2 * Copyright (c) 2016 Google Inc.
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 #include "libavutil/aarch64/asm.S"
23 // All public functions in this file have the following signature:
24 // typedef void (*vp9_mc_func)(uint8_t *dst, ptrdiff_t dst_stride,
25 // const uint8_t *ref, ptrdiff_t ref_stride,
26 // int h, int mx, int my);
28 function ff_vp9_avg64_neon, export=1
31 ld1 {v4.16b, v5.16b, v6.16b, v7.16b}, [x2], x3
32 ld1 {v0.16b, v1.16b, v2.16b, v3.16b}, [x0], x1
33 ld1 {v20.16b, v21.16b, v22.16b, v23.16b}, [x2], x3
34 urhadd v0.16b, v0.16b, v4.16b
35 urhadd v1.16b, v1.16b, v5.16b
36 ld1 {v16.16b, v17.16b, v18.16b, v19.16b}, [x0], x1
37 urhadd v2.16b, v2.16b, v6.16b
38 urhadd v3.16b, v3.16b, v7.16b
40 urhadd v16.16b, v16.16b, v20.16b
41 urhadd v17.16b, v17.16b, v21.16b
42 st1 {v0.16b, v1.16b, v2.16b, v3.16b}, [x5], x1
43 urhadd v18.16b, v18.16b, v22.16b
44 urhadd v19.16b, v19.16b, v23.16b
45 st1 {v16.16b, v17.16b, v18.16b, v19.16b}, [x5], x1
50 function ff_vp9_avg32_neon, export=1
52 ld1 {v2.16b, v3.16b}, [x2], x3
53 ld1 {v0.16b, v1.16b}, [x0]
54 urhadd v0.16b, v0.16b, v2.16b
55 urhadd v1.16b, v1.16b, v3.16b
57 st1 {v0.16b, v1.16b}, [x0], x1
62 function ff_vp9_copy16_neon, export=1
68 ld1 {v0.16b}, [x2], x3
69 ld1 {v1.16b}, [x6], x3
70 ld1 {v2.16b}, [x2], x3
71 ld1 {v3.16b}, [x6], x3
73 st1 {v0.16b}, [x0], x1
74 st1 {v1.16b}, [x5], x1
75 st1 {v2.16b}, [x0], x1
76 st1 {v3.16b}, [x5], x1
81 function ff_vp9_avg16_neon, export=1
84 ld1 {v2.16b}, [x2], x3
85 ld1 {v0.16b}, [x0], x1
86 ld1 {v3.16b}, [x2], x3
87 urhadd v0.16b, v0.16b, v2.16b
88 ld1 {v1.16b}, [x0], x1
89 urhadd v1.16b, v1.16b, v3.16b
91 st1 {v0.16b}, [x5], x1
92 st1 {v1.16b}, [x5], x1
97 function ff_vp9_copy8_neon, export=1
100 ld1 {v1.8b}, [x2], x3
102 st1 {v0.8b}, [x0], x1
103 st1 {v1.8b}, [x0], x1
108 function ff_vp9_avg8_neon, export=1
111 ld1 {v2.8b}, [x2], x3
112 ld1 {v0.8b}, [x0], x1
113 ld1 {v3.8b}, [x2], x3
114 urhadd v0.8b, v0.8b, v2.8b
115 ld1 {v1.8b}, [x0], x1
116 urhadd v1.8b, v1.8b, v3.8b
118 st1 {v0.8b}, [x5], x1
119 st1 {v1.8b}, [x5], x1
124 function ff_vp9_copy4_neon, export=1
126 ld1 {v0.s}[0], [x2], x3
127 ld1 {v1.s}[0], [x2], x3
128 st1 {v0.s}[0], [x0], x1
129 ld1 {v2.s}[0], [x2], x3
130 st1 {v1.s}[0], [x0], x1
131 ld1 {v3.s}[0], [x2], x3
133 st1 {v2.s}[0], [x0], x1
134 st1 {v3.s}[0], [x0], x1
139 function ff_vp9_avg4_neon, export=1
142 ld1 {v2.s}[0], [x2], x3
143 ld1 {v0.s}[0], [x0], x1
144 ld1 {v2.s}[1], [x2], x3
145 ld1 {v0.s}[1], [x0], x1
146 ld1 {v3.s}[0], [x2], x3
147 ld1 {v1.s}[0], [x0], x1
148 ld1 {v3.s}[1], [x2], x3
149 ld1 {v1.s}[1], [x0], x1
151 urhadd v0.8b, v0.8b, v2.8b
152 urhadd v1.8b, v1.8b, v3.8b
153 st1 {v0.s}[0], [x5], x1
154 st1 {v0.s}[1], [x5], x1
155 st1 {v1.s}[0], [x5], x1
156 st1 {v1.s}[1], [x5], x1
162 // Extract a vector from src1-src2 and src4-src5 (src1-src3 and src4-src6
163 // for size >= 16), and multiply-accumulate into dst1 and dst3 (or
164 // dst1-dst2 and dst3-dst4 for size >= 16)
165 .macro extmla dst1, dst2, dst3, dst4, src1, src2, src3, src4, src5, src6, offset, size
166 ext v20.16b, \src1\().16b, \src2\().16b, #(2*\offset)
167 ext v22.16b, \src4\().16b, \src5\().16b, #(2*\offset)
169 mla \dst1\().8h, v20.8h, v0.h[\offset]
170 ext v21.16b, \src2\().16b, \src3\().16b, #(2*\offset)
171 mla \dst3\().8h, v22.8h, v0.h[\offset]
172 ext v23.16b, \src5\().16b, \src6\().16b, #(2*\offset)
173 mla \dst2\().8h, v21.8h, v0.h[\offset]
174 mla \dst4\().8h, v23.8h, v0.h[\offset]
176 mla \dst1\().8h, v20.8h, v0.h[\offset]
177 mla \dst3\().8h, v22.8h, v0.h[\offset]
179 mla \dst1\().4h, v20.4h, v0.h[\offset]
180 mla \dst3\().4h, v22.4h, v0.h[\offset]
183 // The same as above, but don't accumulate straight into the
184 // destination, but use a temp register and accumulate with saturation.
185 .macro extmulqadd dst1, dst2, dst3, dst4, src1, src2, src3, src4, src5, src6, offset, size
186 ext v20.16b, \src1\().16b, \src2\().16b, #(2*\offset)
187 ext v22.16b, \src4\().16b, \src5\().16b, #(2*\offset)
189 mul v20.8h, v20.8h, v0.h[\offset]
190 ext v21.16b, \src2\().16b, \src3\().16b, #(2*\offset)
191 mul v22.8h, v22.8h, v0.h[\offset]
192 ext v23.16b, \src5\().16b, \src6\().16b, #(2*\offset)
193 mul v21.8h, v21.8h, v0.h[\offset]
194 mul v23.8h, v23.8h, v0.h[\offset]
196 mul v20.8h, v20.8h, v0.h[\offset]
197 mul v22.8h, v22.8h, v0.h[\offset]
199 mul v20.4h, v20.4h, v0.h[\offset]
200 mul v22.4h, v22.4h, v0.h[\offset]
203 sqadd \dst1\().4h, \dst1\().4h, v20.4h
204 sqadd \dst3\().4h, \dst3\().4h, v22.4h
206 sqadd \dst1\().8h, \dst1\().8h, v20.8h
207 sqadd \dst3\().8h, \dst3\().8h, v22.8h
209 sqadd \dst2\().8h, \dst2\().8h, v21.8h
210 sqadd \dst4\().8h, \dst4\().8h, v23.8h
216 // Instantiate a horizontal filter function for the given size.
217 // This can work on 4, 8 or 16 pixels in parallel; for larger
218 // widths it will do 16 pixels at a time and loop horizontally.
219 // The actual width is passed in x5, the height in w4 and the
220 // filter coefficients in x9. idx2 is the index of the largest
221 // filter coefficient (3 or 4) and idx1 is the other one of them.
222 .macro do_8tap_h type, size, idx1, idx2
223 function \type\()_8tap_\size\()h_\idx1\idx2
229 // Only size >= 16 loops horizontally and needs
230 // reduced dst stride
234 // size >= 16 loads two qwords and increments x2,
235 // for size 4/8 it's enough with one qword and no
241 // Load the filter vector
249 ld1 {v4.8b, v5.8b, v6.8b}, [x2], #24
250 ld1 {v16.8b, v17.8b, v18.8b}, [x7], #24
252 ld1 {v4.8b, v5.8b}, [x2]
253 ld1 {v16.8b, v17.8b}, [x7]
265 // Accumulate, adding idx2 last with a separate
266 // saturating add. The positive filter coefficients
267 // for all indices except idx2 must add up to less
268 // than 127 for this not to overflow.
269 mul v1.8h, v4.8h, v0.h[0]
270 mul v24.8h, v16.8h, v0.h[0]
272 mul v2.8h, v5.8h, v0.h[0]
273 mul v25.8h, v17.8h, v0.h[0]
275 extmla v1, v2, v24, v25, v4, v5, v6, v16, v17, v18, 1, \size
276 extmla v1, v2, v24, v25, v4, v5, v6, v16, v17, v18, 2, \size
277 extmla v1, v2, v24, v25, v4, v5, v6, v16, v17, v18, \idx1, \size
278 extmla v1, v2, v24, v25, v4, v5, v6, v16, v17, v18, 5, \size
279 extmla v1, v2, v24, v25, v4, v5, v6, v16, v17, v18, 6, \size
280 extmla v1, v2, v24, v25, v4, v5, v6, v16, v17, v18, 7, \size
281 extmulqadd v1, v2, v24, v25, v4, v5, v6, v16, v17, v18, \idx2, \size
283 // Round, shift and saturate
284 sqrshrun v1.8b, v1.8h, #7
285 sqrshrun v24.8b, v24.8h, #7
287 sqrshrun2 v1.16b, v2.8h, #7
288 sqrshrun2 v24.16b, v25.8h, #7
295 urhadd v1.16b, v1.16b, v2.16b
296 urhadd v24.16b, v24.16b, v3.16b
300 urhadd v1.8b, v1.8b, v2.8b
301 urhadd v24.8b, v24.8b, v3.8b
305 urhadd v1.8b, v1.8b, v2.8b
306 urhadd v24.8b, v24.8b, v3.8b
309 // Store and loop horizontally (for size >= 16)
312 st1 {v1.16b}, [x0], #16
313 st1 {v24.16b}, [x6], #16
317 ld1 {v6.16b}, [x2], #16
318 ld1 {v18.16b}, [x7], #16
322 uxtl2 v18.8h, v18.16b
343 .macro do_8tap_h_size size
344 do_8tap_h put, \size, 3, 4
345 do_8tap_h avg, \size, 3, 4
346 do_8tap_h put, \size, 4, 3
347 do_8tap_h avg, \size, 4, 3
354 .macro do_8tap_h_func type, filter, offset, size
355 function ff_vp9_\type\()_\filter\()\size\()_h_neon, export=1
356 movrel x6, X(ff_vp9_subpel_filters), 256*\offset
358 add x9, x6, w5, uxtw #4
361 b.ge \type\()_8tap_16h_34
362 b \type\()_8tap_16h_43
364 b.ge \type\()_8tap_\size\()h_34
365 b \type\()_8tap_\size\()h_43
370 .macro do_8tap_h_filters size
371 do_8tap_h_func put, regular, 1, \size
372 do_8tap_h_func avg, regular, 1, \size
373 do_8tap_h_func put, sharp, 2, \size
374 do_8tap_h_func avg, sharp, 2, \size
375 do_8tap_h_func put, smooth, 0, \size
376 do_8tap_h_func avg, smooth, 0, \size
388 // Round, shift and saturate and store reg1-reg2 over 4 lines
389 .macro do_store4 reg1, reg2, tmp1, tmp2, type
390 sqrshrun \reg1\().8b, \reg1\().8h, #7
391 sqrshrun \reg2\().8b, \reg2\().8h, #7
393 ld1 {\tmp1\().s}[0], [x7], x1
394 ld1 {\tmp2\().s}[0], [x7], x1
395 ld1 {\tmp1\().s}[1], [x7], x1
396 ld1 {\tmp2\().s}[1], [x7], x1
397 urhadd \reg1\().8b, \reg1\().8b, \tmp1\().8b
398 urhadd \reg2\().8b, \reg2\().8b, \tmp2\().8b
400 st1 {\reg1\().s}[0], [x0], x1
401 st1 {\reg2\().s}[0], [x0], x1
402 st1 {\reg1\().s}[1], [x0], x1
403 st1 {\reg2\().s}[1], [x0], x1
406 // Round, shift and saturate and store reg1-4
407 .macro do_store reg1, reg2, reg3, reg4, tmp1, tmp2, tmp3, tmp4, type
408 sqrshrun \reg1\().8b, \reg1\().8h, #7
409 sqrshrun \reg2\().8b, \reg2\().8h, #7
410 sqrshrun \reg3\().8b, \reg3\().8h, #7
411 sqrshrun \reg4\().8b, \reg4\().8h, #7
413 ld1 {\tmp1\().8b}, [x7], x1
414 ld1 {\tmp2\().8b}, [x7], x1
415 ld1 {\tmp3\().8b}, [x7], x1
416 ld1 {\tmp4\().8b}, [x7], x1
417 urhadd \reg1\().8b, \reg1\().8b, \tmp1\().8b
418 urhadd \reg2\().8b, \reg2\().8b, \tmp2\().8b
419 urhadd \reg3\().8b, \reg3\().8b, \tmp3\().8b
420 urhadd \reg4\().8b, \reg4\().8b, \tmp4\().8b
422 st1 {\reg1\().8b}, [x0], x1
423 st1 {\reg2\().8b}, [x0], x1
424 st1 {\reg3\().8b}, [x0], x1
425 st1 {\reg4\().8b}, [x0], x1
428 // Evaluate the filter twice in parallel, from the inputs src1-src9 into dst1-dst2
429 // (src1-src8 into dst1, src2-src9 into dst2), adding idx2 separately
430 // at the end with saturation. Indices 0 and 7 always have negative or zero
431 // coefficients, so they can be accumulated into tmp1-tmp2 together with the
432 // largest coefficient.
433 .macro convolve dst1, dst2, src1, src2, src3, src4, src5, src6, src7, src8, src9, idx1, idx2, tmp1, tmp2
434 mul \dst1\().8h, \src2\().8h, v0.h[1]
435 mul \dst2\().8h, \src3\().8h, v0.h[1]
436 mul \tmp1\().8h, \src1\().8h, v0.h[0]
437 mul \tmp2\().8h, \src2\().8h, v0.h[0]
438 mla \dst1\().8h, \src3\().8h, v0.h[2]
439 mla \dst2\().8h, \src4\().8h, v0.h[2]
441 mla \dst1\().8h, \src4\().8h, v0.h[3]
442 mla \dst2\().8h, \src5\().8h, v0.h[3]
444 mla \dst1\().8h, \src5\().8h, v0.h[4]
445 mla \dst2\().8h, \src6\().8h, v0.h[4]
447 mla \dst1\().8h, \src6\().8h, v0.h[5]
448 mla \dst2\().8h, \src7\().8h, v0.h[5]
449 mla \tmp1\().8h, \src8\().8h, v0.h[7]
450 mla \tmp2\().8h, \src9\().8h, v0.h[7]
451 mla \dst1\().8h, \src7\().8h, v0.h[6]
452 mla \dst2\().8h, \src8\().8h, v0.h[6]
454 mla \tmp1\().8h, \src4\().8h, v0.h[3]
455 mla \tmp2\().8h, \src5\().8h, v0.h[3]
457 mla \tmp1\().8h, \src5\().8h, v0.h[4]
458 mla \tmp2\().8h, \src6\().8h, v0.h[4]
460 sqadd \dst1\().8h, \dst1\().8h, \tmp1\().8h
461 sqadd \dst2\().8h, \dst2\().8h, \tmp2\().8h
464 // Load pixels and extend them to 16 bit
465 .macro loadl dst1, dst2, dst3, dst4
466 ld1 {v1.8b}, [x2], x3
467 ld1 {v2.8b}, [x2], x3
468 ld1 {v3.8b}, [x2], x3
470 ld1 {v4.8b}, [x2], x3
472 uxtl \dst1\().8h, v1.8b
473 uxtl \dst2\().8h, v2.8b
474 uxtl \dst3\().8h, v3.8b
476 uxtl \dst4\().8h, v4.8b
480 // Instantiate a vertical filter function for filtering 8 pixels at a time.
481 // The height is passed in x4, the width in x5 and the filter coefficients
482 // in x6. idx2 is the index of the largest filter coefficient (3 or 4)
483 // and idx1 is the other one of them.
484 .macro do_8tap_8v type, idx1, idx2
485 function \type\()_8tap_8v_\idx1\idx2
486 sub x2, x2, x3, lsl #1
497 loadl v20, v21, v22, v23
499 loadl v24, v25, v26, v27
500 convolve v1, v2, v17, v18, v19, v20, v21, v22, v23, v24, v25, \idx1, \idx2, v5, v6
501 convolve v3, v4, v19, v20, v21, v22, v23, v24, v25, v26, v27, \idx1, \idx2, v5, v6
502 do_store v1, v2, v3, v4, v5, v6, v7, v28, \type
507 loadl v16, v17, v18, v19
508 convolve v1, v2, v21, v22, v23, v24, v25, v26, v27, v16, v17, \idx1, \idx2, v5, v6
509 convolve v3, v4, v23, v24, v25, v26, v27, v16, v17, v18, v19, \idx1, \idx2, v5, v6
510 do_store v1, v2, v3, v4, v5, v6, v7, v28, \type
515 loadl v20, v21, v22, v23
516 convolve v1, v2, v25, v26, v27, v16, v17, v18, v19, v20, v21, \idx1, \idx2, v5, v6
517 convolve v3, v4, v27, v16, v17, v18, v19, v20, v21, v22, v23, \idx1, \idx2, v5, v6
518 do_store v1, v2, v3, v4, v5, v6, v7, v28, \type
526 // x0 -= h * dst_stride
528 // x2 -= h * src_stride
530 // x2 -= 8 * src_stride
531 sub x2, x2, x3, lsl #3
532 // x2 += 1 * src_stride
548 // Instantiate a vertical filter function for filtering a 4 pixels wide
549 // slice. The first half of the registers contain one row, while the second
550 // half of a register contains the second-next row (also stored in the first
551 // half of the register two steps ahead). The convolution does two outputs
552 // at a time; the output of v17-v24 into one, and v18-v25 into another one.
553 // The first half of first output is the first output row, the first half
554 // of the other output is the second output row. The second halves of the
555 // registers are rows 3 and 4.
556 // This only is designed to work for 4 or 8 output lines.
557 .macro do_8tap_4v type, idx1, idx2
558 function \type\()_8tap_4v_\idx1\idx2
559 sub x2, x2, x3, lsl #1
566 ld1 {v1.s}[0], [x2], x3
567 ld1 {v2.s}[0], [x2], x3
568 ld1 {v3.s}[0], [x2], x3
569 ld1 {v4.s}[0], [x2], x3
570 ld1 {v5.s}[0], [x2], x3
571 ld1 {v6.s}[0], [x2], x3
572 trn1 v1.2s, v1.2s, v3.2s
573 ld1 {v7.s}[0], [x2], x3
574 trn1 v2.2s, v2.2s, v4.2s
575 ld1 {v26.s}[0], [x2], x3
577 trn1 v3.2s, v3.2s, v5.2s
578 ld1 {v27.s}[0], [x2], x3
580 trn1 v4.2s, v4.2s, v6.2s
581 ld1 {v28.s}[0], [x2], x3
583 trn1 v5.2s, v5.2s, v7.2s
584 ld1 {v29.s}[0], [x2], x3
586 trn1 v6.2s, v6.2s, v26.2s
588 trn1 v7.2s, v7.2s, v27.2s
590 trn1 v26.2s, v26.2s, v28.2s
592 trn1 v27.2s, v27.2s, v29.2s
596 convolve v1, v2, v17, v18, v19, v20, v21, v22, v23, v24, v25, \idx1, \idx2, v3, v4
597 do_store4 v1, v2, v5, v6, \type
602 ld1 {v1.s}[0], [x2], x3
603 ld1 {v2.s}[0], [x2], x3
604 trn1 v28.2s, v28.2s, v1.2s
605 trn1 v29.2s, v29.2s, v2.2s
606 ld1 {v1.s}[1], [x2], x3
608 ld1 {v2.s}[1], [x2], x3
613 convolve v1, v2, v21, v22, v23, v24, v25, v26, v27, v28, v29, \idx1, \idx2, v3, v4
614 do_store4 v1, v2, v5, v6, \type
627 .macro do_8tap_v_func type, filter, offset, size
628 function ff_vp9_\type\()_\filter\()\size\()_v_neon, export=1
630 movrel x5, X(ff_vp9_subpel_filters), 256*\offset
632 add x6, x5, w6, uxtw #4
635 b.ge \type\()_8tap_8v_34
636 b \type\()_8tap_8v_43
638 b.ge \type\()_8tap_4v_34
639 b \type\()_8tap_4v_43
644 .macro do_8tap_v_filters size
645 do_8tap_v_func put, regular, 1, \size
646 do_8tap_v_func avg, regular, 1, \size
647 do_8tap_v_func put, sharp, 2, \size
648 do_8tap_v_func avg, sharp, 2, \size
649 do_8tap_v_func put, smooth, 0, \size
650 do_8tap_v_func avg, smooth, 0, \size