253 	s32 workspace1[8], workspace2[8];  in fwht()  local
263 			workspace1[0]  = tmp[0] + tmp[1] - add;  in fwht()
264 			workspace1[1]  = tmp[0] - tmp[1];  in fwht()
266 			workspace1[2]  = tmp[2] + tmp[3] - add;  in fwht()
267 			workspace1[3]  = tmp[2] - tmp[3];  in fwht()
269 			workspace1[4]  = tmp[4] + tmp[5] - add;  in fwht()
270 			workspace1[5]  = tmp[4] - tmp[5];  in fwht()
272 			workspace1[6]  = tmp[6] + tmp[7] - add;  in fwht()
273 			workspace1[7]  = tmp[6] - tmp[7];  in fwht()
276 			workspace1[0]  = tmp[0] + tmp[2] - add;  in fwht()
277 			workspace1[1]  = tmp[0] - tmp[2];  in fwht()
279 			workspace1[2]  = tmp[4] + tmp[6] - add;  in fwht()
280 			workspace1[3]  = tmp[4] - tmp[6];  in fwht()
282 			workspace1[4]  = tmp[8] + tmp[10] - add;  in fwht()
283 			workspace1[5]  = tmp[8] - tmp[10];  in fwht()
285 			workspace1[6]  = tmp[12] + tmp[14] - add;  in fwht()
286 			workspace1[7]  = tmp[12] - tmp[14];  in fwht()
289 			workspace1[0]  = tmp[0] + tmp[3] - add;  in fwht()
290 			workspace1[1]  = tmp[0] - tmp[3];  in fwht()
292 			workspace1[2]  = tmp[6] + tmp[9] - add;  in fwht()
293 			workspace1[3]  = tmp[6] - tmp[9];  in fwht()
295 			workspace1[4]  = tmp[12] + tmp[15] - add;  in fwht()
296 			workspace1[5]  = tmp[12] - tmp[15];  in fwht()
298 			workspace1[6]  = tmp[18] + tmp[21] - add;  in fwht()
299 			workspace1[7]  = tmp[18] - tmp[21];  in fwht()
302 			workspace1[0]  = tmp[0] + tmp[4] - add;  in fwht()
303 			workspace1[1]  = tmp[0] - tmp[4];  in fwht()
305 			workspace1[2]  = tmp[8] + tmp[12] - add;  in fwht()
306 			workspace1[3]  = tmp[8] - tmp[12];  in fwht()
308 			workspace1[4]  = tmp[16] + tmp[20] - add;  in fwht()
309 			workspace1[5]  = tmp[16] - tmp[20];  in fwht()
311 			workspace1[6]  = tmp[24] + tmp[28] - add;  in fwht()
312 			workspace1[7]  = tmp[24] - tmp[28];  in fwht()
317 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht()
318 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht()
319 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht()
320 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht()
322 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht()
323 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht()
324 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht()
325 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht()
342 		workspace1[0]  = out[0] + out[1 * 8];  in fwht()
343 		workspace1[1]  = out[0] - out[1 * 8];  in fwht()
345 		workspace1[2]  = out[2 * 8] + out[3 * 8];  in fwht()
346 		workspace1[3]  = out[2 * 8] - out[3 * 8];  in fwht()
348 		workspace1[4]  = out[4 * 8] + out[5 * 8];  in fwht()
349 		workspace1[5]  = out[4 * 8] - out[5 * 8];  in fwht()
351 		workspace1[6]  = out[6 * 8] + out[7 * 8];  in fwht()
352 		workspace1[7]  = out[6 * 8] - out[7 * 8];  in fwht()
355 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht()
356 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht()
357 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht()
358 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht()
360 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht()
361 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht()
362 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht()
363 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht()
386 	s32 workspace1[8], workspace2[8];  in fwht16()  local
393 		workspace1[0]  = tmp[0] + tmp[1];  in fwht16()
394 		workspace1[1]  = tmp[0] - tmp[1];  in fwht16()
396 		workspace1[2]  = tmp[2] + tmp[3];  in fwht16()
397 		workspace1[3]  = tmp[2] - tmp[3];  in fwht16()
399 		workspace1[4]  = tmp[4] + tmp[5];  in fwht16()
400 		workspace1[5]  = tmp[4] - tmp[5];  in fwht16()
402 		workspace1[6]  = tmp[6] + tmp[7];  in fwht16()
403 		workspace1[7]  = tmp[6] - tmp[7];  in fwht16()
406 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht16()
407 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht16()
408 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht16()
409 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht16()
411 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht16()
412 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht16()
413 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht16()
414 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht16()
431 		workspace1[0]  = out[0] + out[1*8];  in fwht16()
432 		workspace1[1]  = out[0] - out[1*8];  in fwht16()
434 		workspace1[2]  = out[2*8] + out[3*8];  in fwht16()
435 		workspace1[3]  = out[2*8] - out[3*8];  in fwht16()
437 		workspace1[4]  = out[4*8] + out[5*8];  in fwht16()
438 		workspace1[5]  = out[4*8] - out[5*8];  in fwht16()
440 		workspace1[6]  = out[6*8] + out[7*8];  in fwht16()
441 		workspace1[7]  = out[6*8] - out[7*8];  in fwht16()
444 		workspace2[0] = workspace1[0] + workspace1[2];  in fwht16()
445 		workspace2[1] = workspace1[0] - workspace1[2];  in fwht16()
446 		workspace2[2] = workspace1[1] - workspace1[3];  in fwht16()
447 		workspace2[3] = workspace1[1] + workspace1[3];  in fwht16()
449 		workspace2[4] = workspace1[4] + workspace1[6];  in fwht16()
450 		workspace2[5] = workspace1[4] - workspace1[6];  in fwht16()
451 		workspace2[6] = workspace1[5] - workspace1[7];  in fwht16()
452 		workspace2[7] = workspace1[5] + workspace1[7];  in fwht16()
473 	int workspace1[8], workspace2[8];  in ifwht()  local
481 		workspace1[0]  = tmp[0] + tmp[1];  in ifwht()
482 		workspace1[1]  = tmp[0] - tmp[1];  in ifwht()
484 		workspace1[2]  = tmp[2] + tmp[3];  in ifwht()
485 		workspace1[3]  = tmp[2] - tmp[3];  in ifwht()
487 		workspace1[4]  = tmp[4] + tmp[5];  in ifwht()
488 		workspace1[5]  = tmp[4] - tmp[5];  in ifwht()
490 		workspace1[6]  = tmp[6] + tmp[7];  in ifwht()
491 		workspace1[7]  = tmp[6] - tmp[7];  in ifwht()
494 		workspace2[0] = workspace1[0] + workspace1[2];  in ifwht()
495 		workspace2[1] = workspace1[0] - workspace1[2];  in ifwht()
496 		workspace2[2] = workspace1[1] - workspace1[3];  in ifwht()
497 		workspace2[3] = workspace1[1] + workspace1[3];  in ifwht()
499 		workspace2[4] = workspace1[4] + workspace1[6];  in ifwht()
500 		workspace2[5] = workspace1[4] - workspace1[6];  in ifwht()
501 		workspace2[6] = workspace1[5] - workspace1[7];  in ifwht()
502 		workspace2[7] = workspace1[5] + workspace1[7];  in ifwht()
519 		workspace1[0]  = out[0] + out[1 * 8];  in ifwht()
520 		workspace1[1]  = out[0] - out[1 * 8];  in ifwht()
522 		workspace1[2]  = out[2 * 8] + out[3 * 8];  in ifwht()
523 		workspace1[3]  = out[2 * 8] - out[3 * 8];  in ifwht()
525 		workspace1[4]  = out[4 * 8] + out[5 * 8];  in ifwht()
526 		workspace1[5]  = out[4 * 8] - out[5 * 8];  in ifwht()
528 		workspace1[6]  = out[6 * 8] + out[7 * 8];  in ifwht()
529 		workspace1[7]  = out[6 * 8] - out[7 * 8];  in ifwht()
532 		workspace2[0] = workspace1[0] + workspace1[2];  in ifwht()
533 		workspace2[1] = workspace1[0] - workspace1[2];  in ifwht()
534 		workspace2[2] = workspace1[1] - workspace1[3];  in ifwht()
535 		workspace2[3] = workspace1[1] + workspace1[3];  in ifwht()
537 		workspace2[4] = workspace1[4] + workspace1[6];  in ifwht()
538 		workspace2[5] = workspace1[4] - workspace1[6];  in ifwht()
539 		workspace2[6] = workspace1[5] - workspace1[7];  in ifwht()
540 		workspace2[7] = workspace1[5] + workspace1[7];  in ifwht()