238         __m128i lo, hi;  in opj_mct_encode_real()  local
244         lo = r;  in opj_mct_encode_real()
246         lo = _mm_mul_epi32(lo, ry);  in opj_mct_encode_real()
248         lo = _mm_add_epi64(lo, mulround);  in opj_mct_encode_real()
250         lo = _mm_srli_epi64(lo, 13);  in opj_mct_encode_real()
252         y = _mm_blend_epi16(lo, hi, 0xCC);  in opj_mct_encode_real()
254         lo = g;  in opj_mct_encode_real()
256         lo = _mm_mul_epi32(lo, gy);  in opj_mct_encode_real()
258         lo = _mm_add_epi64(lo, mulround);  in opj_mct_encode_real()
260         lo = _mm_srli_epi64(lo, 13);  in opj_mct_encode_real()
262         y = _mm_add_epi32(y, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
264         lo = b;  in opj_mct_encode_real()
266         lo = _mm_mul_epi32(lo, by);  in opj_mct_encode_real()
268         lo = _mm_add_epi64(lo, mulround);  in opj_mct_encode_real()
270         lo = _mm_srli_epi64(lo, 13);  in opj_mct_encode_real()
272         y = _mm_add_epi32(y, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
279         lo = _mm_cvtepi32_epi64(_mm_shuffle_epi32(b, _MM_SHUFFLE(3, 2, 2, 0)));  in opj_mct_encode_real()
281         lo = _mm_slli_epi64(lo, 12);  in opj_mct_encode_real()
283         lo = _mm_add_epi64(lo, mulround);  in opj_mct_encode_real()
285         lo = _mm_srli_epi64(lo, 13);  in opj_mct_encode_real()
287         u = _mm_blend_epi16(lo, hi, 0xCC);  in opj_mct_encode_real()
289         lo = r;  in opj_mct_encode_real()
291         lo = _mm_mul_epi32(lo, ru);  in opj_mct_encode_real()
293         lo = _mm_add_epi64(lo, mulround);  in opj_mct_encode_real()
295         lo = _mm_srli_epi64(lo, 13);  in opj_mct_encode_real()
297         u = _mm_sub_epi32(u, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
299         lo = g;  in opj_mct_encode_real()
301         lo = _mm_mul_epi32(lo, gu);  in opj_mct_encode_real()
303         lo = _mm_add_epi64(lo, mulround);  in opj_mct_encode_real()
305         lo = _mm_srli_epi64(lo, 13);  in opj_mct_encode_real()
307         u = _mm_sub_epi32(u, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
314         lo = _mm_cvtepi32_epi64(_mm_shuffle_epi32(r, _MM_SHUFFLE(3, 2, 2, 0)));  in opj_mct_encode_real()
316         lo = _mm_slli_epi64(lo, 12);  in opj_mct_encode_real()
318         lo = _mm_add_epi64(lo, mulround);  in opj_mct_encode_real()
320         lo = _mm_srli_epi64(lo, 13);  in opj_mct_encode_real()
322         v = _mm_blend_epi16(lo, hi, 0xCC);  in opj_mct_encode_real()
324         lo = g;  in opj_mct_encode_real()
326         lo = _mm_mul_epi32(lo, gv);  in opj_mct_encode_real()
328         lo = _mm_add_epi64(lo, mulround);  in opj_mct_encode_real()
330         lo = _mm_srli_epi64(lo, 13);  in opj_mct_encode_real()
332         v = _mm_sub_epi32(v, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()
334         lo = b;  in opj_mct_encode_real()
336         lo = _mm_mul_epi32(lo, bv);  in opj_mct_encode_real()
338         lo = _mm_add_epi64(lo, mulround);  in opj_mct_encode_real()
340         lo = _mm_srli_epi64(lo, 13);  in opj_mct_encode_real()
342         v = _mm_sub_epi32(v, _mm_blend_epi16(lo, hi, 0xCC));  in opj_mct_encode_real()