2 * Blackfin Pixel Operations
3 * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 #include "config_bfin.h"
23 DEFUN(put_pixels_clamped,mL1,
24 (DCTELEM *block, uint8_t *dest, int line_size)):
31 R2 += -4; // line_size
37 LSETUP (ppc$0,ppc$1) LC0=P0;
38 ppc$0: R2 = MIN(R2, R5) (V);
40 R3 = MIN(R3, R5) (V) || R0 = [I0++];
41 R6 = BYTEPACK (R2,R3) || R1 = [I0++];
42 R2 = MAX(R0, R4) (V) || [I1++] = R6;
45 R3 = MIN(R3, R5) (V) || R0 = [I0++];
46 R6 = BYTEPACK (R2,R3) || R1 = [I0++];
47 ppc$1: R2 = Max(R0, R4) (V) || [I1++M1] = R6;
51 DEFUN_END(put_pixels_clamped)
53 DEFUN(add_pixels_clamped,mL1,
54 (DCTELEM *block, uint8_t *dest, int line_size)):
58 R2 += -4; // line_size
65 R0 = [I3++] || R2 = [I1];
66 R2 = R2 << 8 || R0.H = W[I3--] || R3 = [I1++];
67 R3 = R3 >> 8 || R1.L = W[I3] || I3 += 4;
68 R6 = BYTEOP3P(R1:0, R3:2) (LO) || R1.H = W[I3++] || R2 = [I1];
70 LSETUP(apc$2,apc$3) LC1 = P0;
71 apc$2: R7 = BYTEOP3P(R1:0, R3:2) (HI, R) || R0 = [I3++] || R3 = [I1++M0];
72 R2 = R2 << 8 || R0.H = W[I3--];
73 R3 = R3 >> 8 || R1.L = W[I3] || I3 += 4;
74 R6 = R6 + R7 (S) || R1.H = W[I3];
75 R6 = BYTEOP3P(R1:0, R3:2) (LO) || I3+=M3 || [I2++]=R6;
76 R7 = BYTEOP3P(R1:0, R3:2) (HI, R) || R0 = [I3++] || R2 = [I1];
77 R2 = R2 << 8 || R0.H = W[I3--] || R3 = [I1++];
78 R3 = R3 >> 8 || R1.L = W[I3] || I3 += 4;
79 R6 = R6 + R7 (S) || R1.H = W[I3++];
80 apc$3: R6 = BYTEOP3P(R1:0, R3:2) (LO) || [I2++M0] = R6 || R2 = [I1];
84 DEFUN_END(add_pixels_clamped)
91 * Halfpel motion compensation with rounding (a+b+1)>>1.
92 * This is an array[4][4] of motion compensation funcions for 4
93 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
94 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
95 * @param block destination where the result is stored
96 * @param pixels source
97 * @param line_size number of bytes in a horizontal line of block
102 DEFUN(put_pixels8uc,mL1,
103 (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
104 int dest_size, int line_size, int h)):
108 r0=[sp+12]; // dest_size
109 r2=[sp+16]; // line_size
116 LSETUP(pp8$0,pp8$1) LC0=P0;
117 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
119 pp8$0: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
120 R6 = BYTEOP1P(R1:0,R3:2) || R0 = [I0++M0]|| R2 =[I1++M0];
121 R7 = BYTEOP1P(R1:0,R3:2)(R) || R0 = [I0++] || [I3++] = R6 ;
122 pp8$1: DISALGNEXCPT || R2 = [I1++] || [I3++M3] = R7;
126 DEFUN_END(put_pixels8uc)
128 DEFUN(put_pixels16uc,mL1,
129 (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
130 int dest_size, int line_size, int h)):
136 r0=[fp+20]; // dest_size
137 r2=[fp+24]; // line_size
146 LSETUP(pp16$0,pp16$1) LC0=P0;
147 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
149 pp16$0: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
150 R6 = BYTEOP1P(R1:0,R3:2) || R0 = [I0++] || R2 =[I1++];
151 R7 = BYTEOP1P(R1:0,R3:2)(R) || R1 = [I0++] || R3 =[I1++];
153 R6 = BYTEOP1P(R1:0,R3:2) || R0 = [I0++M0] || R2 =[I1++M0];
154 R7 = BYTEOP1P(R1:0,R3:2)(R) || R0 = [I0++] || [I3++] = R7 ;
156 pp16$1: DISALGNEXCPT || R2 = [I1++] || [I3++M3] = R7;
161 DEFUN_END(put_pixels16uc)
168 DEFUN(put_pixels8uc_nornd,mL1,
169 (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
170 int line_size, int h)):
174 r2=[sp+12]; // line_size
181 LSETUP(pp8$2,pp8$3) LC0=P0;
182 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
184 pp8$2: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
185 R6 = BYTEOP1P(R1:0,R3:2)(T) || R0 = [I0++M0]|| R2 =[I1++M0];
186 R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R0 = [I0++] || [I3++] = R6 ;
187 pp8$3: DISALGNEXCPT || R2 = [I1++] || [I3++M3] = R7;
191 DEFUN_END(put_pixels8uc_nornd)
193 DEFUN(put_pixels16uc_nornd,mL1,
194 (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
195 int line_size, int h)):
199 r2=[sp+12]; // line_size
208 LSETUP(pp16$2,pp16$3) LC0=P0;
209 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
212 DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
213 R6 = BYTEOP1P(R1:0,R3:2)(T) || R0 = [I0++] || R2 =[I1++];
214 R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R1 = [I0++] || R3 =[I1++];
217 R6 = BYTEOP1P(R1:0,R3:2)(T) || R0 = [I0++M0] || R2 =[I1++M0];
218 R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R0 = [I0++] || [I3++] = R7 ;
220 pp16$3: DISALGNEXCPT || R2 = [I1++] || [I3++M3] = R7;
225 DEFUN_END(put_pixels16uc_nornd)
227 DEFUN(z_put_pixels16_xy2,mL1,
228 (uint8_t *block, const uint8_t *s0,
229 int dest_size, int line_size, int h)):
233 i0=r1; // src0--> pixels
234 i1=r1; // src1--> pixels + line_size
236 m2=r2; // m2=dest_width-4
241 i1+=m3; /* src1 + line_size */
242 m0=r2; /* line-size - 20 */
248 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
250 LSETUP(LS$16E,LE$16E) LC0=P0;
251 LS$16E: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
252 R4 = BYTEOP2P (R3:2,R1:0) (RNDL) || R0 = [I0++] || R2 =[I1++];
253 R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R1 = [I0++] || [I3++] = R4 ;
254 DISALGNEXCPT || R3 = [I1++] || [I3++] = R5;
255 R4 = BYTEOP2P (R3:2,R1:0) (RNDL) || R0 = [I0++M0]|| R2 = [I1++M0];
256 R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R0 = [I0++] || [I3++] = R4 ;
257 LE$16E: DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
267 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
268 LSETUP(LS$16O,LE$16O) LC0=P0;
269 LS$16O: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
270 R4 = BYTEOP2P (R3:2,R1:0) (RNDH) || R0 = [I0++] || R2 =[I1++];
271 R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R1 = [I0++] || R6 =[I3++];
272 R4 = R4 +|+ R6 || R7 = [I3--];
273 R5 = R5 +|+ R7 || [I3++] = R4;
274 DISALGNEXCPT || R3 =[I1++] || [I3++] = R5;
275 R4 = BYTEOP2P (R3:2,R1:0) (RNDH) || R0 = [I0++M0]|| R2 = [I1++M0];
276 R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R0 = [I0++] || R6 = [I3++];
277 R4 = R4 +|+ R6 || R7 = [I3--];
278 R5 = R5 +|+ R7 || [I3++] = R4;
279 LE$16O: DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
284 DEFUN_END(z_put_pixels16_xy2)
286 DEFUN(put_pixels16_xy2_nornd,mL1,
287 (uint8_t *block, const uint8_t *s0,
288 int line_size, int h)):
292 i0=r1; // src0--> pixels
293 i1=r1; // src1--> pixels + line_size
298 i1+=m3; /* src1 + line_size */
299 m0=r2; /* line-size - 20 */
306 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
308 LSETUP(LS$16ET,LE$16ET) LC0=P0;
309 LS$16ET:DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
310 R4 = BYTEOP2P (R3:2,R1:0) (TL) || R0 = [I0++] || R2 =[I1++];
311 R5 = BYTEOP2P (R3:2,R1:0) (TL,R) || R1 = [I0++] || [I3++] = R4 ;
312 DISALGNEXCPT || R3 = [I1++] || [I3++] = R5;
313 R4 = BYTEOP2P (R3:2,R1:0) (TL) || R0 = [I0++M0]|| R2 = [I1++M0];
314 R5 = BYTEOP2P (R3:2,R1:0) (TL,R) || R0 = [I0++] || [I3++] = R4 ;
315 LE$16ET:DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
325 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
326 LSETUP(LS$16OT,LE$16OT) LC0=P0;
327 LS$16OT:DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
328 R4 = BYTEOP2P (R3:2,R1:0) (TH) || R0 = [I0++] || R2 =[I1++];
329 R5 = BYTEOP2P (R3:2,R1:0) (TH,R) || R1 = [I0++] || R6 =[I3++];
330 R4 = R4 +|+ R6 || R7 = [I3--];
331 R5 = R5 +|+ R7 || [I3++] = R4;
332 DISALGNEXCPT || R3 =[I1++] || [I3++] = R5;
333 R4 = BYTEOP2P (R3:2,R1:0) (TH) || R0 = [I0++M0]|| R2 = [I1++M0];
334 R5 = BYTEOP2P (R3:2,R1:0) (TH,R) || R0 = [I0++] || R6 = [I3++];
335 R4 = R4 +|+ R6 || R7 = [I3--];
336 R5 = R5 +|+ R7 || [I3++] = R4;
337 LE$16OT:DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
342 DEFUN_END(put_pixels16_xy2_nornd)
344 DEFUN(z_put_pixels8_xy2,mL1,
345 (uint8_t *block, const uint8_t *s0,
346 int dest_size, int line_size, int h)):
350 i0=r1; // src0--> pixels
351 i1=r1; // src1--> pixels + line_size
353 m2=r2; // m2=dest_width-4
358 i1+=m3; /* src1 + line_size */
359 m0=r2; /* line-size - 20 */
365 LSETUP(LS$8E,LE$8E) LC0=P0;
366 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
367 LS$8E: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
368 R4 = BYTEOP2P (R3:2,R1:0) (RNDL) || R0 = [I0++M0] || R2 =[I1++M0];
369 R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R0 = [I0++] || [I3++] = R4 ;
370 LE$8E: DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
380 LSETUP(LS$8O,LE$8O) LC0=P0;
381 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
382 LS$8O: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
383 R4 = BYTEOP2P (R3:2,R1:0) (RNDH) || R0 = [I0++M0] || R2 =[I1++M0];
384 R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R0 = [I0++] || R6 =[I3++];
385 R4 = R4 +|+ R6 || R7 = [I3--];
386 R5 = R5 +|+ R7 || [I3++] = R4;
387 LE$8O: DISALGNEXCPT || R2 =[I1++] || [I3++M2] = R5;
392 DEFUN_END(z_put_pixels8_xy2)
394 DEFUN(put_pixels8_xy2_nornd,mL1,
395 (uint8_t *block, const uint8_t *s0, int line_size, int h)):
399 i0=r1; // src0--> pixels
400 i1=r1; // src1--> pixels + line_size
405 i1+=m3; /* src1 + line_size */
406 m0=r2; /* line-size - 20 */
414 LSETUP(LS$8ET,LE$8ET) LC0=P0;
415 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
417 LS$8ET: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
418 R4 = BYTEOP2P (R3:2,R1:0) (TL) || R0 = [I0++M0] || R2 = [I1++M0];
419 R5 = BYTEOP2P (R3:2,R1:0) (TL,R) || R0 = [I0++] || [I3++] = R4 ;
420 LE$8ET: DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
430 LSETUP(LS$8OT,LE$8OT) LC0=P0;
431 DISALGNEXCPT || R0 = [I0++] || R2 = [I1++];
433 LS$8OT: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
434 R4 = BYTEOP2P (R3:2,R1:0) (TH) || R0 = [I0++M0] || R2 = [I1++M0];
435 R5 = BYTEOP2P (R3:2,R1:0) (TH,R) || R0 = [I0++] || R6 = [I3++];
436 R4 = R4 +|+ R6 || R7 = [I3--];
437 R5 = R5 +|+ R7 || [I3++] = R4;
438 LE$8OT: DISALGNEXCPT || R2 =[I1++] || [I3++M2] = R5;
444 DEFUN(diff_pixels,mL1,
445 (DCTELEM *block, uint8_t *s1, uint8_t *s2, int stride)):
452 r2=[fp+20]; // stride
457 LSETUP(.LS0,.LE0) LC0=P0;
458 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
460 .LS0: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
461 (R5,R4) = BYTEOP16M (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M0];
462 (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R0 = [I0++] || [I3++] = R4;
463 DISALGNEXCPT || R2 = [I1++] || [I3++] = R5;
470 DEFUN_END(put_pixels8_xy2_nornd)
473 for (i = 0; i < 16; i++) {
474 for (j = 0; j < 16; j++) {
481 (uint8_t *p, int stride)):
495 LSETUP(LS$PS,LE$PS) LC0=P0;
496 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
498 LS$PS: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
499 (R5,R4) = BYTEOP16P (R3:2,R1:0) || R0 = [I0++] || R2 = [I1++];
502 (R5,R4) = BYTEOP16P (R3:2,R1:0) (R)|| R1 = [I0++] || R3 = [I1++];
505 (R5,R4) = BYTEOP16P (R3:2,R1:0) || R0 = [I0++m0] || R2 = [I1++m0];
508 (R5,R4) = BYTEOP16P (R3:2,R1:0) (R)|| R0 = [I0++] || R2 = [I1++];
520 DEFUN(get_pixels,mL1,
521 (DCTELEM *restrict block, const uint8_t *pixels, int line_size)):
528 LSETUP(gp8$0,gp8$1) LC0=P0;
530 DISALGNEXCPT || R0 = [I0++];
531 DISALGNEXCPT || R1 = [I0++];
533 gp8$0: (R7,R6) = byteunpack R1:0 || R0 = [I0++M0];
534 (R5,R4) = byteunpack R1:0 (R) || R0 = [I0++] || [I3++]=R6;
535 DISALGNEXCPT || R1 = [I0++] || [I3++]=R7;
542 DEFUN_END(get_pixels)
545 /* sad = sad16x16 (ubyte *mb, ubyte *refwin, srcwidth, refwinwidth, h) */
547 DEFUN(z_sad16x16,mL1,
548 (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h)):
554 R0 = [sp+20]; // rwidth
555 P2 = [sp+24]; // height
562 DISALGNEXCPT || R0 = [I0++] || R2 = [I1++];
563 LSETUP (s$16, e$16) LC0=P2;
564 s$16: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
565 SAA (R1:0,R3:2) || R0 = [I0++] || R2 = [I1++];
566 SAA (R1:0,R3:2) (R) || R1 = [I0++] || R3 = [I1++];
567 SAA (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M1];
568 e$16: SAA (R1:0,R3:2) (R) || R0 = [I0++] || R2 = [I1++];
570 R3=A1.L+A1.H, R2=A0.L+A0.H ;
574 DEFUN_END(z_sad16x16)
576 /* sad = sad8x8 (ubyte *mb, ubyte *refwin, int srcwidth, int refwinwidth, int h) */
579 (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h)):
584 r0 = [sp+12]; // rwidth
585 P2 = [sp+16]; //height
592 LSETUP (s$8, e$8) LC0=P2;
593 DISALGNEXCPT || R0 = [I0++] || R2 = [I1++];
594 DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
595 s$8: SAA (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M1];
596 SAA (R1:0,R3:2) (R) || R0 = [I0++] || R2 = [I1++];
597 e$8: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
599 R3=A1.L+A1.H, R2=A0.L+A0.H ;
605 (uint8_t * pix, int line_size)):
608 // Fetch the input arguments.
610 P0 = R1; // line_size
611 P5 = 16; // loop ctr.
613 M0 = P0; // M0 = line_size-16;
614 // Now for the real work.
616 lsetup(_pix_norm1_blkfn_loopStart, _pix_norm1_blkfn_loopEnd) LC1 = P5;
618 DISALGNEXCPT || r0 = [i0++];
620 _pix_norm1_blkfn_loopStart:
621 // following unpacks pix1[0..15] pix1+line_size[0..15]
622 DISALGNEXCPT || r1 = [i0++];
624 (r5, r4) = byteunpack r1:0 || r0 = [i0++];
625 a1 += r5.h * r5.h, a0 += r5.l * r5.l (is);
626 a1 += r4.h * r4.h, a0 += r4.l * r4.l (is);
627 (r5, r4) = byteunpack r1:0(r) || r1 = [i0++];
628 a1 += r5.h * r5.h, a0 += r5.l * r5.l (is);
629 a1 += r4.h * r4.h, a0 += r4.l * r4.l (is);
630 (r5, r4) = byteunpack r1:0 || r0 = [i0++M0];
631 a1 += r5.h * r5.h, a0 += r5.l * r5.l (is);
632 a1 += r4.h * r4.h, a0 += r4.l * r4.l (is);
633 (r5, r4) = byteunpack r1:0(r) || r0 = [i0++];
634 a1 += r5.h * r5.h, a0 += r5.l * r5.l (is);
635 _pix_norm1_blkfn_loopEnd:
636 a1 += r4.h * r4.h, a0 += r4.l * r4.l (is);
639 // Clean up at the end:
649 (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)):
655 r2=[fp+20]; // line_size
660 LSETUP(.S40,.E40) LC0=P0;
661 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
663 .S40: DISALGNEXCPT || R1 = [I0++M0] || R3 = [I1++M0];
664 (R7,R6) = BYTEOP16M (R1:0,R3:2);
665 a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
666 .E40: a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
676 (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)):
682 r2=[fp+20]; // line_size
687 LSETUP(.S80,.E80) LC0=P0;
688 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
690 .S80: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
691 (R7,R6) = BYTEOP16M (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M0];
692 a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
693 a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
694 (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R0 = [I0++] || R2 = [I1++];
695 a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
696 .E80: a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
706 (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)):
712 r2=[fp+20]; // line_size
717 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
718 LSETUP(.S160,.E160) LC0=P0;
720 .S160: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
721 (R7,R6) = BYTEOP16M (R1:0,R3:2) || R0 = [I0++] || R2 = [I1++];
722 a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
723 a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
724 (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R1 = [I0++] || R3 = [I1++];
725 a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
726 a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
727 (R7,R6) = BYTEOP16M (R1:0,R3:2) || R0 = [I0++M0] || R2 = [I1++M0];
728 a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
729 a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
730 (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R0 = [I0++] || R2 = [I1++];
731 a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
732 .E160: a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);