2 * VP9 compatible video decoder
4 * Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>
5 * Copyright (C) 2013 Clément Bœsch <u pkh me>
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 #include "libavutil/avassert.h"
32 static av_always_inline int check_intra_mode(VP9TileData *td, int mode, uint8_t **a,
33 uint8_t *dst_edge, ptrdiff_t stride_edge,
34 uint8_t *dst_inner, ptrdiff_t stride_inner,
35 uint8_t *l, int col, int x, int w,
36 int row, int y, enum TxfmMode tx,
37 int p, int ss_h, int ss_v, int bytesperpixel)
39 VP9Context *s = td->s;
40 int have_top = row > 0 || y > 0;
41 int have_left = col > td->tile_col_start || x > 0;
42 int have_right = x < w - 1;
44 static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {
45 [VERT_PRED] = { { DC_127_PRED, VERT_PRED },
46 { DC_127_PRED, VERT_PRED } },
47 [HOR_PRED] = { { DC_129_PRED, DC_129_PRED },
48 { HOR_PRED, HOR_PRED } },
49 [DC_PRED] = { { DC_128_PRED, TOP_DC_PRED },
50 { LEFT_DC_PRED, DC_PRED } },
51 [DIAG_DOWN_LEFT_PRED] = { { DC_127_PRED, DIAG_DOWN_LEFT_PRED },
52 { DC_127_PRED, DIAG_DOWN_LEFT_PRED } },
53 [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },
54 { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },
55 [VERT_RIGHT_PRED] = { { VERT_RIGHT_PRED, VERT_RIGHT_PRED },
56 { VERT_RIGHT_PRED, VERT_RIGHT_PRED } },
57 [HOR_DOWN_PRED] = { { HOR_DOWN_PRED, HOR_DOWN_PRED },
58 { HOR_DOWN_PRED, HOR_DOWN_PRED } },
59 [VERT_LEFT_PRED] = { { DC_127_PRED, VERT_LEFT_PRED },
60 { DC_127_PRED, VERT_LEFT_PRED } },
61 [HOR_UP_PRED] = { { DC_129_PRED, DC_129_PRED },
62 { HOR_UP_PRED, HOR_UP_PRED } },
63 [TM_VP8_PRED] = { { DC_129_PRED, VERT_PRED },
64 { HOR_PRED, TM_VP8_PRED } },
69 uint8_t needs_topleft:1;
70 uint8_t needs_topright:1;
71 uint8_t invert_left:1;
72 } edges[N_INTRA_PRED_MODES] = {
73 [VERT_PRED] = { .needs_top = 1 },
74 [HOR_PRED] = { .needs_left = 1 },
75 [DC_PRED] = { .needs_top = 1, .needs_left = 1 },
76 [DIAG_DOWN_LEFT_PRED] = { .needs_top = 1, .needs_topright = 1 },
77 [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,
79 [VERT_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,
81 [HOR_DOWN_PRED] = { .needs_left = 1, .needs_top = 1,
83 [VERT_LEFT_PRED] = { .needs_top = 1, .needs_topright = 1 },
84 [HOR_UP_PRED] = { .needs_left = 1, .invert_left = 1 },
85 [TM_VP8_PRED] = { .needs_left = 1, .needs_top = 1,
87 [LEFT_DC_PRED] = { .needs_left = 1 },
88 [TOP_DC_PRED] = { .needs_top = 1 },
89 [DC_128_PRED] = { 0 },
90 [DC_127_PRED] = { 0 },
94 av_assert2(mode >= 0 && mode < 10);
95 mode = mode_conv[mode][have_left][have_top];
96 if (edges[mode].needs_top) {
97 uint8_t *top, *topleft;
98 int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !ss_h) - x) * 4;
101 if (tx == TX_4X4 && edges[mode].needs_topright && have_right)
104 // if top of sb64-row, use s->intra_pred_data[] instead of
105 // dst[-stride] for intra prediction (it contains pre- instead of
106 // post-loopfilter data)
108 top = !(row & 7) && !y ?
109 s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :
110 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];
112 topleft = !(row & 7) && !y ?
113 s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :
114 y == 0 || x == 0 ? &dst_edge[-stride_edge] :
115 &dst_inner[-stride_inner];
119 (!edges[mode].needs_topleft || (have_left && top == topleft)) &&
120 (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&
121 n_px_need + n_px_need_tr <= n_px_have) {
125 if (n_px_need <= n_px_have) {
126 memcpy(*a, top, n_px_need * bytesperpixel);
128 #define memset_bpp(c, i1, v, i2, num) do { \
129 if (bytesperpixel == 1) { \
130 memset(&(c)[(i1)], (v)[(i2)], (num)); \
132 int n, val = AV_RN16A(&(v)[(i2) * 2]); \
133 for (n = 0; n < (num); n++) { \
134 AV_WN16A(&(c)[((i1) + n) * 2], val); \
138 memcpy(*a, top, n_px_have * bytesperpixel);
139 memset_bpp(*a, n_px_have, (*a), n_px_have - 1, n_px_need - n_px_have);
142 #define memset_val(c, val, num) do { \
143 if (bytesperpixel == 1) { \
144 memset((c), (val), (num)); \
147 for (n = 0; n < (num); n++) { \
148 AV_WN16A(&(c)[n * 2], (val)); \
152 memset_val(*a, (128 << (bpp - 8)) - 1, n_px_need);
154 if (edges[mode].needs_topleft) {
155 if (have_left && have_top) {
156 #define assign_bpp(c, i1, v, i2) do { \
157 if (bytesperpixel == 1) { \
158 (c)[(i1)] = (v)[(i2)]; \
160 AV_COPY16(&(c)[(i1) * 2], &(v)[(i2) * 2]); \
163 assign_bpp(*a, -1, topleft, -1);
165 #define assign_val(c, i, v) do { \
166 if (bytesperpixel == 1) { \
169 AV_WN16A(&(c)[(i) * 2], (v)); \
172 assign_val((*a), -1, (128 << (bpp - 8)) + (have_top ? +1 : -1));
175 if (tx == TX_4X4 && edges[mode].needs_topright) {
176 if (have_top && have_right &&
177 n_px_need + n_px_need_tr <= n_px_have) {
178 memcpy(&(*a)[4 * bytesperpixel], &top[4 * bytesperpixel], 4 * bytesperpixel);
180 memset_bpp(*a, 4, *a, 3, 4);
185 if (edges[mode].needs_left) {
187 int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !ss_v) - y) * 4;
188 uint8_t *dst = x == 0 ? dst_edge : dst_inner;
189 ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;
191 if (edges[mode].invert_left) {
192 if (n_px_need <= n_px_have) {
193 for (i = 0; i < n_px_need; i++)
194 assign_bpp(l, i, &dst[i * stride], -1);
196 for (i = 0; i < n_px_have; i++)
197 assign_bpp(l, i, &dst[i * stride], -1);
198 memset_bpp(l, n_px_have, l, n_px_have - 1, n_px_need - n_px_have);
201 if (n_px_need <= n_px_have) {
202 for (i = 0; i < n_px_need; i++)
203 assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);
205 for (i = 0; i < n_px_have; i++)
206 assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);
207 memset_bpp(l, 0, l, n_px_need - n_px_have, n_px_need - n_px_have);
211 memset_val(l, (128 << (bpp - 8)) + 1, 4 << tx);
218 static av_always_inline void intra_recon(VP9TileData *td, ptrdiff_t y_off,
219 ptrdiff_t uv_off, int bytesperpixel)
221 VP9Context *s = td->s;
223 int row = td->row, col = td->col;
224 int w4 = ff_vp9_bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;
225 int h4 = ff_vp9_bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);
226 int end_x = FFMIN(2 * (s->cols - col), w4);
227 int end_y = FFMIN(2 * (s->rows - row), h4);
228 int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;
229 int uvstep1d = 1 << b->uvtx, p;
230 uint8_t *dst = td->dst[0], *dst_r = s->s.frames[CUR_FRAME].tf.f->data[0] + y_off;
231 LOCAL_ALIGNED_32(uint8_t, a_buf, [96]);
232 LOCAL_ALIGNED_32(uint8_t, l, [64]);
234 for (n = 0, y = 0; y < end_y; y += step1d) {
235 uint8_t *ptr = dst, *ptr_r = dst_r;
236 for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d * bytesperpixel,
237 ptr_r += 4 * step1d * bytesperpixel, n += step) {
238 int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?
240 uint8_t *a = &a_buf[32];
241 enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];
242 int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&td->eob[n]) : td->eob[n];
244 mode = check_intra_mode(td, mode, &a, ptr_r,
245 s->s.frames[CUR_FRAME].tf.f->linesize[0],
246 ptr, td->y_stride, l,
247 col, x, w4, row, y, b->tx, 0, 0, 0, bytesperpixel);
248 s->dsp.intra_pred[b->tx][mode](ptr, td->y_stride, l, a);
250 s->dsp.itxfm_add[tx][txtp](ptr, td->y_stride,
251 td->block + 16 * n * bytesperpixel, eob);
253 dst_r += 4 * step1d * s->s.frames[CUR_FRAME].tf.f->linesize[0];
254 dst += 4 * step1d * td->y_stride;
261 step = 1 << (b->uvtx * 2);
262 for (p = 0; p < 2; p++) {
263 dst = td->dst[1 + p];
264 dst_r = s->s.frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;
265 for (n = 0, y = 0; y < end_y; y += uvstep1d) {
266 uint8_t *ptr = dst, *ptr_r = dst_r;
267 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d * bytesperpixel,
268 ptr_r += 4 * uvstep1d * bytesperpixel, n += step) {
269 int mode = b->uvmode;
270 uint8_t *a = &a_buf[32];
271 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&td->uveob[p][n]) : td->uveob[p][n];
273 mode = check_intra_mode(td, mode, &a, ptr_r,
274 s->s.frames[CUR_FRAME].tf.f->linesize[1],
275 ptr, td->uv_stride, l, col, x, w4, row, y,
276 b->uvtx, p + 1, s->ss_h, s->ss_v, bytesperpixel);
277 s->dsp.intra_pred[b->uvtx][mode](ptr, td->uv_stride, l, a);
279 s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, td->uv_stride,
280 td->uvblock[p] + 16 * n * bytesperpixel, eob);
282 dst_r += 4 * uvstep1d * s->s.frames[CUR_FRAME].tf.f->linesize[1];
283 dst += 4 * uvstep1d * td->uv_stride;
288 void ff_vp9_intra_recon_8bpp(VP9TileData *td, ptrdiff_t y_off, ptrdiff_t uv_off)
290 intra_recon(td, y_off, uv_off, 1);
293 void ff_vp9_intra_recon_16bpp(VP9TileData *td, ptrdiff_t y_off, ptrdiff_t uv_off)
295 intra_recon(td, y_off, uv_off, 2);
298 static av_always_inline void mc_luma_unscaled(VP9TileData *td, vp9_mc_func (*mc)[2],
299 uint8_t *dst, ptrdiff_t dst_stride,
300 const uint8_t *ref, ptrdiff_t ref_stride,
301 ThreadFrame *ref_frame,
302 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,
303 int bw, int bh, int w, int h, int bytesperpixel)
305 VP9Context *s = td->s;
306 int mx = mv->x, my = mv->y, th;
310 ref += y * ref_stride + x * bytesperpixel;
313 // FIXME bilinear filter only needs 0/1 pixels, not 3/4
314 // we use +7 because the last 7 pixels of each sbrow can be changed in
315 // the longest loopfilter of the next sbrow
316 th = (y + bh + 4 * !!my + 7) >> 6;
317 ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
318 // The arm/aarch64 _hv filters read one more row than what actually is
319 // needed, so switch to emulated edge one pixel sooner vertically
320 // (!!my * 5) than horizontally (!!mx * 4).
321 if (x < !!mx * 3 || y < !!my * 3 ||
322 x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
323 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
324 ref - !!my * 3 * ref_stride - !!mx * 3 * bytesperpixel,
326 bw + !!mx * 7, bh + !!my * 7,
327 x - !!mx * 3, y - !!my * 3, w, h);
328 ref = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
331 mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);
334 static av_always_inline void mc_chroma_unscaled(VP9TileData *td, vp9_mc_func (*mc)[2],
335 uint8_t *dst_u, uint8_t *dst_v,
336 ptrdiff_t dst_stride,
337 const uint8_t *ref_u, ptrdiff_t src_stride_u,
338 const uint8_t *ref_v, ptrdiff_t src_stride_v,
339 ThreadFrame *ref_frame,
340 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,
341 int bw, int bh, int w, int h, int bytesperpixel)
343 VP9Context *s = td->s;
344 int mx = mv->x * (1 << !s->ss_h), my = mv->y * (1 << !s->ss_v), th;
348 ref_u += y * src_stride_u + x * bytesperpixel;
349 ref_v += y * src_stride_v + x * bytesperpixel;
352 // FIXME bilinear filter only needs 0/1 pixels, not 3/4
353 // we use +7 because the last 7 pixels of each sbrow can be changed in
354 // the longest loopfilter of the next sbrow
355 th = (y + bh + 4 * !!my + 7) >> (6 - s->ss_v);
356 ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
357 // The arm/aarch64 _hv filters read one more row than what actually is
358 // needed, so switch to emulated edge one pixel sooner vertically
359 // (!!my * 5) than horizontally (!!mx * 4).
360 if (x < !!mx * 3 || y < !!my * 3 ||
361 x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
362 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
363 ref_u - !!my * 3 * src_stride_u - !!mx * 3 * bytesperpixel,
365 bw + !!mx * 7, bh + !!my * 7,
366 x - !!mx * 3, y - !!my * 3, w, h);
367 ref_u = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
368 mc[!!mx][!!my](dst_u, dst_stride, ref_u, 160, bh, mx, my);
370 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
371 ref_v - !!my * 3 * src_stride_v - !!mx * 3 * bytesperpixel,
373 bw + !!mx * 7, bh + !!my * 7,
374 x - !!mx * 3, y - !!my * 3, w, h);
375 ref_v = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
376 mc[!!mx][!!my](dst_v, dst_stride, ref_v, 160, bh, mx, my);
378 mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);
379 mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);
383 #define mc_luma_dir(td, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
384 px, py, pw, ph, bw, bh, w, h, i) \
385 mc_luma_unscaled(td, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
386 mv, bw, bh, w, h, bytesperpixel)
387 #define mc_chroma_dir(td, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
388 row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \
389 mc_chroma_unscaled(td, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
390 row, col, mv, bw, bh, w, h, bytesperpixel)
392 #define FN(x) x##_8bpp
393 #define BYTES_PER_PIXEL 1
394 #include "vp9_mc_template.c"
396 #undef BYTES_PER_PIXEL
397 #define FN(x) x##_16bpp
398 #define BYTES_PER_PIXEL 2
399 #include "vp9_mc_template.c"
403 #undef BYTES_PER_PIXEL
406 static av_always_inline void mc_luma_scaled(VP9TileData *td, vp9_scaled_mc_func smc,
407 vp9_mc_func (*mc)[2],
408 uint8_t *dst, ptrdiff_t dst_stride,
409 const uint8_t *ref, ptrdiff_t ref_stride,
410 ThreadFrame *ref_frame,
411 ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,
412 int px, int py, int pw, int ph,
413 int bw, int bh, int w, int h, int bytesperpixel,
414 const uint16_t *scale, const uint8_t *step)
416 VP9Context *s = td->s;
417 if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&
418 s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {
419 mc_luma_unscaled(td, mc, dst, dst_stride, ref, ref_stride, ref_frame,
420 y, x, in_mv, bw, bh, w, h, bytesperpixel);
422 #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)
424 int refbw_m1, refbh_m1;
428 mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);
429 mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);
430 // BUG libvpx seems to scale the two components separately. This introduces
431 // rounding errors but we have to reproduce them to be exactly compatible
432 // with the output from libvpx...
433 mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);
434 my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);
438 ref += y * ref_stride + x * bytesperpixel;
441 refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;
442 refbh_m1 = ((bh - 1) * step[1] + my) >> 4;
443 // FIXME bilinear filter only needs 0/1 pixels, not 3/4
444 // we use +7 because the last 7 pixels of each sbrow can be changed in
445 // the longest loopfilter of the next sbrow
446 th = (y + refbh_m1 + 4 + 7) >> 6;
447 ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
448 // The arm/aarch64 _hv filters read one more row than what actually is
449 // needed, so switch to emulated edge one pixel sooner vertically
450 // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).
451 if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {
452 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
453 ref - 3 * ref_stride - 3 * bytesperpixel,
455 refbw_m1 + 8, refbh_m1 + 8,
457 ref = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
460 smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);
464 static av_always_inline void mc_chroma_scaled(VP9TileData *td, vp9_scaled_mc_func smc,
465 vp9_mc_func (*mc)[2],
466 uint8_t *dst_u, uint8_t *dst_v,
467 ptrdiff_t dst_stride,
468 const uint8_t *ref_u, ptrdiff_t src_stride_u,
469 const uint8_t *ref_v, ptrdiff_t src_stride_v,
470 ThreadFrame *ref_frame,
471 ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,
472 int px, int py, int pw, int ph,
473 int bw, int bh, int w, int h, int bytesperpixel,
474 const uint16_t *scale, const uint8_t *step)
476 VP9Context *s = td->s;
477 if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&
478 s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {
479 mc_chroma_unscaled(td, mc, dst_u, dst_v, dst_stride, ref_u, src_stride_u,
480 ref_v, src_stride_v, ref_frame,
481 y, x, in_mv, bw, bh, w, h, bytesperpixel);
484 int refbw_m1, refbh_m1;
489 // BUG https://code.google.com/p/webm/issues/detail?id=820
490 mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 16, (s->cols * 4 - x + px + 3) * 16);
491 mx = scale_mv(mv.x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);
493 mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);
494 mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);
497 // BUG https://code.google.com/p/webm/issues/detail?id=820
498 mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 16, (s->rows * 4 - y + py + 3) * 16);
499 my = scale_mv(mv.y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);
501 mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);
502 my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);
507 ref_u += y * src_stride_u + x * bytesperpixel;
508 ref_v += y * src_stride_v + x * bytesperpixel;
511 refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;
512 refbh_m1 = ((bh - 1) * step[1] + my) >> 4;
513 // FIXME bilinear filter only needs 0/1 pixels, not 3/4
514 // we use +7 because the last 7 pixels of each sbrow can be changed in
515 // the longest loopfilter of the next sbrow
516 th = (y + refbh_m1 + 4 + 7) >> (6 - s->ss_v);
517 ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
518 // The arm/aarch64 _hv filters read one more row than what actually is
519 // needed, so switch to emulated edge one pixel sooner vertically
520 // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).
521 if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {
522 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
523 ref_u - 3 * src_stride_u - 3 * bytesperpixel,
525 refbw_m1 + 8, refbh_m1 + 8,
527 ref_u = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
528 smc(dst_u, dst_stride, ref_u, 288, bh, mx, my, step[0], step[1]);
530 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
531 ref_v - 3 * src_stride_v - 3 * bytesperpixel,
533 refbw_m1 + 8, refbh_m1 + 8,
535 ref_v = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
536 smc(dst_v, dst_stride, ref_v, 288, bh, mx, my, step[0], step[1]);
538 smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);
539 smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);
544 #define mc_luma_dir(td, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
545 px, py, pw, ph, bw, bh, w, h, i) \
546 mc_luma_scaled(td, s->dsp.s##mc, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
547 mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \
548 s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])
549 #define mc_chroma_dir(td, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
550 row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \
551 mc_chroma_scaled(td, s->dsp.s##mc, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
552 row, col, mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \
553 s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])
555 #define FN(x) x##_scaled_8bpp
556 #define BYTES_PER_PIXEL 1
557 #include "vp9_mc_template.c"
559 #undef BYTES_PER_PIXEL
560 #define FN(x) x##_scaled_16bpp
561 #define BYTES_PER_PIXEL 2
562 #include "vp9_mc_template.c"
566 #undef BYTES_PER_PIXEL
569 static av_always_inline void inter_recon(VP9TileData *td, int bytesperpixel)
571 VP9Context *s = td->s;
573 int row = td->row, col = td->col;
575 if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {
576 if (bytesperpixel == 1) {
577 inter_pred_scaled_8bpp(td);
579 inter_pred_scaled_16bpp(td);
582 if (bytesperpixel == 1) {
585 inter_pred_16bpp(td);
590 /* mostly copied intra_recon() */
592 int w4 = ff_vp9_bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;
593 int h4 = ff_vp9_bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);
594 int end_x = FFMIN(2 * (s->cols - col), w4);
595 int end_y = FFMIN(2 * (s->rows - row), h4);
596 int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;
597 int uvstep1d = 1 << b->uvtx, p;
598 uint8_t *dst = td->dst[0];
601 for (n = 0, y = 0; y < end_y; y += step1d) {
603 for (x = 0; x < end_x; x += step1d,
604 ptr += 4 * step1d * bytesperpixel, n += step) {
605 int eob = b->tx > TX_8X8 ? AV_RN16A(&td->eob[n]) : td->eob[n];
608 s->dsp.itxfm_add[tx][DCT_DCT](ptr, td->y_stride,
609 td->block + 16 * n * bytesperpixel, eob);
611 dst += 4 * td->y_stride * step1d;
617 step = 1 << (b->uvtx * 2);
618 for (p = 0; p < 2; p++) {
619 dst = td->dst[p + 1];
620 for (n = 0, y = 0; y < end_y; y += uvstep1d) {
622 for (x = 0; x < end_x; x += uvstep1d,
623 ptr += 4 * uvstep1d * bytesperpixel, n += step) {
624 int eob = b->uvtx > TX_8X8 ? AV_RN16A(&td->uveob[p][n]) : td->uveob[p][n];
627 s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, td->uv_stride,
628 td->uvblock[p] + 16 * n * bytesperpixel, eob);
630 dst += 4 * uvstep1d * td->uv_stride;
636 void ff_vp9_inter_recon_8bpp(VP9TileData *td)
641 void ff_vp9_inter_recon_16bpp(VP9TileData *td)