225 	s32 workspace1[8], workspace2[8];  in fwht()  local
262 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht()
263 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht()
264 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht()
265 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht()
267 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht()
268 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht()
269 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht()
270 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht()
273 		out[0] = workspace2[0] + workspace2[4];  in fwht()
274 		out[1] = workspace2[0] - workspace2[4];  in fwht()
275 		out[2] = workspace2[1] - workspace2[5];  in fwht()
276 		out[3] = workspace2[1] + workspace2[5];  in fwht()
277 		out[4] = workspace2[2] + workspace2[6];  in fwht()
278 		out[5] = workspace2[2] - workspace2[6];  in fwht()
279 		out[6] = workspace2[3] - workspace2[7];  in fwht()
280 		out[7] = workspace2[3] + workspace2[7];  in fwht()
300 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht()
301 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht()
302 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht()
303 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht()
305 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht()
306 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht()
307 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht()
308 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht()
310 		out[0 * 8] = workspace2[0] + workspace2[4];  in fwht()
311 		out[1 * 8] = workspace2[0] - workspace2[4];  in fwht()
312 		out[2 * 8] = workspace2[1] - workspace2[5];  in fwht()
313 		out[3 * 8] = workspace2[1] + workspace2[5];  in fwht()
314 		out[4 * 8] = workspace2[2] + workspace2[6];  in fwht()
315 		out[5 * 8] = workspace2[2] - workspace2[6];  in fwht()
316 		out[6 * 8] = workspace2[3] - workspace2[7];  in fwht()
317 		out[7 * 8] = workspace2[3] + workspace2[7];  in fwht()
330 	s32 workspace1[8], workspace2[8];  in fwht16()  local
350 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht16()
351 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht16()
352 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht16()
353 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht16()
355 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht16()
356 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht16()
357 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht16()
358 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht16()
361 		out[0] = workspace2[0] + workspace2[4];  in fwht16()
362 		out[1] = workspace2[0] - workspace2[4];  in fwht16()
363 		out[2] = workspace2[1] - workspace2[5];  in fwht16()
364 		out[3] = workspace2[1] + workspace2[5];  in fwht16()
365 		out[4] = workspace2[2] + workspace2[6];  in fwht16()
366 		out[5] = workspace2[2] - workspace2[6];  in fwht16()
367 		out[6] = workspace2[3] - workspace2[7];  in fwht16()
368 		out[7] = workspace2[3] + workspace2[7];  in fwht16()
388 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht16()
389 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht16()
390 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht16()
391 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht16()
393 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht16()
394 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht16()
395 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht16()
396 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht16()
399 		out[0*8] = workspace2[0] + workspace2[4];  in fwht16()
400 		out[1*8] = workspace2[0] - workspace2[4];  in fwht16()
401 		out[2*8] = workspace2[1] - workspace2[5];  in fwht16()
402 		out[3*8] = workspace2[1] + workspace2[5];  in fwht16()
403 		out[4*8] = workspace2[2] + workspace2[6];  in fwht16()
404 		out[5*8] = workspace2[2] - workspace2[6];  in fwht16()
405 		out[6*8] = workspace2[3] - workspace2[7];  in fwht16()
406 		out[7*8] = workspace2[3] + workspace2[7];  in fwht16()
416 	int workspace1[8], workspace2[8];  in ifwht()  local
437 		workspace2[0] = workspace1[0] + workspace1[2];  in ifwht()
438 		workspace2[1] = workspace1[0] - workspace1[2];  in ifwht()
439 		workspace2[2] = workspace1[1] - workspace1[3];  in ifwht()
440 		workspace2[3] = workspace1[1] + workspace1[3];  in ifwht()
442 		workspace2[4] = workspace1[4] + workspace1[6];  in ifwht()
443 		workspace2[5] = workspace1[4] - workspace1[6];  in ifwht()
444 		workspace2[6] = workspace1[5] - workspace1[7];  in ifwht()
445 		workspace2[7] = workspace1[5] + workspace1[7];  in ifwht()
448 		out[0] = workspace2[0] + workspace2[4];  in ifwht()
449 		out[1] = workspace2[0] - workspace2[4];  in ifwht()
450 		out[2] = workspace2[1] - workspace2[5];  in ifwht()
451 		out[3] = workspace2[1] + workspace2[5];  in ifwht()
452 		out[4] = workspace2[2] + workspace2[6];  in ifwht()
453 		out[5] = workspace2[2] - workspace2[6];  in ifwht()
454 		out[6] = workspace2[3] - workspace2[7];  in ifwht()
455 		out[7] = workspace2[3] + workspace2[7];  in ifwht()
475 		workspace2[0] = workspace1[0] + workspace1[2];  in ifwht()
476 		workspace2[1] = workspace1[0] - workspace1[2];  in ifwht()
477 		workspace2[2] = workspace1[1] - workspace1[3];  in ifwht()
478 		workspace2[3] = workspace1[1] + workspace1[3];  in ifwht()
480 		workspace2[4] = workspace1[4] + workspace1[6];  in ifwht()
481 		workspace2[5] = workspace1[4] - workspace1[6];  in ifwht()
482 		workspace2[6] = workspace1[5] - workspace1[7];  in ifwht()
483 		workspace2[7] = workspace1[5] + workspace1[7];  in ifwht()
489 			out[0 * 8] = workspace2[0] + workspace2[4];  in ifwht()
490 			out[1 * 8] = workspace2[0] - workspace2[4];  in ifwht()
491 			out[2 * 8] = workspace2[1] - workspace2[5];  in ifwht()
492 			out[3 * 8] = workspace2[1] + workspace2[5];  in ifwht()
493 			out[4 * 8] = workspace2[2] + workspace2[6];  in ifwht()
494 			out[5 * 8] = workspace2[2] - workspace2[6];  in ifwht()
495 			out[6 * 8] = workspace2[3] - workspace2[7];  in ifwht()
496 			out[7 * 8] = workspace2[3] + workspace2[7];  in ifwht()
503 			out[0 * 8] = workspace2[0] + workspace2[4];  in ifwht()
504 			out[1 * 8] = workspace2[0] - workspace2[4];  in ifwht()
505 			out[2 * 8] = workspace2[1] - workspace2[5];  in ifwht()
506 			out[3 * 8] = workspace2[1] + workspace2[5];  in ifwht()
507 			out[4 * 8] = workspace2[2] + workspace2[6];  in ifwht()
508 			out[5 * 8] = workspace2[2] - workspace2[6];  in ifwht()
509 			out[6 * 8] = workspace2[3] - workspace2[7];  in ifwht()
510 			out[7 * 8] = workspace2[3] + workspace2[7];  in ifwht()