recon_tmpl.c 73.8 KB
Newer Older
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
/*
 * Copyright © 2018, VideoLAN and dav1d authors
 * Copyright © 2018, Two Orioles, LLC
 * 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 "config.h"

#include <string.h>
#include <stdio.h>

#include "common/attributes.h"
#include "common/bitdepth.h"
#include "common/dump.h"
#include "common/intops.h"
#include "common/mem.h"

#include "src/cdef_apply.h"
Ronald S. Bultje's avatar
Ronald S. Bultje committed
40
#include "src/ctx.h"
41 42 43 44 45 46 47 48 49 50 51 52
#include "src/ipred_prepare.h"
#include "src/lf_apply.h"
#include "src/lr_apply.h"
#include "src/recon.h"
#include "src/scan.h"
#include "src/tables.h"
#include "src/wedge.h"

static unsigned read_golomb(MsacContext *const msac) {
    int len = 0;
    unsigned val = 1;

Rostislav Pehlivanov's avatar
Rostislav Pehlivanov committed
53 54
    while (!msac_decode_bool(msac, EC_BOOL_EPROB) && len < 32) len++;
    while (len--) val = (val << 1) | msac_decode_bool(msac, EC_BOOL_EPROB);
55 56 57 58 59 60 61 62 63 64 65 66 67 68

    return val - 1;
}

static int decode_coefs(Dav1dTileContext *const t,
                        uint8_t *const a, uint8_t *const l,
                        const enum RectTxfmSize tx, const enum BlockSize bs,
                        const Av1Block *const b, const int intra,
                        const int plane, coef *cf,
                        enum TxfmType *const txtp, uint8_t *res_ctx)
{
    Dav1dTileState *const ts = t->ts;
    const int chroma = !!plane;
    const Dav1dFrameContext *const f = t->f;
69
    const TxfmInfo *const t_dim = &dav1d_txfm_dimensions[tx];
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
    const int dbg = DEBUG_BLOCK_INFO && plane && 0;

    if (dbg) printf("Start: r=%d\n", ts->msac.rng);

    // does this block have any non-zero coefficients
    const int sctx = get_coef_skip_ctx(t_dim, bs, a, l, chroma, f->cur.p.p.layout);
    const int all_skip =
        msac_decode_bool_adapt(&ts->msac, ts->cdf.coef.skip[t_dim->ctx][sctx]);
    if (dbg)
    printf("Post-non-zero[%d][%d][%d]: r=%d\n",
           t_dim->ctx, sctx, all_skip, ts->msac.rng);
    if (all_skip) {
        *res_ctx = 0x40;
        *txtp = f->frame_hdr.segmentation.lossless[b->seg_id] ? WHT_WHT :
                                                                DCT_DCT;
        return -1;
    }

    // transform type (chroma: derived, luma: explicitly coded)
    if (chroma) {
        if (intra) {
            *txtp = get_uv_intra_txtp(b->uv_mode, tx, &f->frame_hdr, b->seg_id);
        } else {
            const enum TxfmType y_txtp = *txtp;
            *txtp = get_uv_inter_txtp(t_dim, y_txtp, &f->frame_hdr, b->seg_id);
        }
    } else {
        const enum TxfmTypeSet set = get_ext_txtp_set(tx, !intra,
                                                      &f->frame_hdr, b->seg_id);
99
        const unsigned set_cnt = dav1d_tx_type_count[set];
100 101 102 103
        unsigned idx;
        if (set_cnt == 1) {
            idx = 0;
        } else {
104
            const int set_idx = dav1d_tx_type_set_index[!intra][set];
105 106
            const enum IntraPredMode y_mode_nofilt = intra ? b->y_mode == FILTER_PRED ?
                dav1d_filter_mode_to_y_mode[b->y_angle] : b->y_mode : 0;
107 108 109 110 111 112
            uint16_t *const txtp_cdf = intra ?
                       ts->cdf.m.txtp_intra[set_idx][t_dim->min][y_mode_nofilt] :
                       ts->cdf.m.txtp_inter[set_idx][t_dim->min];
            idx = msac_decode_symbol_adapt(&ts->msac, txtp_cdf, set_cnt);
            if (dbg)
            printf("Post-txtp[%d->%d][%d->%d][%d][%d->%d]: r=%d\n",
113
                   set, set_idx, tx, t_dim->min, intra ? (int)y_mode_nofilt : -1,
114
                   idx, dav1d_tx_types_per_set[set][idx], ts->msac.rng);
115
        }
116
        *txtp = dav1d_tx_types_per_set[set][idx];
117 118 119 120 121
    }

    // find end-of-block (eob)
    int eob_bin;
    const int tx2dszctx = imin(t_dim->lw, TX_32X32) + imin(t_dim->lh, TX_32X32);
122
    const enum TxClass tx_class = dav1d_tx_type_class[*txtp];
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
    const int is_1d = tx_class != TX_CLASS_2D;
    switch (tx2dszctx) {
#define case_sz(sz, bin) \
    case sz: { \
        uint16_t *const eob_bin_cdf = ts->cdf.coef.eob_bin_##bin[chroma][is_1d]; \
        eob_bin = msac_decode_symbol_adapt(&ts->msac, eob_bin_cdf, 5 + sz); \
        break; \
    }
    case_sz(0,   16);
    case_sz(1,   32);
    case_sz(2,   64);
    case_sz(3,  128);
    case_sz(4,  256);
    case_sz(5,  512);
    case_sz(6, 1024);
#undef case_sz
    }
    if (dbg)
    printf("Post-eob_bin_%d[%d][%d][%d]: r=%d\n",
           16 << tx2dszctx, chroma, is_1d, eob_bin, ts->msac.rng);
    int eob;
    if (eob_bin > 1) {
        eob = 1 << (eob_bin - 1);
        uint16_t *const eob_hi_bit_cdf =
            ts->cdf.coef.eob_hi_bit[t_dim->ctx][chroma][eob_bin];
        const int eob_hi_bit = msac_decode_bool_adapt(&ts->msac, eob_hi_bit_cdf);
        if (dbg)
        printf("Post-eob_hi_bit[%d][%d][%d][%d]: r=%d\n",
               t_dim->ctx, chroma, eob_bin, eob_hi_bit, ts->msac.rng);
        unsigned mask = eob >> 1;
        if (eob_hi_bit) eob |= mask;
        for (mask >>= 1; mask; mask >>= 1) {
Rostislav Pehlivanov's avatar
Rostislav Pehlivanov committed
155
            const int eob_bit = msac_decode_bool(&ts->msac, EC_BOOL_EPROB);
156 157 158 159 160 161 162 163 164 165 166
            if (eob_bit) eob |= mask;
        }
        if (dbg)
        printf("Post-eob[%d]: r=%d\n", eob, ts->msac.rng);
    } else {
        eob = eob_bin;
    }

    // base tokens
    uint16_t (*const br_cdf)[5] =
        ts->cdf.coef.br_tok[imin(t_dim->ctx, 3)][chroma];
167
    const int16_t *const scan = dav1d_scans[tx][tx_class];
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
    uint8_t levels[36 * 36];
    ptrdiff_t stride = 4 * (imin(t_dim->h, 8) + 1);
    memset(levels, 0, stride * 4 * (imin(t_dim->w, 8) + 1));
    const int shift = 2 + imin(t_dim->lh, 3), mask = 4 * imin(t_dim->h, 8) - 1;
    unsigned cul_level = 0;
    for (int i = eob, is_last = 1; i >= 0; i--, is_last = 0) {
        const int rc = scan[i], x = rc >> shift, y = rc & mask;

        // lo tok
        const int ctx = get_coef_nz_ctx(levels, i, rc, is_last, tx, tx_class);
        uint16_t *const lo_cdf = is_last ?
            ts->cdf.coef.eob_base_tok[t_dim->ctx][chroma][ctx] :
            ts->cdf.coef.base_tok[t_dim->ctx][chroma][ctx];
        int tok = msac_decode_symbol_adapt(&ts->msac, lo_cdf,
                                           4 - is_last) + is_last;
        if (dbg)
        printf("Post-lo_tok[%d][%d][%d][%d=%d=%d]: r=%d\n",
               t_dim->ctx, chroma, ctx, i, rc, tok, ts->msac.rng);
        if (!tok) continue;

        // hi tok
        if (tok == 3) {
            const int br_ctx = get_br_ctx(levels, rc, tx, tx_class);
            do {
                const int tok_br =
                    msac_decode_symbol_adapt(&ts->msac, br_cdf[br_ctx], 4);
                if (dbg)
                printf("Post-hi_tok[%d][%d][%d][%d=%d=%d->%d]: r=%d\n",
                       imin(t_dim->ctx, 3), chroma, br_ctx,
                       i, rc, tok_br, tok, ts->msac.rng);
                tok += tok_br;
                if (tok_br < 3) break;
            } while (tok < 15);
        }

        levels[x * stride + y] = cf[rc] = tok;
    }

    // residual and sign
    int dc_sign = 1;
    const uint16_t *const dq_tbl = ts->dq[b->seg_id][plane];
    const uint8_t *const qm_tbl = f->qm[is_1d || *txtp == IDTX][tx][plane];
    const int dq_shift = imax(0, t_dim->ctx - 2);
    for (int i = 0; i <= eob; i++) {
        const int rc = scan[i];
        int tok = cf[rc];
        if (!tok) continue;
        int dq;

        // sign
        int sign;
        if (i == 0) {
            const int dc_sign_ctx = get_dc_sign_ctx(t_dim, a, l);
            uint16_t *const dc_sign_cdf =
                ts->cdf.coef.dc_sign[chroma][dc_sign_ctx];
            sign = msac_decode_bool_adapt(&ts->msac, dc_sign_cdf);
            if (dbg)
            printf("Post-dc_sign[%d][%d][%d]: r=%d\n",
                   chroma, dc_sign_ctx, sign, ts->msac.rng);
            dc_sign = sign ? 0 : 2;
            dq = (dq_tbl[0] * qm_tbl[0] + 16) >> 5;
        } else {
Rostislav Pehlivanov's avatar
Rostislav Pehlivanov committed
230
            sign = msac_decode_bool(&ts->msac, EC_BOOL_EPROB);
231 232 233 234 235 236 237 238 239 240 241 242 243
            if (dbg)
            printf("Post-sign[%d=%d=%d]: r=%d\n", i, rc, sign, ts->msac.rng);
            dq = (dq_tbl[1] * qm_tbl[rc] + 16) >> 5;
        }

        // residual
        if (tok == 15) {
            tok += read_golomb(&ts->msac);
            if (dbg)
            printf("Post-residual[%d=%d=%d->%d]: r=%d\n",
                   i, rc, tok - 15, tok, ts->msac.rng);
        }

244 245 246 247
        // coefficient parsing, see 5.11.39
        tok &= 0xfffff;

        // dequant, see 7.12.3
248
        cul_level += tok;
249 250 251 252
        tok = (((int64_t)dq * tok) & 0xffffff) >> dq_shift;
        cf[rc] = iclip(sign ? -tok : tok,
                       -(1 << (7 + BITDEPTH)),
                       (1 << (7 + BITDEPTH)) - 1);
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269
    }

    // context
    *res_ctx = imin(cul_level, 63) | (dc_sign << 6);

    return eob;
}

static void read_coef_tree(Dav1dTileContext *const t,
                           const enum BlockSize bs, const Av1Block *const b,
                           const enum RectTxfmSize ytx, const int depth,
                           const uint16_t *const tx_split,
                           const int x_off, const int y_off, pixel *dst)
{
    const Dav1dFrameContext *const f = t->f;
    Dav1dTileState *const ts = t->ts;
    const Dav1dDSPContext *const dsp = f->dsp;
270
    const TxfmInfo *const t_dim = &dav1d_txfm_dimensions[ytx];
271 272
    const int txw = t_dim->w, txh = t_dim->h;

273 274 275 276 277
    /* y_off can be larger than 3 since lossless blocks use TX_4X4 but can't
     * be splitted. Aviods an undefined left shift. */
    if (depth < 2 && tx_split[depth] &&
        tx_split[depth] & (1 << (y_off * 4 + x_off)))
    {
278
        const enum RectTxfmSize sub = t_dim->sub;
279
        const TxfmInfo *const sub_t_dim = &dav1d_txfm_dimensions[sub];
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
        const int txsw = sub_t_dim->w, txsh = sub_t_dim->h;

        read_coef_tree(t, bs, b, sub, depth + 1, tx_split,
                       x_off * 2 + 0, y_off * 2 + 0, dst);
        t->bx += txsw;
        if (txw >= txh && t->bx < f->bw)
            read_coef_tree(t, bs, b, sub, depth + 1, tx_split, x_off * 2 + 1,
                           y_off * 2 + 0, dst ? &dst[4 * txsw] : NULL);
        t->bx -= txsw;
        t->by += txsh;
        if (txh >= txw && t->by < f->bh) {
            if (dst)
                dst += 4 * txsh * PXSTRIDE(f->cur.p.stride[0]);
            read_coef_tree(t, bs, b, sub, depth + 1, tx_split,
                           x_off * 2 + 0, y_off * 2 + 1, dst);
            t->bx += txsw;
            if (txw >= txh && t->bx < f->bw)
                read_coef_tree(t, bs, b, sub, depth + 1, tx_split, x_off * 2 + 1,
                               y_off * 2 + 1, dst ? &dst[4 * txsw] : NULL);
            t->bx -= txsw;
        }
        t->by -= txsh;
    } else {
        const int bx4 = t->bx & 31, by4 = t->by & 31;
        enum TxfmType txtp;
        uint8_t cf_ctx;
        int eob;
        coef *cf;
        struct CodedBlockInfo *cbi;

        if (f->frame_thread.pass) {
            cf = ts->frame_thread.cf;
            ts->frame_thread.cf += imin(t_dim->w, 8) * imin(t_dim->h, 8) * 16;
            cbi = &f->frame_thread.cbi[t->by * f->b4_stride + t->bx];
        } else {
            cf = t->cf;
        }
        if (f->frame_thread.pass != 2) {
            eob = decode_coefs(t, &t->a->lcoef[bx4], &t->l.lcoef[by4],
                               ytx, bs, b, 0, 0, cf, &txtp, &cf_ctx);
            if (DEBUG_BLOCK_INFO)
                printf("Post-y-cf-blk[tx=%d,txtp=%d,eob=%d]: r=%d\n",
                       ytx, txtp, eob, ts->msac.rng);
Ronald S. Bultje's avatar
Ronald S. Bultje committed
323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
            rep_macro(type, t->dir lcoef, off, mul * cf_ctx)
#define default_memset(dir, diridx, off, sz) \
            memset(&t->dir lcoef[off], cf_ctx, sz)
            case_set_upto16_with_default(imin(txh, f->bh - t->by), l., 1, by4);
            case_set_upto16_with_default(imin(txw, f->bw - t->bx), a->, 0, bx4);
#undef default_memset
#undef set_ctx
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
            for (int y = 0; y < txh; y++) { \
                rep_macro(type, txtp_map, 0, mul * txtp); \
                txtp_map += 32; \
            }
            uint8_t *txtp_map = &t->txtp_map[by4 * 32 + bx4];
            case_set_upto16(txw,,,);
#undef set_ctx
339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359
            if (f->frame_thread.pass == 1) {
                cbi->eob[0] = eob;
                cbi->txtp[0] = txtp;
            }
        } else {
            eob = cbi->eob[0];
            txtp = cbi->txtp[0];
        }
        if (!(f->frame_thread.pass & 1)) {
            assert(dst);
            if (eob >= 0) {
                if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS)
                    coef_dump(cf, imin(t_dim->h, 8) * 4, imin(t_dim->w, 8) * 4, 3, "dq");
                dsp->itx.itxfm_add[ytx][txtp](dst, f->cur.p.stride[0], cf, eob);
                if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS)
                    hex_dump(dst, f->cur.p.stride[0], t_dim->w * 4, t_dim->h * 4, "recon");
            }
        }
    }
}

360 361
void bytefn(dav1d_read_coef_blocks)(Dav1dTileContext *const t,
                                    const enum BlockSize bs, const Av1Block *const b)
362 363 364 365 366 367
{
    const Dav1dFrameContext *const f = t->f;
    const int ss_ver = f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I420;
    const int ss_hor = f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I444;
    const int bx4 = t->bx & 31, by4 = t->by & 31;
    const int cbx4 = bx4 >> ss_hor, cby4 = by4 >> ss_ver;
368
    const uint8_t *const b_dim = dav1d_block_dimensions[bs];
369 370 371 372 373 374 375
    const int bw4 = b_dim[0], bh4 = b_dim[1];
    const int cbw4 = (bw4 + 1) >> ss_hor, cbh4 = (bh4 + 1) >> ss_ver;
    const int has_chroma = f->seq_hdr.layout != DAV1D_PIXEL_LAYOUT_I400 &&
                           (bw4 > ss_hor || t->bx & 1) &&
                           (bh4 > ss_ver || t->by & 1);

    if (b->skip) {
Ronald S. Bultje's avatar
Ronald S. Bultje committed
376 377 378 379 380 381 382 383 384 385 386 387
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
        rep_macro(type, t->dir lcoef, off, mul * 0x40)
        case_set(bh4, l., 1, by4);
        case_set(bw4, a->, 0, bx4);
#undef set_ctx
        if (has_chroma) {
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
            rep_macro(type, t->dir ccoef[0], off, mul * 0x40); \
            rep_macro(type, t->dir ccoef[1], off, mul * 0x40)
            case_set(cbh4, l., 1, cby4);
            case_set(cbw4, a->, 0, cbx4);
#undef set_ctx
388 389 390 391 392 393 394 395 396
        }
        return;
    }

    Dav1dTileState *const ts = t->ts;
    const int w4 = imin(bw4, f->bw - t->bx), h4 = imin(bh4, f->bh - t->by);
    const int cw4 = (w4 + ss_hor) >> ss_hor, ch4 = (h4 + ss_ver) >> ss_ver;
    assert(f->frame_thread.pass == 1);
    assert(!b->skip);
397 398
    const TxfmInfo *const uv_t_dim = &dav1d_txfm_dimensions[b->uvtx];
    const TxfmInfo *const t_dim = &dav1d_txfm_dimensions[b->intra ? b->tx : b->max_ytx];
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

    for (int init_y = 0; init_y < h4; init_y += 16) {
        for (int init_x = 0; init_x < w4; init_x += 16) {
            const int sub_h4 = imin(h4, 16 + init_y);
            const int sub_w4 = imin(w4, init_x + 16);
            int y_off = !!init_y, y, x;
            for (y = init_y, t->by += init_y; y < sub_h4;
                 y += t_dim->h, t->by += t_dim->h, y_off++)
            {
                struct CodedBlockInfo *const cbi =
                    &f->frame_thread.cbi[t->by * f->b4_stride];
                int x_off = !!init_x;
                for (x = init_x, t->bx += init_x; x < sub_w4;
                     x += t_dim->w, t->bx += t_dim->w, x_off++)
                {
                    if (!b->intra) {
                        read_coef_tree(t, bs, b, b->max_ytx, 0, b->tx_split,
                                       x_off, y_off, NULL);
                    } else {
                        uint8_t cf_ctx = 0x40;
                        enum TxfmType txtp;
                        const int eob = cbi[t->bx].eob[0] =
                            decode_coefs(t, &t->a->lcoef[bx4 + x],
                                         &t->l.lcoef[by4 + y], b->tx, bs, b, 1,
                                         0, ts->frame_thread.cf, &txtp, &cf_ctx);
                        if (DEBUG_BLOCK_INFO)
                            printf("Post-y-cf-blk[tx=%d,txtp=%d,eob=%d]: r=%d\n",
                                   b->tx, txtp, eob, ts->msac.rng);
                        cbi[t->bx].txtp[0] = txtp;
                        ts->frame_thread.cf += imin(t_dim->w, 8) * imin(t_dim->h, 8) * 16;
Ronald S. Bultje's avatar
Ronald S. Bultje committed
429 430 431 432 433 434 435 436 437 438
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
                        rep_macro(type, t->dir lcoef, off, mul * cf_ctx)
#define default_memset(dir, diridx, off, sz) \
                        memset(&t->dir lcoef[off], cf_ctx, sz)
                        case_set_upto16_with_default(imin(t_dim->h, f->bh - t->by),
                                                     l., 1, by4 + y);
                        case_set_upto16_with_default(imin(t_dim->w, f->bw - t->bx),
                                                     a->, 0, bx4 + x);
#undef default_memset
#undef set_ctx
439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473
                    }
                }
                t->bx -= x;
            }
            t->by -= y;

            if (!has_chroma) continue;

            const int sub_ch4 = imin(ch4, (init_y + 16) >> ss_ver);
            const int sub_cw4 = imin(cw4, (init_x + 16) >> ss_hor);
            for (int pl = 0; pl < 2; pl++) {
                for (y = init_y >> ss_ver, t->by += init_y; y < sub_ch4;
                     y += uv_t_dim->h, t->by += uv_t_dim->h << ss_ver)
                {
                    struct CodedBlockInfo *const cbi =
                        &f->frame_thread.cbi[t->by * f->b4_stride];
                    for (x = init_x >> ss_hor, t->bx += init_x; x < sub_cw4;
                         x += uv_t_dim->w, t->bx += uv_t_dim->w << ss_hor)
                    {
                        uint8_t cf_ctx = 0x40;
                        enum TxfmType txtp;
                        if (!b->intra)
                            txtp = t->txtp_map[(by4 + (y << ss_ver)) * 32 +
                                                bx4 + (x << ss_hor)];
                        const int eob = cbi[t->bx].eob[1 + pl] =
                            decode_coefs(t, &t->a->ccoef[pl][cbx4 + x],
                                         &t->l.ccoef[pl][cby4 + y], b->uvtx, bs,
                                         b, b->intra, 1 + pl, ts->frame_thread.cf,
                                         &txtp, &cf_ctx);
                        if (DEBUG_BLOCK_INFO)
                            printf("Post-uv-cf-blk[pl=%d,tx=%d,"
                                   "txtp=%d,eob=%d]: r=%d\n",
                                   pl, b->uvtx, txtp, eob, ts->msac.rng);
                        cbi[t->bx].txtp[1 + pl] = txtp;
                        ts->frame_thread.cf += uv_t_dim->w * uv_t_dim->h * 16;
Ronald S. Bultje's avatar
Ronald S. Bultje committed
474 475 476 477 478 479 480 481 482 483 484 485
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
                        rep_macro(type, t->dir ccoef[pl], off, mul * cf_ctx)
#define default_memset(dir, diridx, off, sz) \
                        memset(&t->dir ccoef[pl][off], cf_ctx, sz)
                        case_set_upto16_with_default( \
                                 imin(uv_t_dim->h, (f->bh - t->by + ss_ver) >> ss_ver),
                                 l., 1, cby4 + y);
                        case_set_upto16_with_default( \
                                 imin(uv_t_dim->w, (f->bw - t->bx + ss_hor) >> ss_hor),
                                 a->, 0, cbx4 + x);
#undef default_memset
#undef set_ctx
486 487 488 489 490 491 492 493 494
                    }
                    t->bx -= x << ss_hor;
                }
                t->by -= y << ss_ver;
            }
        }
    }
}

495 496 497 498 499 500
static int mc(Dav1dTileContext *const t,
              pixel *const dst8, coef *const dst16, const ptrdiff_t dst_stride,
              const int bw4, const int bh4,
              const int bx, const int by, const int pl,
              const mv mv, const Dav1dThreadPicture *const refp,
              const enum Filter2d filter_2d)
501 502 503 504 505 506 507 508 509 510 511 512
{
    assert((dst8 != NULL) ^ (dst16 != NULL));
    const Dav1dFrameContext *const f = t->f;
    const int ss_ver = !!pl && f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I420;
    const int ss_hor = !!pl && f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I444;
    const int h_mul = 4 >> ss_hor, v_mul = 4 >> ss_ver;
    const int mvx = mv.x, mvy = mv.y;
    const int mx = mvx & (15 >> !ss_hor), my = mvy & (15 >> !ss_ver);
    const int dx = bx * h_mul + (mvx >> (3 + ss_hor));
    const int dy = by * v_mul + (mvy >> (3 + ss_ver));
    ptrdiff_t ref_stride = refp->p.stride[!!pl];
    const pixel *ref;
513
    int w, h;
514

515
    if (refp != &f->cur) { // i.e. not for intrabc
516 517 518 519 520
        if (dav1d_thread_picture_wait(refp, dy + bh4 * v_mul + !!my * 4,
                                      PLANE_TYPE_Y + !!pl))
        {
            return -1;
        }
521 522 523 524 525 526
        w = (f->cur.p.p.w + ss_hor) >> ss_hor;
        h = (f->cur.p.p.h + ss_ver) >> ss_ver;
    } else {
        w = f->bw * 4 >> ss_hor;
        h = f->bh * 4 >> ss_ver;
    }
527
    if (dx < !!mx * 3 || dy < !!my * 3 ||
528 529
        dx + bw4 * h_mul + !!mx * 4 > w ||
        dy + bh4 * v_mul + !!my * 4 > h)
530
    {
531 532
        f->dsp->mc.emu_edge(bw4 * h_mul + !!mx * 7, bh4 * v_mul + !!my * 7,
                            w, h, dx - !!mx * 3, dy - !!my * 3,
533
                            t->emu_edge, 192 * sizeof(pixel),
534
                            refp->p.data[pl], ref_stride);
535 536
        ref = &t->emu_edge[192 * !!my * 3 + !!mx * 3];
        ref_stride = 192 * sizeof(pixel);
537 538 539 540 541 542 543 544 545 546 547
    } else {
        ref = ((pixel *) refp->p.data[pl]) + PXSTRIDE(ref_stride) * dy + dx;
    }

    if (dst8 != NULL) {
        f->dsp->mc.mc[filter_2d](dst8, dst_stride, ref, ref_stride, bw4 * h_mul,
                                 bh4 * v_mul, mx << !ss_hor, my << !ss_ver);
    } else {
        f->dsp->mc.mct[filter_2d](dst16, ref, ref_stride, bw4 * h_mul,
                                  bh4 * v_mul, mx << !ss_hor, my << !ss_ver);
    }
548
    return 0;
549 550
}

551 552 553 554
static int obmc(Dav1dTileContext *const t,
                pixel *const dst, const ptrdiff_t dst_stride,
                const uint8_t *const b_dim, const int pl,
                const int bx4, const int by4, const int w4, const int h4)
555 556 557 558 559 560 561 562
{
    assert(!(t->bx & 1) && !(t->by & 1));
    const Dav1dFrameContext *const f = t->f;
    const refmvs *const r = &f->mvs[t->by * f->b4_stride + t->bx];
    pixel *const lap = t->scratch.lap;
    const int ss_ver = !!pl && f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I420;
    const int ss_hor = !!pl && f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I444;
    const int h_mul = 4 >> ss_hor, v_mul = 4 >> ss_ver;
563
    int res;
564 565 566 567 568 569 570 571

    if (t->by > t->ts->tiling.row_start &&
        (!pl || b_dim[0] * h_mul + b_dim[1] * v_mul >= 16))
    {
        for (int i = 0, x = 0; x < w4 && i < imin(b_dim[2], 4); ) {
            // only odd blocks are considered for overlap handling, hence +1
            const refmvs *const a_r = &r[x - f->b4_stride + 1];
            const uint8_t *const a_b_dim =
572
                dav1d_block_dimensions[sbtype_to_bs[a_r->sb_type]];
573 574

            if (a_r->ref[0] > 0) {
575 576
                const int ow4 = iclip(a_b_dim[0], 2, b_dim[0]);
                const int oh4 = imin(b_dim[1], 16) >> 1;
577 578 579 580 581
                res = mc(t, lap, NULL, ow4 * h_mul * sizeof(pixel), ow4, oh4,
                         t->bx + x, t->by, pl, a_r->mv[0],
                         &f->refp[a_r->ref[0] - 1],
                         dav1d_filter_2d[t->a->filter[1][bx4 + x + 1]][t->a->filter[0][bx4 + x + 1]]);
                if (res) return res;
Henrik Gramner's avatar
Henrik Gramner committed
582 583
                f->dsp->mc.blend_h(&dst[x * h_mul], dst_stride, lap,
                                   h_mul * ow4, v_mul * oh4);
584 585 586 587 588 589 590 591 592 593 594
                i++;
            }
            x += imax(a_b_dim[0], 2);
        }
    }

    if (t->bx > t->ts->tiling.col_start)
        for (int i = 0, y = 0; y < h4 && i < imin(b_dim[3], 4); ) {
            // only odd blocks are considered for overlap handling, hence +1
            const refmvs *const l_r = &r[(y + 1) * f->b4_stride - 1];
            const uint8_t *const l_b_dim =
595
                dav1d_block_dimensions[sbtype_to_bs[l_r->sb_type]];
596 597

            if (l_r->ref[0] > 0) {
598 599
                const int ow4 = imin(b_dim[0], 16) >> 1;
                const int oh4 = iclip(l_b_dim[1], 2, b_dim[1]);
600 601 602 603 604
                res = mc(t, lap, NULL, h_mul * ow4 * sizeof(pixel), ow4, oh4,
                         t->bx, t->by + y, pl, l_r->mv[0],
                         &f->refp[l_r->ref[0] - 1],
                         dav1d_filter_2d[t->l.filter[1][by4 + y + 1]][t->l.filter[0][by4 + y + 1]]);
                if (res) return res;
Henrik Gramner's avatar
Henrik Gramner committed
605 606
                f->dsp->mc.blend_v(&dst[y * v_mul * PXSTRIDE(dst_stride)],
                                   dst_stride, lap, h_mul * ow4, v_mul * oh4);
607 608 609 610
                i++;
            }
            y += imax(l_b_dim[1], 2);
        }
611
    return 0;
612 613
}

614 615 616 617 618
static int warp_affine(Dav1dTileContext *const t,
                       pixel *dst8, coef *dst16, const ptrdiff_t dstride,
                       const uint8_t *const b_dim, const int pl,
                       const Dav1dThreadPicture *const refp,
                       const WarpedMotionParams *const wmp)
619 620 621 622 623 624 625 626 627
{
    assert((dst8 != NULL) ^ (dst16 != NULL));
    const Dav1dFrameContext *const f = t->f;
    const Dav1dDSPContext *const dsp = f->dsp;
    const int ss_ver = !!pl && f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I420;
    const int ss_hor = !!pl && f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I444;
    const int h_mul = 4 >> ss_hor, v_mul = 4 >> ss_ver;
    assert(!((b_dim[0] * h_mul) & 7) && !((b_dim[1] * v_mul) & 7));
    const int32_t *const mat = wmp->matrix;
628 629
    const int width = (f->cur.p.p.w + ss_hor) >> ss_hor;
    const int height = (f->cur.p.p.h + ss_ver) >> ss_ver;
630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649

    for (int y = 0; y < b_dim[1] * v_mul; y += 8) {
        for (int x = 0; x < b_dim[0] * h_mul; x += 8) {
            // calculate transformation relative to center of 8x8 block in
            // luma pixel units
            const int src_x = t->bx * 4 + ((x + 4) << ss_hor);
            const int src_y = t->by * 4 + ((y + 4) << ss_ver);
            const int mvx = (mat[2] * src_x + mat[3] * src_y + mat[0]) >> ss_hor;
            const int mvy = (mat[4] * src_x + mat[5] * src_y + mat[1]) >> ss_ver;

            const int dx = (mvx >> 16) - 4;
            const int mx = ((mvx & 0xffff) - wmp->alpha * 4 -
                                             wmp->beta  * 7) & ~0x3f;
            const int dy = (mvy >> 16) - 4;
            const int my = ((mvy & 0xffff) - wmp->gamma * 4 -
                                             wmp->delta * 4) & ~0x3f;

            const pixel *ref_ptr;
            ptrdiff_t ref_stride = refp->p.stride[!!pl];

650 651 652 653 654
            if (dav1d_thread_picture_wait(refp, dy + 4 + 8,
                                          PLANE_TYPE_Y + !!pl))
            {
                return -1;
            }
655
            if (dx < 3 || dx + 8 + 4 > width || dy < 3 || dy + 8 + 4 > height) {
656
                f->dsp->mc.emu_edge(15, 15, width, height, dx - 3, dy - 3,
657
                                    t->emu_edge, 192 * sizeof(pixel),
658
                                    refp->p.data[pl], ref_stride);
659 660
                ref_ptr = &t->emu_edge[192 * 3 + 3];
                ref_stride = 192 * sizeof(pixel);
661 662 663 664 665 666 667 668 669 670 671 672 673
            } else {
                ref_ptr = ((pixel *) refp->p.data[pl]) + PXSTRIDE(ref_stride) * dy + dx;
            }
            if (dst16 != NULL)
                dsp->mc.warp8x8t(&dst16[x], dstride, ref_ptr, ref_stride,
                                 wmp->abcd, mx, my);
            else
                dsp->mc.warp8x8(&dst8[x], dstride, ref_ptr, ref_stride,
                                wmp->abcd, mx, my);
        }
        if (dst8) dst8  += 8 * PXSTRIDE(dstride);
        else      dst16 += 8 * dstride;
    }
674
    return 0;
675 676
}

677 678 679
void bytefn(dav1d_recon_b_intra)(Dav1dTileContext *const t, const enum BlockSize bs,
                                 const enum EdgeFlags intra_edge_flags,
                                 const Av1Block *const b)
680 681 682 683 684 685 686 687
{
    Dav1dTileState *const ts = t->ts;
    const Dav1dFrameContext *const f = t->f;
    const Dav1dDSPContext *const dsp = f->dsp;
    const int bx4 = t->bx & 31, by4 = t->by & 31;
    const int ss_ver = f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I420;
    const int ss_hor = f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I444;
    const int cbx4 = bx4 >> ss_hor, cby4 = by4 >> ss_ver;
688
    const uint8_t *const b_dim = dav1d_block_dimensions[bs];
689 690 691 692 693 694
    const int bw4 = b_dim[0], bh4 = b_dim[1];
    const int w4 = imin(bw4, f->bw - t->bx), h4 = imin(bh4, f->bh - t->by);
    const int cw4 = (w4 + ss_hor) >> ss_hor, ch4 = (h4 + ss_ver) >> ss_ver;
    const int has_chroma = f->seq_hdr.layout != DAV1D_PIXEL_LAYOUT_I400 &&
                           (bw4 > ss_hor || t->bx & 1) &&
                           (bh4 > ss_ver || t->by & 1);
695 696
    const TxfmInfo *const t_dim = &dav1d_txfm_dimensions[b->tx];
    const TxfmInfo *const uv_t_dim = &dav1d_txfm_dimensions[b->uvtx];
697 698

    // coefficient coding
699 700
    ALIGN_STK_32(pixel, edge_buf, 257,);
    pixel *const edge = edge_buf + 128;
701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756
    const int cbw4 = (bw4 + ss_hor) >> ss_hor, cbh4 = (bh4 + ss_ver) >> ss_ver;

    for (int init_y = 0; init_y < h4; init_y += 16) {
        for (int init_x = 0; init_x < w4; init_x += 16) {
            if (b->pal_sz[0]) {
                pixel *dst = ((pixel *) f->cur.p.data[0]) +
                             4 * (t->by * PXSTRIDE(f->cur.p.stride[0]) + t->bx);
                const uint8_t *pal_idx;
                if (f->frame_thread.pass) {
                    pal_idx = ts->frame_thread.pal_idx;
                    ts->frame_thread.pal_idx += bw4 * bh4 * 16;
                } else {
                    pal_idx = t->scratch.pal_idx;
                }
                const uint16_t *const pal = f->frame_thread.pass ?
                    f->frame_thread.pal[((t->by >> 1) + (t->bx & 1)) * (f->b4_stride >> 1) +
                                        ((t->bx >> 1) + (t->by & 1))][0] : t->pal[0];
                f->dsp->ipred.pal_pred(dst, f->cur.p.stride[0], pal,
                                       pal_idx, bw4 * 4, bh4 * 4);
                if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS)
                    hex_dump(dst, PXSTRIDE(f->cur.p.stride[0]),
                             bw4 * 4, bh4 * 4, "y-pal-pred");
            }

            const int sm_fl = sm_flag(t->a, bx4) | sm_flag(&t->l, by4);
            const int sb_has_tr = init_x + 16 < w4 ? 1 : init_y ? 0 :
                              intra_edge_flags & EDGE_I444_TOP_HAS_RIGHT;
            const int sb_has_bl = init_x ? 0 : init_y + 16 < h4 ? 1 :
                              intra_edge_flags & EDGE_I444_LEFT_HAS_BOTTOM;
            int y, x;
            const int sub_h4 = imin(h4, 16 + init_y);
            const int sub_w4 = imin(w4, init_x + 16);
            for (y = init_y, t->by += init_y; y < sub_h4;
                 y += t_dim->h, t->by += t_dim->h)
            {
                pixel *dst = ((pixel *) f->cur.p.data[0]) +
                               4 * (t->by * PXSTRIDE(f->cur.p.stride[0]) +
                                    t->bx + init_x);
                for (x = init_x, t->bx += init_x; x < sub_w4;
                     x += t_dim->w, t->bx += t_dim->w)
                {
                    if (b->pal_sz[0]) goto skip_y_pred;

                    int angle = b->y_angle;
                    const enum EdgeFlags edge_flags =
                        (((y > init_y || !sb_has_tr) && (x + t_dim->w >= sub_w4)) ?
                             0 : EDGE_I444_TOP_HAS_RIGHT) |
                        ((x > init_x || (!sb_has_bl && y + t_dim->h >= sub_h4)) ?
                             0 : EDGE_I444_LEFT_HAS_BOTTOM);
                    const pixel *top_sb_edge = NULL;
                    if (!(t->by & (f->sb_step - 1))) {
                        top_sb_edge = f->ipred_edge[0];
                        const int sby = t->by >> f->sb_shift;
                        top_sb_edge += f->sb128w * 128 * (sby - 1);
                    }
                    const enum IntraPredMode m =
757 758 759 760 761 762 763 764 765 766
                        bytefn(dav1d_prepare_intra_edges)(t->bx,
                                                          t->bx > ts->tiling.col_start,
                                                          t->by,
                                                          t->by > ts->tiling.row_start,
                                                          ts->tiling.col_end,
                                                          ts->tiling.row_end,
                                                          edge_flags, dst,
                                                          f->cur.p.stride[0], top_sb_edge,
                                                          b->y_mode, &angle,
                                                          t_dim->w, t_dim->h, edge);
767 768 769
                    dsp->ipred.intra_pred[m](dst, f->cur.p.stride[0], edge,
                                             t_dim->w * 4, t_dim->h * 4,
                                             angle | sm_fl);
770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801

                    if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS) {
                        hex_dump(edge - t_dim->h * 4, t_dim->h * 4,
                                 t_dim->h * 4, 2, "l");
                        hex_dump(edge, 0, 1, 1, "tl");
                        hex_dump(edge + 1, t_dim->w * 4,
                                 t_dim->w * 4, 2, "t");
                        hex_dump(dst, f->cur.p.stride[0],
                                 t_dim->w * 4, t_dim->h * 4, "y-intra-pred");
                    }

                skip_y_pred: {}
                    if (!b->skip) {
                        coef *cf;
                        int eob;
                        enum TxfmType txtp;
                        if (f->frame_thread.pass) {
                            cf = ts->frame_thread.cf;
                            ts->frame_thread.cf += imin(t_dim->w, 8) * imin(t_dim->h, 8) * 16;
                            const struct CodedBlockInfo *const cbi =
                                &f->frame_thread.cbi[t->by * f->b4_stride + t->bx];
                            eob = cbi->eob[0];
                            txtp = cbi->txtp[0];
                        } else {
                            uint8_t cf_ctx;
                            cf = t->cf;
                            eob = decode_coefs(t, &t->a->lcoef[bx4 + x],
                                               &t->l.lcoef[by4 + y], b->tx, bs,
                                               b, 1, 0, cf, &txtp, &cf_ctx);
                            if (DEBUG_BLOCK_INFO)
                                printf("Post-y-cf-blk[tx=%d,txtp=%d,eob=%d]: r=%d\n",
                                       b->tx, txtp, eob, ts->msac.rng);
Ronald S. Bultje's avatar
Ronald S. Bultje committed
802 803 804 805 806 807 808 809 810 811
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
                            rep_macro(type, t->dir lcoef, off, mul * cf_ctx)
#define default_memset(dir, diridx, off, sz) \
                            memset(&t->dir lcoef[off], cf_ctx, sz)
                            case_set_upto16_with_default(imin(t_dim->h, f->bh - t->by), \
                                                         l., 1, by4 + y);
                            case_set_upto16_with_default(imin(t_dim->w, f->bw - t->bx), \
                                                         a->, 0, bx4 + x);
#undef default_memset
#undef set_ctx
812 813 814 815 816 817 818 819 820 821 822 823 824 825
                        }
                        if (eob >= 0) {
                            if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS)
                                coef_dump(cf, imin(t_dim->h, 8) * 4,
                                          imin(t_dim->w, 8) * 4, 3, "dq");
                            dsp->itx.itxfm_add[b->tx]
                                              [txtp](dst,
                                                     f->cur.p.stride[0],
                                                     cf, eob);
                            if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS)
                                hex_dump(dst, f->cur.p.stride[0],
                                         t_dim->w * 4, t_dim->h * 4, "recon");
                        }
                    } else if (!f->frame_thread.pass) {
Ronald S. Bultje's avatar
Ronald S. Bultje committed
826 827 828 829 830
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
                        rep_macro(type, t->dir lcoef, off, mul * 0x40)
                        case_set_upto16(t_dim->h, l., 1, by4 + y);
                        case_set_upto16(t_dim->w, a->, 0, bx4 + x);
#undef set_ctx
831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852
                    }
                    dst += 4 * t_dim->w;
                }
                t->bx -= x;
            }
            t->by -= y;

            if (!has_chroma) continue;

            const ptrdiff_t stride = f->cur.p.stride[1];

            if (b->uv_mode == CFL_PRED) {
                assert(!init_x && !init_y);

                int16_t *const ac = t->scratch.ac;
                pixel *y_src = ((pixel *) f->cur.p.data[0]) + 4 * (t->bx & ~ss_hor) +
                                 4 * (t->by & ~ss_ver) * PXSTRIDE(f->cur.p.stride[0]);
                const ptrdiff_t uv_off = 4 * ((t->bx >> ss_hor) +
                                              (t->by >> ss_ver) * PXSTRIDE(stride));
                pixel *const uv_dst[2] = { ((pixel *) f->cur.p.data[1]) + uv_off,
                                           ((pixel *) f->cur.p.data[2]) + uv_off };

853 854 855 856 857
                const int furthest_r =
                    ((cw4 << ss_hor) + t_dim->w - 1) & ~(t_dim->w - 1);
                const int furthest_b =
                    ((ch4 << ss_ver) + t_dim->h - 1) & ~(t_dim->h - 1);
                dsp->ipred.cfl_ac[f->cur.p.p.layout - 1]
858 859 860
                                 [b->uvtx](ac, y_src, f->cur.p.stride[0],
                                           cbw4 - (furthest_r >> ss_hor),
                                           cbh4 - (furthest_b >> ss_ver));
861
                for (int pl = 0; pl < 2; pl++) {
862
                    if (!b->cfl_alpha[pl]) continue;
863 864 865 866 867 868 869
                    int angle = 0;
                    const pixel *top_sb_edge = NULL;
                    if (!((t->by & ~ss_ver) & (f->sb_step - 1))) {
                        top_sb_edge = f->ipred_edge[pl + 1];
                        const int sby = t->by >> f->sb_shift;
                        top_sb_edge += f->sb128w * 128 * (sby - 1);
                    }
870 871 872
                    const int xpos = t->bx >> ss_hor, ypos = t->by >> ss_ver;
                    const int xstart = ts->tiling.col_start >> ss_hor;
                    const int ystart = ts->tiling.row_start >> ss_ver;
873
                    const enum IntraPredMode m =
874 875
                        bytefn(dav1d_prepare_intra_edges)(xpos, xpos > xstart,
                                                          ypos, ypos > ystart,
876 877 878 879
                                                          ts->tiling.col_end >> ss_hor,
                                                          ts->tiling.row_end >> ss_ver,
                                                          0, uv_dst[pl], stride,
                                                          top_sb_edge, DC_PRED, &angle,
880 881
                                                          uv_t_dim->w,
                                                          uv_t_dim->h, edge);
882 883 884
                    dsp->ipred.cfl_pred[m](uv_dst[pl], stride, edge,
                                           uv_t_dim->w * 4,
                                           uv_t_dim->h * 4,
885
                                           ac, b->cfl_alpha[pl]);
886
                }
887
                if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS) {
888
                    ac_dump(ac, 4*cbw4, 4*cbh4, "ac");
889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943
                    hex_dump(uv_dst[0], stride, cbw4 * 4, cbh4 * 4, "u-cfl-pred");
                    hex_dump(uv_dst[1], stride, cbw4 * 4, cbh4 * 4, "v-cfl-pred");
                }
            } else if (b->pal_sz[1]) {
                ptrdiff_t uv_dstoff = 4 * ((t->bx >> ss_hor) +
                                           (t->by >> ss_ver) * PXSTRIDE(f->cur.p.stride[1]));
                const uint8_t *pal_idx;
                if (f->frame_thread.pass) {
                    pal_idx = ts->frame_thread.pal_idx;
                    ts->frame_thread.pal_idx += cbw4 * cbh4 * 16;
                } else {
                    pal_idx = &t->scratch.pal_idx[bw4 * bh4 * 16];
                }
                const uint16_t *const pal_u = f->frame_thread.pass ?
                    f->frame_thread.pal[((t->by >> 1) + (t->bx & 1)) * (f->b4_stride >> 1) +
                                        ((t->bx >> 1) + (t->by & 1))][1] : t->pal[1];
                f->dsp->ipred.pal_pred(((pixel *) f->cur.p.data[1]) + uv_dstoff,
                                       f->cur.p.stride[1], pal_u,
                                       pal_idx, cbw4 * 4, cbh4 * 4);
                const uint16_t *const pal_v = f->frame_thread.pass ?
                    f->frame_thread.pal[((t->by >> 1) + (t->bx & 1)) * (f->b4_stride >> 1) +
                                        ((t->bx >> 1) + (t->by & 1))][2] : t->pal[2];
                f->dsp->ipred.pal_pred(((pixel *) f->cur.p.data[2]) + uv_dstoff,
                                       f->cur.p.stride[1], pal_v,
                                       pal_idx, cbw4 * 4, cbh4 * 4);
                if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS) {
                    hex_dump(((pixel *) f->cur.p.data[1]) + uv_dstoff,
                             PXSTRIDE(f->cur.p.stride[1]),
                             cbw4 * 4, cbh4 * 4, "u-pal-pred");
                    hex_dump(((pixel *) f->cur.p.data[2]) + uv_dstoff,
                             PXSTRIDE(f->cur.p.stride[1]),
                             cbw4 * 4, cbh4 * 4, "v-pal-pred");
                }
            }

            const int sm_uv_fl = sm_uv_flag(t->a, cbx4) |
                                 sm_uv_flag(&t->l, cby4);
            const int uv_sb_has_tr =
                ((init_x + 16) >> ss_hor) < cw4 ? 1 : init_y ? 0 :
                intra_edge_flags & (EDGE_I420_TOP_HAS_RIGHT >> (f->cur.p.p.layout - 1));
            const int uv_sb_has_bl =
                init_x ? 0 : ((init_y + 16) >> ss_ver) < ch4 ? 1 :
                intra_edge_flags & (EDGE_I420_LEFT_HAS_BOTTOM >> (f->cur.p.p.layout - 1));
            const int sub_ch4 = imin(ch4, (init_y + 16) >> ss_ver);
            const int sub_cw4 = imin(cw4, (init_x + 16) >> ss_hor);
            for (int pl = 0; pl < 2; pl++) {
                for (y = init_y >> ss_ver, t->by += init_y; y < sub_ch4;
                     y += uv_t_dim->h, t->by += uv_t_dim->h << ss_ver)
                {
                    pixel *dst = ((pixel *) f->cur.p.data[1 + pl]) +
                                   4 * ((t->by >> ss_ver) * PXSTRIDE(stride) +
                                        ((t->bx + init_x) >> ss_hor));
                    for (x = init_x >> ss_hor, t->bx += init_x; x < sub_cw4;
                         x += uv_t_dim->w, t->bx += uv_t_dim->w << ss_hor)
                    {
944 945 946
                        if ((b->uv_mode == CFL_PRED && b->cfl_alpha[pl]) ||
                            b->pal_sz[1])
                        {
947
                            goto skip_uv_pred;
948
                        }
949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966

                        int angle = b->uv_angle;
                        // this probably looks weird because we're using
                        // luma flags in a chroma loop, but that's because
                        // prepare_intra_edges() expects luma flags as input
                        const enum EdgeFlags edge_flags =
                            (((y > (init_y >> ss_ver) || !uv_sb_has_tr) &&
                              (x + uv_t_dim->w >= sub_cw4)) ?
                                 0 : EDGE_I444_TOP_HAS_RIGHT) |
                            ((x > (init_x >> ss_hor) ||
                              (!uv_sb_has_bl && y + uv_t_dim->h >= sub_ch4)) ?
                                 0 : EDGE_I444_LEFT_HAS_BOTTOM);
                        const pixel *top_sb_edge = NULL;
                        if (!((t->by & ~ss_ver) & (f->sb_step - 1))) {
                            top_sb_edge = f->ipred_edge[1 + pl];
                            const int sby = t->by >> f->sb_shift;
                            top_sb_edge += f->sb128w * 128 * (sby - 1);
                        }
967 968
                        const enum IntraPredMode uv_mode =
                             b->uv_mode == CFL_PRED ? DC_PRED : b->uv_mode;
969 970 971
                        const int xpos = t->bx >> ss_hor, ypos = t->by >> ss_ver;
                        const int xstart = ts->tiling.col_start >> ss_hor;
                        const int ystart = ts->tiling.row_start >> ss_ver;
972
                        const enum IntraPredMode m =
973 974
                            bytefn(dav1d_prepare_intra_edges)(xpos, xpos > xstart,
                                                              ypos, ypos > ystart,
975 976 977
                                                              ts->tiling.col_end >> ss_hor,
                                                              ts->tiling.row_end >> ss_ver,
                                                              edge_flags, dst, stride,
978
                                                              top_sb_edge, uv_mode,
979 980
                                                              &angle, uv_t_dim->w,
                                                              uv_t_dim->h, edge);
981 982 983 984
                        dsp->ipred.intra_pred[m](dst, stride, edge,
                                                 uv_t_dim->w * 4,
                                                 uv_t_dim->h * 4,
                                                 angle | sm_uv_fl);
985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017
                        if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS) {
                            hex_dump(edge - uv_t_dim->h * 4, uv_t_dim->h * 4,
                                     uv_t_dim->h * 4, 2, "l");
                            hex_dump(edge, 0, 1, 1, "tl");
                            hex_dump(edge + 1, uv_t_dim->w * 4,
                                     uv_t_dim->w * 4, 2, "t");
                            hex_dump(dst, stride, uv_t_dim->w * 4,
                                     uv_t_dim->h * 4, pl ? "v-intra-pred" : "u-intra-pred");
                        }

                    skip_uv_pred: {}
                        if (!b->skip) {
                            enum TxfmType txtp;
                            int eob;
                            coef *cf;
                            if (f->frame_thread.pass) {
                                cf = ts->frame_thread.cf;
                                ts->frame_thread.cf += uv_t_dim->w * uv_t_dim->h * 16;
                                const struct CodedBlockInfo *const cbi =
                                    &f->frame_thread.cbi[t->by * f->b4_stride + t->bx];
                                eob = cbi->eob[pl + 1];
                                txtp = cbi->txtp[pl + 1];
                            } else {
                                uint8_t cf_ctx;
                                cf = t->cf;
                                eob = decode_coefs(t, &t->a->ccoef[pl][cbx4 + x],
                                                   &t->l.ccoef[pl][cby4 + y],
                                                   b->uvtx, bs, b, 1, 1 + pl, cf,
                                                   &txtp, &cf_ctx);
                                if (DEBUG_BLOCK_INFO)
                                    printf("Post-uv-cf-blk[pl=%d,tx=%d,"
                                           "txtp=%d,eob=%d]: r=%d [x=%d,cbx4=%d]\n",
                                           pl, b->uvtx, txtp, eob, ts->msac.rng, x, cbx4);
Ronald S. Bultje's avatar
Ronald S. Bultje committed
1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
                                rep_macro(type, t->dir ccoef[pl], off, mul * cf_ctx)
#define default_memset(dir, diridx, off, sz) \
                                memset(&t->dir ccoef[pl][off], cf_ctx, sz)
                                case_set_upto16_with_default( \
                                         imin(uv_t_dim->h, (f->bh - t->by + ss_ver) >> ss_ver),
                                         l., 1, cby4 + y);
                                case_set_upto16_with_default( \
                                         imin(uv_t_dim->w, (f->bw - t->bx + ss_hor) >> ss_hor),
                                         a->, 0, cbx4 + x);
#undef default_memset
#undef set_ctx
1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042
                            }
                            if (eob >= 0) {
                                if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS)
                                    coef_dump(cf, uv_t_dim->h * 4,
                                              uv_t_dim->w * 4, 3, "dq");
                                dsp->itx.itxfm_add[b->uvtx]
                                                  [txtp](dst, stride,
                                                         cf, eob);
                                if (DEBUG_BLOCK_INFO && DEBUG_B_PIXELS)
                                    hex_dump(dst, stride, uv_t_dim->w * 4,
                                             uv_t_dim->h * 4, "recon");
                            }
                        } else if (!f->frame_thread.pass) {
Ronald S. Bultje's avatar
Ronald S. Bultje committed
1043 1044 1045 1046 1047
#define set_ctx(type, dir, diridx, off, mul, rep_macro) \
                            rep_macro(type, t->dir ccoef[pl], off, mul * 0x40)
                            case_set_upto16(uv_t_dim->h, l., 1, cby4 + y);
                            case_set_upto16(uv_t_dim->w, a->, 0, cbx4 + x);
#undef set_ctx
1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058
                        }
                        dst += uv_t_dim->w * 4;
                    }
                    t->bx -= x << ss_hor;
                }
                t->by -= y << ss_ver;
            }
        }
    }
}

1059
int bytefn(dav1d_recon_b_inter)(Dav1dTileContext *const t, const enum BlockSize bs,
1060
                                 const Av1Block *const b)
1061 1062 1063 1064 1065 1066 1067 1068
{
    Dav1dTileState *const ts = t->ts;
    const Dav1dFrameContext *const f = t->f;
    const Dav1dDSPContext *const dsp = f->dsp;
    const int bx4 = t->bx & 31, by4 = t->by & 31;
    const int ss_ver = f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I420;
    const int ss_hor = f->cur.p.p.layout != DAV1D_PIXEL_LAYOUT_I444;
    const int cbx4 = bx4 >> ss_hor, cby4 = by4 >> ss_ver;
1069
    const uint8_t *const b_dim = dav1d_block_dimensions[bs];
1070 1071 1072 1073 1074 1075 1076
    const int bw4 = b_dim[0], bh4 = b_dim[1];
    const int w4 = imin(bw4, f->bw - t->bx), h4 = imin(bh4, f->bh - t->by);
    const int has_chroma = f->seq_hdr.layout != DAV1D_PIXEL_LAYOUT_I400 &&
                           (bw4 > ss_hor || t->bx & 1) &&
                           (bh4 > ss_ver || t->by & 1);
    const int chr_layout_idx = f->cur.p.p.layout == DAV1D_PIXEL_LAYOUT_I400 ? 0 :
                               DAV1D_PIXEL_LAYOUT_I444 - f->cur.p.p.layout;
1077
    int res;
1078 1079 1080 1081 1082 1083 1084 1085 1086

    // prediction
    const int cbh4 = (bh4 + ss_ver) >> ss_ver, cbw4 = (bw4 + ss_hor) >> ss_hor;
    pixel *dst = ((pixel *) f->cur.p.data[0]) +
        4 * (t->by * PXSTRIDE(f->cur.p.stride[0]) + t->bx);
    const ptrdiff_t uvdstoff =
        4 * ((t->bx >> ss_hor) + (t->by >> ss_ver) * PXSTRIDE(f->cur.p.stride[1]));
    if (!(f->frame_hdr.frame_type & 1)) {
        // intrabc
1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
        res = mc(t, dst, NULL, f->cur.p.stride[0],
                 bw4, bh4, t->bx, t->by, 0, b->mv[0], &f->cur, FILTER_2D_BILINEAR);
        if (res) return res;
        if (has_chroma) for (int pl = 1; pl < 3; pl++) {
            res = mc(t, ((pixel *)f->cur.p.data[pl]) + uvdstoff, NULL, f->cur.p.stride[1],
                     bw4 << (bw4 == ss_hor), bh4 << (bh4 == ss_ver),
                     t->bx & ~ss_hor, t->by & ~ss_ver,
                     pl, b->mv[0], &f->cur, FILTER_2D_BILINEAR);
            if (res) return res;
        }
1097 1098 1099 1100
    } else if (b->comp_type == COMP_INTER_NONE) {
        const Dav1dThreadPicture *const refp = &f->refp[b->ref[0]];
        const enum Filter2d filter_2d = b->filter2d;

1101
        if (imin(bw4, bh4) > 1 && !f->frame_hdr.force_integer_mv &&
1102 1103 1104 1105 1106
            ((b->inter_mode == GLOBALMV &&
              f->frame_hdr.gmv[b->ref[0]].type > WM_TYPE_TRANSLATION) ||
             (b->motion_mode == MM_WARP &&
              t->warpmv.type > WM_TYPE_TRANSLATION)))
        {
1107 1108 1109 1110
            res = warp_affine(t, dst, NULL, f->cur.p.stride[0], b_dim, 0, refp,
                              b->motion_mode == MM_WARP ? &t->warpmv :
                                  &f->frame_hdr.gmv[b->ref[0]]);
            if (res) return res;
1111
        } else {
1112 1113 1114 1115 1116 1117 1118
            res = mc(t, dst, NULL, f->cur.p.stride[0],
                     bw4, bh4, t->bx, t->by, 0, b->mv[0], refp, filter_2d);
            if (res) return res;
            if (b->motion_mode == MM_OBMC) {
                res = obmc(t, dst, f->cur.p.stride[0], b_dim, 0, bx4, by4, w4, h4);
                if (res) return res;
            }
1119 1120
        }
        if (b->interintra_type) {
1121 1122
            ALIGN_STK_32(pixel, tl_edge_buf, 65,);
            pixel *const tl_edge = tl_edge_buf + 32;
1123 1124 1125 1126 1127 1128 1129 1130 1131 1132
            enum IntraPredMode m = b->interintra_mode == II_SMOOTH_PRED ?
                                   SMOOTH_PRED : b->interintra_mode;
            pixel *const tmp = t->scratch.interintra;
            int angle = 0;
            const pixel *top_sb_edge = NULL;
            if (!(t->by & (f->sb_step - 1))) {
                top_sb_edge = f->ipred_edge[0];
                const int sby = t->by >> f->sb_shift;
                top_sb_edge += f->sb128w * 128 * (sby - 1);
            }
1133 1134 1135 1136 1137
            m = bytefn(dav1d_prepare_intra_edges)(t->bx, t->bx > ts->tiling.col_start,
                                                  t->by, t->by > ts->tiling.row_start,
                                                  ts->tiling.col_end, ts->tiling.row_end,
                                                  0, dst, f->cur.p.stride[0], top_sb_edge,
                                                  m, &angle, bw4, bh4, tl_edge);
1138 1139
            dsp->ipred.intra_pred[m](tmp, 4 * bw4 * sizeof(pixel),
                                     tl_edge, bw4 * 4, bh4 * 4, 0);
1140 1141
            const uint8_t *const ii_mask =
                b->interintra_type == INTER_INTRA_BLEND ?
1142 1143
                     dav1d_ii_masks[bs][0][b->interintra_mode] :
                     dav1d_wedge_masks[bs][0][0][b->wedge_idx];
1144
            dsp->mc.blend(dst, f->cur.p.stride[0], tmp,
Henrik Gramner's avatar
Henrik Gramner committed
1145
                          bw4 * 4, bh4 * 4, ii_mask);
1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166
        }

        if (!has_chroma) goto skip_inter_chroma_pred;

        // sub8x8 derivation
        int is_sub8x8 = bw4 == ss_hor || bh4 == ss_ver;
        refmvs *r;
        if (is_sub8x8) {
            assert(ss_hor == 1);
            r = &f->mvs[t->by * f->b4_stride + t->bx];
            if (bw4 == 1) is_sub8x8 &= r[-1].ref[0] > 0;
            if (bh4 == ss_ver) is_sub8x8 &= r[-f->b4_stride].ref[0] > 0;
            if (bw4 == 1 && bh4 == ss_ver)
                is_sub8x8 &= r[-(1 + f->b4_stride)].ref[0] > 0;
        }

        // chroma prediction
        if (is_sub8x8) {
            assert(ss_hor == 1);
            int h_off = 0, v_off = 0;
            if (bw4 == 1 && bh4 == ss_ver) {
1167 1168 1169 1170 1171 1172 1173 1174 1175 1176
                for (int pl = 0; pl < 2; pl++) {
                    res = mc(t, ((pixel *) f->cur.p.data[1 + pl]) + uvdstoff,
                             NULL, f->cur.p.stride[1],
                             bw4, bh4, t->bx - 1, t->by - 1, 1 + pl,
                             r[-(f->b4_stride + 1)].mv[0],
                             &f->refp[r[-(f->b4_stride + 1)].ref[0] - 1],
                             f->frame_thread.pass != 2 ? t->tl_4x4_filter :
                                 f->frame_thread.b[((t->by - 1) * f->b4_stride) + t->bx - 1].filter2d);
                    if (res) return res;
                }
1177 1178 1179 1180 1181
                v_off = 2 * PXSTRIDE(f->cur.p.stride[1]);
                h_off = 2;
            }
            if (bw4 == 1) {
                const enum Filter2d left_filter_2d =
1182
                    dav1d_filter_2d[t->l.filter[1][by4]][t->l.filter[0][by4]];
1183 1184 1185 1186 1187 1188 1189 1190
                for (int pl = 0; pl < 2; pl++) {
                    res = mc(t, ((pixel *) f->cur.p.data[1 + pl]) + uvdstoff + v_off, NULL,
                             f->cur.p.stride[1], bw4, bh4, t->bx - 1,
                             t->by, 1 + pl, r[-1].mv[0], &f->refp[r[-1].ref[0] - 1],
                             f->frame_thread.pass != 2 ? left_filter_2d :
                                 f->frame_thread.b[(t->by * f->b4_stride) + t->bx - 1].filter2d);
                    if (res) return res;
                }
1191 1192 1193 1194
                h_off = 2;
            }
            if (bh4 == ss_ver) {
                const enum Filter2d top_filter_2d =
1195
                    dav1d_filter_2d[t->a->filter[1][bx4]][t->a->filter[0][bx4]];
1196 1197 1198 1199 1200 1201 1202 1203 1204
                for (int pl = 0; pl < 2; pl++) {
                    res = mc(t, ((pixel *) f->cur.p.data[1 + pl]) + uvdstoff + h_off, NULL,
                             f->cur.p.stride[1], bw4, bh4, t->bx, t->by - 1,
                             1 + pl, r[-f->b4_stride].mv[0],
                             &f->refp[r[-f->b4_stride].ref[0] - 1],
                             f->frame_thread.pass != 2 ? top_filter_2d :
                                 f->frame_thread.b[((t->by - 1) * f->b4_stride) + t->bx].filter2d);
                    if (res) return res;
                }
1205 1206
                v_off = 2 * PXSTRIDE(f->cur.p.stride[1]);
            }
1207 1208 1209 1210 1211
            for (int pl = 0; pl < 2; pl++) {
                res = mc(t, ((pixel *) f->cur.p.data[1 + pl]) + uvdstoff + h_off + v_off, NULL, f->cur.p.stride[1],
                         bw4, bh4, t->bx, t->by, 1 + pl, b->mv[0], refp, filter_2d);
                if (res) return res;
            }
1212
        } else {
1213
            if (imin(cbw4, cbh4) > 1 && !f->frame_hdr.force_integer_mv &&