2 * Copyright (c) 2016 Google Inc.
4 * This file is part of Libav.
6 * Libav 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 * Libav 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 Libav; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 #include "libavutil/arm/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_copy64_neon, export=1
33 vld1.8 {q0, q1}, [r2]!
34 vst1.8 {q0, q1}, [r0, :128]!
35 vld1.8 {q2, q3}, [r2], r3
37 vst1.8 {q2, q3}, [r0, :128], r1
42 function ff_vp9_avg64_neon, export=1
49 vld1.8 {q8, q9}, [r2]!
50 vld1.8 {q0, q1}, [r0, :128]!
51 vld1.8 {q10, q11}, [r2], r3
53 vld1.8 {q2, q3}, [r0, :128], r1
56 vst1.8 {q0, q1}, [lr, :128]!
58 vst1.8 {q2, q3}, [lr, :128], r1
64 function ff_vp9_copy32_neon, export=1
67 vld1.8 {q0, q1}, [r2], r3
69 vst1.8 {q0, q1}, [r0, :128], r1
74 function ff_vp9_avg32_neon, export=1
77 vld1.8 {q2, q3}, [r2], r3
78 vld1.8 {q0, q1}, [r0, :128]
82 vst1.8 {q0, q1}, [r0, :128], r1
87 function ff_vp9_copy16_neon, export=1
98 vst1.8 {q0}, [r0, :128], r1
99 vst1.8 {q1}, [r4, :128], r1
104 function ff_vp9_avg16_neon, export=1
109 vld1.8 {q2}, [r2], r3
110 vld1.8 {q0}, [r0, :128], r1
111 vld1.8 {q3}, [r2], r3
113 vld1.8 {q1}, [r0, :128], r1
116 vst1.8 {q0}, [lr, :128], r1
117 vst1.8 {q1}, [lr, :128], r1
122 function ff_vp9_copy8_neon, export=1
125 vld1.8 {d0}, [r2], r3
126 vld1.8 {d1}, [r2], r3
128 vst1.8 {d0}, [r0, :64], r1
129 vst1.8 {d1}, [r0, :64], r1
134 function ff_vp9_avg8_neon, export=1
137 vld1.8 {d2}, [r2], r3
138 vld1.8 {d0}, [r0, :64], r1
139 vld1.8 {d3}, [r2], r3
141 vld1.8 {d1}, [r0, :64]
145 vst1.8 {d0}, [r0, :64], r1
146 vst1.8 {d1}, [r0, :64], r1
151 function ff_vp9_copy4_neon, export=1
154 vld1.32 {d0[]}, [r2], r3
155 vld1.32 {d1[]}, [r2], r3
156 vst1.32 {d0[0]}, [r0, :32], r1
157 vld1.32 {d2[]}, [r2], r3
158 vst1.32 {d1[0]}, [r0, :32], r1
159 vld1.32 {d3[]}, [r2], r3
161 vst1.32 {d2[0]}, [r0, :32], r1
162 vst1.32 {d3[0]}, [r0, :32], r1
167 function ff_vp9_avg4_neon, export=1
172 vld1.32 {d4[]}, [r2], r3
173 vld1.32 {d0[]}, [r0, :32], r1
174 vld1.32 {d5[]}, [r2], r3
176 vld1.32 {d1[]}, [r0, :32], r1
177 vld1.32 {d6[]}, [r2], r3
179 vld1.32 {d2[]}, [r0, :32], r1
180 vld1.32 {d7[]}, [r2], r3
182 vld1.32 {d3[]}, [r0, :32], r1
184 vst1.32 {d0[0]}, [lr, :32], r1
186 vst1.32 {d1[0]}, [lr, :32], r1
187 vst1.32 {d2[0]}, [lr, :32], r1
188 vst1.32 {d3[0]}, [lr, :32], r1
193 @ Helper macros for vmul/vmla with a constant from either d0 or d1 depending on index
194 .macro vmul_lane dst, src, idx
196 vmul.s16 \dst, \src, d0[\idx]
198 vmul.s16 \dst, \src, d1[\idx - 4]
201 .macro vmla_lane dst, src, idx
203 vmla.s16 \dst, \src, d0[\idx]
205 vmla.s16 \dst, \src, d1[\idx - 4]
209 @ Extract a vector from src1-src2 and src4-src5 (src1-src3 and src4-src6
210 @ for size >= 16), and multiply-accumulate into dst1 and dst3 (or
211 @ dst1-dst2 and dst3-dst4 for size >= 16)
212 .macro extmla dst1, dst2, dst3, dst4, src1, src2, src3, src4, src5, src6, offset, size
213 vext.8 q14, \src1, \src2, #(2*\offset)
214 vext.8 q15, \src4, \src5, #(2*\offset)
216 vmla_lane \dst1, q14, \offset
217 vext.8 q5, \src2, \src3, #(2*\offset)
218 vmla_lane \dst3, q15, \offset
219 vext.8 q6, \src5, \src6, #(2*\offset)
220 vmla_lane \dst2, q5, \offset
221 vmla_lane \dst4, q6, \offset
223 vmla_lane \dst1, q14, \offset
224 vmla_lane \dst3, q15, \offset
227 @ The same as above, but don't accumulate straight into the
228 @ destination, but use a temp register and accumulate with saturation.
229 .macro extmulqadd dst1, dst2, dst3, dst4, src1, src2, src3, src4, src5, src6, offset, size
230 vext.8 q14, \src1, \src2, #(2*\offset)
231 vext.8 q15, \src4, \src5, #(2*\offset)
233 vmul_lane q14, q14, \offset
234 vext.8 q5, \src2, \src3, #(2*\offset)
235 vmul_lane q15, q15, \offset
236 vext.8 q6, \src5, \src6, #(2*\offset)
237 vmul_lane q5, q5, \offset
238 vmul_lane q6, q6, \offset
240 vmul_lane q14, q14, \offset
241 vmul_lane q15, q15, \offset
243 vqadd.s16 \dst1, \dst1, q14
244 vqadd.s16 \dst3, \dst3, q15
246 vqadd.s16 \dst2, \dst2, q5
247 vqadd.s16 \dst4, \dst4, q6
252 @ Instantiate a horizontal filter function for the given size.
253 @ This can work on 4, 8 or 16 pixels in parallel; for larger
254 @ widths it will do 16 pixels at a time and loop horizontally.
255 @ The actual width is passed in r5, the height in r4 and
256 @ the filter coefficients in r12. idx2 is the index of the largest
257 @ filter coefficient (3 or 4) and idx1 is the other one of them.
258 .macro do_8tap_h type, size, idx1, idx2
259 function \type\()_8tap_\size\()h_\idx1\idx2
265 @ Only size >= 16 loops horizontally and needs
270 @ size >= 16 loads two qwords and increments r2,
271 @ for size 4/8 it's enough with one qword and no
277 @ Load the filter vector
278 vld1.8 {d0}, [r12,:64]
286 vld1.8 {d18, d19, d20}, [r2]!
287 vld1.8 {d24, d25, d26}, [r7]!
302 @ Accumulate, adding idx2 last with a separate
303 @ saturating add. The positive filter coefficients
304 @ for all indices except idx2 must add up to less
305 @ than 127 for this not to overflow.
306 vmul.s16 q1, q8, d0[0]
307 vmul.s16 q3, q11, d0[0]
309 vmul.s16 q2, q9, d0[0]
310 vmul.s16 q4, q12, d0[0]
312 extmla q1, q2, q3, q4, q8, q9, q10, q11, q12, q13, 1, \size
313 extmla q1, q2, q3, q4, q8, q9, q10, q11, q12, q13, 2, \size
314 extmla q1, q2, q3, q4, q8, q9, q10, q11, q12, q13, \idx1, \size
315 extmla q1, q2, q3, q4, q8, q9, q10, q11, q12, q13, 5, \size
316 extmla q1, q2, q3, q4, q8, q9, q10, q11, q12, q13, 6, \size
317 extmla q1, q2, q3, q4, q8, q9, q10, q11, q12, q13, 7, \size
318 extmulqadd q1, q2, q3, q4, q8, q9, q10, q11, q12, q13, \idx2, \size
320 @ Round, shift and saturate
321 vqrshrun.s16 d2, q1, #7
322 vqrshrun.s16 d6, q3, #7
324 vqrshrun.s16 d3, q2, #7
325 vqrshrun.s16 d7, q4, #7
330 vld1.8 {q14}, [r0,:128]
331 vld1.8 {q15}, [r6,:128]
332 vrhadd.u8 q1, q1, q14
333 vrhadd.u8 q3, q3, q15
335 vld1.8 {d28}, [r0,:64]
336 vld1.8 {d30}, [r6,:64]
337 vrhadd.u8 d2, d2, d28
338 vrhadd.u8 d6, d6, d30
340 @ We only need d28[0], but [] is faster on some cores
341 vld1.32 {d28[]}, [r0,:32]
342 vld1.32 {d30[]}, [r6,:32]
343 vrhadd.u8 d2, d2, d28
344 vrhadd.u8 d6, d6, d30
347 @ Store and loop horizontally (for size >= 16)
350 vst1.8 {q1}, [r0,:128]!
351 vst1.8 {q3}, [r6,:128]!
363 vst1.8 {d2}, [r0,:64]
364 vst1.8 {d6}, [r6,:64]
366 vst1.32 {d2[0]}, [r0,:32]
367 vst1.32 {d6[0]}, [r6,:32]
385 .macro do_8tap_h_size size
386 do_8tap_h put, \size, 3, 4
387 do_8tap_h avg, \size, 3, 4
388 do_8tap_h put, \size, 4, 3
389 do_8tap_h avg, \size, 4, 3
396 .macro do_8tap_h_func type, filter, offset, size
397 function ff_vp9_\type\()_\filter\()\size\()_h_neon, export=1
407 movrelx r12, X(ff_vp9_subpel_filters), r6
408 add r12, r12, 120*\offset - 8
410 add r12, r12, r5, lsl #3
413 bge \type\()_8tap_16h_34
414 b \type\()_8tap_16h_43
416 bge \type\()_8tap_\size\()h_34
417 b \type\()_8tap_\size\()h_43
422 .macro do_8tap_h_filters size
423 do_8tap_h_func put, regular, 1, \size
424 do_8tap_h_func avg, regular, 1, \size
425 do_8tap_h_func put, sharp, 2, \size
426 do_8tap_h_func avg, sharp, 2, \size
427 do_8tap_h_func put, smooth, 0, \size
428 do_8tap_h_func avg, smooth, 0, \size
441 @ Round, shift and saturate and store qreg1-2 over 4 lines
442 .macro do_store4 qreg1, dreg1, qreg2, dreg2, tmp1, tmp2, type
443 vqrshrun.s16 \dreg1, \qreg1, #7
444 vqrshrun.s16 \dreg2, \qreg2, #7
446 vld1.32 {\tmp1[]}, [r0,:32], r1
447 vld1.32 {\tmp2[]}, [r0,:32], r1
448 vld1.32 {\tmp1[1]}, [r0,:32], r1
449 vld1.32 {\tmp2[1]}, [r0,:32], r1
450 vrhadd.u8 \dreg1, \dreg1, \tmp1
451 vrhadd.u8 \dreg2, \dreg2, \tmp2
452 sub r0, r0, r1, lsl #2
454 vst1.32 {\dreg1[0]}, [r0,:32], r1
455 vst1.32 {\dreg2[0]}, [r0,:32], r1
456 vst1.32 {\dreg1[1]}, [r0,:32], r1
457 vst1.32 {\dreg2[1]}, [r0,:32], r1
460 @ Round, shift and saturate and store qreg1-4
461 .macro do_store qreg1, dreg1, qreg2, dreg2, qreg3, dreg3, qreg4, dreg4, tmp1, tmp2, tmp3, tmp4, type
462 vqrshrun.s16 \dreg1, \qreg1, #7
463 vqrshrun.s16 \dreg2, \qreg2, #7
464 vqrshrun.s16 \dreg3, \qreg3, #7
465 vqrshrun.s16 \dreg4, \qreg4, #7
467 vld1.8 {\tmp1}, [r0,:64], r1
468 vld1.8 {\tmp2}, [r0,:64], r1
469 vld1.8 {\tmp3}, [r0,:64], r1
470 vld1.8 {\tmp4}, [r0,:64], r1
471 vrhadd.u8 \dreg1, \dreg1, \tmp1
472 vrhadd.u8 \dreg2, \dreg2, \tmp2
473 vrhadd.u8 \dreg3, \dreg3, \tmp3
474 vrhadd.u8 \dreg4, \dreg4, \tmp4
475 sub r0, r0, r1, lsl #2
477 vst1.8 {\dreg1}, [r0,:64], r1
478 vst1.8 {\dreg2}, [r0,:64], r1
479 vst1.8 {\dreg3}, [r0,:64], r1
480 vst1.8 {\dreg4}, [r0,:64], r1
483 @ Evaluate the filter twice in parallel, from the inputs src1-src9 into dst1-dst2
484 @ (src1-src8 into dst1, src2-src9 into dst2), adding idx2 separately
485 @ at the end with saturation. Indices 0 and 7 always have negative or zero
486 @ coefficients, so they can be accumulated into tmp1-tmp2 together with the
487 @ largest coefficient.
488 .macro convolve dst1, dst2, src1, src2, src3, src4, src5, src6, src7, src8, src9, idx1, idx2, tmp1, tmp2
489 vmul.s16 \dst1, \src2, d0[1]
490 vmul.s16 \dst2, \src3, d0[1]
491 vmul.s16 \tmp1, \src1, d0[0]
492 vmul.s16 \tmp2, \src2, d0[0]
493 vmla.s16 \dst1, \src3, d0[2]
494 vmla.s16 \dst2, \src4, d0[2]
496 vmla.s16 \dst1, \src4, d0[3]
497 vmla.s16 \dst2, \src5, d0[3]
499 vmla.s16 \dst1, \src5, d1[0]
500 vmla.s16 \dst2, \src6, d1[0]
502 vmla.s16 \dst1, \src6, d1[1]
503 vmla.s16 \dst2, \src7, d1[1]
504 vmla.s16 \tmp1, \src8, d1[3]
505 vmla.s16 \tmp2, \src9, d1[3]
506 vmla.s16 \dst1, \src7, d1[2]
507 vmla.s16 \dst2, \src8, d1[2]
509 vmla.s16 \tmp1, \src4, d0[3]
510 vmla.s16 \tmp2, \src5, d0[3]
512 vmla.s16 \tmp1, \src5, d1[0]
513 vmla.s16 \tmp2, \src6, d1[0]
515 vqadd.s16 \dst1, \dst1, \tmp1
516 vqadd.s16 \dst2, \dst2, \tmp2
519 @ Load pixels and extend them to 16 bit
520 .macro loadl dst1, dst2, dst3, dst4
521 vld1.8 {d2}, [r2], r3
522 vld1.8 {d3}, [r2], r3
523 vld1.8 {d4}, [r2], r3
525 vld1.8 {d5}, [r2], r3
535 @ Instantiate a vertical filter function for filtering 8 pixels at a time.
536 @ The height is passed in r4, the width in r5 and the filter coefficients
537 @ in r12. idx2 is the index of the largest filter coefficient (3 or 4)
538 @ and idx1 is the other one of them.
539 .macro do_8tap_8v type, idx1, idx2
540 function \type\()_8tap_8v_\idx1\idx2
541 sub r2, r2, r3, lsl #1
543 vld1.8 {d0}, [r12, :64]
549 loadl q8, q9, q10, q11
551 loadl q12, q13, q14, q15
552 convolve q1, q2, q5, q6, q7, q8, q9, q10, q11, q12, q13, \idx1, \idx2, q4, q5
553 convolve q3, q4, q7, q8, q9, q10, q11, q12, q13, q14, q15, \idx1, \idx2, q5, q6
554 do_store q1, d2, q2, d4, q3, d6, q4, d8, d3, d5, d7, d9, \type
560 convolve q1, q2, q9, q10, q11, q12, q13, q14, q15, q4, q5, \idx1, \idx2, q8, q9
561 convolve q3, q8, q11, q12, q13, q14, q15, q4, q5, q6, q7, \idx1, \idx2, q9, q10
562 do_store q1, d2, q2, d4, q3, d6, q8, d16, d3, d5, d7, d17, \type
567 loadl q8, q9, q10, q11
568 convolve q1, q2, q13, q14, q15, q4, q5, q6, q7, q8, q9, \idx1, \idx2, q12, q13
569 convolve q3, q12, q15, q4, q5, q6, q7, q8, q9, q10, q11, \idx1, \idx2, q13, q14
570 do_store q1, d2, q2, d4, q3, d6, q12, d24, d3, d5, d7, d25, \type
578 @ r0 -= h * dst_stride
580 @ r2 -= h * src_stride
582 @ r2 -= 8 * src_stride
583 sub r2, r2, r3, lsl #3
584 @ r2 += 1 * src_stride
601 @ Instantiate a vertical filter function for filtering a 4 pixels wide
602 @ slice. The first half of the registers contain one row, while the second
603 @ half of a register contains the second-next row (also stored in the first
604 @ half of the register two steps ahead). The convolution does two outputs
605 @ at a time; the output of q5-q12 into one, and q4-q13 into another one.
606 @ The first half of first output is the first output row, the first half
607 @ of the other output is the second output row. The second halves of the
608 @ registers are rows 3 and 4.
609 @ This only is designed to work for 4 or 8 output lines.
610 .macro do_8tap_4v type, idx1, idx2
611 function \type\()_8tap_4v_\idx1\idx2
612 sub r2, r2, r3, lsl #1
614 vld1.8 {d0}, [r12, :64]
617 vld1.32 {d2[]}, [r2], r3
618 vld1.32 {d3[]}, [r2], r3
619 vld1.32 {d4[]}, [r2], r3
620 vld1.32 {d5[]}, [r2], r3
621 vld1.32 {d6[]}, [r2], r3
622 vld1.32 {d7[]}, [r2], r3
623 vext.8 d2, d2, d4, #4
624 vld1.32 {d8[]}, [r2], r3
625 vext.8 d3, d3, d5, #4
626 vld1.32 {d9[]}, [r2], r3
628 vext.8 d4, d4, d6, #4
629 vld1.32 {d28[]}, [r2], r3
631 vext.8 d5, d5, d7, #4
632 vld1.32 {d29[]}, [r2], r3
634 vext.8 d6, d6, d8, #4
635 vld1.32 {d30[]}, [r2], r3
637 vext.8 d7, d7, d9, #4
639 vext.8 d8, d8, d28, #4
641 vext.8 d9, d9, d29, #4
643 vext.8 d28, d28, d30, #4
647 convolve q1, q2, q5, q6, q7, q8, q9, q10, q11, q12, q13, \idx1, \idx2, q4, q3
648 do_store4 q1, d2, q2, d4, d3, d5, \type
652 vld1.32 {d2[]}, [r2], r3
653 vld1.32 {d3[]}, [r2], r3
654 vext.8 d29, d29, d2, #4
655 vext.8 d30, d30, d3, #4
656 vld1.32 {d2[1]}, [r2], r3
658 vld1.32 {d3[1]}, [r2], r3
663 convolve q1, q2, q9, q10, q11, q12, q13, q14, q15, q5, q6, \idx1, \idx2, q4, q3
664 do_store4 q1, d2, q2, d4, d3, d5, \type
678 .macro do_8tap_v_func type, filter, offset, size
679 function ff_vp9_\type\()_\filter\()\size\()_v_neon, export=1
683 movrelx r12, X(ff_vp9_subpel_filters), r5
685 add r12, r12, 120*\offset - 8
686 add r12, r12, r5, lsl #3
690 bge \type\()_8tap_8v_34
691 b \type\()_8tap_8v_43
693 bge \type\()_8tap_4v_34
694 b \type\()_8tap_4v_43
699 .macro do_8tap_v_filters size
700 do_8tap_v_func put, regular, 1, \size
701 do_8tap_v_func avg, regular, 1, \size
702 do_8tap_v_func put, sharp, 2, \size
703 do_8tap_v_func avg, sharp, 2, \size
704 do_8tap_v_func put, smooth, 0, \size
705 do_8tap_v_func avg, smooth, 0, \size