1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
|
/*
* Copyright © 2018, VideoLAN and dav1d authors
* Copyright © 2019, Martin Storsjo
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
* ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
* LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
* ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "src/arm/asm.S"
#include "util.S"
#define SUM_STRIDE (384+16)
// void dav1d_sgr_box3_v_neon(int32_t *sumsq, int16_t *sum,
// const int w, const int h,
// const enum LrEdgeFlags edges);
function sgr_box3_v_neon, export=1
push {r4-r9,lr}
ldr r4, [sp, #28]
add r12, r3, #2 // Number of output rows to move back
mov lr, r3 // Number of input rows to move back
add r2, r2, #2 // Actual summed width
mov r7, #(4*SUM_STRIDE) // sumsq stride
mov r8, #(2*SUM_STRIDE) // sum stride
sub r0, r0, #(4*SUM_STRIDE) // sumsq -= stride
sub r1, r1, #(2*SUM_STRIDE) // sum -= stride
tst r4, #4 // LR_HAVE_TOP
beq 0f
// If have top, read from row -2.
sub r5, r0, #(4*SUM_STRIDE)
sub r6, r1, #(2*SUM_STRIDE)
add lr, lr, #2
b 1f
0:
// !LR_HAVE_TOP
// If we don't have top, read from row 0 even if
// we start writing to row -1.
add r5, r0, #(4*SUM_STRIDE)
add r6, r1, #(2*SUM_STRIDE)
1:
tst r4, #8 // LR_HAVE_BOTTOM
beq 1f
// LR_HAVE_BOTTOM
add r3, r3, #2 // Sum all h+2 lines with the main loop
add lr, lr, #2
1:
mov r9, r3 // Backup of h for next loops
1:
// Start of horizontal loop; start one vertical filter slice.
// Start loading rows into q8-q13 and q0-q2 taking top
// padding into consideration.
tst r4, #4 // LR_HAVE_TOP
vld1.32 {q8, q9}, [r5, :128], r7
vld1.16 {q0}, [r6, :128], r8
beq 2f
// LR_HAVE_TOP
vld1.32 {q10, q11}, [r5, :128], r7
vld1.16 {q1}, [r6, :128], r8
vld1.32 {q12, q13}, [r5, :128], r7
vld1.16 {q2}, [r6, :128], r8
b 3f
2: // !LR_HAVE_TOP
vmov q10, q8
vmov q11, q9
vmov q1, q0
vmov q12, q8
vmov q13, q9
vmov q2, q0
3:
subs r3, r3, #1
.macro add3
vadd.i32 q8, q8, q10
vadd.i32 q9, q9, q11
vadd.i16 q0, q0, q1
vadd.i32 q8, q8, q12
vadd.i32 q9, q9, q13
vadd.i16 q0, q0, q2
vst1.32 {q8, q9}, [r0, :128], r7
vst1.16 {q0}, [r1, :128], r8
.endm
add3
vmov q8, q10
vmov q9, q11
vmov q0, q1
vmov q10, q12
vmov q11, q13
vmov q1, q2
ble 4f
vld1.32 {q12, q13}, [r5, :128], r7
vld1.16 {q2}, [r6, :128], r8
b 3b
4:
tst r4, #8 // LR_HAVE_BOTTOM
bne 5f
// !LR_HAVE_BOTTOM
// Produce two more rows, extending the already loaded rows.
add3
vmov q8, q10
vmov q9, q11
vmov q0, q1
add3
5: // End of one vertical slice.
subs r2, r2, #8
ble 0f
// Move pointers back up to the top and loop horizontally.
// Input pointers
mls r5, r7, lr, r5
mls r6, r8, lr, r6
// Output pointers
mls r0, r7, r12, r0
mls r1, r8, r12, r1
add r0, r0, #32
add r1, r1, #16
add r5, r5, #32
add r6, r6, #16
mov r3, r9
b 1b
0:
pop {r4-r9,pc}
.purgem add3
endfunc
// void dav1d_sgr_box5_v_neon(int32_t *sumsq, int16_t *sum,
// const int w, const int h,
// const enum LrEdgeFlags edges);
function sgr_box5_v_neon, export=1
push {r4-r9,lr}
vpush {q5-q7}
ldr r4, [sp, #76]
add r12, r3, #2 // Number of output rows to move back
mov lr, r3 // Number of input rows to move back
add r2, r2, #8 // Actual summed width
mov r7, #(4*SUM_STRIDE) // sumsq stride
mov r8, #(2*SUM_STRIDE) // sum stride
sub r0, r0, #(4*SUM_STRIDE) // sumsq -= stride
sub r1, r1, #(2*SUM_STRIDE) // sum -= stride
tst r4, #4 // LR_HAVE_TOP
beq 0f
// If have top, read from row -2.
sub r5, r0, #(4*SUM_STRIDE)
sub r6, r1, #(2*SUM_STRIDE)
add lr, lr, #2
b 1f
0:
// !LR_HAVE_TOP
// If we don't have top, read from row 0 even if
// we start writing to row -1.
add r5, r0, #(4*SUM_STRIDE)
add r6, r1, #(2*SUM_STRIDE)
1:
tst r4, #8 // LR_HAVE_BOTTOM
beq 0f
// LR_HAVE_BOTTOM
add r3, r3, #2 // Handle h+2 lines with the main loop
add lr, lr, #2
b 1f
0:
// !LR_HAVE_BOTTOM
sub r3, r3, #1 // Handle h-1 lines with the main loop
1:
mov r9, r3 // Backup of h for next loops
1:
// Start of horizontal loop; start one vertical filter slice.
// Start loading rows into q6-q15 and q0-q3,q5 taking top
// padding into consideration.
tst r4, #4 // LR_HAVE_TOP
vld1.32 {q6, q7}, [r5, :128], r7
vld1.16 {q0}, [r6, :128], r8
beq 2f
// LR_HAVE_TOP
vld1.32 {q10, q11}, [r5, :128], r7
vld1.16 {q2}, [r6, :128], r8
vmov q8, q6
vmov q9, q7
vmov q1, q0
vld1.32 {q12, q13}, [r5, :128], r7
vld1.16 {q3}, [r6, :128], r8
b 3f
2: // !LR_HAVE_TOP
vmov q8, q6
vmov q9, q7
vmov q1, q0
vmov q10, q6
vmov q11, q7
vmov q2, q0
vmov q12, q6
vmov q13, q7
vmov q3, q0
3:
cmp r3, #0
beq 4f
vld1.32 {q14, q15}, [r5, :128], r7
vld1.16 {q5}, [r6, :128], r8
3:
// Start of vertical loop
subs r3, r3, #2
.macro add5
vadd.i32 q6, q6, q8
vadd.i32 q7, q7, q9
vadd.i16 q0, q0, q1
vadd.i32 q6, q6, q10
vadd.i32 q7, q7, q11
vadd.i16 q0, q0, q2
vadd.i32 q6, q6, q12
vadd.i32 q7, q7, q13
vadd.i16 q0, q0, q3
vadd.i32 q6, q6, q14
vadd.i32 q7, q7, q15
vadd.i16 q0, q0, q5
vst1.32 {q6, q7}, [r0, :128], r7
vst1.16 {q0}, [r1, :128], r8
.endm
add5
.macro shift2
vmov q6, q10
vmov q7, q11
vmov q0, q2
vmov q8, q12
vmov q9, q13
vmov q1, q3
vmov q10, q14
vmov q11, q15
vmov q2, q5
.endm
shift2
add r0, r0, r7
add r1, r1, r8
ble 5f
vld1.32 {q12, q13}, [r5, :128], r7
vld1.16 {q3}, [r6, :128], r8
vld1.32 {q14, q15}, [r5, :128], r7
vld1.16 {q5}, [r6, :128], r8
b 3b
4:
// h == 1, !LR_HAVE_BOTTOM.
// Pad the last row with the only content row, and add.
vmov q14, q12
vmov q15, q13
vmov q5, q3
add5
shift2
add r0, r0, r7
add r1, r1, r8
add5
b 6f
5:
tst r4, #8 // LR_HAVE_BOTTOM
bne 6f
// !LR_HAVE_BOTTOM
cmp r3, #0
bne 5f
// The intended three edge rows left; output the one at h-2 and
// the past edge one at h.
vld1.32 {q12, q13}, [r5, :128], r7
vld1.16 {q3}, [r6, :128], r8
// Pad the past-edge row from the last content row.
vmov q14, q12
vmov q15, q13
vmov q5, q3
add5
shift2
add r0, r0, r7
add r1, r1, r8
// The last two rows are already padded properly here.
add5
b 6f
5:
// r3 == -1, two rows left, output one.
// Pad the last two rows from the mid one.
vmov q12, q10
vmov q13, q11
vmov q3, q2
vmov q14, q10
vmov q15, q11
vmov q5, q2
add5
add r0, r0, r7
add r1, r1, r8
b 6f
6: // End of one vertical slice.
subs r2, r2, #8
ble 0f
// Move pointers back up to the top and loop horizontally.
// Input pointers
mls r5, r7, lr, r5
mls r6, r8, lr, r6
// Output pointers
mls r0, r7, r12, r0
mls r1, r8, r12, r1
add r0, r0, #32
add r1, r1, #16
add r5, r5, #32
add r6, r6, #16
mov r3, r9
b 1b
0:
vpop {q5-q7}
pop {r4-r9,pc}
.purgem add5
endfunc
// void dav1d_sgr_calc_ab1_neon(int32_t *a, int16_t *b,
// const int w, const int h, const int strength,
// const int bitdepth_max);
// void dav1d_sgr_calc_ab2_neon(int32_t *a, int16_t *b,
// const int w, const int h, const int strength,
// const int bitdepth_max);
function sgr_calc_ab1_neon, export=1
push {r4-r7,lr}
vpush {q4-q7}
ldrd r4, r5, [sp, #84]
add r3, r3, #2 // h += 2
clz r6, r5
vmov.i32 q15, #9 // n
movw r5, #455
mov lr, #SUM_STRIDE
b sgr_calc_ab_neon
endfunc
function sgr_calc_ab2_neon, export=1
push {r4-r7,lr}
vpush {q4-q7}
ldrd r4, r5, [sp, #84]
add r3, r3, #3 // h += 3
clz r6, r5
asr r3, r3, #1 // h /= 2
vmov.i32 q15, #25 // n
mov r5, #164
mov lr, #(2*SUM_STRIDE)
endfunc
function sgr_calc_ab_neon
movrel r12, X(sgr_x_by_x)
sub r6, r6, #24 // -bitdepth_min_8
vld1.8 {q8, q9}, [r12, :128]!
add r7, r6, r6 // -2*bitdepth_min_8
vmov.i8 q11, #5
vmov.i8 d10, #55 // idx of last 5
vld1.8 {q10}, [r12, :128]
vmov.i8 d11, #72 // idx of last 4
vmov.i8 d12, #101 // idx of last 3
vmov.i8 d13, #169 // idx of last 2
vmov.i8 d14, #254 // idx of last 1
vmov.i8 d15, #32 // elements consumed in first vtbl
add r2, r2, #2 // w += 2
add r12, r2, #7
bic r12, r12, #7 // aligned w
sub r12, lr, r12 // increment between rows
vdup.32 q12, r4
sub r0, r0, #(4*(SUM_STRIDE))
sub r1, r1, #(2*(SUM_STRIDE))
mov r4, r2 // backup of w
vsub.i8 q8, q8, q11
vsub.i8 q9, q9, q11
vsub.i8 q10, q10, q11
1:
vld1.32 {q0, q1}, [r0, :128] // a
vld1.16 {q2}, [r1, :128] // b
vdup.32 q13, r7 // -2*bitdepth_min_8
vdup.16 q14, r6 // -bitdepth_min_8
subs r2, r2, #8
vrshl.s32 q0, q0, q13
vrshl.s32 q1, q1, q13
vrshl.s16 q4, q2, q14
vmul.i32 q0, q0, q15 // a * n
vmul.i32 q1, q1, q15 // a * n
vmull.u16 q3, d8, d8 // b * b
vmull.u16 q4, d9, d9 // b * b
vqsub.u32 q0, q0, q3 // imax(a * n - b * b, 0)
vqsub.u32 q1, q1, q4 // imax(a * n - b * b, 0)
vmul.i32 q0, q0, q12 // p * s
vmul.i32 q1, q1, q12 // p * s
vqshrn.u32 d0, q0, #16
vqshrn.u32 d1, q1, #16
vqrshrn.u16 d0, q0, #4 // imin(z, 255)
vcgt.u8 d2, d0, d10 // = -1 if sgr_x_by_x[d0] < 5
vcgt.u8 d3, d0, d11 // = -1 if sgr_x_by_x[d0] < 4
vtbl.8 d1, {q8, q9}, d0
vcgt.u8 d6, d0, d12 // = -1 if sgr_x_by_x[d0] < 3
vsub.i8 d9, d0, d15 // indices for vtbx
vcgt.u8 d7, d0, d13 // = -1 if sgr_x_by_x[d0] < 2
vadd.i8 d2, d2, d3
vtbx.8 d1, {q10}, d9
vcgt.u8 d8, d0, d14 // = -1 if sgr_x_by_x[d0] < 1
vadd.i8 d6, d6, d7
vadd.i8 d8, d8, d22
vadd.i8 d2, d2, d6
vadd.i8 d1, d1, d8
vadd.i8 d1, d1, d2
vmovl.u8 q0, d1 // x
vmov.i16 q13, #256
vdup.32 q14, r5 // one_by_x
vmull.u16 q1, d0, d4 // x * BB[i]
vmull.u16 q2, d1, d5 // x * BB[i]
vmul.i32 q1, q1, q14 // x * BB[i] * sgr_one_by_x
vmul.i32 q2, q2, q14 // x * BB[i] * sgr_one_by_x
vrshr.s32 q1, q1, #12 // AA[i]
vrshr.s32 q2, q2, #12 // AA[i]
vsub.i16 q0, q13, q0 // 256 - x
vst1.32 {q1, q2}, [r0, :128]!
vst1.16 {q0}, [r1, :128]!
bgt 1b
subs r3, r3, #1
ble 0f
add r0, r0, r12, lsl #2
add r1, r1, r12, lsl #1
mov r2, r4
b 1b
0:
vpop {q4-q7}
pop {r4-r7,pc}
endfunc
|
