41       "vdup.32 q4, %[input_range_min]\n"  in Transform()
42       "vdup.32 q5, %[output_range_min]\n"  in Transform()
43       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
44       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
45       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
46       "vsub.f32 q4, q4, q5\n"  in Transform()
49       "subs %[count], %[count], #16\n"  in Transform()
52       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
53       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
54       "pld [%[input], #64]\n"  in Transform()
55       "vcvt.f32.s32 q0, q0\n"  in Transform()
56       "vcvt.f32.s32 q1, q1\n"  in Transform()
57       "vcvt.f32.s32 q2, q2\n"  in Transform()
58       "vcvt.f32.s32 q3, q3\n"  in Transform()
59       "vsub.f32 q0, q0, q6\n"  in Transform()
60       "vsub.f32 q1, q1, q6\n"  in Transform()
61       "vsub.f32 q2, q2, q6\n"  in Transform()
62       "vsub.f32 q3, q3, q6\n"  in Transform()
63       "vmul.f32 q0, q0, q7\n"  in Transform()
64       "vmul.f32 q1, q1, q7\n"  in Transform()
65       "vmul.f32 q2, q2, q7\n"  in Transform()
66       "vmul.f32 q3, q3, q7\n"  in Transform()
67       "vadd.f32 q0, q0, q4\n"  in Transform()
68       "vadd.f32 q1, q1, q4\n"  in Transform()
69       "vadd.f32 q2, q2, q4\n"  in Transform()
70       "vadd.f32 q3, q3, q4\n"  in Transform()
71       "vmul.f32 q0, q0, q8\n"  in Transform()
72       "vmul.f32 q1, q1, q8\n"  in Transform()
73       "vmul.f32 q2, q2, q8\n"  in Transform()
74       "vmul.f32 q3, q3, q8\n"  in Transform()
75       "vcvt.s32.f32 q0, q0\n"  in Transform()
76       "vcvt.s32.f32 q1, q1\n"  in Transform()
77       "vcvt.s32.f32 q2, q2\n"  in Transform()
78       "vcvt.s32.f32 q3, q3\n"  in Transform()
79       "vqmovn.s32 d0, q0\n"  in Transform()
80       "vqmovn.s32 d1, q1\n"  in Transform()
81       "vqmovn.s32 d4, q2\n"  in Transform()
82       "vqmovn.s32 d5, q3\n"  in Transform()
83       "vqmovun.s16 d0, q0\n"  in Transform()
84       "vqmovun.s16 d1, q2\n"  in Transform()
86       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
87       "pld [%[output]]\n"  in Transform()
89       "bne 1b\n"  in Transform()
116       "vdup.32 q4, %[input_range_min]\n"  in Transform()
117       "vdup.32 q5, %[output_range_min]\n"  in Transform()
118       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
119       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
120       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
121       "vsub.f32 q4, q4, q5\n"  in Transform()
124       "subs %[count], %[count], #1\n"  in Transform()
125       "beq 2f\n"  in Transform()
128       "subs %[count], %[count], #16\n"  in Transform()
131       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
132       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
133       "pld [%[input], #64]\n"  in Transform()
134       "vcvt.f32.s32 q0, q0\n"  in Transform()
135       "vcvt.f32.s32 q1, q1\n"  in Transform()
136       "vcvt.f32.s32 q2, q2\n"  in Transform()
137       "vcvt.f32.s32 q3, q3\n"  in Transform()
138       "vsub.f32 q0, q0, q6\n"  in Transform()
139       "vsub.f32 q1, q1, q6\n"  in Transform()
140       "vsub.f32 q2, q2, q6\n"  in Transform()
141       "vsub.f32 q3, q3, q6\n"  in Transform()
142       "vmul.f32 q0, q0, q7\n"  in Transform()
143       "vmul.f32 q1, q1, q7\n"  in Transform()
144       "vmul.f32 q2, q2, q7\n"  in Transform()
145       "vmul.f32 q3, q3, q7\n"  in Transform()
146       "vadd.f32 q0, q0, q4\n"  in Transform()
147       "vadd.f32 q1, q1, q4\n"  in Transform()
148       "vadd.f32 q2, q2, q4\n"  in Transform()
149       "vadd.f32 q3, q3, q4\n"  in Transform()
150       "vmul.f32 q0, q0, q8\n"  in Transform()
151       "vmul.f32 q1, q1, q8\n"  in Transform()
152       "vmul.f32 q2, q2, q8\n"  in Transform()
153       "vmul.f32 q3, q3, q8\n"  in Transform()
154       "vcvt.s32.f32 q0, q0\n"  in Transform()
155       "vcvt.s32.f32 q1, q1\n"  in Transform()
156       "vcvt.s32.f32 q2, q2\n"  in Transform()
157       "vcvt.s32.f32 q3, q3\n"  in Transform()
158       "vqmovn.s32 d0, q0\n"  in Transform()
159       "vqmovn.s32 d1, q1\n"  in Transform()
160       "vqmovn.s32 d4, q2\n"  in Transform()
161       "vqmovn.s32 d5, q3\n"  in Transform()
162       "vqmovun.s16 d0, q0\n"  in Transform()
163       "vqmovun.s16 d1, q2\n"  in Transform()
165       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
166       "pld [%[output]]\n"  in Transform()
168       "bne 1b\n"  in Transform()
174       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
175       "pld [%[input], #64]\n"  in Transform()
176       "vcvt.f32.s32 q0, q0\n"  in Transform()
177       "vsub.f32 q0, q0, q6\n"  in Transform()
178       "vmul.f32 q0, q0, q7\n"  in Transform()
179       "vadd.f32 q0, q0, q4\n"  in Transform()
180       "vmul.f32 q0, q0, q8\n"  in Transform()
181       "vcvt.s32.f32 q0, q0\n"  in Transform()
182       "vqmovn.s32 d0, q0\n"  in Transform()
183       "vqmovun.s16 d0, q0\n"  in Transform()
185       "vst1.8 {d0[0]}, [%[output]]!\n"  in Transform()
186       "pld [%[output]]\n"  in Transform()
213       "vdup.32 q4, %[input_range_min]\n"  in Transform()
214       "vdup.32 q5, %[output_range_min]\n"  in Transform()
215       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
216       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
217       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
218       "vsub.f32 q4, q4, q5\n"  in Transform()
221       "subs %[count], %[count], #2\n"  in Transform()
222       "beq 2f\n"  in Transform()
225       "subs %[count], %[count], #16\n"  in Transform()
228       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
229       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
230       "pld [%[input], #64]\n"  in Transform()
231       "vcvt.f32.s32 q0, q0\n"  in Transform()
232       "vcvt.f32.s32 q1, q1\n"  in Transform()
233       "vcvt.f32.s32 q2, q2\n"  in Transform()
234       "vcvt.f32.s32 q3, q3\n"  in Transform()
235       "vsub.f32 q0, q0, q6\n"  in Transform()
236       "vsub.f32 q1, q1, q6\n"  in Transform()
237       "vsub.f32 q2, q2, q6\n"  in Transform()
238       "vsub.f32 q3, q3, q6\n"  in Transform()
239       "vmul.f32 q0, q0, q7\n"  in Transform()
240       "vmul.f32 q1, q1, q7\n"  in Transform()
241       "vmul.f32 q2, q2, q7\n"  in Transform()
242       "vmul.f32 q3, q3, q7\n"  in Transform()
243       "vadd.f32 q0, q0, q4\n"  in Transform()
244       "vadd.f32 q1, q1, q4\n"  in Transform()
245       "vadd.f32 q2, q2, q4\n"  in Transform()
246       "vadd.f32 q3, q3, q4\n"  in Transform()
247       "vmul.f32 q0, q0, q8\n"  in Transform()
248       "vmul.f32 q1, q1, q8\n"  in Transform()
249       "vmul.f32 q2, q2, q8\n"  in Transform()
250       "vmul.f32 q3, q3, q8\n"  in Transform()
251       "vcvt.s32.f32 q0, q0\n"  in Transform()
252       "vcvt.s32.f32 q1, q1\n"  in Transform()
253       "vcvt.s32.f32 q2, q2\n"  in Transform()
254       "vcvt.s32.f32 q3, q3\n"  in Transform()
255       "vqmovn.s32 d0, q0\n"  in Transform()
256       "vqmovn.s32 d1, q1\n"  in Transform()
257       "vqmovn.s32 d4, q2\n"  in Transform()
258       "vqmovn.s32 d5, q3\n"  in Transform()
259       "vqmovun.s16 d0, q0\n"  in Transform()
260       "vqmovun.s16 d1, q2\n"  in Transform()
262       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
263       "pld [%[output]]\n"  in Transform()
265       "bne 1b\n"  in Transform()
271       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
272       "pld [%[input], #64]\n"  in Transform()
273       "vcvt.f32.s32 q0, q0\n"  in Transform()
274       "vsub.f32 q0, q0, q6\n"  in Transform()
275       "vmul.f32 q0, q0, q7\n"  in Transform()
276       "vadd.f32 q0, q0, q4\n"  in Transform()
277       "vmul.f32 q0, q0, q8\n"  in Transform()
278       "vcvt.s32.f32 q0, q0\n"  in Transform()
279       "vqmovn.s32 d0, q0\n"  in Transform()
280       "vqmovun.s16 d0, q0\n"  in Transform()
282       "vst1.16 {d0[0]}, [%[output]]!\n"  in Transform()
283       "pld [%[output]]\n"  in Transform()
310       "vdup.32 q4, %[input_range_min]\n"  in Transform()
311       "vdup.32 q5, %[output_range_min]\n"  in Transform()
312       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
313       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
314       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
315       "vsub.f32 q4, q4, q5\n"  in Transform()
318       "subs %[count], %[count], #3\n"  in Transform()
319       "beq 2f\n"  in Transform()
322       "subs %[count], %[count], #16\n"  in Transform()
325       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
326       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
327       "pld [%[input], #64]\n"  in Transform()
328       "vcvt.f32.s32 q0, q0\n"  in Transform()
329       "vcvt.f32.s32 q1, q1\n"  in Transform()
330       "vcvt.f32.s32 q2, q2\n"  in Transform()
331       "vcvt.f32.s32 q3, q3\n"  in Transform()
332       "vsub.f32 q0, q0, q6\n"  in Transform()
333       "vsub.f32 q1, q1, q6\n"  in Transform()
334       "vsub.f32 q2, q2, q6\n"  in Transform()
335       "vsub.f32 q3, q3, q6\n"  in Transform()
336       "vmul.f32 q0, q0, q7\n"  in Transform()
337       "vmul.f32 q1, q1, q7\n"  in Transform()
338       "vmul.f32 q2, q2, q7\n"  in Transform()
339       "vmul.f32 q3, q3, q7\n"  in Transform()
340       "vadd.f32 q0, q0, q4\n"  in Transform()
341       "vadd.f32 q1, q1, q4\n"  in Transform()
342       "vadd.f32 q2, q2, q4\n"  in Transform()
343       "vadd.f32 q3, q3, q4\n"  in Transform()
344       "vmul.f32 q0, q0, q8\n"  in Transform()
345       "vmul.f32 q1, q1, q8\n"  in Transform()
346       "vmul.f32 q2, q2, q8\n"  in Transform()
347       "vmul.f32 q3, q3, q8\n"  in Transform()
348       "vcvt.s32.f32 q0, q0\n"  in Transform()
349       "vcvt.s32.f32 q1, q1\n"  in Transform()
350       "vcvt.s32.f32 q2, q2\n"  in Transform()
351       "vcvt.s32.f32 q3, q3\n"  in Transform()
352       "vqmovn.s32 d0, q0\n"  in Transform()
353       "vqmovn.s32 d1, q1\n"  in Transform()
354       "vqmovn.s32 d4, q2\n"  in Transform()
355       "vqmovn.s32 d5, q3\n"  in Transform()
356       "vqmovun.s16 d0, q0\n"  in Transform()
357       "vqmovun.s16 d1, q2\n"  in Transform()
359       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
360       "pld [%[output]]\n"  in Transform()
362       "bne 1b\n"  in Transform()
368       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
369       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
370       "pld [%[input], #64]\n"  in Transform()
371       "vcvt.f32.s32 q0, q0\n"  in Transform()
372       "vsub.f32 q0, q0, q6\n"  in Transform()
373       "vmul.f32 q0, q0, q7\n"  in Transform()
374       "vadd.f32 q0, q0, q4\n"  in Transform()
375       "vmul.f32 q0, q0, q8\n"  in Transform()
376       "vcvt.s32.f32 q0, q0\n"  in Transform()
377       "vqmovn.s32 d0, q0\n"  in Transform()
378       "vqmovun.s16 d0, q0\n"  in Transform()
380       "vst1.16 {d0[0]}, [%[output]]!\n"  in Transform()
381       "vst1.8 {d0[2]}, [%[output]]!\n"  in Transform()
382       "pld [%[output]]\n"  in Transform()
409       "vdup.32 q4, %[input_range_min]\n"  in Transform()
410       "vdup.32 q5, %[output_range_min]\n"  in Transform()
411       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
412       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
413       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
414       "vsub.f32 q4, q4, q5\n"  in Transform()
417       "subs %[count], %[count], #4\n"  in Transform()
418       "beq 2f\n"  in Transform()
421       "subs %[count], %[count], #16\n"  in Transform()
424       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
425       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
426       "pld [%[input], #64]\n"  in Transform()
427       "vcvt.f32.s32 q0, q0\n"  in Transform()
428       "vcvt.f32.s32 q1, q1\n"  in Transform()
429       "vcvt.f32.s32 q2, q2\n"  in Transform()
430       "vcvt.f32.s32 q3, q3\n"  in Transform()
431       "vsub.f32 q0, q0, q6\n"  in Transform()
432       "vsub.f32 q1, q1, q6\n"  in Transform()
433       "vsub.f32 q2, q2, q6\n"  in Transform()
434       "vsub.f32 q3, q3, q6\n"  in Transform()
435       "vmul.f32 q0, q0, q7\n"  in Transform()
436       "vmul.f32 q1, q1, q7\n"  in Transform()
437       "vmul.f32 q2, q2, q7\n"  in Transform()
438       "vmul.f32 q3, q3, q7\n"  in Transform()
439       "vadd.f32 q0, q0, q4\n"  in Transform()
440       "vadd.f32 q1, q1, q4\n"  in Transform()
441       "vadd.f32 q2, q2, q4\n"  in Transform()
442       "vadd.f32 q3, q3, q4\n"  in Transform()
443       "vmul.f32 q0, q0, q8\n"  in Transform()
444       "vmul.f32 q1, q1, q8\n"  in Transform()
445       "vmul.f32 q2, q2, q8\n"  in Transform()
446       "vmul.f32 q3, q3, q8\n"  in Transform()
447       "vcvt.s32.f32 q0, q0\n"  in Transform()
448       "vcvt.s32.f32 q1, q1\n"  in Transform()
449       "vcvt.s32.f32 q2, q2\n"  in Transform()
450       "vcvt.s32.f32 q3, q3\n"  in Transform()
451       "vqmovn.s32 d0, q0\n"  in Transform()
452       "vqmovn.s32 d1, q1\n"  in Transform()
453       "vqmovn.s32 d4, q2\n"  in Transform()
454       "vqmovn.s32 d5, q3\n"  in Transform()
455       "vqmovun.s16 d0, q0\n"  in Transform()
456       "vqmovun.s16 d1, q2\n"  in Transform()
458       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
459       "pld [%[output]]\n"  in Transform()
461       "bne 1b\n"  in Transform()
467       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
468       "pld [%[input], #64]\n"  in Transform()
469       "vcvt.f32.s32 q0, q0\n"  in Transform()
470       "vsub.f32 q0, q0, q6\n"  in Transform()
471       "vmul.f32 q0, q0, q7\n"  in Transform()
472       "vadd.f32 q0, q0, q4\n"  in Transform()
473       "vmul.f32 q0, q0, q8\n"  in Transform()
474       "vcvt.s32.f32 q0, q0\n"  in Transform()
475       "vqmovn.s32 d0, q0\n"  in Transform()
476       "vqmovun.s16 d0, q0\n"  in Transform()
478       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
479       "pld [%[output]]\n"  in Transform()
506       "vdup.32 q4, %[input_range_min]\n"  in Transform()
507       "vdup.32 q5, %[output_range_min]\n"  in Transform()
508       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
509       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
510       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
511       "vsub.f32 q4, q4, q5\n"  in Transform()
514       "subs %[count], %[count], #5\n"  in Transform()
515       "beq 2f\n"  in Transform()
518       "subs %[count], %[count], #16\n"  in Transform()
521       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
522       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
523       "pld [%[input], #64]\n"  in Transform()
524       "vcvt.f32.s32 q0, q0\n"  in Transform()
525       "vcvt.f32.s32 q1, q1\n"  in Transform()
526       "vcvt.f32.s32 q2, q2\n"  in Transform()
527       "vcvt.f32.s32 q3, q3\n"  in Transform()
528       "vsub.f32 q0, q0, q6\n"  in Transform()
529       "vsub.f32 q1, q1, q6\n"  in Transform()
530       "vsub.f32 q2, q2, q6\n"  in Transform()
531       "vsub.f32 q3, q3, q6\n"  in Transform()
532       "vmul.f32 q0, q0, q7\n"  in Transform()
533       "vmul.f32 q1, q1, q7\n"  in Transform()
534       "vmul.f32 q2, q2, q7\n"  in Transform()
535       "vmul.f32 q3, q3, q7\n"  in Transform()
536       "vadd.f32 q0, q0, q4\n"  in Transform()
537       "vadd.f32 q1, q1, q4\n"  in Transform()
538       "vadd.f32 q2, q2, q4\n"  in Transform()
539       "vadd.f32 q3, q3, q4\n"  in Transform()
540       "vmul.f32 q0, q0, q8\n"  in Transform()
541       "vmul.f32 q1, q1, q8\n"  in Transform()
542       "vmul.f32 q2, q2, q8\n"  in Transform()
543       "vmul.f32 q3, q3, q8\n"  in Transform()
544       "vcvt.s32.f32 q0, q0\n"  in Transform()
545       "vcvt.s32.f32 q1, q1\n"  in Transform()
546       "vcvt.s32.f32 q2, q2\n"  in Transform()
547       "vcvt.s32.f32 q3, q3\n"  in Transform()
548       "vqmovn.s32 d0, q0\n"  in Transform()
549       "vqmovn.s32 d1, q1\n"  in Transform()
550       "vqmovn.s32 d4, q2\n"  in Transform()
551       "vqmovn.s32 d5, q3\n"  in Transform()
552       "vqmovun.s16 d0, q0\n"  in Transform()
553       "vqmovun.s16 d1, q2\n"  in Transform()
555       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
556       "pld [%[output]]\n"  in Transform()
558       "bne 1b\n"  in Transform()
564       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
565       "vld1.32 {d2[0]}, [%[input]]!\n"  in Transform()
566       "pld [%[input], #64]\n"  in Transform()
567       "vcvt.f32.s32 q0, q0\n"  in Transform()
568       "vcvt.f32.s32 q1, q1\n"  in Transform()
569       "vsub.f32 q0, q0, q6\n"  in Transform()
570       "vsub.f32 q1, q1, q6\n"  in Transform()
571       "vmul.f32 q0, q0, q7\n"  in Transform()
572       "vmul.f32 q1, q1, q7\n"  in Transform()
573       "vadd.f32 q0, q0, q4\n"  in Transform()
574       "vadd.f32 q1, q1, q4\n"  in Transform()
575       "vmul.f32 q0, q0, q8\n"  in Transform()
576       "vmul.f32 q1, q1, q8\n"  in Transform()
577       "vcvt.s32.f32 q0, q0\n"  in Transform()
578       "vcvt.s32.f32 q1, q1\n"  in Transform()
579       "vqmovn.s32 d0, q0\n"  in Transform()
580       "vqmovn.s32 d1, q1\n"  in Transform()
581       "vqmovun.s16 d0, q0\n"  in Transform()
583       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
584       "vst1.8 {d0[4]}, [%[output]]!\n"  in Transform()
585       "pld [%[output]]\n"  in Transform()
612       "vdup.32 q4, %[input_range_min]\n"  in Transform()
613       "vdup.32 q5, %[output_range_min]\n"  in Transform()
614       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
615       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
616       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
617       "vsub.f32 q4, q4, q5\n"  in Transform()
620       "subs %[count], %[count], #6\n"  in Transform()
621       "beq 2f\n"  in Transform()
624       "subs %[count], %[count], #16\n"  in Transform()
627       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
628       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
629       "pld [%[input], #64]\n"  in Transform()
630       "vcvt.f32.s32 q0, q0\n"  in Transform()
631       "vcvt.f32.s32 q1, q1\n"  in Transform()
632       "vcvt.f32.s32 q2, q2\n"  in Transform()
633       "vcvt.f32.s32 q3, q3\n"  in Transform()
634       "vsub.f32 q0, q0, q6\n"  in Transform()
635       "vsub.f32 q1, q1, q6\n"  in Transform()
636       "vsub.f32 q2, q2, q6\n"  in Transform()
637       "vsub.f32 q3, q3, q6\n"  in Transform()
638       "vmul.f32 q0, q0, q7\n"  in Transform()
639       "vmul.f32 q1, q1, q7\n"  in Transform()
640       "vmul.f32 q2, q2, q7\n"  in Transform()
641       "vmul.f32 q3, q3, q7\n"  in Transform()
642       "vadd.f32 q0, q0, q4\n"  in Transform()
643       "vadd.f32 q1, q1, q4\n"  in Transform()
644       "vadd.f32 q2, q2, q4\n"  in Transform()
645       "vadd.f32 q3, q3, q4\n"  in Transform()
646       "vmul.f32 q0, q0, q8\n"  in Transform()
647       "vmul.f32 q1, q1, q8\n"  in Transform()
648       "vmul.f32 q2, q2, q8\n"  in Transform()
649       "vmul.f32 q3, q3, q8\n"  in Transform()
650       "vcvt.s32.f32 q0, q0\n"  in Transform()
651       "vcvt.s32.f32 q1, q1\n"  in Transform()
652       "vcvt.s32.f32 q2, q2\n"  in Transform()
653       "vcvt.s32.f32 q3, q3\n"  in Transform()
654       "vqmovn.s32 d0, q0\n"  in Transform()
655       "vqmovn.s32 d1, q1\n"  in Transform()
656       "vqmovn.s32 d4, q2\n"  in Transform()
657       "vqmovn.s32 d5, q3\n"  in Transform()
658       "vqmovun.s16 d0, q0\n"  in Transform()
659       "vqmovun.s16 d1, q2\n"  in Transform()
661       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
662       "pld [%[output]]\n"  in Transform()
664       "bne 1b\n"  in Transform()
670       "vld1.32 {d0, d1, d2}, [%[input]]!\n"  in Transform()
671       "pld [%[input], #64]\n"  in Transform()
672       "vcvt.f32.s32 q0, q0\n"  in Transform()
673       "vcvt.f32.s32 q1, q1\n"  in Transform()
674       "vsub.f32 q0, q0, q6\n"  in Transform()
675       "vsub.f32 q1, q1, q6\n"  in Transform()
676       "vmul.f32 q0, q0, q7\n"  in Transform()
677       "vmul.f32 q1, q1, q7\n"  in Transform()
678       "vadd.f32 q0, q0, q4\n"  in Transform()
679       "vadd.f32 q1, q1, q4\n"  in Transform()
680       "vmul.f32 q0, q0, q8\n"  in Transform()
681       "vmul.f32 q1, q1, q8\n"  in Transform()
682       "vcvt.s32.f32 q0, q0\n"  in Transform()
683       "vcvt.s32.f32 q1, q1\n"  in Transform()
684       "vqmovn.s32 d0, q0\n"  in Transform()
685       "vqmovn.s32 d1, q1\n"  in Transform()
686       "vqmovun.s16 d0, q0\n"  in Transform()
688       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
689       "vst1.16 {d0[2]}, [%[output]]!\n"  in Transform()
690       "pld [%[output]]\n"  in Transform()
717       "vdup.32 q4, %[input_range_min]\n"  in Transform()
718       "vdup.32 q5, %[output_range_min]\n"  in Transform()
719       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
720       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
721       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
722       "vsub.f32 q4, q4, q5\n"  in Transform()
725       "subs %[count], %[count], #7\n"  in Transform()
726       "beq 2f\n"  in Transform()
729       "subs %[count], %[count], #16\n"  in Transform()
732       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
733       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
734       "pld [%[input], #64]\n"  in Transform()
735       "vcvt.f32.s32 q0, q0\n"  in Transform()
736       "vcvt.f32.s32 q1, q1\n"  in Transform()
737       "vcvt.f32.s32 q2, q2\n"  in Transform()
738       "vcvt.f32.s32 q3, q3\n"  in Transform()
739       "vsub.f32 q0, q0, q6\n"  in Transform()
740       "vsub.f32 q1, q1, q6\n"  in Transform()
741       "vsub.f32 q2, q2, q6\n"  in Transform()
742       "vsub.f32 q3, q3, q6\n"  in Transform()
743       "vmul.f32 q0, q0, q7\n"  in Transform()
744       "vmul.f32 q1, q1, q7\n"  in Transform()
745       "vmul.f32 q2, q2, q7\n"  in Transform()
746       "vmul.f32 q3, q3, q7\n"  in Transform()
747       "vadd.f32 q0, q0, q4\n"  in Transform()
748       "vadd.f32 q1, q1, q4\n"  in Transform()
749       "vadd.f32 q2, q2, q4\n"  in Transform()
750       "vadd.f32 q3, q3, q4\n"  in Transform()
751       "vmul.f32 q0, q0, q8\n"  in Transform()
752       "vmul.f32 q1, q1, q8\n"  in Transform()
753       "vmul.f32 q2, q2, q8\n"  in Transform()
754       "vmul.f32 q3, q3, q8\n"  in Transform()
755       "vcvt.s32.f32 q0, q0\n"  in Transform()
756       "vcvt.s32.f32 q1, q1\n"  in Transform()
757       "vcvt.s32.f32 q2, q2\n"  in Transform()
758       "vcvt.s32.f32 q3, q3\n"  in Transform()
759       "vqmovn.s32 d0, q0\n"  in Transform()
760       "vqmovn.s32 d1, q1\n"  in Transform()
761       "vqmovn.s32 d4, q2\n"  in Transform()
762       "vqmovn.s32 d5, q3\n"  in Transform()
763       "vqmovun.s16 d0, q0\n"  in Transform()
764       "vqmovun.s16 d1, q2\n"  in Transform()
766       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
767       "pld [%[output]]\n"  in Transform()
769       "bne 1b\n"  in Transform()
775       "vld1.32 {d0, d1, d2}, [%[input]]!\n"  in Transform()
776       "vld1.32 {d3[0]}, [%[input]]!\n"  in Transform()
777       "pld [%[input], #64]\n"  in Transform()
778       "vcvt.f32.s32 q0, q0\n"  in Transform()
779       "vcvt.f32.s32 q1, q1\n"  in Transform()
780       "vsub.f32 q0, q0, q6\n"  in Transform()
781       "vsub.f32 q1, q1, q6\n"  in Transform()
782       "vmul.f32 q0, q0, q7\n"  in Transform()
783       "vmul.f32 q1, q1, q7\n"  in Transform()
784       "vadd.f32 q0, q0, q4\n"  in Transform()
785       "vadd.f32 q1, q1, q4\n"  in Transform()
786       "vmul.f32 q0, q0, q8\n"  in Transform()
787       "vmul.f32 q1, q1, q8\n"  in Transform()
788       "vcvt.s32.f32 q0, q0\n"  in Transform()
789       "vcvt.s32.f32 q1, q1\n"  in Transform()
790       "vqmovn.s32 d0, q0\n"  in Transform()
791       "vqmovn.s32 d1, q1\n"  in Transform()
792       "vqmovun.s16 d0, q0\n"  in Transform()
794       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
795       "vst1.16 {d0[2]}, [%[output]]!\n"  in Transform()
796       "vst1.8 {d0[6]}, [%[output]]!\n"  in Transform()
797       "pld [%[output]]\n"  in Transform()
824       "vdup.32 q4, %[input_range_min]\n"  in Transform()
825       "vdup.32 q5, %[output_range_min]\n"  in Transform()
826       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
827       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
828       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
829       "vsub.f32 q4, q4, q5\n"  in Transform()
832       "subs %[count], %[count], #8\n"  in Transform()
833       "beq 2f\n"  in Transform()
836       "subs %[count], %[count], #16\n"  in Transform()
839       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
840       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
841       "pld [%[input], #64]\n"  in Transform()
842       "vcvt.f32.s32 q0, q0\n"  in Transform()
843       "vcvt.f32.s32 q1, q1\n"  in Transform()
844       "vcvt.f32.s32 q2, q2\n"  in Transform()
845       "vcvt.f32.s32 q3, q3\n"  in Transform()
846       "vsub.f32 q0, q0, q6\n"  in Transform()
847       "vsub.f32 q1, q1, q6\n"  in Transform()
848       "vsub.f32 q2, q2, q6\n"  in Transform()
849       "vsub.f32 q3, q3, q6\n"  in Transform()
850       "vmul.f32 q0, q0, q7\n"  in Transform()
851       "vmul.f32 q1, q1, q7\n"  in Transform()
852       "vmul.f32 q2, q2, q7\n"  in Transform()
853       "vmul.f32 q3, q3, q7\n"  in Transform()
854       "vadd.f32 q0, q0, q4\n"  in Transform()
855       "vadd.f32 q1, q1, q4\n"  in Transform()
856       "vadd.f32 q2, q2, q4\n"  in Transform()
857       "vadd.f32 q3, q3, q4\n"  in Transform()
858       "vmul.f32 q0, q0, q8\n"  in Transform()
859       "vmul.f32 q1, q1, q8\n"  in Transform()
860       "vmul.f32 q2, q2, q8\n"  in Transform()
861       "vmul.f32 q3, q3, q8\n"  in Transform()
862       "vcvt.s32.f32 q0, q0\n"  in Transform()
863       "vcvt.s32.f32 q1, q1\n"  in Transform()
864       "vcvt.s32.f32 q2, q2\n"  in Transform()
865       "vcvt.s32.f32 q3, q3\n"  in Transform()
866       "vqmovn.s32 d0, q0\n"  in Transform()
867       "vqmovn.s32 d1, q1\n"  in Transform()
868       "vqmovn.s32 d4, q2\n"  in Transform()
869       "vqmovn.s32 d5, q3\n"  in Transform()
870       "vqmovun.s16 d0, q0\n"  in Transform()
871       "vqmovun.s16 d1, q2\n"  in Transform()
873       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
874       "pld [%[output]]\n"  in Transform()
876       "bne 1b\n"  in Transform()
882       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
883       "pld [%[input], #64]\n"  in Transform()
884       "vcvt.f32.s32 q0, q0\n"  in Transform()
885       "vcvt.f32.s32 q1, q1\n"  in Transform()
886       "vsub.f32 q0, q0, q6\n"  in Transform()
887       "vsub.f32 q1, q1, q6\n"  in Transform()
888       "vmul.f32 q0, q0, q7\n"  in Transform()
889       "vmul.f32 q1, q1, q7\n"  in Transform()
890       "vadd.f32 q0, q0, q4\n"  in Transform()
891       "vadd.f32 q1, q1, q4\n"  in Transform()
892       "vmul.f32 q0, q0, q8\n"  in Transform()
893       "vmul.f32 q1, q1, q8\n"  in Transform()
894       "vcvt.s32.f32 q0, q0\n"  in Transform()
895       "vcvt.s32.f32 q1, q1\n"  in Transform()
896       "vqmovn.s32 d0, q0\n"  in Transform()
897       "vqmovn.s32 d1, q1\n"  in Transform()
898       "vqmovun.s16 d0, q0\n"  in Transform()
900       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
901       "pld [%[output]]\n"  in Transform()
928       "vdup.32 q4, %[input_range_min]\n"  in Transform()
929       "vdup.32 q5, %[output_range_min]\n"  in Transform()
930       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
931       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
932       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
933       "vsub.f32 q4, q4, q5\n"  in Transform()
936       "subs %[count], %[count], #9\n"  in Transform()
937       "beq 2f\n"  in Transform()
940       "subs %[count], %[count], #16\n"  in Transform()
943       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
944       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
945       "pld [%[input], #64]\n"  in Transform()
946       "vcvt.f32.s32 q0, q0\n"  in Transform()
947       "vcvt.f32.s32 q1, q1\n"  in Transform()
948       "vcvt.f32.s32 q2, q2\n"  in Transform()
949       "vcvt.f32.s32 q3, q3\n"  in Transform()
950       "vsub.f32 q0, q0, q6\n"  in Transform()
951       "vsub.f32 q1, q1, q6\n"  in Transform()
952       "vsub.f32 q2, q2, q6\n"  in Transform()
953       "vsub.f32 q3, q3, q6\n"  in Transform()
954       "vmul.f32 q0, q0, q7\n"  in Transform()
955       "vmul.f32 q1, q1, q7\n"  in Transform()
956       "vmul.f32 q2, q2, q7\n"  in Transform()
957       "vmul.f32 q3, q3, q7\n"  in Transform()
958       "vadd.f32 q0, q0, q4\n"  in Transform()
959       "vadd.f32 q1, q1, q4\n"  in Transform()
960       "vadd.f32 q2, q2, q4\n"  in Transform()
961       "vadd.f32 q3, q3, q4\n"  in Transform()
962       "vmul.f32 q0, q0, q8\n"  in Transform()
963       "vmul.f32 q1, q1, q8\n"  in Transform()
964       "vmul.f32 q2, q2, q8\n"  in Transform()
965       "vmul.f32 q3, q3, q8\n"  in Transform()
966       "vcvt.s32.f32 q0, q0\n"  in Transform()
967       "vcvt.s32.f32 q1, q1\n"  in Transform()
968       "vcvt.s32.f32 q2, q2\n"  in Transform()
969       "vcvt.s32.f32 q3, q3\n"  in Transform()
970       "vqmovn.s32 d0, q0\n"  in Transform()
971       "vqmovn.s32 d1, q1\n"  in Transform()
972       "vqmovn.s32 d4, q2\n"  in Transform()
973       "vqmovn.s32 d5, q3\n"  in Transform()
974       "vqmovun.s16 d0, q0\n"  in Transform()
975       "vqmovun.s16 d1, q2\n"  in Transform()
977       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
978       "pld [%[output]]\n"  in Transform()
980       "bne 1b\n"  in Transform()
986       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
987       "vld1.32 {d4[0]}, [%[input]]!\n"  in Transform()
988       "pld [%[input], #64]\n"  in Transform()
989       "vcvt.f32.s32 q0, q0\n"  in Transform()
990       "vcvt.f32.s32 q1, q1\n"  in Transform()
991       "vcvt.f32.s32 q2, q2\n"  in Transform()
992       "vsub.f32 q0, q0, q6\n"  in Transform()
993       "vsub.f32 q1, q1, q6\n"  in Transform()
994       "vsub.f32 q2, q2, q6\n"  in Transform()
995       "vmul.f32 q0, q0, q7\n"  in Transform()
996       "vmul.f32 q1, q1, q7\n"  in Transform()
997       "vmul.f32 q2, q2, q7\n"  in Transform()
998       "vadd.f32 q0, q0, q4\n"  in Transform()
999       "vadd.f32 q1, q1, q4\n"  in Transform()
1000       "vadd.f32 q2, q2, q4\n"  in Transform()
1001       "vmul.f32 q0, q0, q8\n"  in Transform()
1002       "vmul.f32 q1, q1, q8\n"  in Transform()
1003       "vmul.f32 q2, q2, q8\n"  in Transform()
1004       "vcvt.s32.f32 q0, q0\n"  in Transform()
1005       "vcvt.s32.f32 q1, q1\n"  in Transform()
1006       "vcvt.s32.f32 q2, q2\n"  in Transform()
1007       "vqmovn.s32 d0, q0\n"  in Transform()
1008       "vqmovn.s32 d1, q1\n"  in Transform()
1009       "vqmovn.s32 d4, q2\n"  in Transform()
1010       "vqmovun.s16 d0, q0\n"  in Transform()
1011       "vqmovun.s16 d1, q2\n"  in Transform()
1013       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
1014       "vst1.8 {d1[0]}, [%[output]]!\n"  in Transform()
1015       "pld [%[output]]\n"  in Transform()
1042       "vdup.32 q4, %[input_range_min]\n"  in Transform()
1043       "vdup.32 q5, %[output_range_min]\n"  in Transform()
1044       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
1045       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
1046       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
1047       "vsub.f32 q4, q4, q5\n"  in Transform()
1050       "subs %[count], %[count], #10\n"  in Transform()
1051       "beq 2f\n"  in Transform()
1054       "subs %[count], %[count], #16\n"  in Transform()
1057       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1058       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1059       "pld [%[input], #64]\n"  in Transform()
1060       "vcvt.f32.s32 q0, q0\n"  in Transform()
1061       "vcvt.f32.s32 q1, q1\n"  in Transform()
1062       "vcvt.f32.s32 q2, q2\n"  in Transform()
1063       "vcvt.f32.s32 q3, q3\n"  in Transform()
1064       "vsub.f32 q0, q0, q6\n"  in Transform()
1065       "vsub.f32 q1, q1, q6\n"  in Transform()
1066       "vsub.f32 q2, q2, q6\n"  in Transform()
1067       "vsub.f32 q3, q3, q6\n"  in Transform()
1068       "vmul.f32 q0, q0, q7\n"  in Transform()
1069       "vmul.f32 q1, q1, q7\n"  in Transform()
1070       "vmul.f32 q2, q2, q7\n"  in Transform()
1071       "vmul.f32 q3, q3, q7\n"  in Transform()
1072       "vadd.f32 q0, q0, q4\n"  in Transform()
1073       "vadd.f32 q1, q1, q4\n"  in Transform()
1074       "vadd.f32 q2, q2, q4\n"  in Transform()
1075       "vadd.f32 q3, q3, q4\n"  in Transform()
1076       "vmul.f32 q0, q0, q8\n"  in Transform()
1077       "vmul.f32 q1, q1, q8\n"  in Transform()
1078       "vmul.f32 q2, q2, q8\n"  in Transform()
1079       "vmul.f32 q3, q3, q8\n"  in Transform()
1080       "vcvt.s32.f32 q0, q0\n"  in Transform()
1081       "vcvt.s32.f32 q1, q1\n"  in Transform()
1082       "vcvt.s32.f32 q2, q2\n"  in Transform()
1083       "vcvt.s32.f32 q3, q3\n"  in Transform()
1084       "vqmovn.s32 d0, q0\n"  in Transform()
1085       "vqmovn.s32 d1, q1\n"  in Transform()
1086       "vqmovn.s32 d4, q2\n"  in Transform()
1087       "vqmovn.s32 d5, q3\n"  in Transform()
1088       "vqmovun.s16 d0, q0\n"  in Transform()
1089       "vqmovun.s16 d1, q2\n"  in Transform()
1091       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
1092       "pld [%[output]]\n"  in Transform()
1094       "bne 1b\n"  in Transform()
1100       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1101       "vld1.32 {d4}, [%[input]]!\n"  in Transform()
1102       "pld [%[input], #64]\n"  in Transform()
1103       "vcvt.f32.s32 q0, q0\n"  in Transform()
1104       "vcvt.f32.s32 q1, q1\n"  in Transform()
1105       "vcvt.f32.s32 q2, q2\n"  in Transform()
1106       "vsub.f32 q0, q0, q6\n"  in Transform()
1107       "vsub.f32 q1, q1, q6\n"  in Transform()
1108       "vsub.f32 q2, q2, q6\n"  in Transform()
1109       "vmul.f32 q0, q0, q7\n"  in Transform()
1110       "vmul.f32 q1, q1, q7\n"  in Transform()
1111       "vmul.f32 q2, q2, q7\n"  in Transform()
1112       "vadd.f32 q0, q0, q4\n"  in Transform()
1113       "vadd.f32 q1, q1, q4\n"  in Transform()
1114       "vadd.f32 q2, q2, q4\n"  in Transform()
1115       "vmul.f32 q0, q0, q8\n"  in Transform()
1116       "vmul.f32 q1, q1, q8\n"  in Transform()
1117       "vmul.f32 q2, q2, q8\n"  in Transform()
1118       "vcvt.s32.f32 q0, q0\n"  in Transform()
1119       "vcvt.s32.f32 q1, q1\n"  in Transform()
1120       "vcvt.s32.f32 q2, q2\n"  in Transform()
1121       "vqmovn.s32 d0, q0\n"  in Transform()
1122       "vqmovn.s32 d1, q1\n"  in Transform()
1123       "vqmovn.s32 d4, q2\n"  in Transform()
1124       "vqmovun.s16 d0, q0\n"  in Transform()
1125       "vqmovun.s16 d1, q2\n"  in Transform()
1127       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
1128       "vst1.16 {d1[0]}, [%[output]]!\n"  in Transform()
1129       "pld [%[output]]\n"  in Transform()
1156       "vdup.32 q4, %[input_range_min]\n"  in Transform()
1157       "vdup.32 q5, %[output_range_min]\n"  in Transform()
1158       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
1159       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
1160       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
1161       "vsub.f32 q4, q4, q5\n"  in Transform()
1164       "subs %[count], %[count], #11\n"  in Transform()
1165       "beq 2f\n"  in Transform()
1168       "subs %[count], %[count], #16\n"  in Transform()
1171       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1172       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1173       "pld [%[input], #64]\n"  in Transform()
1174       "vcvt.f32.s32 q0, q0\n"  in Transform()
1175       "vcvt.f32.s32 q1, q1\n"  in Transform()
1176       "vcvt.f32.s32 q2, q2\n"  in Transform()
1177       "vcvt.f32.s32 q3, q3\n"  in Transform()
1178       "vsub.f32 q0, q0, q6\n"  in Transform()
1179       "vsub.f32 q1, q1, q6\n"  in Transform()
1180       "vsub.f32 q2, q2, q6\n"  in Transform()
1181       "vsub.f32 q3, q3, q6\n"  in Transform()
1182       "vmul.f32 q0, q0, q7\n"  in Transform()
1183       "vmul.f32 q1, q1, q7\n"  in Transform()
1184       "vmul.f32 q2, q2, q7\n"  in Transform()
1185       "vmul.f32 q3, q3, q7\n"  in Transform()
1186       "vadd.f32 q0, q0, q4\n"  in Transform()
1187       "vadd.f32 q1, q1, q4\n"  in Transform()
1188       "vadd.f32 q2, q2, q4\n"  in Transform()
1189       "vadd.f32 q3, q3, q4\n"  in Transform()
1190       "vmul.f32 q0, q0, q8\n"  in Transform()
1191       "vmul.f32 q1, q1, q8\n"  in Transform()
1192       "vmul.f32 q2, q2, q8\n"  in Transform()
1193       "vmul.f32 q3, q3, q8\n"  in Transform()
1194       "vcvt.s32.f32 q0, q0\n"  in Transform()
1195       "vcvt.s32.f32 q1, q1\n"  in Transform()
1196       "vcvt.s32.f32 q2, q2\n"  in Transform()
1197       "vcvt.s32.f32 q3, q3\n"  in Transform()
1198       "vqmovn.s32 d0, q0\n"  in Transform()
1199       "vqmovn.s32 d1, q1\n"  in Transform()
1200       "vqmovn.s32 d4, q2\n"  in Transform()
1201       "vqmovn.s32 d5, q3\n"  in Transform()
1202       "vqmovun.s16 d0, q0\n"  in Transform()
1203       "vqmovun.s16 d1, q2\n"  in Transform()
1205       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
1206       "pld [%[output]]\n"  in Transform()
1208       "bne 1b\n"  in Transform()
1214       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1215       "vld1.32 {d4}, [%[input]]!\n"  in Transform()
1216       "vld1.32 {d5[0]}, [%[input]]!\n"  in Transform()
1217       "pld [%[input], #64]\n"  in Transform()
1218       "vcvt.f32.s32 q0, q0\n"  in Transform()
1219       "vcvt.f32.s32 q1, q1\n"  in Transform()
1220       "vcvt.f32.s32 q2, q2\n"  in Transform()
1221       "vsub.f32 q0, q0, q6\n"  in Transform()
1222       "vsub.f32 q1, q1, q6\n"  in Transform()
1223       "vsub.f32 q2, q2, q6\n"  in Transform()
1224       "vmul.f32 q0, q0, q7\n"  in Transform()
1225       "vmul.f32 q1, q1, q7\n"  in Transform()
1226       "vmul.f32 q2, q2, q7\n"  in Transform()
1227       "vadd.f32 q0, q0, q4\n"  in Transform()
1228       "vadd.f32 q1, q1, q4\n"  in Transform()
1229       "vadd.f32 q2, q2, q4\n"  in Transform()
1230       "vmul.f32 q0, q0, q8\n"  in Transform()
1231       "vmul.f32 q1, q1, q8\n"  in Transform()
1232       "vmul.f32 q2, q2, q8\n"  in Transform()
1233       "vcvt.s32.f32 q0, q0\n"  in Transform()
1234       "vcvt.s32.f32 q1, q1\n"  in Transform()
1235       "vcvt.s32.f32 q2, q2\n"  in Transform()
1236       "vqmovn.s32 d0, q0\n"  in Transform()
1237       "vqmovn.s32 d1, q1\n"  in Transform()
1238       "vqmovn.s32 d4, q2\n"  in Transform()
1239       "vqmovun.s16 d0, q0\n"  in Transform()
1240       "vqmovun.s16 d1, q2\n"  in Transform()
1242       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
1243       "vst1.16 {d1[0]}, [%[output]]!\n"  in Transform()
1244       "vst1.8 {d1[2]}, [%[output]]!\n"  in Transform()
1245       "pld [%[output]]\n"  in Transform()
1272       "vdup.32 q4, %[input_range_min]\n"  in Transform()
1273       "vdup.32 q5, %[output_range_min]\n"  in Transform()
1274       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
1275       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
1276       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
1277       "vsub.f32 q4, q4, q5\n"  in Transform()
1280       "subs %[count], %[count], #12\n"  in Transform()
1281       "beq 2f\n"  in Transform()
1284       "subs %[count], %[count], #16\n"  in Transform()
1287       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1288       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1289       "pld [%[input], #64]\n"  in Transform()
1290       "vcvt.f32.s32 q0, q0\n"  in Transform()
1291       "vcvt.f32.s32 q1, q1\n"  in Transform()
1292       "vcvt.f32.s32 q2, q2\n"  in Transform()
1293       "vcvt.f32.s32 q3, q3\n"  in Transform()
1294       "vsub.f32 q0, q0, q6\n"  in Transform()
1295       "vsub.f32 q1, q1, q6\n"  in Transform()
1296       "vsub.f32 q2, q2, q6\n"  in Transform()
1297       "vsub.f32 q3, q3, q6\n"  in Transform()
1298       "vmul.f32 q0, q0, q7\n"  in Transform()
1299       "vmul.f32 q1, q1, q7\n"  in Transform()
1300       "vmul.f32 q2, q2, q7\n"  in Transform()
1301       "vmul.f32 q3, q3, q7\n"  in Transform()
1302       "vadd.f32 q0, q0, q4\n"  in Transform()
1303       "vadd.f32 q1, q1, q4\n"  in Transform()
1304       "vadd.f32 q2, q2, q4\n"  in Transform()
1305       "vadd.f32 q3, q3, q4\n"  in Transform()
1306       "vmul.f32 q0, q0, q8\n"  in Transform()
1307       "vmul.f32 q1, q1, q8\n"  in Transform()
1308       "vmul.f32 q2, q2, q8\n"  in Transform()
1309       "vmul.f32 q3, q3, q8\n"  in Transform()
1310       "vcvt.s32.f32 q0, q0\n"  in Transform()
1311       "vcvt.s32.f32 q1, q1\n"  in Transform()
1312       "vcvt.s32.f32 q2, q2\n"  in Transform()
1313       "vcvt.s32.f32 q3, q3\n"  in Transform()
1314       "vqmovn.s32 d0, q0\n"  in Transform()
1315       "vqmovn.s32 d1, q1\n"  in Transform()
1316       "vqmovn.s32 d4, q2\n"  in Transform()
1317       "vqmovn.s32 d5, q3\n"  in Transform()
1318       "vqmovun.s16 d0, q0\n"  in Transform()
1319       "vqmovun.s16 d1, q2\n"  in Transform()
1321       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
1322       "pld [%[output]]\n"  in Transform()
1324       "bne 1b\n"  in Transform()
1330       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1331       "vld1.32 {d4, d5}, [%[input]]!\n"  in Transform()
1332       "pld [%[input], #64]\n"  in Transform()
1333       "vcvt.f32.s32 q0, q0\n"  in Transform()
1334       "vcvt.f32.s32 q1, q1\n"  in Transform()
1335       "vcvt.f32.s32 q2, q2\n"  in Transform()
1336       "vsub.f32 q0, q0, q6\n"  in Transform()
1337       "vsub.f32 q1, q1, q6\n"  in Transform()
1338       "vsub.f32 q2, q2, q6\n"  in Transform()
1339       "vmul.f32 q0, q0, q7\n"  in Transform()
1340       "vmul.f32 q1, q1, q7\n"  in Transform()
1341       "vmul.f32 q2, q2, q7\n"  in Transform()
1342       "vadd.f32 q0, q0, q4\n"  in Transform()
1343       "vadd.f32 q1, q1, q4\n"  in Transform()
1344       "vadd.f32 q2, q2, q4\n"  in Transform()
1345       "vmul.f32 q0, q0, q8\n"  in Transform()
1346       "vmul.f32 q1, q1, q8\n"  in Transform()
1347       "vmul.f32 q2, q2, q8\n"  in Transform()
1348       "vcvt.s32.f32 q0, q0\n"  in Transform()
1349       "vcvt.s32.f32 q1, q1\n"  in Transform()
1350       "vcvt.s32.f32 q2, q2\n"  in Transform()
1351       "vqmovn.s32 d0, q0\n"  in Transform()
1352       "vqmovn.s32 d1, q1\n"  in Transform()
1353       "vqmovn.s32 d4, q2\n"  in Transform()
1354       "vqmovun.s16 d0, q0\n"  in Transform()
1355       "vqmovun.s16 d1, q2\n"  in Transform()
1357       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
1358       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
1359       "pld [%[output]]\n"  in Transform()
1386       "vdup.32 q4, %[input_range_min]\n"  in Transform()
1387       "vdup.32 q5, %[output_range_min]\n"  in Transform()
1388       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
1389       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
1390       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
1391       "vsub.f32 q4, q4, q5\n"  in Transform()
1394       "subs %[count], %[count], #13\n"  in Transform()
1395       "beq 2f\n"  in Transform()
1398       "subs %[count], %[count], #16\n"  in Transform()
1401       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1402       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1403       "pld [%[input], #64]\n"  in Transform()
1404       "vcvt.f32.s32 q0, q0\n"  in Transform()
1405       "vcvt.f32.s32 q1, q1\n"  in Transform()
1406       "vcvt.f32.s32 q2, q2\n"  in Transform()
1407       "vcvt.f32.s32 q3, q3\n"  in Transform()
1408       "vsub.f32 q0, q0, q6\n"  in Transform()
1409       "vsub.f32 q1, q1, q6\n"  in Transform()
1410       "vsub.f32 q2, q2, q6\n"  in Transform()
1411       "vsub.f32 q3, q3, q6\n"  in Transform()
1412       "vmul.f32 q0, q0, q7\n"  in Transform()
1413       "vmul.f32 q1, q1, q7\n"  in Transform()
1414       "vmul.f32 q2, q2, q7\n"  in Transform()
1415       "vmul.f32 q3, q3, q7\n"  in Transform()
1416       "vadd.f32 q0, q0, q4\n"  in Transform()
1417       "vadd.f32 q1, q1, q4\n"  in Transform()
1418       "vadd.f32 q2, q2, q4\n"  in Transform()
1419       "vadd.f32 q3, q3, q4\n"  in Transform()
1420       "vmul.f32 q0, q0, q8\n"  in Transform()
1421       "vmul.f32 q1, q1, q8\n"  in Transform()
1422       "vmul.f32 q2, q2, q8\n"  in Transform()
1423       "vmul.f32 q3, q3, q8\n"  in Transform()
1424       "vcvt.s32.f32 q0, q0\n"  in Transform()
1425       "vcvt.s32.f32 q1, q1\n"  in Transform()
1426       "vcvt.s32.f32 q2, q2\n"  in Transform()
1427       "vcvt.s32.f32 q3, q3\n"  in Transform()
1428       "vqmovn.s32 d0, q0\n"  in Transform()
1429       "vqmovn.s32 d1, q1\n"  in Transform()
1430       "vqmovn.s32 d4, q2\n"  in Transform()
1431       "vqmovn.s32 d5, q3\n"  in Transform()
1432       "vqmovun.s16 d0, q0\n"  in Transform()
1433       "vqmovun.s16 d1, q2\n"  in Transform()
1435       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
1436       "pld [%[output]]\n"  in Transform()
1438       "bne 1b\n"  in Transform()
1444       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1445       "vld1.32 {d4, d5}, [%[input]]!\n"  in Transform()
1446       "vld1.32 {d6[0]}, [%[input]]!\n"  in Transform()
1447       "pld [%[input], #64]\n"  in Transform()
1448       "vcvt.f32.s32 q0, q0\n"  in Transform()
1449       "vcvt.f32.s32 q1, q1\n"  in Transform()
1450       "vcvt.f32.s32 q2, q2\n"  in Transform()
1451       "vcvt.f32.s32 q3, q3\n"  in Transform()
1452       "vsub.f32 q0, q0, q6\n"  in Transform()
1453       "vsub.f32 q1, q1, q6\n"  in Transform()
1454       "vsub.f32 q2, q2, q6\n"  in Transform()
1455       "vsub.f32 q3, q3, q6\n"  in Transform()
1456       "vmul.f32 q0, q0, q7\n"  in Transform()
1457       "vmul.f32 q1, q1, q7\n"  in Transform()
1458       "vmul.f32 q2, q2, q7\n"  in Transform()
1459       "vmul.f32 q3, q3, q7\n"  in Transform()
1460       "vadd.f32 q0, q0, q4\n"  in Transform()
1461       "vadd.f32 q1, q1, q4\n"  in Transform()
1462       "vadd.f32 q2, q2, q4\n"  in Transform()
1463       "vadd.f32 q3, q3, q4\n"  in Transform()
1464       "vmul.f32 q0, q0, q8\n"  in Transform()
1465       "vmul.f32 q1, q1, q8\n"  in Transform()
1466       "vmul.f32 q2, q2, q8\n"  in Transform()
1467       "vmul.f32 q3, q3, q8\n"  in Transform()
1468       "vcvt.s32.f32 q0, q0\n"  in Transform()
1469       "vcvt.s32.f32 q1, q1\n"  in Transform()
1470       "vcvt.s32.f32 q2, q2\n"  in Transform()
1471       "vcvt.s32.f32 q3, q3\n"  in Transform()
1472       "vqmovn.s32 d0, q0\n"  in Transform()
1473       "vqmovn.s32 d1, q1\n"  in Transform()
1474       "vqmovn.s32 d4, q2\n"  in Transform()
1475       "vqmovn.s32 d5, q3\n"  in Transform()
1476       "vqmovun.s16 d0, q0\n"  in Transform()
1477       "vqmovun.s16 d1, q2\n"  in Transform()
1479       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
1480       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
1481       "vst1.8 {d1[4]}, [%[output]]!\n"  in Transform()
1482       "pld [%[output]]\n"  in Transform()
1509       "vdup.32 q4, %[input_range_min]\n"  in Transform()
1510       "vdup.32 q5, %[output_range_min]\n"  in Transform()
1511       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
1512       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
1513       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
1514       "vsub.f32 q4, q4, q5\n"  in Transform()
1517       "subs %[count], %[count], #14\n"  in Transform()
1518       "beq 2f\n"  in Transform()
1521       "subs %[count], %[count], #16\n"  in Transform()
1524       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1525       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1526       "pld [%[input], #64]\n"  in Transform()
1527       "vcvt.f32.s32 q0, q0\n"  in Transform()
1528       "vcvt.f32.s32 q1, q1\n"  in Transform()
1529       "vcvt.f32.s32 q2, q2\n"  in Transform()
1530       "vcvt.f32.s32 q3, q3\n"  in Transform()
1531       "vsub.f32 q0, q0, q6\n"  in Transform()
1532       "vsub.f32 q1, q1, q6\n"  in Transform()
1533       "vsub.f32 q2, q2, q6\n"  in Transform()
1534       "vsub.f32 q3, q3, q6\n"  in Transform()
1535       "vmul.f32 q0, q0, q7\n"  in Transform()
1536       "vmul.f32 q1, q1, q7\n"  in Transform()
1537       "vmul.f32 q2, q2, q7\n"  in Transform()
1538       "vmul.f32 q3, q3, q7\n"  in Transform()
1539       "vadd.f32 q0, q0, q4\n"  in Transform()
1540       "vadd.f32 q1, q1, q4\n"  in Transform()
1541       "vadd.f32 q2, q2, q4\n"  in Transform()
1542       "vadd.f32 q3, q3, q4\n"  in Transform()
1543       "vmul.f32 q0, q0, q8\n"  in Transform()
1544       "vmul.f32 q1, q1, q8\n"  in Transform()
1545       "vmul.f32 q2, q2, q8\n"  in Transform()
1546       "vmul.f32 q3, q3, q8\n"  in Transform()
1547       "vcvt.s32.f32 q0, q0\n"  in Transform()
1548       "vcvt.s32.f32 q1, q1\n"  in Transform()
1549       "vcvt.s32.f32 q2, q2\n"  in Transform()
1550       "vcvt.s32.f32 q3, q3\n"  in Transform()
1551       "vqmovn.s32 d0, q0\n"  in Transform()
1552       "vqmovn.s32 d1, q1\n"  in Transform()
1553       "vqmovn.s32 d4, q2\n"  in Transform()
1554       "vqmovn.s32 d5, q3\n"  in Transform()
1555       "vqmovun.s16 d0, q0\n"  in Transform()
1556       "vqmovun.s16 d1, q2\n"  in Transform()
1558       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
1559       "pld [%[output]]\n"  in Transform()
1561       "bne 1b\n"  in Transform()
1567       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1568       "vld1.32 {d4, d5, d6}, [%[input]]!\n"  in Transform()
1569       "pld [%[input], #64]\n"  in Transform()
1570       "vcvt.f32.s32 q0, q0\n"  in Transform()
1571       "vcvt.f32.s32 q1, q1\n"  in Transform()
1572       "vcvt.f32.s32 q2, q2\n"  in Transform()
1573       "vcvt.f32.s32 q3, q3\n"  in Transform()
1574       "vsub.f32 q0, q0, q6\n"  in Transform()
1575       "vsub.f32 q1, q1, q6\n"  in Transform()
1576       "vsub.f32 q2, q2, q6\n"  in Transform()
1577       "vsub.f32 q3, q3, q6\n"  in Transform()
1578       "vmul.f32 q0, q0, q7\n"  in Transform()
1579       "vmul.f32 q1, q1, q7\n"  in Transform()
1580       "vmul.f32 q2, q2, q7\n"  in Transform()
1581       "vmul.f32 q3, q3, q7\n"  in Transform()
1582       "vadd.f32 q0, q0, q4\n"  in Transform()
1583       "vadd.f32 q1, q1, q4\n"  in Transform()
1584       "vadd.f32 q2, q2, q4\n"  in Transform()
1585       "vadd.f32 q3, q3, q4\n"  in Transform()
1586       "vmul.f32 q0, q0, q8\n"  in Transform()
1587       "vmul.f32 q1, q1, q8\n"  in Transform()
1588       "vmul.f32 q2, q2, q8\n"  in Transform()
1589       "vmul.f32 q3, q3, q8\n"  in Transform()
1590       "vcvt.s32.f32 q0, q0\n"  in Transform()
1591       "vcvt.s32.f32 q1, q1\n"  in Transform()
1592       "vcvt.s32.f32 q2, q2\n"  in Transform()
1593       "vcvt.s32.f32 q3, q3\n"  in Transform()
1594       "vqmovn.s32 d0, q0\n"  in Transform()
1595       "vqmovn.s32 d1, q1\n"  in Transform()
1596       "vqmovn.s32 d4, q2\n"  in Transform()
1597       "vqmovn.s32 d5, q3\n"  in Transform()
1598       "vqmovun.s16 d0, q0\n"  in Transform()
1599       "vqmovun.s16 d1, q2\n"  in Transform()
1601       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
1602       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
1603       "vst1.16 {d1[2]}, [%[output]]!\n"  in Transform()
1604       "pld [%[output]]\n"  in Transform()
1631       "vdup.32 q4, %[input_range_min]\n"  in Transform()
1632       "vdup.32 q5, %[output_range_min]\n"  in Transform()
1633       "vdup.32 q6, %[input_range_offset]\n"  in Transform()
1634       "vdup.32 q7, %[input_range_scale]\n"  in Transform()
1635       "vdup.32 q8, %[one_over_output_range_scale]\n"  in Transform()
1636       "vsub.f32 q4, q4, q5\n"  in Transform()
1639       "subs %[count], %[count], #15\n"  in Transform()
1640       "beq 2f\n"  in Transform()
1643       "subs %[count], %[count], #16\n"  in Transform()
1646       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1647       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1648       "pld [%[input], #64]\n"  in Transform()
1649       "vcvt.f32.s32 q0, q0\n"  in Transform()
1650       "vcvt.f32.s32 q1, q1\n"  in Transform()
1651       "vcvt.f32.s32 q2, q2\n"  in Transform()
1652       "vcvt.f32.s32 q3, q3\n"  in Transform()
1653       "vsub.f32 q0, q0, q6\n"  in Transform()
1654       "vsub.f32 q1, q1, q6\n"  in Transform()
1655       "vsub.f32 q2, q2, q6\n"  in Transform()
1656       "vsub.f32 q3, q3, q6\n"  in Transform()
1657       "vmul.f32 q0, q0, q7\n"  in Transform()
1658       "vmul.f32 q1, q1, q7\n"  in Transform()
1659       "vmul.f32 q2, q2, q7\n"  in Transform()
1660       "vmul.f32 q3, q3, q7\n"  in Transform()
1661       "vadd.f32 q0, q0, q4\n"  in Transform()
1662       "vadd.f32 q1, q1, q4\n"  in Transform()
1663       "vadd.f32 q2, q2, q4\n"  in Transform()
1664       "vadd.f32 q3, q3, q4\n"  in Transform()
1665       "vmul.f32 q0, q0, q8\n"  in Transform()
1666       "vmul.f32 q1, q1, q8\n"  in Transform()
1667       "vmul.f32 q2, q2, q8\n"  in Transform()
1668       "vmul.f32 q3, q3, q8\n"  in Transform()
1669       "vcvt.s32.f32 q0, q0\n"  in Transform()
1670       "vcvt.s32.f32 q1, q1\n"  in Transform()
1671       "vcvt.s32.f32 q2, q2\n"  in Transform()
1672       "vcvt.s32.f32 q3, q3\n"  in Transform()
1673       "vqmovn.s32 d0, q0\n"  in Transform()
1674       "vqmovn.s32 d1, q1\n"  in Transform()
1675       "vqmovn.s32 d4, q2\n"  in Transform()
1676       "vqmovn.s32 d5, q3\n"  in Transform()
1677       "vqmovun.s16 d0, q0\n"  in Transform()
1678       "vqmovun.s16 d1, q2\n"  in Transform()
1680       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
1681       "pld [%[output]]\n"  in Transform()
1683       "bne 1b\n"  in Transform()
1689       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1690       "vld1.32 {d4, d5, d6}, [%[input]]!\n"  in Transform()
1691       "vld1.32 {d7[0]}, [%[input]]!\n"  in Transform()
1692       "pld [%[input], #64]\n"  in Transform()
1693       "vcvt.f32.s32 q0, q0\n"  in Transform()
1694       "vcvt.f32.s32 q1, q1\n"  in Transform()
1695       "vcvt.f32.s32 q2, q2\n"  in Transform()
1696       "vcvt.f32.s32 q3, q3\n"  in Transform()
1697       "vsub.f32 q0, q0, q6\n"  in Transform()
1698       "vsub.f32 q1, q1, q6\n"  in Transform()
1699       "vsub.f32 q2, q2, q6\n"  in Transform()
1700       "vsub.f32 q3, q3, q6\n"  in Transform()
1701       "vmul.f32 q0, q0, q7\n"  in Transform()
1702       "vmul.f32 q1, q1, q7\n"  in Transform()
1703       "vmul.f32 q2, q2, q7\n"  in Transform()
1704       "vmul.f32 q3, q3, q7\n"  in Transform()
1705       "vadd.f32 q0, q0, q4\n"  in Transform()
1706       "vadd.f32 q1, q1, q4\n"  in Transform()
1707       "vadd.f32 q2, q2, q4\n"  in Transform()
1708       "vadd.f32 q3, q3, q4\n"  in Transform()
1709       "vmul.f32 q0, q0, q8\n"  in Transform()
1710       "vmul.f32 q1, q1, q8\n"  in Transform()
1711       "vmul.f32 q2, q2, q8\n"  in Transform()
1712       "vmul.f32 q3, q3, q8\n"  in Transform()
1713       "vcvt.s32.f32 q0, q0\n"  in Transform()
1714       "vcvt.s32.f32 q1, q1\n"  in Transform()
1715       "vcvt.s32.f32 q2, q2\n"  in Transform()
1716       "vcvt.s32.f32 q3, q3\n"  in Transform()
1717       "vqmovn.s32 d0, q0\n"  in Transform()
1718       "vqmovn.s32 d1, q1\n"  in Transform()
1719       "vqmovn.s32 d4, q2\n"  in Transform()
1720       "vqmovn.s32 d5, q3\n"  in Transform()
1721       "vqmovun.s16 d0, q0\n"  in Transform()
1722       "vqmovun.s16 d1, q2\n"  in Transform()
1724       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
1725       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
1726       "vst1.16 {d1[2]}, [%[output]]!\n"  in Transform()
1727       "vst1.8 {d1[6]}, [%[output]]!\n"  in Transform()
1728       "pld [%[output]]\n"  in Transform()
1755       "vdup.32 q4, %[range_min]\n"  in Transform()
1756       "vdup.32 q5, %[range_offset]\n"  in Transform()
1757       "vdup.32 q6, %[range_scale]\n"  in Transform()
1760       "subs %[count], %[count], #16\n"  in Transform()
1763       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1764       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1765       "pld [%[input], #64]\n"  in Transform()
1766       "vsub.f32 q0, q0, q4\n"  in Transform()
1767       "vsub.f32 q1, q1, q4\n"  in Transform()
1768       "vsub.f32 q2, q2, q4\n"  in Transform()
1769       "vsub.f32 q3, q3, q4\n"  in Transform()
1770       "vmul.f32 q0, q0, q6\n"  in Transform()
1771       "vmul.f32 q1, q1, q6\n"  in Transform()
1772       "vmul.f32 q2, q2, q6\n"  in Transform()
1773       "vmul.f32 q3, q3, q6\n"  in Transform()
1774       "vadd.f32 q0, q0, q5\n"  in Transform()
1775       "vadd.f32 q1, q1, q5\n"  in Transform()
1776       "vadd.f32 q2, q2, q5\n"  in Transform()
1777       "vadd.f32 q3, q3, q5\n"  in Transform()
1778       "vcvt.s32.f32 q0, q0\n"  in Transform()
1779       "vcvt.s32.f32 q1, q1\n"  in Transform()
1780       "vcvt.s32.f32 q2, q2\n"  in Transform()
1781       "vcvt.s32.f32 q3, q3\n"  in Transform()
1782       "vqmovn.s32 d0, q0\n"  in Transform()
1783       "vqmovn.s32 d1, q1\n"  in Transform()
1784       "vqmovn.s32 d4, q2\n"  in Transform()
1785       "vqmovn.s32 d5, q3\n"  in Transform()
1786       "vqmovun.s16 d0, q0\n"  in Transform()
1787       "vqmovun.s16 d1, q2\n"  in Transform()
1789       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
1790       "pld [%[output]]\n"  in Transform()
1792       "bne 1b\n"  in Transform()
1816       "vdup.32 q4, %[range_min]\n"  in Transform()
1817       "vdup.32 q5, %[range_offset]\n"  in Transform()
1818       "vdup.32 q6, %[range_scale]\n"  in Transform()
1821       "subs %[count], %[count], #1\n"  in Transform()
1822       "beq 2f\n"  in Transform()
1825       "subs %[count], %[count], #16\n"  in Transform()
1828       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1829       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1830       "pld [%[input], #64]\n"  in Transform()
1831       "vsub.f32 q0, q0, q4\n"  in Transform()
1832       "vsub.f32 q1, q1, q4\n"  in Transform()
1833       "vsub.f32 q2, q2, q4\n"  in Transform()
1834       "vsub.f32 q3, q3, q4\n"  in Transform()
1835       "vmul.f32 q0, q0, q6\n"  in Transform()
1836       "vmul.f32 q1, q1, q6\n"  in Transform()
1837       "vmul.f32 q2, q2, q6\n"  in Transform()
1838       "vmul.f32 q3, q3, q6\n"  in Transform()
1839       "vadd.f32 q0, q0, q5\n"  in Transform()
1840       "vadd.f32 q1, q1, q5\n"  in Transform()
1841       "vadd.f32 q2, q2, q5\n"  in Transform()
1842       "vadd.f32 q3, q3, q5\n"  in Transform()
1843       "vcvt.s32.f32 q0, q0\n"  in Transform()
1844       "vcvt.s32.f32 q1, q1\n"  in Transform()
1845       "vcvt.s32.f32 q2, q2\n"  in Transform()
1846       "vcvt.s32.f32 q3, q3\n"  in Transform()
1847       "vqmovn.s32 d0, q0\n"  in Transform()
1848       "vqmovn.s32 d1, q1\n"  in Transform()
1849       "vqmovn.s32 d4, q2\n"  in Transform()
1850       "vqmovn.s32 d5, q3\n"  in Transform()
1851       "vqmovun.s16 d0, q0\n"  in Transform()
1852       "vqmovun.s16 d1, q2\n"  in Transform()
1854       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
1855       "pld [%[output]]\n"  in Transform()
1857       "bne 1b\n"  in Transform()
1863       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
1864       "pld [%[input], #64]\n"  in Transform()
1865       "vsub.f32 q0, q0, q4\n"  in Transform()
1866       "vmul.f32 q0, q0, q6\n"  in Transform()
1867       "vadd.f32 q0, q0, q5\n"  in Transform()
1868       "vcvt.s32.f32 q0, q0\n"  in Transform()
1869       "vqmovn.s32 d0, q0\n"  in Transform()
1870       "vqmovun.s16 d0, q0\n"  in Transform()
1872       "vst1.8 {d0[0]}, [%[output]]!\n"  in Transform()
1873       "pld [%[output]]\n"  in Transform()
1897       "vdup.32 q4, %[range_min]\n"  in Transform()
1898       "vdup.32 q5, %[range_offset]\n"  in Transform()
1899       "vdup.32 q6, %[range_scale]\n"  in Transform()
1902       "subs %[count], %[count], #2\n"  in Transform()
1903       "beq 2f\n"  in Transform()
1906       "subs %[count], %[count], #16\n"  in Transform()
1909       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1910       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1911       "pld [%[input], #64]\n"  in Transform()
1912       "vsub.f32 q0, q0, q4\n"  in Transform()
1913       "vsub.f32 q1, q1, q4\n"  in Transform()
1914       "vsub.f32 q2, q2, q4\n"  in Transform()
1915       "vsub.f32 q3, q3, q4\n"  in Transform()
1916       "vmul.f32 q0, q0, q6\n"  in Transform()
1917       "vmul.f32 q1, q1, q6\n"  in Transform()
1918       "vmul.f32 q2, q2, q6\n"  in Transform()
1919       "vmul.f32 q3, q3, q6\n"  in Transform()
1920       "vadd.f32 q0, q0, q5\n"  in Transform()
1921       "vadd.f32 q1, q1, q5\n"  in Transform()
1922       "vadd.f32 q2, q2, q5\n"  in Transform()
1923       "vadd.f32 q3, q3, q5\n"  in Transform()
1924       "vcvt.s32.f32 q0, q0\n"  in Transform()
1925       "vcvt.s32.f32 q1, q1\n"  in Transform()
1926       "vcvt.s32.f32 q2, q2\n"  in Transform()
1927       "vcvt.s32.f32 q3, q3\n"  in Transform()
1928       "vqmovn.s32 d0, q0\n"  in Transform()
1929       "vqmovn.s32 d1, q1\n"  in Transform()
1930       "vqmovn.s32 d4, q2\n"  in Transform()
1931       "vqmovn.s32 d5, q3\n"  in Transform()
1932       "vqmovun.s16 d0, q0\n"  in Transform()
1933       "vqmovun.s16 d1, q2\n"  in Transform()
1935       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
1936       "pld [%[output]]\n"  in Transform()
1938       "bne 1b\n"  in Transform()
1944       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
1945       "pld [%[input], #64]\n"  in Transform()
1946       "vsub.f32 q0, q0, q4\n"  in Transform()
1947       "vmul.f32 q0, q0, q6\n"  in Transform()
1948       "vadd.f32 q0, q0, q5\n"  in Transform()
1949       "vcvt.s32.f32 q0, q0\n"  in Transform()
1950       "vqmovn.s32 d0, q0\n"  in Transform()
1951       "vqmovun.s16 d0, q0\n"  in Transform()
1953       "vst1.16 {d0[0]}, [%[output]]!\n"  in Transform()
1954       "pld [%[output]]\n"  in Transform()
1978       "vdup.32 q4, %[range_min]\n"  in Transform()
1979       "vdup.32 q5, %[range_offset]\n"  in Transform()
1980       "vdup.32 q6, %[range_scale]\n"  in Transform()
1983       "subs %[count], %[count], #3\n"  in Transform()
1984       "beq 2f\n"  in Transform()
1987       "subs %[count], %[count], #16\n"  in Transform()
1990       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
1991       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
1992       "pld [%[input], #64]\n"  in Transform()
1993       "vsub.f32 q0, q0, q4\n"  in Transform()
1994       "vsub.f32 q1, q1, q4\n"  in Transform()
1995       "vsub.f32 q2, q2, q4\n"  in Transform()
1996       "vsub.f32 q3, q3, q4\n"  in Transform()
1997       "vmul.f32 q0, q0, q6\n"  in Transform()
1998       "vmul.f32 q1, q1, q6\n"  in Transform()
1999       "vmul.f32 q2, q2, q6\n"  in Transform()
2000       "vmul.f32 q3, q3, q6\n"  in Transform()
2001       "vadd.f32 q0, q0, q5\n"  in Transform()
2002       "vadd.f32 q1, q1, q5\n"  in Transform()
2003       "vadd.f32 q2, q2, q5\n"  in Transform()
2004       "vadd.f32 q3, q3, q5\n"  in Transform()
2005       "vcvt.s32.f32 q0, q0\n"  in Transform()
2006       "vcvt.s32.f32 q1, q1\n"  in Transform()
2007       "vcvt.s32.f32 q2, q2\n"  in Transform()
2008       "vcvt.s32.f32 q3, q3\n"  in Transform()
2009       "vqmovn.s32 d0, q0\n"  in Transform()
2010       "vqmovn.s32 d1, q1\n"  in Transform()
2011       "vqmovn.s32 d4, q2\n"  in Transform()
2012       "vqmovn.s32 d5, q3\n"  in Transform()
2013       "vqmovun.s16 d0, q0\n"  in Transform()
2014       "vqmovun.s16 d1, q2\n"  in Transform()
2016       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2017       "pld [%[output]]\n"  in Transform()
2019       "bne 1b\n"  in Transform()
2025       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
2026       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
2027       "pld [%[input], #64]\n"  in Transform()
2028       "vsub.f32 q0, q0, q4\n"  in Transform()
2029       "vmul.f32 q0, q0, q6\n"  in Transform()
2030       "vadd.f32 q0, q0, q5\n"  in Transform()
2031       "vcvt.s32.f32 q0, q0\n"  in Transform()
2032       "vqmovn.s32 d0, q0\n"  in Transform()
2033       "vqmovun.s16 d0, q0\n"  in Transform()
2035       "vst1.16 {d0[0]}, [%[output]]!\n"  in Transform()
2036       "vst1.8 {d0[2]}, [%[output]]!\n"  in Transform()
2037       "pld [%[output]]\n"  in Transform()
2061       "vdup.32 q4, %[range_min]\n"  in Transform()
2062       "vdup.32 q5, %[range_offset]\n"  in Transform()
2063       "vdup.32 q6, %[range_scale]\n"  in Transform()
2066       "subs %[count], %[count], #4\n"  in Transform()
2067       "beq 2f\n"  in Transform()
2070       "subs %[count], %[count], #16\n"  in Transform()
2073       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2074       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2075       "pld [%[input], #64]\n"  in Transform()
2076       "vsub.f32 q0, q0, q4\n"  in Transform()
2077       "vsub.f32 q1, q1, q4\n"  in Transform()
2078       "vsub.f32 q2, q2, q4\n"  in Transform()
2079       "vsub.f32 q3, q3, q4\n"  in Transform()
2080       "vmul.f32 q0, q0, q6\n"  in Transform()
2081       "vmul.f32 q1, q1, q6\n"  in Transform()
2082       "vmul.f32 q2, q2, q6\n"  in Transform()
2083       "vmul.f32 q3, q3, q6\n"  in Transform()
2084       "vadd.f32 q0, q0, q5\n"  in Transform()
2085       "vadd.f32 q1, q1, q5\n"  in Transform()
2086       "vadd.f32 q2, q2, q5\n"  in Transform()
2087       "vadd.f32 q3, q3, q5\n"  in Transform()
2088       "vcvt.s32.f32 q0, q0\n"  in Transform()
2089       "vcvt.s32.f32 q1, q1\n"  in Transform()
2090       "vcvt.s32.f32 q2, q2\n"  in Transform()
2091       "vcvt.s32.f32 q3, q3\n"  in Transform()
2092       "vqmovn.s32 d0, q0\n"  in Transform()
2093       "vqmovn.s32 d1, q1\n"  in Transform()
2094       "vqmovn.s32 d4, q2\n"  in Transform()
2095       "vqmovn.s32 d5, q3\n"  in Transform()
2096       "vqmovun.s16 d0, q0\n"  in Transform()
2097       "vqmovun.s16 d1, q2\n"  in Transform()
2099       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2100       "pld [%[output]]\n"  in Transform()
2102       "bne 1b\n"  in Transform()
2108       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
2109       "pld [%[input], #64]\n"  in Transform()
2110       "vsub.f32 q0, q0, q4\n"  in Transform()
2111       "vmul.f32 q0, q0, q6\n"  in Transform()
2112       "vadd.f32 q0, q0, q5\n"  in Transform()
2113       "vcvt.s32.f32 q0, q0\n"  in Transform()
2114       "vqmovn.s32 d0, q0\n"  in Transform()
2115       "vqmovun.s16 d0, q0\n"  in Transform()
2117       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
2118       "pld [%[output]]\n"  in Transform()
2142       "vdup.32 q4, %[range_min]\n"  in Transform()
2143       "vdup.32 q5, %[range_offset]\n"  in Transform()
2144       "vdup.32 q6, %[range_scale]\n"  in Transform()
2147       "subs %[count], %[count], #5\n"  in Transform()
2148       "beq 2f\n"  in Transform()
2151       "subs %[count], %[count], #16\n"  in Transform()
2154       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2155       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2156       "pld [%[input], #64]\n"  in Transform()
2157       "vsub.f32 q0, q0, q4\n"  in Transform()
2158       "vsub.f32 q1, q1, q4\n"  in Transform()
2159       "vsub.f32 q2, q2, q4\n"  in Transform()
2160       "vsub.f32 q3, q3, q4\n"  in Transform()
2161       "vmul.f32 q0, q0, q6\n"  in Transform()
2162       "vmul.f32 q1, q1, q6\n"  in Transform()
2163       "vmul.f32 q2, q2, q6\n"  in Transform()
2164       "vmul.f32 q3, q3, q6\n"  in Transform()
2165       "vadd.f32 q0, q0, q5\n"  in Transform()
2166       "vadd.f32 q1, q1, q5\n"  in Transform()
2167       "vadd.f32 q2, q2, q5\n"  in Transform()
2168       "vadd.f32 q3, q3, q5\n"  in Transform()
2169       "vcvt.s32.f32 q0, q0\n"  in Transform()
2170       "vcvt.s32.f32 q1, q1\n"  in Transform()
2171       "vcvt.s32.f32 q2, q2\n"  in Transform()
2172       "vcvt.s32.f32 q3, q3\n"  in Transform()
2173       "vqmovn.s32 d0, q0\n"  in Transform()
2174       "vqmovn.s32 d1, q1\n"  in Transform()
2175       "vqmovn.s32 d4, q2\n"  in Transform()
2176       "vqmovn.s32 d5, q3\n"  in Transform()
2177       "vqmovun.s16 d0, q0\n"  in Transform()
2178       "vqmovun.s16 d1, q2\n"  in Transform()
2180       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2181       "pld [%[output]]\n"  in Transform()
2183       "bne 1b\n"  in Transform()
2189       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
2190       "vld1.32 {d2[0]}, [%[input]]!\n"  in Transform()
2191       "pld [%[input], #64]\n"  in Transform()
2192       "vsub.f32 q0, q0, q4\n"  in Transform()
2193       "vsub.f32 q1, q1, q4\n"  in Transform()
2194       "vmul.f32 q0, q0, q6\n"  in Transform()
2195       "vmul.f32 q1, q1, q6\n"  in Transform()
2196       "vadd.f32 q0, q0, q5\n"  in Transform()
2197       "vadd.f32 q1, q1, q5\n"  in Transform()
2198       "vcvt.s32.f32 q0, q0\n"  in Transform()
2199       "vcvt.s32.f32 q1, q1\n"  in Transform()
2200       "vqmovn.s32 d0, q0\n"  in Transform()
2201       "vqmovn.s32 d1, q1\n"  in Transform()
2202       "vqmovun.s16 d0, q0\n"  in Transform()
2204       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
2205       "vst1.8 {d0[4]}, [%[output]]!\n"  in Transform()
2206       "pld [%[output]]\n"  in Transform()
2230       "vdup.32 q4, %[range_min]\n"  in Transform()
2231       "vdup.32 q5, %[range_offset]\n"  in Transform()
2232       "vdup.32 q6, %[range_scale]\n"  in Transform()
2235       "subs %[count], %[count], #6\n"  in Transform()
2236       "beq 2f\n"  in Transform()
2239       "subs %[count], %[count], #16\n"  in Transform()
2242       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2243       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2244       "pld [%[input], #64]\n"  in Transform()
2245       "vsub.f32 q0, q0, q4\n"  in Transform()
2246       "vsub.f32 q1, q1, q4\n"  in Transform()
2247       "vsub.f32 q2, q2, q4\n"  in Transform()
2248       "vsub.f32 q3, q3, q4\n"  in Transform()
2249       "vmul.f32 q0, q0, q6\n"  in Transform()
2250       "vmul.f32 q1, q1, q6\n"  in Transform()
2251       "vmul.f32 q2, q2, q6\n"  in Transform()
2252       "vmul.f32 q3, q3, q6\n"  in Transform()
2253       "vadd.f32 q0, q0, q5\n"  in Transform()
2254       "vadd.f32 q1, q1, q5\n"  in Transform()
2255       "vadd.f32 q2, q2, q5\n"  in Transform()
2256       "vadd.f32 q3, q3, q5\n"  in Transform()
2257       "vcvt.s32.f32 q0, q0\n"  in Transform()
2258       "vcvt.s32.f32 q1, q1\n"  in Transform()
2259       "vcvt.s32.f32 q2, q2\n"  in Transform()
2260       "vcvt.s32.f32 q3, q3\n"  in Transform()
2261       "vqmovn.s32 d0, q0\n"  in Transform()
2262       "vqmovn.s32 d1, q1\n"  in Transform()
2263       "vqmovn.s32 d4, q2\n"  in Transform()
2264       "vqmovn.s32 d5, q3\n"  in Transform()
2265       "vqmovun.s16 d0, q0\n"  in Transform()
2266       "vqmovun.s16 d1, q2\n"  in Transform()
2268       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2269       "pld [%[output]]\n"  in Transform()
2271       "bne 1b\n"  in Transform()
2277       "vld1.32 {d0, d1, d2}, [%[input]]!\n"  in Transform()
2278       "pld [%[input], #64]\n"  in Transform()
2279       "vsub.f32 q0, q0, q4\n"  in Transform()
2280       "vsub.f32 q1, q1, q4\n"  in Transform()
2281       "vmul.f32 q0, q0, q6\n"  in Transform()
2282       "vmul.f32 q1, q1, q6\n"  in Transform()
2283       "vadd.f32 q0, q0, q5\n"  in Transform()
2284       "vadd.f32 q1, q1, q5\n"  in Transform()
2285       "vcvt.s32.f32 q0, q0\n"  in Transform()
2286       "vcvt.s32.f32 q1, q1\n"  in Transform()
2287       "vqmovn.s32 d0, q0\n"  in Transform()
2288       "vqmovn.s32 d1, q1\n"  in Transform()
2289       "vqmovun.s16 d0, q0\n"  in Transform()
2291       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
2292       "vst1.16 {d0[2]}, [%[output]]!\n"  in Transform()
2293       "pld [%[output]]\n"  in Transform()
2317       "vdup.32 q4, %[range_min]\n"  in Transform()
2318       "vdup.32 q5, %[range_offset]\n"  in Transform()
2319       "vdup.32 q6, %[range_scale]\n"  in Transform()
2322       "subs %[count], %[count], #7\n"  in Transform()
2323       "beq 2f\n"  in Transform()
2326       "subs %[count], %[count], #16\n"  in Transform()
2329       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2330       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2331       "pld [%[input], #64]\n"  in Transform()
2332       "vsub.f32 q0, q0, q4\n"  in Transform()
2333       "vsub.f32 q1, q1, q4\n"  in Transform()
2334       "vsub.f32 q2, q2, q4\n"  in Transform()
2335       "vsub.f32 q3, q3, q4\n"  in Transform()
2336       "vmul.f32 q0, q0, q6\n"  in Transform()
2337       "vmul.f32 q1, q1, q6\n"  in Transform()
2338       "vmul.f32 q2, q2, q6\n"  in Transform()
2339       "vmul.f32 q3, q3, q6\n"  in Transform()
2340       "vadd.f32 q0, q0, q5\n"  in Transform()
2341       "vadd.f32 q1, q1, q5\n"  in Transform()
2342       "vadd.f32 q2, q2, q5\n"  in Transform()
2343       "vadd.f32 q3, q3, q5\n"  in Transform()
2344       "vcvt.s32.f32 q0, q0\n"  in Transform()
2345       "vcvt.s32.f32 q1, q1\n"  in Transform()
2346       "vcvt.s32.f32 q2, q2\n"  in Transform()
2347       "vcvt.s32.f32 q3, q3\n"  in Transform()
2348       "vqmovn.s32 d0, q0\n"  in Transform()
2349       "vqmovn.s32 d1, q1\n"  in Transform()
2350       "vqmovn.s32 d4, q2\n"  in Transform()
2351       "vqmovn.s32 d5, q3\n"  in Transform()
2352       "vqmovun.s16 d0, q0\n"  in Transform()
2353       "vqmovun.s16 d1, q2\n"  in Transform()
2355       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2356       "pld [%[output]]\n"  in Transform()
2358       "bne 1b\n"  in Transform()
2364       "vld1.32 {d0, d1, d2}, [%[input]]!\n"  in Transform()
2365       "vld1.32 {d3[0]}, [%[input]]!\n"  in Transform()
2366       "pld [%[input], #64]\n"  in Transform()
2367       "vsub.f32 q0, q0, q4\n"  in Transform()
2368       "vsub.f32 q1, q1, q4\n"  in Transform()
2369       "vmul.f32 q0, q0, q6\n"  in Transform()
2370       "vmul.f32 q1, q1, q6\n"  in Transform()
2371       "vadd.f32 q0, q0, q5\n"  in Transform()
2372       "vadd.f32 q1, q1, q5\n"  in Transform()
2373       "vcvt.s32.f32 q0, q0\n"  in Transform()
2374       "vcvt.s32.f32 q1, q1\n"  in Transform()
2375       "vqmovn.s32 d0, q0\n"  in Transform()
2376       "vqmovn.s32 d1, q1\n"  in Transform()
2377       "vqmovun.s16 d0, q0\n"  in Transform()
2379       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
2380       "vst1.16 {d0[2]}, [%[output]]!\n"  in Transform()
2381       "vst1.8 {d0[6]}, [%[output]]!\n"  in Transform()
2382       "pld [%[output]]\n"  in Transform()
2406       "vdup.32 q4, %[range_min]\n"  in Transform()
2407       "vdup.32 q5, %[range_offset]\n"  in Transform()
2408       "vdup.32 q6, %[range_scale]\n"  in Transform()
2411       "subs %[count], %[count], #8\n"  in Transform()
2412       "beq 2f\n"  in Transform()
2415       "subs %[count], %[count], #16\n"  in Transform()
2418       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2419       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2420       "pld [%[input], #64]\n"  in Transform()
2421       "vsub.f32 q0, q0, q4\n"  in Transform()
2422       "vsub.f32 q1, q1, q4\n"  in Transform()
2423       "vsub.f32 q2, q2, q4\n"  in Transform()
2424       "vsub.f32 q3, q3, q4\n"  in Transform()
2425       "vmul.f32 q0, q0, q6\n"  in Transform()
2426       "vmul.f32 q1, q1, q6\n"  in Transform()
2427       "vmul.f32 q2, q2, q6\n"  in Transform()
2428       "vmul.f32 q3, q3, q6\n"  in Transform()
2429       "vadd.f32 q0, q0, q5\n"  in Transform()
2430       "vadd.f32 q1, q1, q5\n"  in Transform()
2431       "vadd.f32 q2, q2, q5\n"  in Transform()
2432       "vadd.f32 q3, q3, q5\n"  in Transform()
2433       "vcvt.s32.f32 q0, q0\n"  in Transform()
2434       "vcvt.s32.f32 q1, q1\n"  in Transform()
2435       "vcvt.s32.f32 q2, q2\n"  in Transform()
2436       "vcvt.s32.f32 q3, q3\n"  in Transform()
2437       "vqmovn.s32 d0, q0\n"  in Transform()
2438       "vqmovn.s32 d1, q1\n"  in Transform()
2439       "vqmovn.s32 d4, q2\n"  in Transform()
2440       "vqmovn.s32 d5, q3\n"  in Transform()
2441       "vqmovun.s16 d0, q0\n"  in Transform()
2442       "vqmovun.s16 d1, q2\n"  in Transform()
2444       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2445       "pld [%[output]]\n"  in Transform()
2447       "bne 1b\n"  in Transform()
2453       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2454       "pld [%[input], #64]\n"  in Transform()
2455       "vsub.f32 q0, q0, q4\n"  in Transform()
2456       "vsub.f32 q1, q1, q4\n"  in Transform()
2457       "vmul.f32 q0, q0, q6\n"  in Transform()
2458       "vmul.f32 q1, q1, q6\n"  in Transform()
2459       "vadd.f32 q0, q0, q5\n"  in Transform()
2460       "vadd.f32 q1, q1, q5\n"  in Transform()
2461       "vcvt.s32.f32 q0, q0\n"  in Transform()
2462       "vcvt.s32.f32 q1, q1\n"  in Transform()
2463       "vqmovn.s32 d0, q0\n"  in Transform()
2464       "vqmovn.s32 d1, q1\n"  in Transform()
2465       "vqmovun.s16 d0, q0\n"  in Transform()
2467       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
2468       "pld [%[output]]\n"  in Transform()
2492       "vdup.32 q4, %[range_min]\n"  in Transform()
2493       "vdup.32 q5, %[range_offset]\n"  in Transform()
2494       "vdup.32 q6, %[range_scale]\n"  in Transform()
2497       "subs %[count], %[count], #9\n"  in Transform()
2498       "beq 2f\n"  in Transform()
2501       "subs %[count], %[count], #16\n"  in Transform()
2504       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2505       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2506       "pld [%[input], #64]\n"  in Transform()
2507       "vsub.f32 q0, q0, q4\n"  in Transform()
2508       "vsub.f32 q1, q1, q4\n"  in Transform()
2509       "vsub.f32 q2, q2, q4\n"  in Transform()
2510       "vsub.f32 q3, q3, q4\n"  in Transform()
2511       "vmul.f32 q0, q0, q6\n"  in Transform()
2512       "vmul.f32 q1, q1, q6\n"  in Transform()
2513       "vmul.f32 q2, q2, q6\n"  in Transform()
2514       "vmul.f32 q3, q3, q6\n"  in Transform()
2515       "vadd.f32 q0, q0, q5\n"  in Transform()
2516       "vadd.f32 q1, q1, q5\n"  in Transform()
2517       "vadd.f32 q2, q2, q5\n"  in Transform()
2518       "vadd.f32 q3, q3, q5\n"  in Transform()
2519       "vcvt.s32.f32 q0, q0\n"  in Transform()
2520       "vcvt.s32.f32 q1, q1\n"  in Transform()
2521       "vcvt.s32.f32 q2, q2\n"  in Transform()
2522       "vcvt.s32.f32 q3, q3\n"  in Transform()
2523       "vqmovn.s32 d0, q0\n"  in Transform()
2524       "vqmovn.s32 d1, q1\n"  in Transform()
2525       "vqmovn.s32 d4, q2\n"  in Transform()
2526       "vqmovn.s32 d5, q3\n"  in Transform()
2527       "vqmovun.s16 d0, q0\n"  in Transform()
2528       "vqmovun.s16 d1, q2\n"  in Transform()
2530       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2531       "pld [%[output]]\n"  in Transform()
2533       "bne 1b\n"  in Transform()
2539       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2540       "vld1.32 {d4[0]}, [%[input]]!\n"  in Transform()
2541       "pld [%[input], #64]\n"  in Transform()
2542       "vsub.f32 q0, q0, q4\n"  in Transform()
2543       "vsub.f32 q1, q1, q4\n"  in Transform()
2544       "vsub.f32 q2, q2, q4\n"  in Transform()
2545       "vmul.f32 q0, q0, q6\n"  in Transform()
2546       "vmul.f32 q1, q1, q6\n"  in Transform()
2547       "vmul.f32 q2, q2, q6\n"  in Transform()
2548       "vadd.f32 q0, q0, q5\n"  in Transform()
2549       "vadd.f32 q1, q1, q5\n"  in Transform()
2550       "vadd.f32 q2, q2, q5\n"  in Transform()
2551       "vcvt.s32.f32 q0, q0\n"  in Transform()
2552       "vcvt.s32.f32 q1, q1\n"  in Transform()
2553       "vcvt.s32.f32 q2, q2\n"  in Transform()
2554       "vqmovn.s32 d0, q0\n"  in Transform()
2555       "vqmovn.s32 d1, q1\n"  in Transform()
2556       "vqmovn.s32 d4, q2\n"  in Transform()
2557       "vqmovun.s16 d0, q0\n"  in Transform()
2558       "vqmovun.s16 d1, q2\n"  in Transform()
2560       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
2561       "vst1.8 {d1[0]}, [%[output]]!\n"  in Transform()
2562       "pld [%[output]]\n"  in Transform()
2586       "vdup.32 q4, %[range_min]\n"  in Transform()
2587       "vdup.32 q5, %[range_offset]\n"  in Transform()
2588       "vdup.32 q6, %[range_scale]\n"  in Transform()
2591       "subs %[count], %[count], #10\n"  in Transform()
2592       "beq 2f\n"  in Transform()
2595       "subs %[count], %[count], #16\n"  in Transform()
2598       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2599       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2600       "pld [%[input], #64]\n"  in Transform()
2601       "vsub.f32 q0, q0, q4\n"  in Transform()
2602       "vsub.f32 q1, q1, q4\n"  in Transform()
2603       "vsub.f32 q2, q2, q4\n"  in Transform()
2604       "vsub.f32 q3, q3, q4\n"  in Transform()
2605       "vmul.f32 q0, q0, q6\n"  in Transform()
2606       "vmul.f32 q1, q1, q6\n"  in Transform()
2607       "vmul.f32 q2, q2, q6\n"  in Transform()
2608       "vmul.f32 q3, q3, q6\n"  in Transform()
2609       "vadd.f32 q0, q0, q5\n"  in Transform()
2610       "vadd.f32 q1, q1, q5\n"  in Transform()
2611       "vadd.f32 q2, q2, q5\n"  in Transform()
2612       "vadd.f32 q3, q3, q5\n"  in Transform()
2613       "vcvt.s32.f32 q0, q0\n"  in Transform()
2614       "vcvt.s32.f32 q1, q1\n"  in Transform()
2615       "vcvt.s32.f32 q2, q2\n"  in Transform()
2616       "vcvt.s32.f32 q3, q3\n"  in Transform()
2617       "vqmovn.s32 d0, q0\n"  in Transform()
2618       "vqmovn.s32 d1, q1\n"  in Transform()
2619       "vqmovn.s32 d4, q2\n"  in Transform()
2620       "vqmovn.s32 d5, q3\n"  in Transform()
2621       "vqmovun.s16 d0, q0\n"  in Transform()
2622       "vqmovun.s16 d1, q2\n"  in Transform()
2624       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2625       "pld [%[output]]\n"  in Transform()
2627       "bne 1b\n"  in Transform()
2633       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2634       "vld1.32 {d4}, [%[input]]!\n"  in Transform()
2635       "pld [%[input], #64]\n"  in Transform()
2636       "vsub.f32 q0, q0, q4\n"  in Transform()
2637       "vsub.f32 q1, q1, q4\n"  in Transform()
2638       "vsub.f32 q2, q2, q4\n"  in Transform()
2639       "vmul.f32 q0, q0, q6\n"  in Transform()
2640       "vmul.f32 q1, q1, q6\n"  in Transform()
2641       "vmul.f32 q2, q2, q6\n"  in Transform()
2642       "vadd.f32 q0, q0, q5\n"  in Transform()
2643       "vadd.f32 q1, q1, q5\n"  in Transform()
2644       "vadd.f32 q2, q2, q5\n"  in Transform()
2645       "vcvt.s32.f32 q0, q0\n"  in Transform()
2646       "vcvt.s32.f32 q1, q1\n"  in Transform()
2647       "vcvt.s32.f32 q2, q2\n"  in Transform()
2648       "vqmovn.s32 d0, q0\n"  in Transform()
2649       "vqmovn.s32 d1, q1\n"  in Transform()
2650       "vqmovn.s32 d4, q2\n"  in Transform()
2651       "vqmovun.s16 d0, q0\n"  in Transform()
2652       "vqmovun.s16 d1, q2\n"  in Transform()
2654       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
2655       "vst1.16 {d1[0]}, [%[output]]!\n"  in Transform()
2656       "pld [%[output]]\n"  in Transform()
2680       "vdup.32 q4, %[range_min]\n"  in Transform()
2681       "vdup.32 q5, %[range_offset]\n"  in Transform()
2682       "vdup.32 q6, %[range_scale]\n"  in Transform()
2685       "subs %[count], %[count], #11\n"  in Transform()
2686       "beq 2f\n"  in Transform()
2689       "subs %[count], %[count], #16\n"  in Transform()
2692       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2693       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2694       "pld [%[input], #64]\n"  in Transform()
2695       "vsub.f32 q0, q0, q4\n"  in Transform()
2696       "vsub.f32 q1, q1, q4\n"  in Transform()
2697       "vsub.f32 q2, q2, q4\n"  in Transform()
2698       "vsub.f32 q3, q3, q4\n"  in Transform()
2699       "vmul.f32 q0, q0, q6\n"  in Transform()
2700       "vmul.f32 q1, q1, q6\n"  in Transform()
2701       "vmul.f32 q2, q2, q6\n"  in Transform()
2702       "vmul.f32 q3, q3, q6\n"  in Transform()
2703       "vadd.f32 q0, q0, q5\n"  in Transform()
2704       "vadd.f32 q1, q1, q5\n"  in Transform()
2705       "vadd.f32 q2, q2, q5\n"  in Transform()
2706       "vadd.f32 q3, q3, q5\n"  in Transform()
2707       "vcvt.s32.f32 q0, q0\n"  in Transform()
2708       "vcvt.s32.f32 q1, q1\n"  in Transform()
2709       "vcvt.s32.f32 q2, q2\n"  in Transform()
2710       "vcvt.s32.f32 q3, q3\n"  in Transform()
2711       "vqmovn.s32 d0, q0\n"  in Transform()
2712       "vqmovn.s32 d1, q1\n"  in Transform()
2713       "vqmovn.s32 d4, q2\n"  in Transform()
2714       "vqmovn.s32 d5, q3\n"  in Transform()
2715       "vqmovun.s16 d0, q0\n"  in Transform()
2716       "vqmovun.s16 d1, q2\n"  in Transform()
2718       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2719       "pld [%[output]]\n"  in Transform()
2721       "bne 1b\n"  in Transform()
2727       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2728       "vld1.32 {d4}, [%[input]]!\n"  in Transform()
2729       "vld1.32 {d5[0]}, [%[input]]!\n"  in Transform()
2730       "pld [%[input], #64]\n"  in Transform()
2731       "vsub.f32 q0, q0, q4\n"  in Transform()
2732       "vsub.f32 q1, q1, q4\n"  in Transform()
2733       "vsub.f32 q2, q2, q4\n"  in Transform()
2734       "vmul.f32 q0, q0, q6\n"  in Transform()
2735       "vmul.f32 q1, q1, q6\n"  in Transform()
2736       "vmul.f32 q2, q2, q6\n"  in Transform()
2737       "vadd.f32 q0, q0, q5\n"  in Transform()
2738       "vadd.f32 q1, q1, q5\n"  in Transform()
2739       "vadd.f32 q2, q2, q5\n"  in Transform()
2740       "vcvt.s32.f32 q0, q0\n"  in Transform()
2741       "vcvt.s32.f32 q1, q1\n"  in Transform()
2742       "vcvt.s32.f32 q2, q2\n"  in Transform()
2743       "vqmovn.s32 d0, q0\n"  in Transform()
2744       "vqmovn.s32 d1, q1\n"  in Transform()
2745       "vqmovn.s32 d4, q2\n"  in Transform()
2746       "vqmovun.s16 d0, q0\n"  in Transform()
2747       "vqmovun.s16 d1, q2\n"  in Transform()
2749       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
2750       "vst1.16 {d1[0]}, [%[output]]!\n"  in Transform()
2751       "vst1.8 {d1[2]}, [%[output]]!\n"  in Transform()
2752       "pld [%[output]]\n"  in Transform()
2776       "vdup.32 q4, %[range_min]\n"  in Transform()
2777       "vdup.32 q5, %[range_offset]\n"  in Transform()
2778       "vdup.32 q6, %[range_scale]\n"  in Transform()
2781       "subs %[count], %[count], #12\n"  in Transform()
2782       "beq 2f\n"  in Transform()
2785       "subs %[count], %[count], #16\n"  in Transform()
2788       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2789       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2790       "pld [%[input], #64]\n"  in Transform()
2791       "vsub.f32 q0, q0, q4\n"  in Transform()
2792       "vsub.f32 q1, q1, q4\n"  in Transform()
2793       "vsub.f32 q2, q2, q4\n"  in Transform()
2794       "vsub.f32 q3, q3, q4\n"  in Transform()
2795       "vmul.f32 q0, q0, q6\n"  in Transform()
2796       "vmul.f32 q1, q1, q6\n"  in Transform()
2797       "vmul.f32 q2, q2, q6\n"  in Transform()
2798       "vmul.f32 q3, q3, q6\n"  in Transform()
2799       "vadd.f32 q0, q0, q5\n"  in Transform()
2800       "vadd.f32 q1, q1, q5\n"  in Transform()
2801       "vadd.f32 q2, q2, q5\n"  in Transform()
2802       "vadd.f32 q3, q3, q5\n"  in Transform()
2803       "vcvt.s32.f32 q0, q0\n"  in Transform()
2804       "vcvt.s32.f32 q1, q1\n"  in Transform()
2805       "vcvt.s32.f32 q2, q2\n"  in Transform()
2806       "vcvt.s32.f32 q3, q3\n"  in Transform()
2807       "vqmovn.s32 d0, q0\n"  in Transform()
2808       "vqmovn.s32 d1, q1\n"  in Transform()
2809       "vqmovn.s32 d4, q2\n"  in Transform()
2810       "vqmovn.s32 d5, q3\n"  in Transform()
2811       "vqmovun.s16 d0, q0\n"  in Transform()
2812       "vqmovun.s16 d1, q2\n"  in Transform()
2814       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2815       "pld [%[output]]\n"  in Transform()
2817       "bne 1b\n"  in Transform()
2823       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2824       "vld1.32 {d4, d5}, [%[input]]!\n"  in Transform()
2825       "pld [%[input], #64]\n"  in Transform()
2826       "vsub.f32 q0, q0, q4\n"  in Transform()
2827       "vsub.f32 q1, q1, q4\n"  in Transform()
2828       "vsub.f32 q2, q2, q4\n"  in Transform()
2829       "vmul.f32 q0, q0, q6\n"  in Transform()
2830       "vmul.f32 q1, q1, q6\n"  in Transform()
2831       "vmul.f32 q2, q2, q6\n"  in Transform()
2832       "vadd.f32 q0, q0, q5\n"  in Transform()
2833       "vadd.f32 q1, q1, q5\n"  in Transform()
2834       "vadd.f32 q2, q2, q5\n"  in Transform()
2835       "vcvt.s32.f32 q0, q0\n"  in Transform()
2836       "vcvt.s32.f32 q1, q1\n"  in Transform()
2837       "vcvt.s32.f32 q2, q2\n"  in Transform()
2838       "vqmovn.s32 d0, q0\n"  in Transform()
2839       "vqmovn.s32 d1, q1\n"  in Transform()
2840       "vqmovn.s32 d4, q2\n"  in Transform()
2841       "vqmovun.s16 d0, q0\n"  in Transform()
2842       "vqmovun.s16 d1, q2\n"  in Transform()
2844       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
2845       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
2846       "pld [%[output]]\n"  in Transform()
2870       "vdup.32 q4, %[range_min]\n"  in Transform()
2871       "vdup.32 q5, %[range_offset]\n"  in Transform()
2872       "vdup.32 q6, %[range_scale]\n"  in Transform()
2875       "subs %[count], %[count], #13\n"  in Transform()
2876       "beq 2f\n"  in Transform()
2879       "subs %[count], %[count], #16\n"  in Transform()
2882       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2883       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2884       "pld [%[input], #64]\n"  in Transform()
2885       "vsub.f32 q0, q0, q4\n"  in Transform()
2886       "vsub.f32 q1, q1, q4\n"  in Transform()
2887       "vsub.f32 q2, q2, q4\n"  in Transform()
2888       "vsub.f32 q3, q3, q4\n"  in Transform()
2889       "vmul.f32 q0, q0, q6\n"  in Transform()
2890       "vmul.f32 q1, q1, q6\n"  in Transform()
2891       "vmul.f32 q2, q2, q6\n"  in Transform()
2892       "vmul.f32 q3, q3, q6\n"  in Transform()
2893       "vadd.f32 q0, q0, q5\n"  in Transform()
2894       "vadd.f32 q1, q1, q5\n"  in Transform()
2895       "vadd.f32 q2, q2, q5\n"  in Transform()
2896       "vadd.f32 q3, q3, q5\n"  in Transform()
2897       "vcvt.s32.f32 q0, q0\n"  in Transform()
2898       "vcvt.s32.f32 q1, q1\n"  in Transform()
2899       "vcvt.s32.f32 q2, q2\n"  in Transform()
2900       "vcvt.s32.f32 q3, q3\n"  in Transform()
2901       "vqmovn.s32 d0, q0\n"  in Transform()
2902       "vqmovn.s32 d1, q1\n"  in Transform()
2903       "vqmovn.s32 d4, q2\n"  in Transform()
2904       "vqmovn.s32 d5, q3\n"  in Transform()
2905       "vqmovun.s16 d0, q0\n"  in Transform()
2906       "vqmovun.s16 d1, q2\n"  in Transform()
2908       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
2909       "pld [%[output]]\n"  in Transform()
2911       "bne 1b\n"  in Transform()
2917       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2918       "vld1.32 {d4, d5}, [%[input]]!\n"  in Transform()
2919       "vld1.32 {d6[0]}, [%[input]]!\n"  in Transform()
2920       "pld [%[input], #64]\n"  in Transform()
2921       "vsub.f32 q0, q0, q4\n"  in Transform()
2922       "vsub.f32 q1, q1, q4\n"  in Transform()
2923       "vsub.f32 q2, q2, q4\n"  in Transform()
2924       "vsub.f32 q3, q3, q4\n"  in Transform()
2925       "vmul.f32 q0, q0, q6\n"  in Transform()
2926       "vmul.f32 q1, q1, q6\n"  in Transform()
2927       "vmul.f32 q2, q2, q6\n"  in Transform()
2928       "vmul.f32 q3, q3, q6\n"  in Transform()
2929       "vadd.f32 q0, q0, q5\n"  in Transform()
2930       "vadd.f32 q1, q1, q5\n"  in Transform()
2931       "vadd.f32 q2, q2, q5\n"  in Transform()
2932       "vadd.f32 q3, q3, q5\n"  in Transform()
2933       "vcvt.s32.f32 q0, q0\n"  in Transform()
2934       "vcvt.s32.f32 q1, q1\n"  in Transform()
2935       "vcvt.s32.f32 q2, q2\n"  in Transform()
2936       "vcvt.s32.f32 q3, q3\n"  in Transform()
2937       "vqmovn.s32 d0, q0\n"  in Transform()
2938       "vqmovn.s32 d1, q1\n"  in Transform()
2939       "vqmovn.s32 d4, q2\n"  in Transform()
2940       "vqmovn.s32 d5, q3\n"  in Transform()
2941       "vqmovun.s16 d0, q0\n"  in Transform()
2942       "vqmovun.s16 d1, q2\n"  in Transform()
2944       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
2945       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
2946       "vst1.8 {d1[4]}, [%[output]]!\n"  in Transform()
2947       "pld [%[output]]\n"  in Transform()
2971       "vdup.32 q4, %[range_min]\n"  in Transform()
2972       "vdup.32 q5, %[range_offset]\n"  in Transform()
2973       "vdup.32 q6, %[range_scale]\n"  in Transform()
2976       "subs %[count], %[count], #14\n"  in Transform()
2977       "beq 2f\n"  in Transform()
2980       "subs %[count], %[count], #16\n"  in Transform()
2983       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
2984       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
2985       "pld [%[input], #64]\n"  in Transform()
2986       "vsub.f32 q0, q0, q4\n"  in Transform()
2987       "vsub.f32 q1, q1, q4\n"  in Transform()
2988       "vsub.f32 q2, q2, q4\n"  in Transform()
2989       "vsub.f32 q3, q3, q4\n"  in Transform()
2990       "vmul.f32 q0, q0, q6\n"  in Transform()
2991       "vmul.f32 q1, q1, q6\n"  in Transform()
2992       "vmul.f32 q2, q2, q6\n"  in Transform()
2993       "vmul.f32 q3, q3, q6\n"  in Transform()
2994       "vadd.f32 q0, q0, q5\n"  in Transform()
2995       "vadd.f32 q1, q1, q5\n"  in Transform()
2996       "vadd.f32 q2, q2, q5\n"  in Transform()
2997       "vadd.f32 q3, q3, q5\n"  in Transform()
2998       "vcvt.s32.f32 q0, q0\n"  in Transform()
2999       "vcvt.s32.f32 q1, q1\n"  in Transform()
3000       "vcvt.s32.f32 q2, q2\n"  in Transform()
3001       "vcvt.s32.f32 q3, q3\n"  in Transform()
3002       "vqmovn.s32 d0, q0\n"  in Transform()
3003       "vqmovn.s32 d1, q1\n"  in Transform()
3004       "vqmovn.s32 d4, q2\n"  in Transform()
3005       "vqmovn.s32 d5, q3\n"  in Transform()
3006       "vqmovun.s16 d0, q0\n"  in Transform()
3007       "vqmovun.s16 d1, q2\n"  in Transform()
3009       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
3010       "pld [%[output]]\n"  in Transform()
3012       "bne 1b\n"  in Transform()
3018       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
3019       "vld1.32 {d4, d5, d6}, [%[input]]!\n"  in Transform()
3020       "pld [%[input], #64]\n"  in Transform()
3021       "vsub.f32 q0, q0, q4\n"  in Transform()
3022       "vsub.f32 q1, q1, q4\n"  in Transform()
3023       "vsub.f32 q2, q2, q4\n"  in Transform()
3024       "vsub.f32 q3, q3, q4\n"  in Transform()
3025       "vmul.f32 q0, q0, q6\n"  in Transform()
3026       "vmul.f32 q1, q1, q6\n"  in Transform()
3027       "vmul.f32 q2, q2, q6\n"  in Transform()
3028       "vmul.f32 q3, q3, q6\n"  in Transform()
3029       "vadd.f32 q0, q0, q5\n"  in Transform()
3030       "vadd.f32 q1, q1, q5\n"  in Transform()
3031       "vadd.f32 q2, q2, q5\n"  in Transform()
3032       "vadd.f32 q3, q3, q5\n"  in Transform()
3033       "vcvt.s32.f32 q0, q0\n"  in Transform()
3034       "vcvt.s32.f32 q1, q1\n"  in Transform()
3035       "vcvt.s32.f32 q2, q2\n"  in Transform()
3036       "vcvt.s32.f32 q3, q3\n"  in Transform()
3037       "vqmovn.s32 d0, q0\n"  in Transform()
3038       "vqmovn.s32 d1, q1\n"  in Transform()
3039       "vqmovn.s32 d4, q2\n"  in Transform()
3040       "vqmovn.s32 d5, q3\n"  in Transform()
3041       "vqmovun.s16 d0, q0\n"  in Transform()
3042       "vqmovun.s16 d1, q2\n"  in Transform()
3044       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
3045       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
3046       "vst1.16 {d1[2]}, [%[output]]!\n"  in Transform()
3047       "pld [%[output]]\n"  in Transform()
3071       "vdup.32 q4, %[range_min]\n"  in Transform()
3072       "vdup.32 q5, %[range_offset]\n"  in Transform()
3073       "vdup.32 q6, %[range_scale]\n"  in Transform()
3076       "subs %[count], %[count], #15\n"  in Transform()
3077       "beq 2f\n"  in Transform()
3080       "subs %[count], %[count], #16\n"  in Transform()
3083       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
3084       "vld1.32 {d4, d5, d6, d7}, [%[input]]!\n"  in Transform()
3085       "pld [%[input], #64]\n"  in Transform()
3086       "vsub.f32 q0, q0, q4\n"  in Transform()
3087       "vsub.f32 q1, q1, q4\n"  in Transform()
3088       "vsub.f32 q2, q2, q4\n"  in Transform()
3089       "vsub.f32 q3, q3, q4\n"  in Transform()
3090       "vmul.f32 q0, q0, q6\n"  in Transform()
3091       "vmul.f32 q1, q1, q6\n"  in Transform()
3092       "vmul.f32 q2, q2, q6\n"  in Transform()
3093       "vmul.f32 q3, q3, q6\n"  in Transform()
3094       "vadd.f32 q0, q0, q5\n"  in Transform()
3095       "vadd.f32 q1, q1, q5\n"  in Transform()
3096       "vadd.f32 q2, q2, q5\n"  in Transform()
3097       "vadd.f32 q3, q3, q5\n"  in Transform()
3098       "vcvt.s32.f32 q0, q0\n"  in Transform()
3099       "vcvt.s32.f32 q1, q1\n"  in Transform()
3100       "vcvt.s32.f32 q2, q2\n"  in Transform()
3101       "vcvt.s32.f32 q3, q3\n"  in Transform()
3102       "vqmovn.s32 d0, q0\n"  in Transform()
3103       "vqmovn.s32 d1, q1\n"  in Transform()
3104       "vqmovn.s32 d4, q2\n"  in Transform()
3105       "vqmovn.s32 d5, q3\n"  in Transform()
3106       "vqmovun.s16 d0, q0\n"  in Transform()
3107       "vqmovun.s16 d1, q2\n"  in Transform()
3109       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
3110       "pld [%[output]]\n"  in Transform()
3112       "bne 1b\n"  in Transform()
3118       "vld1.32 {d0, d1, d2, d3}, [%[input]]!\n"  in Transform()
3119       "vld1.32 {d4, d5, d6}, [%[input]]!\n"  in Transform()
3120       "vld1.32 {d7[0]}, [%[input]]!\n"  in Transform()
3121       "pld [%[input], #64]\n"  in Transform()
3122       "vsub.f32 q0, q0, q4\n"  in Transform()
3123       "vsub.f32 q1, q1, q4\n"  in Transform()
3124       "vsub.f32 q2, q2, q4\n"  in Transform()
3125       "vsub.f32 q3, q3, q4\n"  in Transform()
3126       "vmul.f32 q0, q0, q6\n"  in Transform()
3127       "vmul.f32 q1, q1, q6\n"  in Transform()
3128       "vmul.f32 q2, q2, q6\n"  in Transform()
3129       "vmul.f32 q3, q3, q6\n"  in Transform()
3130       "vadd.f32 q0, q0, q5\n"  in Transform()
3131       "vadd.f32 q1, q1, q5\n"  in Transform()
3132       "vadd.f32 q2, q2, q5\n"  in Transform()
3133       "vadd.f32 q3, q3, q5\n"  in Transform()
3134       "vcvt.s32.f32 q0, q0\n"  in Transform()
3135       "vcvt.s32.f32 q1, q1\n"  in Transform()
3136       "vcvt.s32.f32 q2, q2\n"  in Transform()
3137       "vcvt.s32.f32 q3, q3\n"  in Transform()
3138       "vqmovn.s32 d0, q0\n"  in Transform()
3139       "vqmovn.s32 d1, q1\n"  in Transform()
3140       "vqmovn.s32 d4, q2\n"  in Transform()
3141       "vqmovn.s32 d5, q3\n"  in Transform()
3142       "vqmovun.s16 d0, q0\n"  in Transform()
3143       "vqmovun.s16 d1, q2\n"  in Transform()
3145       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
3146       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
3147       "vst1.16 {d1[2]}, [%[output]]!\n"  in Transform()
3148       "vst1.8 {d1[6]}, [%[output]]!\n"  in Transform()
3149       "pld [%[output]]\n"  in Transform()
3173       "vdup.32 q4, %[range_min]\n"  in Transform()
3174       "vdup.32 q5, %[range_offset]\n"  in Transform()
3175       "vdup.32 q6, %[range_scale]\n"  in Transform()
3178       "subs %[count], %[count], #16\n"  in Transform()
3181       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3182       "pld [%[input], #32]\n"  in Transform()
3183       "vmovl.u8 q1, d1\n"  in Transform()
3184       "vmovl.u8 q0, d0\n"  in Transform()
3185       "vmovl.s16 q3, d3\n"  in Transform()
3186       "vmovl.s16 q2, d2\n"  in Transform()
3187       "vmovl.s16 q1, d1\n"  in Transform()
3188       "vmovl.s16 q0, d0\n"  in Transform()
3189       "vcvt.f32.s32 q0, q0\n"  in Transform()
3190       "vcvt.f32.s32 q1, q1\n"  in Transform()
3191       "vcvt.f32.s32 q2, q2\n"  in Transform()
3192       "vcvt.f32.s32 q3, q3\n"  in Transform()
3193       "vsub.f32 q0, q0, q5\n"  in Transform()
3194       "vsub.f32 q1, q1, q5\n"  in Transform()
3195       "vsub.f32 q2, q2, q5\n"  in Transform()
3196       "vsub.f32 q3, q3, q5\n"  in Transform()
3197       "vmul.f32 q0, q0, q6\n"  in Transform()
3198       "vmul.f32 q1, q1, q6\n"  in Transform()
3199       "vmul.f32 q2, q2, q6\n"  in Transform()
3200       "vmul.f32 q3, q3, q6\n"  in Transform()
3201       "vadd.f32 q0, q0, q4\n"  in Transform()
3202       "vadd.f32 q1, q1, q4\n"  in Transform()
3203       "vadd.f32 q2, q2, q4\n"  in Transform()
3204       "vadd.f32 q3, q3, q4\n"  in Transform()
3206       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3207       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3208       "pld [%[output]]\n"  in Transform()
3210       "bne 1b\n"  in Transform()
3234       "vdup.32 q4, %[range_min]\n"  in Transform()
3235       "vdup.32 q5, %[range_offset]\n"  in Transform()
3236       "vdup.32 q6, %[range_scale]\n"  in Transform()
3239       "subs %[count], %[count], #1\n"  in Transform()
3240       "beq 2f\n"  in Transform()
3243       "subs %[count], %[count], #16\n"  in Transform()
3246       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3247       "pld [%[input], #32]\n"  in Transform()
3248       "vmovl.u8 q1, d1\n"  in Transform()
3249       "vmovl.u8 q0, d0\n"  in Transform()
3250       "vmovl.s16 q3, d3\n"  in Transform()
3251       "vmovl.s16 q2, d2\n"  in Transform()
3252       "vmovl.s16 q1, d1\n"  in Transform()
3253       "vmovl.s16 q0, d0\n"  in Transform()
3254       "vcvt.f32.s32 q0, q0\n"  in Transform()
3255       "vcvt.f32.s32 q1, q1\n"  in Transform()
3256       "vcvt.f32.s32 q2, q2\n"  in Transform()
3257       "vcvt.f32.s32 q3, q3\n"  in Transform()
3258       "vsub.f32 q0, q0, q5\n"  in Transform()
3259       "vsub.f32 q1, q1, q5\n"  in Transform()
3260       "vsub.f32 q2, q2, q5\n"  in Transform()
3261       "vsub.f32 q3, q3, q5\n"  in Transform()
3262       "vmul.f32 q0, q0, q6\n"  in Transform()
3263       "vmul.f32 q1, q1, q6\n"  in Transform()
3264       "vmul.f32 q2, q2, q6\n"  in Transform()
3265       "vmul.f32 q3, q3, q6\n"  in Transform()
3266       "vadd.f32 q0, q0, q4\n"  in Transform()
3267       "vadd.f32 q1, q1, q4\n"  in Transform()
3268       "vadd.f32 q2, q2, q4\n"  in Transform()
3269       "vadd.f32 q3, q3, q4\n"  in Transform()
3271       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3272       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3273       "pld [%[output]]\n"  in Transform()
3275       "bne 1b\n"  in Transform()
3281       "vld1.8 {d0[0]}, [%[input]]!\n"  in Transform()
3282       "pld [%[input], #32]\n"  in Transform()
3283       "vmovl.u8 q0, d0\n"  in Transform()
3284       "vmovl.s16 q0, d0\n"  in Transform()
3285       "vcvt.f32.s32 q0, q0\n"  in Transform()
3286       "vsub.f32 q0, q0, q5\n"  in Transform()
3287       "vmul.f32 q0, q0, q6\n"  in Transform()
3288       "vadd.f32 q0, q0, q4\n"  in Transform()
3290       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
3291       "pld [%[output]]\n"  in Transform()
3315       "vdup.32 q4, %[range_min]\n"  in Transform()
3316       "vdup.32 q5, %[range_offset]\n"  in Transform()
3317       "vdup.32 q6, %[range_scale]\n"  in Transform()
3320       "subs %[count], %[count], #2\n"  in Transform()
3321       "beq 2f\n"  in Transform()
3324       "subs %[count], %[count], #16\n"  in Transform()
3327       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3328       "pld [%[input], #32]\n"  in Transform()
3329       "vmovl.u8 q1, d1\n"  in Transform()
3330       "vmovl.u8 q0, d0\n"  in Transform()
3331       "vmovl.s16 q3, d3\n"  in Transform()
3332       "vmovl.s16 q2, d2\n"  in Transform()
3333       "vmovl.s16 q1, d1\n"  in Transform()
3334       "vmovl.s16 q0, d0\n"  in Transform()
3335       "vcvt.f32.s32 q0, q0\n"  in Transform()
3336       "vcvt.f32.s32 q1, q1\n"  in Transform()
3337       "vcvt.f32.s32 q2, q2\n"  in Transform()
3338       "vcvt.f32.s32 q3, q3\n"  in Transform()
3339       "vsub.f32 q0, q0, q5\n"  in Transform()
3340       "vsub.f32 q1, q1, q5\n"  in Transform()
3341       "vsub.f32 q2, q2, q5\n"  in Transform()
3342       "vsub.f32 q3, q3, q5\n"  in Transform()
3343       "vmul.f32 q0, q0, q6\n"  in Transform()
3344       "vmul.f32 q1, q1, q6\n"  in Transform()
3345       "vmul.f32 q2, q2, q6\n"  in Transform()
3346       "vmul.f32 q3, q3, q6\n"  in Transform()
3347       "vadd.f32 q0, q0, q4\n"  in Transform()
3348       "vadd.f32 q1, q1, q4\n"  in Transform()
3349       "vadd.f32 q2, q2, q4\n"  in Transform()
3350       "vadd.f32 q3, q3, q4\n"  in Transform()
3352       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3353       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3354       "pld [%[output]]\n"  in Transform()
3356       "bne 1b\n"  in Transform()
3362       "vld1.16 {d0[0]}, [%[input]]!\n"  in Transform()
3363       "pld [%[input], #32]\n"  in Transform()
3364       "vmovl.u8 q0, d0\n"  in Transform()
3365       "vmovl.s16 q0, d0\n"  in Transform()
3366       "vcvt.f32.s32 q0, q0\n"  in Transform()
3367       "vsub.f32 q0, q0, q5\n"  in Transform()
3368       "vmul.f32 q0, q0, q6\n"  in Transform()
3369       "vadd.f32 q0, q0, q4\n"  in Transform()
3371       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
3372       "pld [%[output]]\n"  in Transform()
3396       "vdup.32 q4, %[range_min]\n"  in Transform()
3397       "vdup.32 q5, %[range_offset]\n"  in Transform()
3398       "vdup.32 q6, %[range_scale]\n"  in Transform()
3401       "subs %[count], %[count], #3\n"  in Transform()
3402       "beq 2f\n"  in Transform()
3405       "subs %[count], %[count], #16\n"  in Transform()
3408       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3409       "pld [%[input], #32]\n"  in Transform()
3410       "vmovl.u8 q1, d1\n"  in Transform()
3411       "vmovl.u8 q0, d0\n"  in Transform()
3412       "vmovl.s16 q3, d3\n"  in Transform()
3413       "vmovl.s16 q2, d2\n"  in Transform()
3414       "vmovl.s16 q1, d1\n"  in Transform()
3415       "vmovl.s16 q0, d0\n"  in Transform()
3416       "vcvt.f32.s32 q0, q0\n"  in Transform()
3417       "vcvt.f32.s32 q1, q1\n"  in Transform()
3418       "vcvt.f32.s32 q2, q2\n"  in Transform()
3419       "vcvt.f32.s32 q3, q3\n"  in Transform()
3420       "vsub.f32 q0, q0, q5\n"  in Transform()
3421       "vsub.f32 q1, q1, q5\n"  in Transform()
3422       "vsub.f32 q2, q2, q5\n"  in Transform()
3423       "vsub.f32 q3, q3, q5\n"  in Transform()
3424       "vmul.f32 q0, q0, q6\n"  in Transform()
3425       "vmul.f32 q1, q1, q6\n"  in Transform()
3426       "vmul.f32 q2, q2, q6\n"  in Transform()
3427       "vmul.f32 q3, q3, q6\n"  in Transform()
3428       "vadd.f32 q0, q0, q4\n"  in Transform()
3429       "vadd.f32 q1, q1, q4\n"  in Transform()
3430       "vadd.f32 q2, q2, q4\n"  in Transform()
3431       "vadd.f32 q3, q3, q4\n"  in Transform()
3433       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3434       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3435       "pld [%[output]]\n"  in Transform()
3437       "bne 1b\n"  in Transform()
3443       "vld1.16 {d0[0]}, [%[input]]!\n"  in Transform()
3444       "vld1.8 {d0[2]}, [%[input]]!\n"  in Transform()
3445       "pld [%[input], #32]\n"  in Transform()
3446       "vmovl.u8 q0, d0\n"  in Transform()
3447       "vmovl.s16 q0, d0\n"  in Transform()
3448       "vcvt.f32.s32 q0, q0\n"  in Transform()
3449       "vsub.f32 q0, q0, q5\n"  in Transform()
3450       "vmul.f32 q0, q0, q6\n"  in Transform()
3451       "vadd.f32 q0, q0, q4\n"  in Transform()
3453       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
3454       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
3455       "pld [%[output]]\n"  in Transform()
3479       "vdup.32 q4, %[range_min]\n"  in Transform()
3480       "vdup.32 q5, %[range_offset]\n"  in Transform()
3481       "vdup.32 q6, %[range_scale]\n"  in Transform()
3484       "subs %[count], %[count], #4\n"  in Transform()
3485       "beq 2f\n"  in Transform()
3488       "subs %[count], %[count], #16\n"  in Transform()
3491       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3492       "pld [%[input], #32]\n"  in Transform()
3493       "vmovl.u8 q1, d1\n"  in Transform()
3494       "vmovl.u8 q0, d0\n"  in Transform()
3495       "vmovl.s16 q3, d3\n"  in Transform()
3496       "vmovl.s16 q2, d2\n"  in Transform()
3497       "vmovl.s16 q1, d1\n"  in Transform()
3498       "vmovl.s16 q0, d0\n"  in Transform()
3499       "vcvt.f32.s32 q0, q0\n"  in Transform()
3500       "vcvt.f32.s32 q1, q1\n"  in Transform()
3501       "vcvt.f32.s32 q2, q2\n"  in Transform()
3502       "vcvt.f32.s32 q3, q3\n"  in Transform()
3503       "vsub.f32 q0, q0, q5\n"  in Transform()
3504       "vsub.f32 q1, q1, q5\n"  in Transform()
3505       "vsub.f32 q2, q2, q5\n"  in Transform()
3506       "vsub.f32 q3, q3, q5\n"  in Transform()
3507       "vmul.f32 q0, q0, q6\n"  in Transform()
3508       "vmul.f32 q1, q1, q6\n"  in Transform()
3509       "vmul.f32 q2, q2, q6\n"  in Transform()
3510       "vmul.f32 q3, q3, q6\n"  in Transform()
3511       "vadd.f32 q0, q0, q4\n"  in Transform()
3512       "vadd.f32 q1, q1, q4\n"  in Transform()
3513       "vadd.f32 q2, q2, q4\n"  in Transform()
3514       "vadd.f32 q3, q3, q4\n"  in Transform()
3516       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3517       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3518       "pld [%[output]]\n"  in Transform()
3520       "bne 1b\n"  in Transform()
3526       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
3527       "pld [%[input], #32]\n"  in Transform()
3528       "vmovl.u8 q0, d0\n"  in Transform()
3529       "vmovl.s16 q0, d0\n"  in Transform()
3530       "vcvt.f32.s32 q0, q0\n"  in Transform()
3531       "vsub.f32 q0, q0, q5\n"  in Transform()
3532       "vmul.f32 q0, q0, q6\n"  in Transform()
3533       "vadd.f32 q0, q0, q4\n"  in Transform()
3535       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
3536       "pld [%[output]]\n"  in Transform()
3560       "vdup.32 q4, %[range_min]\n"  in Transform()
3561       "vdup.32 q5, %[range_offset]\n"  in Transform()
3562       "vdup.32 q6, %[range_scale]\n"  in Transform()
3565       "subs %[count], %[count], #5\n"  in Transform()
3566       "beq 2f\n"  in Transform()
3569       "subs %[count], %[count], #16\n"  in Transform()
3572       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3573       "pld [%[input], #32]\n"  in Transform()
3574       "vmovl.u8 q1, d1\n"  in Transform()
3575       "vmovl.u8 q0, d0\n"  in Transform()
3576       "vmovl.s16 q3, d3\n"  in Transform()
3577       "vmovl.s16 q2, d2\n"  in Transform()
3578       "vmovl.s16 q1, d1\n"  in Transform()
3579       "vmovl.s16 q0, d0\n"  in Transform()
3580       "vcvt.f32.s32 q0, q0\n"  in Transform()
3581       "vcvt.f32.s32 q1, q1\n"  in Transform()
3582       "vcvt.f32.s32 q2, q2\n"  in Transform()
3583       "vcvt.f32.s32 q3, q3\n"  in Transform()
3584       "vsub.f32 q0, q0, q5\n"  in Transform()
3585       "vsub.f32 q1, q1, q5\n"  in Transform()
3586       "vsub.f32 q2, q2, q5\n"  in Transform()
3587       "vsub.f32 q3, q3, q5\n"  in Transform()
3588       "vmul.f32 q0, q0, q6\n"  in Transform()
3589       "vmul.f32 q1, q1, q6\n"  in Transform()
3590       "vmul.f32 q2, q2, q6\n"  in Transform()
3591       "vmul.f32 q3, q3, q6\n"  in Transform()
3592       "vadd.f32 q0, q0, q4\n"  in Transform()
3593       "vadd.f32 q1, q1, q4\n"  in Transform()
3594       "vadd.f32 q2, q2, q4\n"  in Transform()
3595       "vadd.f32 q3, q3, q4\n"  in Transform()
3597       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3598       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3599       "pld [%[output]]\n"  in Transform()
3601       "bne 1b\n"  in Transform()
3607       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
3608       "vld1.8 {d0[4]}, [%[input]]!\n"  in Transform()
3609       "pld [%[input], #32]\n"  in Transform()
3610       "vmovl.u8 q0, d0\n"  in Transform()
3611       "vmovl.s16 q1, d1\n"  in Transform()
3612       "vmovl.s16 q0, d0\n"  in Transform()
3613       "vcvt.f32.s32 q0, q0\n"  in Transform()
3614       "vcvt.f32.s32 q1, q1\n"  in Transform()
3615       "vsub.f32 q0, q0, q5\n"  in Transform()
3616       "vsub.f32 q1, q1, q5\n"  in Transform()
3617       "vmul.f32 q0, q0, q6\n"  in Transform()
3618       "vmul.f32 q1, q1, q6\n"  in Transform()
3619       "vadd.f32 q0, q0, q4\n"  in Transform()
3620       "vadd.f32 q1, q1, q4\n"  in Transform()
3622       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
3623       "vst1.32 {d2[0]}, [%[output]]!\n"  in Transform()
3624       "pld [%[output]]\n"  in Transform()
3648       "vdup.32 q4, %[range_min]\n"  in Transform()
3649       "vdup.32 q5, %[range_offset]\n"  in Transform()
3650       "vdup.32 q6, %[range_scale]\n"  in Transform()
3653       "subs %[count], %[count], #6\n"  in Transform()
3654       "beq 2f\n"  in Transform()
3657       "subs %[count], %[count], #16\n"  in Transform()
3660       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3661       "pld [%[input], #32]\n"  in Transform()
3662       "vmovl.u8 q1, d1\n"  in Transform()
3663       "vmovl.u8 q0, d0\n"  in Transform()
3664       "vmovl.s16 q3, d3\n"  in Transform()
3665       "vmovl.s16 q2, d2\n"  in Transform()
3666       "vmovl.s16 q1, d1\n"  in Transform()
3667       "vmovl.s16 q0, d0\n"  in Transform()
3668       "vcvt.f32.s32 q0, q0\n"  in Transform()
3669       "vcvt.f32.s32 q1, q1\n"  in Transform()
3670       "vcvt.f32.s32 q2, q2\n"  in Transform()
3671       "vcvt.f32.s32 q3, q3\n"  in Transform()
3672       "vsub.f32 q0, q0, q5\n"  in Transform()
3673       "vsub.f32 q1, q1, q5\n"  in Transform()
3674       "vsub.f32 q2, q2, q5\n"  in Transform()
3675       "vsub.f32 q3, q3, q5\n"  in Transform()
3676       "vmul.f32 q0, q0, q6\n"  in Transform()
3677       "vmul.f32 q1, q1, q6\n"  in Transform()
3678       "vmul.f32 q2, q2, q6\n"  in Transform()
3679       "vmul.f32 q3, q3, q6\n"  in Transform()
3680       "vadd.f32 q0, q0, q4\n"  in Transform()
3681       "vadd.f32 q1, q1, q4\n"  in Transform()
3682       "vadd.f32 q2, q2, q4\n"  in Transform()
3683       "vadd.f32 q3, q3, q4\n"  in Transform()
3685       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3686       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3687       "pld [%[output]]\n"  in Transform()
3689       "bne 1b\n"  in Transform()
3695       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
3696       "vld1.16 {d0[2]}, [%[input]]!\n"  in Transform()
3697       "pld [%[input], #32]\n"  in Transform()
3698       "vmovl.u8 q0, d0\n"  in Transform()
3699       "vmovl.s16 q1, d1\n"  in Transform()
3700       "vmovl.s16 q0, d0\n"  in Transform()
3701       "vcvt.f32.s32 q0, q0\n"  in Transform()
3702       "vcvt.f32.s32 q1, q1\n"  in Transform()
3703       "vsub.f32 q0, q0, q5\n"  in Transform()
3704       "vsub.f32 q1, q1, q5\n"  in Transform()
3705       "vmul.f32 q0, q0, q6\n"  in Transform()
3706       "vmul.f32 q1, q1, q6\n"  in Transform()
3707       "vadd.f32 q0, q0, q4\n"  in Transform()
3708       "vadd.f32 q1, q1, q4\n"  in Transform()
3710       "vst1.32 {d0, d1, d2}, [%[output]]!\n"  in Transform()
3711       "pld [%[output]]\n"  in Transform()
3735       "vdup.32 q4, %[range_min]\n"  in Transform()
3736       "vdup.32 q5, %[range_offset]\n"  in Transform()
3737       "vdup.32 q6, %[range_scale]\n"  in Transform()
3740       "subs %[count], %[count], #7\n"  in Transform()
3741       "beq 2f\n"  in Transform()
3744       "subs %[count], %[count], #16\n"  in Transform()
3747       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3748       "pld [%[input], #32]\n"  in Transform()
3749       "vmovl.u8 q1, d1\n"  in Transform()
3750       "vmovl.u8 q0, d0\n"  in Transform()
3751       "vmovl.s16 q3, d3\n"  in Transform()
3752       "vmovl.s16 q2, d2\n"  in Transform()
3753       "vmovl.s16 q1, d1\n"  in Transform()
3754       "vmovl.s16 q0, d0\n"  in Transform()
3755       "vcvt.f32.s32 q0, q0\n"  in Transform()
3756       "vcvt.f32.s32 q1, q1\n"  in Transform()
3757       "vcvt.f32.s32 q2, q2\n"  in Transform()
3758       "vcvt.f32.s32 q3, q3\n"  in Transform()
3759       "vsub.f32 q0, q0, q5\n"  in Transform()
3760       "vsub.f32 q1, q1, q5\n"  in Transform()
3761       "vsub.f32 q2, q2, q5\n"  in Transform()
3762       "vsub.f32 q3, q3, q5\n"  in Transform()
3763       "vmul.f32 q0, q0, q6\n"  in Transform()
3764       "vmul.f32 q1, q1, q6\n"  in Transform()
3765       "vmul.f32 q2, q2, q6\n"  in Transform()
3766       "vmul.f32 q3, q3, q6\n"  in Transform()
3767       "vadd.f32 q0, q0, q4\n"  in Transform()
3768       "vadd.f32 q1, q1, q4\n"  in Transform()
3769       "vadd.f32 q2, q2, q4\n"  in Transform()
3770       "vadd.f32 q3, q3, q4\n"  in Transform()
3772       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3773       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3774       "pld [%[output]]\n"  in Transform()
3776       "bne 1b\n"  in Transform()
3782       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
3783       "vld1.16 {d0[2]}, [%[input]]!\n"  in Transform()
3784       "vld1.8 {d0[6]}, [%[input]]!\n"  in Transform()
3785       "pld [%[input], #32]\n"  in Transform()
3786       "vmovl.u8 q0, d0\n"  in Transform()
3787       "vmovl.s16 q1, d1\n"  in Transform()
3788       "vmovl.s16 q0, d0\n"  in Transform()
3789       "vcvt.f32.s32 q0, q0\n"  in Transform()
3790       "vcvt.f32.s32 q1, q1\n"  in Transform()
3791       "vsub.f32 q0, q0, q5\n"  in Transform()
3792       "vsub.f32 q1, q1, q5\n"  in Transform()
3793       "vmul.f32 q0, q0, q6\n"  in Transform()
3794       "vmul.f32 q1, q1, q6\n"  in Transform()
3795       "vadd.f32 q0, q0, q4\n"  in Transform()
3796       "vadd.f32 q1, q1, q4\n"  in Transform()
3798       "vst1.32 {d0, d1, d2}, [%[output]]!\n"  in Transform()
3799       "vst1.32 {d3[0]}, [%[output]]!\n"  in Transform()
3800       "pld [%[output]]\n"  in Transform()
3824       "vdup.32 q4, %[range_min]\n"  in Transform()
3825       "vdup.32 q5, %[range_offset]\n"  in Transform()
3826       "vdup.32 q6, %[range_scale]\n"  in Transform()
3829       "subs %[count], %[count], #8\n"  in Transform()
3830       "beq 2f\n"  in Transform()
3833       "subs %[count], %[count], #16\n"  in Transform()
3836       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3837       "pld [%[input], #32]\n"  in Transform()
3838       "vmovl.u8 q1, d1\n"  in Transform()
3839       "vmovl.u8 q0, d0\n"  in Transform()
3840       "vmovl.s16 q3, d3\n"  in Transform()
3841       "vmovl.s16 q2, d2\n"  in Transform()
3842       "vmovl.s16 q1, d1\n"  in Transform()
3843       "vmovl.s16 q0, d0\n"  in Transform()
3844       "vcvt.f32.s32 q0, q0\n"  in Transform()
3845       "vcvt.f32.s32 q1, q1\n"  in Transform()
3846       "vcvt.f32.s32 q2, q2\n"  in Transform()
3847       "vcvt.f32.s32 q3, q3\n"  in Transform()
3848       "vsub.f32 q0, q0, q5\n"  in Transform()
3849       "vsub.f32 q1, q1, q5\n"  in Transform()
3850       "vsub.f32 q2, q2, q5\n"  in Transform()
3851       "vsub.f32 q3, q3, q5\n"  in Transform()
3852       "vmul.f32 q0, q0, q6\n"  in Transform()
3853       "vmul.f32 q1, q1, q6\n"  in Transform()
3854       "vmul.f32 q2, q2, q6\n"  in Transform()
3855       "vmul.f32 q3, q3, q6\n"  in Transform()
3856       "vadd.f32 q0, q0, q4\n"  in Transform()
3857       "vadd.f32 q1, q1, q4\n"  in Transform()
3858       "vadd.f32 q2, q2, q4\n"  in Transform()
3859       "vadd.f32 q3, q3, q4\n"  in Transform()
3861       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3862       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3863       "pld [%[output]]\n"  in Transform()
3865       "bne 1b\n"  in Transform()
3871       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
3872       "pld [%[input], #32]\n"  in Transform()
3873       "vmovl.u8 q0, d0\n"  in Transform()
3874       "vmovl.s16 q1, d1\n"  in Transform()
3875       "vmovl.s16 q0, d0\n"  in Transform()
3876       "vcvt.f32.s32 q0, q0\n"  in Transform()
3877       "vcvt.f32.s32 q1, q1\n"  in Transform()
3878       "vsub.f32 q0, q0, q5\n"  in Transform()
3879       "vsub.f32 q1, q1, q5\n"  in Transform()
3880       "vmul.f32 q0, q0, q6\n"  in Transform()
3881       "vmul.f32 q1, q1, q6\n"  in Transform()
3882       "vadd.f32 q0, q0, q4\n"  in Transform()
3883       "vadd.f32 q1, q1, q4\n"  in Transform()
3885       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3886       "pld [%[output]]\n"  in Transform()
3910       "vdup.32 q4, %[range_min]\n"  in Transform()
3911       "vdup.32 q5, %[range_offset]\n"  in Transform()
3912       "vdup.32 q6, %[range_scale]\n"  in Transform()
3915       "subs %[count], %[count], #9\n"  in Transform()
3916       "beq 2f\n"  in Transform()
3919       "subs %[count], %[count], #16\n"  in Transform()
3922       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
3923       "pld [%[input], #32]\n"  in Transform()
3924       "vmovl.u8 q1, d1\n"  in Transform()
3925       "vmovl.u8 q0, d0\n"  in Transform()
3926       "vmovl.s16 q3, d3\n"  in Transform()
3927       "vmovl.s16 q2, d2\n"  in Transform()
3928       "vmovl.s16 q1, d1\n"  in Transform()
3929       "vmovl.s16 q0, d0\n"  in Transform()
3930       "vcvt.f32.s32 q0, q0\n"  in Transform()
3931       "vcvt.f32.s32 q1, q1\n"  in Transform()
3932       "vcvt.f32.s32 q2, q2\n"  in Transform()
3933       "vcvt.f32.s32 q3, q3\n"  in Transform()
3934       "vsub.f32 q0, q0, q5\n"  in Transform()
3935       "vsub.f32 q1, q1, q5\n"  in Transform()
3936       "vsub.f32 q2, q2, q5\n"  in Transform()
3937       "vsub.f32 q3, q3, q5\n"  in Transform()
3938       "vmul.f32 q0, q0, q6\n"  in Transform()
3939       "vmul.f32 q1, q1, q6\n"  in Transform()
3940       "vmul.f32 q2, q2, q6\n"  in Transform()
3941       "vmul.f32 q3, q3, q6\n"  in Transform()
3942       "vadd.f32 q0, q0, q4\n"  in Transform()
3943       "vadd.f32 q1, q1, q4\n"  in Transform()
3944       "vadd.f32 q2, q2, q4\n"  in Transform()
3945       "vadd.f32 q3, q3, q4\n"  in Transform()
3947       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3948       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
3949       "pld [%[output]]\n"  in Transform()
3951       "bne 1b\n"  in Transform()
3957       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
3958       "vld1.8 {d1[0]}, [%[input]]!\n"  in Transform()
3959       "pld [%[input], #32]\n"  in Transform()
3960       "vmovl.u8 q1, d1\n"  in Transform()
3961       "vmovl.u8 q0, d0\n"  in Transform()
3962       "vmovl.s16 q2, d2\n"  in Transform()
3963       "vmovl.s16 q1, d1\n"  in Transform()
3964       "vmovl.s16 q0, d0\n"  in Transform()
3965       "vcvt.f32.s32 q0, q0\n"  in Transform()
3966       "vcvt.f32.s32 q1, q1\n"  in Transform()
3967       "vcvt.f32.s32 q2, q2\n"  in Transform()
3968       "vsub.f32 q0, q0, q5\n"  in Transform()
3969       "vsub.f32 q1, q1, q5\n"  in Transform()
3970       "vsub.f32 q2, q2, q5\n"  in Transform()
3971       "vmul.f32 q0, q0, q6\n"  in Transform()
3972       "vmul.f32 q1, q1, q6\n"  in Transform()
3973       "vmul.f32 q2, q2, q6\n"  in Transform()
3974       "vadd.f32 q0, q0, q4\n"  in Transform()
3975       "vadd.f32 q1, q1, q4\n"  in Transform()
3976       "vadd.f32 q2, q2, q4\n"  in Transform()
3978       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
3979       "vst1.32 {d4[0]}, [%[output]]!\n"  in Transform()
3980       "pld [%[output]]\n"  in Transform()
4004       "vdup.32 q4, %[range_min]\n"  in Transform()
4005       "vdup.32 q5, %[range_offset]\n"  in Transform()
4006       "vdup.32 q6, %[range_scale]\n"  in Transform()
4009       "subs %[count], %[count], #10\n"  in Transform()
4010       "beq 2f\n"  in Transform()
4013       "subs %[count], %[count], #16\n"  in Transform()
4016       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4017       "pld [%[input], #32]\n"  in Transform()
4018       "vmovl.u8 q1, d1\n"  in Transform()
4019       "vmovl.u8 q0, d0\n"  in Transform()
4020       "vmovl.s16 q3, d3\n"  in Transform()
4021       "vmovl.s16 q2, d2\n"  in Transform()
4022       "vmovl.s16 q1, d1\n"  in Transform()
4023       "vmovl.s16 q0, d0\n"  in Transform()
4024       "vcvt.f32.s32 q0, q0\n"  in Transform()
4025       "vcvt.f32.s32 q1, q1\n"  in Transform()
4026       "vcvt.f32.s32 q2, q2\n"  in Transform()
4027       "vcvt.f32.s32 q3, q3\n"  in Transform()
4028       "vsub.f32 q0, q0, q5\n"  in Transform()
4029       "vsub.f32 q1, q1, q5\n"  in Transform()
4030       "vsub.f32 q2, q2, q5\n"  in Transform()
4031       "vsub.f32 q3, q3, q5\n"  in Transform()
4032       "vmul.f32 q0, q0, q6\n"  in Transform()
4033       "vmul.f32 q1, q1, q6\n"  in Transform()
4034       "vmul.f32 q2, q2, q6\n"  in Transform()
4035       "vmul.f32 q3, q3, q6\n"  in Transform()
4036       "vadd.f32 q0, q0, q4\n"  in Transform()
4037       "vadd.f32 q1, q1, q4\n"  in Transform()
4038       "vadd.f32 q2, q2, q4\n"  in Transform()
4039       "vadd.f32 q3, q3, q4\n"  in Transform()
4041       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4042       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
4043       "pld [%[output]]\n"  in Transform()
4045       "bne 1b\n"  in Transform()
4051       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
4052       "vld1.16 {d1[0]}, [%[input]]!\n"  in Transform()
4053       "pld [%[input], #32]\n"  in Transform()
4054       "vmovl.u8 q1, d1\n"  in Transform()
4055       "vmovl.u8 q0, d0\n"  in Transform()
4056       "vmovl.s16 q2, d2\n"  in Transform()
4057       "vmovl.s16 q1, d1\n"  in Transform()
4058       "vmovl.s16 q0, d0\n"  in Transform()
4059       "vcvt.f32.s32 q0, q0\n"  in Transform()
4060       "vcvt.f32.s32 q1, q1\n"  in Transform()
4061       "vcvt.f32.s32 q2, q2\n"  in Transform()
4062       "vsub.f32 q0, q0, q5\n"  in Transform()
4063       "vsub.f32 q1, q1, q5\n"  in Transform()
4064       "vsub.f32 q2, q2, q5\n"  in Transform()
4065       "vmul.f32 q0, q0, q6\n"  in Transform()
4066       "vmul.f32 q1, q1, q6\n"  in Transform()
4067       "vmul.f32 q2, q2, q6\n"  in Transform()
4068       "vadd.f32 q0, q0, q4\n"  in Transform()
4069       "vadd.f32 q1, q1, q4\n"  in Transform()
4070       "vadd.f32 q2, q2, q4\n"  in Transform()
4072       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4073       "vst1.32 {d4}, [%[output]]!\n"  in Transform()
4074       "pld [%[output]]\n"  in Transform()
4098       "vdup.32 q4, %[range_min]\n"  in Transform()
4099       "vdup.32 q5, %[range_offset]\n"  in Transform()
4100       "vdup.32 q6, %[range_scale]\n"  in Transform()
4103       "subs %[count], %[count], #11\n"  in Transform()
4104       "beq 2f\n"  in Transform()
4107       "subs %[count], %[count], #16\n"  in Transform()
4110       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4111       "pld [%[input], #32]\n"  in Transform()
4112       "vmovl.u8 q1, d1\n"  in Transform()
4113       "vmovl.u8 q0, d0\n"  in Transform()
4114       "vmovl.s16 q3, d3\n"  in Transform()
4115       "vmovl.s16 q2, d2\n"  in Transform()
4116       "vmovl.s16 q1, d1\n"  in Transform()
4117       "vmovl.s16 q0, d0\n"  in Transform()
4118       "vcvt.f32.s32 q0, q0\n"  in Transform()
4119       "vcvt.f32.s32 q1, q1\n"  in Transform()
4120       "vcvt.f32.s32 q2, q2\n"  in Transform()
4121       "vcvt.f32.s32 q3, q3\n"  in Transform()
4122       "vsub.f32 q0, q0, q5\n"  in Transform()
4123       "vsub.f32 q1, q1, q5\n"  in Transform()
4124       "vsub.f32 q2, q2, q5\n"  in Transform()
4125       "vsub.f32 q3, q3, q5\n"  in Transform()
4126       "vmul.f32 q0, q0, q6\n"  in Transform()
4127       "vmul.f32 q1, q1, q6\n"  in Transform()
4128       "vmul.f32 q2, q2, q6\n"  in Transform()
4129       "vmul.f32 q3, q3, q6\n"  in Transform()
4130       "vadd.f32 q0, q0, q4\n"  in Transform()
4131       "vadd.f32 q1, q1, q4\n"  in Transform()
4132       "vadd.f32 q2, q2, q4\n"  in Transform()
4133       "vadd.f32 q3, q3, q4\n"  in Transform()
4135       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4136       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
4137       "pld [%[output]]\n"  in Transform()
4139       "bne 1b\n"  in Transform()
4145       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
4146       "vld1.16 {d1[0]}, [%[input]]!\n"  in Transform()
4147       "vld1.8 {d1[2]}, [%[input]]!\n"  in Transform()
4148       "pld [%[input], #32]\n"  in Transform()
4149       "vmovl.u8 q1, d1\n"  in Transform()
4150       "vmovl.u8 q0, d0\n"  in Transform()
4151       "vmovl.s16 q2, d2\n"  in Transform()
4152       "vmovl.s16 q1, d1\n"  in Transform()
4153       "vmovl.s16 q0, d0\n"  in Transform()
4154       "vcvt.f32.s32 q0, q0\n"  in Transform()
4155       "vcvt.f32.s32 q1, q1\n"  in Transform()
4156       "vcvt.f32.s32 q2, q2\n"  in Transform()
4157       "vsub.f32 q0, q0, q5\n"  in Transform()
4158       "vsub.f32 q1, q1, q5\n"  in Transform()
4159       "vsub.f32 q2, q2, q5\n"  in Transform()
4160       "vmul.f32 q0, q0, q6\n"  in Transform()
4161       "vmul.f32 q1, q1, q6\n"  in Transform()
4162       "vmul.f32 q2, q2, q6\n"  in Transform()
4163       "vadd.f32 q0, q0, q4\n"  in Transform()
4164       "vadd.f32 q1, q1, q4\n"  in Transform()
4165       "vadd.f32 q2, q2, q4\n"  in Transform()
4167       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4168       "vst1.32 {d4}, [%[output]]!\n"  in Transform()
4169       "vst1.32 {d5[0]}, [%[output]]!\n"  in Transform()
4170       "pld [%[output]]\n"  in Transform()
4194       "vdup.32 q4, %[range_min]\n"  in Transform()
4195       "vdup.32 q5, %[range_offset]\n"  in Transform()
4196       "vdup.32 q6, %[range_scale]\n"  in Transform()
4199       "subs %[count], %[count], #12\n"  in Transform()
4200       "beq 2f\n"  in Transform()
4203       "subs %[count], %[count], #16\n"  in Transform()
4206       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4207       "pld [%[input], #32]\n"  in Transform()
4208       "vmovl.u8 q1, d1\n"  in Transform()
4209       "vmovl.u8 q0, d0\n"  in Transform()
4210       "vmovl.s16 q3, d3\n"  in Transform()
4211       "vmovl.s16 q2, d2\n"  in Transform()
4212       "vmovl.s16 q1, d1\n"  in Transform()
4213       "vmovl.s16 q0, d0\n"  in Transform()
4214       "vcvt.f32.s32 q0, q0\n"  in Transform()
4215       "vcvt.f32.s32 q1, q1\n"  in Transform()
4216       "vcvt.f32.s32 q2, q2\n"  in Transform()
4217       "vcvt.f32.s32 q3, q3\n"  in Transform()
4218       "vsub.f32 q0, q0, q5\n"  in Transform()
4219       "vsub.f32 q1, q1, q5\n"  in Transform()
4220       "vsub.f32 q2, q2, q5\n"  in Transform()
4221       "vsub.f32 q3, q3, q5\n"  in Transform()
4222       "vmul.f32 q0, q0, q6\n"  in Transform()
4223       "vmul.f32 q1, q1, q6\n"  in Transform()
4224       "vmul.f32 q2, q2, q6\n"  in Transform()
4225       "vmul.f32 q3, q3, q6\n"  in Transform()
4226       "vadd.f32 q0, q0, q4\n"  in Transform()
4227       "vadd.f32 q1, q1, q4\n"  in Transform()
4228       "vadd.f32 q2, q2, q4\n"  in Transform()
4229       "vadd.f32 q3, q3, q4\n"  in Transform()
4231       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4232       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
4233       "pld [%[output]]\n"  in Transform()
4235       "bne 1b\n"  in Transform()
4241       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
4242       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
4243       "pld [%[input], #32]\n"  in Transform()
4244       "vmovl.u8 q1, d1\n"  in Transform()
4245       "vmovl.u8 q0, d0\n"  in Transform()
4246       "vmovl.s16 q2, d2\n"  in Transform()
4247       "vmovl.s16 q1, d1\n"  in Transform()
4248       "vmovl.s16 q0, d0\n"  in Transform()
4249       "vcvt.f32.s32 q0, q0\n"  in Transform()
4250       "vcvt.f32.s32 q1, q1\n"  in Transform()
4251       "vcvt.f32.s32 q2, q2\n"  in Transform()
4252       "vsub.f32 q0, q0, q5\n"  in Transform()
4253       "vsub.f32 q1, q1, q5\n"  in Transform()
4254       "vsub.f32 q2, q2, q5\n"  in Transform()
4255       "vmul.f32 q0, q0, q6\n"  in Transform()
4256       "vmul.f32 q1, q1, q6\n"  in Transform()
4257       "vmul.f32 q2, q2, q6\n"  in Transform()
4258       "vadd.f32 q0, q0, q4\n"  in Transform()
4259       "vadd.f32 q1, q1, q4\n"  in Transform()
4260       "vadd.f32 q2, q2, q4\n"  in Transform()
4262       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4263       "vst1.32 {d4, d5}, [%[output]]!\n"  in Transform()
4264       "pld [%[output]]\n"  in Transform()
4288       "vdup.32 q4, %[range_min]\n"  in Transform()
4289       "vdup.32 q5, %[range_offset]\n"  in Transform()
4290       "vdup.32 q6, %[range_scale]\n"  in Transform()
4293       "subs %[count], %[count], #13\n"  in Transform()
4294       "beq 2f\n"  in Transform()
4297       "subs %[count], %[count], #16\n"  in Transform()
4300       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4301       "pld [%[input], #32]\n"  in Transform()
4302       "vmovl.u8 q1, d1\n"  in Transform()
4303       "vmovl.u8 q0, d0\n"  in Transform()
4304       "vmovl.s16 q3, d3\n"  in Transform()
4305       "vmovl.s16 q2, d2\n"  in Transform()
4306       "vmovl.s16 q1, d1\n"  in Transform()
4307       "vmovl.s16 q0, d0\n"  in Transform()
4308       "vcvt.f32.s32 q0, q0\n"  in Transform()
4309       "vcvt.f32.s32 q1, q1\n"  in Transform()
4310       "vcvt.f32.s32 q2, q2\n"  in Transform()
4311       "vcvt.f32.s32 q3, q3\n"  in Transform()
4312       "vsub.f32 q0, q0, q5\n"  in Transform()
4313       "vsub.f32 q1, q1, q5\n"  in Transform()
4314       "vsub.f32 q2, q2, q5\n"  in Transform()
4315       "vsub.f32 q3, q3, q5\n"  in Transform()
4316       "vmul.f32 q0, q0, q6\n"  in Transform()
4317       "vmul.f32 q1, q1, q6\n"  in Transform()
4318       "vmul.f32 q2, q2, q6\n"  in Transform()
4319       "vmul.f32 q3, q3, q6\n"  in Transform()
4320       "vadd.f32 q0, q0, q4\n"  in Transform()
4321       "vadd.f32 q1, q1, q4\n"  in Transform()
4322       "vadd.f32 q2, q2, q4\n"  in Transform()
4323       "vadd.f32 q3, q3, q4\n"  in Transform()
4325       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4326       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
4327       "pld [%[output]]\n"  in Transform()
4329       "bne 1b\n"  in Transform()
4335       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
4336       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
4337       "vld1.8 {d1[4]}, [%[input]]!\n"  in Transform()
4338       "pld [%[input], #32]\n"  in Transform()
4339       "vmovl.u8 q1, d1\n"  in Transform()
4340       "vmovl.u8 q0, d0\n"  in Transform()
4341       "vmovl.s16 q3, d3\n"  in Transform()
4342       "vmovl.s16 q2, d2\n"  in Transform()
4343       "vmovl.s16 q1, d1\n"  in Transform()
4344       "vmovl.s16 q0, d0\n"  in Transform()
4345       "vcvt.f32.s32 q0, q0\n"  in Transform()
4346       "vcvt.f32.s32 q1, q1\n"  in Transform()
4347       "vcvt.f32.s32 q2, q2\n"  in Transform()
4348       "vcvt.f32.s32 q3, q3\n"  in Transform()
4349       "vsub.f32 q0, q0, q5\n"  in Transform()
4350       "vsub.f32 q1, q1, q5\n"  in Transform()
4351       "vsub.f32 q2, q2, q5\n"  in Transform()
4352       "vsub.f32 q3, q3, q5\n"  in Transform()
4353       "vmul.f32 q0, q0, q6\n"  in Transform()
4354       "vmul.f32 q1, q1, q6\n"  in Transform()
4355       "vmul.f32 q2, q2, q6\n"  in Transform()
4356       "vmul.f32 q3, q3, q6\n"  in Transform()
4357       "vadd.f32 q0, q0, q4\n"  in Transform()
4358       "vadd.f32 q1, q1, q4\n"  in Transform()
4359       "vadd.f32 q2, q2, q4\n"  in Transform()
4360       "vadd.f32 q3, q3, q4\n"  in Transform()
4362       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4363       "vst1.32 {d4, d5}, [%[output]]!\n"  in Transform()
4364       "vst1.32 {d6[0]}, [%[output]]!\n"  in Transform()
4365       "pld [%[output]]\n"  in Transform()
4389       "vdup.32 q4, %[range_min]\n"  in Transform()
4390       "vdup.32 q5, %[range_offset]\n"  in Transform()
4391       "vdup.32 q6, %[range_scale]\n"  in Transform()
4394       "subs %[count], %[count], #14\n"  in Transform()
4395       "beq 2f\n"  in Transform()
4398       "subs %[count], %[count], #16\n"  in Transform()
4401       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4402       "pld [%[input], #32]\n"  in Transform()
4403       "vmovl.u8 q1, d1\n"  in Transform()
4404       "vmovl.u8 q0, d0\n"  in Transform()
4405       "vmovl.s16 q3, d3\n"  in Transform()
4406       "vmovl.s16 q2, d2\n"  in Transform()
4407       "vmovl.s16 q1, d1\n"  in Transform()
4408       "vmovl.s16 q0, d0\n"  in Transform()
4409       "vcvt.f32.s32 q0, q0\n"  in Transform()
4410       "vcvt.f32.s32 q1, q1\n"  in Transform()
4411       "vcvt.f32.s32 q2, q2\n"  in Transform()
4412       "vcvt.f32.s32 q3, q3\n"  in Transform()
4413       "vsub.f32 q0, q0, q5\n"  in Transform()
4414       "vsub.f32 q1, q1, q5\n"  in Transform()
4415       "vsub.f32 q2, q2, q5\n"  in Transform()
4416       "vsub.f32 q3, q3, q5\n"  in Transform()
4417       "vmul.f32 q0, q0, q6\n"  in Transform()
4418       "vmul.f32 q1, q1, q6\n"  in Transform()
4419       "vmul.f32 q2, q2, q6\n"  in Transform()
4420       "vmul.f32 q3, q3, q6\n"  in Transform()
4421       "vadd.f32 q0, q0, q4\n"  in Transform()
4422       "vadd.f32 q1, q1, q4\n"  in Transform()
4423       "vadd.f32 q2, q2, q4\n"  in Transform()
4424       "vadd.f32 q3, q3, q4\n"  in Transform()
4426       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4427       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
4428       "pld [%[output]]\n"  in Transform()
4430       "bne 1b\n"  in Transform()
4436       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
4437       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
4438       "vld1.16 {d1[2]}, [%[input]]!\n"  in Transform()
4439       "pld [%[input], #32]\n"  in Transform()
4440       "vmovl.u8 q1, d1\n"  in Transform()
4441       "vmovl.u8 q0, d0\n"  in Transform()
4442       "vmovl.s16 q3, d3\n"  in Transform()
4443       "vmovl.s16 q2, d2\n"  in Transform()
4444       "vmovl.s16 q1, d1\n"  in Transform()
4445       "vmovl.s16 q0, d0\n"  in Transform()
4446       "vcvt.f32.s32 q0, q0\n"  in Transform()
4447       "vcvt.f32.s32 q1, q1\n"  in Transform()
4448       "vcvt.f32.s32 q2, q2\n"  in Transform()
4449       "vcvt.f32.s32 q3, q3\n"  in Transform()
4450       "vsub.f32 q0, q0, q5\n"  in Transform()
4451       "vsub.f32 q1, q1, q5\n"  in Transform()
4452       "vsub.f32 q2, q2, q5\n"  in Transform()
4453       "vsub.f32 q3, q3, q5\n"  in Transform()
4454       "vmul.f32 q0, q0, q6\n"  in Transform()
4455       "vmul.f32 q1, q1, q6\n"  in Transform()
4456       "vmul.f32 q2, q2, q6\n"  in Transform()
4457       "vmul.f32 q3, q3, q6\n"  in Transform()
4458       "vadd.f32 q0, q0, q4\n"  in Transform()
4459       "vadd.f32 q1, q1, q4\n"  in Transform()
4460       "vadd.f32 q2, q2, q4\n"  in Transform()
4461       "vadd.f32 q3, q3, q4\n"  in Transform()
4463       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4464       "vst1.32 {d4, d5, d6}, [%[output]]!\n"  in Transform()
4465       "pld [%[output]]\n"  in Transform()
4489       "vdup.32 q4, %[range_min]\n"  in Transform()
4490       "vdup.32 q5, %[range_offset]\n"  in Transform()
4491       "vdup.32 q6, %[range_scale]\n"  in Transform()
4494       "subs %[count], %[count], #15\n"  in Transform()
4495       "beq 2f\n"  in Transform()
4498       "subs %[count], %[count], #16\n"  in Transform()
4501       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4502       "pld [%[input], #32]\n"  in Transform()
4503       "vmovl.u8 q1, d1\n"  in Transform()
4504       "vmovl.u8 q0, d0\n"  in Transform()
4505       "vmovl.s16 q3, d3\n"  in Transform()
4506       "vmovl.s16 q2, d2\n"  in Transform()
4507       "vmovl.s16 q1, d1\n"  in Transform()
4508       "vmovl.s16 q0, d0\n"  in Transform()
4509       "vcvt.f32.s32 q0, q0\n"  in Transform()
4510       "vcvt.f32.s32 q1, q1\n"  in Transform()
4511       "vcvt.f32.s32 q2, q2\n"  in Transform()
4512       "vcvt.f32.s32 q3, q3\n"  in Transform()
4513       "vsub.f32 q0, q0, q5\n"  in Transform()
4514       "vsub.f32 q1, q1, q5\n"  in Transform()
4515       "vsub.f32 q2, q2, q5\n"  in Transform()
4516       "vsub.f32 q3, q3, q5\n"  in Transform()
4517       "vmul.f32 q0, q0, q6\n"  in Transform()
4518       "vmul.f32 q1, q1, q6\n"  in Transform()
4519       "vmul.f32 q2, q2, q6\n"  in Transform()
4520       "vmul.f32 q3, q3, q6\n"  in Transform()
4521       "vadd.f32 q0, q0, q4\n"  in Transform()
4522       "vadd.f32 q1, q1, q4\n"  in Transform()
4523       "vadd.f32 q2, q2, q4\n"  in Transform()
4524       "vadd.f32 q3, q3, q4\n"  in Transform()
4526       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4527       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
4528       "pld [%[output]]\n"  in Transform()
4530       "bne 1b\n"  in Transform()
4536       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
4537       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
4538       "vld1.16 {d1[2]}, [%[input]]!\n"  in Transform()
4539       "vld1.8 {d1[6]}, [%[input]]!\n"  in Transform()
4540       "pld [%[input], #32]\n"  in Transform()
4541       "vmovl.u8 q1, d1\n"  in Transform()
4542       "vmovl.u8 q0, d0\n"  in Transform()
4543       "vmovl.s16 q3, d3\n"  in Transform()
4544       "vmovl.s16 q2, d2\n"  in Transform()
4545       "vmovl.s16 q1, d1\n"  in Transform()
4546       "vmovl.s16 q0, d0\n"  in Transform()
4547       "vcvt.f32.s32 q0, q0\n"  in Transform()
4548       "vcvt.f32.s32 q1, q1\n"  in Transform()
4549       "vcvt.f32.s32 q2, q2\n"  in Transform()
4550       "vcvt.f32.s32 q3, q3\n"  in Transform()
4551       "vsub.f32 q0, q0, q5\n"  in Transform()
4552       "vsub.f32 q1, q1, q5\n"  in Transform()
4553       "vsub.f32 q2, q2, q5\n"  in Transform()
4554       "vsub.f32 q3, q3, q5\n"  in Transform()
4555       "vmul.f32 q0, q0, q6\n"  in Transform()
4556       "vmul.f32 q1, q1, q6\n"  in Transform()
4557       "vmul.f32 q2, q2, q6\n"  in Transform()
4558       "vmul.f32 q3, q3, q6\n"  in Transform()
4559       "vadd.f32 q0, q0, q4\n"  in Transform()
4560       "vadd.f32 q1, q1, q4\n"  in Transform()
4561       "vadd.f32 q2, q2, q4\n"  in Transform()
4562       "vadd.f32 q3, q3, q4\n"  in Transform()
4564       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
4565       "vst1.32 {d4, d5, d6}, [%[output]]!\n"  in Transform()
4566       "vst1.32 {d7[0]}, [%[output]]!\n"  in Transform()
4567       "pld [%[output]]\n"  in Transform()
4594       "vdup.8 q4, %[min]\n"  in Transform()
4595       "vdup.8 q5, %[max]\n"  in Transform()
4598       "subs %[count], %[count], #16\n"  in Transform()
4601       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4602       "pld [%[input], #16]\n"  in Transform()
4603       "vmax.u8 q0, q0, q4\n"  in Transform()
4604       "vmin.u8 q0, q0, q5\n"  in Transform()
4606       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
4607       "pld [%[output]]\n"  in Transform()
4609       "bne 1b\n"  in Transform()
4634       "vdup.8 q4, %[min]\n"  in Transform()
4635       "vdup.8 q5, %[max]\n"  in Transform()
4638       "subs %[count], %[count], #1\n"  in Transform()
4639       "beq 2f\n"  in Transform()
4642       "subs %[count], %[count], #16\n"  in Transform()
4645       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4646       "pld [%[input], #16]\n"  in Transform()
4647       "vmax.u8 q0, q0, q4\n"  in Transform()
4648       "vmin.u8 q0, q0, q5\n"  in Transform()
4650       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
4651       "pld [%[output]]\n"  in Transform()
4653       "bne 1b\n"  in Transform()
4659       "vld1.8 {d0[0]}, [%[input]]!\n"  in Transform()
4660       "pld [%[input], #16]\n"  in Transform()
4661       "vmax.u8 q0, q0, q4\n"  in Transform()
4662       "vmin.u8 q0, q0, q5\n"  in Transform()
4664       "vst1.8 {d0[0]}, [%[output]]!\n"  in Transform()
4665       "pld [%[output]]\n"  in Transform()
4690       "vdup.8 q4, %[min]\n"  in Transform()
4691       "vdup.8 q5, %[max]\n"  in Transform()
4694       "subs %[count], %[count], #2\n"  in Transform()
4695       "beq 2f\n"  in Transform()
4698       "subs %[count], %[count], #16\n"  in Transform()
4701       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4702       "pld [%[input], #16]\n"  in Transform()
4703       "vmax.u8 q0, q0, q4\n"  in Transform()
4704       "vmin.u8 q0, q0, q5\n"  in Transform()
4706       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
4707       "pld [%[output]]\n"  in Transform()
4709       "bne 1b\n"  in Transform()
4715       "vld1.16 {d0[0]}, [%[input]]!\n"  in Transform()
4716       "pld [%[input], #16]\n"  in Transform()
4717       "vmax.u8 q0, q0, q4\n"  in Transform()
4718       "vmin.u8 q0, q0, q5\n"  in Transform()
4720       "vst1.16 {d0[0]}, [%[output]]!\n"  in Transform()
4721       "pld [%[output]]\n"  in Transform()
4746       "vdup.8 q4, %[min]\n"  in Transform()
4747       "vdup.8 q5, %[max]\n"  in Transform()
4750       "subs %[count], %[count], #3\n"  in Transform()
4751       "beq 2f\n"  in Transform()
4754       "subs %[count], %[count], #16\n"  in Transform()
4757       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4758       "pld [%[input], #16]\n"  in Transform()
4759       "vmax.u8 q0, q0, q4\n"  in Transform()
4760       "vmin.u8 q0, q0, q5\n"  in Transform()
4762       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
4763       "pld [%[output]]\n"  in Transform()
4765       "bne 1b\n"  in Transform()
4771       "vld1.16 {d0[0]}, [%[input]]!\n"  in Transform()
4772       "vld1.8 {d0[2]}, [%[input]]!\n"  in Transform()
4773       "pld [%[input], #16]\n"  in Transform()
4774       "vmax.u8 q0, q0, q4\n"  in Transform()
4775       "vmin.u8 q0, q0, q5\n"  in Transform()
4777       "vst1.16 {d0[0]}, [%[output]]!\n"  in Transform()
4778       "vst1.8 {d0[2]}, [%[output]]!\n"  in Transform()
4779       "pld [%[output]]\n"  in Transform()
4804       "vdup.8 q4, %[min]\n"  in Transform()
4805       "vdup.8 q5, %[max]\n"  in Transform()
4808       "subs %[count], %[count], #4\n"  in Transform()
4809       "beq 2f\n"  in Transform()
4812       "subs %[count], %[count], #16\n"  in Transform()
4815       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4816       "pld [%[input], #16]\n"  in Transform()
4817       "vmax.u8 q0, q0, q4\n"  in Transform()
4818       "vmin.u8 q0, q0, q5\n"  in Transform()
4820       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
4821       "pld [%[output]]\n"  in Transform()
4823       "bne 1b\n"  in Transform()
4829       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
4830       "pld [%[input], #16]\n"  in Transform()
4831       "vmax.u8 q0, q0, q4\n"  in Transform()
4832       "vmin.u8 q0, q0, q5\n"  in Transform()
4834       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
4835       "pld [%[output]]\n"  in Transform()
4860       "vdup.8 q4, %[min]\n"  in Transform()
4861       "vdup.8 q5, %[max]\n"  in Transform()
4864       "subs %[count], %[count], #5\n"  in Transform()
4865       "beq 2f\n"  in Transform()
4868       "subs %[count], %[count], #16\n"  in Transform()
4871       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4872       "pld [%[input], #16]\n"  in Transform()
4873       "vmax.u8 q0, q0, q4\n"  in Transform()
4874       "vmin.u8 q0, q0, q5\n"  in Transform()
4876       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
4877       "pld [%[output]]\n"  in Transform()
4879       "bne 1b\n"  in Transform()
4885       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
4886       "vld1.8 {d0[4]}, [%[input]]!\n"  in Transform()
4887       "pld [%[input], #16]\n"  in Transform()
4888       "vmax.u8 q0, q0, q4\n"  in Transform()
4889       "vmin.u8 q0, q0, q5\n"  in Transform()
4891       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
4892       "vst1.8 {d0[4]}, [%[output]]!\n"  in Transform()
4893       "pld [%[output]]\n"  in Transform()
4918       "vdup.8 q4, %[min]\n"  in Transform()
4919       "vdup.8 q5, %[max]\n"  in Transform()
4922       "subs %[count], %[count], #6\n"  in Transform()
4923       "beq 2f\n"  in Transform()
4926       "subs %[count], %[count], #16\n"  in Transform()
4929       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4930       "pld [%[input], #16]\n"  in Transform()
4931       "vmax.u8 q0, q0, q4\n"  in Transform()
4932       "vmin.u8 q0, q0, q5\n"  in Transform()
4934       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
4935       "pld [%[output]]\n"  in Transform()
4937       "bne 1b\n"  in Transform()
4943       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
4944       "vld1.16 {d0[2]}, [%[input]]!\n"  in Transform()
4945       "pld [%[input], #16]\n"  in Transform()
4946       "vmax.u8 q0, q0, q4\n"  in Transform()
4947       "vmin.u8 q0, q0, q5\n"  in Transform()
4949       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
4950       "vst1.16 {d0[2]}, [%[output]]!\n"  in Transform()
4951       "pld [%[output]]\n"  in Transform()
4976       "vdup.8 q4, %[min]\n"  in Transform()
4977       "vdup.8 q5, %[max]\n"  in Transform()
4980       "subs %[count], %[count], #7\n"  in Transform()
4981       "beq 2f\n"  in Transform()
4984       "subs %[count], %[count], #16\n"  in Transform()
4987       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
4988       "pld [%[input], #16]\n"  in Transform()
4989       "vmax.u8 q0, q0, q4\n"  in Transform()
4990       "vmin.u8 q0, q0, q5\n"  in Transform()
4992       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
4993       "pld [%[output]]\n"  in Transform()
4995       "bne 1b\n"  in Transform()
5001       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
5002       "vld1.16 {d0[2]}, [%[input]]!\n"  in Transform()
5003       "vld1.8 {d0[6]}, [%[input]]!\n"  in Transform()
5004       "pld [%[input], #16]\n"  in Transform()
5005       "vmax.u8 q0, q0, q4\n"  in Transform()
5006       "vmin.u8 q0, q0, q5\n"  in Transform()
5008       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
5009       "vst1.16 {d0[2]}, [%[output]]!\n"  in Transform()
5010       "vst1.8 {d0[6]}, [%[output]]!\n"  in Transform()
5011       "pld [%[output]]\n"  in Transform()
5036       "vdup.8 q4, %[min]\n"  in Transform()
5037       "vdup.8 q5, %[max]\n"  in Transform()
5040       "subs %[count], %[count], #8\n"  in Transform()
5041       "beq 2f\n"  in Transform()
5044       "subs %[count], %[count], #16\n"  in Transform()
5047       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5048       "pld [%[input], #16]\n"  in Transform()
5049       "vmax.u8 q0, q0, q4\n"  in Transform()
5050       "vmin.u8 q0, q0, q5\n"  in Transform()
5052       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
5053       "pld [%[output]]\n"  in Transform()
5055       "bne 1b\n"  in Transform()
5061       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
5062       "pld [%[input], #16]\n"  in Transform()
5063       "vmax.u8 q0, q0, q4\n"  in Transform()
5064       "vmin.u8 q0, q0, q5\n"  in Transform()
5066       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
5067       "pld [%[output]]\n"  in Transform()
5092       "vdup.8 q4, %[min]\n"  in Transform()
5093       "vdup.8 q5, %[max]\n"  in Transform()
5096       "subs %[count], %[count], #9\n"  in Transform()
5097       "beq 2f\n"  in Transform()
5100       "subs %[count], %[count], #16\n"  in Transform()
5103       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5104       "pld [%[input], #16]\n"  in Transform()
5105       "vmax.u8 q0, q0, q4\n"  in Transform()
5106       "vmin.u8 q0, q0, q5\n"  in Transform()
5108       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
5109       "pld [%[output]]\n"  in Transform()
5111       "bne 1b\n"  in Transform()
5117       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
5118       "vld1.8 {d1[0]}, [%[input]]!\n"  in Transform()
5119       "pld [%[input], #16]\n"  in Transform()
5120       "vmax.u8 q0, q0, q4\n"  in Transform()
5121       "vmin.u8 q0, q0, q5\n"  in Transform()
5123       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
5124       "vst1.8 {d1[0]}, [%[output]]!\n"  in Transform()
5125       "pld [%[output]]\n"  in Transform()
5150       "vdup.8 q4, %[min]\n"  in Transform()
5151       "vdup.8 q5, %[max]\n"  in Transform()
5154       "subs %[count], %[count], #10\n"  in Transform()
5155       "beq 2f\n"  in Transform()
5158       "subs %[count], %[count], #16\n"  in Transform()
5161       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5162       "pld [%[input], #16]\n"  in Transform()
5163       "vmax.u8 q0, q0, q4\n"  in Transform()
5164       "vmin.u8 q0, q0, q5\n"  in Transform()
5166       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
5167       "pld [%[output]]\n"  in Transform()
5169       "bne 1b\n"  in Transform()
5175       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
5176       "vld1.16 {d1[0]}, [%[input]]!\n"  in Transform()
5177       "pld [%[input], #16]\n"  in Transform()
5178       "vmax.u8 q0, q0, q4\n"  in Transform()
5179       "vmin.u8 q0, q0, q5\n"  in Transform()
5181       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
5182       "vst1.16 {d1[0]}, [%[output]]!\n"  in Transform()
5183       "pld [%[output]]\n"  in Transform()
5208       "vdup.8 q4, %[min]\n"  in Transform()
5209       "vdup.8 q5, %[max]\n"  in Transform()
5212       "subs %[count], %[count], #11\n"  in Transform()
5213       "beq 2f\n"  in Transform()
5216       "subs %[count], %[count], #16\n"  in Transform()
5219       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5220       "pld [%[input], #16]\n"  in Transform()
5221       "vmax.u8 q0, q0, q4\n"  in Transform()
5222       "vmin.u8 q0, q0, q5\n"  in Transform()
5224       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
5225       "pld [%[output]]\n"  in Transform()
5227       "bne 1b\n"  in Transform()
5233       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
5234       "vld1.16 {d1[0]}, [%[input]]!\n"  in Transform()
5235       "vld1.8 {d1[2]}, [%[input]]!\n"  in Transform()
5236       "pld [%[input], #16]\n"  in Transform()
5237       "vmax.u8 q0, q0, q4\n"  in Transform()
5238       "vmin.u8 q0, q0, q5\n"  in Transform()
5240       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
5241       "vst1.16 {d1[0]}, [%[output]]!\n"  in Transform()
5242       "vst1.8 {d1[2]}, [%[output]]!\n"  in Transform()
5243       "pld [%[output]]\n"  in Transform()
5268       "vdup.8 q4, %[min]\n"  in Transform()
5269       "vdup.8 q5, %[max]\n"  in Transform()
5272       "subs %[count], %[count], #12\n"  in Transform()
5273       "beq 2f\n"  in Transform()
5276       "subs %[count], %[count], #16\n"  in Transform()
5279       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5280       "pld [%[input], #16]\n"  in Transform()
5281       "vmax.u8 q0, q0, q4\n"  in Transform()
5282       "vmin.u8 q0, q0, q5\n"  in Transform()
5284       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
5285       "pld [%[output]]\n"  in Transform()
5287       "bne 1b\n"  in Transform()
5293       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
5294       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
5295       "pld [%[input], #16]\n"  in Transform()
5296       "vmax.u8 q0, q0, q4\n"  in Transform()
5297       "vmin.u8 q0, q0, q5\n"  in Transform()
5299       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
5300       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
5301       "pld [%[output]]\n"  in Transform()
5326       "vdup.8 q4, %[min]\n"  in Transform()
5327       "vdup.8 q5, %[max]\n"  in Transform()
5330       "subs %[count], %[count], #13\n"  in Transform()
5331       "beq 2f\n"  in Transform()
5334       "subs %[count], %[count], #16\n"  in Transform()
5337       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5338       "pld [%[input], #16]\n"  in Transform()
5339       "vmax.u8 q0, q0, q4\n"  in Transform()
5340       "vmin.u8 q0, q0, q5\n"  in Transform()
5342       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
5343       "pld [%[output]]\n"  in Transform()
5345       "bne 1b\n"  in Transform()
5351       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
5352       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
5353       "vld1.8 {d1[4]}, [%[input]]!\n"  in Transform()
5354       "pld [%[input], #16]\n"  in Transform()
5355       "vmax.u8 q0, q0, q4\n"  in Transform()
5356       "vmin.u8 q0, q0, q5\n"  in Transform()
5358       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
5359       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
5360       "vst1.8 {d1[4]}, [%[output]]!\n"  in Transform()
5361       "pld [%[output]]\n"  in Transform()
5386       "vdup.8 q4, %[min]\n"  in Transform()
5387       "vdup.8 q5, %[max]\n"  in Transform()
5390       "subs %[count], %[count], #14\n"  in Transform()
5391       "beq 2f\n"  in Transform()
5394       "subs %[count], %[count], #16\n"  in Transform()
5397       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5398       "pld [%[input], #16]\n"  in Transform()
5399       "vmax.u8 q0, q0, q4\n"  in Transform()
5400       "vmin.u8 q0, q0, q5\n"  in Transform()
5402       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
5403       "pld [%[output]]\n"  in Transform()
5405       "bne 1b\n"  in Transform()
5411       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
5412       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
5413       "vld1.16 {d1[2]}, [%[input]]!\n"  in Transform()
5414       "pld [%[input], #16]\n"  in Transform()
5415       "vmax.u8 q0, q0, q4\n"  in Transform()
5416       "vmin.u8 q0, q0, q5\n"  in Transform()
5418       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
5419       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
5420       "vst1.16 {d1[2]}, [%[output]]!\n"  in Transform()
5421       "pld [%[output]]\n"  in Transform()
5446       "vdup.8 q4, %[min]\n"  in Transform()
5447       "vdup.8 q5, %[max]\n"  in Transform()
5450       "subs %[count], %[count], #15\n"  in Transform()
5451       "beq 2f\n"  in Transform()
5454       "subs %[count], %[count], #16\n"  in Transform()
5457       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5458       "pld [%[input], #16]\n"  in Transform()
5459       "vmax.u8 q0, q0, q4\n"  in Transform()
5460       "vmin.u8 q0, q0, q5\n"  in Transform()
5462       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
5463       "pld [%[output]]\n"  in Transform()
5465       "bne 1b\n"  in Transform()
5471       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
5472       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
5473       "vld1.16 {d1[2]}, [%[input]]!\n"  in Transform()
5474       "vld1.8 {d1[6]}, [%[input]]!\n"  in Transform()
5475       "pld [%[input], #16]\n"  in Transform()
5476       "vmax.u8 q0, q0, q4\n"  in Transform()
5477       "vmin.u8 q0, q0, q5\n"  in Transform()
5479       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
5480       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
5481       "vst1.16 {d1[2]}, [%[output]]!\n"  in Transform()
5482       "vst1.8 {d1[6]}, [%[output]]!\n"  in Transform()
5483       "pld [%[output]]\n"  in Transform()
5506       "ldr r0, %[input_range_min]\n"  in Transform()
5507       "vdup.32 q8, r0\n"  in Transform()
5508       "ldr r0, %[input_range_scale]\n"  in Transform()
5509       "vdup.32 q9, r0\n"  in Transform()
5510       "ldr r0, %[bias_range_min]\n"  in Transform()
5511       "vdup.32 q10, r0\n"  in Transform()
5512       "ldr r0, %[bias_range_scale]\n"  in Transform()
5513       "vdup.32 q11, r0\n"  in Transform()
5514       "ldr r0, %[output_range_min]\n"  in Transform()
5515       "vdup.32 q12, r0\n"  in Transform()
5516       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
5517       "vdup.32 q13, r0\n"  in Transform()
5518       "ldr r0, %[output_range_offset]\n"  in Transform()
5519       "vdup.32 q14, r0\n"  in Transform()
5521       "mov r0, %[count]\n"  in Transform()
5522       "mov r1, %[bias]\n"  in Transform()
5524       "subs r0, r0, #16\n"  in Transform()
5527       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5528       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
5529       "pld [%[input], #32]\n"  in Transform()
5530       "vmovl.u8 q1, d1\n"  in Transform()
5531       "vmovl.u8 q0, d0\n"  in Transform()
5532       "vmovl.u8 q5, d9\n"  in Transform()
5533       "vmovl.u8 q4, d8\n"  in Transform()
5534       "vmovl.s16 q3, d3\n"  in Transform()
5535       "vmovl.s16 q2, d2\n"  in Transform()
5536       "vmovl.s16 q7, d11\n"  in Transform()
5537       "vmovl.s16 q6, d10\n"  in Transform()
5538       "vmovl.s16 q1, d1\n"  in Transform()
5539       "vmovl.s16 q0, d0\n"  in Transform()
5540       "vmovl.s16 q5, d9\n"  in Transform()
5541       "vmovl.s16 q4, d8\n"  in Transform()
5542       "vcvt.f32.s32 q0, q0\n"  in Transform()
5543       "vcvt.f32.s32 q1, q1\n"  in Transform()
5544       "vcvt.f32.s32 q2, q2\n"  in Transform()
5545       "vcvt.f32.s32 q3, q3\n"  in Transform()
5546       "vcvt.f32.s32 q4, q4\n"  in Transform()
5547       "vcvt.f32.s32 q5, q5\n"  in Transform()
5548       "vcvt.f32.s32 q6, q6\n"  in Transform()
5549       "vcvt.f32.s32 q7, q7\n"  in Transform()
5550       "vmul.f32 q0, q0, q9\n"  in Transform()
5551       "vmul.f32 q1, q1, q9\n"  in Transform()
5552       "vmul.f32 q2, q2, q9\n"  in Transform()
5553       "vmul.f32 q3, q3, q9\n"  in Transform()
5554       "vmul.f32 q4, q4, q11\n"  in Transform()
5555       "vmul.f32 q5, q5, q11\n"  in Transform()
5556       "vmul.f32 q6, q6, q11\n"  in Transform()
5557       "vmul.f32 q7, q7, q11\n"  in Transform()
5558       "vadd.f32 q0, q0, q8\n"  in Transform()
5559       "vadd.f32 q1, q1, q8\n"  in Transform()
5560       "vadd.f32 q2, q2, q8\n"  in Transform()
5561       "vadd.f32 q3, q3, q8\n"  in Transform()
5562       "vadd.f32 q4, q4, q10\n"  in Transform()
5563       "vadd.f32 q5, q5, q10\n"  in Transform()
5564       "vadd.f32 q6, q6, q10\n"  in Transform()
5565       "vadd.f32 q7, q7, q10\n"  in Transform()
5566       "vadd.f32 q0, q0, q4\n"  in Transform()
5567       "vadd.f32 q1, q1, q5\n"  in Transform()
5568       "vadd.f32 q2, q2, q6\n"  in Transform()
5569       "vadd.f32 q3, q3, q7\n"  in Transform()
5570       "vsub.f32 q0, q0, q12\n"  in Transform()
5571       "vsub.f32 q1, q1, q12\n"  in Transform()
5572       "vsub.f32 q2, q2, q12\n"  in Transform()
5573       "vsub.f32 q3, q3, q12\n"  in Transform()
5574       "vmul.f32 q0, q0, q13\n"  in Transform()
5575       "vmul.f32 q1, q1, q13\n"  in Transform()
5576       "vmul.f32 q2, q2, q13\n"  in Transform()
5577       "vmul.f32 q3, q3, q13\n"  in Transform()
5578       "vadd.f32 q0, q0, q14\n"  in Transform()
5579       "vadd.f32 q1, q1, q14\n"  in Transform()
5580       "vadd.f32 q2, q2, q14\n"  in Transform()
5581       "vadd.f32 q3, q3, q14\n"  in Transform()
5582       "vcvt.s32.f32 q0, q0\n"  in Transform()
5583       "vcvt.s32.f32 q1, q1\n"  in Transform()
5584       "vcvt.s32.f32 q2, q2\n"  in Transform()
5585       "vcvt.s32.f32 q3, q3\n"  in Transform()
5587       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
5588       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
5589       "pld [%[output]]\n"  in Transform()
5590       "bne 2b\n"  in Transform()
5591       "subs %[rows], %[rows], #1\n"  in Transform()
5592       "bne 1b\n"  in Transform()
5624       "ldr r0, %[input_range_min]\n"  in Transform()
5625       "vdup.32 q8, r0\n"  in Transform()
5626       "ldr r0, %[input_range_scale]\n"  in Transform()
5627       "vdup.32 q9, r0\n"  in Transform()
5628       "ldr r0, %[bias_range_min]\n"  in Transform()
5629       "vdup.32 q10, r0\n"  in Transform()
5630       "ldr r0, %[bias_range_scale]\n"  in Transform()
5631       "vdup.32 q11, r0\n"  in Transform()
5632       "ldr r0, %[output_range_min]\n"  in Transform()
5633       "vdup.32 q12, r0\n"  in Transform()
5634       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
5635       "vdup.32 q13, r0\n"  in Transform()
5636       "ldr r0, %[output_range_offset]\n"  in Transform()
5637       "vdup.32 q14, r0\n"  in Transform()
5639       "mov r0, %[count]\n"  in Transform()
5640       "mov r1, %[bias]\n"  in Transform()
5641       "subs r0, r0, #1\n"  in Transform()
5642       "beq 3f\n"  in Transform()
5644       "subs r0, r0, #16\n"  in Transform()
5647       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5648       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
5649       "pld [%[input], #32]\n"  in Transform()
5650       "vmovl.u8 q1, d1\n"  in Transform()
5651       "vmovl.u8 q0, d0\n"  in Transform()
5652       "vmovl.u8 q5, d9\n"  in Transform()
5653       "vmovl.u8 q4, d8\n"  in Transform()
5654       "vmovl.s16 q3, d3\n"  in Transform()
5655       "vmovl.s16 q2, d2\n"  in Transform()
5656       "vmovl.s16 q7, d11\n"  in Transform()
5657       "vmovl.s16 q6, d10\n"  in Transform()
5658       "vmovl.s16 q1, d1\n"  in Transform()
5659       "vmovl.s16 q0, d0\n"  in Transform()
5660       "vmovl.s16 q5, d9\n"  in Transform()
5661       "vmovl.s16 q4, d8\n"  in Transform()
5662       "vcvt.f32.s32 q0, q0\n"  in Transform()
5663       "vcvt.f32.s32 q1, q1\n"  in Transform()
5664       "vcvt.f32.s32 q2, q2\n"  in Transform()
5665       "vcvt.f32.s32 q3, q3\n"  in Transform()
5666       "vcvt.f32.s32 q4, q4\n"  in Transform()
5667       "vcvt.f32.s32 q5, q5\n"  in Transform()
5668       "vcvt.f32.s32 q6, q6\n"  in Transform()
5669       "vcvt.f32.s32 q7, q7\n"  in Transform()
5670       "vmul.f32 q0, q0, q9\n"  in Transform()
5671       "vmul.f32 q1, q1, q9\n"  in Transform()
5672       "vmul.f32 q2, q2, q9\n"  in Transform()
5673       "vmul.f32 q3, q3, q9\n"  in Transform()
5674       "vmul.f32 q4, q4, q11\n"  in Transform()
5675       "vmul.f32 q5, q5, q11\n"  in Transform()
5676       "vmul.f32 q6, q6, q11\n"  in Transform()
5677       "vmul.f32 q7, q7, q11\n"  in Transform()
5678       "vadd.f32 q0, q0, q8\n"  in Transform()
5679       "vadd.f32 q1, q1, q8\n"  in Transform()
5680       "vadd.f32 q2, q2, q8\n"  in Transform()
5681       "vadd.f32 q3, q3, q8\n"  in Transform()
5682       "vadd.f32 q4, q4, q10\n"  in Transform()
5683       "vadd.f32 q5, q5, q10\n"  in Transform()
5684       "vadd.f32 q6, q6, q10\n"  in Transform()
5685       "vadd.f32 q7, q7, q10\n"  in Transform()
5686       "vadd.f32 q0, q0, q4\n"  in Transform()
5687       "vadd.f32 q1, q1, q5\n"  in Transform()
5688       "vadd.f32 q2, q2, q6\n"  in Transform()
5689       "vadd.f32 q3, q3, q7\n"  in Transform()
5690       "vsub.f32 q0, q0, q12\n"  in Transform()
5691       "vsub.f32 q1, q1, q12\n"  in Transform()
5692       "vsub.f32 q2, q2, q12\n"  in Transform()
5693       "vsub.f32 q3, q3, q12\n"  in Transform()
5694       "vmul.f32 q0, q0, q13\n"  in Transform()
5695       "vmul.f32 q1, q1, q13\n"  in Transform()
5696       "vmul.f32 q2, q2, q13\n"  in Transform()
5697       "vmul.f32 q3, q3, q13\n"  in Transform()
5698       "vadd.f32 q0, q0, q14\n"  in Transform()
5699       "vadd.f32 q1, q1, q14\n"  in Transform()
5700       "vadd.f32 q2, q2, q14\n"  in Transform()
5701       "vadd.f32 q3, q3, q14\n"  in Transform()
5702       "vcvt.s32.f32 q0, q0\n"  in Transform()
5703       "vcvt.s32.f32 q1, q1\n"  in Transform()
5704       "vcvt.s32.f32 q2, q2\n"  in Transform()
5705       "vcvt.s32.f32 q3, q3\n"  in Transform()
5707       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
5708       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
5709       "pld [%[output]]\n"  in Transform()
5710       "bne 2b\n"  in Transform()
5714       "vld1.8 {d0[0]}, [%[input]]!\n"  in Transform()
5715       "vld1.8 {d2[0]}, [r1]!\n"  in Transform()
5716       "pld [%[input], #32]\n"  in Transform()
5717       "vmovl.u8 q0, d0\n"  in Transform()
5718       "vmovl.u8 q1, d2\n"  in Transform()
5719       "vmovl.s16 q0, d0\n"  in Transform()
5720       "vmovl.s16 q1, d2\n"  in Transform()
5721       "vcvt.f32.s32 q0, q0\n"  in Transform()
5722       "vcvt.f32.s32 q1, q1\n"  in Transform()
5723       "vmul.f32 q0, q0, q9\n"  in Transform()
5724       "vmul.f32 q1, q1, q11\n"  in Transform()
5725       "vadd.f32 q0, q0, q8\n"  in Transform()
5726       "vadd.f32 q1, q1, q10\n"  in Transform()
5727       "vadd.f32 q0, q0, q1\n"  in Transform()
5728       "vsub.f32 q0, q0, q12\n"  in Transform()
5729       "vmul.f32 q0, q0, q13\n"  in Transform()
5730       "vadd.f32 q0, q0, q14\n"  in Transform()
5731       "vcvt.s32.f32 q0, q0\n"  in Transform()
5733       "vst1.32 {d0[0]}, [%[output]]!\n"  in Transform()
5734       "pld [%[output]]\n"  in Transform()
5735       "subs %[rows], %[rows], #1\n"  in Transform()
5736       "bne 1b\n"  in Transform()
5768       "ldr r0, %[input_range_min]\n"  in Transform()
5769       "vdup.32 q8, r0\n"  in Transform()
5770       "ldr r0, %[input_range_scale]\n"  in Transform()
5771       "vdup.32 q9, r0\n"  in Transform()
5772       "ldr r0, %[bias_range_min]\n"  in Transform()
5773       "vdup.32 q10, r0\n"  in Transform()
5774       "ldr r0, %[bias_range_scale]\n"  in Transform()
5775       "vdup.32 q11, r0\n"  in Transform()
5776       "ldr r0, %[output_range_min]\n"  in Transform()
5777       "vdup.32 q12, r0\n"  in Transform()
5778       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
5779       "vdup.32 q13, r0\n"  in Transform()
5780       "ldr r0, %[output_range_offset]\n"  in Transform()
5781       "vdup.32 q14, r0\n"  in Transform()
5783       "mov r0, %[count]\n"  in Transform()
5784       "mov r1, %[bias]\n"  in Transform()
5785       "subs r0, r0, #2\n"  in Transform()
5786       "beq 3f\n"  in Transform()
5788       "subs r0, r0, #16\n"  in Transform()
5791       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5792       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
5793       "pld [%[input], #32]\n"  in Transform()
5794       "vmovl.u8 q1, d1\n"  in Transform()
5795       "vmovl.u8 q0, d0\n"  in Transform()
5796       "vmovl.u8 q5, d9\n"  in Transform()
5797       "vmovl.u8 q4, d8\n"  in Transform()
5798       "vmovl.s16 q3, d3\n"  in Transform()
5799       "vmovl.s16 q2, d2\n"  in Transform()
5800       "vmovl.s16 q7, d11\n"  in Transform()
5801       "vmovl.s16 q6, d10\n"  in Transform()
5802       "vmovl.s16 q1, d1\n"  in Transform()
5803       "vmovl.s16 q0, d0\n"  in Transform()
5804       "vmovl.s16 q5, d9\n"  in Transform()
5805       "vmovl.s16 q4, d8\n"  in Transform()
5806       "vcvt.f32.s32 q0, q0\n"  in Transform()
5807       "vcvt.f32.s32 q1, q1\n"  in Transform()
5808       "vcvt.f32.s32 q2, q2\n"  in Transform()
5809       "vcvt.f32.s32 q3, q3\n"  in Transform()
5810       "vcvt.f32.s32 q4, q4\n"  in Transform()
5811       "vcvt.f32.s32 q5, q5\n"  in Transform()
5812       "vcvt.f32.s32 q6, q6\n"  in Transform()
5813       "vcvt.f32.s32 q7, q7\n"  in Transform()
5814       "vmul.f32 q0, q0, q9\n"  in Transform()
5815       "vmul.f32 q1, q1, q9\n"  in Transform()
5816       "vmul.f32 q2, q2, q9\n"  in Transform()
5817       "vmul.f32 q3, q3, q9\n"  in Transform()
5818       "vmul.f32 q4, q4, q11\n"  in Transform()
5819       "vmul.f32 q5, q5, q11\n"  in Transform()
5820       "vmul.f32 q6, q6, q11\n"  in Transform()
5821       "vmul.f32 q7, q7, q11\n"  in Transform()
5822       "vadd.f32 q0, q0, q8\n"  in Transform()
5823       "vadd.f32 q1, q1, q8\n"  in Transform()
5824       "vadd.f32 q2, q2, q8\n"  in Transform()
5825       "vadd.f32 q3, q3, q8\n"  in Transform()
5826       "vadd.f32 q4, q4, q10\n"  in Transform()
5827       "vadd.f32 q5, q5, q10\n"  in Transform()
5828       "vadd.f32 q6, q6, q10\n"  in Transform()
5829       "vadd.f32 q7, q7, q10\n"  in Transform()
5830       "vadd.f32 q0, q0, q4\n"  in Transform()
5831       "vadd.f32 q1, q1, q5\n"  in Transform()
5832       "vadd.f32 q2, q2, q6\n"  in Transform()
5833       "vadd.f32 q3, q3, q7\n"  in Transform()
5834       "vsub.f32 q0, q0, q12\n"  in Transform()
5835       "vsub.f32 q1, q1, q12\n"  in Transform()
5836       "vsub.f32 q2, q2, q12\n"  in Transform()
5837       "vsub.f32 q3, q3, q12\n"  in Transform()
5838       "vmul.f32 q0, q0, q13\n"  in Transform()
5839       "vmul.f32 q1, q1, q13\n"  in Transform()
5840       "vmul.f32 q2, q2, q13\n"  in Transform()
5841       "vmul.f32 q3, q3, q13\n"  in Transform()
5842       "vadd.f32 q0, q0, q14\n"  in Transform()
5843       "vadd.f32 q1, q1, q14\n"  in Transform()
5844       "vadd.f32 q2, q2, q14\n"  in Transform()
5845       "vadd.f32 q3, q3, q14\n"  in Transform()
5846       "vcvt.s32.f32 q0, q0\n"  in Transform()
5847       "vcvt.s32.f32 q1, q1\n"  in Transform()
5848       "vcvt.s32.f32 q2, q2\n"  in Transform()
5849       "vcvt.s32.f32 q3, q3\n"  in Transform()
5851       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
5852       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
5853       "pld [%[output]]\n"  in Transform()
5854       "bne 2b\n"  in Transform()
5858       "vld1.16 {d0[0]}, [%[input]]!\n"  in Transform()
5859       "vld1.16 {d2[0]}, [r1]!\n"  in Transform()
5860       "pld [%[input], #32]\n"  in Transform()
5861       "vmovl.u8 q0, d0\n"  in Transform()
5862       "vmovl.u8 q1, d2\n"  in Transform()
5863       "vmovl.s16 q0, d0\n"  in Transform()
5864       "vmovl.s16 q1, d2\n"  in Transform()
5865       "vcvt.f32.s32 q0, q0\n"  in Transform()
5866       "vcvt.f32.s32 q1, q1\n"  in Transform()
5867       "vmul.f32 q0, q0, q9\n"  in Transform()
5868       "vmul.f32 q1, q1, q11\n"  in Transform()
5869       "vadd.f32 q0, q0, q8\n"  in Transform()
5870       "vadd.f32 q1, q1, q10\n"  in Transform()
5871       "vadd.f32 q0, q0, q1\n"  in Transform()
5872       "vsub.f32 q0, q0, q12\n"  in Transform()
5873       "vmul.f32 q0, q0, q13\n"  in Transform()
5874       "vadd.f32 q0, q0, q14\n"  in Transform()
5875       "vcvt.s32.f32 q0, q0\n"  in Transform()
5877       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
5878       "pld [%[output]]\n"  in Transform()
5879       "subs %[rows], %[rows], #1\n"  in Transform()
5880       "bne 1b\n"  in Transform()
5912       "ldr r0, %[input_range_min]\n"  in Transform()
5913       "vdup.32 q8, r0\n"  in Transform()
5914       "ldr r0, %[input_range_scale]\n"  in Transform()
5915       "vdup.32 q9, r0\n"  in Transform()
5916       "ldr r0, %[bias_range_min]\n"  in Transform()
5917       "vdup.32 q10, r0\n"  in Transform()
5918       "ldr r0, %[bias_range_scale]\n"  in Transform()
5919       "vdup.32 q11, r0\n"  in Transform()
5920       "ldr r0, %[output_range_min]\n"  in Transform()
5921       "vdup.32 q12, r0\n"  in Transform()
5922       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
5923       "vdup.32 q13, r0\n"  in Transform()
5924       "ldr r0, %[output_range_offset]\n"  in Transform()
5925       "vdup.32 q14, r0\n"  in Transform()
5927       "mov r0, %[count]\n"  in Transform()
5928       "mov r1, %[bias]\n"  in Transform()
5929       "subs r0, r0, #3\n"  in Transform()
5930       "beq 3f\n"  in Transform()
5932       "subs r0, r0, #16\n"  in Transform()
5935       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
5936       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
5937       "pld [%[input], #32]\n"  in Transform()
5938       "vmovl.u8 q1, d1\n"  in Transform()
5939       "vmovl.u8 q0, d0\n"  in Transform()
5940       "vmovl.u8 q5, d9\n"  in Transform()
5941       "vmovl.u8 q4, d8\n"  in Transform()
5942       "vmovl.s16 q3, d3\n"  in Transform()
5943       "vmovl.s16 q2, d2\n"  in Transform()
5944       "vmovl.s16 q7, d11\n"  in Transform()
5945       "vmovl.s16 q6, d10\n"  in Transform()
5946       "vmovl.s16 q1, d1\n"  in Transform()
5947       "vmovl.s16 q0, d0\n"  in Transform()
5948       "vmovl.s16 q5, d9\n"  in Transform()
5949       "vmovl.s16 q4, d8\n"  in Transform()
5950       "vcvt.f32.s32 q0, q0\n"  in Transform()
5951       "vcvt.f32.s32 q1, q1\n"  in Transform()
5952       "vcvt.f32.s32 q2, q2\n"  in Transform()
5953       "vcvt.f32.s32 q3, q3\n"  in Transform()
5954       "vcvt.f32.s32 q4, q4\n"  in Transform()
5955       "vcvt.f32.s32 q5, q5\n"  in Transform()
5956       "vcvt.f32.s32 q6, q6\n"  in Transform()
5957       "vcvt.f32.s32 q7, q7\n"  in Transform()
5958       "vmul.f32 q0, q0, q9\n"  in Transform()
5959       "vmul.f32 q1, q1, q9\n"  in Transform()
5960       "vmul.f32 q2, q2, q9\n"  in Transform()
5961       "vmul.f32 q3, q3, q9\n"  in Transform()
5962       "vmul.f32 q4, q4, q11\n"  in Transform()
5963       "vmul.f32 q5, q5, q11\n"  in Transform()
5964       "vmul.f32 q6, q6, q11\n"  in Transform()
5965       "vmul.f32 q7, q7, q11\n"  in Transform()
5966       "vadd.f32 q0, q0, q8\n"  in Transform()
5967       "vadd.f32 q1, q1, q8\n"  in Transform()
5968       "vadd.f32 q2, q2, q8\n"  in Transform()
5969       "vadd.f32 q3, q3, q8\n"  in Transform()
5970       "vadd.f32 q4, q4, q10\n"  in Transform()
5971       "vadd.f32 q5, q5, q10\n"  in Transform()
5972       "vadd.f32 q6, q6, q10\n"  in Transform()
5973       "vadd.f32 q7, q7, q10\n"  in Transform()
5974       "vadd.f32 q0, q0, q4\n"  in Transform()
5975       "vadd.f32 q1, q1, q5\n"  in Transform()
5976       "vadd.f32 q2, q2, q6\n"  in Transform()
5977       "vadd.f32 q3, q3, q7\n"  in Transform()
5978       "vsub.f32 q0, q0, q12\n"  in Transform()
5979       "vsub.f32 q1, q1, q12\n"  in Transform()
5980       "vsub.f32 q2, q2, q12\n"  in Transform()
5981       "vsub.f32 q3, q3, q12\n"  in Transform()
5982       "vmul.f32 q0, q0, q13\n"  in Transform()
5983       "vmul.f32 q1, q1, q13\n"  in Transform()
5984       "vmul.f32 q2, q2, q13\n"  in Transform()
5985       "vmul.f32 q3, q3, q13\n"  in Transform()
5986       "vadd.f32 q0, q0, q14\n"  in Transform()
5987       "vadd.f32 q1, q1, q14\n"  in Transform()
5988       "vadd.f32 q2, q2, q14\n"  in Transform()
5989       "vadd.f32 q3, q3, q14\n"  in Transform()
5990       "vcvt.s32.f32 q0, q0\n"  in Transform()
5991       "vcvt.s32.f32 q1, q1\n"  in Transform()
5992       "vcvt.s32.f32 q2, q2\n"  in Transform()
5993       "vcvt.s32.f32 q3, q3\n"  in Transform()
5995       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
5996       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
5997       "pld [%[output]]\n"  in Transform()
5998       "bne 2b\n"  in Transform()
6002       "vld1.16 {d0[0]}, [%[input]]!\n"  in Transform()
6003       "vld1.8 {d0[2]}, [%[input]]!\n"  in Transform()
6004       "vld1.16 {d2[0]}, [r1]!\n"  in Transform()
6005       "vld1.8 {d2[2]}, [r1]!\n"  in Transform()
6006       "pld [%[input], #32]\n"  in Transform()
6007       "vmovl.u8 q0, d0\n"  in Transform()
6008       "vmovl.u8 q1, d2\n"  in Transform()
6009       "vmovl.s16 q0, d0\n"  in Transform()
6010       "vmovl.s16 q1, d2\n"  in Transform()
6011       "vcvt.f32.s32 q0, q0\n"  in Transform()
6012       "vcvt.f32.s32 q1, q1\n"  in Transform()
6013       "vmul.f32 q0, q0, q9\n"  in Transform()
6014       "vmul.f32 q1, q1, q11\n"  in Transform()
6015       "vadd.f32 q0, q0, q8\n"  in Transform()
6016       "vadd.f32 q1, q1, q10\n"  in Transform()
6017       "vadd.f32 q0, q0, q1\n"  in Transform()
6018       "vsub.f32 q0, q0, q12\n"  in Transform()
6019       "vmul.f32 q0, q0, q13\n"  in Transform()
6020       "vadd.f32 q0, q0, q14\n"  in Transform()
6021       "vcvt.s32.f32 q0, q0\n"  in Transform()
6023       "vst1.32 {d0}, [%[output]]!\n"  in Transform()
6024       "vst1.32 {d1[0]}, [%[output]]!\n"  in Transform()
6025       "pld [%[output]]\n"  in Transform()
6026       "subs %[rows], %[rows], #1\n"  in Transform()
6027       "bne 1b\n"  in Transform()
6059       "ldr r0, %[input_range_min]\n"  in Transform()
6060       "vdup.32 q8, r0\n"  in Transform()
6061       "ldr r0, %[input_range_scale]\n"  in Transform()
6062       "vdup.32 q9, r0\n"  in Transform()
6063       "ldr r0, %[bias_range_min]\n"  in Transform()
6064       "vdup.32 q10, r0\n"  in Transform()
6065       "ldr r0, %[bias_range_scale]\n"  in Transform()
6066       "vdup.32 q11, r0\n"  in Transform()
6067       "ldr r0, %[output_range_min]\n"  in Transform()
6068       "vdup.32 q12, r0\n"  in Transform()
6069       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
6070       "vdup.32 q13, r0\n"  in Transform()
6071       "ldr r0, %[output_range_offset]\n"  in Transform()
6072       "vdup.32 q14, r0\n"  in Transform()
6074       "mov r0, %[count]\n"  in Transform()
6075       "mov r1, %[bias]\n"  in Transform()
6076       "subs r0, r0, #4\n"  in Transform()
6077       "beq 3f\n"  in Transform()
6079       "subs r0, r0, #16\n"  in Transform()
6082       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
6083       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
6084       "pld [%[input], #32]\n"  in Transform()
6085       "vmovl.u8 q1, d1\n"  in Transform()
6086       "vmovl.u8 q0, d0\n"  in Transform()
6087       "vmovl.u8 q5, d9\n"  in Transform()
6088       "vmovl.u8 q4, d8\n"  in Transform()
6089       "vmovl.s16 q3, d3\n"  in Transform()
6090       "vmovl.s16 q2, d2\n"  in Transform()
6091       "vmovl.s16 q7, d11\n"  in Transform()
6092       "vmovl.s16 q6, d10\n"  in Transform()
6093       "vmovl.s16 q1, d1\n"  in Transform()
6094       "vmovl.s16 q0, d0\n"  in Transform()
6095       "vmovl.s16 q5, d9\n"  in Transform()
6096       "vmovl.s16 q4, d8\n"  in Transform()
6097       "vcvt.f32.s32 q0, q0\n"  in Transform()
6098       "vcvt.f32.s32 q1, q1\n"  in Transform()
6099       "vcvt.f32.s32 q2, q2\n"  in Transform()
6100       "vcvt.f32.s32 q3, q3\n"  in Transform()
6101       "vcvt.f32.s32 q4, q4\n"  in Transform()
6102       "vcvt.f32.s32 q5, q5\n"  in Transform()
6103       "vcvt.f32.s32 q6, q6\n"  in Transform()
6104       "vcvt.f32.s32 q7, q7\n"  in Transform()
6105       "vmul.f32 q0, q0, q9\n"  in Transform()
6106       "vmul.f32 q1, q1, q9\n"  in Transform()
6107       "vmul.f32 q2, q2, q9\n"  in Transform()
6108       "vmul.f32 q3, q3, q9\n"  in Transform()
6109       "vmul.f32 q4, q4, q11\n"  in Transform()
6110       "vmul.f32 q5, q5, q11\n"  in Transform()
6111       "vmul.f32 q6, q6, q11\n"  in Transform()
6112       "vmul.f32 q7, q7, q11\n"  in Transform()
6113       "vadd.f32 q0, q0, q8\n"  in Transform()
6114       "vadd.f32 q1, q1, q8\n"  in Transform()
6115       "vadd.f32 q2, q2, q8\n"  in Transform()
6116       "vadd.f32 q3, q3, q8\n"  in Transform()
6117       "vadd.f32 q4, q4, q10\n"  in Transform()
6118       "vadd.f32 q5, q5, q10\n"  in Transform()
6119       "vadd.f32 q6, q6, q10\n"  in Transform()
6120       "vadd.f32 q7, q7, q10\n"  in Transform()
6121       "vadd.f32 q0, q0, q4\n"  in Transform()
6122       "vadd.f32 q1, q1, q5\n"  in Transform()
6123       "vadd.f32 q2, q2, q6\n"  in Transform()
6124       "vadd.f32 q3, q3, q7\n"  in Transform()
6125       "vsub.f32 q0, q0, q12\n"  in Transform()
6126       "vsub.f32 q1, q1, q12\n"  in Transform()
6127       "vsub.f32 q2, q2, q12\n"  in Transform()
6128       "vsub.f32 q3, q3, q12\n"  in Transform()
6129       "vmul.f32 q0, q0, q13\n"  in Transform()
6130       "vmul.f32 q1, q1, q13\n"  in Transform()
6131       "vmul.f32 q2, q2, q13\n"  in Transform()
6132       "vmul.f32 q3, q3, q13\n"  in Transform()
6133       "vadd.f32 q0, q0, q14\n"  in Transform()
6134       "vadd.f32 q1, q1, q14\n"  in Transform()
6135       "vadd.f32 q2, q2, q14\n"  in Transform()
6136       "vadd.f32 q3, q3, q14\n"  in Transform()
6137       "vcvt.s32.f32 q0, q0\n"  in Transform()
6138       "vcvt.s32.f32 q1, q1\n"  in Transform()
6139       "vcvt.s32.f32 q2, q2\n"  in Transform()
6140       "vcvt.s32.f32 q3, q3\n"  in Transform()
6142       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
6143       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
6144       "pld [%[output]]\n"  in Transform()
6145       "bne 2b\n"  in Transform()
6149       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
6150       "vld1.32 {d2[0]}, [r1]!\n"  in Transform()
6151       "pld [%[input], #32]\n"  in Transform()
6152       "vmovl.u8 q0, d0\n"  in Transform()
6153       "vmovl.u8 q1, d2\n"  in Transform()
6154       "vmovl.s16 q0, d0\n"  in Transform()
6155       "vmovl.s16 q1, d2\n"  in Transform()
6156       "vcvt.f32.s32 q0, q0\n"  in Transform()
6157       "vcvt.f32.s32 q1, q1\n"  in Transform()
6158       "vmul.f32 q0, q0, q9\n"  in Transform()
6159       "vmul.f32 q1, q1, q11\n"  in Transform()
6160       "vadd.f32 q0, q0, q8\n"  in Transform()
6161       "vadd.f32 q1, q1, q10\n"  in Transform()
6162       "vadd.f32 q0, q0, q1\n"  in Transform()
6163       "vsub.f32 q0, q0, q12\n"  in Transform()
6164       "vmul.f32 q0, q0, q13\n"  in Transform()
6165       "vadd.f32 q0, q0, q14\n"  in Transform()
6166       "vcvt.s32.f32 q0, q0\n"  in Transform()
6168       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
6169       "pld [%[output]]\n"  in Transform()
6170       "subs %[rows], %[rows], #1\n"  in Transform()
6171       "bne 1b\n"  in Transform()
6203       "ldr r0, %[input_range_min]\n"  in Transform()
6204       "vdup.32 q8, r0\n"  in Transform()
6205       "ldr r0, %[input_range_scale]\n"  in Transform()
6206       "vdup.32 q9, r0\n"  in Transform()
6207       "ldr r0, %[bias_range_min]\n"  in Transform()
6208       "vdup.32 q10, r0\n"  in Transform()
6209       "ldr r0, %[bias_range_scale]\n"  in Transform()
6210       "vdup.32 q11, r0\n"  in Transform()
6211       "ldr r0, %[output_range_min]\n"  in Transform()
6212       "vdup.32 q12, r0\n"  in Transform()
6213       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
6214       "vdup.32 q13, r0\n"  in Transform()
6215       "ldr r0, %[output_range_offset]\n"  in Transform()
6216       "vdup.32 q14, r0\n"  in Transform()
6218       "mov r0, %[count]\n"  in Transform()
6219       "mov r1, %[bias]\n"  in Transform()
6220       "subs r0, r0, #5\n"  in Transform()
6221       "beq 3f\n"  in Transform()
6223       "subs r0, r0, #16\n"  in Transform()
6226       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
6227       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
6228       "pld [%[input], #32]\n"  in Transform()
6229       "vmovl.u8 q1, d1\n"  in Transform()
6230       "vmovl.u8 q0, d0\n"  in Transform()
6231       "vmovl.u8 q5, d9\n"  in Transform()
6232       "vmovl.u8 q4, d8\n"  in Transform()
6233       "vmovl.s16 q3, d3\n"  in Transform()
6234       "vmovl.s16 q2, d2\n"  in Transform()
6235       "vmovl.s16 q7, d11\n"  in Transform()
6236       "vmovl.s16 q6, d10\n"  in Transform()
6237       "vmovl.s16 q1, d1\n"  in Transform()
6238       "vmovl.s16 q0, d0\n"  in Transform()
6239       "vmovl.s16 q5, d9\n"  in Transform()
6240       "vmovl.s16 q4, d8\n"  in Transform()
6241       "vcvt.f32.s32 q0, q0\n"  in Transform()
6242       "vcvt.f32.s32 q1, q1\n"  in Transform()
6243       "vcvt.f32.s32 q2, q2\n"  in Transform()
6244       "vcvt.f32.s32 q3, q3\n"  in Transform()
6245       "vcvt.f32.s32 q4, q4\n"  in Transform()
6246       "vcvt.f32.s32 q5, q5\n"  in Transform()
6247       "vcvt.f32.s32 q6, q6\n"  in Transform()
6248       "vcvt.f32.s32 q7, q7\n"  in Transform()
6249       "vmul.f32 q0, q0, q9\n"  in Transform()
6250       "vmul.f32 q1, q1, q9\n"  in Transform()
6251       "vmul.f32 q2, q2, q9\n"  in Transform()
6252       "vmul.f32 q3, q3, q9\n"  in Transform()
6253       "vmul.f32 q4, q4, q11\n"  in Transform()
6254       "vmul.f32 q5, q5, q11\n"  in Transform()
6255       "vmul.f32 q6, q6, q11\n"  in Transform()
6256       "vmul.f32 q7, q7, q11\n"  in Transform()
6257       "vadd.f32 q0, q0, q8\n"  in Transform()
6258       "vadd.f32 q1, q1, q8\n"  in Transform()
6259       "vadd.f32 q2, q2, q8\n"  in Transform()
6260       "vadd.f32 q3, q3, q8\n"  in Transform()
6261       "vadd.f32 q4, q4, q10\n"  in Transform()
6262       "vadd.f32 q5, q5, q10\n"  in Transform()
6263       "vadd.f32 q6, q6, q10\n"  in Transform()
6264       "vadd.f32 q7, q7, q10\n"  in Transform()
6265       "vadd.f32 q0, q0, q4\n"  in Transform()
6266       "vadd.f32 q1, q1, q5\n"  in Transform()
6267       "vadd.f32 q2, q2, q6\n"  in Transform()
6268       "vadd.f32 q3, q3, q7\n"  in Transform()
6269       "vsub.f32 q0, q0, q12\n"  in Transform()
6270       "vsub.f32 q1, q1, q12\n"  in Transform()
6271       "vsub.f32 q2, q2, q12\n"  in Transform()
6272       "vsub.f32 q3, q3, q12\n"  in Transform()
6273       "vmul.f32 q0, q0, q13\n"  in Transform()
6274       "vmul.f32 q1, q1, q13\n"  in Transform()
6275       "vmul.f32 q2, q2, q13\n"  in Transform()
6276       "vmul.f32 q3, q3, q13\n"  in Transform()
6277       "vadd.f32 q0, q0, q14\n"  in Transform()
6278       "vadd.f32 q1, q1, q14\n"  in Transform()
6279       "vadd.f32 q2, q2, q14\n"  in Transform()
6280       "vadd.f32 q3, q3, q14\n"  in Transform()
6281       "vcvt.s32.f32 q0, q0\n"  in Transform()
6282       "vcvt.s32.f32 q1, q1\n"  in Transform()
6283       "vcvt.s32.f32 q2, q2\n"  in Transform()
6284       "vcvt.s32.f32 q3, q3\n"  in Transform()
6286       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
6287       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
6288       "pld [%[output]]\n"  in Transform()
6289       "bne 2b\n"  in Transform()
6293       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
6294       "vld1.8 {d0[4]}, [%[input]]!\n"  in Transform()
6295       "vld1.32 {d4[0]}, [r1]!\n"  in Transform()
6296       "vld1.8 {d4[4]}, [r1]!\n"  in Transform()
6297       "pld [%[input], #32]\n"  in Transform()
6298       "vmovl.u8 q0, d0\n"  in Transform()
6299       "vmovl.u8 q2, d4\n"  in Transform()
6300       "vmovl.s16 q1, d1\n"  in Transform()
6301       "vmovl.s16 q0, d0\n"  in Transform()
6302       "vmovl.s16 q3, d5\n"  in Transform()
6303       "vmovl.s16 q2, d4\n"  in Transform()
6304       "vcvt.f32.s32 q0, q0\n"  in Transform()
6305       "vcvt.f32.s32 q1, q1\n"  in Transform()
6306       "vcvt.f32.s32 q2, q2\n"  in Transform()
6307       "vcvt.f32.s32 q3, q3\n"  in Transform()
6308       "vmul.f32 q0, q0, q9\n"  in Transform()
6309       "vmul.f32 q1, q1, q9\n"  in Transform()
6310       "vmul.f32 q2, q2, q11\n"  in Transform()
6311       "vmul.f32 q3, q3, q11\n"  in Transform()
6312       "vadd.f32 q0, q0, q8\n"  in Transform()
6313       "vadd.f32 q1, q1, q8\n"  in Transform()
6314       "vadd.f32 q2, q2, q10\n"  in Transform()
6315       "vadd.f32 q3, q3, q10\n"  in Transform()
6316       "vadd.f32 q0, q0, q2\n"  in Transform()
6317       "vadd.f32 q1, q1, q3\n"  in Transform()
6318       "vsub.f32 q0, q0, q12\n"  in Transform()
6319       "vsub.f32 q1, q1, q12\n"  in Transform()
6320       "vmul.f32 q0, q0, q13\n"  in Transform()
6321       "vmul.f32 q1, q1, q13\n"  in Transform()
6322       "vadd.f32 q0, q0, q14\n"  in Transform()
6323       "vadd.f32 q1, q1, q14\n"  in Transform()
6324       "vcvt.s32.f32 q0, q0\n"  in Transform()
6325       "vcvt.s32.f32 q1, q1\n"  in Transform()
6327       "vst1.32 {d0, d1}, [%[output]]!\n"  in Transform()
6328       "vst1.32 {d2[0]}, [%[output]]!\n"  in Transform()
6329       "pld [%[output]]\n"  in Transform()
6330       "subs %[rows], %[rows], #1\n"  in Transform()
6331       "bne 1b\n"  in Transform()
6363       "ldr r0, %[input_range_min]\n"  in Transform()
6364       "vdup.32 q8, r0\n"  in Transform()
6365       "ldr r0, %[input_range_scale]\n"  in Transform()
6366       "vdup.32 q9, r0\n"  in Transform()
6367       "ldr r0, %[bias_range_min]\n"  in Transform()
6368       "vdup.32 q10, r0\n"  in Transform()
6369       "ldr r0, %[bias_range_scale]\n"  in Transform()
6370       "vdup.32 q11, r0\n"  in Transform()
6371       "ldr r0, %[output_range_min]\n"  in Transform()
6372       "vdup.32 q12, r0\n"  in Transform()
6373       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
6374       "vdup.32 q13, r0\n"  in Transform()
6375       "ldr r0, %[output_range_offset]\n"  in Transform()
6376       "vdup.32 q14, r0\n"  in Transform()
6378       "mov r0, %[count]\n"  in Transform()
6379       "mov r1, %[bias]\n"  in Transform()
6380       "subs r0, r0, #6\n"  in Transform()
6381       "beq 3f\n"  in Transform()
6383       "subs r0, r0, #16\n"  in Transform()
6386       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
6387       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
6388       "pld [%[input], #32]\n"  in Transform()
6389       "vmovl.u8 q1, d1\n"  in Transform()
6390       "vmovl.u8 q0, d0\n"  in Transform()
6391       "vmovl.u8 q5, d9\n"  in Transform()
6392       "vmovl.u8 q4, d8\n"  in Transform()
6393       "vmovl.s16 q3, d3\n"  in Transform()
6394       "vmovl.s16 q2, d2\n"  in Transform()
6395       "vmovl.s16 q7, d11\n"  in Transform()
6396       "vmovl.s16 q6, d10\n"  in Transform()
6397       "vmovl.s16 q1, d1\n"  in Transform()
6398       "vmovl.s16 q0, d0\n"  in Transform()
6399       "vmovl.s16 q5, d9\n"  in Transform()
6400       "vmovl.s16 q4, d8\n"  in Transform()
6401       "vcvt.f32.s32 q0, q0\n"  in Transform()
6402       "vcvt.f32.s32 q1, q1\n"  in Transform()
6403       "vcvt.f32.s32 q2, q2\n"  in Transform()
6404       "vcvt.f32.s32 q3, q3\n"  in Transform()
6405       "vcvt.f32.s32 q4, q4\n"  in Transform()
6406       "vcvt.f32.s32 q5, q5\n"  in Transform()
6407       "vcvt.f32.s32 q6, q6\n"  in Transform()
6408       "vcvt.f32.s32 q7, q7\n"  in Transform()
6409       "vmul.f32 q0, q0, q9\n"  in Transform()
6410       "vmul.f32 q1, q1, q9\n"  in Transform()
6411       "vmul.f32 q2, q2, q9\n"  in Transform()
6412       "vmul.f32 q3, q3, q9\n"  in Transform()
6413       "vmul.f32 q4, q4, q11\n"  in Transform()
6414       "vmul.f32 q5, q5, q11\n"  in Transform()
6415       "vmul.f32 q6, q6, q11\n"  in Transform()
6416       "vmul.f32 q7, q7, q11\n"  in Transform()
6417       "vadd.f32 q0, q0, q8\n"  in Transform()
6418       "vadd.f32 q1, q1, q8\n"  in Transform()
6419       "vadd.f32 q2, q2, q8\n"  in Transform()
6420       "vadd.f32 q3, q3, q8\n"  in Transform()
6421       "vadd.f32 q4, q4, q10\n"  in Transform()
6422       "vadd.f32 q5, q5, q10\n"  in Transform()
6423       "vadd.f32 q6, q6, q10\n"  in Transform()
6424       "vadd.f32 q7, q7, q10\n"  in Transform()
6425       "vadd.f32 q0, q0, q4\n"  in Transform()
6426       "vadd.f32 q1, q1, q5\n"  in Transform()
6427       "vadd.f32 q2, q2, q6\n"  in Transform()
6428       "vadd.f32 q3, q3, q7\n"  in Transform()
6429       "vsub.f32 q0, q0, q12\n"  in Transform()
6430       "vsub.f32 q1, q1, q12\n"  in Transform()
6431       "vsub.f32 q2, q2, q12\n"  in Transform()
6432       "vsub.f32 q3, q3, q12\n"  in Transform()
6433       "vmul.f32 q0, q0, q13\n"  in Transform()
6434       "vmul.f32 q1, q1, q13\n"  in Transform()
6435       "vmul.f32 q2, q2, q13\n"  in Transform()
6436       "vmul.f32 q3, q3, q13\n"  in Transform()
6437       "vadd.f32 q0, q0, q14\n"  in Transform()
6438       "vadd.f32 q1, q1, q14\n"  in Transform()
6439       "vadd.f32 q2, q2, q14\n"  in Transform()
6440       "vadd.f32 q3, q3, q14\n"  in Transform()
6441       "vcvt.s32.f32 q0, q0\n"  in Transform()
6442       "vcvt.s32.f32 q1, q1\n"  in Transform()
6443       "vcvt.s32.f32 q2, q2\n"  in Transform()
6444       "vcvt.s32.f32 q3, q3\n"  in Transform()
6446       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
6447       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
6448       "pld [%[output]]\n"  in Transform()
6449       "bne 2b\n"  in Transform()
6453       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
6454       "vld1.16 {d0[2]}, [%[input]]!\n"  in Transform()
6455       "vld1.32 {d4[0]}, [r1]!\n"  in Transform()
6456       "vld1.16 {d4[2]}, [r1]!\n"  in Transform()
6457       "pld [%[input], #32]\n"  in Transform()
6458       "vmovl.u8 q0, d0\n"  in Transform()
6459       "vmovl.u8 q2, d4\n"  in Transform()
6460       "vmovl.s16 q1, d1\n"  in Transform()
6461       "vmovl.s16 q0, d0\n"  in Transform()
6462       "vmovl.s16 q3, d5\n"  in Transform()
6463       "vmovl.s16 q2, d4\n"  in Transform()
6464       "vcvt.f32.s32 q0, q0\n"  in Transform()
6465       "vcvt.f32.s32 q1, q1\n"  in Transform()
6466       "vcvt.f32.s32 q2, q2\n"  in Transform()
6467       "vcvt.f32.s32 q3, q3\n"  in Transform()
6468       "vmul.f32 q0, q0, q9\n"  in Transform()
6469       "vmul.f32 q1, q1, q9\n"  in Transform()
6470       "vmul.f32 q2, q2, q11\n"  in Transform()
6471       "vmul.f32 q3, q3, q11\n"  in Transform()
6472       "vadd.f32 q0, q0, q8\n"  in Transform()
6473       "vadd.f32 q1, q1, q8\n"  in Transform()
6474       "vadd.f32 q2, q2, q10\n"  in Transform()
6475       "vadd.f32 q3, q3, q10\n"  in Transform()
6476       "vadd.f32 q0, q0, q2\n"  in Transform()
6477       "vadd.f32 q1, q1, q3\n"  in Transform()
6478       "vsub.f32 q0, q0, q12\n"  in Transform()
6479       "vsub.f32 q1, q1, q12\n"  in Transform()
6480       "vmul.f32 q0, q0, q13\n"  in Transform()
6481       "vmul.f32 q1, q1, q13\n"  in Transform()
6482       "vadd.f32 q0, q0, q14\n"  in Transform()
6483       "vadd.f32 q1, q1, q14\n"  in Transform()
6484       "vcvt.s32.f32 q0, q0\n"  in Transform()
6485       "vcvt.s32.f32 q1, q1\n"  in Transform()
6487       "vst1.32 {d0, d1, d2}, [%[output]]!\n"  in Transform()
6488       "pld [%[output]]\n"  in Transform()
6489       "subs %[rows], %[rows], #1\n"  in Transform()
6490       "bne 1b\n"  in Transform()
6522       "ldr r0, %[input_range_min]\n"  in Transform()
6523       "vdup.32 q8, r0\n"  in Transform()
6524       "ldr r0, %[input_range_scale]\n"  in Transform()
6525       "vdup.32 q9, r0\n"  in Transform()
6526       "ldr r0, %[bias_range_min]\n"  in Transform()
6527       "vdup.32 q10, r0\n"  in Transform()
6528       "ldr r0, %[bias_range_scale]\n"  in Transform()
6529       "vdup.32 q11, r0\n"  in Transform()
6530       "ldr r0, %[output_range_min]\n"  in Transform()
6531       "vdup.32 q12, r0\n"  in Transform()
6532       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
6533       "vdup.32 q13, r0\n"  in Transform()
6534       "ldr r0, %[output_range_offset]\n"  in Transform()
6535       "vdup.32 q14, r0\n"  in Transform()
6537       "mov r0, %[count]\n"  in Transform()
6538       "mov r1, %[bias]\n"  in Transform()
6539       "subs r0, r0, #7\n"  in Transform()
6540       "beq 3f\n"  in Transform()
6542       "subs r0, r0, #16\n"  in Transform()
6545       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
6546       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
6547       "pld [%[input], #32]\n"  in Transform()
6548       "vmovl.u8 q1, d1\n"  in Transform()
6549       "vmovl.u8 q0, d0\n"  in Transform()
6550       "vmovl.u8 q5, d9\n"  in Transform()
6551       "vmovl.u8 q4, d8\n"  in Transform()
6552       "vmovl.s16 q3, d3\n"  in Transform()
6553       "vmovl.s16 q2, d2\n"  in Transform()
6554       "vmovl.s16 q7, d11\n"  in Transform()
6555       "vmovl.s16 q6, d10\n"  in Transform()
6556       "vmovl.s16 q1, d1\n"  in Transform()
6557       "vmovl.s16 q0, d0\n"  in Transform()
6558       "vmovl.s16 q5, d9\n"  in Transform()
6559       "vmovl.s16 q4, d8\n"  in Transform()
6560       "vcvt.f32.s32 q0, q0\n"  in Transform()
6561       "vcvt.f32.s32 q1, q1\n"  in Transform()
6562       "vcvt.f32.s32 q2, q2\n"  in Transform()
6563       "vcvt.f32.s32 q3, q3\n"  in Transform()
6564       "vcvt.f32.s32 q4, q4\n"  in Transform()
6565       "vcvt.f32.s32 q5, q5\n"  in Transform()
6566       "vcvt.f32.s32 q6, q6\n"  in Transform()
6567       "vcvt.f32.s32 q7, q7\n"  in Transform()
6568       "vmul.f32 q0, q0, q9\n"  in Transform()
6569       "vmul.f32 q1, q1, q9\n"  in Transform()
6570       "vmul.f32 q2, q2, q9\n"  in Transform()
6571       "vmul.f32 q3, q3, q9\n"  in Transform()
6572       "vmul.f32 q4, q4, q11\n"  in Transform()
6573       "vmul.f32 q5, q5, q11\n"  in Transform()
6574       "vmul.f32 q6, q6, q11\n"  in Transform()
6575       "vmul.f32 q7, q7, q11\n"  in Transform()
6576       "vadd.f32 q0, q0, q8\n"  in Transform()
6577       "vadd.f32 q1, q1, q8\n"  in Transform()
6578       "vadd.f32 q2, q2, q8\n"  in Transform()
6579       "vadd.f32 q3, q3, q8\n"  in Transform()
6580       "vadd.f32 q4, q4, q10\n"  in Transform()
6581       "vadd.f32 q5, q5, q10\n"  in Transform()
6582       "vadd.f32 q6, q6, q10\n"  in Transform()
6583       "vadd.f32 q7, q7, q10\n"  in Transform()
6584       "vadd.f32 q0, q0, q4\n"  in Transform()
6585       "vadd.f32 q1, q1, q5\n"  in Transform()
6586       "vadd.f32 q2, q2, q6\n"  in Transform()
6587       "vadd.f32 q3, q3, q7\n"  in Transform()
6588       "vsub.f32 q0, q0, q12\n"  in Transform()
6589       "vsub.f32 q1, q1, q12\n"  in Transform()
6590       "vsub.f32 q2, q2, q12\n"  in Transform()
6591       "vsub.f32 q3, q3, q12\n"  in Transform()
6592       "vmul.f32 q0, q0, q13\n"  in Transform()
6593       "vmul.f32 q1, q1, q13\n"  in Transform()
6594       "vmul.f32 q2, q2, q13\n"  in Transform()
6595       "vmul.f32 q3, q3, q13\n"  in Transform()
6596       "vadd.f32 q0, q0, q14\n"  in Transform()
6597       "vadd.f32 q1, q1, q14\n"  in Transform()
6598       "vadd.f32 q2, q2, q14\n"  in Transform()
6599       "vadd.f32 q3, q3, q14\n"  in Transform()
6600       "vcvt.s32.f32 q0, q0\n"  in Transform()
6601       "vcvt.s32.f32 q1, q1\n"  in Transform()
6602       "vcvt.s32.f32 q2, q2\n"  in Transform()
6603       "vcvt.s32.f32 q3, q3\n"  in Transform()
6605       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
6606       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
6607       "pld [%[output]]\n"  in Transform()
6608       "bne 2b\n"  in Transform()
6612       "vld1.32 {d0[0]}, [%[input]]!\n"  in Transform()
6613       "vld1.16 {d0[2]}, [%[input]]!\n"  in Transform()
6614       "vld1.8 {d0[6]}, [%[input]]!\n"  in Transform()
6615       "vld1.32 {d4[0]}, [r1]!\n"  in Transform()
6616       "vld1.16 {d4[2]}, [r1]!\n"  in Transform()
6617       "vld1.8 {d4[6]}, [r1]!\n"  in Transform()
6618       "pld [%[input], #32]\n"  in Transform()
6619       "vmovl.u8 q0, d0\n"  in Transform()
6620       "vmovl.u8 q2, d4\n"  in Transform()
6621       "vmovl.s16 q1, d1\n"  in Transform()
6622       "vmovl.s16 q0, d0\n"  in Transform()
6623       "vmovl.s16 q3, d5\n"  in Transform()
6624       "vmovl.s16 q2, d4\n"  in Transform()
6625       "vcvt.f32.s32 q0, q0\n"  in Transform()
6626       "vcvt.f32.s32 q1, q1\n"  in Transform()
6627       "vcvt.f32.s32 q2, q2\n"  in Transform()
6628       "vcvt.f32.s32 q3, q3\n"  in Transform()
6629       "vmul.f32 q0, q0, q9\n"  in Transform()
6630       "vmul.f32 q1, q1, q9\n"  in Transform()
6631       "vmul.f32 q2, q2, q11\n"  in Transform()
6632       "vmul.f32 q3, q3, q11\n"  in Transform()
6633       "vadd.f32 q0, q0, q8\n"  in Transform()
6634       "vadd.f32 q1, q1, q8\n"  in Transform()
6635       "vadd.f32 q2, q2, q10\n"  in Transform()
6636       "vadd.f32 q3, q3, q10\n"  in Transform()
6637       "vadd.f32 q0, q0, q2\n"  in Transform()
6638       "vadd.f32 q1, q1, q3\n"  in Transform()
6639       "vsub.f32 q0, q0, q12\n"  in Transform()
6640       "vsub.f32 q1, q1, q12\n"  in Transform()
6641       "vmul.f32 q0, q0, q13\n"  in Transform()
6642       "vmul.f32 q1, q1, q13\n"  in Transform()
6643       "vadd.f32 q0, q0, q14\n"  in Transform()
6644       "vadd.f32 q1, q1, q14\n"  in Transform()
6645       "vcvt.s32.f32 q0, q0\n"  in Transform()
6646       "vcvt.s32.f32 q1, q1\n"  in Transform()
6648       "vst1.32 {d0, d1, d2}, [%[output]]!\n"  in Transform()
6649       "vst1.32 {d3[0]}, [%[output]]!\n"  in Transform()
6650       "pld [%[output]]\n"  in Transform()
6651       "subs %[rows], %[rows], #1\n"  in Transform()
6652       "bne 1b\n"  in Transform()
6684       "ldr r0, %[input_range_min]\n"  in Transform()
6685       "vdup.32 q8, r0\n"  in Transform()
6686       "ldr r0, %[input_range_scale]\n"  in Transform()
6687       "vdup.32 q9, r0\n"  in Transform()
6688       "ldr r0, %[bias_range_min]\n"  in Transform()
6689       "vdup.32 q10, r0\n"  in Transform()
6690       "ldr r0, %[bias_range_scale]\n"  in Transform()
6691       "vdup.32 q11, r0\n"  in Transform()
6692       "ldr r0, %[output_range_min]\n"  in Transform()
6693       "vdup.32 q12, r0\n"  in Transform()
6694       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
6695       "vdup.32 q13, r0\n"  in Transform()
6696       "ldr r0, %[output_range_offset]\n"  in Transform()
6697       "vdup.32 q14, r0\n"  in Transform()
6699       "mov r0, %[count]\n"  in Transform()
6700       "mov r1, %[bias]\n"  in Transform()
6701       "subs r0, r0, #8\n"  in Transform()
6702       "beq 3f\n"  in Transform()
6704       "subs r0, r0, #16\n"  in Transform()
6707       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
6708       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
6709       "pld [%[input], #32]\n"  in Transform()
6710       "vmovl.u8 q1, d1\n"  in Transform()
6711       "vmovl.u8 q0, d0\n"  in Transform()
6712       "vmovl.u8 q5, d9\n"  in Transform()
6713       "vmovl.u8 q4, d8\n"  in Transform()
6714       "vmovl.s16 q3, d3\n"  in Transform()
6715       "vmovl.s16 q2, d2\n"  in Transform()
6716       "vmovl.s16 q7, d11\n"  in Transform()
6717       "vmovl.s16 q6, d10\n"  in Transform()
6718       "vmovl.s16 q1, d1\n"  in Transform()
6719       "vmovl.s16 q0, d0\n"  in Transform()
6720       "vmovl.s16 q5, d9\n"  in Transform()
6721       "vmovl.s16 q4, d8\n"  in Transform()
6722       "vcvt.f32.s32 q0, q0\n"  in Transform()
6723       "vcvt.f32.s32 q1, q1\n"  in Transform()
6724       "vcvt.f32.s32 q2, q2\n"  in Transform()
6725       "vcvt.f32.s32 q3, q3\n"  in Transform()
6726       "vcvt.f32.s32 q4, q4\n"  in Transform()
6727       "vcvt.f32.s32 q5, q5\n"  in Transform()
6728       "vcvt.f32.s32 q6, q6\n"  in Transform()
6729       "vcvt.f32.s32 q7, q7\n"  in Transform()
6730       "vmul.f32 q0, q0, q9\n"  in Transform()
6731       "vmul.f32 q1, q1, q9\n"  in Transform()
6732       "vmul.f32 q2, q2, q9\n"  in Transform()
6733       "vmul.f32 q3, q3, q9\n"  in Transform()
6734       "vmul.f32 q4, q4, q11\n"  in Transform()
6735       "vmul.f32 q5, q5, q11\n"  in Transform()
6736       "vmul.f32 q6, q6, q11\n"  in Transform()
6737       "vmul.f32 q7, q7, q11\n"  in Transform()
6738       "vadd.f32 q0, q0, q8\n"  in Transform()
6739       "vadd.f32 q1, q1, q8\n"  in Transform()
6740       "vadd.f32 q2, q2, q8\n"  in Transform()
6741       "vadd.f32 q3, q3, q8\n"  in Transform()
6742       "vadd.f32 q4, q4, q10\n"  in Transform()
6743       "vadd.f32 q5, q5, q10\n"  in Transform()
6744       "vadd.f32 q6, q6, q10\n"  in Transform()
6745       "vadd.f32 q7, q7, q10\n"  in Transform()
6746       "vadd.f32 q0, q0, q4\n"  in Transform()
6747       "vadd.f32 q1, q1, q5\n"  in Transform()
6748       "vadd.f32 q2, q2, q6\n"  in Transform()
6749       "vadd.f32 q3, q3, q7\n"  in Transform()
6750       "vsub.f32 q0, q0, q12\n"  in Transform()
6751       "vsub.f32 q1, q1, q12\n"  in Transform()
6752       "vsub.f32 q2, q2, q12\n"  in Transform()
6753       "vsub.f32 q3, q3, q12\n"  in Transform()
6754       "vmul.f32 q0, q0, q13\n"  in Transform()
6755       "vmul.f32 q1, q1, q13\n"  in Transform()
6756       "vmul.f32 q2, q2, q13\n"  in Transform()
6757       "vmul.f32 q3, q3, q13\n"  in Transform()
6758       "vadd.f32 q0, q0, q14\n"  in Transform()
6759       "vadd.f32 q1, q1, q14\n"  in Transform()
6760       "vadd.f32 q2, q2, q14\n"  in Transform()
6761       "vadd.f32 q3, q3, q14\n"  in Transform()
6762       "vcvt.s32.f32 q0, q0\n"  in Transform()
6763       "vcvt.s32.f32 q1, q1\n"  in Transform()
6764       "vcvt.s32.f32 q2, q2\n"  in Transform()
6765       "vcvt.s32.f32 q3, q3\n"  in Transform()
6767       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
6768       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
6769       "pld [%[output]]\n"  in Transform()
6770       "bne 2b\n"  in Transform()
6774       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
6775       "vld1.32 {d4}, [r1]!\n"  in Transform()
6776       "pld [%[input], #32]\n"  in Transform()
6777       "vmovl.u8 q0, d0\n"  in Transform()
6778       "vmovl.u8 q2, d4\n"  in Transform()
6779       "vmovl.s16 q1, d1\n"  in Transform()
6780       "vmovl.s16 q0, d0\n"  in Transform()
6781       "vmovl.s16 q3, d5\n"  in Transform()
6782       "vmovl.s16 q2, d4\n"  in Transform()
6783       "vcvt.f32.s32 q0, q0\n"  in Transform()
6784       "vcvt.f32.s32 q1, q1\n"  in Transform()
6785       "vcvt.f32.s32 q2, q2\n"  in Transform()
6786       "vcvt.f32.s32 q3, q3\n"  in Transform()
6787       "vmul.f32 q0, q0, q9\n"  in Transform()
6788       "vmul.f32 q1, q1, q9\n"  in Transform()
6789       "vmul.f32 q2, q2, q11\n"  in Transform()
6790       "vmul.f32 q3, q3, q11\n"  in Transform()
6791       "vadd.f32 q0, q0, q8\n"  in Transform()
6792       "vadd.f32 q1, q1, q8\n"  in Transform()
6793       "vadd.f32 q2, q2, q10\n"  in Transform()
6794       "vadd.f32 q3, q3, q10\n"  in Transform()
6795       "vadd.f32 q0, q0, q2\n"  in Transform()
6796       "vadd.f32 q1, q1, q3\n"  in Transform()
6797       "vsub.f32 q0, q0, q12\n"  in Transform()
6798       "vsub.f32 q1, q1, q12\n"  in Transform()
6799       "vmul.f32 q0, q0, q13\n"  in Transform()
6800       "vmul.f32 q1, q1, q13\n"  in Transform()
6801       "vadd.f32 q0, q0, q14\n"  in Transform()
6802       "vadd.f32 q1, q1, q14\n"  in Transform()
6803       "vcvt.s32.f32 q0, q0\n"  in Transform()
6804       "vcvt.s32.f32 q1, q1\n"  in Transform()
6806       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
6807       "pld [%[output]]\n"  in Transform()
6808       "subs %[rows], %[rows], #1\n"  in Transform()
6809       "bne 1b\n"  in Transform()
6841       "ldr r0, %[input_range_min]\n"  in Transform()
6842       "vdup.32 q8, r0\n"  in Transform()
6843       "ldr r0, %[input_range_scale]\n"  in Transform()
6844       "vdup.32 q9, r0\n"  in Transform()
6845       "ldr r0, %[bias_range_min]\n"  in Transform()
6846       "vdup.32 q10, r0\n"  in Transform()
6847       "ldr r0, %[bias_range_scale]\n"  in Transform()
6848       "vdup.32 q11, r0\n"  in Transform()
6849       "ldr r0, %[output_range_min]\n"  in Transform()
6850       "vdup.32 q12, r0\n"  in Transform()
6851       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
6852       "vdup.32 q13, r0\n"  in Transform()
6853       "ldr r0, %[output_range_offset]\n"  in Transform()
6854       "vdup.32 q14, r0\n"  in Transform()
6856       "mov r0, %[count]\n"  in Transform()
6857       "mov r1, %[bias]\n"  in Transform()
6858       "subs r0, r0, #9\n"  in Transform()
6859       "beq 3f\n"  in Transform()
6861       "subs r0, r0, #16\n"  in Transform()
6864       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
6865       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
6866       "pld [%[input], #32]\n"  in Transform()
6867       "vmovl.u8 q1, d1\n"  in Transform()
6868       "vmovl.u8 q0, d0\n"  in Transform()
6869       "vmovl.u8 q5, d9\n"  in Transform()
6870       "vmovl.u8 q4, d8\n"  in Transform()
6871       "vmovl.s16 q3, d3\n"  in Transform()
6872       "vmovl.s16 q2, d2\n"  in Transform()
6873       "vmovl.s16 q7, d11\n"  in Transform()
6874       "vmovl.s16 q6, d10\n"  in Transform()
6875       "vmovl.s16 q1, d1\n"  in Transform()
6876       "vmovl.s16 q0, d0\n"  in Transform()
6877       "vmovl.s16 q5, d9\n"  in Transform()
6878       "vmovl.s16 q4, d8\n"  in Transform()
6879       "vcvt.f32.s32 q0, q0\n"  in Transform()
6880       "vcvt.f32.s32 q1, q1\n"  in Transform()
6881       "vcvt.f32.s32 q2, q2\n"  in Transform()
6882       "vcvt.f32.s32 q3, q3\n"  in Transform()
6883       "vcvt.f32.s32 q4, q4\n"  in Transform()
6884       "vcvt.f32.s32 q5, q5\n"  in Transform()
6885       "vcvt.f32.s32 q6, q6\n"  in Transform()
6886       "vcvt.f32.s32 q7, q7\n"  in Transform()
6887       "vmul.f32 q0, q0, q9\n"  in Transform()
6888       "vmul.f32 q1, q1, q9\n"  in Transform()
6889       "vmul.f32 q2, q2, q9\n"  in Transform()
6890       "vmul.f32 q3, q3, q9\n"  in Transform()
6891       "vmul.f32 q4, q4, q11\n"  in Transform()
6892       "vmul.f32 q5, q5, q11\n"  in Transform()
6893       "vmul.f32 q6, q6, q11\n"  in Transform()
6894       "vmul.f32 q7, q7, q11\n"  in Transform()
6895       "vadd.f32 q0, q0, q8\n"  in Transform()
6896       "vadd.f32 q1, q1, q8\n"  in Transform()
6897       "vadd.f32 q2, q2, q8\n"  in Transform()
6898       "vadd.f32 q3, q3, q8\n"  in Transform()
6899       "vadd.f32 q4, q4, q10\n"  in Transform()
6900       "vadd.f32 q5, q5, q10\n"  in Transform()
6901       "vadd.f32 q6, q6, q10\n"  in Transform()
6902       "vadd.f32 q7, q7, q10\n"  in Transform()
6903       "vadd.f32 q0, q0, q4\n"  in Transform()
6904       "vadd.f32 q1, q1, q5\n"  in Transform()
6905       "vadd.f32 q2, q2, q6\n"  in Transform()
6906       "vadd.f32 q3, q3, q7\n"  in Transform()
6907       "vsub.f32 q0, q0, q12\n"  in Transform()
6908       "vsub.f32 q1, q1, q12\n"  in Transform()
6909       "vsub.f32 q2, q2, q12\n"  in Transform()
6910       "vsub.f32 q3, q3, q12\n"  in Transform()
6911       "vmul.f32 q0, q0, q13\n"  in Transform()
6912       "vmul.f32 q1, q1, q13\n"  in Transform()
6913       "vmul.f32 q2, q2, q13\n"  in Transform()
6914       "vmul.f32 q3, q3, q13\n"  in Transform()
6915       "vadd.f32 q0, q0, q14\n"  in Transform()
6916       "vadd.f32 q1, q1, q14\n"  in Transform()
6917       "vadd.f32 q2, q2, q14\n"  in Transform()
6918       "vadd.f32 q3, q3, q14\n"  in Transform()
6919       "vcvt.s32.f32 q0, q0\n"  in Transform()
6920       "vcvt.s32.f32 q1, q1\n"  in Transform()
6921       "vcvt.s32.f32 q2, q2\n"  in Transform()
6922       "vcvt.s32.f32 q3, q3\n"  in Transform()
6924       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
6925       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
6926       "pld [%[output]]\n"  in Transform()
6927       "bne 2b\n"  in Transform()
6931       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
6932       "vld1.8 {d1[0]}, [%[input]]!\n"  in Transform()
6933       "vld1.32 {d6}, [r1]!\n"  in Transform()
6934       "vld1.8 {d7[0]}, [r1]!\n"  in Transform()
6935       "pld [%[input], #32]\n"  in Transform()
6936       "vmovl.u8 q1, d1\n"  in Transform()
6937       "vmovl.u8 q0, d0\n"  in Transform()
6938       "vmovl.u8 q4, d7\n"  in Transform()
6939       "vmovl.u8 q3, d6\n"  in Transform()
6940       "vmovl.s16 q2, d2\n"  in Transform()
6941       "vmovl.s16 q5, d8\n"  in Transform()
6942       "vmovl.s16 q1, d1\n"  in Transform()
6943       "vmovl.s16 q0, d0\n"  in Transform()
6944       "vmovl.s16 q4, d7\n"  in Transform()
6945       "vmovl.s16 q3, d6\n"  in Transform()
6946       "vcvt.f32.s32 q0, q0\n"  in Transform()
6947       "vcvt.f32.s32 q1, q1\n"  in Transform()
6948       "vcvt.f32.s32 q2, q2\n"  in Transform()
6949       "vcvt.f32.s32 q3, q3\n"  in Transform()
6950       "vcvt.f32.s32 q4, q4\n"  in Transform()
6951       "vcvt.f32.s32 q5, q5\n"  in Transform()
6952       "vmul.f32 q0, q0, q9\n"  in Transform()
6953       "vmul.f32 q1, q1, q9\n"  in Transform()
6954       "vmul.f32 q2, q2, q9\n"  in Transform()
6955       "vmul.f32 q3, q3, q11\n"  in Transform()
6956       "vmul.f32 q4, q4, q11\n"  in Transform()
6957       "vmul.f32 q5, q5, q11\n"  in Transform()
6958       "vadd.f32 q0, q0, q8\n"  in Transform()
6959       "vadd.f32 q1, q1, q8\n"  in Transform()
6960       "vadd.f32 q2, q2, q8\n"  in Transform()
6961       "vadd.f32 q3, q3, q10\n"  in Transform()
6962       "vadd.f32 q4, q4, q10\n"  in Transform()
6963       "vadd.f32 q5, q5, q10\n"  in Transform()
6964       "vadd.f32 q0, q0, q3\n"  in Transform()
6965       "vadd.f32 q1, q1, q4\n"  in Transform()
6966       "vadd.f32 q2, q2, q5\n"  in Transform()
6967       "vsub.f32 q0, q0, q12\n"  in Transform()
6968       "vsub.f32 q1, q1, q12\n"  in Transform()
6969       "vsub.f32 q2, q2, q12\n"  in Transform()
6970       "vmul.f32 q0, q0, q13\n"  in Transform()
6971       "vmul.f32 q1, q1, q13\n"  in Transform()
6972       "vmul.f32 q2, q2, q13\n"  in Transform()
6973       "vadd.f32 q0, q0, q14\n"  in Transform()
6974       "vadd.f32 q1, q1, q14\n"  in Transform()
6975       "vadd.f32 q2, q2, q14\n"  in Transform()
6976       "vcvt.s32.f32 q0, q0\n"  in Transform()
6977       "vcvt.s32.f32 q1, q1\n"  in Transform()
6978       "vcvt.s32.f32 q2, q2\n"  in Transform()
6980       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
6981       "vst1.32 {d4[0]}, [%[output]]!\n"  in Transform()
6982       "pld [%[output]]\n"  in Transform()
6983       "subs %[rows], %[rows], #1\n"  in Transform()
6984       "bne 1b\n"  in Transform()
7016       "ldr r0, %[input_range_min]\n"  in Transform()
7017       "vdup.32 q8, r0\n"  in Transform()
7018       "ldr r0, %[input_range_scale]\n"  in Transform()
7019       "vdup.32 q9, r0\n"  in Transform()
7020       "ldr r0, %[bias_range_min]\n"  in Transform()
7021       "vdup.32 q10, r0\n"  in Transform()
7022       "ldr r0, %[bias_range_scale]\n"  in Transform()
7023       "vdup.32 q11, r0\n"  in Transform()
7024       "ldr r0, %[output_range_min]\n"  in Transform()
7025       "vdup.32 q12, r0\n"  in Transform()
7026       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
7027       "vdup.32 q13, r0\n"  in Transform()
7028       "ldr r0, %[output_range_offset]\n"  in Transform()
7029       "vdup.32 q14, r0\n"  in Transform()
7031       "mov r0, %[count]\n"  in Transform()
7032       "mov r1, %[bias]\n"  in Transform()
7033       "subs r0, r0, #10\n"  in Transform()
7034       "beq 3f\n"  in Transform()
7036       "subs r0, r0, #16\n"  in Transform()
7039       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
7040       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
7041       "pld [%[input], #32]\n"  in Transform()
7042       "vmovl.u8 q1, d1\n"  in Transform()
7043       "vmovl.u8 q0, d0\n"  in Transform()
7044       "vmovl.u8 q5, d9\n"  in Transform()
7045       "vmovl.u8 q4, d8\n"  in Transform()
7046       "vmovl.s16 q3, d3\n"  in Transform()
7047       "vmovl.s16 q2, d2\n"  in Transform()
7048       "vmovl.s16 q7, d11\n"  in Transform()
7049       "vmovl.s16 q6, d10\n"  in Transform()
7050       "vmovl.s16 q1, d1\n"  in Transform()
7051       "vmovl.s16 q0, d0\n"  in Transform()
7052       "vmovl.s16 q5, d9\n"  in Transform()
7053       "vmovl.s16 q4, d8\n"  in Transform()
7054       "vcvt.f32.s32 q0, q0\n"  in Transform()
7055       "vcvt.f32.s32 q1, q1\n"  in Transform()
7056       "vcvt.f32.s32 q2, q2\n"  in Transform()
7057       "vcvt.f32.s32 q3, q3\n"  in Transform()
7058       "vcvt.f32.s32 q4, q4\n"  in Transform()
7059       "vcvt.f32.s32 q5, q5\n"  in Transform()
7060       "vcvt.f32.s32 q6, q6\n"  in Transform()
7061       "vcvt.f32.s32 q7, q7\n"  in Transform()
7062       "vmul.f32 q0, q0, q9\n"  in Transform()
7063       "vmul.f32 q1, q1, q9\n"  in Transform()
7064       "vmul.f32 q2, q2, q9\n"  in Transform()
7065       "vmul.f32 q3, q3, q9\n"  in Transform()
7066       "vmul.f32 q4, q4, q11\n"  in Transform()
7067       "vmul.f32 q5, q5, q11\n"  in Transform()
7068       "vmul.f32 q6, q6, q11\n"  in Transform()
7069       "vmul.f32 q7, q7, q11\n"  in Transform()
7070       "vadd.f32 q0, q0, q8\n"  in Transform()
7071       "vadd.f32 q1, q1, q8\n"  in Transform()
7072       "vadd.f32 q2, q2, q8\n"  in Transform()
7073       "vadd.f32 q3, q3, q8\n"  in Transform()
7074       "vadd.f32 q4, q4, q10\n"  in Transform()
7075       "vadd.f32 q5, q5, q10\n"  in Transform()
7076       "vadd.f32 q6, q6, q10\n"  in Transform()
7077       "vadd.f32 q7, q7, q10\n"  in Transform()
7078       "vadd.f32 q0, q0, q4\n"  in Transform()
7079       "vadd.f32 q1, q1, q5\n"  in Transform()
7080       "vadd.f32 q2, q2, q6\n"  in Transform()
7081       "vadd.f32 q3, q3, q7\n"  in Transform()
7082       "vsub.f32 q0, q0, q12\n"  in Transform()
7083       "vsub.f32 q1, q1, q12\n"  in Transform()
7084       "vsub.f32 q2, q2, q12\n"  in Transform()
7085       "vsub.f32 q3, q3, q12\n"  in Transform()
7086       "vmul.f32 q0, q0, q13\n"  in Transform()
7087       "vmul.f32 q1, q1, q13\n"  in Transform()
7088       "vmul.f32 q2, q2, q13\n"  in Transform()
7089       "vmul.f32 q3, q3, q13\n"  in Transform()
7090       "vadd.f32 q0, q0, q14\n"  in Transform()
7091       "vadd.f32 q1, q1, q14\n"  in Transform()
7092       "vadd.f32 q2, q2, q14\n"  in Transform()
7093       "vadd.f32 q3, q3, q14\n"  in Transform()
7094       "vcvt.s32.f32 q0, q0\n"  in Transform()
7095       "vcvt.s32.f32 q1, q1\n"  in Transform()
7096       "vcvt.s32.f32 q2, q2\n"  in Transform()
7097       "vcvt.s32.f32 q3, q3\n"  in Transform()
7099       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7100       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
7101       "pld [%[output]]\n"  in Transform()
7102       "bne 2b\n"  in Transform()
7106       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
7107       "vld1.16 {d1[0]}, [%[input]]!\n"  in Transform()
7108       "vld1.32 {d6}, [r1]!\n"  in Transform()
7109       "vld1.16 {d7[0]}, [r1]!\n"  in Transform()
7110       "pld [%[input], #32]\n"  in Transform()
7111       "vmovl.u8 q1, d1\n"  in Transform()
7112       "vmovl.u8 q0, d0\n"  in Transform()
7113       "vmovl.u8 q4, d7\n"  in Transform()
7114       "vmovl.u8 q3, d6\n"  in Transform()
7115       "vmovl.s16 q2, d2\n"  in Transform()
7116       "vmovl.s16 q5, d8\n"  in Transform()
7117       "vmovl.s16 q1, d1\n"  in Transform()
7118       "vmovl.s16 q0, d0\n"  in Transform()
7119       "vmovl.s16 q4, d7\n"  in Transform()
7120       "vmovl.s16 q3, d6\n"  in Transform()
7121       "vcvt.f32.s32 q0, q0\n"  in Transform()
7122       "vcvt.f32.s32 q1, q1\n"  in Transform()
7123       "vcvt.f32.s32 q2, q2\n"  in Transform()
7124       "vcvt.f32.s32 q3, q3\n"  in Transform()
7125       "vcvt.f32.s32 q4, q4\n"  in Transform()
7126       "vcvt.f32.s32 q5, q5\n"  in Transform()
7127       "vmul.f32 q0, q0, q9\n"  in Transform()
7128       "vmul.f32 q1, q1, q9\n"  in Transform()
7129       "vmul.f32 q2, q2, q9\n"  in Transform()
7130       "vmul.f32 q3, q3, q11\n"  in Transform()
7131       "vmul.f32 q4, q4, q11\n"  in Transform()
7132       "vmul.f32 q5, q5, q11\n"  in Transform()
7133       "vadd.f32 q0, q0, q8\n"  in Transform()
7134       "vadd.f32 q1, q1, q8\n"  in Transform()
7135       "vadd.f32 q2, q2, q8\n"  in Transform()
7136       "vadd.f32 q3, q3, q10\n"  in Transform()
7137       "vadd.f32 q4, q4, q10\n"  in Transform()
7138       "vadd.f32 q5, q5, q10\n"  in Transform()
7139       "vadd.f32 q0, q0, q3\n"  in Transform()
7140       "vadd.f32 q1, q1, q4\n"  in Transform()
7141       "vadd.f32 q2, q2, q5\n"  in Transform()
7142       "vsub.f32 q0, q0, q12\n"  in Transform()
7143       "vsub.f32 q1, q1, q12\n"  in Transform()
7144       "vsub.f32 q2, q2, q12\n"  in Transform()
7145       "vmul.f32 q0, q0, q13\n"  in Transform()
7146       "vmul.f32 q1, q1, q13\n"  in Transform()
7147       "vmul.f32 q2, q2, q13\n"  in Transform()
7148       "vadd.f32 q0, q0, q14\n"  in Transform()
7149       "vadd.f32 q1, q1, q14\n"  in Transform()
7150       "vadd.f32 q2, q2, q14\n"  in Transform()
7151       "vcvt.s32.f32 q0, q0\n"  in Transform()
7152       "vcvt.s32.f32 q1, q1\n"  in Transform()
7153       "vcvt.s32.f32 q2, q2\n"  in Transform()
7155       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7156       "vst1.32 {d4}, [%[output]]!\n"  in Transform()
7157       "pld [%[output]]\n"  in Transform()
7158       "subs %[rows], %[rows], #1\n"  in Transform()
7159       "bne 1b\n"  in Transform()
7191       "ldr r0, %[input_range_min]\n"  in Transform()
7192       "vdup.32 q8, r0\n"  in Transform()
7193       "ldr r0, %[input_range_scale]\n"  in Transform()
7194       "vdup.32 q9, r0\n"  in Transform()
7195       "ldr r0, %[bias_range_min]\n"  in Transform()
7196       "vdup.32 q10, r0\n"  in Transform()
7197       "ldr r0, %[bias_range_scale]\n"  in Transform()
7198       "vdup.32 q11, r0\n"  in Transform()
7199       "ldr r0, %[output_range_min]\n"  in Transform()
7200       "vdup.32 q12, r0\n"  in Transform()
7201       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
7202       "vdup.32 q13, r0\n"  in Transform()
7203       "ldr r0, %[output_range_offset]\n"  in Transform()
7204       "vdup.32 q14, r0\n"  in Transform()
7206       "mov r0, %[count]\n"  in Transform()
7207       "mov r1, %[bias]\n"  in Transform()
7208       "subs r0, r0, #11\n"  in Transform()
7209       "beq 3f\n"  in Transform()
7211       "subs r0, r0, #16\n"  in Transform()
7214       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
7215       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
7216       "pld [%[input], #32]\n"  in Transform()
7217       "vmovl.u8 q1, d1\n"  in Transform()
7218       "vmovl.u8 q0, d0\n"  in Transform()
7219       "vmovl.u8 q5, d9\n"  in Transform()
7220       "vmovl.u8 q4, d8\n"  in Transform()
7221       "vmovl.s16 q3, d3\n"  in Transform()
7222       "vmovl.s16 q2, d2\n"  in Transform()
7223       "vmovl.s16 q7, d11\n"  in Transform()
7224       "vmovl.s16 q6, d10\n"  in Transform()
7225       "vmovl.s16 q1, d1\n"  in Transform()
7226       "vmovl.s16 q0, d0\n"  in Transform()
7227       "vmovl.s16 q5, d9\n"  in Transform()
7228       "vmovl.s16 q4, d8\n"  in Transform()
7229       "vcvt.f32.s32 q0, q0\n"  in Transform()
7230       "vcvt.f32.s32 q1, q1\n"  in Transform()
7231       "vcvt.f32.s32 q2, q2\n"  in Transform()
7232       "vcvt.f32.s32 q3, q3\n"  in Transform()
7233       "vcvt.f32.s32 q4, q4\n"  in Transform()
7234       "vcvt.f32.s32 q5, q5\n"  in Transform()
7235       "vcvt.f32.s32 q6, q6\n"  in Transform()
7236       "vcvt.f32.s32 q7, q7\n"  in Transform()
7237       "vmul.f32 q0, q0, q9\n"  in Transform()
7238       "vmul.f32 q1, q1, q9\n"  in Transform()
7239       "vmul.f32 q2, q2, q9\n"  in Transform()
7240       "vmul.f32 q3, q3, q9\n"  in Transform()
7241       "vmul.f32 q4, q4, q11\n"  in Transform()
7242       "vmul.f32 q5, q5, q11\n"  in Transform()
7243       "vmul.f32 q6, q6, q11\n"  in Transform()
7244       "vmul.f32 q7, q7, q11\n"  in Transform()
7245       "vadd.f32 q0, q0, q8\n"  in Transform()
7246       "vadd.f32 q1, q1, q8\n"  in Transform()
7247       "vadd.f32 q2, q2, q8\n"  in Transform()
7248       "vadd.f32 q3, q3, q8\n"  in Transform()
7249       "vadd.f32 q4, q4, q10\n"  in Transform()
7250       "vadd.f32 q5, q5, q10\n"  in Transform()
7251       "vadd.f32 q6, q6, q10\n"  in Transform()
7252       "vadd.f32 q7, q7, q10\n"  in Transform()
7253       "vadd.f32 q0, q0, q4\n"  in Transform()
7254       "vadd.f32 q1, q1, q5\n"  in Transform()
7255       "vadd.f32 q2, q2, q6\n"  in Transform()
7256       "vadd.f32 q3, q3, q7\n"  in Transform()
7257       "vsub.f32 q0, q0, q12\n"  in Transform()
7258       "vsub.f32 q1, q1, q12\n"  in Transform()
7259       "vsub.f32 q2, q2, q12\n"  in Transform()
7260       "vsub.f32 q3, q3, q12\n"  in Transform()
7261       "vmul.f32 q0, q0, q13\n"  in Transform()
7262       "vmul.f32 q1, q1, q13\n"  in Transform()
7263       "vmul.f32 q2, q2, q13\n"  in Transform()
7264       "vmul.f32 q3, q3, q13\n"  in Transform()
7265       "vadd.f32 q0, q0, q14\n"  in Transform()
7266       "vadd.f32 q1, q1, q14\n"  in Transform()
7267       "vadd.f32 q2, q2, q14\n"  in Transform()
7268       "vadd.f32 q3, q3, q14\n"  in Transform()
7269       "vcvt.s32.f32 q0, q0\n"  in Transform()
7270       "vcvt.s32.f32 q1, q1\n"  in Transform()
7271       "vcvt.s32.f32 q2, q2\n"  in Transform()
7272       "vcvt.s32.f32 q3, q3\n"  in Transform()
7274       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7275       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
7276       "pld [%[output]]\n"  in Transform()
7277       "bne 2b\n"  in Transform()
7281       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
7282       "vld1.16 {d1[0]}, [%[input]]!\n"  in Transform()
7283       "vld1.8 {d1[2]}, [%[input]]!\n"  in Transform()
7284       "vld1.32 {d6}, [r1]!\n"  in Transform()
7285       "vld1.16 {d7[0]}, [r1]!\n"  in Transform()
7286       "vld1.8 {d7[2]}, [r1]!\n"  in Transform()
7287       "pld [%[input], #32]\n"  in Transform()
7288       "vmovl.u8 q1, d1\n"  in Transform()
7289       "vmovl.u8 q0, d0\n"  in Transform()
7290       "vmovl.u8 q4, d7\n"  in Transform()
7291       "vmovl.u8 q3, d6\n"  in Transform()
7292       "vmovl.s16 q2, d2\n"  in Transform()
7293       "vmovl.s16 q5, d8\n"  in Transform()
7294       "vmovl.s16 q1, d1\n"  in Transform()
7295       "vmovl.s16 q0, d0\n"  in Transform()
7296       "vmovl.s16 q4, d7\n"  in Transform()
7297       "vmovl.s16 q3, d6\n"  in Transform()
7298       "vcvt.f32.s32 q0, q0\n"  in Transform()
7299       "vcvt.f32.s32 q1, q1\n"  in Transform()
7300       "vcvt.f32.s32 q2, q2\n"  in Transform()
7301       "vcvt.f32.s32 q3, q3\n"  in Transform()
7302       "vcvt.f32.s32 q4, q4\n"  in Transform()
7303       "vcvt.f32.s32 q5, q5\n"  in Transform()
7304       "vmul.f32 q0, q0, q9\n"  in Transform()
7305       "vmul.f32 q1, q1, q9\n"  in Transform()
7306       "vmul.f32 q2, q2, q9\n"  in Transform()
7307       "vmul.f32 q3, q3, q11\n"  in Transform()
7308       "vmul.f32 q4, q4, q11\n"  in Transform()
7309       "vmul.f32 q5, q5, q11\n"  in Transform()
7310       "vadd.f32 q0, q0, q8\n"  in Transform()
7311       "vadd.f32 q1, q1, q8\n"  in Transform()
7312       "vadd.f32 q2, q2, q8\n"  in Transform()
7313       "vadd.f32 q3, q3, q10\n"  in Transform()
7314       "vadd.f32 q4, q4, q10\n"  in Transform()
7315       "vadd.f32 q5, q5, q10\n"  in Transform()
7316       "vadd.f32 q0, q0, q3\n"  in Transform()
7317       "vadd.f32 q1, q1, q4\n"  in Transform()
7318       "vadd.f32 q2, q2, q5\n"  in Transform()
7319       "vsub.f32 q0, q0, q12\n"  in Transform()
7320       "vsub.f32 q1, q1, q12\n"  in Transform()
7321       "vsub.f32 q2, q2, q12\n"  in Transform()
7322       "vmul.f32 q0, q0, q13\n"  in Transform()
7323       "vmul.f32 q1, q1, q13\n"  in Transform()
7324       "vmul.f32 q2, q2, q13\n"  in Transform()
7325       "vadd.f32 q0, q0, q14\n"  in Transform()
7326       "vadd.f32 q1, q1, q14\n"  in Transform()
7327       "vadd.f32 q2, q2, q14\n"  in Transform()
7328       "vcvt.s32.f32 q0, q0\n"  in Transform()
7329       "vcvt.s32.f32 q1, q1\n"  in Transform()
7330       "vcvt.s32.f32 q2, q2\n"  in Transform()
7332       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7333       "vst1.32 {d4}, [%[output]]!\n"  in Transform()
7334       "vst1.32 {d5[0]}, [%[output]]!\n"  in Transform()
7335       "pld [%[output]]\n"  in Transform()
7336       "subs %[rows], %[rows], #1\n"  in Transform()
7337       "bne 1b\n"  in Transform()
7369       "ldr r0, %[input_range_min]\n"  in Transform()
7370       "vdup.32 q8, r0\n"  in Transform()
7371       "ldr r0, %[input_range_scale]\n"  in Transform()
7372       "vdup.32 q9, r0\n"  in Transform()
7373       "ldr r0, %[bias_range_min]\n"  in Transform()
7374       "vdup.32 q10, r0\n"  in Transform()
7375       "ldr r0, %[bias_range_scale]\n"  in Transform()
7376       "vdup.32 q11, r0\n"  in Transform()
7377       "ldr r0, %[output_range_min]\n"  in Transform()
7378       "vdup.32 q12, r0\n"  in Transform()
7379       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
7380       "vdup.32 q13, r0\n"  in Transform()
7381       "ldr r0, %[output_range_offset]\n"  in Transform()
7382       "vdup.32 q14, r0\n"  in Transform()
7384       "mov r0, %[count]\n"  in Transform()
7385       "mov r1, %[bias]\n"  in Transform()
7386       "subs r0, r0, #12\n"  in Transform()
7387       "beq 3f\n"  in Transform()
7389       "subs r0, r0, #16\n"  in Transform()
7392       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
7393       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
7394       "pld [%[input], #32]\n"  in Transform()
7395       "vmovl.u8 q1, d1\n"  in Transform()
7396       "vmovl.u8 q0, d0\n"  in Transform()
7397       "vmovl.u8 q5, d9\n"  in Transform()
7398       "vmovl.u8 q4, d8\n"  in Transform()
7399       "vmovl.s16 q3, d3\n"  in Transform()
7400       "vmovl.s16 q2, d2\n"  in Transform()
7401       "vmovl.s16 q7, d11\n"  in Transform()
7402       "vmovl.s16 q6, d10\n"  in Transform()
7403       "vmovl.s16 q1, d1\n"  in Transform()
7404       "vmovl.s16 q0, d0\n"  in Transform()
7405       "vmovl.s16 q5, d9\n"  in Transform()
7406       "vmovl.s16 q4, d8\n"  in Transform()
7407       "vcvt.f32.s32 q0, q0\n"  in Transform()
7408       "vcvt.f32.s32 q1, q1\n"  in Transform()
7409       "vcvt.f32.s32 q2, q2\n"  in Transform()
7410       "vcvt.f32.s32 q3, q3\n"  in Transform()
7411       "vcvt.f32.s32 q4, q4\n"  in Transform()
7412       "vcvt.f32.s32 q5, q5\n"  in Transform()
7413       "vcvt.f32.s32 q6, q6\n"  in Transform()
7414       "vcvt.f32.s32 q7, q7\n"  in Transform()
7415       "vmul.f32 q0, q0, q9\n"  in Transform()
7416       "vmul.f32 q1, q1, q9\n"  in Transform()
7417       "vmul.f32 q2, q2, q9\n"  in Transform()
7418       "vmul.f32 q3, q3, q9\n"  in Transform()
7419       "vmul.f32 q4, q4, q11\n"  in Transform()
7420       "vmul.f32 q5, q5, q11\n"  in Transform()
7421       "vmul.f32 q6, q6, q11\n"  in Transform()
7422       "vmul.f32 q7, q7, q11\n"  in Transform()
7423       "vadd.f32 q0, q0, q8\n"  in Transform()
7424       "vadd.f32 q1, q1, q8\n"  in Transform()
7425       "vadd.f32 q2, q2, q8\n"  in Transform()
7426       "vadd.f32 q3, q3, q8\n"  in Transform()
7427       "vadd.f32 q4, q4, q10\n"  in Transform()
7428       "vadd.f32 q5, q5, q10\n"  in Transform()
7429       "vadd.f32 q6, q6, q10\n"  in Transform()
7430       "vadd.f32 q7, q7, q10\n"  in Transform()
7431       "vadd.f32 q0, q0, q4\n"  in Transform()
7432       "vadd.f32 q1, q1, q5\n"  in Transform()
7433       "vadd.f32 q2, q2, q6\n"  in Transform()
7434       "vadd.f32 q3, q3, q7\n"  in Transform()
7435       "vsub.f32 q0, q0, q12\n"  in Transform()
7436       "vsub.f32 q1, q1, q12\n"  in Transform()
7437       "vsub.f32 q2, q2, q12\n"  in Transform()
7438       "vsub.f32 q3, q3, q12\n"  in Transform()
7439       "vmul.f32 q0, q0, q13\n"  in Transform()
7440       "vmul.f32 q1, q1, q13\n"  in Transform()
7441       "vmul.f32 q2, q2, q13\n"  in Transform()
7442       "vmul.f32 q3, q3, q13\n"  in Transform()
7443       "vadd.f32 q0, q0, q14\n"  in Transform()
7444       "vadd.f32 q1, q1, q14\n"  in Transform()
7445       "vadd.f32 q2, q2, q14\n"  in Transform()
7446       "vadd.f32 q3, q3, q14\n"  in Transform()
7447       "vcvt.s32.f32 q0, q0\n"  in Transform()
7448       "vcvt.s32.f32 q1, q1\n"  in Transform()
7449       "vcvt.s32.f32 q2, q2\n"  in Transform()
7450       "vcvt.s32.f32 q3, q3\n"  in Transform()
7452       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7453       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
7454       "pld [%[output]]\n"  in Transform()
7455       "bne 2b\n"  in Transform()
7459       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
7460       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
7461       "vld1.32 {d6}, [r1]!\n"  in Transform()
7462       "vld1.32 {d7[0]}, [r1]!\n"  in Transform()
7463       "pld [%[input], #32]\n"  in Transform()
7464       "vmovl.u8 q1, d1\n"  in Transform()
7465       "vmovl.u8 q0, d0\n"  in Transform()
7466       "vmovl.u8 q4, d7\n"  in Transform()
7467       "vmovl.u8 q3, d6\n"  in Transform()
7468       "vmovl.s16 q2, d2\n"  in Transform()
7469       "vmovl.s16 q5, d8\n"  in Transform()
7470       "vmovl.s16 q1, d1\n"  in Transform()
7471       "vmovl.s16 q0, d0\n"  in Transform()
7472       "vmovl.s16 q4, d7\n"  in Transform()
7473       "vmovl.s16 q3, d6\n"  in Transform()
7474       "vcvt.f32.s32 q0, q0\n"  in Transform()
7475       "vcvt.f32.s32 q1, q1\n"  in Transform()
7476       "vcvt.f32.s32 q2, q2\n"  in Transform()
7477       "vcvt.f32.s32 q3, q3\n"  in Transform()
7478       "vcvt.f32.s32 q4, q4\n"  in Transform()
7479       "vcvt.f32.s32 q5, q5\n"  in Transform()
7480       "vmul.f32 q0, q0, q9\n"  in Transform()
7481       "vmul.f32 q1, q1, q9\n"  in Transform()
7482       "vmul.f32 q2, q2, q9\n"  in Transform()
7483       "vmul.f32 q3, q3, q11\n"  in Transform()
7484       "vmul.f32 q4, q4, q11\n"  in Transform()
7485       "vmul.f32 q5, q5, q11\n"  in Transform()
7486       "vadd.f32 q0, q0, q8\n"  in Transform()
7487       "vadd.f32 q1, q1, q8\n"  in Transform()
7488       "vadd.f32 q2, q2, q8\n"  in Transform()
7489       "vadd.f32 q3, q3, q10\n"  in Transform()
7490       "vadd.f32 q4, q4, q10\n"  in Transform()
7491       "vadd.f32 q5, q5, q10\n"  in Transform()
7492       "vadd.f32 q0, q0, q3\n"  in Transform()
7493       "vadd.f32 q1, q1, q4\n"  in Transform()
7494       "vadd.f32 q2, q2, q5\n"  in Transform()
7495       "vsub.f32 q0, q0, q12\n"  in Transform()
7496       "vsub.f32 q1, q1, q12\n"  in Transform()
7497       "vsub.f32 q2, q2, q12\n"  in Transform()
7498       "vmul.f32 q0, q0, q13\n"  in Transform()
7499       "vmul.f32 q1, q1, q13\n"  in Transform()
7500       "vmul.f32 q2, q2, q13\n"  in Transform()
7501       "vadd.f32 q0, q0, q14\n"  in Transform()
7502       "vadd.f32 q1, q1, q14\n"  in Transform()
7503       "vadd.f32 q2, q2, q14\n"  in Transform()
7504       "vcvt.s32.f32 q0, q0\n"  in Transform()
7505       "vcvt.s32.f32 q1, q1\n"  in Transform()
7506       "vcvt.s32.f32 q2, q2\n"  in Transform()
7508       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7509       "vst1.32 {d4, d5}, [%[output]]!\n"  in Transform()
7510       "pld [%[output]]\n"  in Transform()
7511       "subs %[rows], %[rows], #1\n"  in Transform()
7512       "bne 1b\n"  in Transform()
7544       "ldr r0, %[input_range_min]\n"  in Transform()
7545       "vdup.32 q8, r0\n"  in Transform()
7546       "ldr r0, %[input_range_scale]\n"  in Transform()
7547       "vdup.32 q9, r0\n"  in Transform()
7548       "ldr r0, %[bias_range_min]\n"  in Transform()
7549       "vdup.32 q10, r0\n"  in Transform()
7550       "ldr r0, %[bias_range_scale]\n"  in Transform()
7551       "vdup.32 q11, r0\n"  in Transform()
7552       "ldr r0, %[output_range_min]\n"  in Transform()
7553       "vdup.32 q12, r0\n"  in Transform()
7554       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
7555       "vdup.32 q13, r0\n"  in Transform()
7556       "ldr r0, %[output_range_offset]\n"  in Transform()
7557       "vdup.32 q14, r0\n"  in Transform()
7559       "mov r0, %[count]\n"  in Transform()
7560       "mov r1, %[bias]\n"  in Transform()
7561       "subs r0, r0, #13\n"  in Transform()
7562       "beq 3f\n"  in Transform()
7564       "subs r0, r0, #16\n"  in Transform()
7567       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
7568       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
7569       "pld [%[input], #32]\n"  in Transform()
7570       "vmovl.u8 q1, d1\n"  in Transform()
7571       "vmovl.u8 q0, d0\n"  in Transform()
7572       "vmovl.u8 q5, d9\n"  in Transform()
7573       "vmovl.u8 q4, d8\n"  in Transform()
7574       "vmovl.s16 q3, d3\n"  in Transform()
7575       "vmovl.s16 q2, d2\n"  in Transform()
7576       "vmovl.s16 q7, d11\n"  in Transform()
7577       "vmovl.s16 q6, d10\n"  in Transform()
7578       "vmovl.s16 q1, d1\n"  in Transform()
7579       "vmovl.s16 q0, d0\n"  in Transform()
7580       "vmovl.s16 q5, d9\n"  in Transform()
7581       "vmovl.s16 q4, d8\n"  in Transform()
7582       "vcvt.f32.s32 q0, q0\n"  in Transform()
7583       "vcvt.f32.s32 q1, q1\n"  in Transform()
7584       "vcvt.f32.s32 q2, q2\n"  in Transform()
7585       "vcvt.f32.s32 q3, q3\n"  in Transform()
7586       "vcvt.f32.s32 q4, q4\n"  in Transform()
7587       "vcvt.f32.s32 q5, q5\n"  in Transform()
7588       "vcvt.f32.s32 q6, q6\n"  in Transform()
7589       "vcvt.f32.s32 q7, q7\n"  in Transform()
7590       "vmul.f32 q0, q0, q9\n"  in Transform()
7591       "vmul.f32 q1, q1, q9\n"  in Transform()
7592       "vmul.f32 q2, q2, q9\n"  in Transform()
7593       "vmul.f32 q3, q3, q9\n"  in Transform()
7594       "vmul.f32 q4, q4, q11\n"  in Transform()
7595       "vmul.f32 q5, q5, q11\n"  in Transform()
7596       "vmul.f32 q6, q6, q11\n"  in Transform()
7597       "vmul.f32 q7, q7, q11\n"  in Transform()
7598       "vadd.f32 q0, q0, q8\n"  in Transform()
7599       "vadd.f32 q1, q1, q8\n"  in Transform()
7600       "vadd.f32 q2, q2, q8\n"  in Transform()
7601       "vadd.f32 q3, q3, q8\n"  in Transform()
7602       "vadd.f32 q4, q4, q10\n"  in Transform()
7603       "vadd.f32 q5, q5, q10\n"  in Transform()
7604       "vadd.f32 q6, q6, q10\n"  in Transform()
7605       "vadd.f32 q7, q7, q10\n"  in Transform()
7606       "vadd.f32 q0, q0, q4\n"  in Transform()
7607       "vadd.f32 q1, q1, q5\n"  in Transform()
7608       "vadd.f32 q2, q2, q6\n"  in Transform()
7609       "vadd.f32 q3, q3, q7\n"  in Transform()
7610       "vsub.f32 q0, q0, q12\n"  in Transform()
7611       "vsub.f32 q1, q1, q12\n"  in Transform()
7612       "vsub.f32 q2, q2, q12\n"  in Transform()
7613       "vsub.f32 q3, q3, q12\n"  in Transform()
7614       "vmul.f32 q0, q0, q13\n"  in Transform()
7615       "vmul.f32 q1, q1, q13\n"  in Transform()
7616       "vmul.f32 q2, q2, q13\n"  in Transform()
7617       "vmul.f32 q3, q3, q13\n"  in Transform()
7618       "vadd.f32 q0, q0, q14\n"  in Transform()
7619       "vadd.f32 q1, q1, q14\n"  in Transform()
7620       "vadd.f32 q2, q2, q14\n"  in Transform()
7621       "vadd.f32 q3, q3, q14\n"  in Transform()
7622       "vcvt.s32.f32 q0, q0\n"  in Transform()
7623       "vcvt.s32.f32 q1, q1\n"  in Transform()
7624       "vcvt.s32.f32 q2, q2\n"  in Transform()
7625       "vcvt.s32.f32 q3, q3\n"  in Transform()
7627       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7628       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
7629       "pld [%[output]]\n"  in Transform()
7630       "bne 2b\n"  in Transform()
7634       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
7635       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
7636       "vld1.8 {d1[4]}, [%[input]]!\n"  in Transform()
7637       "vld1.32 {d8}, [r1]!\n"  in Transform()
7638       "vld1.32 {d9[0]}, [r1]!\n"  in Transform()
7639       "vld1.8 {d9[4]}, [r1]!\n"  in Transform()
7640       "pld [%[input], #32]\n"  in Transform()
7641       "vmovl.u8 q1, d1\n"  in Transform()
7642       "vmovl.u8 q0, d0\n"  in Transform()
7643       "vmovl.u8 q5, d9\n"  in Transform()
7644       "vmovl.u8 q4, d8\n"  in Transform()
7645       "vmovl.s16 q3, d3\n"  in Transform()
7646       "vmovl.s16 q2, d2\n"  in Transform()
7647       "vmovl.s16 q7, d11\n"  in Transform()
7648       "vmovl.s16 q6, d10\n"  in Transform()
7649       "vmovl.s16 q1, d1\n"  in Transform()
7650       "vmovl.s16 q0, d0\n"  in Transform()
7651       "vmovl.s16 q5, d9\n"  in Transform()
7652       "vmovl.s16 q4, d8\n"  in Transform()
7653       "vcvt.f32.s32 q0, q0\n"  in Transform()
7654       "vcvt.f32.s32 q1, q1\n"  in Transform()
7655       "vcvt.f32.s32 q2, q2\n"  in Transform()
7656       "vcvt.f32.s32 q3, q3\n"  in Transform()
7657       "vcvt.f32.s32 q4, q4\n"  in Transform()
7658       "vcvt.f32.s32 q5, q5\n"  in Transform()
7659       "vcvt.f32.s32 q6, q6\n"  in Transform()
7660       "vcvt.f32.s32 q7, q7\n"  in Transform()
7661       "vmul.f32 q0, q0, q9\n"  in Transform()
7662       "vmul.f32 q1, q1, q9\n"  in Transform()
7663       "vmul.f32 q2, q2, q9\n"  in Transform()
7664       "vmul.f32 q3, q3, q9\n"  in Transform()
7665       "vmul.f32 q4, q4, q11\n"  in Transform()
7666       "vmul.f32 q5, q5, q11\n"  in Transform()
7667       "vmul.f32 q6, q6, q11\n"  in Transform()
7668       "vmul.f32 q7, q7, q11\n"  in Transform()
7669       "vadd.f32 q0, q0, q8\n"  in Transform()
7670       "vadd.f32 q1, q1, q8\n"  in Transform()
7671       "vadd.f32 q2, q2, q8\n"  in Transform()
7672       "vadd.f32 q3, q3, q8\n"  in Transform()
7673       "vadd.f32 q4, q4, q10\n"  in Transform()
7674       "vadd.f32 q5, q5, q10\n"  in Transform()
7675       "vadd.f32 q6, q6, q10\n"  in Transform()
7676       "vadd.f32 q7, q7, q10\n"  in Transform()
7677       "vadd.f32 q0, q0, q4\n"  in Transform()
7678       "vadd.f32 q1, q1, q5\n"  in Transform()
7679       "vadd.f32 q2, q2, q6\n"  in Transform()
7680       "vadd.f32 q3, q3, q7\n"  in Transform()
7681       "vsub.f32 q0, q0, q12\n"  in Transform()
7682       "vsub.f32 q1, q1, q12\n"  in Transform()
7683       "vsub.f32 q2, q2, q12\n"  in Transform()
7684       "vsub.f32 q3, q3, q12\n"  in Transform()
7685       "vmul.f32 q0, q0, q13\n"  in Transform()
7686       "vmul.f32 q1, q1, q13\n"  in Transform()
7687       "vmul.f32 q2, q2, q13\n"  in Transform()
7688       "vmul.f32 q3, q3, q13\n"  in Transform()
7689       "vadd.f32 q0, q0, q14\n"  in Transform()
7690       "vadd.f32 q1, q1, q14\n"  in Transform()
7691       "vadd.f32 q2, q2, q14\n"  in Transform()
7692       "vadd.f32 q3, q3, q14\n"  in Transform()
7693       "vcvt.s32.f32 q0, q0\n"  in Transform()
7694       "vcvt.s32.f32 q1, q1\n"  in Transform()
7695       "vcvt.s32.f32 q2, q2\n"  in Transform()
7696       "vcvt.s32.f32 q3, q3\n"  in Transform()
7698       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7699       "vst1.32 {d4, d5}, [%[output]]!\n"  in Transform()
7700       "vst1.32 {d6[0]}, [%[output]]!\n"  in Transform()
7701       "pld [%[output]]\n"  in Transform()
7702       "subs %[rows], %[rows], #1\n"  in Transform()
7703       "bne 1b\n"  in Transform()
7735       "ldr r0, %[input_range_min]\n"  in Transform()
7736       "vdup.32 q8, r0\n"  in Transform()
7737       "ldr r0, %[input_range_scale]\n"  in Transform()
7738       "vdup.32 q9, r0\n"  in Transform()
7739       "ldr r0, %[bias_range_min]\n"  in Transform()
7740       "vdup.32 q10, r0\n"  in Transform()
7741       "ldr r0, %[bias_range_scale]\n"  in Transform()
7742       "vdup.32 q11, r0\n"  in Transform()
7743       "ldr r0, %[output_range_min]\n"  in Transform()
7744       "vdup.32 q12, r0\n"  in Transform()
7745       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
7746       "vdup.32 q13, r0\n"  in Transform()
7747       "ldr r0, %[output_range_offset]\n"  in Transform()
7748       "vdup.32 q14, r0\n"  in Transform()
7750       "mov r0, %[count]\n"  in Transform()
7751       "mov r1, %[bias]\n"  in Transform()
7752       "subs r0, r0, #14\n"  in Transform()
7753       "beq 3f\n"  in Transform()
7755       "subs r0, r0, #16\n"  in Transform()
7758       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
7759       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
7760       "pld [%[input], #32]\n"  in Transform()
7761       "vmovl.u8 q1, d1\n"  in Transform()
7762       "vmovl.u8 q0, d0\n"  in Transform()
7763       "vmovl.u8 q5, d9\n"  in Transform()
7764       "vmovl.u8 q4, d8\n"  in Transform()
7765       "vmovl.s16 q3, d3\n"  in Transform()
7766       "vmovl.s16 q2, d2\n"  in Transform()
7767       "vmovl.s16 q7, d11\n"  in Transform()
7768       "vmovl.s16 q6, d10\n"  in Transform()
7769       "vmovl.s16 q1, d1\n"  in Transform()
7770       "vmovl.s16 q0, d0\n"  in Transform()
7771       "vmovl.s16 q5, d9\n"  in Transform()
7772       "vmovl.s16 q4, d8\n"  in Transform()
7773       "vcvt.f32.s32 q0, q0\n"  in Transform()
7774       "vcvt.f32.s32 q1, q1\n"  in Transform()
7775       "vcvt.f32.s32 q2, q2\n"  in Transform()
7776       "vcvt.f32.s32 q3, q3\n"  in Transform()
7777       "vcvt.f32.s32 q4, q4\n"  in Transform()
7778       "vcvt.f32.s32 q5, q5\n"  in Transform()
7779       "vcvt.f32.s32 q6, q6\n"  in Transform()
7780       "vcvt.f32.s32 q7, q7\n"  in Transform()
7781       "vmul.f32 q0, q0, q9\n"  in Transform()
7782       "vmul.f32 q1, q1, q9\n"  in Transform()
7783       "vmul.f32 q2, q2, q9\n"  in Transform()
7784       "vmul.f32 q3, q3, q9\n"  in Transform()
7785       "vmul.f32 q4, q4, q11\n"  in Transform()
7786       "vmul.f32 q5, q5, q11\n"  in Transform()
7787       "vmul.f32 q6, q6, q11\n"  in Transform()
7788       "vmul.f32 q7, q7, q11\n"  in Transform()
7789       "vadd.f32 q0, q0, q8\n"  in Transform()
7790       "vadd.f32 q1, q1, q8\n"  in Transform()
7791       "vadd.f32 q2, q2, q8\n"  in Transform()
7792       "vadd.f32 q3, q3, q8\n"  in Transform()
7793       "vadd.f32 q4, q4, q10\n"  in Transform()
7794       "vadd.f32 q5, q5, q10\n"  in Transform()
7795       "vadd.f32 q6, q6, q10\n"  in Transform()
7796       "vadd.f32 q7, q7, q10\n"  in Transform()
7797       "vadd.f32 q0, q0, q4\n"  in Transform()
7798       "vadd.f32 q1, q1, q5\n"  in Transform()
7799       "vadd.f32 q2, q2, q6\n"  in Transform()
7800       "vadd.f32 q3, q3, q7\n"  in Transform()
7801       "vsub.f32 q0, q0, q12\n"  in Transform()
7802       "vsub.f32 q1, q1, q12\n"  in Transform()
7803       "vsub.f32 q2, q2, q12\n"  in Transform()
7804       "vsub.f32 q3, q3, q12\n"  in Transform()
7805       "vmul.f32 q0, q0, q13\n"  in Transform()
7806       "vmul.f32 q1, q1, q13\n"  in Transform()
7807       "vmul.f32 q2, q2, q13\n"  in Transform()
7808       "vmul.f32 q3, q3, q13\n"  in Transform()
7809       "vadd.f32 q0, q0, q14\n"  in Transform()
7810       "vadd.f32 q1, q1, q14\n"  in Transform()
7811       "vadd.f32 q2, q2, q14\n"  in Transform()
7812       "vadd.f32 q3, q3, q14\n"  in Transform()
7813       "vcvt.s32.f32 q0, q0\n"  in Transform()
7814       "vcvt.s32.f32 q1, q1\n"  in Transform()
7815       "vcvt.s32.f32 q2, q2\n"  in Transform()
7816       "vcvt.s32.f32 q3, q3\n"  in Transform()
7818       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7819       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
7820       "pld [%[output]]\n"  in Transform()
7821       "bne 2b\n"  in Transform()
7825       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
7826       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
7827       "vld1.16 {d1[2]}, [%[input]]!\n"  in Transform()
7828       "vld1.32 {d8}, [r1]!\n"  in Transform()
7829       "vld1.32 {d9[0]}, [r1]!\n"  in Transform()
7830       "vld1.16 {d9[2]}, [r1]!\n"  in Transform()
7831       "pld [%[input], #32]\n"  in Transform()
7832       "vmovl.u8 q1, d1\n"  in Transform()
7833       "vmovl.u8 q0, d0\n"  in Transform()
7834       "vmovl.u8 q5, d9\n"  in Transform()
7835       "vmovl.u8 q4, d8\n"  in Transform()
7836       "vmovl.s16 q3, d3\n"  in Transform()
7837       "vmovl.s16 q2, d2\n"  in Transform()
7838       "vmovl.s16 q7, d11\n"  in Transform()
7839       "vmovl.s16 q6, d10\n"  in Transform()
7840       "vmovl.s16 q1, d1\n"  in Transform()
7841       "vmovl.s16 q0, d0\n"  in Transform()
7842       "vmovl.s16 q5, d9\n"  in Transform()
7843       "vmovl.s16 q4, d8\n"  in Transform()
7844       "vcvt.f32.s32 q0, q0\n"  in Transform()
7845       "vcvt.f32.s32 q1, q1\n"  in Transform()
7846       "vcvt.f32.s32 q2, q2\n"  in Transform()
7847       "vcvt.f32.s32 q3, q3\n"  in Transform()
7848       "vcvt.f32.s32 q4, q4\n"  in Transform()
7849       "vcvt.f32.s32 q5, q5\n"  in Transform()
7850       "vcvt.f32.s32 q6, q6\n"  in Transform()
7851       "vcvt.f32.s32 q7, q7\n"  in Transform()
7852       "vmul.f32 q0, q0, q9\n"  in Transform()
7853       "vmul.f32 q1, q1, q9\n"  in Transform()
7854       "vmul.f32 q2, q2, q9\n"  in Transform()
7855       "vmul.f32 q3, q3, q9\n"  in Transform()
7856       "vmul.f32 q4, q4, q11\n"  in Transform()
7857       "vmul.f32 q5, q5, q11\n"  in Transform()
7858       "vmul.f32 q6, q6, q11\n"  in Transform()
7859       "vmul.f32 q7, q7, q11\n"  in Transform()
7860       "vadd.f32 q0, q0, q8\n"  in Transform()
7861       "vadd.f32 q1, q1, q8\n"  in Transform()
7862       "vadd.f32 q2, q2, q8\n"  in Transform()
7863       "vadd.f32 q3, q3, q8\n"  in Transform()
7864       "vadd.f32 q4, q4, q10\n"  in Transform()
7865       "vadd.f32 q5, q5, q10\n"  in Transform()
7866       "vadd.f32 q6, q6, q10\n"  in Transform()
7867       "vadd.f32 q7, q7, q10\n"  in Transform()
7868       "vadd.f32 q0, q0, q4\n"  in Transform()
7869       "vadd.f32 q1, q1, q5\n"  in Transform()
7870       "vadd.f32 q2, q2, q6\n"  in Transform()
7871       "vadd.f32 q3, q3, q7\n"  in Transform()
7872       "vsub.f32 q0, q0, q12\n"  in Transform()
7873       "vsub.f32 q1, q1, q12\n"  in Transform()
7874       "vsub.f32 q2, q2, q12\n"  in Transform()
7875       "vsub.f32 q3, q3, q12\n"  in Transform()
7876       "vmul.f32 q0, q0, q13\n"  in Transform()
7877       "vmul.f32 q1, q1, q13\n"  in Transform()
7878       "vmul.f32 q2, q2, q13\n"  in Transform()
7879       "vmul.f32 q3, q3, q13\n"  in Transform()
7880       "vadd.f32 q0, q0, q14\n"  in Transform()
7881       "vadd.f32 q1, q1, q14\n"  in Transform()
7882       "vadd.f32 q2, q2, q14\n"  in Transform()
7883       "vadd.f32 q3, q3, q14\n"  in Transform()
7884       "vcvt.s32.f32 q0, q0\n"  in Transform()
7885       "vcvt.s32.f32 q1, q1\n"  in Transform()
7886       "vcvt.s32.f32 q2, q2\n"  in Transform()
7887       "vcvt.s32.f32 q3, q3\n"  in Transform()
7889       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
7890       "vst1.32 {d4, d5, d6}, [%[output]]!\n"  in Transform()
7891       "pld [%[output]]\n"  in Transform()
7892       "subs %[rows], %[rows], #1\n"  in Transform()
7893       "bne 1b\n"  in Transform()
7925       "ldr r0, %[input_range_min]\n"  in Transform()
7926       "vdup.32 q8, r0\n"  in Transform()
7927       "ldr r0, %[input_range_scale]\n"  in Transform()
7928       "vdup.32 q9, r0\n"  in Transform()
7929       "ldr r0, %[bias_range_min]\n"  in Transform()
7930       "vdup.32 q10, r0\n"  in Transform()
7931       "ldr r0, %[bias_range_scale]\n"  in Transform()
7932       "vdup.32 q11, r0\n"  in Transform()
7933       "ldr r0, %[output_range_min]\n"  in Transform()
7934       "vdup.32 q12, r0\n"  in Transform()
7935       "ldr r0, %[one_over_output_range_scale]\n"  in Transform()
7936       "vdup.32 q13, r0\n"  in Transform()
7937       "ldr r0, %[output_range_offset]\n"  in Transform()
7938       "vdup.32 q14, r0\n"  in Transform()
7940       "mov r0, %[count]\n"  in Transform()
7941       "mov r1, %[bias]\n"  in Transform()
7942       "subs r0, r0, #15\n"  in Transform()
7943       "beq 3f\n"  in Transform()
7945       "subs r0, r0, #16\n"  in Transform()
7948       "vld1.32 {d0, d1}, [%[input]]!\n"  in Transform()
7949       "vld1.32 {d8, d9}, [r1]!\n"  in Transform()
7950       "pld [%[input], #32]\n"  in Transform()
7951       "vmovl.u8 q1, d1\n"  in Transform()
7952       "vmovl.u8 q0, d0\n"  in Transform()
7953       "vmovl.u8 q5, d9\n"  in Transform()
7954       "vmovl.u8 q4, d8\n"  in Transform()
7955       "vmovl.s16 q3, d3\n"  in Transform()
7956       "vmovl.s16 q2, d2\n"  in Transform()
7957       "vmovl.s16 q7, d11\n"  in Transform()
7958       "vmovl.s16 q6, d10\n"  in Transform()
7959       "vmovl.s16 q1, d1\n"  in Transform()
7960       "vmovl.s16 q0, d0\n"  in Transform()
7961       "vmovl.s16 q5, d9\n"  in Transform()
7962       "vmovl.s16 q4, d8\n"  in Transform()
7963       "vcvt.f32.s32 q0, q0\n"  in Transform()
7964       "vcvt.f32.s32 q1, q1\n"  in Transform()
7965       "vcvt.f32.s32 q2, q2\n"  in Transform()
7966       "vcvt.f32.s32 q3, q3\n"  in Transform()
7967       "vcvt.f32.s32 q4, q4\n"  in Transform()
7968       "vcvt.f32.s32 q5, q5\n"  in Transform()
7969       "vcvt.f32.s32 q6, q6\n"  in Transform()
7970       "vcvt.f32.s32 q7, q7\n"  in Transform()
7971       "vmul.f32 q0, q0, q9\n"  in Transform()
7972       "vmul.f32 q1, q1, q9\n"  in Transform()
7973       "vmul.f32 q2, q2, q9\n"  in Transform()
7974       "vmul.f32 q3, q3, q9\n"  in Transform()
7975       "vmul.f32 q4, q4, q11\n"  in Transform()
7976       "vmul.f32 q5, q5, q11\n"  in Transform()
7977       "vmul.f32 q6, q6, q11\n"  in Transform()
7978       "vmul.f32 q7, q7, q11\n"  in Transform()
7979       "vadd.f32 q0, q0, q8\n"  in Transform()
7980       "vadd.f32 q1, q1, q8\n"  in Transform()
7981       "vadd.f32 q2, q2, q8\n"  in Transform()
7982       "vadd.f32 q3, q3, q8\n"  in Transform()
7983       "vadd.f32 q4, q4, q10\n"  in Transform()
7984       "vadd.f32 q5, q5, q10\n"  in Transform()
7985       "vadd.f32 q6, q6, q10\n"  in Transform()
7986       "vadd.f32 q7, q7, q10\n"  in Transform()
7987       "vadd.f32 q0, q0, q4\n"  in Transform()
7988       "vadd.f32 q1, q1, q5\n"  in Transform()
7989       "vadd.f32 q2, q2, q6\n"  in Transform()
7990       "vadd.f32 q3, q3, q7\n"  in Transform()
7991       "vsub.f32 q0, q0, q12\n"  in Transform()
7992       "vsub.f32 q1, q1, q12\n"  in Transform()
7993       "vsub.f32 q2, q2, q12\n"  in Transform()
7994       "vsub.f32 q3, q3, q12\n"  in Transform()
7995       "vmul.f32 q0, q0, q13\n"  in Transform()
7996       "vmul.f32 q1, q1, q13\n"  in Transform()
7997       "vmul.f32 q2, q2, q13\n"  in Transform()
7998       "vmul.f32 q3, q3, q13\n"  in Transform()
7999       "vadd.f32 q0, q0, q14\n"  in Transform()
8000       "vadd.f32 q1, q1, q14\n"  in Transform()
8001       "vadd.f32 q2, q2, q14\n"  in Transform()
8002       "vadd.f32 q3, q3, q14\n"  in Transform()
8003       "vcvt.s32.f32 q0, q0\n"  in Transform()
8004       "vcvt.s32.f32 q1, q1\n"  in Transform()
8005       "vcvt.s32.f32 q2, q2\n"  in Transform()
8006       "vcvt.s32.f32 q3, q3\n"  in Transform()
8008       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
8009       "vst1.32 {d4, d5, d6, d7}, [%[output]]!\n"  in Transform()
8010       "pld [%[output]]\n"  in Transform()
8011       "bne 2b\n"  in Transform()
8015       "vld1.32 {d0}, [%[input]]!\n"  in Transform()
8016       "vld1.32 {d1[0]}, [%[input]]!\n"  in Transform()
8017       "vld1.16 {d1[2]}, [%[input]]!\n"  in Transform()
8018       "vld1.8 {d1[6]}, [%[input]]!\n"  in Transform()
8019       "vld1.32 {d8}, [r1]!\n"  in Transform()
8020       "vld1.32 {d9[0]}, [r1]!\n"  in Transform()
8021       "vld1.16 {d9[2]}, [r1]!\n"  in Transform()
8022       "vld1.8 {d9[6]}, [r1]!\n"  in Transform()
8023       "pld [%[input], #32]\n"  in Transform()
8024       "vmovl.u8 q1, d1\n"  in Transform()
8025       "vmovl.u8 q0, d0\n"  in Transform()
8026       "vmovl.u8 q5, d9\n"  in Transform()
8027       "vmovl.u8 q4, d8\n"  in Transform()
8028       "vmovl.s16 q3, d3\n"  in Transform()
8029       "vmovl.s16 q2, d2\n"  in Transform()
8030       "vmovl.s16 q7, d11\n"  in Transform()
8031       "vmovl.s16 q6, d10\n"  in Transform()
8032       "vmovl.s16 q1, d1\n"  in Transform()
8033       "vmovl.s16 q0, d0\n"  in Transform()
8034       "vmovl.s16 q5, d9\n"  in Transform()
8035       "vmovl.s16 q4, d8\n"  in Transform()
8036       "vcvt.f32.s32 q0, q0\n"  in Transform()
8037       "vcvt.f32.s32 q1, q1\n"  in Transform()
8038       "vcvt.f32.s32 q2, q2\n"  in Transform()
8039       "vcvt.f32.s32 q3, q3\n"  in Transform()
8040       "vcvt.f32.s32 q4, q4\n"  in Transform()
8041       "vcvt.f32.s32 q5, q5\n"  in Transform()
8042       "vcvt.f32.s32 q6, q6\n"  in Transform()
8043       "vcvt.f32.s32 q7, q7\n"  in Transform()
8044       "vmul.f32 q0, q0, q9\n"  in Transform()
8045       "vmul.f32 q1, q1, q9\n"  in Transform()
8046       "vmul.f32 q2, q2, q9\n"  in Transform()
8047       "vmul.f32 q3, q3, q9\n"  in Transform()
8048       "vmul.f32 q4, q4, q11\n"  in Transform()
8049       "vmul.f32 q5, q5, q11\n"  in Transform()
8050       "vmul.f32 q6, q6, q11\n"  in Transform()
8051       "vmul.f32 q7, q7, q11\n"  in Transform()
8052       "vadd.f32 q0, q0, q8\n"  in Transform()
8053       "vadd.f32 q1, q1, q8\n"  in Transform()
8054       "vadd.f32 q2, q2, q8\n"  in Transform()
8055       "vadd.f32 q3, q3, q8\n"  in Transform()
8056       "vadd.f32 q4, q4, q10\n"  in Transform()
8057       "vadd.f32 q5, q5, q10\n"  in Transform()
8058       "vadd.f32 q6, q6, q10\n"  in Transform()
8059       "vadd.f32 q7, q7, q10\n"  in Transform()
8060       "vadd.f32 q0, q0, q4\n"  in Transform()
8061       "vadd.f32 q1, q1, q5\n"  in Transform()
8062       "vadd.f32 q2, q2, q6\n"  in Transform()
8063       "vadd.f32 q3, q3, q7\n"  in Transform()
8064       "vsub.f32 q0, q0, q12\n"  in Transform()
8065       "vsub.f32 q1, q1, q12\n"  in Transform()
8066       "vsub.f32 q2, q2, q12\n"  in Transform()
8067       "vsub.f32 q3, q3, q12\n"  in Transform()
8068       "vmul.f32 q0, q0, q13\n"  in Transform()
8069       "vmul.f32 q1, q1, q13\n"  in Transform()
8070       "vmul.f32 q2, q2, q13\n"  in Transform()
8071       "vmul.f32 q3, q3, q13\n"  in Transform()
8072       "vadd.f32 q0, q0, q14\n"  in Transform()
8073       "vadd.f32 q1, q1, q14\n"  in Transform()
8074       "vadd.f32 q2, q2, q14\n"  in Transform()
8075       "vadd.f32 q3, q3, q14\n"  in Transform()
8076       "vcvt.s32.f32 q0, q0\n"  in Transform()
8077       "vcvt.s32.f32 q1, q1\n"  in Transform()
8078       "vcvt.s32.f32 q2, q2\n"  in Transform()
8079       "vcvt.s32.f32 q3, q3\n"  in Transform()
8081       "vst1.32 {d0, d1, d2, d3}, [%[output]]!\n"  in Transform()
8082       "vst1.32 {d4, d5, d6}, [%[output]]!\n"  in Transform()
8083       "vst1.32 {d7[0]}, [%[output]]!\n"  in Transform()
8084       "pld [%[output]]\n"  in Transform()
8085       "subs %[rows], %[rows], #1\n"  in Transform()
8086       "bne 1b\n"  in Transform()