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"
25 #include "libavutil/mem_internal.h"
33 static av_always_inline int check_intra_mode(VP9TileData *td, int mode, uint8_t **a,
34 uint8_t *dst_edge, ptrdiff_t stride_edge,
35 uint8_t *dst_inner, ptrdiff_t stride_inner,
36 uint8_t *l, int col, int x, int w,
37 int row, int y, enum TxfmMode tx,
38 int p, int ss_h, int ss_v, int bytesperpixel)
40 VP9Context *s = td->s;
41 int have_top = row > 0 || y > 0;
42 int have_left = col > td->tile_col_start || x > 0;
43 int have_right = x < w - 1;
45 static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {
46 [VERT_PRED] = { { DC_127_PRED, VERT_PRED },
47 { DC_127_PRED, VERT_PRED } },
48 [HOR_PRED] = { { DC_129_PRED, DC_129_PRED },
49 { HOR_PRED, HOR_PRED } },
50 [DC_PRED] = { { DC_128_PRED, TOP_DC_PRED },
51 { LEFT_DC_PRED, DC_PRED } },
52 [DIAG_DOWN_LEFT_PRED] = { { DC_127_PRED, DIAG_DOWN_LEFT_PRED },
53 { DC_127_PRED, DIAG_DOWN_LEFT_PRED } },
54 [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },
55 { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },
56 [VERT_RIGHT_PRED] = { { VERT_RIGHT_PRED, VERT_RIGHT_PRED },
57 { VERT_RIGHT_PRED, VERT_RIGHT_PRED } },
58 [HOR_DOWN_PRED] = { { HOR_DOWN_PRED, HOR_DOWN_PRED },
59 { HOR_DOWN_PRED, HOR_DOWN_PRED } },
60 [VERT_LEFT_PRED] = { { DC_127_PRED, VERT_LEFT_PRED },
61 { DC_127_PRED, VERT_LEFT_PRED } },
62 [HOR_UP_PRED] = { { DC_129_PRED, DC_129_PRED },
63 { HOR_UP_PRED, HOR_UP_PRED } },
64 [TM_VP8_PRED] = { { DC_129_PRED, VERT_PRED },
65 { HOR_PRED, TM_VP8_PRED } },
70 uint8_t needs_topleft:1;
71 uint8_t needs_topright:1;
72 uint8_t invert_left:1;
73 } edges[N_INTRA_PRED_MODES] = {
74 [VERT_PRED] = { .needs_top = 1 },
75 [HOR_PRED] = { .needs_left = 1 },
76 [DC_PRED] = { .needs_top = 1, .needs_left = 1 },
77 [DIAG_DOWN_LEFT_PRED] = { .needs_top = 1, .needs_topright = 1 },
78 [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,
80 [VERT_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,
82 [HOR_DOWN_PRED] = { .needs_left = 1, .needs_top = 1,
84 [VERT_LEFT_PRED] = { .needs_top = 1, .needs_topright = 1 },
85 [HOR_UP_PRED] = { .needs_left = 1, .invert_left = 1 },
86 [TM_VP8_PRED] = { .needs_left = 1, .needs_top = 1,
88 [LEFT_DC_PRED] = { .needs_left = 1 },
89 [TOP_DC_PRED] = { .needs_top = 1 },
90 [DC_128_PRED] = { 0 },
91 [DC_127_PRED] = { 0 },
95 av_assert2(mode >= 0 && mode < 10);
96 mode = mode_conv[mode][have_left][have_top];
97 if (edges[mode].needs_top) {
98 uint8_t *top, *topleft;
99 int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !ss_h) - x) * 4;
100 int n_px_need_tr = 0;
102 if (tx == TX_4X4 && edges[mode].needs_topright && have_right)
105 // if top of sb64-row, use s->intra_pred_data[] instead of
106 // dst[-stride] for intra prediction (it contains pre- instead of
107 // post-loopfilter data)
109 top = !(row & 7) && !y ?
110 s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :
111 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];
113 topleft = !(row & 7) && !y ?
114 s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :
115 y == 0 || x == 0 ? &dst_edge[-stride_edge] :
116 &dst_inner[-stride_inner];
120 (!edges[mode].needs_topleft || (have_left && top == topleft)) &&
121 (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&
122 n_px_need + n_px_need_tr <= n_px_have) {
126 if (n_px_need <= n_px_have) {
127 memcpy(*a, top, n_px_need * bytesperpixel);
129 #define memset_bpp(c, i1, v, i2, num) do { \
130 if (bytesperpixel == 1) { \
131 memset(&(c)[(i1)], (v)[(i2)], (num)); \
133 int n, val = AV_RN16A(&(v)[(i2) * 2]); \
134 for (n = 0; n < (num); n++) { \
135 AV_WN16A(&(c)[((i1) + n) * 2], val); \
139 memcpy(*a, top, n_px_have * bytesperpixel);
140 memset_bpp(*a, n_px_have, (*a), n_px_have - 1, n_px_need - n_px_have);
143 #define memset_val(c, val, num) do { \
144 if (bytesperpixel == 1) { \
145 memset((c), (val), (num)); \
148 for (n = 0; n < (num); n++) { \
149 AV_WN16A(&(c)[n * 2], (val)); \
153 memset_val(*a, (128 << (bpp - 8)) - 1, n_px_need);
155 if (edges[mode].needs_topleft) {
156 if (have_left && have_top) {
157 #define assign_bpp(c, i1, v, i2) do { \
158 if (bytesperpixel == 1) { \
159 (c)[(i1)] = (v)[(i2)]; \
161 AV_COPY16(&(c)[(i1) * 2], &(v)[(i2) * 2]); \
164 assign_bpp(*a, -1, topleft, -1);
166 #define assign_val(c, i, v) do { \
167 if (bytesperpixel == 1) { \
170 AV_WN16A(&(c)[(i) * 2], (v)); \
173 assign_val((*a), -1, (128 << (bpp - 8)) + (have_top ? +1 : -1));
176 if (tx == TX_4X4 && edges[mode].needs_topright) {
177 if (have_top && have_right &&
178 n_px_need + n_px_need_tr <= n_px_have) {
179 memcpy(&(*a)[4 * bytesperpixel], &top[4 * bytesperpixel], 4 * bytesperpixel);
181 memset_bpp(*a, 4, *a, 3, 4);
186 if (edges[mode].needs_left) {
188 int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !ss_v) - y) * 4;
189 uint8_t *dst = x == 0 ? dst_edge : dst_inner;
190 ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;
192 if (edges[mode].invert_left) {
193 if (n_px_need <= n_px_have) {
194 for (i = 0; i < n_px_need; i++)
195 assign_bpp(l, i, &dst[i * stride], -1);
197 for (i = 0; i < n_px_have; i++)
198 assign_bpp(l, i, &dst[i * stride], -1);
199 memset_bpp(l, n_px_have, l, n_px_have - 1, n_px_need - n_px_have);
202 if (n_px_need <= n_px_have) {
203 for (i = 0; i < n_px_need; i++)
204 assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);
206 for (i = 0; i < n_px_have; i++)
207 assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);
208 memset_bpp(l, 0, l, n_px_need - n_px_have, n_px_need - n_px_have);
212 memset_val(l, (128 << (bpp - 8)) + 1, 4 << tx);
219 static av_always_inline void intra_recon(VP9TileData *td, ptrdiff_t y_off,
220 ptrdiff_t uv_off, int bytesperpixel)
222 VP9Context *s = td->s;
224 int row = td->row, col = td->col;
225 int w4 = ff_vp9_bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;
226 int h4 = ff_vp9_bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);
227 int end_x = FFMIN(2 * (s->cols - col), w4);
228 int end_y = FFMIN(2 * (s->rows - row), h4);
229 int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;
230 int uvstep1d = 1 << b->uvtx, p;
231 uint8_t *dst = td->dst[0], *dst_r = s->s.frames[CUR_FRAME].tf.f->data[0] + y_off;
232 LOCAL_ALIGNED_32(uint8_t, a_buf, [96]);
233 LOCAL_ALIGNED_32(uint8_t, l, [64]);
235 for (n = 0, y = 0; y < end_y; y += step1d) {
236 uint8_t *ptr = dst, *ptr_r = dst_r;
237 for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d * bytesperpixel,
238 ptr_r += 4 * step1d * bytesperpixel, n += step) {
239 int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?
241 uint8_t *a = &a_buf[32];
242 enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];
243 int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&td->eob[n]) : td->eob[n];
245 mode = check_intra_mode(td, mode, &a, ptr_r,
246 s->s.frames[CUR_FRAME].tf.f->linesize[0],
247 ptr, td->y_stride, l,
248 col, x, w4, row, y, b->tx, 0, 0, 0, bytesperpixel);
249 s->dsp.intra_pred[b->tx][mode](ptr, td->y_stride, l, a);
251 s->dsp.itxfm_add[tx][txtp](ptr, td->y_stride,
252 td->block + 16 * n * bytesperpixel, eob);
254 dst_r += 4 * step1d * s->s.frames[CUR_FRAME].tf.f->linesize[0];
255 dst += 4 * step1d * td->y_stride;
262 step = 1 << (b->uvtx * 2);
263 for (p = 0; p < 2; p++) {
264 dst = td->dst[1 + p];
265 dst_r = s->s.frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;
266 for (n = 0, y = 0; y < end_y; y += uvstep1d) {
267 uint8_t *ptr = dst, *ptr_r = dst_r;
268 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d * bytesperpixel,
269 ptr_r += 4 * uvstep1d * bytesperpixel, n += step) {
270 int mode = b->uvmode;
271 uint8_t *a = &a_buf[32];
272 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&td->uveob[p][n]) : td->uveob[p][n];
274 mode = check_intra_mode(td, mode, &a, ptr_r,
275 s->s.frames[CUR_FRAME].tf.f->linesize[1],
276 ptr, td->uv_stride, l, col, x, w4, row, y,
277 b->uvtx, p + 1, s->ss_h, s->ss_v, bytesperpixel);
278 s->dsp.intra_pred[b->uvtx][mode](ptr, td->uv_stride, l, a);
280 s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, td->uv_stride,
281 td->uvblock[p] + 16 * n * bytesperpixel, eob);
283 dst_r += 4 * uvstep1d * s->s.frames[CUR_FRAME].tf.f->linesize[1];
284 dst += 4 * uvstep1d * td->uv_stride;
289 void ff_vp9_intra_recon_8bpp(VP9TileData *td, ptrdiff_t y_off, ptrdiff_t uv_off)
291 intra_recon(td, y_off, uv_off, 1);
294 void ff_vp9_intra_recon_16bpp(VP9TileData *td, ptrdiff_t y_off, ptrdiff_t uv_off)
296 intra_recon(td, y_off, uv_off, 2);
299 static av_always_inline void mc_luma_unscaled(VP9TileData *td, vp9_mc_func (*mc)[2],
300 uint8_t *dst, ptrdiff_t dst_stride,
301 const uint8_t *ref, ptrdiff_t ref_stride,
302 ThreadFrame *ref_frame,
303 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,
304 int bw, int bh, int w, int h, int bytesperpixel)
306 VP9Context *s = td->s;
307 int mx = mv->x, my = mv->y, th;
311 ref += y * ref_stride + x * bytesperpixel;
314 // FIXME bilinear filter only needs 0/1 pixels, not 3/4
315 // we use +7 because the last 7 pixels of each sbrow can be changed in
316 // the longest loopfilter of the next sbrow
317 th = (y + bh + 4 * !!my + 7) >> 6;
318 ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
319 // The arm/aarch64 _hv filters read one more row than what actually is
320 // needed, so switch to emulated edge one pixel sooner vertically
321 // (!!my * 5) than horizontally (!!mx * 4).
322 if (x < !!mx * 3 || y < !!my * 3 ||
323 x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
324 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
325 ref - !!my * 3 * ref_stride - !!mx * 3 * bytesperpixel,
327 bw + !!mx * 7, bh + !!my * 7,
328 x - !!mx * 3, y - !!my * 3, w, h);
329 ref = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
332 mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);
335 static av_always_inline void mc_chroma_unscaled(VP9TileData *td, vp9_mc_func (*mc)[2],
336 uint8_t *dst_u, uint8_t *dst_v,
337 ptrdiff_t dst_stride,
338 const uint8_t *ref_u, ptrdiff_t src_stride_u,
339 const uint8_t *ref_v, ptrdiff_t src_stride_v,
340 ThreadFrame *ref_frame,
341 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,
342 int bw, int bh, int w, int h, int bytesperpixel)
344 VP9Context *s = td->s;
345 int mx = mv->x * (1 << !s->ss_h), my = mv->y * (1 << !s->ss_v), th;
349 ref_u += y * src_stride_u + x * bytesperpixel;
350 ref_v += y * src_stride_v + x * bytesperpixel;
353 // FIXME bilinear filter only needs 0/1 pixels, not 3/4
354 // we use +7 because the last 7 pixels of each sbrow can be changed in
355 // the longest loopfilter of the next sbrow
356 th = (y + bh + 4 * !!my + 7) >> (6 - s->ss_v);
357 ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
358 // The arm/aarch64 _hv filters read one more row than what actually is
359 // needed, so switch to emulated edge one pixel sooner vertically
360 // (!!my * 5) than horizontally (!!mx * 4).
361 if (x < !!mx * 3 || y < !!my * 3 ||
362 x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
363 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
364 ref_u - !!my * 3 * src_stride_u - !!mx * 3 * bytesperpixel,
366 bw + !!mx * 7, bh + !!my * 7,
367 x - !!mx * 3, y - !!my * 3, w, h);
368 ref_u = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
369 mc[!!mx][!!my](dst_u, dst_stride, ref_u, 160, bh, mx, my);
371 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
372 ref_v - !!my * 3 * src_stride_v - !!mx * 3 * bytesperpixel,
374 bw + !!mx * 7, bh + !!my * 7,
375 x - !!mx * 3, y - !!my * 3, w, h);
376 ref_v = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
377 mc[!!mx][!!my](dst_v, dst_stride, ref_v, 160, bh, mx, my);
379 mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);
380 mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);
384 #define mc_luma_dir(td, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
385 px, py, pw, ph, bw, bh, w, h, i) \
386 mc_luma_unscaled(td, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
387 mv, bw, bh, w, h, bytesperpixel)
388 #define mc_chroma_dir(td, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
389 row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \
390 mc_chroma_unscaled(td, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
391 row, col, mv, bw, bh, w, h, bytesperpixel)
393 #define FN(x) x##_8bpp
394 #define BYTES_PER_PIXEL 1
395 #include "vp9_mc_template.c"
397 #undef BYTES_PER_PIXEL
398 #define FN(x) x##_16bpp
399 #define BYTES_PER_PIXEL 2
400 #include "vp9_mc_template.c"
404 #undef BYTES_PER_PIXEL
407 static av_always_inline void mc_luma_scaled(VP9TileData *td, vp9_scaled_mc_func smc,
408 vp9_mc_func (*mc)[2],
409 uint8_t *dst, ptrdiff_t dst_stride,
410 const uint8_t *ref, ptrdiff_t ref_stride,
411 ThreadFrame *ref_frame,
412 ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,
413 int px, int py, int pw, int ph,
414 int bw, int bh, int w, int h, int bytesperpixel,
415 const uint16_t *scale, const uint8_t *step)
417 VP9Context *s = td->s;
418 if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&
419 s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {
420 mc_luma_unscaled(td, mc, dst, dst_stride, ref, ref_stride, ref_frame,
421 y, x, in_mv, bw, bh, w, h, bytesperpixel);
423 #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)
425 int refbw_m1, refbh_m1;
429 mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);
430 mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);
431 // BUG libvpx seems to scale the two components separately. This introduces
432 // rounding errors but we have to reproduce them to be exactly compatible
433 // with the output from libvpx...
434 mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);
435 my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);
439 ref += y * ref_stride + x * bytesperpixel;
442 refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;
443 refbh_m1 = ((bh - 1) * step[1] + my) >> 4;
444 // FIXME bilinear filter only needs 0/1 pixels, not 3/4
445 // we use +7 because the last 7 pixels of each sbrow can be changed in
446 // the longest loopfilter of the next sbrow
447 th = (y + refbh_m1 + 4 + 7) >> 6;
448 ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
449 // The arm/aarch64 _hv filters read one more row than what actually is
450 // needed, so switch to emulated edge one pixel sooner vertically
451 // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).
452 if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {
453 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
454 ref - 3 * ref_stride - 3 * bytesperpixel,
456 refbw_m1 + 8, refbh_m1 + 8,
458 ref = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
461 smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);
465 static av_always_inline void mc_chroma_scaled(VP9TileData *td, vp9_scaled_mc_func smc,
466 vp9_mc_func (*mc)[2],
467 uint8_t *dst_u, uint8_t *dst_v,
468 ptrdiff_t dst_stride,
469 const uint8_t *ref_u, ptrdiff_t src_stride_u,
470 const uint8_t *ref_v, ptrdiff_t src_stride_v,
471 ThreadFrame *ref_frame,
472 ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,
473 int px, int py, int pw, int ph,
474 int bw, int bh, int w, int h, int bytesperpixel,
475 const uint16_t *scale, const uint8_t *step)
477 VP9Context *s = td->s;
478 if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&
479 s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {
480 mc_chroma_unscaled(td, mc, dst_u, dst_v, dst_stride, ref_u, src_stride_u,
481 ref_v, src_stride_v, ref_frame,
482 y, x, in_mv, bw, bh, w, h, bytesperpixel);
485 int refbw_m1, refbh_m1;
490 // BUG https://code.google.com/p/webm/issues/detail?id=820
491 mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 16, (s->cols * 4 - x + px + 3) * 16);
492 mx = scale_mv(mv.x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);
494 mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);
495 mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);
498 // BUG https://code.google.com/p/webm/issues/detail?id=820
499 mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 16, (s->rows * 4 - y + py + 3) * 16);
500 my = scale_mv(mv.y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);
502 mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);
503 my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);
508 ref_u += y * src_stride_u + x * bytesperpixel;
509 ref_v += y * src_stride_v + x * bytesperpixel;
512 refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;
513 refbh_m1 = ((bh - 1) * step[1] + my) >> 4;
514 // FIXME bilinear filter only needs 0/1 pixels, not 3/4
515 // we use +7 because the last 7 pixels of each sbrow can be changed in
516 // the longest loopfilter of the next sbrow
517 th = (y + refbh_m1 + 4 + 7) >> (6 - s->ss_v);
518 ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);
519 // The arm/aarch64 _hv filters read one more row than what actually is
520 // needed, so switch to emulated edge one pixel sooner vertically
521 // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).
522 if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {
523 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
524 ref_u - 3 * src_stride_u - 3 * bytesperpixel,
526 refbw_m1 + 8, refbh_m1 + 8,
528 ref_u = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
529 smc(dst_u, dst_stride, ref_u, 288, bh, mx, my, step[0], step[1]);
531 s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
532 ref_v - 3 * src_stride_v - 3 * bytesperpixel,
534 refbw_m1 + 8, refbh_m1 + 8,
536 ref_v = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
537 smc(dst_v, dst_stride, ref_v, 288, bh, mx, my, step[0], step[1]);
539 smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);
540 smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);
545 #define mc_luma_dir(td, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
546 px, py, pw, ph, bw, bh, w, h, i) \
547 mc_luma_scaled(td, s->dsp.s##mc, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
548 mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \
549 s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])
550 #define mc_chroma_dir(td, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
551 row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \
552 mc_chroma_scaled(td, s->dsp.s##mc, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
553 row, col, mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \
554 s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])
556 #define FN(x) x##_scaled_8bpp
557 #define BYTES_PER_PIXEL 1
558 #include "vp9_mc_template.c"
560 #undef BYTES_PER_PIXEL
561 #define FN(x) x##_scaled_16bpp
562 #define BYTES_PER_PIXEL 2
563 #include "vp9_mc_template.c"
567 #undef BYTES_PER_PIXEL
570 static av_always_inline void inter_recon(VP9TileData *td, int bytesperpixel)
572 VP9Context *s = td->s;
574 int row = td->row, col = td->col;
576 if (s->mvscale[b->ref[0]][0] == REF_INVALID_SCALE ||
577 (b->comp && s->mvscale[b->ref[1]][0] == REF_INVALID_SCALE)) {
578 if (!s->td->error_info) {
579 s->td->error_info = AVERROR_INVALIDDATA;
580 av_log(NULL, AV_LOG_ERROR, "Bitstream not supported, "
581 "reference frame has invalid dimensions\n");
586 if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {
587 if (bytesperpixel == 1) {
588 inter_pred_scaled_8bpp(td);
590 inter_pred_scaled_16bpp(td);
593 if (bytesperpixel == 1) {
596 inter_pred_16bpp(td);
601 /* mostly copied intra_recon() */
603 int w4 = ff_vp9_bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;
604 int h4 = ff_vp9_bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);
605 int end_x = FFMIN(2 * (s->cols - col), w4);
606 int end_y = FFMIN(2 * (s->rows - row), h4);
607 int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;
608 int uvstep1d = 1 << b->uvtx, p;
609 uint8_t *dst = td->dst[0];
612 for (n = 0, y = 0; y < end_y; y += step1d) {
614 for (x = 0; x < end_x; x += step1d,
615 ptr += 4 * step1d * bytesperpixel, n += step) {
616 int eob = b->tx > TX_8X8 ? AV_RN16A(&td->eob[n]) : td->eob[n];
619 s->dsp.itxfm_add[tx][DCT_DCT](ptr, td->y_stride,
620 td->block + 16 * n * bytesperpixel, eob);
622 dst += 4 * td->y_stride * step1d;
628 step = 1 << (b->uvtx * 2);
629 for (p = 0; p < 2; p++) {
630 dst = td->dst[p + 1];
631 for (n = 0, y = 0; y < end_y; y += uvstep1d) {
633 for (x = 0; x < end_x; x += uvstep1d,
634 ptr += 4 * uvstep1d * bytesperpixel, n += step) {
635 int eob = b->uvtx > TX_8X8 ? AV_RN16A(&td->uveob[p][n]) : td->uveob[p][n];
638 s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, td->uv_stride,
639 td->uvblock[p] + 16 * n * bytesperpixel, eob);
641 dst += 4 * uvstep1d * td->uv_stride;
647 void ff_vp9_inter_recon_8bpp(VP9TileData *td)
652 void ff_vp9_inter_recon_16bpp(VP9TileData *td)