shuffle_epi32(g, _MM_SHUFFLE(3, 3, 1, 1));  in opj_mct_encode_real()
293 		hi = _mm_mul_epi32(hi, gu);  in opj_mct_encode_real()
295 		hi = _mm_add_epi64(hi, mulround);  in opj_mct_encode_real()
297 		hi = _mm_slli_epi64(hi, 32-13);  in opj_mct_encode_real()
298 		u = _mm_sub_epi32(u, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
306 		hi = _mm_cvtepi32_epi64(_mm_shuffle_epi32(r, _MM_SHUFFLE(3, 2, 3, 1)));  in opj_mct_encode_real()
308 		hi = _mm_slli_epi64(hi, 12);  in opj_mct_encode_real()
310 		hi = _mm_add_epi64(hi, mulround);  in opj_mct_encode_real()
312 		hi = _mm_slli_epi64(hi, 32-13);  in opj_mct_encode_real()
313 		v = _mm_blend_epi16(lo, hi, 0xCC);  in opj_mct_encode_real()
316 		hi = _mm_shuffle_epi32(g, _MM_SHUFFLE(3, 3, 1, 1));  in opj_mct_encode_real()
318 		hi = _mm_mul_epi32(hi, gv);  in opj_mct_encode_real()
320 		hi = _mm_add_epi64(hi, mulround);  in opj_mct_encode_real()
322 		hi = _mm_slli_epi64(hi, 32-13);  in opj_mct_encode_real()
323 		v = _mm_sub_epi32(v, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
326 		hi = _mm_shuffle_epi32(b, _MM_SHUFFLE(3, 3, 1, 1));  in opj_mct_encode_real()
328 		hi = _mm_mul_epi32(hi, bv);  in opj_mct_encode_real()
330 		hi = _mm_add_epi64(hi, mulround);  in opj_mct_encode_real()
332 		hi = _mm_slli_epi64(hi, 32-13);  in opj_mct_encode_real()
333 		v = _mm_sub_epi32(v, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()